xref: /openbmc/qemu/tcg/arm/tcg-target.c.inc (revision 4d137ff819bae33d045f13bb9186e3a2c71cb7e4)
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_ext8s(TCGContext *s, TCGType t, TCGReg rd, TCGReg rn)
925{
926    /* sxtb */
927    tcg_out32(s, 0x06af0070 | (COND_AL << 28) | (rd << 12) | rn);
928}
929
930static void tcg_out_ext8u(TCGContext *s, TCGReg rd, TCGReg rn)
931{
932    tcg_out_dat_imm(s, COND_AL, ARITH_AND, rd, rn, 0xff);
933}
934
935static void tcg_out_ext16s(TCGContext *s, TCGType t, TCGReg rd, TCGReg rn)
936{
937    /* sxth */
938    tcg_out32(s, 0x06bf0070 | (COND_AL << 28) | (rd << 12) | rn);
939}
940
941static void tcg_out_ext16u(TCGContext *s, TCGReg rd, TCGReg rn)
942{
943    /* uxth */
944    tcg_out32(s, 0x06ff0070 | (COND_AL << 28) | (rd << 12) | rn);
945}
946
947static void tcg_out_ext32s(TCGContext *s, TCGReg rd, TCGReg rn)
948{
949    g_assert_not_reached();
950}
951
952static void tcg_out_ext32u(TCGContext *s, TCGReg rd, TCGReg rn)
953{
954    g_assert_not_reached();
955}
956
957static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
958{
959    g_assert_not_reached();
960}
961
962static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
963{
964    g_assert_not_reached();
965}
966
967static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg rd, TCGReg rn)
968{
969    g_assert_not_reached();
970}
971
972static void tgen_deposit(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1,
973                         TCGReg a2, unsigned ofs, unsigned len)
974{
975    /* bfi/bfc */
976    tcg_out32(s, 0x07c00010 | (COND_AL << 28) | (a0 << 12) | a1
977              | (ofs << 7) | ((ofs + len - 1) << 16));
978}
979
980static void tgen_depositi(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1,
981                          tcg_target_long a2, unsigned ofs, unsigned len)
982{
983    /* bfi becomes bfc with rn == 15.  */
984    tgen_deposit(s, type, a0, a1, 15, ofs, len);
985}
986
987static const TCGOutOpDeposit outop_deposit = {
988    .base.static_constraint = C_O1_I2(r, 0, rZ),
989    .out_rrr = tgen_deposit,
990    .out_rri = tgen_depositi,
991};
992
993static void tgen_extract(TCGContext *s, TCGType type, TCGReg rd, TCGReg rn,
994                         unsigned ofs, unsigned len)
995{
996    /* According to gcc, AND can be faster. */
997    if (ofs == 0 && len <= 8) {
998        tcg_out_dat_imm(s, COND_AL, ARITH_AND, rd, rn,
999                        encode_imm_nofail((1 << len) - 1));
1000        return;
1001    }
1002
1003    if (use_armv7_instructions) {
1004        /* ubfx */
1005        tcg_out32(s, 0x07e00050 | (COND_AL << 28) | (rd << 12) | rn
1006                  | (ofs << 7) | ((len - 1) << 16));
1007        return;
1008    }
1009
1010    assert(ofs % 8 == 0);
1011    switch (len) {
1012    case 8:
1013        /* uxtb */
1014        tcg_out32(s, 0x06ef0070 | (COND_AL << 28) |
1015                  (rd << 12) | (ofs << 7) | rn);
1016        break;
1017    case 16:
1018        /* uxth */
1019        tcg_out32(s, 0x06ff0070 | (COND_AL << 28) |
1020                  (rd << 12) | (ofs << 7) | rn);
1021        break;
1022    default:
1023        g_assert_not_reached();
1024    }
1025}
1026
1027static const TCGOutOpExtract outop_extract = {
1028    .base.static_constraint = C_O1_I1(r, r),
1029    .out_rr = tgen_extract,
1030};
1031
1032static void tgen_sextract(TCGContext *s, TCGType type, TCGReg rd, TCGReg rn,
1033                          unsigned ofs, unsigned len)
1034{
1035    if (use_armv7_instructions) {
1036        /* sbfx */
1037        tcg_out32(s, 0x07a00050 | (COND_AL << 28) | (rd << 12) | rn
1038                  | (ofs << 7) | ((len - 1) << 16));
1039        return;
1040    }
1041
1042    assert(ofs % 8 == 0);
1043    switch (len) {
1044    case 8:
1045        /* sxtb */
1046        tcg_out32(s, 0x06af0070 | (COND_AL << 28) |
1047                  (rd << 12) | (ofs << 7) | rn);
1048        break;
1049    case 16:
1050        /* sxth */
1051        tcg_out32(s, 0x06bf0070 | (COND_AL << 28) |
1052                  (rd << 12) | (ofs << 7) | rn);
1053        break;
1054    default:
1055        g_assert_not_reached();
1056    }
1057}
1058
1059static const TCGOutOpExtract outop_sextract = {
1060    .base.static_constraint = C_O1_I1(r, r),
1061    .out_rr = tgen_sextract,
1062};
1063
1064
1065static void tcg_out_ld32u(TCGContext *s, ARMCond cond,
1066                          TCGReg rd, TCGReg rn, int32_t offset)
1067{
1068    if (offset > 0xfff || offset < -0xfff) {
1069        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1070        tcg_out_ld32_r(s, cond, rd, rn, TCG_REG_TMP);
1071    } else
1072        tcg_out_ld32_12(s, cond, rd, rn, offset);
1073}
1074
1075static void tcg_out_st32(TCGContext *s, ARMCond cond,
1076                         TCGReg rd, TCGReg rn, int32_t offset)
1077{
1078    if (offset > 0xfff || offset < -0xfff) {
1079        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1080        tcg_out_st32_r(s, cond, rd, rn, TCG_REG_TMP);
1081    } else
1082        tcg_out_st32_12(s, cond, rd, rn, offset);
1083}
1084
1085static void tcg_out_ld16u(TCGContext *s, ARMCond cond,
1086                          TCGReg rd, TCGReg rn, int32_t offset)
1087{
1088    if (offset > 0xff || offset < -0xff) {
1089        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1090        tcg_out_ld16u_r(s, cond, rd, rn, TCG_REG_TMP);
1091    } else
1092        tcg_out_ld16u_8(s, cond, rd, rn, offset);
1093}
1094
1095static void tcg_out_ld16s(TCGContext *s, ARMCond cond,
1096                          TCGReg rd, TCGReg rn, int32_t offset)
1097{
1098    if (offset > 0xff || offset < -0xff) {
1099        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1100        tcg_out_ld16s_r(s, cond, rd, rn, TCG_REG_TMP);
1101    } else
1102        tcg_out_ld16s_8(s, cond, rd, rn, offset);
1103}
1104
1105static void tcg_out_st16(TCGContext *s, ARMCond cond,
1106                         TCGReg rd, TCGReg rn, int32_t offset)
1107{
1108    if (offset > 0xff || offset < -0xff) {
1109        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1110        tcg_out_st16_r(s, cond, rd, rn, TCG_REG_TMP);
1111    } else
1112        tcg_out_st16_8(s, cond, rd, rn, offset);
1113}
1114
1115static void tcg_out_ld8u(TCGContext *s, ARMCond cond,
1116                         TCGReg rd, TCGReg rn, int32_t offset)
1117{
1118    if (offset > 0xfff || offset < -0xfff) {
1119        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1120        tcg_out_ld8_r(s, cond, rd, rn, TCG_REG_TMP);
1121    } else
1122        tcg_out_ld8_12(s, cond, rd, rn, offset);
1123}
1124
1125static void tcg_out_ld8s(TCGContext *s, ARMCond cond,
1126                         TCGReg rd, TCGReg rn, int32_t offset)
1127{
1128    if (offset > 0xff || offset < -0xff) {
1129        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1130        tcg_out_ld8s_r(s, cond, rd, rn, TCG_REG_TMP);
1131    } else
1132        tcg_out_ld8s_8(s, cond, rd, rn, offset);
1133}
1134
1135static void tcg_out_st8(TCGContext *s, ARMCond cond,
1136                        TCGReg rd, TCGReg rn, int32_t offset)
1137{
1138    if (offset > 0xfff || offset < -0xfff) {
1139        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1140        tcg_out_st8_r(s, cond, rd, rn, TCG_REG_TMP);
1141    } else
1142        tcg_out_st8_12(s, cond, rd, rn, offset);
1143}
1144
1145/*
1146 * The _goto case is normally between TBs within the same code buffer, and
1147 * with the code buffer limited to 16MB we wouldn't need the long case.
1148 * But we also use it for the tail-call to the qemu_ld/st helpers, which does.
1149 */
1150static void tcg_out_goto(TCGContext *s, ARMCond cond, const tcg_insn_unit *addr)
1151{
1152    intptr_t addri = (intptr_t)addr;
1153    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1154    bool arm_mode = !(addri & 1);
1155
1156    if (arm_mode && disp - 8 < 0x01fffffd && disp - 8 > -0x01fffffd) {
1157        tcg_out_b_imm(s, cond, disp);
1158        return;
1159    }
1160
1161    /* LDR is interworking from v5t. */
1162    tcg_out_movi_pool(s, cond, TCG_REG_PC, addri);
1163}
1164
1165/*
1166 * The call case is mostly used for helpers - so it's not unreasonable
1167 * for them to be beyond branch range.
1168 */
1169static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *addr)
1170{
1171    intptr_t addri = (intptr_t)addr;
1172    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1173    bool arm_mode = !(addri & 1);
1174
1175    if (disp - 8 < 0x02000000 && disp - 8 >= -0x02000000) {
1176        if (arm_mode) {
1177            tcg_out_bl_imm(s, COND_AL, disp);
1178        } else {
1179            tcg_out_blx_imm(s, disp);
1180        }
1181        return;
1182    }
1183
1184    tcg_out_movi32(s, COND_AL, TCG_REG_TMP, addri);
1185    tcg_out_blx_reg(s, COND_AL, TCG_REG_TMP);
1186}
1187
1188static void tcg_out_call(TCGContext *s, const tcg_insn_unit *addr,
1189                         const TCGHelperInfo *info)
1190{
1191    tcg_out_call_int(s, addr);
1192}
1193
1194static void tcg_out_goto_label(TCGContext *s, ARMCond cond, TCGLabel *l)
1195{
1196    if (l->has_value) {
1197        tcg_out_goto(s, cond, l->u.value_ptr);
1198    } else {
1199        tcg_out_reloc(s, s->code_ptr, R_ARM_PC24, l, 0);
1200        tcg_out_b_imm(s, cond, 0);
1201    }
1202}
1203
1204static void tcg_out_mb(TCGContext *s, TCGArg a0)
1205{
1206    if (use_armv7_instructions) {
1207        tcg_out32(s, INSN_DMB_ISH);
1208    } else {
1209        tcg_out32(s, INSN_DMB_MCR);
1210    }
1211}
1212
1213static TCGCond tgen_cmp(TCGContext *s, TCGCond cond, TCGReg a, TCGReg b)
1214{
1215    if (is_tst_cond(cond)) {
1216        tcg_out_dat_reg(s, COND_AL, ARITH_TST, 0, a, b, SHIFT_IMM_LSL(0));
1217        return tcg_tst_eqne_cond(cond);
1218    }
1219    tcg_out_dat_reg(s, COND_AL, ARITH_CMP, 0, a, b, SHIFT_IMM_LSL(0));
1220    return cond;
1221}
1222
1223static TCGCond tgen_cmpi(TCGContext *s, TCGCond cond, TCGReg a, TCGArg b)
1224{
1225    int imm12;
1226
1227    if (!is_tst_cond(cond)) {
1228        tcg_out_dat_IN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0, a, b);
1229        return cond;
1230    }
1231
1232    /*
1233     * The compare constraints allow rIN, but TST does not support N.
