xref: /openbmc/qemu/tcg/arm/tcg-target.c.inc (revision 0cadc1eda1a3120c37c713ab6d6b7a02da0d2e6f)
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#include "../tcg-ldst.c.inc"
27#include "../tcg-pool.c.inc"
28
29int arm_arch = __ARM_ARCH;
30
31#ifndef use_idiv_instructions
32bool use_idiv_instructions;
33#endif
34#ifndef use_neon_instructions
35bool use_neon_instructions;
36#endif
37
38#ifdef CONFIG_DEBUG_TCG
39static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
40    "%r0",  "%r1",  "%r2",  "%r3",  "%r4",  "%r5",  "%r6",  "%r7",
41    "%r8",  "%r9",  "%r10", "%r11", "%r12", "%sp",  "%r14", "%pc",
42    "%q0",  "%q1",  "%q2",  "%q3",  "%q4",  "%q5",  "%q6",  "%q7",
43    "%q8",  "%q9",  "%q10", "%q11", "%q12", "%q13", "%q14", "%q15",
44};
45#endif
46
47static const int tcg_target_reg_alloc_order[] = {
48    TCG_REG_R4,
49    TCG_REG_R5,
50    TCG_REG_R6,
51    TCG_REG_R7,
52    TCG_REG_R8,
53    TCG_REG_R9,
54    TCG_REG_R10,
55    TCG_REG_R11,
56    TCG_REG_R13,
57    TCG_REG_R0,
58    TCG_REG_R1,
59    TCG_REG_R2,
60    TCG_REG_R3,
61    TCG_REG_R12,
62    TCG_REG_R14,
63
64    TCG_REG_Q0,
65    TCG_REG_Q1,
66    TCG_REG_Q2,
67    TCG_REG_Q3,
68    /* Q4 - Q7 are call-saved, and skipped. */
69    TCG_REG_Q8,
70    TCG_REG_Q9,
71    TCG_REG_Q10,
72    TCG_REG_Q11,
73    TCG_REG_Q12,
74    TCG_REG_Q13,
75    TCG_REG_Q14,
76    TCG_REG_Q15,
77};
78
79static const int tcg_target_call_iarg_regs[4] = {
80    TCG_REG_R0, TCG_REG_R1, TCG_REG_R2, TCG_REG_R3
81};
82
83static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
84{
85    tcg_debug_assert(kind == TCG_CALL_RET_NORMAL);
86    tcg_debug_assert(slot >= 0 && slot <= 3);
87    return TCG_REG_R0 + slot;
88}
89
90#define TCG_REG_TMP  TCG_REG_R12
91#define TCG_VEC_TMP  TCG_REG_Q15
92#ifndef CONFIG_SOFTMMU
93#define TCG_REG_GUEST_BASE  TCG_REG_R11
94#endif
95
96typedef enum {
97    COND_EQ = 0x0,
98    COND_NE = 0x1,
99    COND_CS = 0x2,	/* Unsigned greater or equal */
100    COND_CC = 0x3,	/* Unsigned less than */
101    COND_MI = 0x4,	/* Negative */
102    COND_PL = 0x5,	/* Zero or greater */
103    COND_VS = 0x6,	/* Overflow */
104    COND_VC = 0x7,	/* No overflow */
105    COND_HI = 0x8,	/* Unsigned greater than */
106    COND_LS = 0x9,	/* Unsigned less or equal */
107    COND_GE = 0xa,
108    COND_LT = 0xb,
109    COND_GT = 0xc,
110    COND_LE = 0xd,
111    COND_AL = 0xe,
112} ARMCond;
113
114#define TO_CPSR (1 << 20)
115
116#define SHIFT_IMM_LSL(im)	(((im) << 7) | 0x00)
117#define SHIFT_IMM_LSR(im)	(((im) << 7) | 0x20)
118#define SHIFT_IMM_ASR(im)	(((im) << 7) | 0x40)
119#define SHIFT_IMM_ROR(im)	(((im) << 7) | 0x60)
120#define SHIFT_REG_LSL(rs)	(((rs) << 8) | 0x10)
121#define SHIFT_REG_LSR(rs)	(((rs) << 8) | 0x30)
122#define SHIFT_REG_ASR(rs)	(((rs) << 8) | 0x50)
123#define SHIFT_REG_ROR(rs)	(((rs) << 8) | 0x70)
124
125typedef enum {
126    ARITH_AND = 0x0 << 21,
127    ARITH_EOR = 0x1 << 21,
128    ARITH_SUB = 0x2 << 21,
129    ARITH_RSB = 0x3 << 21,
130    ARITH_ADD = 0x4 << 21,
131    ARITH_ADC = 0x5 << 21,
132    ARITH_SBC = 0x6 << 21,
133    ARITH_RSC = 0x7 << 21,
134    ARITH_TST = 0x8 << 21 | TO_CPSR,
135    ARITH_CMP = 0xa << 21 | TO_CPSR,
136    ARITH_CMN = 0xb << 21 | TO_CPSR,
137    ARITH_ORR = 0xc << 21,
138    ARITH_MOV = 0xd << 21,
139    ARITH_BIC = 0xe << 21,
140    ARITH_MVN = 0xf << 21,
141
142    INSN_B         = 0x0a000000,
143
144    INSN_CLZ       = 0x016f0f10,
145    INSN_RBIT      = 0x06ff0f30,
146
147    INSN_LDMIA     = 0x08b00000,
148    INSN_STMDB     = 0x09200000,
149
150    INSN_LDR_IMM   = 0x04100000,
151    INSN_LDR_REG   = 0x06100000,
152    INSN_STR_IMM   = 0x04000000,
153    INSN_STR_REG   = 0x06000000,
154
155    INSN_LDRH_IMM  = 0x005000b0,
156    INSN_LDRH_REG  = 0x001000b0,
157    INSN_LDRSH_IMM = 0x005000f0,
158    INSN_LDRSH_REG = 0x001000f0,
159    INSN_STRH_IMM  = 0x004000b0,
160    INSN_STRH_REG  = 0x000000b0,
161
162    INSN_LDRB_IMM  = 0x04500000,
163    INSN_LDRB_REG  = 0x06500000,
164    INSN_LDRSB_IMM = 0x005000d0,
165    INSN_LDRSB_REG = 0x001000d0,
166    INSN_STRB_IMM  = 0x04400000,
167    INSN_STRB_REG  = 0x06400000,
168
169    INSN_LDRD_IMM  = 0x004000d0,
170    INSN_LDRD_REG  = 0x000000d0,
171    INSN_STRD_IMM  = 0x004000f0,
172    INSN_STRD_REG  = 0x000000f0,
173
174    INSN_DMB_ISH   = 0xf57ff05b,
175    INSN_DMB_MCR   = 0xee070fba,
176
177    /* Architected nop introduced in v6k.  */
178    /* ??? This is an MSR (imm) 0,0,0 insn.  Anyone know if this
179       also Just So Happened to do nothing on pre-v6k so that we
180       don't need to conditionalize it?  */
181    INSN_NOP_v6k   = 0xe320f000,
182    /* Otherwise the assembler uses mov r0,r0 */
183    INSN_NOP_v4    = (COND_AL << 28) | ARITH_MOV,
184
185    INSN_VADD      = 0xf2000800,
186    INSN_VAND      = 0xf2000110,
187    INSN_VBIC      = 0xf2100110,
188    INSN_VEOR      = 0xf3000110,
189    INSN_VORN      = 0xf2300110,
190    INSN_VORR      = 0xf2200110,
191    INSN_VSUB      = 0xf3000800,
192    INSN_VMUL      = 0xf2000910,
193    INSN_VQADD     = 0xf2000010,
194    INSN_VQADD_U   = 0xf3000010,
195    INSN_VQSUB     = 0xf2000210,
196    INSN_VQSUB_U   = 0xf3000210,
197    INSN_VMAX      = 0xf2000600,
198    INSN_VMAX_U    = 0xf3000600,
199    INSN_VMIN      = 0xf2000610,
200    INSN_VMIN_U    = 0xf3000610,
201
202    INSN_VABS      = 0xf3b10300,
203    INSN_VMVN      = 0xf3b00580,
204    INSN_VNEG      = 0xf3b10380,
205
206    INSN_VCEQ0     = 0xf3b10100,
207    INSN_VCGT0     = 0xf3b10000,
208    INSN_VCGE0     = 0xf3b10080,
209    INSN_VCLE0     = 0xf3b10180,
210    INSN_VCLT0     = 0xf3b10200,
211
212    INSN_VCEQ      = 0xf3000810,
213    INSN_VCGE      = 0xf2000310,
214    INSN_VCGT      = 0xf2000300,
215    INSN_VCGE_U    = 0xf3000310,
216    INSN_VCGT_U    = 0xf3000300,
217
218    INSN_VSHLI     = 0xf2800510,  /* VSHL (immediate) */
219    INSN_VSARI     = 0xf2800010,  /* VSHR.S */
220    INSN_VSHRI     = 0xf3800010,  /* VSHR.U */
221    INSN_VSLI      = 0xf3800510,
222    INSN_VSHL_S    = 0xf2000400,  /* VSHL.S (register) */
223    INSN_VSHL_U    = 0xf3000400,  /* VSHL.U (register) */
224
225    INSN_VBSL      = 0xf3100110,
226    INSN_VBIT      = 0xf3200110,
227    INSN_VBIF      = 0xf3300110,
228
229    INSN_VTST      = 0xf2000810,
230
231    INSN_VDUP_G    = 0xee800b10,  /* VDUP (ARM core register) */
232    INSN_VDUP_S    = 0xf3b00c00,  /* VDUP (scalar) */
233    INSN_VLDR_D    = 0xed100b00,  /* VLDR.64 */
234    INSN_VLD1      = 0xf4200000,  /* VLD1 (multiple single elements) */
235    INSN_VLD1R     = 0xf4a00c00,  /* VLD1 (single element to all lanes) */
236    INSN_VST1      = 0xf4000000,  /* VST1 (multiple single elements) */
237    INSN_VMOVI     = 0xf2800010,  /* VMOV (immediate) */
238} ARMInsn;
239
240#define INSN_NOP   (use_armv7_instructions ? INSN_NOP_v6k : INSN_NOP_v4)
241
242static const uint8_t tcg_cond_to_arm_cond[] = {
243    [TCG_COND_EQ] = COND_EQ,
244    [TCG_COND_NE] = COND_NE,
245    [TCG_COND_LT] = COND_LT,
246    [TCG_COND_GE] = COND_GE,
247    [TCG_COND_LE] = COND_LE,
248    [TCG_COND_GT] = COND_GT,
249    /* unsigned */
250    [TCG_COND_LTU] = COND_CC,
251    [TCG_COND_GEU] = COND_CS,
252    [TCG_COND_LEU] = COND_LS,
253    [TCG_COND_GTU] = COND_HI,
254};
255
256static int encode_imm(uint32_t imm);
257
258/* TCG private relocation type: add with pc+imm8 */
259#define R_ARM_PC8  11
260
261/* TCG private relocation type: vldr with imm8 << 2 */
262#define R_ARM_PC11 12
263
264static bool reloc_pc24(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
265{
266    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
267    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) >> 2;
268
269    if (offset == sextract32(offset, 0, 24)) {
270        *src_rw = deposit32(*src_rw, 0, 24, offset);
271        return true;
272    }
273    return false;
274}
275
276static bool reloc_pc13(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
277{
278    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
279    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
280
281    if (offset >= -0xfff && offset <= 0xfff) {
282        tcg_insn_unit insn = *src_rw;
283        bool u = (offset >= 0);
284        if (!u) {
285            offset = -offset;
286        }
287        insn = deposit32(insn, 23, 1, u);
288        insn = deposit32(insn, 0, 12, offset);
289        *src_rw = insn;
290        return true;
291    }
292    return false;
293}
294
295static bool reloc_pc11(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
296{
297    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
298    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) / 4;
299
300    if (offset >= -0xff && offset <= 0xff) {
301        tcg_insn_unit insn = *src_rw;
302        bool u = (offset >= 0);
303        if (!u) {
304            offset = -offset;
305        }
306        insn = deposit32(insn, 23, 1, u);
307        insn = deposit32(insn, 0, 8, offset);
308        *src_rw = insn;
309        return true;
310    }
311    return false;
312}
313
314static bool reloc_pc8(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
315{
316    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
317    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
318    int imm12 = encode_imm(offset);
319
320    if (imm12 >= 0) {
321        *src_rw = deposit32(*src_rw, 0, 12, imm12);
322        return true;
323    }
324    return false;
325}
326
327static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
328                        intptr_t value, intptr_t addend)
329{
330    tcg_debug_assert(addend == 0);
331    switch (type) {
332    case R_ARM_PC24:
333        return reloc_pc24(code_ptr, (const tcg_insn_unit *)value);
334    case R_ARM_PC13:
335        return reloc_pc13(code_ptr, (const tcg_insn_unit *)value);
336    case R_ARM_PC11:
337        return reloc_pc11(code_ptr, (const tcg_insn_unit *)value);
338    case R_ARM_PC8:
339        return reloc_pc8(code_ptr, (const tcg_insn_unit *)value);
340    default:
341        g_assert_not_reached();
342    }
343}
344
345#define TCG_CT_CONST_ARM  0x100
346#define TCG_CT_CONST_INV  0x200
347#define TCG_CT_CONST_NEG  0x400
348#define TCG_CT_CONST_ZERO 0x800
349#define TCG_CT_CONST_ORRI 0x1000
350#define TCG_CT_CONST_ANDI 0x2000
351
352#define ALL_GENERAL_REGS  0xffffu
353#define ALL_VECTOR_REGS   0xffff0000u
354
355/*
356 * r0-r2 will be overwritten when reading the tlb entry (softmmu only)
357 * and r0-r1 doing the byte swapping, so don't use these.
