xref: /openbmc/qemu/tcg/arm/tcg-target.c.inc (revision a363e1e179445102d7940e92d394d6c00c126f13)
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 tcg_out_bswap16(TCGContext *s, ARMCond cond,
973                            TCGReg rd, TCGReg rn, int flags)
974{
975    if (flags & TCG_BSWAP_OS) {
976        /* revsh */
977        tcg_out32(s, 0x06ff0fb0 | (cond << 28) | (rd << 12) | rn);
978        return;
979    }
980
981    /* rev16 */
982    tcg_out32(s, 0x06bf0fb0 | (cond << 28) | (rd << 12) | rn);
983    if ((flags & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
984        /* uxth */
985        tcg_out32(s, 0x06ff0070 | (cond << 28) | (rd << 12) | rd);
986    }
987}
988
989static void tcg_out_bswap32(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rn)
990{
991    /* rev */
992    tcg_out32(s, 0x06bf0f30 | (cond << 28) | (rd << 12) | rn);
993}
994
995static void tcg_out_deposit(TCGContext *s, ARMCond cond, TCGReg rd,
996                            TCGArg a1, int ofs, int len, bool const_a1)
997{
998    if (const_a1) {
999        /* bfi becomes bfc with rn == 15.  */
1000        a1 = 15;
1001    }
1002    /* bfi/bfc */
1003    tcg_out32(s, 0x07c00010 | (cond << 28) | (rd << 12) | a1
1004              | (ofs << 7) | ((ofs + len - 1) << 16));
1005}
1006
1007static void tcg_out_extract(TCGContext *s, ARMCond cond, TCGReg rd,
1008                            TCGReg rn, int ofs, int len)
1009{
1010    /* According to gcc, AND can be faster. */
1011    if (ofs == 0 && len <= 8) {
1012        tcg_out_dat_imm(s, cond, ARITH_AND, rd, rn,
1013                        encode_imm_nofail((1 << len) - 1));
1014        return;
1015    }
1016
1017    if (use_armv7_instructions) {
1018        /* ubfx */
1019        tcg_out32(s, 0x07e00050 | (cond << 28) | (rd << 12) | rn
1020                  | (ofs << 7) | ((len - 1) << 16));
1021        return;
1022    }
1023
1024    assert(ofs % 8 == 0);
1025    switch (len) {
1026    case 8:
1027        /* uxtb */
1028        tcg_out32(s, 0x06ef0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1029        break;
1030    case 16:
1031        /* uxth */
1032        tcg_out32(s, 0x06ff0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1033        break;
1034    default:
1035        g_assert_not_reached();
1036    }
1037}
1038
1039static void tcg_out_sextract(TCGContext *s, ARMCond cond, TCGReg rd,
1040                             TCGReg rn, int ofs, int len)
1041{
1042    if (use_armv7_instructions) {
1043        /* sbfx */
1044        tcg_out32(s, 0x07a00050 | (cond << 28) | (rd << 12) | rn
1045                  | (ofs << 7) | ((len - 1) << 16));
1046        return;
1047    }
1048
1049    assert(ofs % 8 == 0);
1050    switch (len) {
1051    case 8:
1052        /* sxtb */
1053        tcg_out32(s, 0x06af0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1054        break;
1055    case 16:
1056        /* sxth */
1057        tcg_out32(s, 0x06bf0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1058        break;
1059    default:
1060        g_assert_not_reached();
1061    }
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, const TCGArg *args,
1258                            const int *const_args)
1259{
1260    TCGReg al = args[0];
1261    TCGReg ah = args[1];
1262    TCGArg bl = args[2];
1263    TCGArg bh = args[3];
1264    TCGCond cond = args[4];
1265    int const_bl = const_args[2];
1266    int const_bh = const_args[3];
1267
1268    switch (cond) {
1269    case TCG_COND_EQ:
1270    case TCG_COND_NE:
1271    case TCG_COND_LTU:
1272    case TCG_COND_LEU:
1273    case TCG_COND_GTU:
1274    case TCG_COND_GEU:
1275        /*
1276         * We perform a conditional comparison.  If the high half is
1277         * equal, then overwrite the flags with the comparison of the
1278         * low half.  The resulting flags cover the whole.
1279         */
1280        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, ah, bh, const_bh);
1281        tcg_out_dat_rI(s, COND_EQ, ARITH_CMP, 0, al, bl, const_bl);
1282        return cond;
1283
1284    case TCG_COND_TSTEQ:
1285    case TCG_COND_TSTNE:
1286        /* Similar, but with TST instead of CMP. */
1287        tcg_out_dat_rI(s, COND_AL, ARITH_TST, 0, ah, bh, const_bh);
1288        tcg_out_dat_rI(s, COND_EQ, ARITH_TST, 0, al, bl, const_bl);
1289        return tcg_tst_eqne_cond(cond);
1290
1291    case TCG_COND_LT:
1292    case TCG_COND_GE:
1293        /* We perform a double-word subtraction and examine the result.
1294           We do not actually need the result of the subtract, so the
1295           low part "subtract" is a compare.  For the high half we have
1296           no choice but to compute into a temporary.  */
1297        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, al, bl, const_bl);
1298        tcg_out_dat_rI(s, COND_AL, ARITH_SBC | TO_CPSR,
1299                       TCG_REG_TMP, ah, bh, const_bh);
1300        return cond;
1301
1302    case TCG_COND_LE:
1303    case TCG_COND_GT:
1304        /* Similar, but with swapped arguments, via reversed subtract.  */
1305        tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR,
1306                       TCG_REG_TMP, al, bl, const_bl);
1307        tcg_out_dat_rI(s, COND_AL, ARITH_RSC | TO_CPSR,
1308                       TCG_REG_TMP, ah, bh, const_bh);
1309        return tcg_swap_cond(cond);
1310
1311    default:
1312        g_assert_not_reached();
1313    }
1314}
1315
1316/*
1317 * Note that TCGReg references Q-registers.
1318 * Q-regno = 2 * D-regno, so shift left by 1 while inserting.
1319 */
1320static uint32_t encode_vd(TCGReg rd)
1321{
1322    tcg_debug_assert(rd >= TCG_REG_Q0);
1323    return (extract32(rd, 3, 1) << 22) | (extract32(rd, 0, 3) << 13);
1324}
1325
1326static uint32_t encode_vn(TCGReg rn)
1327{
1328    tcg_debug_assert(rn >= TCG_REG_Q0);
1329    return (extract32(rn, 3, 1) << 7) | (extract32(rn, 0, 3) << 17);
1330}
1331
1332static uint32_t encode_vm(TCGReg rm)
1333{
1334    tcg_debug_assert(rm >= TCG_REG_Q0);
1335    return (extract32(rm, 3, 1) << 5) | (extract32(rm, 0, 3) << 1);
1336}
1337
1338static void tcg_out_vreg2(TCGContext *s, ARMInsn insn, int q, int vece,
1339                          TCGReg d, TCGReg m)
1340{
1341    tcg_out32(s, insn | (vece << 18) | (q << 6) |
1342              encode_vd(d) | encode_vm(m));
1343}
1344
1345static void tcg_out_vreg3(TCGContext *s, ARMInsn insn, int q, int vece,
1346                          TCGReg d, TCGReg n, TCGReg m)
1347{
1348    tcg_out32(s, insn | (vece << 20) | (q << 6) |
1349              encode_vd(d) | encode_vn(n) | encode_vm(m));
1350}
1351
1352static void tcg_out_vmovi(TCGContext *s, TCGReg rd,
1353                          int q, int op, int cmode, uint8_t imm8)
1354{
1355    tcg_out32(s, INSN_VMOVI | encode_vd(rd) | (q << 6) | (op << 5)
1356              | (cmode << 8) | extract32(imm8, 0, 4)
1357              | (extract32(imm8, 4, 3) << 16)
1358              | (extract32(imm8, 7, 1) << 24));
1359}
1360
1361static void tcg_out_vshifti(TCGContext *s, ARMInsn insn, int q,
1362                            TCGReg rd, TCGReg rm, int l_imm6)
1363{
1364    tcg_out32(s, insn | (q << 6) | encode_vd(rd) | encode_vm(rm) |
1365              (extract32(l_imm6, 6, 1) << 7) |
1366              (extract32(l_imm6, 0, 6) << 16));
1367}
1368
1369static void tcg_out_vldst(TCGContext *s, ARMInsn insn,
1370                          TCGReg rd, TCGReg rn, int offset)
1371{
1372    if (offset != 0) {
1373        if (check_fit_imm(offset) || check_fit_imm(-offset)) {
1374            tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
1375                            TCG_REG_TMP, rn, offset, true);
1376        } else {
1377            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP, offset);
1378            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1379                            TCG_REG_TMP, TCG_REG_TMP, rn, 0);
1380        }
1381        rn = TCG_REG_TMP;
1382    }
1383    tcg_out32(s, insn | (rn << 16) | encode_vd(rd) | 0xf);
1384}
1385
1386typedef struct {
1387    ARMCond cond;
1388    TCGReg base;
1389    int index;
1390    bool index_scratch;
1391    TCGAtomAlign aa;
1392} HostAddress;
1393
1394bool tcg_target_has_memory_bswap(MemOp memop)
1395{
1396    return false;
1397}
1398
1399static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
1400{
1401    /* We arrive at the slow path via "BLNE", so R14 contains l->raddr. */
1402    return TCG_REG_R14;
1403}
1404
1405static const TCGLdstHelperParam ldst_helper_param = {
1406    .ra_gen = ldst_ra_gen,
1407    .ntmp = 1,
1408    .tmp = { TCG_REG_TMP },
1409};
1410
1411static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1412{
1413    MemOp opc = get_memop(lb->oi);
1414
1415    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1416        return false;
1417    }
1418
1419    tcg_out_ld_helper_args(s, lb, &ldst_helper_param);
1420    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE]);
1421    tcg_out_ld_helper_ret(s, lb, false, &ldst_helper_param);
1422
1423    tcg_out_goto(s, COND_AL, lb->raddr);
1424    return true;
1425}
1426
1427static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1428{
1429    MemOp opc = get_memop(lb->oi);
1430
1431    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1432        return false;
1433    }
1434
1435    tcg_out_st_helper_args(s, lb, &ldst_helper_param);
1436
1437    /* Tail-call to the helper, which will return to the fast path.  */
1438    tcg_out_goto(s, COND_AL, qemu_st_helpers[opc & MO_SIZE]);
1439    return true;
1440}
1441
1442/* We expect to use an 9-bit sign-magnitude negative offset from ENV.  */
1443#define MIN_TLB_MASK_TABLE_OFS  -256
1444
1445static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
1446                                           TCGReg addr, MemOpIdx oi, bool is_ld)
1447{
1448    TCGLabelQemuLdst *ldst = NULL;
1449    MemOp opc = get_memop(oi);
1450    unsigned a_mask;
1451
1452    if (tcg_use_softmmu) {
1453        *h = (HostAddress){
1454            .cond = COND_AL,
1455            .base = addr,
1456            .index = TCG_REG_R1,
1457            .index_scratch = true,
1458        };
1459    } else {
1460        *h = (HostAddress){
1461            .cond = COND_AL,
1462            .base = addr,
1463            .index = guest_base ? TCG_REG_GUEST_BASE : -1,
1464            .index_scratch = false,
1465        };
1466    }
1467
1468    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, false);
1469    a_mask = (1 << h->aa.align) - 1;
1470
1471    if (tcg_use_softmmu) {
1472        int mem_index = get_mmuidx(oi);
1473        int cmp_off = is_ld ? offsetof(CPUTLBEntry, addr_read)
1474                            : offsetof(CPUTLBEntry, addr_write);
1475        int fast_off = tlb_mask_table_ofs(s, mem_index);
1476        unsigned s_mask = (1 << (opc & MO_SIZE)) - 1;
1477        TCGReg t_addr;
1478
1479        ldst = new_ldst_label(s);
1480        ldst->is_ld = is_ld;
1481        ldst->oi = oi;
1482        ldst->addr_reg = addr;
1483
1484        /* Load cpu->neg.tlb.f[mmu_idx].{mask,table} into {r0,r1}.  */
1485        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, mask) != 0);
1486        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, table) != 4);
1487        tcg_out_ldrd_8(s, COND_AL, TCG_REG_R0, TCG_AREG0, fast_off);
1488
1489        /* Extract the tlb index from the address into R0.  */
1490        tcg_out_dat_reg(s, COND_AL, ARITH_AND, TCG_REG_R0, TCG_REG_R0, addr,
1491                        SHIFT_IMM_LSR(s->page_bits - CPU_TLB_ENTRY_BITS));
1492
1493        /*
1494         * Add the tlb_table pointer, creating the CPUTLBEntry address in R1.