1234     * Be prepared to load the constant into a scratch register.
1235     */
1236    imm12 = encode_imm(b);
1237    if (imm12 >= 0) {
1238        tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, a, imm12);
1239    } else {
1240        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, b);
1241        tcg_out_dat_reg(s, COND_AL, ARITH_TST, 0,
1242                        a, TCG_REG_TMP, SHIFT_IMM_LSL(0));
1243    }
1244    return tcg_tst_eqne_cond(cond);
1245}
1246
1247static TCGCond tcg_out_cmp(TCGContext *s, TCGCond cond, TCGReg a,
1248                           TCGArg b, int b_const)
1249{
1250    if (b_const) {
1251        return tgen_cmpi(s, cond, a, b);
1252    } else {
1253        return tgen_cmp(s, cond, a, b);
1254    }
1255}
1256
1257static TCGCond tcg_out_cmp2(TCGContext *s, TCGCond cond, TCGReg al, TCGReg ah,
1258                            TCGArg bl, bool const_bl, TCGArg bh, bool const_bh)
1259{
1260    switch (cond) {
1261    case TCG_COND_EQ:
1262    case TCG_COND_NE:
1263    case TCG_COND_LTU:
1264    case TCG_COND_LEU:
1265    case TCG_COND_GTU:
1266    case TCG_COND_GEU:
1267        /*
1268         * We perform a conditional comparison.  If the high half is
1269         * equal, then overwrite the flags with the comparison of the
1270         * low half.  The resulting flags cover the whole.
1271         */
1272        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, ah, bh, const_bh);
1273        tcg_out_dat_rI(s, COND_EQ, ARITH_CMP, 0, al, bl, const_bl);
1274        return cond;
1275
1276    case TCG_COND_TSTEQ:
1277    case TCG_COND_TSTNE:
1278        /* Similar, but with TST instead of CMP. */
1279        tcg_out_dat_rI(s, COND_AL, ARITH_TST, 0, ah, bh, const_bh);
1280        tcg_out_dat_rI(s, COND_EQ, ARITH_TST, 0, al, bl, const_bl);
1281        return tcg_tst_eqne_cond(cond);
1282
1283    case TCG_COND_LT:
1284    case TCG_COND_GE:
1285        /* We perform a double-word subtraction and examine the result.
1286           We do not actually need the result of the subtract, so the
1287           low part "subtract" is a compare.  For the high half we have
1288           no choice but to compute into a temporary.  */
1289        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, al, bl, const_bl);
1290        tcg_out_dat_rI(s, COND_AL, ARITH_SBC | TO_CPSR,
1291                       TCG_REG_TMP, ah, bh, const_bh);
1292        return cond;
1293
1294    case TCG_COND_LE:
1295    case TCG_COND_GT:
1296        /* Similar, but with swapped arguments, via reversed subtract.  */
1297        tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR,
1298                       TCG_REG_TMP, al, bl, const_bl);
1299        tcg_out_dat_rI(s, COND_AL, ARITH_RSC | TO_CPSR,
1300                       TCG_REG_TMP, ah, bh, const_bh);
1301        return tcg_swap_cond(cond);
1302
1303    default:
1304        g_assert_not_reached();
1305    }
1306}
1307
1308/*
1309 * Note that TCGReg references Q-registers.
1310 * Q-regno = 2 * D-regno, so shift left by 1 while inserting.
1311 */
1312static uint32_t encode_vd(TCGReg rd)
1313{
1314    tcg_debug_assert(rd >= TCG_REG_Q0);
1315    return (extract32(rd, 3, 1) << 22) | (extract32(rd, 0, 3) << 13);
1316}
1317
1318static uint32_t encode_vn(TCGReg rn)
1319{
1320    tcg_debug_assert(rn >= TCG_REG_Q0);
1321    return (extract32(rn, 3, 1) << 7) | (extract32(rn, 0, 3) << 17);
1322}
1323
1324static uint32_t encode_vm(TCGReg rm)
1325{
1326    tcg_debug_assert(rm >= TCG_REG_Q0);
1327    return (extract32(rm, 3, 1) << 5) | (extract32(rm, 0, 3) << 1);
1328}
1329
1330static void tcg_out_vreg2(TCGContext *s, ARMInsn insn, int q, int vece,
1331                          TCGReg d, TCGReg m)
1332{
1333    tcg_out32(s, insn | (vece << 18) | (q << 6) |
1334              encode_vd(d) | encode_vm(m));
1335}
1336
1337static void tcg_out_vreg3(TCGContext *s, ARMInsn insn, int q, int vece,
1338                          TCGReg d, TCGReg n, TCGReg m)
1339{
1340    tcg_out32(s, insn | (vece << 20) | (q << 6) |
1341              encode_vd(d) | encode_vn(n) | encode_vm(m));
1342}
1343
1344static void tcg_out_vmovi(TCGContext *s, TCGReg rd,
1345                          int q, int op, int cmode, uint8_t imm8)
1346{
1347    tcg_out32(s, INSN_VMOVI | encode_vd(rd) | (q << 6) | (op << 5)
1348              | (cmode << 8) | extract32(imm8, 0, 4)
1349              | (extract32(imm8, 4, 3) << 16)
1350              | (extract32(imm8, 7, 1) << 24));
1351}
1352
1353static void tcg_out_vshifti(TCGContext *s, ARMInsn insn, int q,
1354                            TCGReg rd, TCGReg rm, int l_imm6)
1355{
1356    tcg_out32(s, insn | (q << 6) | encode_vd(rd) | encode_vm(rm) |
1357              (extract32(l_imm6, 6, 1) << 7) |
1358              (extract32(l_imm6, 0, 6) << 16));
1359}
1360
1361static void tcg_out_vldst(TCGContext *s, ARMInsn insn,
1362                          TCGReg rd, TCGReg rn, int offset)
1363{
1364    if (offset != 0) {
1365        if (check_fit_imm(offset) || check_fit_imm(-offset)) {
1366            tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
1367                            TCG_REG_TMP, rn, offset, true);
1368        } else {
1369            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP, offset);
1370            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1371                            TCG_REG_TMP, TCG_REG_TMP, rn, 0);
1372        }
1373        rn = TCG_REG_TMP;
1374    }
1375    tcg_out32(s, insn | (rn << 16) | encode_vd(rd) | 0xf);
1376}
1377
1378typedef struct {
1379    ARMCond cond;
1380    TCGReg base;
1381    int index;
1382    bool index_scratch;
1383    TCGAtomAlign aa;
1384} HostAddress;
1385
1386bool tcg_target_has_memory_bswap(MemOp memop)
1387{
1388    return false;
1389}
1390
1391static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
1392{
1393    /* We arrive at the slow path via "BLNE", so R14 contains l->raddr. */
1394    return TCG_REG_R14;
1395}
1396
1397static const TCGLdstHelperParam ldst_helper_param = {
1398    .ra_gen = ldst_ra_gen,
1399    .ntmp = 1,
1400    .tmp = { TCG_REG_TMP },
1401};
1402
1403static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1404{
1405    MemOp opc = get_memop(lb->oi);
1406
1407    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1408        return false;
1409    }
1410
1411    tcg_out_ld_helper_args(s, lb, &ldst_helper_param);
1412    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE]);
1413    tcg_out_ld_helper_ret(s, lb, false, &ldst_helper_param);
1414
1415    tcg_out_goto(s, COND_AL, lb->raddr);
1416    return true;
1417}
1418
1419static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1420{
1421    MemOp opc = get_memop(lb->oi);
1422
1423    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1424        return false;
1425    }
1426
1427    tcg_out_st_helper_args(s, lb, &ldst_helper_param);
1428
1429    /* Tail-call to the helper, which will return to the fast path.  */
1430    tcg_out_goto(s, COND_AL, qemu_st_helpers[opc & MO_SIZE]);
1431    return true;
1432}
1433
1434/* We expect to use an 9-bit sign-magnitude negative offset from ENV.  */
1435#define MIN_TLB_MASK_TABLE_OFS  -256
1436
1437static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
1438                                           TCGReg addr, MemOpIdx oi, bool is_ld)
1439{
1440    TCGLabelQemuLdst *ldst = NULL;
1441    MemOp opc = get_memop(oi);
1442    unsigned a_mask;
1443
1444    if (tcg_use_softmmu) {
1445        *h = (HostAddress){
1446            .cond = COND_AL,
1447            .base = addr,
1448            .index = TCG_REG_R1,
1449            .index_scratch = true,
1450        };
1451    } else {
1452        *h = (HostAddress){
1453            .cond = COND_AL,
1454            .base = addr,
1455            .index = guest_base ? TCG_REG_GUEST_BASE : -1,
1456            .index_scratch = false,
1457        };
1458    }
1459
1460    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, false);
1461    a_mask = (1 << h->aa.align) - 1;
1462
1463    if (tcg_use_softmmu) {
1464        int mem_index = get_mmuidx(oi);
1465        int cmp_off = is_ld ? offsetof(CPUTLBEntry, addr_read)
1466                            : offsetof(CPUTLBEntry, addr_write);
1467        int fast_off = tlb_mask_table_ofs(s, mem_index);
1468        unsigned s_mask = (1 << (opc & MO_SIZE)) - 1;
1469        TCGReg t_addr;
1470
1471        ldst = new_ldst_label(s);
1472        ldst->is_ld = is_ld;
1473        ldst->oi = oi;
1474        ldst->addr_reg = addr;
1475
1476        /* Load cpu->neg.tlb.f[mmu_idx].{mask,table} into {r0,r1}.  */
1477        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, mask) != 0);
1478        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, table) != 4);
1479        tcg_out_ldrd_8(s, COND_AL, TCG_REG_R0, TCG_AREG0, fast_off);
1480
1481        /* Extract the tlb index from the address into R0.  */
1482        tcg_out_dat_reg(s, COND_AL, ARITH_AND, TCG_REG_R0, TCG_REG_R0, addr,
1483                        SHIFT_IMM_LSR(s->page_bits - CPU_TLB_ENTRY_BITS));
1484
1485        /*
1486         * Add the tlb_table pointer, creating the CPUTLBEntry address in R1.
1487         * Load the tlb comparator into R2 and the fast path addend into R1.
1488         */
1489        if (cmp_off == 0) {
1490            tcg_out_ld32_rwb(s, COND_AL, TCG_REG_R2, TCG_REG_R1, TCG_REG_R0);
1491        } else {
1492            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1493                            TCG_REG_R1, TCG_REG_R1, TCG_REG_R0, 0);
1494            tcg_out_ld32_12(s, COND_AL, TCG_REG_R2, TCG_REG_R1, cmp_off);
1495        }
1496
1497        /* Load the tlb addend.  */
1498        tcg_out_ld32_12(s, COND_AL, TCG_REG_R1, TCG_REG_R1,
1499                        offsetof(CPUTLBEntry, addend));
1500
1501        /*
1502         * Check alignment, check comparators.
1503         * Do this in 2-4 insns.  Use MOVW for v7, if possible,
1504         * to reduce the number of sequential conditional instructions.
1505         * Almost all guests have at least 4k pages, which means that we need
1506         * to clear at least 9 bits even for an 8-byte memory, which means it
1507         * isn't worth checking for an immediate operand for BIC.
1508         *
1509         * For unaligned accesses, test the page of the last unit of alignment.
1510         * This leaves the least significant alignment bits unchanged, and of
1511         * course must be zero.