358 * r3 is removed for softmmu to avoid clashes with helper arguments.
359 */
360#ifdef CONFIG_SOFTMMU
361#define ALL_QLOAD_REGS \
362    (ALL_GENERAL_REGS & ~((1 << TCG_REG_R0) | (1 << TCG_REG_R1) | \
363                          (1 << TCG_REG_R2) | (1 << TCG_REG_R3) | \
364                          (1 << TCG_REG_R14)))
365#define ALL_QSTORE_REGS \
366    (ALL_GENERAL_REGS & ~((1 << TCG_REG_R0) | (1 << TCG_REG_R1) | \
367                          (1 << TCG_REG_R2) | (1 << TCG_REG_R14) | \
368                          ((TARGET_LONG_BITS == 64) << TCG_REG_R3)))
369#else
370#define ALL_QLOAD_REGS   ALL_GENERAL_REGS
371#define ALL_QSTORE_REGS \
372    (ALL_GENERAL_REGS & ~((1 << TCG_REG_R0) | (1 << TCG_REG_R1)))
373#endif
374
375/*
376 * ARM immediates for ALU instructions are made of an unsigned 8-bit
377 * right-rotated by an even amount between 0 and 30.
378 *
379 * Return < 0 if @imm cannot be encoded, else the entire imm12 field.
380 */
381static int encode_imm(uint32_t imm)
382{
383    uint32_t rot, imm8;
384
385    /* Simple case, no rotation required. */
386    if ((imm & ~0xff) == 0) {
387        return imm;
388    }
389
390    /* Next, try a simple even shift.  */
391    rot = ctz32(imm) & ~1;
392    imm8 = imm >> rot;
393    rot = 32 - rot;
394    if ((imm8 & ~0xff) == 0) {
395        goto found;
396    }
397
398    /*
399     * Finally, try harder with rotations.
400     * The ctz test above will have taken care of rotates >= 8.
401     */
402    for (rot = 2; rot < 8; rot += 2) {
403        imm8 = rol32(imm, rot);
404        if ((imm8 & ~0xff) == 0) {
405            goto found;
406        }
407    }
408    /* Fail: imm cannot be encoded. */
409    return -1;
410
411 found:
412    /* Note that rot is even, and we discard bit 0 by shifting by 7. */
413    return rot << 7 | imm8;
414}
415
416static int encode_imm_nofail(uint32_t imm)
417{
418    int ret = encode_imm(imm);
419    tcg_debug_assert(ret >= 0);
420    return ret;
421}
422
423static bool check_fit_imm(uint32_t imm)
424{
425    return encode_imm(imm) >= 0;
426}
427
428/* Return true if v16 is a valid 16-bit shifted immediate.  */
429static bool is_shimm16(uint16_t v16, int *cmode, int *imm8)
430{
431    if (v16 == (v16 & 0xff)) {
432        *cmode = 0x8;
433        *imm8 = v16 & 0xff;
434        return true;
435    } else if (v16 == (v16 & 0xff00)) {
436        *cmode = 0xa;
437        *imm8 = v16 >> 8;
438        return true;
439    }
440    return false;
441}
442
443/* Return true if v32 is a valid 32-bit shifted immediate.  */
444static bool is_shimm32(uint32_t v32, int *cmode, int *imm8)
445{
446    if (v32 == (v32 & 0xff)) {
447        *cmode = 0x0;
448        *imm8 = v32 & 0xff;
449        return true;
450    } else if (v32 == (v32 & 0xff00)) {
451        *cmode = 0x2;
452        *imm8 = (v32 >> 8) & 0xff;
453        return true;
454    } else if (v32 == (v32 & 0xff0000)) {
455        *cmode = 0x4;
456        *imm8 = (v32 >> 16) & 0xff;
457        return true;
458    } else if (v32 == (v32 & 0xff000000)) {
459        *cmode = 0x6;
460        *imm8 = v32 >> 24;
461        return true;
462    }
463    return false;
464}
465
466/* Return true if v32 is a valid 32-bit shifting ones immediate.  */
467static bool is_soimm32(uint32_t v32, int *cmode, int *imm8)
468{
469    if ((v32 & 0xffff00ff) == 0xff) {
470        *cmode = 0xc;
471        *imm8 = (v32 >> 8) & 0xff;
472        return true;
473    } else if ((v32 & 0xff00ffff) == 0xffff) {
474        *cmode = 0xd;
475        *imm8 = (v32 >> 16) & 0xff;
476        return true;
477    }
478    return false;
479}
480
481/*
482 * Return non-zero if v32 can be formed by MOVI+ORR.
483 * Place the parameters for MOVI in (cmode, imm8).
484 * Return the cmode for ORR; the imm8 can be had via extraction from v32.
485 */
486static int is_shimm32_pair(uint32_t v32, int *cmode, int *imm8)
487{
488    int i;
489
490    for (i = 6; i > 0; i -= 2) {
491        /* Mask out one byte we can add with ORR.  */
492        uint32_t tmp = v32 & ~(0xffu << (i * 4));
493        if (is_shimm32(tmp, cmode, imm8) ||
494            is_soimm32(tmp, cmode, imm8)) {
495            break;
496        }
497    }
498    return i;
499}
500
501/* Return true if V is a valid 16-bit or 32-bit shifted immediate.  */
502static bool is_shimm1632(uint32_t v32, int *cmode, int *imm8)
503{
504    if (v32 == deposit32(v32, 16, 16, v32)) {
505        return is_shimm16(v32, cmode, imm8);
506    } else {
507        return is_shimm32(v32, cmode, imm8);
508    }
509}
510
511/* Test if a constant matches the constraint.
512 * TODO: define constraints for:
513 *
514 * ldr/str offset:   between -0xfff and 0xfff
515 * ldrh/strh offset: between -0xff and 0xff
516 * mov operand2:     values represented with x << (2 * y), x < 0x100
517 * add, sub, eor...: ditto
518 */
519static bool tcg_target_const_match(int64_t val, TCGType type, int ct)
520{
521    if (ct & TCG_CT_CONST) {
522        return 1;
523    } else if ((ct & TCG_CT_CONST_ARM) && check_fit_imm(val)) {
524        return 1;
525    } else if ((ct & TCG_CT_CONST_INV) && check_fit_imm(~val)) {
526        return 1;
527    } else if ((ct & TCG_CT_CONST_NEG) && check_fit_imm(-val)) {
528        return 1;
529    } else if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
530        return 1;
531    }
532
533    switch (ct & (TCG_CT_CONST_ORRI | TCG_CT_CONST_ANDI)) {
534    case 0:
535        break;
536    case TCG_CT_CONST_ANDI:
537        val = ~val;
538        /* fallthru */
539    case TCG_CT_CONST_ORRI:
540        if (val == deposit64(val, 32, 32, val)) {
541            int cmode, imm8;
542            return is_shimm1632(val, &cmode, &imm8);
543        }
544        break;
545    default:
546        /* Both bits should not be set for the same insn.  */
547        g_assert_not_reached();
548    }
549
550    return 0;
551}
552
553static void tcg_out_b_imm(TCGContext *s, ARMCond cond, int32_t offset)
554{
555    tcg_out32(s, (cond << 28) | INSN_B |
556                    (((offset - 8) >> 2) & 0x00ffffff));
557}
558
559static void tcg_out_bl_imm(TCGContext *s, ARMCond cond, int32_t offset)
560{
561    tcg_out32(s, (cond << 28) | 0x0b000000 |
562                    (((offset - 8) >> 2) & 0x00ffffff));
563}
564
565static void tcg_out_blx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
566{
567    tcg_out32(s, (cond << 28) | 0x012fff30 | rn);
568}
569
570static void tcg_out_blx_imm(TCGContext *s, int32_t offset)
571{
572    tcg_out32(s, 0xfa000000 | ((offset & 2) << 23) |
573                (((offset - 8) >> 2) & 0x00ffffff));
574}
575
576static void tcg_out_dat_reg(TCGContext *s, ARMCond cond, ARMInsn opc,
577                            TCGReg rd, TCGReg rn, TCGReg rm, int shift)
578{
579    tcg_out32(s, (cond << 28) | (0 << 25) | opc |
580                    (rn << 16) | (rd << 12) | shift | rm);
581}
582
583static void tcg_out_mov_reg(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rm)
584{
585    /* Simple reg-reg move, optimising out the 'do nothing' case */
586    if (rd != rm) {
587        tcg_out_dat_reg(s, cond, ARITH_MOV, rd, 0, rm, SHIFT_IMM_LSL(0));
588    }
589}
590
591static void tcg_out_bx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
592{
593    tcg_out32(s, (cond << 28) | 0x012fff10 | rn);
594}
595
596static void tcg_out_b_reg(TCGContext *s, ARMCond cond, TCGReg rn)
597{
598    /*
599     * Unless the C portion of QEMU is compiled as thumb, we don't need
600     * true BX semantics; merely a branch to an address held in a register.