1495         * Load the tlb comparator into R2 and the fast path addend into R1.
1496         */
1497        if (cmp_off == 0) {
1498            tcg_out_ld32_rwb(s, COND_AL, TCG_REG_R2, TCG_REG_R1, TCG_REG_R0);
1499        } else {
1500            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1501                            TCG_REG_R1, TCG_REG_R1, TCG_REG_R0, 0);
1502            tcg_out_ld32_12(s, COND_AL, TCG_REG_R2, TCG_REG_R1, cmp_off);
1503        }
1504
1505        /* Load the tlb addend.  */
1506        tcg_out_ld32_12(s, COND_AL, TCG_REG_R1, TCG_REG_R1,
1507                        offsetof(CPUTLBEntry, addend));
1508
1509        /*
1510         * Check alignment, check comparators.
1511         * Do this in 2-4 insns.  Use MOVW for v7, if possible,
1512         * to reduce the number of sequential conditional instructions.
1513         * Almost all guests have at least 4k pages, which means that we need
1514         * to clear at least 9 bits even for an 8-byte memory, which means it
1515         * isn't worth checking for an immediate operand for BIC.
1516         *
1517         * For unaligned accesses, test the page of the last unit of alignment.
1518         * This leaves the least significant alignment bits unchanged, and of
1519         * course must be zero.
1520         */
1521        t_addr = addr;
1522        if (a_mask < s_mask) {
1523            t_addr = TCG_REG_R0;
1524            tcg_out_dat_imm(s, COND_AL, ARITH_ADD, t_addr,
1525                            addr, s_mask - a_mask);
1526        }
1527        if (use_armv7_instructions && s->page_bits <= 16) {
1528            tcg_out_movi32(s, COND_AL, TCG_REG_TMP, ~(s->page_mask | a_mask));
1529            tcg_out_dat_reg(s, COND_AL, ARITH_BIC, TCG_REG_TMP,
1530                            t_addr, TCG_REG_TMP, 0);
1531            tcg_out_dat_reg(s, COND_AL, ARITH_CMP, 0,
1532                            TCG_REG_R2, TCG_REG_TMP, 0);
1533        } else {
1534            if (a_mask) {
1535                tcg_debug_assert(a_mask <= 0xff);
1536                tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addr, a_mask);
1537            }
1538            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0, t_addr,
1539                            SHIFT_IMM_LSR(s->page_bits));
1540            tcg_out_dat_reg(s, (a_mask ? COND_EQ : COND_AL), ARITH_CMP,
1541                            0, TCG_REG_R2, TCG_REG_TMP,
1542                            SHIFT_IMM_LSL(s->page_bits));
1543        }
1544    } else if (a_mask) {
1545        ldst = new_ldst_label(s);
1546        ldst->is_ld = is_ld;
1547        ldst->oi = oi;
1548        ldst->addr_reg = addr;
1549
1550        /* We are expecting alignment to max out at 7 */
1551        tcg_debug_assert(a_mask <= 0xff);
1552        /* tst addr, #mask */
1553        tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addr, a_mask);
1554    }
1555
1556    return ldst;
1557}
1558
1559static void tcg_out_qemu_ld_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1560                                   TCGReg datahi, HostAddress h)
1561{
1562    TCGReg base;
1563
1564    /* Byte swapping is left to middle-end expansion. */
1565    tcg_debug_assert((opc & MO_BSWAP) == 0);
1566
1567    switch (opc & MO_SSIZE) {
1568    case MO_UB:
1569        if (h.index < 0) {
1570            tcg_out_ld8_12(s, h.cond, datalo, h.base, 0);
1571        } else {
1572            tcg_out_ld8_r(s, h.cond, datalo, h.base, h.index);
1573        }
1574        break;
1575    case MO_SB:
1576        if (h.index < 0) {
1577            tcg_out_ld8s_8(s, h.cond, datalo, h.base, 0);
1578        } else {
1579            tcg_out_ld8s_r(s, h.cond, datalo, h.base, h.index);
1580        }
1581        break;
1582    case MO_UW:
1583        if (h.index < 0) {
1584            tcg_out_ld16u_8(s, h.cond, datalo, h.base, 0);
1585        } else {
1586            tcg_out_ld16u_r(s, h.cond, datalo, h.base, h.index);
1587        }
1588        break;
1589    case MO_SW:
1590        if (h.index < 0) {
1591            tcg_out_ld16s_8(s, h.cond, datalo, h.base, 0);
1592        } else {
1593            tcg_out_ld16s_r(s, h.cond, datalo, h.base, h.index);
1594        }
1595        break;
1596    case MO_UL:
1597        if (h.index < 0) {
1598            tcg_out_ld32_12(s, h.cond, datalo, h.base, 0);
1599        } else {
1600            tcg_out_ld32_r(s, h.cond, datalo, h.base, h.index);
1601        }
1602        break;
1603    case MO_UQ:
1604        /* We used pair allocation for datalo, so already should be aligned. */
1605        tcg_debug_assert((datalo & 1) == 0);
1606        tcg_debug_assert(datahi == datalo + 1);
1607        /* LDRD requires alignment; double-check that. */
1608        if (memop_alignment_bits(opc) >= MO_64) {
1609            if (h.index < 0) {
1610                tcg_out_ldrd_8(s, h.cond, datalo, h.base, 0);
1611                break;
1612            }
1613            /*
1614             * Rm (the second address op) must not overlap Rt or Rt + 1.
1615             * Since datalo is aligned, we can simplify the test via alignment.
1616             * Flip the two address arguments if that works.
1617             */
1618            if ((h.index & ~1) != datalo) {
1619                tcg_out_ldrd_r(s, h.cond, datalo, h.base, h.index);
1620                break;
1621            }
1622            if ((h.base & ~1) != datalo) {
1623                tcg_out_ldrd_r(s, h.cond, datalo, h.index, h.base);
1624                break;
1625            }
1626        }
1627        if (h.index < 0) {
1628            base = h.base;
1629            if (datalo == h.base) {
1630                tcg_out_mov_reg(s, h.cond, TCG_REG_TMP, base);
1631                base = TCG_REG_TMP;
1632            }
1633        } else if (h.index_scratch) {
1634            tcg_out_ld32_rwb(s, h.cond, datalo, h.index, h.base);
1635            tcg_out_ld32_12(s, h.cond, datahi, h.index, 4);
1636            break;
1637        } else {
1638            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1639                            h.base, h.index, SHIFT_IMM_LSL(0));
1640            base = TCG_REG_TMP;
1641        }
1642        tcg_out_ld32_12(s, h.cond, datalo, base, 0);
1643        tcg_out_ld32_12(s, h.cond, datahi, base, 4);
1644        break;
1645    default:
1646        g_assert_not_reached();
1647    }
1648}
1649
1650static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
1651                            TCGReg addr, MemOpIdx oi, TCGType data_type)
1652{
1653    MemOp opc = get_memop(oi);
1654    TCGLabelQemuLdst *ldst;
1655    HostAddress h;
1656
1657    ldst = prepare_host_addr(s, &h, addr, oi, true);
1658    if (ldst) {
1659        ldst->type = data_type;
1660        ldst->datalo_reg = datalo;
1661        ldst->datahi_reg = datahi;
1662
1663        /*
1664         * This a conditional BL only to load a pointer within this
1665         * opcode into LR for the slow path.  We will not be using
1666         * the value for a tail call.