1512         */
1513        t_addr = addr;
1514        if (a_mask < s_mask) {
1515            t_addr = TCG_REG_R0;
1516            tcg_out_dat_imm(s, COND_AL, ARITH_ADD, t_addr,
1517                            addr, s_mask - a_mask);
1518        }
1519        if (use_armv7_instructions && s->page_bits <= 16) {
1520            tcg_out_movi32(s, COND_AL, TCG_REG_TMP, ~(s->page_mask | a_mask));
1521            tcg_out_dat_reg(s, COND_AL, ARITH_BIC, TCG_REG_TMP,
1522                            t_addr, TCG_REG_TMP, 0);
1523            tcg_out_dat_reg(s, COND_AL, ARITH_CMP, 0,
1524                            TCG_REG_R2, TCG_REG_TMP, 0);
1525        } else {
1526            if (a_mask) {
1527                tcg_debug_assert(a_mask <= 0xff);
1528                tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addr, a_mask);
1529            }
1530            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0, t_addr,
1531                            SHIFT_IMM_LSR(s->page_bits));
1532            tcg_out_dat_reg(s, (a_mask ? COND_EQ : COND_AL), ARITH_CMP,
1533                            0, TCG_REG_R2, TCG_REG_TMP,
1534                            SHIFT_IMM_LSL(s->page_bits));
1535        }
1536    } else if (a_mask) {
1537        ldst = new_ldst_label(s);
1538        ldst->is_ld = is_ld;
1539        ldst->oi = oi;
1540        ldst->addr_reg = addr;
1541
1542        /* We are expecting alignment to max out at 7 */
1543        tcg_debug_assert(a_mask <= 0xff);
1544        /* tst addr, #mask */
1545        tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addr, a_mask);
1546    }
1547
1548    return ldst;
1549}
1550
1551static void tcg_out_qemu_ld_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1552                                   TCGReg datahi, HostAddress h)
1553{
1554    TCGReg base;
1555
1556    /* Byte swapping is left to middle-end expansion. */
1557    tcg_debug_assert((opc & MO_BSWAP) == 0);
1558
1559    switch (opc & MO_SSIZE) {
1560    case MO_UB:
1561        if (h.index < 0) {
1562            tcg_out_ld8_12(s, h.cond, datalo, h.base, 0);
1563        } else {
1564            tcg_out_ld8_r(s, h.cond, datalo, h.base, h.index);
1565        }
1566        break;
1567    case MO_SB:
1568        if (h.index < 0) {
1569            tcg_out_ld8s_8(s, h.cond, datalo, h.base, 0);
1570        } else {
1571            tcg_out_ld8s_r(s, h.cond, datalo, h.base, h.index);
1572        }
1573        break;
1574    case MO_UW:
1575        if (h.index < 0) {
1576            tcg_out_ld16u_8(s, h.cond, datalo, h.base, 0);
1577        } else {
1578            tcg_out_ld16u_r(s, h.cond, datalo, h.base, h.index);
1579        }
1580        break;
1581    case MO_SW:
1582        if (h.index < 0) {
1583            tcg_out_ld16s_8(s, h.cond, datalo, h.base, 0);
1584        } else {
1585            tcg_out_ld16s_r(s, h.cond, datalo, h.base, h.index);
1586        }
1587        break;
1588    case MO_UL:
1589        if (h.index < 0) {
1590            tcg_out_ld32_12(s, h.cond, datalo, h.base, 0);
1591        } else {
1592            tcg_out_ld32_r(s, h.cond, datalo, h.base, h.index);
1593        }
1594        break;
1595    case MO_UQ:
1596        /* We used pair allocation for datalo, so already should be aligned. */
1597        tcg_debug_assert((datalo & 1) == 0);
1598        tcg_debug_assert(datahi == datalo + 1);
1599        /* LDRD requires alignment; double-check that. */
1600        if (memop_alignment_bits(opc) >= MO_64) {
1601            if (h.index < 0) {
1602                tcg_out_ldrd_8(s, h.cond, datalo, h.base, 0);
1603                break;
1604            }
1605            /*
1606             * Rm (the second address op) must not overlap Rt or Rt + 1.
1607             * Since datalo is aligned, we can simplify the test via alignment.
1608             * Flip the two address arguments if that works.
1609             */
1610            if ((h.index & ~1) != datalo) {
1611                tcg_out_ldrd_r(s, h.cond, datalo, h.base, h.index);
1612                break;
1613            }
1614            if ((h.base & ~1) != datalo) {
1615                tcg_out_ldrd_r(s, h.cond, datalo, h.index, h.base);
1616                break;
1617            }
1618        }
1619        if (h.index < 0) {
1620            base = h.base;
1621            if (datalo == h.base) {
1622                tcg_out_mov_reg(s, h.cond, TCG_REG_TMP, base);
1623                base = TCG_REG_TMP;
1624            }
1625        } else if (h.index_scratch) {
1626            tcg_out_ld32_rwb(s, h.cond, datalo, h.index, h.base);
1627            tcg_out_ld32_12(s, h.cond, datahi, h.index, 4);
1628            break;
1629        } else {
1630            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1631                            h.base, h.index, SHIFT_IMM_LSL(0));
1632            base = TCG_REG_TMP;
1633        }
1634        tcg_out_ld32_12(s, h.cond, datalo, base, 0);
1635        tcg_out_ld32_12(s, h.cond, datahi, base, 4);
1636        break;
1637    default:
1638        g_assert_not_reached();
1639    }
1640}
1641
1642static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
1643                            TCGReg addr, MemOpIdx oi, TCGType data_type)
1644{
1645    MemOp opc = get_memop(oi);
1646    TCGLabelQemuLdst *ldst;
1647    HostAddress h;
1648
1649    ldst = prepare_host_addr(s, &h, addr, oi, true);
1650    if (ldst) {
1651        ldst->type = data_type;
1652        ldst->datalo_reg = datalo;
1653        ldst->datahi_reg = datahi;
1654
1655        /*
1656         * This a conditional BL only to load a pointer within this
1657         * opcode into LR for the slow path.  We will not be using
1658         * the value for a tail call.
1659         */
1660        ldst->label_ptr[0] = s->code_ptr;
1661        tcg_out_bl_imm(s, COND_NE, 0);
1662
1663        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1664        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1665    } else {
1666        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1667    }
1668}
1669
1670static void tcg_out_qemu_st_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1671                                   TCGReg datahi, HostAddress h)
1672{
1673    /* Byte swapping is left to middle-end expansion. */
1674    tcg_debug_assert((opc & MO_BSWAP) == 0);
1675
1676    switch (opc & MO_SIZE) {
1677    case MO_8:
1678        if (h.index < 0) {
1679            tcg_out_st8_12(s, h.cond, datalo, h.base, 0);
1680        } else {
1681            tcg_out_st8_r(s, h.cond, datalo, h.base, h.index);
1682        }
1683        break;
1684    case MO_16:
1685        if (h.index < 0) {
1686            tcg_out_st16_8(s, h.cond, datalo, h.base, 0);
1687        } else {
1688            tcg_out_st16_r(s, h.cond, datalo, h.base, h.index);
1689        }
1690        break;
1691    case MO_32:
1692        if (h.index < 0) {
1693            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1694        } else {
1695            tcg_out_st32_r(s, h.cond, datalo, h.base, h.index);
1696        }
1697        break;
1698    case MO_64:
1699        /* We used pair allocation for datalo, so already should be aligned. */
1700        tcg_debug_assert((datalo & 1) == 0);
1701        tcg_debug_assert(datahi == datalo + 1);
1702        /* STRD requires alignment; double-check that. */
1703        if (memop_alignment_bits(opc) >= MO_64) {
1704            if (h.index < 0) {
1705                tcg_out_strd_8(s, h.cond, datalo, h.base, 0);
1706            } else {
1707                tcg_out_strd_r(s, h.cond, datalo, h.base, h.index);
1708            }
1709        } else if (h.index < 0) {
1710            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1711            tcg_out_st32_12(s, h.cond, datahi, h.base, 4);
1712        } else if (h.index_scratch) {
1713            tcg_out_st32_rwb(s, h.cond, datalo, h.index, h.base);
1714            tcg_out_st32_12(s, h.cond, datahi, h.index, 4);
1715        } else {
1716            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1717                            h.base, h.index, SHIFT_IMM_LSL(0));
1718            tcg_out_st32_12(s, h.cond, datalo, TCG_REG_TMP, 0);
1719            tcg_out_st32_12(s, h.cond, datahi, TCG_REG_TMP, 4);
1720        }
1721        break;
1722    default:
1723        g_assert_not_reached();
1724    }
1725}
1726
1727static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
1728                            TCGReg addr, MemOpIdx oi, TCGType data_type)
1729{
1730    MemOp opc = get_memop(oi);
1731    TCGLabelQemuLdst *ldst;
1732    HostAddress h;
1733
1734    ldst = prepare_host_addr(s, &h, addr, oi, false);
1735    if (ldst) {
1736        ldst->type = data_type;
1737        ldst->datalo_reg = datalo;
1738        ldst->datahi_reg = datahi;
1739
1740        h.cond = COND_EQ;
1741        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1742
1743        /* The conditional call is last, as we're going to return here. */
1744        ldst->label_ptr[0] = s->code_ptr;
1745        tcg_out_bl_imm(s, COND_NE, 0);
1746        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1747    } else {
1748        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1749    }
1750}
1751
1752static void tcg_out_epilogue(TCGContext *s);
1753
1754static void tcg_out_exit_tb(TCGContext *s, uintptr_t arg)
1755{
1756    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, arg);
1757    tcg_out_epilogue(s);
1758}
1759
1760static void tcg_out_goto_tb(TCGContext *s, int which)
1761{
1762    uintptr_t i_addr;
1763    intptr_t i_disp;
1764
1765    /* Direct branch will be patched by tb_target_set_jmp_target. */
1766    set_jmp_insn_offset(s, which);
1767    tcg_out32(s, INSN_NOP);
1768
1769    /* When branch is out of range, fall through to indirect. */
1770    i_addr = get_jmp_target_addr(s, which);
1771    i_disp = tcg_pcrel_diff(s, (void *)i_addr) - 8;
1772    tcg_debug_assert(i_disp < 0);
1773    if (i_disp >= -0xfff) {
1774        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_PC, i_disp);
1775    } else {
1776        /*
1777         * The TB is close, but outside the 12 bits addressable by
1778         * the load.  We can extend this to 20 bits with a sub of a
1779         * shifted immediate from pc.