601     */
602    tcg_out_bx_reg(s, cond, rn);
603}
604
605static void tcg_out_dat_imm(TCGContext *s, ARMCond cond, ARMInsn opc,
606                            TCGReg rd, TCGReg rn, int im)
607{
608    tcg_out32(s, (cond << 28) | (1 << 25) | opc |
609                    (rn << 16) | (rd << 12) | im);
610}
611
612static void tcg_out_ldstm(TCGContext *s, ARMCond cond, ARMInsn opc,
613                          TCGReg rn, uint16_t mask)
614{
615    tcg_out32(s, (cond << 28) | opc | (rn << 16) | mask);
616}
617
618/* Note that this routine is used for both LDR and LDRH formats, so we do
619   not wish to include an immediate shift at this point.  */
620static void tcg_out_memop_r(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
621                            TCGReg rn, TCGReg rm, bool u, bool p, bool w)
622{
623    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24)
624              | (w << 21) | (rn << 16) | (rt << 12) | rm);
625}
626
627static void tcg_out_memop_8(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
628                            TCGReg rn, int imm8, bool p, bool w)
629{
630    bool u = 1;
631    if (imm8 < 0) {
632        imm8 = -imm8;
633        u = 0;
634    }
635    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
636              (rn << 16) | (rt << 12) | ((imm8 & 0xf0) << 4) | (imm8 & 0xf));
637}
638
639static void tcg_out_memop_12(TCGContext *s, ARMCond cond, ARMInsn opc,
640                             TCGReg rt, TCGReg rn, int imm12, bool p, bool w)
641{
642    bool u = 1;
643    if (imm12 < 0) {
644        imm12 = -imm12;
645        u = 0;
646    }
647    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
648              (rn << 16) | (rt << 12) | imm12);
649}
650
651static void tcg_out_ld32_12(TCGContext *s, ARMCond cond, TCGReg rt,
652                            TCGReg rn, int imm12)
653{
654    tcg_out_memop_12(s, cond, INSN_LDR_IMM, rt, rn, imm12, 1, 0);
655}
656
657static void tcg_out_st32_12(TCGContext *s, ARMCond cond, TCGReg rt,
658                            TCGReg rn, int imm12)
659{
660    tcg_out_memop_12(s, cond, INSN_STR_IMM, rt, rn, imm12, 1, 0);
661}
662
663static void tcg_out_ld32_r(TCGContext *s, ARMCond cond, TCGReg rt,
664                           TCGReg rn, TCGReg rm)
665{
666    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 0);
667}
668
669static void tcg_out_st32_r(TCGContext *s, ARMCond cond, TCGReg rt,
670                           TCGReg rn, TCGReg rm)
671{
672    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 0);
673}
674
675static void tcg_out_ldrd_8(TCGContext *s, ARMCond cond, TCGReg rt,
676                           TCGReg rn, int imm8)
677{
678    tcg_out_memop_8(s, cond, INSN_LDRD_IMM, rt, rn, imm8, 1, 0);
679}
680
681static void tcg_out_ldrd_r(TCGContext *s, ARMCond cond, TCGReg rt,
682                           TCGReg rn, TCGReg rm)
683{
684    tcg_out_memop_r(s, cond, INSN_LDRD_REG, rt, rn, rm, 1, 1, 0);
685}
686
687static void __attribute__((unused))
688tcg_out_ldrd_rwb(TCGContext *s, ARMCond cond, TCGReg rt, TCGReg rn, TCGReg rm)
689{
690    tcg_out_memop_r(s, cond, INSN_LDRD_REG, rt, rn, rm, 1, 1, 1);
691}
692
693static void __attribute__((unused))
694tcg_out_strd_8(TCGContext *s, ARMCond cond, TCGReg rt, TCGReg rn, int imm8)
695{
696    tcg_out_memop_8(s, cond, INSN_STRD_IMM, rt, rn, imm8, 1, 0);
697}
698
699static void tcg_out_strd_r(TCGContext *s, ARMCond cond, TCGReg rt,
700                           TCGReg rn, TCGReg rm)
701{
702    tcg_out_memop_r(s, cond, INSN_STRD_REG, rt, rn, rm, 1, 1, 0);
703}
704
705/* Register pre-increment with base writeback.  */
706static void tcg_out_ld32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
707                             TCGReg rn, TCGReg rm)
708{
709    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 1);
710}
711
712static void tcg_out_st32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
713                             TCGReg rn, TCGReg rm)
714{
715    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 1);
716}
717
718static void tcg_out_ld16u_8(TCGContext *s, ARMCond cond, TCGReg rt,
719                            TCGReg rn, int imm8)
720{
721    tcg_out_memop_8(s, cond, INSN_LDRH_IMM, rt, rn, imm8, 1, 0);
722}
723
724static void tcg_out_st16_8(TCGContext *s, ARMCond cond, TCGReg rt,
725                           TCGReg rn, int imm8)
726{
727    tcg_out_memop_8(s, cond, INSN_STRH_IMM, rt, rn, imm8, 1, 0);
728}
729
730static void tcg_out_ld16u_r(TCGContext *s, ARMCond cond, TCGReg rt,
731                            TCGReg rn, TCGReg rm)
732{
733    tcg_out_memop_r(s, cond, INSN_LDRH_REG, rt, rn, rm, 1, 1, 0);
734}
735
736static void tcg_out_st16_r(TCGContext *s, ARMCond cond, TCGReg rt,
737                           TCGReg rn, TCGReg rm)
738{
739    tcg_out_memop_r(s, cond, INSN_STRH_REG, rt, rn, rm, 1, 1, 0);
740}
741
742static void tcg_out_ld16s_8(TCGContext *s, ARMCond cond, TCGReg rt,
743                            TCGReg rn, int imm8)
744{
745    tcg_out_memop_8(s, cond, INSN_LDRSH_IMM, rt, rn, imm8, 1, 0);
746}
747
748static void tcg_out_ld16s_r(TCGContext *s, ARMCond cond, TCGReg rt,
749                            TCGReg rn, TCGReg rm)
750{
751    tcg_out_memop_r(s, cond, INSN_LDRSH_REG, rt, rn, rm, 1, 1, 0);
752}
753
754static void tcg_out_ld8_12(TCGContext *s, ARMCond cond, TCGReg rt,
755                           TCGReg rn, int imm12)
756{
757    tcg_out_memop_12(s, cond, INSN_LDRB_IMM, rt, rn, imm12, 1, 0);
758}
759
760static void tcg_out_st8_12(TCGContext *s, ARMCond cond, TCGReg rt,
761                           TCGReg rn, int imm12)
762{
763    tcg_out_memop_12(s, cond, INSN_STRB_IMM, rt, rn, imm12, 1, 0);
764}
765
766static void tcg_out_ld8_r(TCGContext *s, ARMCond cond, TCGReg rt,
767                          TCGReg rn, TCGReg rm)
768{
769    tcg_out_memop_r(s, cond, INSN_LDRB_REG, rt, rn, rm, 1, 1, 0);
770}
771
772static void tcg_out_st8_r(TCGContext *s, ARMCond cond, TCGReg rt,
773                          TCGReg rn, TCGReg rm)
774{
775    tcg_out_memop_r(s, cond, INSN_STRB_REG, rt, rn, rm, 1, 1, 0);
776}
777
778static void tcg_out_ld8s_8(TCGContext *s, ARMCond cond, TCGReg rt,
779                           TCGReg rn, int imm8)
780{
781    tcg_out_memop_8(s, cond, INSN_LDRSB_IMM, rt, rn, imm8, 1, 0);
782}
783
784static void tcg_out_ld8s_r(TCGContext *s, ARMCond cond, TCGReg rt,
785                           TCGReg rn, TCGReg rm)
786{
787    tcg_out_memop_r(s, cond, INSN_LDRSB_REG, rt, rn, rm, 1, 1, 0);
788}
789
790static void tcg_out_movi_pool(TCGContext *s, ARMCond cond,
791                              TCGReg rd, uint32_t arg)
792{
793    new_pool_label(s, arg, R_ARM_PC13, s->code_ptr, 0);
794    tcg_out_ld32_12(s, cond, rd, TCG_REG_PC, 0);
795}
796
797static void tcg_out_movi32(TCGContext *s, ARMCond cond,
798                           TCGReg rd, uint32_t arg)
799{
800    int imm12, diff, opc, sh1, sh2;
801    uint32_t tt0, tt1, tt2;
802
803    /* Check a single MOV/MVN before anything else.  */
804    imm12 = encode_imm(arg);
805    if (imm12 >= 0) {
806        tcg_out_dat_imm(s, cond, ARITH_MOV, rd, 0, imm12);
807        return;
808    }
809    imm12 = encode_imm(~arg);
810    if (imm12 >= 0) {
811        tcg_out_dat_imm(s, cond, ARITH_MVN, rd, 0, imm12);
812        return;
813    }
814
815    /* Check for a pc-relative address.  This will usually be the TB,
816       or within the TB, which is immediately before the code block.  */
817    diff = tcg_pcrel_diff(s, (void *)arg) - 8;
818    if (diff >= 0) {
819        imm12 = encode_imm(diff);
820        if (imm12 >= 0) {
821            tcg_out_dat_imm(s, cond, ARITH_ADD, rd, TCG_REG_PC, imm12);
822            return;
823        }
824    } else {
825        imm12 = encode_imm(-diff);
826        if (imm12 >= 0) {
827            tcg_out_dat_imm(s, cond, ARITH_SUB, rd, TCG_REG_PC, imm12);
828            return;
829        }
830    }
831
832    /* Use movw + movt.  */
833    if (use_armv7_instructions) {
834        /* movw */
835        tcg_out32(s, (cond << 28) | 0x03000000 | (rd << 12)
836                  | ((arg << 4) & 0x000f0000) | (arg & 0xfff));
837        if (arg & 0xffff0000) {
838            /* movt */
839            tcg_out32(s, (cond << 28) | 0x03400000 | (rd << 12)
840                      | ((arg >> 12) & 0x000f0000) | ((arg >> 16) & 0xfff));
841        }
842        return;
843    }
844
845    /* Look for sequences of two insns.  If we have lots of 1's, we can
846       shorten the sequence by beginning with mvn and then clearing
847       higher bits with eor.  */
848    tt0 = arg;
849    opc = ARITH_MOV;
850    if (ctpop32(arg) > 16) {
851        tt0 = ~arg;
852        opc = ARITH_MVN;
853    }
854    sh1 = ctz32(tt0) & ~1;
855    tt1 = tt0 & ~(0xff << sh1);
856    sh2 = ctz32(tt1) & ~1;
857    tt2 = tt1 & ~(0xff << sh2);
858    if (tt2 == 0) {
859        int rot;
860
861        rot = ((32 - sh1) << 7) & 0xf00;
862        tcg_out_dat_imm(s, cond, opc, rd,  0, ((tt0 >> sh1) & 0xff) | rot);
863        rot = ((32 - sh2) << 7) & 0xf00;
864        tcg_out_dat_imm(s, cond, ARITH_EOR, rd, rd,
865                        ((tt0 >> sh2) & 0xff) | rot);
866        return;
867    }
868
869    /* Otherwise, drop it into the constant pool.  */
870    tcg_out_movi_pool(s, cond, rd, arg);
871}
872
873/*
874 * Emit either the reg,imm or reg,reg form of a data-processing insn.