1667         */
1668        ldst->label_ptr[0] = s->code_ptr;
1669        tcg_out_bl_imm(s, COND_NE, 0);
1670
1671        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1672        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1673    } else {
1674        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1675    }
1676}
1677
1678static void tcg_out_qemu_st_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1679                                   TCGReg datahi, HostAddress h)
1680{
1681    /* Byte swapping is left to middle-end expansion. */
1682    tcg_debug_assert((opc & MO_BSWAP) == 0);
1683
1684    switch (opc & MO_SIZE) {
1685    case MO_8:
1686        if (h.index < 0) {
1687            tcg_out_st8_12(s, h.cond, datalo, h.base, 0);
1688        } else {
1689            tcg_out_st8_r(s, h.cond, datalo, h.base, h.index);
1690        }
1691        break;
1692    case MO_16:
1693        if (h.index < 0) {
1694            tcg_out_st16_8(s, h.cond, datalo, h.base, 0);
1695        } else {
1696            tcg_out_st16_r(s, h.cond, datalo, h.base, h.index);
1697        }
1698        break;
1699    case MO_32:
1700        if (h.index < 0) {
1701            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1702        } else {
1703            tcg_out_st32_r(s, h.cond, datalo, h.base, h.index);
1704        }
1705        break;
1706    case MO_64:
1707        /* We used pair allocation for datalo, so already should be aligned. */
1708        tcg_debug_assert((datalo & 1) == 0);
1709        tcg_debug_assert(datahi == datalo + 1);
1710        /* STRD requires alignment; double-check that. */
1711        if (memop_alignment_bits(opc) >= MO_64) {
1712            if (h.index < 0) {
1713                tcg_out_strd_8(s, h.cond, datalo, h.base, 0);
1714            } else {
1715                tcg_out_strd_r(s, h.cond, datalo, h.base, h.index);
1716            }
1717        } else if (h.index < 0) {
1718            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1719            tcg_out_st32_12(s, h.cond, datahi, h.base, 4);
1720        } else if (h.index_scratch) {
1721            tcg_out_st32_rwb(s, h.cond, datalo, h.index, h.base);
1722            tcg_out_st32_12(s, h.cond, datahi, h.index, 4);
1723        } else {
1724            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1725                            h.base, h.index, SHIFT_IMM_LSL(0));
1726            tcg_out_st32_12(s, h.cond, datalo, TCG_REG_TMP, 0);
1727            tcg_out_st32_12(s, h.cond, datahi, TCG_REG_TMP, 4);
1728        }
1729        break;
1730    default:
1731        g_assert_not_reached();
1732    }
1733}
1734
1735static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
1736                            TCGReg addr, MemOpIdx oi, TCGType data_type)
1737{
1738    MemOp opc = get_memop(oi);
1739    TCGLabelQemuLdst *ldst;
1740    HostAddress h;
1741
1742    ldst = prepare_host_addr(s, &h, addr, oi, false);
1743    if (ldst) {
1744        ldst->type = data_type;
1745        ldst->datalo_reg = datalo;
1746        ldst->datahi_reg = datahi;
1747
1748        h.cond = COND_EQ;
1749        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1750
1751        /* The conditional call is last, as we're going to return here. */
1752        ldst->label_ptr[0] = s->code_ptr;
1753        tcg_out_bl_imm(s, COND_NE, 0);
1754        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1755    } else {
1756        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1757    }
1758}
1759
1760static void tcg_out_epilogue(TCGContext *s);
1761
1762static void tcg_out_exit_tb(TCGContext *s, uintptr_t arg)
1763{
1764    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, arg);
1765    tcg_out_epilogue(s);
1766}
1767
1768static void tcg_out_goto_tb(TCGContext *s, int which)
1769{
1770    uintptr_t i_addr;
1771    intptr_t i_disp;
1772
1773    /* Direct branch will be patched by tb_target_set_jmp_target. */
1774    set_jmp_insn_offset(s, which);
1775    tcg_out32(s, INSN_NOP);
1776
1777    /* When branch is out of range, fall through to indirect. */
1778    i_addr = get_jmp_target_addr(s, which);
1779    i_disp = tcg_pcrel_diff(s, (void *)i_addr) - 8;
1780    tcg_debug_assert(i_disp < 0);
1781    if (i_disp >= -0xfff) {
1782        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_PC, i_disp);
1783    } else {
1784        /*
1785         * The TB is close, but outside the 12 bits addressable by
1786         * the load.  We can extend this to 20 bits with a sub of a
1787         * shifted immediate from pc.
1788         */
1789        int h = -i_disp;
1790        int l = -(h & 0xfff);
1791
1792        h = encode_imm_nofail(h + l);
1793        tcg_out_dat_imm(s, COND_AL, ARITH_SUB, TCG_REG_R0, TCG_REG_PC, h);
1794        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_R0, l);
1795    }
1796    set_jmp_reset_offset(s, which);
1797}
1798
1799void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1800                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1801{
1802    uintptr_t addr = tb->jmp_target_addr[n];
1803    ptrdiff_t offset = addr - (jmp_rx + 8);
1804    tcg_insn_unit insn;
1805
1806    /* Either directly branch, or fall through to indirect branch. */
1807    if (offset == sextract64(offset, 0, 26)) {
1808        /* B <addr> */
1809        insn = deposit32((COND_AL << 28) | INSN_B, 0, 24, offset >> 2);
1810    } else {
1811        insn = INSN_NOP;
1812    }
1813
1814    qatomic_set((uint32_t *)jmp_rw, insn);
1815    flush_idcache_range(jmp_rx, jmp_rw, 4);
1816}
1817
1818
1819static void tgen_add(TCGContext *s, TCGType type,
1820                     TCGReg a0, TCGReg a1, TCGReg a2)
1821{
1822    tcg_out_dat_reg(s, COND_AL, ARITH_ADD, a0, a1, a2, SHIFT_IMM_LSL(0));
1823}
1824
1825static void tgen_addi(TCGContext *s, TCGType type,
1826                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1827{
1828    tcg_out_dat_IN(s, COND_AL, ARITH_ADD, ARITH_SUB, a0, a1, a2);
1829}
1830
1831static const TCGOutOpBinary outop_add = {
1832    .base.static_constraint = C_O1_I2(r, r, rIN),
1833    .out_rrr = tgen_add,
1834    .out_rri = tgen_addi,
1835};
1836
1837static void tgen_and(TCGContext *s, TCGType type,
1838                     TCGReg a0, TCGReg a1, TCGReg a2)
1839{
1840    tcg_out_dat_reg(s, COND_AL, ARITH_AND, a0, a1, a2, SHIFT_IMM_LSL(0));
1841}
1842
1843static void tgen_andi(TCGContext *s, TCGType type,
1844                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1845{
1846    tcg_out_dat_IK(s, COND_AL, ARITH_AND, ARITH_BIC, a0, a1, a2);
1847}
1848
1849static const TCGOutOpBinary outop_and = {
1850    .base.static_constraint = C_O1_I2(r, r, rIK),
1851    .out_rrr = tgen_and,
1852    .out_rri = tgen_andi,
1853};
1854
1855static void tgen_andc(TCGContext *s, TCGType type,
1856                      TCGReg a0, TCGReg a1, TCGReg a2)
1857{
1858    tcg_out_dat_reg(s, COND_AL, ARITH_BIC, a0, a1, a2, SHIFT_IMM_LSL(0));
1859}
1860
1861static const TCGOutOpBinary outop_andc = {
1862    .base.static_constraint = C_O1_I2(r, r, r),
1863    .out_rrr = tgen_andc,
1864};
1865
1866static void tgen_clz(TCGContext *s, TCGType type,
1867                     TCGReg a0, TCGReg a1, TCGReg a2)
1868{
1869    tcg_out_dat_imm(s, COND_AL, ARITH_CMP, 0, a1, 0);
1870    tcg_out_dat_reg(s, COND_NE, INSN_CLZ, a0, 0, a1, 0);
1871    tcg_out_mov_reg(s, COND_EQ, a0, a2);
1872}
1873
1874static void tgen_clzi(TCGContext *s, TCGType type,
1875                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1876{
1877    if (a2 == 32) {
1878        tcg_out_dat_reg(s, COND_AL, INSN_CLZ, a0, 0, a1, 0);
1879    } else {
1880        tcg_out_dat_imm(s, COND_AL, ARITH_CMP, 0, a1, 0);
1881        tcg_out_dat_reg(s, COND_NE, INSN_CLZ, a0, 0, a1, 0);
1882        tcg_out_movi32(s, COND_EQ, a0, a2);
1883    }
1884}
1885
1886static const TCGOutOpBinary outop_clz = {
1887    .base.static_constraint = C_O1_I2(r, r, rIK),
1888    .out_rrr = tgen_clz,
1889    .out_rri = tgen_clzi,
1890};
1891
1892static const TCGOutOpUnary outop_ctpop = {
1893    .base.static_constraint = C_NotImplemented,
1894};
1895
1896static void tgen_ctz(TCGContext *s, TCGType type,
1897                     TCGReg a0, TCGReg a1, TCGReg a2)
1898{
1899    tcg_out_dat_reg(s, COND_AL, INSN_RBIT, TCG_REG_TMP, 0, a1, 0);
1900    tgen_clz(s, TCG_TYPE_I32, a0, TCG_REG_TMP, a2);
1901}
1902
1903static void tgen_ctzi(TCGContext *s, TCGType type,
1904                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1905{
1906    tcg_out_dat_reg(s, COND_AL, INSN_RBIT, TCG_REG_TMP, 0, a1, 0);
1907    tgen_clzi(s, TCG_TYPE_I32, a0, TCG_REG_TMP, a2);
1908}
1909
1910static TCGConstraintSetIndex cset_ctz(TCGType type, unsigned flags)
1911{
1912    return use_armv7_instructions ? C_O1_I2(r, r, rIK) : C_NotImplemented;
1913}
1914
1915static const TCGOutOpBinary outop_ctz = {
1916    .base.static_constraint = C_Dynamic,
1917    .base.dynamic_constraint = cset_ctz,
1918    .out_rrr = tgen_ctz,
1919    .out_rri = tgen_ctzi,
1920};
1921
1922static TCGConstraintSetIndex cset_idiv(TCGType type, unsigned flags)
1923{
1924    return use_idiv_instructions ? C_O1_I2(r, r, r) : C_NotImplemented;
1925}
1926
1927static void tgen_divs(TCGContext *s, TCGType type,
1928                      TCGReg a0, TCGReg a1, TCGReg a2)
1929{
1930    /* sdiv */
1931    tcg_out32(s, 0x0710f010 | (COND_AL << 28) | (a0 << 16) | a1 | (a2 << 8));
1932}
1933
1934static const TCGOutOpBinary outop_divs = {
1935    .base.static_constraint = C_Dynamic,
1936    .base.dynamic_constraint = cset_idiv,
1937    .out_rrr = tgen_divs,
1938};
1939
1940static const TCGOutOpDivRem outop_divs2 = {
1941    .base.static_constraint = C_NotImplemented,
1942};
1943
1944static void tgen_divu(TCGContext *s, TCGType type,
1945                      TCGReg a0, TCGReg a1, TCGReg a2)
1946{
1947    /* udiv */
1948    tcg_out32(s, 0x0730f010 | (COND_AL << 28) | (a0 << 16) | a1 | (a2 << 8));
1949}
1950
1951static const TCGOutOpBinary outop_divu = {
1952    .base.static_constraint = C_Dynamic,
1953    .base.dynamic_constraint = cset_idiv,
1954    .out_rrr = tgen_divu,
1955};
1956
1957static const TCGOutOpDivRem outop_divu2 = {
1958    .base.static_constraint = C_NotImplemented,
1959};
1960
1961static const TCGOutOpBinary outop_eqv = {
1962    .base.static_constraint = C_NotImplemented,