1780         */
1781        int h = -i_disp;
1782        int l = -(h & 0xfff);
1783
1784        h = encode_imm_nofail(h + l);
1785        tcg_out_dat_imm(s, COND_AL, ARITH_SUB, TCG_REG_R0, TCG_REG_PC, h);
1786        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_R0, l);
1787    }
1788    set_jmp_reset_offset(s, which);
1789}
1790
1791void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1792                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1793{
1794    uintptr_t addr = tb->jmp_target_addr[n];
1795    ptrdiff_t offset = addr - (jmp_rx + 8);
1796    tcg_insn_unit insn;
1797
1798    /* Either directly branch, or fall through to indirect branch. */
1799    if (offset == sextract64(offset, 0, 26)) {
1800        /* B <addr> */
1801        insn = deposit32((COND_AL << 28) | INSN_B, 0, 24, offset >> 2);
1802    } else {
1803        insn = INSN_NOP;
1804    }
1805
1806    qatomic_set((uint32_t *)jmp_rw, insn);
1807    flush_idcache_range(jmp_rx, jmp_rw, 4);
1808}
1809
1810
1811static void tgen_add(TCGContext *s, TCGType type,
1812                     TCGReg a0, TCGReg a1, TCGReg a2)
1813{
1814    tcg_out_dat_reg(s, COND_AL, ARITH_ADD, a0, a1, a2, SHIFT_IMM_LSL(0));
1815}
1816
1817static void tgen_addi(TCGContext *s, TCGType type,
1818                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1819{
1820    tcg_out_dat_IN(s, COND_AL, ARITH_ADD, ARITH_SUB, a0, a1, a2);
1821}
1822
1823static const TCGOutOpBinary outop_add = {
1824    .base.static_constraint = C_O1_I2(r, r, rIN),
1825    .out_rrr = tgen_add,
1826    .out_rri = tgen_addi,
1827};
1828
1829static void tgen_and(TCGContext *s, TCGType type,
1830                     TCGReg a0, TCGReg a1, TCGReg a2)
1831{
1832    tcg_out_dat_reg(s, COND_AL, ARITH_AND, a0, a1, a2, SHIFT_IMM_LSL(0));
1833}
1834
1835static void tgen_andi(TCGContext *s, TCGType type,
1836                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1837{
1838    tcg_out_dat_IK(s, COND_AL, ARITH_AND, ARITH_BIC, a0, a1, a2);
1839}
1840
1841static const TCGOutOpBinary outop_and = {
1842    .base.static_constraint = C_O1_I2(r, r, rIK),
1843    .out_rrr = tgen_and,
1844    .out_rri = tgen_andi,
1845};
1846
1847static void tgen_andc(TCGContext *s, TCGType type,
1848                      TCGReg a0, TCGReg a1, TCGReg a2)
1849{
1850    tcg_out_dat_reg(s, COND_AL, ARITH_BIC, a0, a1, a2, SHIFT_IMM_LSL(0));
1851}
1852
1853static const TCGOutOpBinary outop_andc = {
1854    .base.static_constraint = C_O1_I2(r, r, r),
1855    .out_rrr = tgen_andc,
1856};
1857
1858static void tgen_clz(TCGContext *s, TCGType type,
1859                     TCGReg a0, TCGReg a1, TCGReg a2)
1860{
1861    tcg_out_dat_imm(s, COND_AL, ARITH_CMP, 0, a1, 0);
1862    tcg_out_dat_reg(s, COND_NE, INSN_CLZ, a0, 0, a1, 0);
1863    tcg_out_mov_reg(s, COND_EQ, a0, a2);
1864}
1865
1866static void tgen_clzi(TCGContext *s, TCGType type,
1867                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1868{
1869    if (a2 == 32) {
1870        tcg_out_dat_reg(s, COND_AL, INSN_CLZ, a0, 0, a1, 0);
1871    } else {
1872        tcg_out_dat_imm(s, COND_AL, ARITH_CMP, 0, a1, 0);
1873        tcg_out_dat_reg(s, COND_NE, INSN_CLZ, a0, 0, a1, 0);
1874        tcg_out_movi32(s, COND_EQ, a0, a2);
1875    }
1876}
1877
1878static const TCGOutOpBinary outop_clz = {
1879    .base.static_constraint = C_O1_I2(r, r, rIK),
1880    .out_rrr = tgen_clz,
1881    .out_rri = tgen_clzi,
1882};
1883
1884static const TCGOutOpUnary outop_ctpop = {
1885    .base.static_constraint = C_NotImplemented,
1886};
1887
1888static void tgen_ctz(TCGContext *s, TCGType type,
1889                     TCGReg a0, TCGReg a1, TCGReg a2)
1890{
1891    tcg_out_dat_reg(s, COND_AL, INSN_RBIT, TCG_REG_TMP, 0, a1, 0);
1892    tgen_clz(s, TCG_TYPE_I32, a0, TCG_REG_TMP, a2);
1893}
1894
1895static void tgen_ctzi(TCGContext *s, TCGType type,
1896                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1897{
1898    tcg_out_dat_reg(s, COND_AL, INSN_RBIT, TCG_REG_TMP, 0, a1, 0);
1899    tgen_clzi(s, TCG_TYPE_I32, a0, TCG_REG_TMP, a2);
1900}
1901
1902static TCGConstraintSetIndex cset_ctz(TCGType type, unsigned flags)
1903{
1904    return use_armv7_instructions ? C_O1_I2(r, r, rIK) : C_NotImplemented;
1905}
1906
1907static const TCGOutOpBinary outop_ctz = {
1908    .base.static_constraint = C_Dynamic,
1909    .base.dynamic_constraint = cset_ctz,
1910    .out_rrr = tgen_ctz,
1911    .out_rri = tgen_ctzi,
1912};
1913
1914static TCGConstraintSetIndex cset_idiv(TCGType type, unsigned flags)
1915{
1916    return use_idiv_instructions ? C_O1_I2(r, r, r) : C_NotImplemented;
1917}
1918
1919static void tgen_divs(TCGContext *s, TCGType type,
1920                      TCGReg a0, TCGReg a1, TCGReg a2)
1921{
1922    /* sdiv */
1923    tcg_out32(s, 0x0710f010 | (COND_AL << 28) | (a0 << 16) | a1 | (a2 << 8));
1924}
1925
1926static const TCGOutOpBinary outop_divs = {
1927    .base.static_constraint = C_Dynamic,
1928    .base.dynamic_constraint = cset_idiv,
1929    .out_rrr = tgen_divs,
1930};
1931
1932static const TCGOutOpDivRem outop_divs2 = {
1933    .base.static_constraint = C_NotImplemented,
1934};
1935
1936static void tgen_divu(TCGContext *s, TCGType type,
1937                      TCGReg a0, TCGReg a1, TCGReg a2)
1938{
1939    /* udiv */
1940    tcg_out32(s, 0x0730f010 | (COND_AL << 28) | (a0 << 16) | a1 | (a2 << 8));
1941}
1942
1943static const TCGOutOpBinary outop_divu = {
1944    .base.static_constraint = C_Dynamic,
1945    .base.dynamic_constraint = cset_idiv,
1946    .out_rrr = tgen_divu,
1947};
1948
1949static const TCGOutOpDivRem outop_divu2 = {
1950    .base.static_constraint = C_NotImplemented,
1951};
1952
1953static const TCGOutOpBinary outop_eqv = {
1954    .base.static_constraint = C_NotImplemented,
1955};
1956
1957static void tgen_mul(TCGContext *s, TCGType type,
1958                     TCGReg a0, TCGReg a1, TCGReg a2)
1959{
1960    /* mul */
1961    tcg_out32(s, (COND_AL << 28) | 0x90 | (a0 << 16) | (a1 << 8) | a2);
1962}
1963
1964static const TCGOutOpBinary outop_mul = {
1965    .base.static_constraint = C_O1_I2(r, r, r),
1966    .out_rrr = tgen_mul,
1967};
1968
1969static void tgen_muls2(TCGContext *s, TCGType type,
1970                       TCGReg rd0, TCGReg rd1, TCGReg rn, TCGReg rm)
1971{
1972    /* smull */
1973    tcg_out32(s, (COND_AL << 28) | 0x00c00090 |
1974              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
1975}
1976
1977static const TCGOutOpMul2 outop_muls2 = {
1978    .base.static_constraint = C_O2_I2(r, r, r, r),
1979    .out_rrrr = tgen_muls2,
1980};
1981
1982static const TCGOutOpBinary outop_mulsh = {
1983    .base.static_constraint = C_NotImplemented,
1984};
1985
1986static void tgen_mulu2(TCGContext *s, TCGType type,
1987                       TCGReg rd0, TCGReg rd1, TCGReg rn, TCGReg rm)
1988{
1989    /* umull */
1990    tcg_out32(s, (COND_AL << 28) | 0x00800090 |
1991              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
1992}
1993
1994static const TCGOutOpMul2 outop_mulu2 = {
1995    .base.static_constraint = C_O2_I2(r, r, r, r),
1996    .out_rrrr = tgen_mulu2,
1997};
1998
1999static const TCGOutOpBinary outop_muluh = {
2000    .base.static_constraint = C_NotImplemented,
2001};
2002
2003static const TCGOutOpBinary outop_nand = {
2004    .base.static_constraint = C_NotImplemented,
2005};
2006
2007static const TCGOutOpBinary outop_nor = {
2008    .base.static_constraint = C_NotImplemented,
2009};
2010
2011static void tgen_or(TCGContext *s, TCGType type,
2012                     TCGReg a0, TCGReg a1, TCGReg a2)
2013{
2014    tcg_out_dat_reg(s, COND_AL, ARITH_ORR, a0, a1, a2, SHIFT_IMM_LSL(0));
2015}
2016
2017static void tgen_ori(TCGContext *s, TCGType type,
2018                     TCGReg a0, TCGReg a1, tcg_target_long a2)
2019{
2020    tcg_out_dat_imm(s, COND_AL, ARITH_ORR, a0, a1, encode_imm_nofail(a2));
2021}
2022
2023static const TCGOutOpBinary outop_or = {
2024    .base.static_constraint = C_O1_I2(r, r, rI),
2025    .out_rrr = tgen_or,
2026    .out_rri = tgen_ori,
2027};
2028
2029static const TCGOutOpBinary outop_orc = {
2030    .base.static_constraint = C_NotImplemented,
2031};
2032
2033static const TCGOutOpBinary outop_rems = {
2034    .base.static_constraint = C_NotImplemented,
2035};
2036
2037static const TCGOutOpBinary outop_remu = {
2038    .base.static_constraint = C_NotImplemented,
2039};
2040
2041static const TCGOutOpBinary outop_rotl = {
2042    .base.static_constraint = C_NotImplemented,
2043};
2044
2045static void tgen_rotr(TCGContext *s, TCGType type,
2046                      TCGReg a0, TCGReg a1, TCGReg a2)
2047{
2048    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1, SHIFT_REG_ROR(a2));
2049}
2050
2051static void tgen_rotri(TCGContext *s, TCGType type,
2052                       TCGReg a0, TCGReg a1, tcg_target_long a2)
2053{
2054    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1, SHIFT_IMM_ROR(a2 & 0x1f));
2055}
2056
2057static const TCGOutOpBinary outop_rotr = {
2058    .base.static_constraint = C_O1_I2(r, r, ri),
2059    .out_rrr = tgen_rotr,
2060    .out_rri = tgen_rotri,
2061};
2062
2063static void tgen_sar(TCGContext *s, TCGType type,
2064                     TCGReg a0, TCGReg a1, TCGReg a2)
2065{
2066    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1, SHIFT_REG_ASR(a2));
2067}
2068
2069static void tgen_sari(TCGContext *s, TCGType type,
2070                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2071{
2072    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1,
2073                    SHIFT_IMM_ASR(a2 & 0x1f));
2074}
2075
2076static const TCGOutOpBinary outop_sar = {
2077    .base.static_constraint = C_O1_I2(r, r, ri),
2078    .out_rrr = tgen_sar,
2079    .out_rri = tgen_sari,
2080};
2081
2082static void tgen_shl(TCGContext *s, TCGType type,
2083                     TCGReg a0, TCGReg a1, TCGReg a2)
2084{
2085    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1, SHIFT_REG_LSL(a2));
2086}
2087
2088static void tgen_shli(TCGContext *s, TCGType type,
2089                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2090{
2091    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1,
2092                    SHIFT_IMM_LSL(a2 & 0x1f));
2093}
2094
2095static const TCGOutOpBinary outop_shl = {
2096    .base.static_constraint = C_O1_I2(r, r, ri),
2097    .out_rrr = tgen_shl,
2098    .out_rri = tgen_shli,
2099};
2100
2101static void tgen_shr(TCGContext *s, TCGType type,
2102                     TCGReg a0, TCGReg a1, TCGReg a2)
2103{
2104    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1, SHIFT_REG_LSR(a2));
2105}
2106
2107static void tgen_shri(TCGContext *s, TCGType type,
2108                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2109{
2110    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1,
2111                    SHIFT_IMM_LSR(a2 & 0x1f));
2112}
2113
2114static const TCGOutOpBinary outop_shr = {
2115    .base.static_constraint = C_O1_I2(r, r, ri),
2116    .out_rrr = tgen_shr,
2117    .out_rri = tgen_shri,
2118};
2119
2120static void tgen_sub(TCGContext *s, TCGType type,
2121                     TCGReg a0, TCGReg a1, TCGReg a2)
2122{
2123    tcg_out_dat_reg(s, COND_AL, ARITH_SUB, a0, a1, a2, SHIFT_IMM_LSL(0));
2124}
2125
2126static void tgen_subfi(TCGContext *s, TCGType type,
2127                       TCGReg a0, tcg_target_long a1, TCGReg a2)
2128{