875 * rhs must satisfy the "rI" constraint.
876 */
877static void tcg_out_dat_rI(TCGContext *s, ARMCond cond, ARMInsn opc,
878                           TCGReg dst, TCGReg lhs, TCGArg rhs, int rhs_is_const)
879{
880    if (rhs_is_const) {
881        tcg_out_dat_imm(s, cond, opc, dst, lhs, encode_imm_nofail(rhs));
882    } else {
883        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
884    }
885}
886
887/*
888 * Emit either the reg,imm or reg,reg form of a data-processing insn.
889 * rhs must satisfy the "rIK" constraint.
890 */
891static void tcg_out_dat_rIK(TCGContext *s, ARMCond cond, ARMInsn opc,
892                            ARMInsn opinv, TCGReg dst, TCGReg lhs, TCGArg rhs,
893                            bool rhs_is_const)
894{
895    if (rhs_is_const) {
896        int imm12 = encode_imm(rhs);
897        if (imm12 < 0) {
898            imm12 = encode_imm_nofail(~rhs);
899            opc = opinv;
900        }
901        tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
902    } else {
903        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
904    }
905}
906
907static void tcg_out_dat_rIN(TCGContext *s, ARMCond cond, ARMInsn opc,
908                            ARMInsn opneg, TCGReg dst, TCGReg lhs, TCGArg rhs,
909                            bool rhs_is_const)
910{
911    /* Emit either the reg,imm or reg,reg form of a data-processing insn.
912     * rhs must satisfy the "rIN" constraint.
913     */
914    if (rhs_is_const) {
915        int imm12 = encode_imm(rhs);
916        if (imm12 < 0) {
917            imm12 = encode_imm_nofail(-rhs);
918            opc = opneg;
919        }
920        tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
921    } else {
922        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
923    }
924}
925
926static void tcg_out_mul32(TCGContext *s, ARMCond cond, TCGReg rd,
927                          TCGReg rn, TCGReg rm)
928{
929    /* mul */
930    tcg_out32(s, (cond << 28) | 0x90 | (rd << 16) | (rm << 8) | rn);
931}
932
933static void tcg_out_umull32(TCGContext *s, ARMCond cond, TCGReg rd0,
934                            TCGReg rd1, TCGReg rn, TCGReg rm)
935{
936    /* umull */
937    tcg_out32(s, (cond << 28) | 0x00800090 |
938              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
939}
940
941static void tcg_out_smull32(TCGContext *s, ARMCond cond, TCGReg rd0,
942                            TCGReg rd1, TCGReg rn, TCGReg rm)
943{
944    /* smull */
945    tcg_out32(s, (cond << 28) | 0x00c00090 |
946              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
947}
948
949static void tcg_out_sdiv(TCGContext *s, ARMCond cond,
950                         TCGReg rd, TCGReg rn, TCGReg rm)
951{
952    tcg_out32(s, 0x0710f010 | (cond << 28) | (rd << 16) | rn | (rm << 8));
953}
954
955static void tcg_out_udiv(TCGContext *s, ARMCond cond,
956                         TCGReg rd, TCGReg rn, TCGReg rm)
957{
958    tcg_out32(s, 0x0730f010 | (cond << 28) | (rd << 16) | rn | (rm << 8));
959}
960
961static void tcg_out_ext8s(TCGContext *s, TCGType t, TCGReg rd, TCGReg rn)
962{
963    /* sxtb */
964    tcg_out32(s, 0x06af0070 | (COND_AL << 28) | (rd << 12) | rn);
965}
966
967static void tcg_out_ext8u(TCGContext *s, TCGReg rd, TCGReg rn)
968{
969    tcg_out_dat_imm(s, COND_AL, ARITH_AND, rd, rn, 0xff);
970}
971
972static void tcg_out_ext16s(TCGContext *s, TCGType t, TCGReg rd, TCGReg rn)
973{
974    /* sxth */
975    tcg_out32(s, 0x06bf0070 | (COND_AL << 28) | (rd << 12) | rn);
976}
977
978static void tcg_out_ext16u(TCGContext *s, TCGReg rd, TCGReg rn)
979{
980    /* uxth */
981    tcg_out32(s, 0x06ff0070 | (COND_AL << 28) | (rd << 12) | rn);
982}
983
984static void tcg_out_ext32s(TCGContext *s, TCGReg rd, TCGReg rn)
985{
986    g_assert_not_reached();
987}
988
989static void tcg_out_ext32u(TCGContext *s, TCGReg rd, TCGReg rn)
990{
991    g_assert_not_reached();
992}
993
994static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
995{
996    g_assert_not_reached();
997}
998
999static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
1000{
1001    g_assert_not_reached();
1002}
1003
1004static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg rd, TCGReg rn)
1005{
1006    g_assert_not_reached();
1007}
1008
1009static void tcg_out_bswap16(TCGContext *s, ARMCond cond,
1010                            TCGReg rd, TCGReg rn, int flags)
1011{
1012    if (flags & TCG_BSWAP_OS) {
1013        /* revsh */
1014        tcg_out32(s, 0x06ff0fb0 | (cond << 28) | (rd << 12) | rn);
1015        return;
1016    }
1017
1018    /* rev16 */
1019    tcg_out32(s, 0x06bf0fb0 | (cond << 28) | (rd << 12) | rn);
1020    if ((flags & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1021        /* uxth */
1022        tcg_out32(s, 0x06ff0070 | (cond << 28) | (rd << 12) | rd);
1023    }
1024}
1025
1026static void tcg_out_bswap32(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rn)
1027{
1028    /* rev */
1029    tcg_out32(s, 0x06bf0f30 | (cond << 28) | (rd << 12) | rn);
1030}
1031
1032static void tcg_out_deposit(TCGContext *s, ARMCond cond, TCGReg rd,
1033                            TCGArg a1, int ofs, int len, bool const_a1)
1034{
1035    if (const_a1) {
1036        /* bfi becomes bfc with rn == 15.  */
1037        a1 = 15;
1038    }
1039    /* bfi/bfc */
1040    tcg_out32(s, 0x07c00010 | (cond << 28) | (rd << 12) | a1
1041              | (ofs << 7) | ((ofs + len - 1) << 16));
1042}
1043
1044static void tcg_out_extract(TCGContext *s, ARMCond cond, TCGReg rd,
1045                            TCGReg rn, int ofs, int len)
1046{
1047    /* ubfx */
1048    tcg_out32(s, 0x07e00050 | (cond << 28) | (rd << 12) | rn
1049              | (ofs << 7) | ((len - 1) << 16));
1050}
1051
1052static void tcg_out_sextract(TCGContext *s, ARMCond cond, TCGReg rd,
1053                             TCGReg rn, int ofs, int len)
1054{
1055    /* sbfx */
1056    tcg_out32(s, 0x07a00050 | (cond << 28) | (rd << 12) | rn
1057              | (ofs << 7) | ((len - 1) << 16));
1058}
1059
1060static void tcg_out_ld32u(TCGContext *s, ARMCond cond,
1061                          TCGReg rd, TCGReg rn, int32_t offset)
1062{
1063    if (offset > 0xfff || offset < -0xfff) {
1064        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1065        tcg_out_ld32_r(s, cond, rd, rn, TCG_REG_TMP);
1066    } else
1067        tcg_out_ld32_12(s, cond, rd, rn, offset);
1068}
1069
1070static void tcg_out_st32(TCGContext *s, ARMCond cond,
1071                         TCGReg rd, TCGReg rn, int32_t offset)
1072{
1073    if (offset > 0xfff || offset < -0xfff) {
1074        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1075        tcg_out_st32_r(s, cond, rd, rn, TCG_REG_TMP);
1076    } else
1077        tcg_out_st32_12(s, cond, rd, rn, offset);
1078}
1079
1080static void tcg_out_ld16u(TCGContext *s, ARMCond cond,
1081                          TCGReg rd, TCGReg rn, int32_t offset)
1082{
1083    if (offset > 0xff || offset < -0xff) {
1084        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1085        tcg_out_ld16u_r(s, cond, rd, rn, TCG_REG_TMP);
1086    } else
1087        tcg_out_ld16u_8(s, cond, rd, rn, offset);
1088}
1089
1090static void tcg_out_ld16s(TCGContext *s, ARMCond cond,
1091                          TCGReg rd, TCGReg rn, int32_t offset)
1092{
1093    if (offset > 0xff || offset < -0xff) {
1094        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1095        tcg_out_ld16s_r(s, cond, rd, rn, TCG_REG_TMP);
1096    } else
1097        tcg_out_ld16s_8(s, cond, rd, rn, offset);
1098}
1099
1100static void tcg_out_st16(TCGContext *s, ARMCond cond,
1101                         TCGReg rd, TCGReg rn, int32_t offset)
1102{
1103    if (offset > 0xff || offset < -0xff) {
1104        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1105        tcg_out_st16_r(s, cond, rd, rn, TCG_REG_TMP);
1106    } else
1107        tcg_out_st16_8(s, cond, rd, rn, offset);
1108}
1109
1110static void tcg_out_ld8u(TCGContext *s, ARMCond cond,
1111                         TCGReg rd, TCGReg rn, int32_t offset)
1112{
1113    if (offset > 0xfff || offset < -0xfff) {
1114        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1115        tcg_out_ld8_r(s, cond, rd, rn, TCG_REG_TMP);
1116    } else
1117        tcg_out_ld8_12(s, cond, rd, rn, offset);
1118}
1119
1120static void tcg_out_ld8s(TCGContext *s, ARMCond cond,
1121                         TCGReg rd, TCGReg rn, int32_t offset)
1122{
1123    if (offset > 0xff || offset < -0xff) {
1124        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1125        tcg_out_ld8s_r(s, cond, rd, rn, TCG_REG_TMP);
1126    } else
1127        tcg_out_ld8s_8(s, cond, rd, rn, offset);
1128}
1129
1130static void tcg_out_st8(TCGContext *s, ARMCond cond,
1131                        TCGReg rd, TCGReg rn, int32_t offset)
1132{
1133    if (offset > 0xfff || offset < -0xfff) {
1134        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1135        tcg_out_st8_r(s, cond, rd, rn, TCG_REG_TMP);
1136    } else
1137        tcg_out_st8_12(s, cond, rd, rn, offset);
1138}
1139
1140/*
1141 * The _goto case is normally between TBs within the same code buffer, and
1142 * with the code buffer limited to 16MB we wouldn't need the long case.