1963};
1964
1965static void tgen_mul(TCGContext *s, TCGType type,
1966                     TCGReg a0, TCGReg a1, TCGReg a2)
1967{
1968    /* mul */
1969    tcg_out32(s, (COND_AL << 28) | 0x90 | (a0 << 16) | (a1 << 8) | a2);
1970}
1971
1972static const TCGOutOpBinary outop_mul = {
1973    .base.static_constraint = C_O1_I2(r, r, r),
1974    .out_rrr = tgen_mul,
1975};
1976
1977static void tgen_muls2(TCGContext *s, TCGType type,
1978                       TCGReg rd0, TCGReg rd1, TCGReg rn, TCGReg rm)
1979{
1980    /* smull */
1981    tcg_out32(s, (COND_AL << 28) | 0x00c00090 |
1982              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
1983}
1984
1985static const TCGOutOpMul2 outop_muls2 = {
1986    .base.static_constraint = C_O2_I2(r, r, r, r),
1987    .out_rrrr = tgen_muls2,
1988};
1989
1990static const TCGOutOpBinary outop_mulsh = {
1991    .base.static_constraint = C_NotImplemented,
1992};
1993
1994static void tgen_mulu2(TCGContext *s, TCGType type,
1995                       TCGReg rd0, TCGReg rd1, TCGReg rn, TCGReg rm)
1996{
1997    /* umull */
1998    tcg_out32(s, (COND_AL << 28) | 0x00800090 |
1999              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
2000}
2001
2002static const TCGOutOpMul2 outop_mulu2 = {
2003    .base.static_constraint = C_O2_I2(r, r, r, r),
2004    .out_rrrr = tgen_mulu2,
2005};
2006
2007static const TCGOutOpBinary outop_muluh = {
2008    .base.static_constraint = C_NotImplemented,
2009};
2010
2011static const TCGOutOpBinary outop_nand = {
2012    .base.static_constraint = C_NotImplemented,
2013};
2014
2015static const TCGOutOpBinary outop_nor = {
2016    .base.static_constraint = C_NotImplemented,
2017};
2018
2019static void tgen_or(TCGContext *s, TCGType type,
2020                     TCGReg a0, TCGReg a1, TCGReg a2)
2021{
2022    tcg_out_dat_reg(s, COND_AL, ARITH_ORR, a0, a1, a2, SHIFT_IMM_LSL(0));
2023}
2024
2025static void tgen_ori(TCGContext *s, TCGType type,
2026                     TCGReg a0, TCGReg a1, tcg_target_long a2)
2027{
2028    tcg_out_dat_imm(s, COND_AL, ARITH_ORR, a0, a1, encode_imm_nofail(a2));
2029}
2030
2031static const TCGOutOpBinary outop_or = {
2032    .base.static_constraint = C_O1_I2(r, r, rI),
2033    .out_rrr = tgen_or,
2034    .out_rri = tgen_ori,
2035};
2036
2037static const TCGOutOpBinary outop_orc = {
2038    .base.static_constraint = C_NotImplemented,
2039};
2040
2041static const TCGOutOpBinary outop_rems = {
2042    .base.static_constraint = C_NotImplemented,
2043};
2044
2045static const TCGOutOpBinary outop_remu = {
2046    .base.static_constraint = C_NotImplemented,
2047};
2048
2049static const TCGOutOpBinary outop_rotl = {
2050    .base.static_constraint = C_NotImplemented,
2051};
2052
2053static void tgen_rotr(TCGContext *s, TCGType type,
2054                      TCGReg a0, TCGReg a1, TCGReg a2)
2055{
2056    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1, SHIFT_REG_ROR(a2));
2057}
2058
2059static void tgen_rotri(TCGContext *s, TCGType type,
2060                       TCGReg a0, TCGReg a1, tcg_target_long a2)
2061{
2062    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1, SHIFT_IMM_ROR(a2 & 0x1f));
2063}
2064
2065static const TCGOutOpBinary outop_rotr = {
2066    .base.static_constraint = C_O1_I2(r, r, ri),
2067    .out_rrr = tgen_rotr,
2068    .out_rri = tgen_rotri,
2069};
2070
2071static void tgen_sar(TCGContext *s, TCGType type,
2072                     TCGReg a0, TCGReg a1, TCGReg a2)
2073{
2074    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1, SHIFT_REG_ASR(a2));
2075}
2076
2077static void tgen_sari(TCGContext *s, TCGType type,
2078                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2079{
2080    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1,
2081                    SHIFT_IMM_ASR(a2 & 0x1f));
2082}
2083
2084static const TCGOutOpBinary outop_sar = {
2085    .base.static_constraint = C_O1_I2(r, r, ri),
2086    .out_rrr = tgen_sar,
2087    .out_rri = tgen_sari,
2088};
2089
2090static void tgen_shl(TCGContext *s, TCGType type,
2091                     TCGReg a0, TCGReg a1, TCGReg a2)
2092{
2093    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1, SHIFT_REG_LSL(a2));
2094}
2095
2096static void tgen_shli(TCGContext *s, TCGType type,
2097                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2098{
2099    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1,
2100                    SHIFT_IMM_LSL(a2 & 0x1f));
2101}
2102
2103static const TCGOutOpBinary outop_shl = {
2104    .base.static_constraint = C_O1_I2(r, r, ri),
2105    .out_rrr = tgen_shl,
2106    .out_rri = tgen_shli,
2107};
2108
2109static void tgen_shr(TCGContext *s, TCGType type,
2110                     TCGReg a0, TCGReg a1, TCGReg a2)
2111{
2112    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1, SHIFT_REG_LSR(a2));
2113}
2114
2115static void tgen_shri(TCGContext *s, TCGType type,
2116                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2117{
2118    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1,
2119                    SHIFT_IMM_LSR(a2 & 0x1f));
2120}
2121
2122static const TCGOutOpBinary outop_shr = {
2123    .base.static_constraint = C_O1_I2(r, r, ri),
2124    .out_rrr = tgen_shr,
2125    .out_rri = tgen_shri,
2126};
2127
2128static void tgen_sub(TCGContext *s, TCGType type,
2129                     TCGReg a0, TCGReg a1, TCGReg a2)
2130{
2131    tcg_out_dat_reg(s, COND_AL, ARITH_SUB, a0, a1, a2, SHIFT_IMM_LSL(0));
2132}
2133
2134static void tgen_subfi(TCGContext *s, TCGType type,
2135                       TCGReg a0, tcg_target_long a1, TCGReg a2)
2136{
2137    tcg_out_dat_imm(s, COND_AL, ARITH_RSB, a0, a2, encode_imm_nofail(a1));
2138}
2139
2140static const TCGOutOpSubtract outop_sub = {
2141    .base.static_constraint = C_O1_I2(r, rI, r),
2142    .out_rrr = tgen_sub,
2143    .out_rir = tgen_subfi,
2144};
2145
2146static void tgen_xor(TCGContext *s, TCGType type,
2147                     TCGReg a0, TCGReg a1, TCGReg a2)
2148{
2149    tcg_out_dat_reg(s, COND_AL, ARITH_EOR, a0, a1, a2, SHIFT_IMM_LSL(0));
2150}
2151
2152static void tgen_xori(TCGContext *s, TCGType type,
2153                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2154{
2155    tcg_out_dat_imm(s, COND_AL, ARITH_EOR, a0, a1, encode_imm_nofail(a2));
2156}
2157
2158static const TCGOutOpBinary outop_xor = {
2159    .base.static_constraint = C_O1_I2(r, r, rI),
2160    .out_rrr = tgen_xor,
2161    .out_rri = tgen_xori,
2162};
2163
2164static void tgen_neg(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
2165{
2166    tgen_subfi(s, type, a0, 0, a1);
2167}
2168
2169static const TCGOutOpUnary outop_neg = {
2170    .base.static_constraint = C_O1_I1(r, r),
2171    .out_rr = tgen_neg,
2172};
2173
2174static void tgen_not(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
2175{
2176    tcg_out_dat_reg(s, COND_AL, ARITH_MVN, a0, 0, a1, SHIFT_IMM_LSL(0));
2177}
2178
2179static const TCGOutOpUnary outop_not = {
2180    .base.static_constraint = C_O1_I1(r, r),
2181    .out_rr = tgen_not,
2182};
2183
2184static void finish_setcond(TCGContext *s, TCGCond cond, TCGReg ret, bool neg)
2185{
2186    tcg_out_movi32(s, tcg_cond_to_arm_cond[tcg_invert_cond(cond)], ret, 0);
2187    tcg_out_movi32(s, tcg_cond_to_arm_cond[cond], ret, neg ? -1 : 1);
2188}
2189
2190static void tgen_setcond(TCGContext *s, TCGType type, TCGCond cond,
2191                         TCGReg a0, TCGReg a1, TCGReg a2)
2192{
2193    cond = tgen_cmp(s, cond, a1, a2);
2194    finish_setcond(s, cond, a0, false);
2195}
2196
2197static void tgen_setcondi(TCGContext *s, TCGType type, TCGCond cond,
2198                          TCGReg a0, TCGReg a1, tcg_target_long a2)
2199{
2200    cond = tgen_cmpi(s, cond, a1, a2);
2201    finish_setcond(s, cond, a0, false);
2202}
2203
2204static const TCGOutOpSetcond outop_setcond = {
2205    .base.static_constraint = C_O1_I2(r, r, rIN),
2206    .out_rrr = tgen_setcond,
2207    .out_rri = tgen_setcondi,
2208};
2209
2210static void tgen_negsetcond(TCGContext *s, TCGType type, TCGCond cond,
2211                            TCGReg a0, TCGReg a1, TCGReg a2)
2212{
2213    cond = tgen_cmp(s, cond, a1, a2);
2214    finish_setcond(s, cond, a0, true);
2215}
2216
2217static void tgen_negsetcondi(TCGContext *s, TCGType type, TCGCond cond,
2218                             TCGReg a0, TCGReg a1, tcg_target_long a2)
2219{
2220    cond = tgen_cmpi(s, cond, a1, a2);
2221    finish_setcond(s, cond, a0, true);
2222}
2223
2224static const TCGOutOpSetcond outop_negsetcond = {
2225    .base.static_constraint = C_O1_I2(r, r, rIN),
2226    .out_rrr = tgen_negsetcond,
2227    .out_rri = tgen_negsetcondi,
2228};
2229
2230
2231static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
2232                       const TCGArg args[TCG_MAX_OP_ARGS],
2233                       const int const_args[TCG_MAX_OP_ARGS])
2234{
2235    TCGArg a0, a1, a2, a3, a4, a5;
2236    int c;
2237
2238    switch (opc) {
2239    case INDEX_op_goto_ptr:
2240        tcg_out_b_reg(s, COND_AL, args[0]);
2241        break;
2242    case INDEX_op_br:
2243        tcg_out_goto_label(s, COND_AL, arg_label(args[0]));
2244        break;
2245
2246    case INDEX_op_ld8u_i32:
2247        tcg_out_ld8u(s, COND_AL, args[0], args[1], args[2]);
2248        break;
2249    case INDEX_op_ld8s_i32:
2250        tcg_out_ld8s(s, COND_AL, args[0], args[1], args[2]);
2251        break;
2252    case INDEX_op_ld16u_i32:
2253        tcg_out_ld16u(s, COND_AL, args[0], args[1], args[2]);
2254        break;
2255    case INDEX_op_ld16s_i32:
2256        tcg_out_ld16s(s, COND_AL, args[0], args[1], args[2]);
2257        break;
2258    case INDEX_op_ld_i32:
2259        tcg_out_ld32u(s, COND_AL, args[0], args[1], args[2]);
2260        break;
2261    case INDEX_op_st8_i32:
2262        tcg_out_st8(s, COND_AL, args[0], args[1], args[2]);
2263        break;
2264    case INDEX_op_st16_i32:
2265        tcg_out_st16(s, COND_AL, args[0], args[1], args[2]);
2266        break;
2267    case INDEX_op_st_i32:
2268        tcg_out_st32(s, COND_AL, args[0], args[1], args[2]);
2269        break;
2270
2271    case INDEX_op_movcond_i32:
2272        /* Constraints mean that v2 is always in the same register as dest,
2273         * so we only need to do "if condition passed, move v1 to dest".