2129    tcg_out_dat_imm(s, COND_AL, ARITH_RSB, a0, a2, encode_imm_nofail(a1));
2130}
2131
2132static const TCGOutOpSubtract outop_sub = {
2133    .base.static_constraint = C_O1_I2(r, rI, r),
2134    .out_rrr = tgen_sub,
2135    .out_rir = tgen_subfi,
2136};
2137
2138static void tgen_xor(TCGContext *s, TCGType type,
2139                     TCGReg a0, TCGReg a1, TCGReg a2)
2140{
2141    tcg_out_dat_reg(s, COND_AL, ARITH_EOR, a0, a1, a2, SHIFT_IMM_LSL(0));
2142}
2143
2144static void tgen_xori(TCGContext *s, TCGType type,
2145                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2146{
2147    tcg_out_dat_imm(s, COND_AL, ARITH_EOR, a0, a1, encode_imm_nofail(a2));
2148}
2149
2150static const TCGOutOpBinary outop_xor = {
2151    .base.static_constraint = C_O1_I2(r, r, rI),
2152    .out_rrr = tgen_xor,
2153    .out_rri = tgen_xori,
2154};
2155
2156static void tgen_bswap16(TCGContext *s, TCGType type,
2157                         TCGReg rd, TCGReg rn, unsigned flags)
2158{
2159    if (flags & TCG_BSWAP_OS) {
2160        /* revsh */
2161        tcg_out32(s, 0x06ff0fb0 | (COND_AL << 28) | (rd << 12) | rn);
2162        return;
2163    }
2164
2165    /* rev16 */
2166    tcg_out32(s, 0x06bf0fb0 | (COND_AL << 28) | (rd << 12) | rn);
2167    if ((flags & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
2168        tcg_out_ext16u(s, rd, rd);
2169    }
2170}
2171
2172static const TCGOutOpBswap outop_bswap16 = {
2173    .base.static_constraint = C_O1_I1(r, r),
2174    .out_rr = tgen_bswap16,
2175};
2176
2177static void tgen_bswap32(TCGContext *s, TCGType type,
2178                         TCGReg rd, TCGReg rn, unsigned flags)
2179{
2180    /* rev */
2181    tcg_out32(s, 0x06bf0f30 | (COND_AL << 28) | (rd << 12) | rn);
2182}
2183
2184static const TCGOutOpBswap outop_bswap32 = {
2185    .base.static_constraint = C_O1_I1(r, r),
2186    .out_rr = tgen_bswap32,
2187};
2188
2189static const TCGOutOpUnary outop_bswap64 = {
2190    .base.static_constraint = C_NotImplemented,
2191};
2192
2193static void tgen_neg(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
2194{
2195    tgen_subfi(s, type, a0, 0, a1);
2196}
2197
2198static const TCGOutOpUnary outop_neg = {
2199    .base.static_constraint = C_O1_I1(r, r),
2200    .out_rr = tgen_neg,
2201};
2202
2203static void tgen_not(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
2204{
2205    tcg_out_dat_reg(s, COND_AL, ARITH_MVN, a0, 0, a1, SHIFT_IMM_LSL(0));
2206}
2207
2208static const TCGOutOpUnary outop_not = {
2209    .base.static_constraint = C_O1_I1(r, r),
2210    .out_rr = tgen_not,
2211};
2212
2213static void tgen_brcond(TCGContext *s, TCGType type, TCGCond cond,
2214                        TCGReg a0, TCGReg a1, TCGLabel *l)
2215{
2216    cond = tgen_cmp(s, cond, a0, a1);
2217    tcg_out_goto_label(s, tcg_cond_to_arm_cond[cond], l);
2218}
2219
2220static void tgen_brcondi(TCGContext *s, TCGType type, TCGCond cond,
2221                         TCGReg a0, tcg_target_long a1, TCGLabel *l)
2222{
2223    cond = tgen_cmpi(s, cond, a0, a1);
2224    tcg_out_goto_label(s, tcg_cond_to_arm_cond[cond], l);
2225}
2226
2227static const TCGOutOpBrcond outop_brcond = {
2228    .base.static_constraint = C_O0_I2(r, rIN),
2229    .out_rr = tgen_brcond,
2230    .out_ri = tgen_brcondi,
2231};
2232
2233static void finish_setcond(TCGContext *s, TCGCond cond, TCGReg ret, bool neg)
2234{
2235    tcg_out_movi32(s, tcg_cond_to_arm_cond[tcg_invert_cond(cond)], ret, 0);
2236    tcg_out_movi32(s, tcg_cond_to_arm_cond[cond], ret, neg ? -1 : 1);
2237}
2238
2239static void tgen_setcond(TCGContext *s, TCGType type, TCGCond cond,
2240                         TCGReg a0, TCGReg a1, TCGReg a2)
2241{
2242    cond = tgen_cmp(s, cond, a1, a2);
2243    finish_setcond(s, cond, a0, false);
2244}
2245
2246static void tgen_setcondi(TCGContext *s, TCGType type, TCGCond cond,
2247                          TCGReg a0, TCGReg a1, tcg_target_long a2)
2248{
2249    cond = tgen_cmpi(s, cond, a1, a2);
2250    finish_setcond(s, cond, a0, false);
2251}
2252
2253static const TCGOutOpSetcond outop_setcond = {
2254    .base.static_constraint = C_O1_I2(r, r, rIN),
2255    .out_rrr = tgen_setcond,
2256    .out_rri = tgen_setcondi,
2257};
2258
2259static void tgen_negsetcond(TCGContext *s, TCGType type, TCGCond cond,
2260                            TCGReg a0, TCGReg a1, TCGReg a2)
2261{
2262    cond = tgen_cmp(s, cond, a1, a2);
2263    finish_setcond(s, cond, a0, true);
2264}
2265
2266static void tgen_negsetcondi(TCGContext *s, TCGType type, TCGCond cond,
2267                             TCGReg a0, TCGReg a1, tcg_target_long a2)
2268{
2269    cond = tgen_cmpi(s, cond, a1, a2);
2270    finish_setcond(s, cond, a0, true);
2271}
2272
2273static const TCGOutOpSetcond outop_negsetcond = {
2274    .base.static_constraint = C_O1_I2(r, r, rIN),
2275    .out_rrr = tgen_negsetcond,
2276    .out_rri = tgen_negsetcondi,
2277};
2278
2279static void tgen_movcond(TCGContext *s, TCGType type, TCGCond cond,
2280                         TCGReg ret, TCGReg c1, TCGArg c2, bool const_c2,
2281                         TCGArg vt, bool const_vt, TCGArg vf, bool consf_vf)
2282{
2283    cond = tcg_out_cmp(s, cond, c1, c2, const_c2);
2284    tcg_out_dat_rIK(s, tcg_cond_to_arm_cond[cond], ARITH_MOV, ARITH_MVN,
2285                    ret, 0, vt, const_vt);
2286}
2287
2288static const TCGOutOpMovcond outop_movcond = {
2289    .base.static_constraint = C_O1_I4(r, r, rIN, rIK, 0),
2290    .out = tgen_movcond,
2291};
2292
2293static void tgen_brcond2(TCGContext *s, TCGCond cond, TCGReg al, TCGReg ah,
2294                         TCGArg bl, bool const_bl, TCGArg bh, bool const_bh,
2295                         TCGLabel *l)
2296{
2297    cond = tcg_out_cmp2(s, cond, al, ah, bl, const_bl, bh, const_bh);
2298    tcg_out_goto_label(s, tcg_cond_to_arm_cond[cond], l);
2299}
2300
2301static const TCGOutOpBrcond2 outop_brcond2 = {
2302    .base.static_constraint = C_O0_I4(r, r, rI, rI),
2303    .out = tgen_brcond2,
2304};
2305
2306static void tgen_setcond2(TCGContext *s, TCGCond cond, TCGReg ret,
2307                          TCGReg al, TCGReg ah,
2308                          TCGArg bl, bool const_bl,
2309                          TCGArg bh, bool const_bh)
2310{
2311    cond = tcg_out_cmp2(s, cond, al, ah, bl, const_bl, bh, const_bh);
2312    finish_setcond(s, cond, ret, false);
2313}
2314
2315static const TCGOutOpSetcond2 outop_setcond2 = {
2316    .base.static_constraint = C_O1_I4(r, r, r, rI, rI),
2317    .out = tgen_setcond2,
2318};
2319
2320static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
2321                       const TCGArg args[TCG_MAX_OP_ARGS],
2322                       const int const_args[TCG_MAX_OP_ARGS])
2323{
2324    TCGArg a0, a1, a2, a3, a4, a5;
2325
2326    switch (opc) {
2327    case INDEX_op_goto_ptr:
2328        tcg_out_b_reg(s, COND_AL, args[0]);
2329        break;
2330    case INDEX_op_br:
2331        tcg_out_goto_label(s, COND_AL, arg_label(args[0]));
2332        break;
2333
2334    case INDEX_op_ld8u_i32:
2335        tcg_out_ld8u(s, COND_AL, args[0], args[1], args[2]);
2336        break;
2337    case INDEX_op_ld8s_i32:
2338        tcg_out_ld8s(s, COND_AL, args[0], args[1], args[2]);
2339        break;
2340    case INDEX_op_ld16u_i32:
2341        tcg_out_ld16u(s, COND_AL, args[0], args[1], args[2]);
2342        break;
2343    case INDEX_op_ld16s_i32:
2344        tcg_out_ld16s(s, COND_AL, args[0], args[1], args[2]);
2345        break;
2346    case INDEX_op_ld_i32:
2347        tcg_out_ld32u(s, COND_AL, args[0], args[1], args[2]);
2348        break;
2349    case INDEX_op_st8_i32:
2350        tcg_out_st8(s, COND_AL, args[0], args[1], args[2]);
2351        break;
2352    case INDEX_op_st16_i32:
2353        tcg_out_st16(s, COND_AL, args[0], args[1], args[2]);
2354        break;
2355    case INDEX_op_st_i32:
2356        tcg_out_st32(s, COND_AL, args[0], args[1], args[2]);
2357        break;
2358
2359    case INDEX_op_add2_i32:
2360        a0 = args[0], a1 = args[1], a2 = args[2];
2361        a3 = args[3], a4 = args[4], a5 = args[5];
2362        if (a0 == a3 || (a0 == a5 && !const_args[5])) {
2363            a0 = TCG_REG_TMP;
2364        }
2365        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD | TO_CPSR, ARITH_SUB | TO_CPSR,
2366                        a0, a2, a4, const_args[4]);
2367        tcg_out_dat_rIK(s, COND_AL, ARITH_ADC, ARITH_SBC,
2368                        a1, a3, a5, const_args[5]);
2369        tcg_out_mov_reg(s, COND_AL, args[0], a0);
2370        break;
2371    case INDEX_op_sub2_i32:
2372        a0 = args[0], a1 = args[1], a2 = args[2];
2373        a3 = args[3], a4 = args[4], a5 = args[5];
2374        if ((a0 == a3 && !const_args[3]) || (a0 == a5 && !const_args[5])) {
2375            a0 = TCG_REG_TMP;
2376        }
2377        if (const_args[2]) {
2378            if (const_args[4]) {
2379                tcg_out_movi32(s, COND_AL, a0, a4);
2380                a4 = a0;
2381            }
2382            tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR, a0, a4, a2, 1);
2383        } else {
2384            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB | TO_CPSR,
2385                            ARITH_ADD | TO_CPSR, a0, a2, a4, const_args[4]);
2386        }
2387        if (const_args[3]) {
2388            if (const_args[5]) {
2389                tcg_out_movi32(s, COND_AL, a1, a5);
2390                a5 = a1;
2391            }
2392            tcg_out_dat_rI(s, COND_AL, ARITH_RSC, a1, a5, a3, 1);
2393        } else {
2394            tcg_out_dat_rIK(s, COND_AL, ARITH_SBC, ARITH_ADC,
2395                            a1, a3, a5, const_args[5]);
2396        }
2397        tcg_out_mov_reg(s, COND_AL, args[0], a0);
2398        break;
2399
2400    case INDEX_op_qemu_ld_i32:
2401        tcg_out_qemu_ld(s, args[0], -1, args[1], args[2], TCG_TYPE_I32);
2402        break;
2403    case INDEX_op_qemu_ld_i64:
2404        tcg_out_qemu_ld(s, args[0], args[1], args[2], args[3], TCG_TYPE_I64);
2405        break;
2406
2407    case INDEX_op_qemu_st_i32:
2408        tcg_out_qemu_st(s, args[0], -1, args[1], args[2], TCG_TYPE_I32);
2409        break;
2410    case INDEX_op_qemu_st_i64:
2411        tcg_out_qemu_st(s, args[0], args[1], args[2], args[3], TCG_TYPE_I64);
2412        break;
2413
2414    case INDEX_op_extract2_i32:
2415        /* ??? These optimization vs zero should be generic.  */
2416        /* ??? But we can't substitute 2 for 1 in the opcode stream yet.  */
2417        if (const_args[1]) {
2418            if (const_args[2]) {
2419                tcg_out_movi(s, TCG_TYPE_REG, args[0], 0);
2420            } else {
2421                tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2422                                args[2], SHIFT_IMM_LSL(32 - args[3]));
2423            }
2424        } else if (const_args[2]) {
2425            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2426                            args[1], SHIFT_IMM_LSR(args[3]));
2427        } else {
2428            /* We can do extract2 in 2 insns, vs the 3 required otherwise.  */
2429            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0,
2430                            args[2], SHIFT_IMM_LSL(32 - args[3]));
2431            tcg_out_dat_reg(s, COND_AL, ARITH_ORR, args[0], TCG_REG_TMP,
2432                            args[1], SHIFT_IMM_LSR(args[3]));
2433        }
2434        break;
2435
2436    case INDEX_op_mb:
2437        tcg_out_mb(s, args[0]);
2438        break;
2439
2440    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
2441    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
2442    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
2443    default:
2444        g_assert_not_reached();
2445    }
2446}
2447
2448static TCGConstraintSetIndex
2449tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
2450{
2451    switch (op) {
2452    case INDEX_op_goto_ptr:
2453        return C_O0_I1(r);
2454
2455    case INDEX_op_ld8u_i32:
2456    case INDEX_op_ld8s_i32:
2457    case INDEX_op_ld16u_i32:
2458    case INDEX_op_ld16s_i32:
2459    case INDEX_op_ld_i32:
2460        return C_O1_I1(r, r);
2461
2462    case INDEX_op_st8_i32:
2463    case INDEX_op_st16_i32:
2464    case INDEX_op_st_i32:
2465        return C_O0_I2(r, r);
2466
2467    case INDEX_op_extract2_i32:
2468        return C_O1_I2(r, rZ, rZ);
2469    case INDEX_op_add2_i32:
2470        return C_O2_I4(r, r, r, r, rIN, rIK);
2471    case INDEX_op_sub2_i32:
2472        return C_O2_I4(r, r, rI, rI, rIN, rIK);
2473    case INDEX_op_qemu_ld_i32:
2474        return C_O1_I1(r, q);
2475    case INDEX_op_qemu_ld_i64:
2476        return C_O2_I1(e, p, q);
2477    case INDEX_op_qemu_st_i32:
2478        return C_O0_I2(q, q);
2479    case INDEX_op_qemu_st_i64:
2480        return C_O0_I3(Q, p, q);
2481
2482    case INDEX_op_st_vec:
2483        return C_O0_I2(w, r);
2484    case INDEX_op_ld_vec:
2485    case INDEX_op_dupm_vec:
2486        return C_O1_I1(w, r);
2487    case INDEX_op_dup_vec:
2488        return C_O1_I1(w, wr);
2489    case INDEX_op_abs_vec:
2490    case INDEX_op_neg_vec:
2491    case INDEX_op_not_vec:
2492    case INDEX_op_shli_vec:
2493    case INDEX_op_shri_vec:
2494    case INDEX_op_sari_vec:
2495        return C_O1_I1(w, w);
2496    case INDEX_op_dup2_vec:
2497    case INDEX_op_add_vec:
2498    case INDEX_op_mul_vec:
2499    case INDEX_op_smax_vec:
2500    case INDEX_op_smin_vec:
2501    case INDEX_op_ssadd_vec:
2502    case INDEX_op_sssub_vec:
2503    case INDEX_op_sub_vec:
2504    case INDEX_op_umax_vec:
2505    case INDEX_op_umin_vec:
2506    case INDEX_op_usadd_vec:
2507    case INDEX_op_ussub_vec:
2508    case INDEX_op_xor_vec:
2509    case INDEX_op_arm_sshl_vec:
2510    case INDEX_op_arm_ushl_vec:
2511        return C_O1_I2(w, w, w);
2512    case INDEX_op_arm_sli_vec:
2513        return C_O1_I2(w, 0, w);
2514    case INDEX_op_or_vec:
2515    case INDEX_op_andc_vec:
2516        return C_O1_I2(w, w, wO);
2517    case INDEX_op_and_vec:
2518    case INDEX_op_orc_vec:
2519        return C_O1_I2(w, w, wV);
2520    case INDEX_op_cmp_vec:
2521        return C_O1_I2(w, w, wZ);
2522    case INDEX_op_bitsel_vec:
2523        return C_O1_I3(w, w, w, w);
2524    default:
2525        return C_NotImplemented;
2526    }
2527}
2528
2529static void tcg_target_init(TCGContext *s)
2530{
2531    /*
2532     * Only probe for the platform and capabilities if we haven't already
2533     * determined maximum values at compile time.
2534     */
2535#if !defined(use_idiv_instructions) || !defined(use_neon_instructions)
2536    {
2537        unsigned long hwcap = qemu_getauxval(AT_HWCAP);
2538#ifndef use_idiv_instructions
2539        use_idiv_instructions = (hwcap & HWCAP_ARM_IDIVA) != 0;
2540#endif
2541#ifndef use_neon_instructions
2542        use_neon_instructions = (hwcap & HWCAP_ARM_NEON) != 0;
2543#endif
2544    }
2545#endif
2546
2547    if (__ARM_ARCH < 7) {
2548        const char *pl = (const char *)qemu_getauxval(AT_PLATFORM);
2549        if (pl != NULL && pl[0] == 'v' && pl[1] >= '4' && pl[1] <= '9') {
2550            arm_arch = pl[1] - '0';
2551        }
2552
2553        if (arm_arch < 6) {
2554            error_report("TCG: ARMv%d is unsupported; exiting", arm_arch);
2555            exit(EXIT_FAILURE);
2556        }
2557    }
2558
2559    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
2560
2561    tcg_target_call_clobber_regs = 0;
2562    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R0);
2563    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R1);
2564    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R2);
2565    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R3);
2566    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R12);
2567    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R14);
2568
2569    if (use_neon_instructions) {
2570        tcg_target_available_regs[TCG_TYPE_V64]  = ALL_VECTOR_REGS;
2571        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
2572
2573        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q0);
2574        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q1);
2575        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q2);
2576        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q3);
2577        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q8);
2578        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q9);
2579        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q10);
2580        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q11);
2581        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q12);
2582        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q13);
2583        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q14);
2584        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q15);
2585    }
2586
2587    s->reserved_regs = 0;
2588    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
2589    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP);
2590    tcg_regset_set_reg(s->reserved_regs, TCG_REG_PC);
2591    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP);
2592}
2593
2594static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
2595                       TCGReg arg1, intptr_t arg2)
2596{
2597    switch (type) {
2598    case TCG_TYPE_I32:
2599        tcg_out_ld32u(s, COND_AL, arg, arg1, arg2);
2600        return;
2601    case TCG_TYPE_V64:
2602        /* regs 1; size 8; align 8 */
2603        tcg_out_vldst(s, INSN_VLD1 | 0x7d0, arg, arg1, arg2);
2604        return;
2605    case TCG_TYPE_V128:
2606        /*
2607         * We have only 8-byte alignment for the stack per the ABI.
2608         * Rather than dynamically re-align the stack, it's easier
2609         * to simply not request alignment beyond that.  So:
2610         * regs 2; size 8; align 8
2611         */
2612        tcg_out_vldst(s, INSN_VLD1 | 0xad0, arg, arg1, arg2);
2613        return;
2614    default:
2615        g_assert_not_reached();
2616    }
2617}
2618
2619static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
2620                       TCGReg arg1, intptr_t arg2)
2621{
2622    switch (type) {
2623    case TCG_TYPE_I32:
2624        tcg_out_st32(s, COND_AL, arg, arg1, arg2);
2625        return;
2626    case TCG_TYPE_V64:
2627        /* regs 1; size 8; align 8 */
2628        tcg_out_vldst(s, INSN_VST1 | 0x7d0, arg, arg1, arg2);
2629        return;
2630    case TCG_TYPE_V128:
2631        /* See tcg_out_ld re alignment: regs 2; size 8; align 8 */
2632        tcg_out_vldst(s, INSN_VST1 | 0xad0, arg, arg1, arg2);
2633        return;
2634    default:
2635        g_assert_not_reached();
2636    }
2637}
2638
2639static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
2640                        TCGReg base, intptr_t ofs)
2641{
2642    return false;
2643}
2644
2645static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
2646{
2647    if (ret == arg) {
2648        return true;
2649    }
2650    switch (type) {
2651    case TCG_TYPE_I32:
2652        if (ret < TCG_REG_Q0 && arg < TCG_REG_Q0) {
2653            tcg_out_mov_reg(s, COND_AL, ret, arg);
2654            return true;
2655        }
2656        return false;
2657
2658    case TCG_TYPE_V64:
2659    case TCG_TYPE_V128:
2660        /* "VMOV D,N" is an alias for "VORR D,N,N". */
2661        tcg_out_vreg3(s, INSN_VORR, type - TCG_TYPE_V64, 0, ret, arg, arg);
2662        return true;
2663
2664    default:
2665        g_assert_not_reached();
2666    }
2667}
2668
2669static void tcg_out_movi(TCGContext *s, TCGType type,
2670                         TCGReg ret, tcg_target_long arg)
2671{
2672    tcg_debug_assert(type == TCG_TYPE_I32);
2673    tcg_debug_assert(ret < TCG_REG_Q0);
2674    tcg_out_movi32(s, COND_AL, ret, arg);
2675}
2676
2677static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
2678{
2679    return false;
2680}
2681
2682static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
2683                             tcg_target_long imm)
2684{
2685    int enc, opc = ARITH_ADD;
2686
2687    /* All of the easiest immediates to encode are positive. */
2688    if (imm < 0) {
2689        imm = -imm;
2690        opc = ARITH_SUB;
2691    }
2692    enc = encode_imm(imm);
2693    if (enc >= 0) {
2694        tcg_out_dat_imm(s, COND_AL, opc, rd, rs, enc);
2695    } else {
2696        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, imm);
2697        tcg_out_dat_reg(s, COND_AL, opc, rd, rs,
2698                        TCG_REG_TMP, SHIFT_IMM_LSL(0));
2699    }
2700}
2701
2702/* Type is always V128, with I64 elements.  */
2703static void tcg_out_dup2_vec(TCGContext *s, TCGReg rd, TCGReg rl, TCGReg rh)
2704{
2705    /* Move high element into place first. */
2706    /* VMOV Dd+1, Ds */
2707    tcg_out_vreg3(s, INSN_VORR | (1 << 12), 0, 0, rd, rh, rh);
2708    /* Move low element into place; tcg_out_mov will check for nop. */
2709    tcg_out_mov(s, TCG_TYPE_V64, rd, rl);
2710}
2711
2712static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
2713                            TCGReg rd, TCGReg rs)
2714{
2715    int q = type - TCG_TYPE_V64;
2716
2717    if (vece == MO_64) {
2718        if (type == TCG_TYPE_V128) {
2719            tcg_out_dup2_vec(s, rd, rs, rs);
2720        } else {
2721            tcg_out_mov(s, TCG_TYPE_V64, rd, rs);
2722        }
2723    } else if (rs < TCG_REG_Q0) {
2724        int b = (vece == MO_8);
2725        int e = (vece == MO_16);
2726        tcg_out32(s, INSN_VDUP_G | (b << 22) | (q << 21) | (e << 5) |
2727                  encode_vn(rd) | (rs << 12));
2728    } else {
2729        int imm4 = 1 << vece;
2730        tcg_out32(s, INSN_VDUP_S | (imm4 << 16) | (q << 6) |
2731                  encode_vd(rd) | encode_vm(rs));
2732    }
2733    return true;
2734}
2735
2736static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
2737                             TCGReg rd, TCGReg base, intptr_t offset)
2738{
2739    if (vece == MO_64) {
2740        tcg_out_ld(s, TCG_TYPE_V64, rd, base, offset);
2741        if (type == TCG_TYPE_V128) {
2742            tcg_out_dup2_vec(s, rd, rd, rd);
2743        }
2744    } else {
2745        int q = type - TCG_TYPE_V64;
2746        tcg_out_vldst(s, INSN_VLD1R | (vece << 6) | (q << 5),
2747                      rd, base, offset);
2748    }
2749    return true;
2750}
2751
2752static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
2753                             TCGReg rd, int64_t v64)
2754{
2755    int q = type - TCG_TYPE_V64;
2756    int cmode, imm8, i;
2757
2758    /* Test all bytes equal first.  */
2759    if (vece == MO_8) {
2760        tcg_out_vmovi(s, rd, q, 0, 0xe, v64);
2761        return;
2762    }
2763
2764    /*
2765     * Test all bytes 0x00 or 0xff second.  This can match cases that
2766     * might otherwise take 2 or 3 insns for MO_16 or MO_32 below.