1143 * But we also use it for the tail-call to the qemu_ld/st helpers, which does.
1144 */
1145static void tcg_out_goto(TCGContext *s, ARMCond cond, const tcg_insn_unit *addr)
1146{
1147    intptr_t addri = (intptr_t)addr;
1148    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1149    bool arm_mode = !(addri & 1);
1150
1151    if (arm_mode && disp - 8 < 0x01fffffd && disp - 8 > -0x01fffffd) {
1152        tcg_out_b_imm(s, cond, disp);
1153        return;
1154    }
1155
1156    /* LDR is interworking from v5t. */
1157    tcg_out_movi_pool(s, cond, TCG_REG_PC, addri);
1158}
1159
1160/*
1161 * The call case is mostly used for helpers - so it's not unreasonable
1162 * for them to be beyond branch range.
1163 */
1164static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *addr)
1165{
1166    intptr_t addri = (intptr_t)addr;
1167    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1168    bool arm_mode = !(addri & 1);
1169
1170    if (disp - 8 < 0x02000000 && disp - 8 >= -0x02000000) {
1171        if (arm_mode) {
1172            tcg_out_bl_imm(s, COND_AL, disp);
1173        } else {
1174            tcg_out_blx_imm(s, disp);
1175        }
1176        return;
1177    }
1178
1179    tcg_out_movi32(s, COND_AL, TCG_REG_TMP, addri);
1180    tcg_out_blx_reg(s, COND_AL, TCG_REG_TMP);
1181}
1182
1183static void tcg_out_call(TCGContext *s, const tcg_insn_unit *addr,
1184                         const TCGHelperInfo *info)
1185{
1186    tcg_out_call_int(s, addr);
1187}
1188
1189static void tcg_out_goto_label(TCGContext *s, ARMCond cond, TCGLabel *l)
1190{
1191    if (l->has_value) {
1192        tcg_out_goto(s, cond, l->u.value_ptr);
1193    } else {
1194        tcg_out_reloc(s, s->code_ptr, R_ARM_PC24, l, 0);
1195        tcg_out_b_imm(s, cond, 0);
1196    }
1197}
1198
1199static void tcg_out_mb(TCGContext *s, TCGArg a0)
1200{
1201    if (use_armv7_instructions) {
1202        tcg_out32(s, INSN_DMB_ISH);
1203    } else {
1204        tcg_out32(s, INSN_DMB_MCR);
1205    }
1206}
1207
1208static TCGCond tcg_out_cmp2(TCGContext *s, const TCGArg *args,
1209                            const int *const_args)
1210{
1211    TCGReg al = args[0];
1212    TCGReg ah = args[1];
1213    TCGArg bl = args[2];
1214    TCGArg bh = args[3];
1215    TCGCond cond = args[4];
1216    int const_bl = const_args[2];
1217    int const_bh = const_args[3];
1218
1219    switch (cond) {
1220    case TCG_COND_EQ:
1221    case TCG_COND_NE:
1222    case TCG_COND_LTU:
1223    case TCG_COND_LEU:
1224    case TCG_COND_GTU:
1225    case TCG_COND_GEU:
1226        /* We perform a conditional comparision.  If the high half is
1227           equal, then overwrite the flags with the comparison of the
1228           low half.  The resulting flags cover the whole.  */
1229        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, ah, bh, const_bh);
1230        tcg_out_dat_rI(s, COND_EQ, ARITH_CMP, 0, al, bl, const_bl);
1231        return cond;
1232
1233    case TCG_COND_LT:
1234    case TCG_COND_GE:
1235        /* We perform a double-word subtraction and examine the result.
1236           We do not actually need the result of the subtract, so the
1237           low part "subtract" is a compare.  For the high half we have
1238           no choice but to compute into a temporary.  */
1239        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, al, bl, const_bl);
1240        tcg_out_dat_rI(s, COND_AL, ARITH_SBC | TO_CPSR,
1241                       TCG_REG_TMP, ah, bh, const_bh);
1242        return cond;
1243
1244    case TCG_COND_LE:
1245    case TCG_COND_GT:
1246        /* Similar, but with swapped arguments, via reversed subtract.  */
1247        tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR,
1248                       TCG_REG_TMP, al, bl, const_bl);
1249        tcg_out_dat_rI(s, COND_AL, ARITH_RSC | TO_CPSR,
1250                       TCG_REG_TMP, ah, bh, const_bh);
1251        return tcg_swap_cond(cond);
1252
1253    default:
1254        g_assert_not_reached();
1255    }
1256}
1257
1258/*
1259 * Note that TCGReg references Q-registers.
1260 * Q-regno = 2 * D-regno, so shift left by 1 whlie inserting.
1261 */
1262static uint32_t encode_vd(TCGReg rd)
1263{
1264    tcg_debug_assert(rd >= TCG_REG_Q0);
1265    return (extract32(rd, 3, 1) << 22) | (extract32(rd, 0, 3) << 13);
1266}
1267
1268static uint32_t encode_vn(TCGReg rn)
1269{
1270    tcg_debug_assert(rn >= TCG_REG_Q0);
1271    return (extract32(rn, 3, 1) << 7) | (extract32(rn, 0, 3) << 17);
1272}
1273
1274static uint32_t encode_vm(TCGReg rm)
1275{
1276    tcg_debug_assert(rm >= TCG_REG_Q0);
1277    return (extract32(rm, 3, 1) << 5) | (extract32(rm, 0, 3) << 1);
1278}
1279
1280static void tcg_out_vreg2(TCGContext *s, ARMInsn insn, int q, int vece,
1281                          TCGReg d, TCGReg m)
1282{
1283    tcg_out32(s, insn | (vece << 18) | (q << 6) |
1284              encode_vd(d) | encode_vm(m));
1285}
1286
1287static void tcg_out_vreg3(TCGContext *s, ARMInsn insn, int q, int vece,
1288                          TCGReg d, TCGReg n, TCGReg m)
1289{
1290    tcg_out32(s, insn | (vece << 20) | (q << 6) |
1291              encode_vd(d) | encode_vn(n) | encode_vm(m));
1292}
1293
1294static void tcg_out_vmovi(TCGContext *s, TCGReg rd,
1295                          int q, int op, int cmode, uint8_t imm8)
1296{
1297    tcg_out32(s, INSN_VMOVI | encode_vd(rd) | (q << 6) | (op << 5)
1298              | (cmode << 8) | extract32(imm8, 0, 4)
1299              | (extract32(imm8, 4, 3) << 16)
1300              | (extract32(imm8, 7, 1) << 24));
1301}
1302
1303static void tcg_out_vshifti(TCGContext *s, ARMInsn insn, int q,
1304                            TCGReg rd, TCGReg rm, int l_imm6)
1305{
1306    tcg_out32(s, insn | (q << 6) | encode_vd(rd) | encode_vm(rm) |
1307              (extract32(l_imm6, 6, 1) << 7) |
1308              (extract32(l_imm6, 0, 6) << 16));
1309}
1310
1311static void tcg_out_vldst(TCGContext *s, ARMInsn insn,
1312                          TCGReg rd, TCGReg rn, int offset)
1313{
1314    if (offset != 0) {
1315        if (check_fit_imm(offset) || check_fit_imm(-offset)) {
1316            tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
1317                            TCG_REG_TMP, rn, offset, true);
1318        } else {
1319            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP, offset);
1320            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1321                            TCG_REG_TMP, TCG_REG_TMP, rn, 0);
1322        }
1323        rn = TCG_REG_TMP;
1324    }
1325    tcg_out32(s, insn | (rn << 16) | encode_vd(rd) | 0xf);
1326}
1327
1328typedef struct {
1329    ARMCond cond;
1330    TCGReg base;
1331    int index;
1332    bool index_scratch;
1333} HostAddress;
1334
1335#ifdef CONFIG_SOFTMMU
1336static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
1337{
1338    /* We arrive at the slow path via "BLNE", so R14 contains l->raddr. */
1339    return TCG_REG_R14;
1340}
1341
1342static const TCGLdstHelperParam ldst_helper_param = {
1343    .ra_gen = ldst_ra_gen,
1344    .ntmp = 1,
1345    .tmp = { TCG_REG_TMP },
1346};
1347
1348static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1349{
1350    MemOp opc = get_memop(lb->oi);
1351
1352    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1353        return false;
1354    }
1355
1356    tcg_out_ld_helper_args(s, lb, &ldst_helper_param);
1357    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE]);
1358    tcg_out_ld_helper_ret(s, lb, false, &ldst_helper_param);
1359
1360    tcg_out_goto(s, COND_AL, lb->raddr);
1361    return true;
1362}
1363
1364static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1365{
1366    MemOp opc = get_memop(lb->oi);
1367
1368    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1369        return false;
1370    }
1371
1372    tcg_out_st_helper_args(s, lb, &ldst_helper_param);
1373
1374    /* Tail-call to the helper, which will return to the fast path.  */
1375    tcg_out_goto(s, COND_AL, qemu_st_helpers[opc & MO_SIZE]);
1376    return true;
1377}
1378#else
1379static bool tcg_out_fail_alignment(TCGContext *s, TCGLabelQemuLdst *l)
1380{
1381    if (!reloc_pc24(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1382        return false;
1383    }
1384
1385    if (TARGET_LONG_BITS == 64) {
1386        /* 64-bit target address is aligned into R2:R3. */
1387        TCGMovExtend ext[2] = {
1388            { .dst = TCG_REG_R2, .dst_type = TCG_TYPE_I32,
1389              .src = l->addrlo_reg,
1390              .src_type = TCG_TYPE_I32, .src_ext = MO_UL },
1391            { .dst = TCG_REG_R3, .dst_type = TCG_TYPE_I32,
1392              .src = l->addrhi_reg,
1393              .src_type = TCG_TYPE_I32, .src_ext = MO_UL },
1394        };
1395        tcg_out_movext2(s, &ext[0], &ext[1], TCG_REG_TMP);
1396    } else {
1397        tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R1, l->addrlo_reg);
1398    }
1399    tcg_out_mov(s, TCG_TYPE_PTR, TCG_REG_R0, TCG_AREG0);
1400
1401    /*
1402     * Tail call to the helper, with the return address back inline,
1403     * just for the clarity of the debugging traceback -- the helper
1404     * cannot return.  We have used BLNE to arrive here, so LR is
1405     * already set.