2274         */
2275        c = tcg_out_cmp(s, args[5], args[1], args[2], const_args[2]);
2276        tcg_out_dat_rIK(s, tcg_cond_to_arm_cond[c], ARITH_MOV,
2277                        ARITH_MVN, args[0], 0, args[3], const_args[3]);
2278        break;
2279    case INDEX_op_add2_i32:
2280        a0 = args[0], a1 = args[1], a2 = args[2];
2281        a3 = args[3], a4 = args[4], a5 = args[5];
2282        if (a0 == a3 || (a0 == a5 && !const_args[5])) {
2283            a0 = TCG_REG_TMP;
2284        }
2285        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD | TO_CPSR, ARITH_SUB | TO_CPSR,
2286                        a0, a2, a4, const_args[4]);
2287        tcg_out_dat_rIK(s, COND_AL, ARITH_ADC, ARITH_SBC,
2288                        a1, a3, a5, const_args[5]);
2289        tcg_out_mov_reg(s, COND_AL, args[0], a0);
2290        break;
2291    case INDEX_op_sub2_i32:
2292        a0 = args[0], a1 = args[1], a2 = args[2];
2293        a3 = args[3], a4 = args[4], a5 = args[5];
2294        if ((a0 == a3 && !const_args[3]) || (a0 == a5 && !const_args[5])) {
2295            a0 = TCG_REG_TMP;
2296        }
2297        if (const_args[2]) {
2298            if (const_args[4]) {
2299                tcg_out_movi32(s, COND_AL, a0, a4);
2300                a4 = a0;
2301            }
2302            tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR, a0, a4, a2, 1);
2303        } else {
2304            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB | TO_CPSR,
2305                            ARITH_ADD | TO_CPSR, a0, a2, a4, const_args[4]);
2306        }
2307        if (const_args[3]) {
2308            if (const_args[5]) {
2309                tcg_out_movi32(s, COND_AL, a1, a5);
2310                a5 = a1;
2311            }
2312            tcg_out_dat_rI(s, COND_AL, ARITH_RSC, a1, a5, a3, 1);
2313        } else {
2314            tcg_out_dat_rIK(s, COND_AL, ARITH_SBC, ARITH_ADC,
2315                            a1, a3, a5, const_args[5]);
2316        }
2317        tcg_out_mov_reg(s, COND_AL, args[0], a0);
2318        break;
2319
2320    case INDEX_op_brcond_i32:
2321        c = tcg_out_cmp(s, args[2], args[0], args[1], const_args[1]);
2322        tcg_out_goto_label(s, tcg_cond_to_arm_cond[c], arg_label(args[3]));
2323        break;
2324
2325    case INDEX_op_brcond2_i32:
2326        c = tcg_out_cmp2(s, args, const_args);
2327        tcg_out_goto_label(s, tcg_cond_to_arm_cond[c], arg_label(args[5]));
2328        break;
2329    case INDEX_op_setcond2_i32:
2330        c = tcg_out_cmp2(s, args + 1, const_args + 1);
2331        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c], ARITH_MOV, args[0], 0, 1);
2332        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2333                        ARITH_MOV, args[0], 0, 0);
2334        break;
2335
2336    case INDEX_op_qemu_ld_i32:
2337        tcg_out_qemu_ld(s, args[0], -1, args[1], args[2], TCG_TYPE_I32);
2338        break;
2339    case INDEX_op_qemu_ld_i64:
2340        tcg_out_qemu_ld(s, args[0], args[1], args[2], args[3], TCG_TYPE_I64);
2341        break;
2342
2343    case INDEX_op_qemu_st_i32:
2344        tcg_out_qemu_st(s, args[0], -1, args[1], args[2], TCG_TYPE_I32);
2345        break;
2346    case INDEX_op_qemu_st_i64:
2347        tcg_out_qemu_st(s, args[0], args[1], args[2], args[3], TCG_TYPE_I64);
2348        break;
2349
2350    case INDEX_op_bswap16_i32:
2351        tcg_out_bswap16(s, COND_AL, args[0], args[1], args[2]);
2352        break;
2353    case INDEX_op_bswap32_i32:
2354        tcg_out_bswap32(s, COND_AL, args[0], args[1]);
2355        break;
2356
2357    case INDEX_op_deposit_i32:
2358        tcg_out_deposit(s, COND_AL, args[0], args[2],
2359                        args[3], args[4], const_args[2]);
2360        break;
2361    case INDEX_op_extract_i32:
2362        tcg_out_extract(s, COND_AL, args[0], args[1], args[2], args[3]);
2363        break;
2364    case INDEX_op_sextract_i32:
2365        tcg_out_sextract(s, COND_AL, args[0], args[1], args[2], args[3]);
2366        break;
2367    case INDEX_op_extract2_i32:
2368        /* ??? These optimization vs zero should be generic.  */
2369        /* ??? But we can't substitute 2 for 1 in the opcode stream yet.  */
2370        if (const_args[1]) {
2371            if (const_args[2]) {
2372                tcg_out_movi(s, TCG_TYPE_REG, args[0], 0);
2373            } else {
2374                tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2375                                args[2], SHIFT_IMM_LSL(32 - args[3]));
2376            }
2377        } else if (const_args[2]) {
2378            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2379                            args[1], SHIFT_IMM_LSR(args[3]));
2380        } else {
2381            /* We can do extract2 in 2 insns, vs the 3 required otherwise.  */
2382            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0,
2383                            args[2], SHIFT_IMM_LSL(32 - args[3]));
2384            tcg_out_dat_reg(s, COND_AL, ARITH_ORR, args[0], TCG_REG_TMP,
2385                            args[1], SHIFT_IMM_LSR(args[3]));
2386        }
2387        break;
2388
2389    case INDEX_op_mb:
2390        tcg_out_mb(s, args[0]);
2391        break;
2392
2393    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
2394    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
2395    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
2396    default:
2397        g_assert_not_reached();
2398    }
2399}
2400
2401static TCGConstraintSetIndex
2402tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
2403{
2404    switch (op) {
2405    case INDEX_op_goto_ptr:
2406        return C_O0_I1(r);
2407
2408    case INDEX_op_ld8u_i32:
2409    case INDEX_op_ld8s_i32:
2410    case INDEX_op_ld16u_i32:
2411    case INDEX_op_ld16s_i32:
2412    case INDEX_op_ld_i32:
2413    case INDEX_op_bswap16_i32:
2414    case INDEX_op_bswap32_i32:
2415    case INDEX_op_extract_i32:
2416    case INDEX_op_sextract_i32:
2417        return C_O1_I1(r, r);
2418
2419    case INDEX_op_st8_i32:
2420    case INDEX_op_st16_i32:
2421    case INDEX_op_st_i32:
2422        return C_O0_I2(r, r);
2423
2424    case INDEX_op_brcond_i32:
2425        return C_O0_I2(r, rIN);
2426    case INDEX_op_deposit_i32:
2427        return C_O1_I2(r, 0, rZ);
2428    case INDEX_op_extract2_i32:
2429        return C_O1_I2(r, rZ, rZ);
2430    case INDEX_op_movcond_i32:
2431        return C_O1_I4(r, r, rIN, rIK, 0);
2432    case INDEX_op_add2_i32:
2433        return C_O2_I4(r, r, r, r, rIN, rIK);
2434    case INDEX_op_sub2_i32:
2435        return C_O2_I4(r, r, rI, rI, rIN, rIK);
2436    case INDEX_op_brcond2_i32:
2437        return C_O0_I4(r, r, rI, rI);
2438    case INDEX_op_setcond2_i32:
2439        return C_O1_I4(r, r, r, rI, rI);
2440
2441    case INDEX_op_qemu_ld_i32:
2442        return C_O1_I1(r, q);
2443    case INDEX_op_qemu_ld_i64:
2444        return C_O2_I1(e, p, q);
2445    case INDEX_op_qemu_st_i32:
2446        return C_O0_I2(q, q);
2447    case INDEX_op_qemu_st_i64:
2448        return C_O0_I3(Q, p, q);
2449
2450    case INDEX_op_st_vec:
2451        return C_O0_I2(w, r);
2452    case INDEX_op_ld_vec:
2453    case INDEX_op_dupm_vec:
2454        return C_O1_I1(w, r);
2455    case INDEX_op_dup_vec:
2456        return C_O1_I1(w, wr);
2457    case INDEX_op_abs_vec:
2458    case INDEX_op_neg_vec:
2459    case INDEX_op_not_vec:
2460    case INDEX_op_shli_vec:
2461    case INDEX_op_shri_vec:
2462    case INDEX_op_sari_vec:
2463        return C_O1_I1(w, w);
2464    case INDEX_op_dup2_vec:
2465    case INDEX_op_add_vec:
2466    case INDEX_op_mul_vec:
2467    case INDEX_op_smax_vec:
2468    case INDEX_op_smin_vec:
2469    case INDEX_op_ssadd_vec:
2470    case INDEX_op_sssub_vec:
2471    case INDEX_op_sub_vec:
2472    case INDEX_op_umax_vec:
2473    case INDEX_op_umin_vec:
2474    case INDEX_op_usadd_vec:
2475    case INDEX_op_ussub_vec:
2476    case INDEX_op_xor_vec:
2477    case INDEX_op_arm_sshl_vec:
2478    case INDEX_op_arm_ushl_vec:
2479        return C_O1_I2(w, w, w);
2480    case INDEX_op_arm_sli_vec:
2481        return C_O1_I2(w, 0, w);
2482    case INDEX_op_or_vec:
2483    case INDEX_op_andc_vec:
2484        return C_O1_I2(w, w, wO);
2485    case INDEX_op_and_vec:
2486    case INDEX_op_orc_vec:
2487        return C_O1_I2(w, w, wV);
2488    case INDEX_op_cmp_vec:
2489        return C_O1_I2(w, w, wZ);
2490    case INDEX_op_bitsel_vec:
2491        return C_O1_I3(w, w, w, w);
2492    default:
2493        return C_NotImplemented;
2494    }
2495}
2496
2497static void tcg_target_init(TCGContext *s)
2498{
2499    /*
2500     * Only probe for the platform and capabilities if we haven't already
2501     * determined maximum values at compile time.