2767     */
2768    for (i = imm8 = 0; i < 8; i++) {
2769        uint8_t byte = v64 >> (i * 8);
2770        if (byte == 0xff) {
2771            imm8 |= 1 << i;
2772        } else if (byte != 0) {
2773            goto fail_bytes;
2774        }
2775    }
2776    tcg_out_vmovi(s, rd, q, 1, 0xe, imm8);
2777    return;
2778 fail_bytes:
2779
2780    /*
2781     * Tests for various replications.  For each element width, if we
2782     * cannot find an expansion there's no point checking a larger
2783     * width because we already know by replication it cannot match.
2784     */
2785    if (vece == MO_16) {
2786        uint16_t v16 = v64;
2787
2788        if (is_shimm16(v16, &cmode, &imm8)) {
2789            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2790            return;
2791        }
2792        if (is_shimm16(~v16, &cmode, &imm8)) {
2793            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2794            return;
2795        }
2796
2797        /*
2798         * Otherwise, all remaining constants can be loaded in two insns:
2799         * rd = v16 & 0xff, rd |= v16 & 0xff00.
2800         */
2801        tcg_out_vmovi(s, rd, q, 0, 0x8, v16 & 0xff);
2802        tcg_out_vmovi(s, rd, q, 0, 0xb, v16 >> 8);   /* VORRI */
2803        return;
2804    }
2805
2806    if (vece == MO_32) {
2807        uint32_t v32 = v64;
2808
2809        if (is_shimm32(v32, &cmode, &imm8) ||
2810            is_soimm32(v32, &cmode, &imm8)) {
2811            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2812            return;
2813        }
2814        if (is_shimm32(~v32, &cmode, &imm8) ||
2815            is_soimm32(~v32, &cmode, &imm8)) {
2816            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2817            return;
2818        }
2819
2820        /*
2821         * Restrict the set of constants to those we can load with
2822         * two instructions.  Others we load from the pool.
2823         */
2824        i = is_shimm32_pair(v32, &cmode, &imm8);
2825        if (i) {
2826            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2827            tcg_out_vmovi(s, rd, q, 0, i | 1, extract32(v32, i * 4, 8));
2828            return;
2829        }
2830        i = is_shimm32_pair(~v32, &cmode, &imm8);
2831        if (i) {
2832            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2833            tcg_out_vmovi(s, rd, q, 1, i | 1, extract32(~v32, i * 4, 8));
2834            return;
2835        }
2836    }
2837
2838    /*
2839     * As a last resort, load from the constant pool.
2840     */
2841    if (!q || vece == MO_64) {
2842        new_pool_l2(s, R_ARM_PC11, s->code_ptr, 0, v64, v64 >> 32);
2843        /* VLDR Dd, [pc + offset] */
2844        tcg_out32(s, INSN_VLDR_D | encode_vd(rd) | (0xf << 16));
2845        if (q) {
2846            tcg_out_dup2_vec(s, rd, rd, rd);
2847        }
2848    } else {
2849        new_pool_label(s, (uint32_t)v64, R_ARM_PC8, s->code_ptr, 0);
2850        /* add tmp, pc, offset */
2851        tcg_out_dat_imm(s, COND_AL, ARITH_ADD, TCG_REG_TMP, TCG_REG_PC, 0);
2852        tcg_out_dupm_vec(s, type, MO_32, rd, TCG_REG_TMP, 0);
2853    }
2854}
2855
2856static const ARMInsn vec_cmp_insn[16] = {
2857    [TCG_COND_EQ] = INSN_VCEQ,
2858    [TCG_COND_GT] = INSN_VCGT,
2859    [TCG_COND_GE] = INSN_VCGE,
2860    [TCG_COND_GTU] = INSN_VCGT_U,
2861    [TCG_COND_GEU] = INSN_VCGE_U,
2862};
2863
2864static const ARMInsn vec_cmp0_insn[16] = {
2865    [TCG_COND_EQ] = INSN_VCEQ0,
2866    [TCG_COND_GT] = INSN_VCGT0,
2867    [TCG_COND_GE] = INSN_VCGE0,
2868    [TCG_COND_LT] = INSN_VCLT0,
2869    [TCG_COND_LE] = INSN_VCLE0,
2870};
2871
2872static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2873                           unsigned vecl, unsigned vece,
2874                           const TCGArg args[TCG_MAX_OP_ARGS],
2875                           const int const_args[TCG_MAX_OP_ARGS])
2876{
2877    TCGType type = vecl + TCG_TYPE_V64;
2878    unsigned q = vecl;
2879    TCGArg a0, a1, a2, a3;
2880    int cmode, imm8;
2881
2882    a0 = args[0];
2883    a1 = args[1];
2884    a2 = args[2];
2885
2886    switch (opc) {
2887    case INDEX_op_ld_vec:
2888        tcg_out_ld(s, type, a0, a1, a2);
2889        return;
2890    case INDEX_op_st_vec:
2891        tcg_out_st(s, type, a0, a1, a2);
2892        return;
2893    case INDEX_op_dupm_vec:
2894        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
2895        return;
2896    case INDEX_op_dup2_vec:
2897        tcg_out_dup2_vec(s, a0, a1, a2);
2898        return;
2899    case INDEX_op_abs_vec:
2900        tcg_out_vreg2(s, INSN_VABS, q, vece, a0, a1);
2901        return;
2902    case INDEX_op_neg_vec:
2903        tcg_out_vreg2(s, INSN_VNEG, q, vece, a0, a1);
2904        return;
2905    case INDEX_op_not_vec:
2906        tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a1);
2907        return;
2908    case INDEX_op_add_vec:
2909        tcg_out_vreg3(s, INSN_VADD, q, vece, a0, a1, a2);
2910        return;
2911    case INDEX_op_mul_vec:
2912        tcg_out_vreg3(s, INSN_VMUL, q, vece, a0, a1, a2);
2913        return;
2914    case INDEX_op_smax_vec:
2915        tcg_out_vreg3(s, INSN_VMAX, q, vece, a0, a1, a2);
2916        return;
2917    case INDEX_op_smin_vec:
2918        tcg_out_vreg3(s, INSN_VMIN, q, vece, a0, a1, a2);
2919        return;
2920    case INDEX_op_sub_vec:
2921        tcg_out_vreg3(s, INSN_VSUB, q, vece, a0, a1, a2);
2922        return;
2923    case INDEX_op_ssadd_vec:
2924        tcg_out_vreg3(s, INSN_VQADD, q, vece, a0, a1, a2);
2925        return;
2926    case INDEX_op_sssub_vec:
2927        tcg_out_vreg3(s, INSN_VQSUB, q, vece, a0, a1, a2);
2928        return;
2929    case INDEX_op_umax_vec:
2930        tcg_out_vreg3(s, INSN_VMAX_U, q, vece, a0, a1, a2);
2931        return;
2932    case INDEX_op_umin_vec:
2933        tcg_out_vreg3(s, INSN_VMIN_U, q, vece, a0, a1, a2);
2934        return;
2935    case INDEX_op_usadd_vec:
2936        tcg_out_vreg3(s, INSN_VQADD_U, q, vece, a0, a1, a2);
2937        return;
2938    case INDEX_op_ussub_vec:
2939        tcg_out_vreg3(s, INSN_VQSUB_U, q, vece, a0, a1, a2);
2940        return;
2941    case INDEX_op_xor_vec:
2942        tcg_out_vreg3(s, INSN_VEOR, q, 0, a0, a1, a2);
2943        return;
2944    case INDEX_op_arm_sshl_vec:
2945        /*
2946         * Note that Vm is the data and Vn is the shift count,
2947         * therefore the arguments appear reversed.