1406     */
1407    tcg_out_goto(s, COND_AL, (const void *)
1408                 (l->is_ld ? helper_unaligned_ld : helper_unaligned_st));
1409    return true;
1410}
1411
1412static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
1413{
1414    return tcg_out_fail_alignment(s, l);
1415}
1416
1417static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
1418{
1419    return tcg_out_fail_alignment(s, l);
1420}
1421#endif /* SOFTMMU */
1422
1423static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
1424                                           TCGReg addrlo, TCGReg addrhi,
1425                                           MemOpIdx oi, bool is_ld)
1426{
1427    TCGLabelQemuLdst *ldst = NULL;
1428    MemOp opc = get_memop(oi);
1429    MemOp a_bits = get_alignment_bits(opc);
1430    unsigned a_mask = (1 << a_bits) - 1;
1431
1432#ifdef CONFIG_SOFTMMU
1433    int mem_index = get_mmuidx(oi);
1434    int cmp_off = is_ld ? offsetof(CPUTLBEntry, addr_read)
1435                        : offsetof(CPUTLBEntry, addr_write);
1436    int fast_off = TLB_MASK_TABLE_OFS(mem_index);
1437    unsigned s_mask = (1 << (opc & MO_SIZE)) - 1;
1438    TCGReg t_addr;
1439
1440    ldst = new_ldst_label(s);
1441    ldst->is_ld = is_ld;
1442    ldst->oi = oi;
1443    ldst->addrlo_reg = addrlo;
1444    ldst->addrhi_reg = addrhi;
1445
1446    /* Load env_tlb(env)->f[mmu_idx].{mask,table} into {r0,r1}.  */
1447    QEMU_BUILD_BUG_ON(TLB_MASK_TABLE_OFS(0) > 0);
1448    QEMU_BUILD_BUG_ON(TLB_MASK_TABLE_OFS(0) < -256);
1449    QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, mask) != 0);
1450    QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, table) != 4);
1451    tcg_out_ldrd_8(s, COND_AL, TCG_REG_R0, TCG_AREG0, fast_off);
1452
1453    /* Extract the tlb index from the address into R0.  */
1454    tcg_out_dat_reg(s, COND_AL, ARITH_AND, TCG_REG_R0, TCG_REG_R0, addrlo,
1455                    SHIFT_IMM_LSR(TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS));
1456
1457    /*
1458     * Add the tlb_table pointer, creating the CPUTLBEntry address in R1.
1459     * Load the tlb comparator into R2/R3 and the fast path addend into R1.
1460     */
1461    if (cmp_off == 0) {
1462        if (TARGET_LONG_BITS == 64) {
1463            tcg_out_ldrd_rwb(s, COND_AL, TCG_REG_R2, TCG_REG_R1, TCG_REG_R0);
1464        } else {
1465            tcg_out_ld32_rwb(s, COND_AL, TCG_REG_R2, TCG_REG_R1, TCG_REG_R0);
1466        }
1467    } else {
1468        tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1469                        TCG_REG_R1, TCG_REG_R1, TCG_REG_R0, 0);
1470        if (TARGET_LONG_BITS == 64) {
1471            tcg_out_ldrd_8(s, COND_AL, TCG_REG_R2, TCG_REG_R1, cmp_off);
1472        } else {
1473            tcg_out_ld32_12(s, COND_AL, TCG_REG_R2, TCG_REG_R1, cmp_off);
1474        }
1475    }
1476
1477    /* Load the tlb addend.  */
1478    tcg_out_ld32_12(s, COND_AL, TCG_REG_R1, TCG_REG_R1,
1479                    offsetof(CPUTLBEntry, addend));
1480
1481    /*
1482     * Check alignment, check comparators.
1483     * Do this in 2-4 insns.  Use MOVW for v7, if possible,
1484     * to reduce the number of sequential conditional instructions.
1485     * Almost all guests have at least 4k pages, which means that we need
1486     * to clear at least 9 bits even for an 8-byte memory, which means it
1487     * isn't worth checking for an immediate operand for BIC.
1488     *
1489     * For unaligned accesses, test the page of the last unit of alignment.
1490     * This leaves the least significant alignment bits unchanged, and of
1491     * course must be zero.
1492     */
1493    t_addr = addrlo;
1494    if (a_mask < s_mask) {
1495        t_addr = TCG_REG_R0;
1496        tcg_out_dat_imm(s, COND_AL, ARITH_ADD, t_addr,
1497                        addrlo, s_mask - a_mask);
1498    }
1499    if (use_armv7_instructions && TARGET_PAGE_BITS <= 16) {
1500        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, ~(TARGET_PAGE_MASK | a_mask));
1501        tcg_out_dat_reg(s, COND_AL, ARITH_BIC, TCG_REG_TMP,
1502                        t_addr, TCG_REG_TMP, 0);
1503        tcg_out_dat_reg(s, COND_AL, ARITH_CMP, 0, TCG_REG_R2, TCG_REG_TMP, 0);
1504    } else {
1505        if (a_mask) {
1506            tcg_debug_assert(a_mask <= 0xff);
1507            tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addrlo, a_mask);
1508        }
1509        tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0, t_addr,
1510                        SHIFT_IMM_LSR(TARGET_PAGE_BITS));
1511        tcg_out_dat_reg(s, (a_mask ? COND_EQ : COND_AL), ARITH_CMP,
1512                        0, TCG_REG_R2, TCG_REG_TMP,
1513                        SHIFT_IMM_LSL(TARGET_PAGE_BITS));
1514    }
1515
1516    if (TARGET_LONG_BITS == 64) {
1517        tcg_out_dat_reg(s, COND_EQ, ARITH_CMP, 0, TCG_REG_R3, addrhi, 0);
1518    }
1519
1520    *h = (HostAddress){
1521        .cond = COND_AL,
1522        .base = addrlo,
1523        .index = TCG_REG_R1,
1524        .index_scratch = true,
1525    };
1526#else
1527    if (a_mask) {
1528        ldst = new_ldst_label(s);
1529        ldst->is_ld = is_ld;
1530        ldst->oi = oi;
1531        ldst->addrlo_reg = addrlo;
1532        ldst->addrhi_reg = addrhi;
1533
1534        /* We are expecting a_bits to max out at 7 */
1535        tcg_debug_assert(a_mask <= 0xff);
1536        /* tst addr, #mask */
1537        tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addrlo, a_mask);
1538    }
1539
1540    *h = (HostAddress){
1541        .cond = COND_AL,
1542        .base = addrlo,
1543        .index = guest_base ? TCG_REG_GUEST_BASE : -1,
1544        .index_scratch = false,
1545    };
1546#endif
1547
1548    return ldst;
1549}
1550
1551static void tcg_out_qemu_ld_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1552                                   TCGReg datahi, HostAddress h)
1553{
1554    TCGReg base;
1555
1556    /* Byte swapping is left to middle-end expansion. */
1557    tcg_debug_assert((opc & MO_BSWAP) == 0);
1558
1559    switch (opc & MO_SSIZE) {
1560    case MO_UB:
1561        if (h.index < 0) {
1562            tcg_out_ld8_12(s, h.cond, datalo, h.base, 0);
1563        } else {
1564            tcg_out_ld8_r(s, h.cond, datalo, h.base, h.index);
1565        }
1566        break;
1567    case MO_SB:
1568        if (h.index < 0) {
1569            tcg_out_ld8s_8(s, h.cond, datalo, h.base, 0);
1570        } else {
1571            tcg_out_ld8s_r(s, h.cond, datalo, h.base, h.index);
1572        }
1573        break;
1574    case MO_UW:
1575        if (h.index < 0) {
1576            tcg_out_ld16u_8(s, h.cond, datalo, h.base, 0);
1577        } else {
1578            tcg_out_ld16u_r(s, h.cond, datalo, h.base, h.index);
1579        }
1580        break;
1581    case MO_SW:
1582        if (h.index < 0) {
1583            tcg_out_ld16s_8(s, h.cond, datalo, h.base, 0);
1584        } else {
1585            tcg_out_ld16s_r(s, h.cond, datalo, h.base, h.index);
1586        }
1587        break;
1588    case MO_UL:
1589        if (h.index < 0) {
1590            tcg_out_ld32_12(s, h.cond, datalo, h.base, 0);
1591        } else {
1592            tcg_out_ld32_r(s, h.cond, datalo, h.base, h.index);
1593        }
1594        break;
1595    case MO_UQ:
1596        /* We used pair allocation for datalo, so already should be aligned. */
1597        tcg_debug_assert((datalo & 1) == 0);
1598        tcg_debug_assert(datahi == datalo + 1);
1599        /* LDRD requires alignment; double-check that. */
1600        if (get_alignment_bits(opc) >= MO_64) {
1601            if (h.index < 0) {
1602                tcg_out_ldrd_8(s, h.cond, datalo, h.base, 0);
1603                break;
1604            }
1605            /*
1606             * Rm (the second address op) must not overlap Rt or Rt + 1.
1607             * Since datalo is aligned, we can simplify the test via alignment.
1608             * Flip the two address arguments if that works.