2502     */
2503#if !defined(use_idiv_instructions) || !defined(use_neon_instructions)
2504    {
2505        unsigned long hwcap = qemu_getauxval(AT_HWCAP);
2506#ifndef use_idiv_instructions
2507        use_idiv_instructions = (hwcap & HWCAP_ARM_IDIVA) != 0;
2508#endif
2509#ifndef use_neon_instructions
2510        use_neon_instructions = (hwcap & HWCAP_ARM_NEON) != 0;
2511#endif
2512    }
2513#endif
2514
2515    if (__ARM_ARCH < 7) {
2516        const char *pl = (const char *)qemu_getauxval(AT_PLATFORM);
2517        if (pl != NULL && pl[0] == 'v' && pl[1] >= '4' && pl[1] <= '9') {
2518            arm_arch = pl[1] - '0';
2519        }
2520
2521        if (arm_arch < 6) {
2522            error_report("TCG: ARMv%d is unsupported; exiting", arm_arch);
2523            exit(EXIT_FAILURE);
2524        }
2525    }
2526
2527    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
2528
2529    tcg_target_call_clobber_regs = 0;
2530    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R0);
2531    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R1);
2532    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R2);
2533    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R3);
2534    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R12);
2535    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R14);
2536
2537    if (use_neon_instructions) {
2538        tcg_target_available_regs[TCG_TYPE_V64]  = ALL_VECTOR_REGS;
2539        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
2540
2541        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q0);
2542        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q1);
2543        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q2);
2544        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q3);
2545        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q8);
2546        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q9);
2547        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q10);
2548        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q11);
2549        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q12);
2550        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q13);
2551        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q14);
2552        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q15);
2553    }
2554
2555    s->reserved_regs = 0;
2556    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
2557    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP);
2558    tcg_regset_set_reg(s->reserved_regs, TCG_REG_PC);
2559    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP);
2560}
2561
2562static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
2563                       TCGReg arg1, intptr_t arg2)
2564{
2565    switch (type) {
2566    case TCG_TYPE_I32:
2567        tcg_out_ld32u(s, COND_AL, arg, arg1, arg2);
2568        return;
2569    case TCG_TYPE_V64:
2570        /* regs 1; size 8; align 8 */
2571        tcg_out_vldst(s, INSN_VLD1 | 0x7d0, arg, arg1, arg2);
2572        return;
2573    case TCG_TYPE_V128:
2574        /*
2575         * We have only 8-byte alignment for the stack per the ABI.
2576         * Rather than dynamically re-align the stack, it's easier
2577         * to simply not request alignment beyond that.  So:
2578         * regs 2; size 8; align 8
2579         */
2580        tcg_out_vldst(s, INSN_VLD1 | 0xad0, arg, arg1, arg2);
2581        return;
2582    default:
2583        g_assert_not_reached();
2584    }
2585}
2586
2587static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
2588                       TCGReg arg1, intptr_t arg2)
2589{
2590    switch (type) {
2591    case TCG_TYPE_I32:
2592        tcg_out_st32(s, COND_AL, arg, arg1, arg2);
2593        return;
2594    case TCG_TYPE_V64:
2595        /* regs 1; size 8; align 8 */
2596        tcg_out_vldst(s, INSN_VST1 | 0x7d0, arg, arg1, arg2);
2597        return;
2598    case TCG_TYPE_V128:
2599        /* See tcg_out_ld re alignment: regs 2; size 8; align 8 */
2600        tcg_out_vldst(s, INSN_VST1 | 0xad0, arg, arg1, arg2);
2601        return;
2602    default:
2603        g_assert_not_reached();
2604    }
2605}
2606
2607static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
2608                        TCGReg base, intptr_t ofs)
2609{
2610    return false;
2611}
2612
2613static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
2614{
2615    if (ret == arg) {
2616        return true;
2617    }
2618    switch (type) {
2619    case TCG_TYPE_I32:
2620        if (ret < TCG_REG_Q0 && arg < TCG_REG_Q0) {
2621            tcg_out_mov_reg(s, COND_AL, ret, arg);
2622            return true;
2623        }
2624        return false;
2625
2626    case TCG_TYPE_V64:
2627    case TCG_TYPE_V128:
2628        /* "VMOV D,N" is an alias for "VORR D,N,N". */
2629        tcg_out_vreg3(s, INSN_VORR, type - TCG_TYPE_V64, 0, ret, arg, arg);
2630        return true;
2631
2632    default:
2633        g_assert_not_reached();
2634    }
2635}
2636
2637static void tcg_out_movi(TCGContext *s, TCGType type,
2638                         TCGReg ret, tcg_target_long arg)
2639{
2640    tcg_debug_assert(type == TCG_TYPE_I32);
2641    tcg_debug_assert(ret < TCG_REG_Q0);
2642    tcg_out_movi32(s, COND_AL, ret, arg);
2643}
2644
2645static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
2646{
2647    return false;
2648}
2649
2650static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
2651                             tcg_target_long imm)
2652{
2653    int enc, opc = ARITH_ADD;
2654
2655    /* All of the easiest immediates to encode are positive. */
2656    if (imm < 0) {
2657        imm = -imm;
2658        opc = ARITH_SUB;
2659    }
2660    enc = encode_imm(imm);
2661    if (enc >= 0) {
2662        tcg_out_dat_imm(s, COND_AL, opc, rd, rs, enc);
2663    } else {
2664        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, imm);
2665        tcg_out_dat_reg(s, COND_AL, opc, rd, rs,
2666                        TCG_REG_TMP, SHIFT_IMM_LSL(0));
2667    }
2668}
2669
2670/* Type is always V128, with I64 elements.  */
2671static void tcg_out_dup2_vec(TCGContext *s, TCGReg rd, TCGReg rl, TCGReg rh)
2672{
2673    /* Move high element into place first. */
2674    /* VMOV Dd+1, Ds */
2675    tcg_out_vreg3(s, INSN_VORR | (1 << 12), 0, 0, rd, rh, rh);
2676    /* Move low element into place; tcg_out_mov will check for nop. */
2677    tcg_out_mov(s, TCG_TYPE_V64, rd, rl);
2678}
2679
2680static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
2681                            TCGReg rd, TCGReg rs)
2682{
2683    int q = type - TCG_TYPE_V64;
2684
2685    if (vece == MO_64) {
2686        if (type == TCG_TYPE_V128) {
2687            tcg_out_dup2_vec(s, rd, rs, rs);
2688        } else {
2689            tcg_out_mov(s, TCG_TYPE_V64, rd, rs);
2690        }
2691    } else if (rs < TCG_REG_Q0) {
2692        int b = (vece == MO_8);
2693        int e = (vece == MO_16);
2694        tcg_out32(s, INSN_VDUP_G | (b << 22) | (q << 21) | (e << 5) |
2695                  encode_vn(rd) | (rs << 12));
2696    } else {
2697        int imm4 = 1 << vece;
2698        tcg_out32(s, INSN_VDUP_S | (imm4 << 16) | (q << 6) |
2699                  encode_vd(rd) | encode_vm(rs));
2700    }
2701    return true;
2702}
2703
2704static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
2705                             TCGReg rd, TCGReg base, intptr_t offset)
2706{
2707    if (vece == MO_64) {
2708        tcg_out_ld(s, TCG_TYPE_V64, rd, base, offset);
2709        if (type == TCG_TYPE_V128) {
2710            tcg_out_dup2_vec(s, rd, rd, rd);
2711        }
2712    } else {
2713        int q = type - TCG_TYPE_V64;
2714        tcg_out_vldst(s, INSN_VLD1R | (vece << 6) | (q << 5),
2715                      rd, base, offset);
2716    }
2717    return true;
2718}
2719
2720static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
2721                             TCGReg rd, int64_t v64)
2722{
2723    int q = type - TCG_TYPE_V64;
2724    int cmode, imm8, i;
2725
2726    /* Test all bytes equal first.  */
2727    if (vece == MO_8) {
2728        tcg_out_vmovi(s, rd, q, 0, 0xe, v64);
2729        return;
2730    }
2731
2732    /*
2733     * Test all bytes 0x00 or 0xff second.  This can match cases that
2734     * might otherwise take 2 or 3 insns for MO_16 or MO_32 below.
2735     */
2736    for (i = imm8 = 0; i < 8; i++) {
2737        uint8_t byte = v64 >> (i * 8);
2738        if (byte == 0xff) {
2739            imm8 |= 1 << i;
2740        } else if (byte != 0) {
2741            goto fail_bytes;
2742        }
2743    }
2744    tcg_out_vmovi(s, rd, q, 1, 0xe, imm8);
2745    return;
2746 fail_bytes:
2747
2748    /*
2749     * Tests for various replications.  For each element width, if we
2750     * cannot find an expansion there's no point checking a larger
2751     * width because we already know by replication it cannot match.
2752     */
2753    if (vece == MO_16) {
2754        uint16_t v16 = v64;
2755
2756        if (is_shimm16(v16, &cmode, &imm8)) {
2757            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2758            return;
2759        }
2760        if (is_shimm16(~v16, &cmode, &imm8)) {
2761            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2762            return;
2763        }
2764
2765        /*
2766         * Otherwise, all remaining constants can be loaded in two insns:
2767         * rd = v16 & 0xff, rd |= v16 & 0xff00.