2948         */
2949        tcg_out_vreg3(s, INSN_VSHL_S, q, vece, a0, a2, a1);
2950        return;
2951    case INDEX_op_arm_ushl_vec:
2952        /* See above. */
2953        tcg_out_vreg3(s, INSN_VSHL_U, q, vece, a0, a2, a1);
2954        return;
2955    case INDEX_op_shli_vec:
2956        tcg_out_vshifti(s, INSN_VSHLI, q, a0, a1, a2 + (8 << vece));
2957        return;
2958    case INDEX_op_shri_vec:
2959        tcg_out_vshifti(s, INSN_VSHRI, q, a0, a1, (16 << vece) - a2);
2960        return;
2961    case INDEX_op_sari_vec:
2962        tcg_out_vshifti(s, INSN_VSARI, q, a0, a1, (16 << vece) - a2);
2963        return;
2964    case INDEX_op_arm_sli_vec:
2965        tcg_out_vshifti(s, INSN_VSLI, q, a0, a2, args[3] + (8 << vece));
2966        return;
2967
2968    case INDEX_op_andc_vec:
2969        if (!const_args[2]) {
2970            tcg_out_vreg3(s, INSN_VBIC, q, 0, a0, a1, a2);
2971            return;
2972        }
2973        a2 = ~a2;
2974        /* fall through */
2975    case INDEX_op_and_vec:
2976        if (const_args[2]) {
2977            is_shimm1632(~a2, &cmode, &imm8);
2978            if (a0 == a1) {
2979                tcg_out_vmovi(s, a0, q, 1, cmode | 1, imm8); /* VBICI */
2980                return;
2981            }
2982            tcg_out_vmovi(s, a0, q, 1, cmode, imm8); /* VMVNI */
2983            a2 = a0;
2984        }
2985        tcg_out_vreg3(s, INSN_VAND, q, 0, a0, a1, a2);
2986        return;
2987
2988    case INDEX_op_orc_vec:
2989        if (!const_args[2]) {
2990            tcg_out_vreg3(s, INSN_VORN, q, 0, a0, a1, a2);
2991            return;
2992        }
2993        a2 = ~a2;
2994        /* fall through */
2995    case INDEX_op_or_vec:
2996        if (const_args[2]) {
2997            is_shimm1632(a2, &cmode, &imm8);
2998            if (a0 == a1) {
2999                tcg_out_vmovi(s, a0, q, 0, cmode | 1, imm8); /* VORRI */
3000                return;
3001            }
3002            tcg_out_vmovi(s, a0, q, 0, cmode, imm8); /* VMOVI */
3003            a2 = a0;
3004        }
3005        tcg_out_vreg3(s, INSN_VORR, q, 0, a0, a1, a2);
3006        return;
3007
3008    case INDEX_op_cmp_vec:
3009        {
3010            TCGCond cond = args[3];
3011            ARMInsn insn;
3012
3013            switch (cond) {
3014            case TCG_COND_NE:
3015                if (const_args[2]) {
3016                    tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a1);
3017                } else {
3018                    tcg_out_vreg3(s, INSN_VCEQ, q, vece, a0, a1, a2);
3019                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
3020                }
3021                break;
3022
3023            case TCG_COND_TSTNE:
3024            case TCG_COND_TSTEQ:
3025                if (const_args[2]) {
3026                    /* (x & 0) == 0 */
3027                    tcg_out_dupi_vec(s, type, MO_8, a0,
3028                                     -(cond == TCG_COND_TSTEQ));
3029                    break;
3030                }
3031                tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a2);
3032                if (cond == TCG_COND_TSTEQ) {
3033                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
3034                }
3035                break;
3036
3037            default:
3038                if (const_args[2]) {
3039                    insn = vec_cmp0_insn[cond];
3040                    if (insn) {
3041                        tcg_out_vreg2(s, insn, q, vece, a0, a1);
3042                        return;
3043                    }
3044                    tcg_out_dupi_vec(s, type, MO_8, TCG_VEC_TMP, 0);
3045                    a2 = TCG_VEC_TMP;
3046                }
3047                insn = vec_cmp_insn[cond];
3048                if (insn == 0) {
3049                    TCGArg t;
3050                    t = a1, a1 = a2, a2 = t;
3051                    cond = tcg_swap_cond(cond);
3052                    insn = vec_cmp_insn[cond];
3053                    tcg_debug_assert(insn != 0);
3054                }
3055                tcg_out_vreg3(s, insn, q, vece, a0, a1, a2);
3056                break;
3057            }
3058        }
3059        return;
3060
3061    case INDEX_op_bitsel_vec:
3062        a3 = args[3];
3063        if (a0 == a3) {
3064            tcg_out_vreg3(s, INSN_VBIT, q, 0, a0, a2, a1);
3065        } else if (a0 == a2) {
3066            tcg_out_vreg3(s, INSN_VBIF, q, 0, a0, a3, a1);
3067        } else {
3068            tcg_out_mov(s, type, a0, a1);
3069            tcg_out_vreg3(s, INSN_VBSL, q, 0, a0, a2, a3);
3070        }
3071        return;
3072
3073    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
3074    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
3075    default:
3076        g_assert_not_reached();
3077    }
3078}
3079
3080int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
3081{
3082    switch (opc) {
3083    case INDEX_op_add_vec:
3084    case INDEX_op_sub_vec:
3085    case INDEX_op_and_vec:
3086    case INDEX_op_andc_vec:
3087    case INDEX_op_or_vec:
3088    case INDEX_op_orc_vec:
3089    case INDEX_op_xor_vec:
3090    case INDEX_op_not_vec:
3091    case INDEX_op_shli_vec:
3092    case INDEX_op_shri_vec:
3093    case INDEX_op_sari_vec:
3094    case INDEX_op_ssadd_vec:
3095    case INDEX_op_sssub_vec:
3096    case INDEX_op_usadd_vec:
3097    case INDEX_op_ussub_vec:
3098    case INDEX_op_bitsel_vec:
3099        return 1;
3100    case INDEX_op_abs_vec:
3101    case INDEX_op_cmp_vec:
3102    case INDEX_op_mul_vec:
3103    case INDEX_op_neg_vec:
3104    case INDEX_op_smax_vec:
3105    case INDEX_op_smin_vec:
3106    case INDEX_op_umax_vec:
3107    case INDEX_op_umin_vec:
3108        return vece < MO_64;
3109    case INDEX_op_shlv_vec:
3110    case INDEX_op_shrv_vec:
3111    case INDEX_op_sarv_vec:
3112    case INDEX_op_rotli_vec:
3113    case INDEX_op_rotlv_vec:
3114    case INDEX_op_rotrv_vec:
3115        return -1;
3116    default:
3117        return 0;
3118    }
3119}
3120
3121void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
3122                       TCGArg a0, ...)
3123{
3124    va_list va;
3125    TCGv_vec v0, v1, v2, t1, t2, c1;
3126    TCGArg a2;
3127
3128    va_start(va, a0);
3129    v0 = temp_tcgv_vec(arg_temp(a0));
3130    v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
3131    a2 = va_arg(va, TCGArg);
3132    va_end(va);
3133
3134    switch (opc) {
3135    case INDEX_op_shlv_vec:
3136        /*
3137         * Merely propagate shlv_vec to arm_ushl_vec.
3138         * In this way we don't set TCG_TARGET_HAS_shv_vec
3139         * because everything is done via expansion.
3140         */
3141        v2 = temp_tcgv_vec(arg_temp(a2));
3142        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
3143                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
3144        break;
3145
3146    case INDEX_op_shrv_vec:
3147    case INDEX_op_sarv_vec:
3148        /* Right shifts are negative left shifts for NEON.  */
3149        v2 = temp_tcgv_vec(arg_temp(a2));
3150        t1 = tcg_temp_new_vec(type);
3151        tcg_gen_neg_vec(vece, t1, v2);
3152        if (opc == INDEX_op_shrv_vec) {
3153            opc = INDEX_op_arm_ushl_vec;
3154        } else {
3155            opc = INDEX_op_arm_sshl_vec;
3156        }
3157        vec_gen_3(opc, type, vece, tcgv_vec_arg(v0),
3158                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
3159        tcg_temp_free_vec(t1);
3160        break;
3161
3162    case INDEX_op_rotli_vec:
3163        t1 = tcg_temp_new_vec(type);
3164        tcg_gen_shri_vec(vece, t1, v1, -a2 & ((8 << vece) - 1));
3165        vec_gen_4(INDEX_op_arm_sli_vec, type, vece,
3166                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(v1), a2);
3167        tcg_temp_free_vec(t1);
3168        break;
3169
3170    case INDEX_op_rotlv_vec:
3171        v2 = temp_tcgv_vec(arg_temp(a2));
3172        t1 = tcg_temp_new_vec(type);
3173        c1 = tcg_constant_vec(type, vece, 8 << vece);
3174        tcg_gen_sub_vec(vece, t1, v2, c1);
3175        /* Right shifts are negative left shifts for NEON.  */
3176        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
3177                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
3178        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
3179                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
3180        tcg_gen_or_vec(vece, v0, v0, t1);
3181        tcg_temp_free_vec(t1);
3182        break;
3183
3184    case INDEX_op_rotrv_vec:
3185        v2 = temp_tcgv_vec(arg_temp(a2));
3186        t1 = tcg_temp_new_vec(type);
3187        t2 = tcg_temp_new_vec(type);
3188        c1 = tcg_constant_vec(type, vece, 8 << vece);
3189        tcg_gen_neg_vec(vece, t1, v2);
3190        tcg_gen_sub_vec(vece, t2, c1, v2);
3191        /* Right shifts are negative left shifts for NEON.  */
3192        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
3193                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
3194        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t2),
3195                  tcgv_vec_arg(v1), tcgv_vec_arg(t2));
3196        tcg_gen_or_vec(vece, v0, t1, t2);
3197        tcg_temp_free_vec(t1);
3198        tcg_temp_free_vec(t2);
3199        break;
3200
3201    default:
3202        g_assert_not_reached();
3203    }
3204}
3205
3206static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
3207{
3208    int i;
3209    for (i = 0; i < count; ++i) {
3210        p[i] = INSN_NOP;
3211    }
3212}
3213
3214/* Compute frame size via macros, to share between tcg_target_qemu_prologue
3215   and tcg_register_jit.  */
3216
3217#define PUSH_SIZE  ((11 - 4 + 1 + 1) * sizeof(tcg_target_long))
3218
3219#define FRAME_SIZE \
3220    ((PUSH_SIZE \
3221      + TCG_STATIC_CALL_ARGS_SIZE \
3222      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
3223      + TCG_TARGET_STACK_ALIGN - 1) \
3224     & -TCG_TARGET_STACK_ALIGN)
3225
3226#define STACK_ADDEND  (FRAME_SIZE - PUSH_SIZE)
3227
3228static void tcg_target_qemu_prologue(TCGContext *s)
3229{
3230    /* Calling convention requires us to save r4-r11 and lr.  */
3231    /* stmdb sp!, { r4 - r11, lr } */
3232    tcg_out_ldstm(s, COND_AL, INSN_STMDB, TCG_REG_CALL_STACK,
3233                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
3234                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3235                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_R14));
3236
3237    /* Reserve callee argument and tcg temp space.  */
3238    tcg_out_dat_rI(s, COND_AL, ARITH_SUB, TCG_REG_CALL_STACK,
3239                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
3240    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
3241                  CPU_TEMP_BUF_NLONGS * sizeof(long));
3242
3243    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
3244
3245    if (!tcg_use_softmmu && guest_base) {
3246        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_GUEST_BASE, guest_base);
3247        tcg_regset_set_reg(s->reserved_regs, TCG_REG_GUEST_BASE);
3248    }
3249
3250    tcg_out_b_reg(s, COND_AL, tcg_target_call_iarg_regs[1]);
3251
3252    /*
3253     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
3254     * and fall through to the rest of the epilogue.
3255     */
3256    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
3257    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, 0);
3258    tcg_out_epilogue(s);
3259}
3260
3261static void tcg_out_epilogue(TCGContext *s)
3262{
3263    /* Release local stack frame.  */
3264    tcg_out_dat_rI(s, COND_AL, ARITH_ADD, TCG_REG_CALL_STACK,
3265                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
3266
3267    /* ldmia sp!, { r4 - r11, pc } */
3268    tcg_out_ldstm(s, COND_AL, INSN_LDMIA, TCG_REG_CALL_STACK,
3269                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
3270                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3271                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_PC));
3272}
3273
3274static void tcg_out_tb_start(TCGContext *s)
3275{
3276    /* nothing to do */
3277}
3278
3279typedef struct {
3280    DebugFrameHeader h;
3281    uint8_t fde_def_cfa[4];
3282    uint8_t fde_reg_ofs[18];
3283} DebugFrame;
3284
3285#define ELF_HOST_MACHINE EM_ARM
3286
3287/* We're expecting a 2 byte uleb128 encoded value.  */
3288QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
3289
3290static const DebugFrame debug_frame = {
3291    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
3292    .h.cie.id = -1,
3293    .h.cie.version = 1,
3294    .h.cie.code_align = 1,
3295    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
3296    .h.cie.return_column = 14,
3297
3298    /* Total FDE size does not include the "len" member.  */
3299    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
3300
3301    .fde_def_cfa = {
3302        12, 13,                         /* DW_CFA_def_cfa sp, ... */
3303        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
3304        (FRAME_SIZE >> 7)
3305    },
3306    .fde_reg_ofs = {
3307        /* The following must match the stmdb in the prologue.  */
3308        0x8e, 1,                        /* DW_CFA_offset, lr, -4 */
3309        0x8b, 2,                        /* DW_CFA_offset, r11, -8 */
3310        0x8a, 3,                        /* DW_CFA_offset, r10, -12 */
3311        0x89, 4,                        /* DW_CFA_offset, r9, -16 */
3312        0x88, 5,                        /* DW_CFA_offset, r8, -20 */
3313        0x87, 6,                        /* DW_CFA_offset, r7, -24 */
3314        0x86, 7,                        /* DW_CFA_offset, r6, -28 */
3315        0x85, 8,                        /* DW_CFA_offset, r5, -32 */
3316        0x84, 9,                        /* DW_CFA_offset, r4, -36 */
3317    }
3318};
3319
3320void tcg_register_jit(const void *buf, size_t buf_size)
3321{
3322    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
3323}
3324