1609             */
1610            if ((h.index & ~1) != datalo) {
1611                tcg_out_ldrd_r(s, h.cond, datalo, h.base, h.index);
1612                break;
1613            }
1614            if ((h.base & ~1) != datalo) {
1615                tcg_out_ldrd_r(s, h.cond, datalo, h.index, h.base);
1616                break;
1617            }
1618        }
1619        if (h.index < 0) {
1620            base = h.base;
1621            if (datalo == h.base) {
1622                tcg_out_mov_reg(s, h.cond, TCG_REG_TMP, base);
1623                base = TCG_REG_TMP;
1624            }
1625        } else if (h.index_scratch) {
1626            tcg_out_ld32_rwb(s, h.cond, datalo, h.index, h.base);
1627            tcg_out_ld32_12(s, h.cond, datahi, h.index, 4);
1628            break;
1629        } else {
1630            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1631                            h.base, h.index, SHIFT_IMM_LSL(0));
1632            base = TCG_REG_TMP;
1633        }
1634        tcg_out_ld32_12(s, h.cond, datalo, base, 0);
1635        tcg_out_ld32_12(s, h.cond, datahi, base, 4);
1636        break;
1637    default:
1638        g_assert_not_reached();
1639    }
1640}
1641
1642static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
1643                            TCGReg addrlo, TCGReg addrhi,
1644                            MemOpIdx oi, TCGType data_type)
1645{
1646    MemOp opc = get_memop(oi);
1647    TCGLabelQemuLdst *ldst;
1648    HostAddress h;
1649
1650    ldst = prepare_host_addr(s, &h, addrlo, addrhi, oi, true);
1651    if (ldst) {
1652        ldst->type = data_type;
1653        ldst->datalo_reg = datalo;
1654        ldst->datahi_reg = datahi;
1655
1656        /*
1657         * This a conditional BL only to load a pointer within this
1658         * opcode into LR for the slow path.  We will not be using
1659         * the value for a tail call.
1660         */
1661        ldst->label_ptr[0] = s->code_ptr;
1662        tcg_out_bl_imm(s, COND_NE, 0);
1663
1664        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1665        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1666    } else {
1667        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1668    }
1669}
1670
1671static void tcg_out_qemu_st_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1672                                   TCGReg datahi, HostAddress h)
1673{
1674    /* Byte swapping is left to middle-end expansion. */
1675    tcg_debug_assert((opc & MO_BSWAP) == 0);
1676
1677    switch (opc & MO_SIZE) {
1678    case MO_8:
1679        if (h.index < 0) {
1680            tcg_out_st8_12(s, h.cond, datalo, h.base, 0);
1681        } else {
1682            tcg_out_st8_r(s, h.cond, datalo, h.base, h.index);
1683        }
1684        break;
1685    case MO_16:
1686        if (h.index < 0) {
1687            tcg_out_st16_8(s, h.cond, datalo, h.base, 0);
1688        } else {
1689            tcg_out_st16_r(s, h.cond, datalo, h.base, h.index);
1690        }
1691        break;
1692    case MO_32:
1693        if (h.index < 0) {
1694            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1695        } else {
1696            tcg_out_st32_r(s, h.cond, datalo, h.base, h.index);
1697        }
1698        break;
1699    case MO_64:
1700        /* We used pair allocation for datalo, so already should be aligned. */
1701        tcg_debug_assert((datalo & 1) == 0);
1702        tcg_debug_assert(datahi == datalo + 1);
1703        /* STRD requires alignment; double-check that. */
1704        if (get_alignment_bits(opc) >= MO_64) {
1705            if (h.index < 0) {
1706                tcg_out_strd_8(s, h.cond, datalo, h.base, 0);
1707            } else {
1708                tcg_out_strd_r(s, h.cond, datalo, h.base, h.index);
1709            }
1710        } else if (h.index_scratch) {
1711            tcg_out_st32_rwb(s, h.cond, datalo, h.index, h.base);
1712            tcg_out_st32_12(s, h.cond, datahi, h.index, 4);
1713        } else {
1714            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1715                            h.base, h.index, SHIFT_IMM_LSL(0));
1716            tcg_out_st32_12(s, h.cond, datalo, TCG_REG_TMP, 0);
1717            tcg_out_st32_12(s, h.cond, datahi, TCG_REG_TMP, 4);
1718        }
1719        break;
1720    default:
1721        g_assert_not_reached();
1722    }
1723}
1724
1725static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
1726                            TCGReg addrlo, TCGReg addrhi,
1727                            MemOpIdx oi, TCGType data_type)
1728{
1729    MemOp opc = get_memop(oi);
1730    TCGLabelQemuLdst *ldst;
1731    HostAddress h;
1732
1733    ldst = prepare_host_addr(s, &h, addrlo, addrhi, oi, false);
1734    if (ldst) {
1735        ldst->type = data_type;
1736        ldst->datalo_reg = datalo;
1737        ldst->datahi_reg = datahi;
1738
1739        h.cond = COND_EQ;
1740        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1741
1742        /* The conditional call is last, as we're going to return here. */
1743        ldst->label_ptr[0] = s->code_ptr;
1744        tcg_out_bl_imm(s, COND_NE, 0);
1745        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1746    } else {
1747        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1748    }
1749}
1750
1751static void tcg_out_epilogue(TCGContext *s);
1752
1753static void tcg_out_exit_tb(TCGContext *s, uintptr_t arg)
1754{
1755    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, arg);
1756    tcg_out_epilogue(s);
1757}
1758
1759static void tcg_out_goto_tb(TCGContext *s, int which)
1760{
1761    uintptr_t i_addr;
1762    intptr_t i_disp;
1763
1764    /* Direct branch will be patched by tb_target_set_jmp_target. */
1765    set_jmp_insn_offset(s, which);
1766    tcg_out32(s, INSN_NOP);
1767
1768    /* When branch is out of range, fall through to indirect. */
1769    i_addr = get_jmp_target_addr(s, which);
1770    i_disp = tcg_pcrel_diff(s, (void *)i_addr) - 8;
1771    tcg_debug_assert(i_disp < 0);
1772    if (i_disp >= -0xfff) {
1773        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_PC, i_disp);
1774    } else {
1775        /*
1776         * The TB is close, but outside the 12 bits addressable by
1777         * the load.  We can extend this to 20 bits with a sub of a
1778         * shifted immediate from pc.
1779         */
1780        int h = -i_disp;
1781        int l = h & 0xfff;
1782
1783        h = encode_imm_nofail(h - l);
1784        tcg_out_dat_imm(s, COND_AL, ARITH_SUB, TCG_REG_R0, TCG_REG_PC, h);
1785        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_R0, l);
1786    }
1787    set_jmp_reset_offset(s, which);
1788}
1789
1790void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1791                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1792{
1793    uintptr_t addr = tb->jmp_target_addr[n];
1794    ptrdiff_t offset = addr - (jmp_rx + 8);
1795    tcg_insn_unit insn;
1796
1797    /* Either directly branch, or fall through to indirect branch. */
1798    if (offset == sextract64(offset, 0, 26)) {
1799        /* B <addr> */
1800        insn = deposit32((COND_AL << 28) | INSN_B, 0, 24, offset >> 2);
1801    } else {
1802        insn = INSN_NOP;
1803    }
1804
1805    qatomic_set((uint32_t *)jmp_rw, insn);
1806    flush_idcache_range(jmp_rx, jmp_rw, 4);
1807}
1808
1809static void tcg_out_op(TCGContext *s, TCGOpcode opc,
1810                       const TCGArg args[TCG_MAX_OP_ARGS],
1811                       const int const_args[TCG_MAX_OP_ARGS])
1812{
1813    TCGArg a0, a1, a2, a3, a4, a5;
1814    int c;
1815
1816    switch (opc) {
1817    case INDEX_op_goto_ptr:
1818        tcg_out_b_reg(s, COND_AL, args[0]);
1819        break;
1820    case INDEX_op_br:
1821        tcg_out_goto_label(s, COND_AL, arg_label(args[0]));
1822        break;
1823
1824    case INDEX_op_ld8u_i32:
1825        tcg_out_ld8u(s, COND_AL, args[0], args[1], args[2]);
1826        break;
1827    case INDEX_op_ld8s_i32:
1828        tcg_out_ld8s(s, COND_AL, args[0], args[1], args[2]);
1829        break;
1830    case INDEX_op_ld16u_i32:
1831        tcg_out_ld16u(s, COND_AL, args[0], args[1], args[2]);
1832        break;
1833    case INDEX_op_ld16s_i32:
1834        tcg_out_ld16s(s, COND_AL, args[0], args[1], args[2]);
1835        break;
1836    case INDEX_op_ld_i32:
1837        tcg_out_ld32u(s, COND_AL, args[0], args[1], args[2]);
1838        break;
1839    case INDEX_op_st8_i32:
1840        tcg_out_st8(s, COND_AL, args[0], args[1], args[2]);
1841        break;
1842    case INDEX_op_st16_i32:
1843        tcg_out_st16(s, COND_AL, args[0], args[1], args[2]);
1844        break;
1845    case INDEX_op_st_i32:
1846        tcg_out_st32(s, COND_AL, args[0], args[1], args[2]);
1847        break;
1848
1849    case INDEX_op_movcond_i32:
1850        /* Constraints mean that v2 is always in the same register as dest,
1851         * so we only need to do "if condition passed, move v1 to dest".