2768         */
2769        tcg_out_vmovi(s, rd, q, 0, 0x8, v16 & 0xff);
2770        tcg_out_vmovi(s, rd, q, 0, 0xb, v16 >> 8);   /* VORRI */
2771        return;
2772    }
2773
2774    if (vece == MO_32) {
2775        uint32_t v32 = v64;
2776
2777        if (is_shimm32(v32, &cmode, &imm8) ||
2778            is_soimm32(v32, &cmode, &imm8)) {
2779            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2780            return;
2781        }
2782        if (is_shimm32(~v32, &cmode, &imm8) ||
2783            is_soimm32(~v32, &cmode, &imm8)) {
2784            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2785            return;
2786        }
2787
2788        /*
2789         * Restrict the set of constants to those we can load with
2790         * two instructions.  Others we load from the pool.
2791         */
2792        i = is_shimm32_pair(v32, &cmode, &imm8);
2793        if (i) {
2794            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2795            tcg_out_vmovi(s, rd, q, 0, i | 1, extract32(v32, i * 4, 8));
2796            return;
2797        }
2798        i = is_shimm32_pair(~v32, &cmode, &imm8);
2799        if (i) {
2800            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2801            tcg_out_vmovi(s, rd, q, 1, i | 1, extract32(~v32, i * 4, 8));
2802            return;
2803        }
2804    }
2805
2806    /*
2807     * As a last resort, load from the constant pool.
2808     */
2809    if (!q || vece == MO_64) {
2810        new_pool_l2(s, R_ARM_PC11, s->code_ptr, 0, v64, v64 >> 32);
2811        /* VLDR Dd, [pc + offset] */
2812        tcg_out32(s, INSN_VLDR_D | encode_vd(rd) | (0xf << 16));
2813        if (q) {
2814            tcg_out_dup2_vec(s, rd, rd, rd);
2815        }
2816    } else {
2817        new_pool_label(s, (uint32_t)v64, R_ARM_PC8, s->code_ptr, 0);
2818        /* add tmp, pc, offset */
2819        tcg_out_dat_imm(s, COND_AL, ARITH_ADD, TCG_REG_TMP, TCG_REG_PC, 0);
2820        tcg_out_dupm_vec(s, type, MO_32, rd, TCG_REG_TMP, 0);
2821    }
2822}
2823
2824static const ARMInsn vec_cmp_insn[16] = {
2825    [TCG_COND_EQ] = INSN_VCEQ,
2826    [TCG_COND_GT] = INSN_VCGT,
2827    [TCG_COND_GE] = INSN_VCGE,
2828    [TCG_COND_GTU] = INSN_VCGT_U,
2829    [TCG_COND_GEU] = INSN_VCGE_U,
2830};
2831
2832static const ARMInsn vec_cmp0_insn[16] = {
2833    [TCG_COND_EQ] = INSN_VCEQ0,
2834    [TCG_COND_GT] = INSN_VCGT0,
2835    [TCG_COND_GE] = INSN_VCGE0,
2836    [TCG_COND_LT] = INSN_VCLT0,
2837    [TCG_COND_LE] = INSN_VCLE0,
2838};
2839
2840static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2841                           unsigned vecl, unsigned vece,
2842                           const TCGArg args[TCG_MAX_OP_ARGS],
2843                           const int const_args[TCG_MAX_OP_ARGS])
2844{
2845    TCGType type = vecl + TCG_TYPE_V64;
2846    unsigned q = vecl;
2847    TCGArg a0, a1, a2, a3;
2848    int cmode, imm8;
2849
2850    a0 = args[0];
2851    a1 = args[1];
2852    a2 = args[2];
2853
2854    switch (opc) {
2855    case INDEX_op_ld_vec:
2856        tcg_out_ld(s, type, a0, a1, a2);
2857        return;
2858    case INDEX_op_st_vec:
2859        tcg_out_st(s, type, a0, a1, a2);
2860        return;
2861    case INDEX_op_dupm_vec:
2862        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
2863        return;
2864    case INDEX_op_dup2_vec:
2865        tcg_out_dup2_vec(s, a0, a1, a2);
2866        return;
2867    case INDEX_op_abs_vec:
2868        tcg_out_vreg2(s, INSN_VABS, q, vece, a0, a1);
2869        return;
2870    case INDEX_op_neg_vec:
2871        tcg_out_vreg2(s, INSN_VNEG, q, vece, a0, a1);
2872        return;
2873    case INDEX_op_not_vec:
2874        tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a1);
2875        return;
2876    case INDEX_op_add_vec:
2877        tcg_out_vreg3(s, INSN_VADD, q, vece, a0, a1, a2);
2878        return;
2879    case INDEX_op_mul_vec:
2880        tcg_out_vreg3(s, INSN_VMUL, q, vece, a0, a1, a2);
2881        return;
2882    case INDEX_op_smax_vec:
2883        tcg_out_vreg3(s, INSN_VMAX, q, vece, a0, a1, a2);
2884        return;
2885    case INDEX_op_smin_vec:
2886        tcg_out_vreg3(s, INSN_VMIN, q, vece, a0, a1, a2);
2887        return;
2888    case INDEX_op_sub_vec:
2889        tcg_out_vreg3(s, INSN_VSUB, q, vece, a0, a1, a2);
2890        return;
2891    case INDEX_op_ssadd_vec:
2892        tcg_out_vreg3(s, INSN_VQADD, q, vece, a0, a1, a2);
2893        return;
2894    case INDEX_op_sssub_vec:
2895        tcg_out_vreg3(s, INSN_VQSUB, q, vece, a0, a1, a2);
2896        return;
2897    case INDEX_op_umax_vec:
2898        tcg_out_vreg3(s, INSN_VMAX_U, q, vece, a0, a1, a2);
2899        return;
2900    case INDEX_op_umin_vec:
2901        tcg_out_vreg3(s, INSN_VMIN_U, q, vece, a0, a1, a2);
2902        return;
2903    case INDEX_op_usadd_vec:
2904        tcg_out_vreg3(s, INSN_VQADD_U, q, vece, a0, a1, a2);
2905        return;
2906    case INDEX_op_ussub_vec:
2907        tcg_out_vreg3(s, INSN_VQSUB_U, q, vece, a0, a1, a2);
2908        return;
2909    case INDEX_op_xor_vec:
2910        tcg_out_vreg3(s, INSN_VEOR, q, 0, a0, a1, a2);
2911        return;
2912    case INDEX_op_arm_sshl_vec:
2913        /*
2914         * Note that Vm is the data and Vn is the shift count,
2915         * therefore the arguments appear reversed.
2916         */
2917        tcg_out_vreg3(s, INSN_VSHL_S, q, vece, a0, a2, a1);
2918        return;
2919    case INDEX_op_arm_ushl_vec:
2920        /* See above. */
2921        tcg_out_vreg3(s, INSN_VSHL_U, q, vece, a0, a2, a1);
2922        return;
2923    case INDEX_op_shli_vec:
2924        tcg_out_vshifti(s, INSN_VSHLI, q, a0, a1, a2 + (8 << vece));
2925        return;
2926    case INDEX_op_shri_vec:
2927        tcg_out_vshifti(s, INSN_VSHRI, q, a0, a1, (16 << vece) - a2);
2928        return;
2929    case INDEX_op_sari_vec:
2930        tcg_out_vshifti(s, INSN_VSARI, q, a0, a1, (16 << vece) - a2);
2931        return;
2932    case INDEX_op_arm_sli_vec:
2933        tcg_out_vshifti(s, INSN_VSLI, q, a0, a2, args[3] + (8 << vece));
2934        return;
2935
2936    case INDEX_op_andc_vec:
2937        if (!const_args[2]) {
2938            tcg_out_vreg3(s, INSN_VBIC, q, 0, a0, a1, a2);
2939            return;
2940        }
2941        a2 = ~a2;
2942        /* fall through */
2943    case INDEX_op_and_vec:
2944        if (const_args[2]) {
2945            is_shimm1632(~a2, &cmode, &imm8);
2946            if (a0 == a1) {
2947                tcg_out_vmovi(s, a0, q, 1, cmode | 1, imm8); /* VBICI */
2948                return;
2949            }
2950            tcg_out_vmovi(s, a0, q, 1, cmode, imm8); /* VMVNI */
2951            a2 = a0;
2952        }
2953        tcg_out_vreg3(s, INSN_VAND, q, 0, a0, a1, a2);
2954        return;
2955
2956    case INDEX_op_orc_vec:
2957        if (!const_args[2]) {
2958            tcg_out_vreg3(s, INSN_VORN, q, 0, a0, a1, a2);
2959            return;
2960        }
2961        a2 = ~a2;
2962        /* fall through */
2963    case INDEX_op_or_vec:
2964        if (const_args[2]) {
2965            is_shimm1632(a2, &cmode, &imm8);
2966            if (a0 == a1) {
2967                tcg_out_vmovi(s, a0, q, 0, cmode | 1, imm8); /* VORRI */
2968                return;
2969            }
2970            tcg_out_vmovi(s, a0, q, 0, cmode, imm8); /* VMOVI */
2971            a2 = a0;
2972        }
2973        tcg_out_vreg3(s, INSN_VORR, q, 0, a0, a1, a2);
2974        return;
2975
2976    case INDEX_op_cmp_vec:
2977        {
2978            TCGCond cond = args[3];
2979            ARMInsn insn;
2980
2981            switch (cond) {
2982            case TCG_COND_NE:
2983                if (const_args[2]) {
2984                    tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a1);
2985                } else {
2986                    tcg_out_vreg3(s, INSN_VCEQ, q, vece, a0, a1, a2);
2987                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
2988                }
2989                break;
2990
2991            case TCG_COND_TSTNE:
2992            case TCG_COND_TSTEQ:
2993                if (const_args[2]) {
2994                    /* (x & 0) == 0 */
2995                    tcg_out_dupi_vec(s, type, MO_8, a0,
2996                                     -(cond == TCG_COND_TSTEQ));
2997                    break;
2998                }
2999                tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a2);
3000                if (cond == TCG_COND_TSTEQ) {
3001                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
3002                }
3003                break;
3004
3005            default:
3006                if (const_args[2]) {
3007                    insn = vec_cmp0_insn[cond];
3008                    if (insn) {
3009                        tcg_out_vreg2(s, insn, q, vece, a0, a1);
3010                        return;
3011                    }
3012                    tcg_out_dupi_vec(s, type, MO_8, TCG_VEC_TMP, 0);
3013                    a2 = TCG_VEC_TMP;
3014                }
3015                insn = vec_cmp_insn[cond];
3016                if (insn == 0) {
3017                    TCGArg t;
3018                    t = a1, a1 = a2, a2 = t;
3019                    cond = tcg_swap_cond(cond);
3020                    insn = vec_cmp_insn[cond];
3021                    tcg_debug_assert(insn != 0);
3022                }
3023                tcg_out_vreg3(s, insn, q, vece, a0, a1, a2);
3024                break;
3025            }
3026        }
3027        return;
3028
3029    case INDEX_op_bitsel_vec:
3030        a3 = args[3];
3031        if (a0 == a3) {
3032            tcg_out_vreg3(s, INSN_VBIT, q, 0, a0, a2, a1);
3033        } else if (a0 == a2) {
3034            tcg_out_vreg3(s, INSN_VBIF, q, 0, a0, a3, a1);
3035        } else {
3036            tcg_out_mov(s, type, a0, a1);
3037            tcg_out_vreg3(s, INSN_VBSL, q, 0, a0, a2, a3);
3038        }
3039        return;
3040
3041    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
3042    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
3043    default:
3044        g_assert_not_reached();
3045    }
3046}
3047
3048int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
3049{
3050    switch (opc) {
3051    case INDEX_op_add_vec:
3052    case INDEX_op_sub_vec:
3053    case INDEX_op_and_vec:
3054    case INDEX_op_andc_vec:
3055    case INDEX_op_or_vec:
3056    case INDEX_op_orc_vec:
3057    case INDEX_op_xor_vec:
3058    case INDEX_op_not_vec:
3059    case INDEX_op_shli_vec:
3060    case INDEX_op_shri_vec:
3061    case INDEX_op_sari_vec:
3062    case INDEX_op_ssadd_vec:
3063    case INDEX_op_sssub_vec:
3064    case INDEX_op_usadd_vec:
3065    case INDEX_op_ussub_vec:
3066    case INDEX_op_bitsel_vec:
3067        return 1;
3068    case INDEX_op_abs_vec:
3069    case INDEX_op_cmp_vec:
3070    case INDEX_op_mul_vec:
3071    case INDEX_op_neg_vec:
3072    case INDEX_op_smax_vec:
3073    case INDEX_op_smin_vec:
3074    case INDEX_op_umax_vec:
3075    case INDEX_op_umin_vec:
3076        return vece < MO_64;
3077    case INDEX_op_shlv_vec:
3078    case INDEX_op_shrv_vec:
3079    case INDEX_op_sarv_vec:
3080    case INDEX_op_rotli_vec:
3081    case INDEX_op_rotlv_vec:
3082    case INDEX_op_rotrv_vec:
3083        return -1;
3084    default:
3085        return 0;
3086    }
3087}
3088
3089void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
3090                       TCGArg a0, ...)
3091{
3092    va_list va;
3093    TCGv_vec v0, v1, v2, t1, t2, c1;
3094    TCGArg a2;
3095
3096    va_start(va, a0);
3097    v0 = temp_tcgv_vec(arg_temp(a0));
3098    v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
3099    a2 = va_arg(va, TCGArg);
3100    va_end(va);
3101
3102    switch (opc) {
3103    case INDEX_op_shlv_vec:
3104        /*
3105         * Merely propagate shlv_vec to arm_ushl_vec.
3106         * In this way we don't set TCG_TARGET_HAS_shv_vec
3107         * because everything is done via expansion.
3108         */
3109        v2 = temp_tcgv_vec(arg_temp(a2));
3110        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
3111                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
3112        break;
3113
3114    case INDEX_op_shrv_vec:
3115    case INDEX_op_sarv_vec:
3116        /* Right shifts are negative left shifts for NEON.  */
3117        v2 = temp_tcgv_vec(arg_temp(a2));
3118        t1 = tcg_temp_new_vec(type);
3119        tcg_gen_neg_vec(vece, t1, v2);
3120        if (opc == INDEX_op_shrv_vec) {
3121            opc = INDEX_op_arm_ushl_vec;
3122        } else {
3123            opc = INDEX_op_arm_sshl_vec;
3124        }
3125        vec_gen_3(opc, type, vece, tcgv_vec_arg(v0),
3126                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
3127        tcg_temp_free_vec(t1);
3128        break;
3129
3130    case INDEX_op_rotli_vec:
3131        t1 = tcg_temp_new_vec(type);
3132        tcg_gen_shri_vec(vece, t1, v1, -a2 & ((8 << vece) - 1));
3133        vec_gen_4(INDEX_op_arm_sli_vec, type, vece,
3134                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(v1), a2);
3135        tcg_temp_free_vec(t1);
3136        break;
3137
3138    case INDEX_op_rotlv_vec:
3139        v2 = temp_tcgv_vec(arg_temp(a2));
3140        t1 = tcg_temp_new_vec(type);
3141        c1 = tcg_constant_vec(type, vece, 8 << vece);
3142        tcg_gen_sub_vec(vece, t1, v2, c1);
3143        /* Right shifts are negative left shifts for NEON.  */
3144        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
3145                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
3146        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
3147                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
3148        tcg_gen_or_vec(vece, v0, v0, t1);
3149        tcg_temp_free_vec(t1);
3150        break;
3151
3152    case INDEX_op_rotrv_vec:
3153        v2 = temp_tcgv_vec(arg_temp(a2));
3154        t1 = tcg_temp_new_vec(type);
3155        t2 = tcg_temp_new_vec(type);
3156        c1 = tcg_constant_vec(type, vece, 8 << vece);
3157        tcg_gen_neg_vec(vece, t1, v2);
3158        tcg_gen_sub_vec(vece, t2, c1, v2);
3159        /* Right shifts are negative left shifts for NEON.  */
3160        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
3161                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
3162        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t2),
3163                  tcgv_vec_arg(v1), tcgv_vec_arg(t2));
3164        tcg_gen_or_vec(vece, v0, t1, t2);
3165        tcg_temp_free_vec(t1);
3166        tcg_temp_free_vec(t2);
3167        break;
3168
3169    default:
3170        g_assert_not_reached();
3171    }
3172}
3173
3174static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
3175{
3176    int i;
3177    for (i = 0; i < count; ++i) {
3178        p[i] = INSN_NOP;
3179    }
3180}
3181
3182/* Compute frame size via macros, to share between tcg_target_qemu_prologue
3183   and tcg_register_jit.  */
3184
3185#define PUSH_SIZE  ((11 - 4 + 1 + 1) * sizeof(tcg_target_long))
3186
3187#define FRAME_SIZE \
3188    ((PUSH_SIZE \
3189      + TCG_STATIC_CALL_ARGS_SIZE \
3190      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
3191      + TCG_TARGET_STACK_ALIGN - 1) \
3192     & -TCG_TARGET_STACK_ALIGN)
3193
3194#define STACK_ADDEND  (FRAME_SIZE - PUSH_SIZE)
3195
3196static void tcg_target_qemu_prologue(TCGContext *s)
3197{
3198    /* Calling convention requires us to save r4-r11 and lr.  */
3199    /* stmdb sp!, { r4 - r11, lr } */
3200    tcg_out_ldstm(s, COND_AL, INSN_STMDB, TCG_REG_CALL_STACK,
3201                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
3202                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3203                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_R14));
3204
3205    /* Reserve callee argument and tcg temp space.  */
3206    tcg_out_dat_rI(s, COND_AL, ARITH_SUB, TCG_REG_CALL_STACK,
3207                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
3208    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
3209                  CPU_TEMP_BUF_NLONGS * sizeof(long));
3210
3211    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
3212
3213    if (!tcg_use_softmmu && guest_base) {
3214        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_GUEST_BASE, guest_base);
3215        tcg_regset_set_reg(s->reserved_regs, TCG_REG_GUEST_BASE);
3216    }
3217
3218    tcg_out_b_reg(s, COND_AL, tcg_target_call_iarg_regs[1]);
3219
3220    /*
3221     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
3222     * and fall through to the rest of the epilogue.
3223     */
3224    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
3225    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, 0);
3226    tcg_out_epilogue(s);
3227}
3228
3229static void tcg_out_epilogue(TCGContext *s)
3230{
3231    /* Release local stack frame.  */
3232    tcg_out_dat_rI(s, COND_AL, ARITH_ADD, TCG_REG_CALL_STACK,
3233                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
3234
3235    /* ldmia sp!, { r4 - r11, pc } */
3236    tcg_out_ldstm(s, COND_AL, INSN_LDMIA, TCG_REG_CALL_STACK,
3237                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
3238                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3239                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_PC));
3240}
3241
3242static void tcg_out_tb_start(TCGContext *s)
3243{
3244    /* nothing to do */
3245}
3246
3247typedef struct {
3248    DebugFrameHeader h;
3249    uint8_t fde_def_cfa[4];
3250    uint8_t fde_reg_ofs[18];
3251} DebugFrame;
3252
3253#define ELF_HOST_MACHINE EM_ARM
3254
3255/* We're expecting a 2 byte uleb128 encoded value.  */
3256QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
3257
3258static const DebugFrame debug_frame = {
3259    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
3260    .h.cie.id = -1,
3261    .h.cie.version = 1,
3262    .h.cie.code_align = 1,
3263    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
3264    .h.cie.return_column = 14,
3265
3266    /* Total FDE size does not include the "len" member.  */
3267    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
3268
3269    .fde_def_cfa = {
3270        12, 13,                         /* DW_CFA_def_cfa sp, ... */
3271        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
3272        (FRAME_SIZE >> 7)
3273    },
3274    .fde_reg_ofs = {
3275        /* The following must match the stmdb in the prologue.  */
3276        0x8e, 1,                        /* DW_CFA_offset, lr, -4 */
3277        0x8b, 2,                        /* DW_CFA_offset, r11, -8 */
3278        0x8a, 3,                        /* DW_CFA_offset, r10, -12 */
3279        0x89, 4,                        /* DW_CFA_offset, r9, -16 */
3280        0x88, 5,                        /* DW_CFA_offset, r8, -20 */
3281        0x87, 6,                        /* DW_CFA_offset, r7, -24 */
3282        0x86, 7,                        /* DW_CFA_offset, r6, -28 */
3283        0x85, 8,                        /* DW_CFA_offset, r5, -32 */
3284        0x84, 9,                        /* DW_CFA_offset, r4, -36 */
3285    }
3286};
3287
3288void tcg_register_jit(const void *buf, size_t buf_size)
3289{
3290    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
3291}
3292