1852         */
1853        tcg_out_dat_rIN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0,
1854                        args[1], args[2], const_args[2]);
1855        tcg_out_dat_rIK(s, tcg_cond_to_arm_cond[args[5]], ARITH_MOV,
1856                        ARITH_MVN, args[0], 0, args[3], const_args[3]);
1857        break;
1858    case INDEX_op_add_i32:
1859        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
1860                        args[0], args[1], args[2], const_args[2]);
1861        break;
1862    case INDEX_op_sub_i32:
1863        if (const_args[1]) {
1864            if (const_args[2]) {
1865                tcg_out_movi32(s, COND_AL, args[0], args[1] - args[2]);
1866            } else {
1867                tcg_out_dat_rI(s, COND_AL, ARITH_RSB,
1868                               args[0], args[2], args[1], 1);
1869            }
1870        } else {
1871            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB, ARITH_ADD,
1872                            args[0], args[1], args[2], const_args[2]);
1873        }
1874        break;
1875    case INDEX_op_and_i32:
1876        tcg_out_dat_rIK(s, COND_AL, ARITH_AND, ARITH_BIC,
1877                        args[0], args[1], args[2], const_args[2]);
1878        break;
1879    case INDEX_op_andc_i32:
1880        tcg_out_dat_rIK(s, COND_AL, ARITH_BIC, ARITH_AND,
1881                        args[0], args[1], args[2], const_args[2]);
1882        break;
1883    case INDEX_op_or_i32:
1884        c = ARITH_ORR;
1885        goto gen_arith;
1886    case INDEX_op_xor_i32:
1887        c = ARITH_EOR;
1888        /* Fall through.  */
1889    gen_arith:
1890        tcg_out_dat_rI(s, COND_AL, c, args[0], args[1], args[2], const_args[2]);
1891        break;
1892    case INDEX_op_add2_i32:
1893        a0 = args[0], a1 = args[1], a2 = args[2];
1894        a3 = args[3], a4 = args[4], a5 = args[5];
1895        if (a0 == a3 || (a0 == a5 && !const_args[5])) {
1896            a0 = TCG_REG_TMP;
1897        }
1898        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD | TO_CPSR, ARITH_SUB | TO_CPSR,
1899                        a0, a2, a4, const_args[4]);
1900        tcg_out_dat_rIK(s, COND_AL, ARITH_ADC, ARITH_SBC,
1901                        a1, a3, a5, const_args[5]);
1902        tcg_out_mov_reg(s, COND_AL, args[0], a0);
1903        break;
1904    case INDEX_op_sub2_i32:
1905        a0 = args[0], a1 = args[1], a2 = args[2];
1906        a3 = args[3], a4 = args[4], a5 = args[5];
1907        if ((a0 == a3 && !const_args[3]) || (a0 == a5 && !const_args[5])) {
1908            a0 = TCG_REG_TMP;
1909        }
1910        if (const_args[2]) {
1911            if (const_args[4]) {
1912                tcg_out_movi32(s, COND_AL, a0, a4);
1913                a4 = a0;
1914            }
1915            tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR, a0, a4, a2, 1);
1916        } else {
1917            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB | TO_CPSR,
1918                            ARITH_ADD | TO_CPSR, a0, a2, a4, const_args[4]);
1919        }
1920        if (const_args[3]) {
1921            if (const_args[5]) {
1922                tcg_out_movi32(s, COND_AL, a1, a5);
1923                a5 = a1;
1924            }
1925            tcg_out_dat_rI(s, COND_AL, ARITH_RSC, a1, a5, a3, 1);
1926        } else {
1927            tcg_out_dat_rIK(s, COND_AL, ARITH_SBC, ARITH_ADC,
1928                            a1, a3, a5, const_args[5]);
1929        }
1930        tcg_out_mov_reg(s, COND_AL, args[0], a0);
1931        break;
1932    case INDEX_op_neg_i32:
1933        tcg_out_dat_imm(s, COND_AL, ARITH_RSB, args[0], args[1], 0);
1934        break;
1935    case INDEX_op_not_i32:
1936        tcg_out_dat_reg(s, COND_AL,
1937                        ARITH_MVN, args[0], 0, args[1], SHIFT_IMM_LSL(0));
1938        break;
1939    case INDEX_op_mul_i32:
1940        tcg_out_mul32(s, COND_AL, args[0], args[1], args[2]);
1941        break;
1942    case INDEX_op_mulu2_i32:
1943        tcg_out_umull32(s, COND_AL, args[0], args[1], args[2], args[3]);
1944        break;
1945    case INDEX_op_muls2_i32:
1946        tcg_out_smull32(s, COND_AL, args[0], args[1], args[2], args[3]);
1947        break;
1948    /* XXX: Perhaps args[2] & 0x1f is wrong */
1949    case INDEX_op_shl_i32:
1950        c = const_args[2] ?
1951                SHIFT_IMM_LSL(args[2] & 0x1f) : SHIFT_REG_LSL(args[2]);
1952        goto gen_shift32;
1953    case INDEX_op_shr_i32:
1954        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_LSR(args[2] & 0x1f) :
1955                SHIFT_IMM_LSL(0) : SHIFT_REG_LSR(args[2]);
1956        goto gen_shift32;
1957    case INDEX_op_sar_i32:
1958        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ASR(args[2] & 0x1f) :
1959                SHIFT_IMM_LSL(0) : SHIFT_REG_ASR(args[2]);
1960        goto gen_shift32;
1961    case INDEX_op_rotr_i32:
1962        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ROR(args[2] & 0x1f) :
1963                SHIFT_IMM_LSL(0) : SHIFT_REG_ROR(args[2]);
1964        /* Fall through.  */
1965    gen_shift32:
1966        tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1], c);
1967        break;
1968
1969    case INDEX_op_rotl_i32:
1970        if (const_args[2]) {
1971            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
1972                            ((0x20 - args[2]) & 0x1f) ?
1973                            SHIFT_IMM_ROR((0x20 - args[2]) & 0x1f) :
1974                            SHIFT_IMM_LSL(0));
1975        } else {
1976            tcg_out_dat_imm(s, COND_AL, ARITH_RSB, TCG_REG_TMP, args[2], 0x20);
1977            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
1978                            SHIFT_REG_ROR(TCG_REG_TMP));
1979        }
1980        break;
1981
1982    case INDEX_op_ctz_i32:
1983        tcg_out_dat_reg(s, COND_AL, INSN_RBIT, TCG_REG_TMP, 0, args[1], 0);
1984        a1 = TCG_REG_TMP;
1985        goto do_clz;
1986
1987    case INDEX_op_clz_i32:
1988        a1 = args[1];
1989    do_clz:
1990        a0 = args[0];
1991        a2 = args[2];
1992        c = const_args[2];
1993        if (c && a2 == 32) {
1994            tcg_out_dat_reg(s, COND_AL, INSN_CLZ, a0, 0, a1, 0);
1995            break;
1996        }
1997        tcg_out_dat_imm(s, COND_AL, ARITH_CMP, 0, a1, 0);
1998        tcg_out_dat_reg(s, COND_NE, INSN_CLZ, a0, 0, a1, 0);
1999        if (c || a0 != a2) {
2000            tcg_out_dat_rIK(s, COND_EQ, ARITH_MOV, ARITH_MVN, a0, 0, a2, c);
2001        }
2002        break;
2003
2004    case INDEX_op_brcond_i32:
2005        tcg_out_dat_rIN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0,
2006                       args[0], args[1], const_args[1]);
2007        tcg_out_goto_label(s, tcg_cond_to_arm_cond[args[2]],
2008                           arg_label(args[3]));
2009        break;
2010    case INDEX_op_setcond_i32:
2011        tcg_out_dat_rIN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0,
2012                        args[1], args[2], const_args[2]);
2013        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[args[3]],
2014                        ARITH_MOV, args[0], 0, 1);
2015        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(args[3])],
2016                        ARITH_MOV, args[0], 0, 0);
2017        break;
2018
2019    case INDEX_op_brcond2_i32:
2020        c = tcg_out_cmp2(s, args, const_args);
2021        tcg_out_goto_label(s, tcg_cond_to_arm_cond[c], arg_label(args[5]));
2022        break;
2023    case INDEX_op_setcond2_i32:
2024        c = tcg_out_cmp2(s, args + 1, const_args + 1);
2025        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c], ARITH_MOV, args[0], 0, 1);
2026        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2027                        ARITH_MOV, args[0], 0, 0);
2028        break;
2029
2030    case INDEX_op_qemu_ld_i32:
2031        if (TARGET_LONG_BITS == 32) {
2032            tcg_out_qemu_ld(s, args[0], -1, args[1], -1,
2033                            args[2], TCG_TYPE_I32);
2034        } else {
2035            tcg_out_qemu_ld(s, args[0], -1, args[1], args[2],
2036                            args[3], TCG_TYPE_I32);
2037        }
2038        break;
2039    case INDEX_op_qemu_ld_i64:
2040        if (TARGET_LONG_BITS == 32) {
2041            tcg_out_qemu_ld(s, args[0], args[1], args[2], -1,
2042                            args[3], TCG_TYPE_I64);
2043        } else {
2044            tcg_out_qemu_ld(s, args[0], args[1], args[2], args[3],
2045                            args[4], TCG_TYPE_I64);
2046        }
2047        break;
2048    case INDEX_op_qemu_st_i32:
2049        if (TARGET_LONG_BITS == 32) {
2050            tcg_out_qemu_st(s, args[0], -1, args[1], -1,
2051                            args[2], TCG_TYPE_I32);
2052        } else {
2053            tcg_out_qemu_st(s, args[0], -1, args[1], args[2],
2054                            args[3], TCG_TYPE_I32);
2055        }
2056        break;
2057    case INDEX_op_qemu_st_i64:
2058        if (TARGET_LONG_BITS == 32) {
2059            tcg_out_qemu_st(s, args[0], args[1], args[2], -1,
2060                            args[3], TCG_TYPE_I64);
2061        } else {
2062            tcg_out_qemu_st(s, args[0], args[1], args[2], args[3],
2063                            args[4], TCG_TYPE_I64);
2064        }
2065        break;
2066
2067    case INDEX_op_bswap16_i32:
2068        tcg_out_bswap16(s, COND_AL, args[0], args[1], args[2]);
2069        break;
2070    case INDEX_op_bswap32_i32:
2071        tcg_out_bswap32(s, COND_AL, args[0], args[1]);
2072        break;
2073
2074    case INDEX_op_deposit_i32:
2075        tcg_out_deposit(s, COND_AL, args[0], args[2],
2076                        args[3], args[4], const_args[2]);
2077        break;
2078    case INDEX_op_extract_i32:
2079        tcg_out_extract(s, COND_AL, args[0], args[1], args[2], args[3]);
2080        break;
2081    case INDEX_op_sextract_i32:
2082        tcg_out_sextract(s, COND_AL, args[0], args[1], args[2], args[3]);
2083        break;
2084    case INDEX_op_extract2_i32:
2085        /* ??? These optimization vs zero should be generic.  */
2086        /* ??? But we can't substitute 2 for 1 in the opcode stream yet.  */
2087        if (const_args[1]) {
2088            if (const_args[2]) {
2089                tcg_out_movi(s, TCG_TYPE_REG, args[0], 0);
2090            } else {
2091                tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2092                                args[2], SHIFT_IMM_LSL(32 - args[3]));
2093            }
2094        } else if (const_args[2]) {
2095            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2096                            args[1], SHIFT_IMM_LSR(args[3]));
2097        } else {
2098            /* We can do extract2 in 2 insns, vs the 3 required otherwise.  */
2099            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0,
2100                            args[2], SHIFT_IMM_LSL(32 - args[3]));
2101            tcg_out_dat_reg(s, COND_AL, ARITH_ORR, args[0], TCG_REG_TMP,
2102                            args[1], SHIFT_IMM_LSR(args[3]));
2103        }
2104        break;
2105
2106    case INDEX_op_div_i32:
2107        tcg_out_sdiv(s, COND_AL, args[0], args[1], args[2]);
2108        break;
2109    case INDEX_op_divu_i32:
2110        tcg_out_udiv(s, COND_AL, args[0], args[1], args[2]);
2111        break;
2112
2113    case INDEX_op_mb:
2114        tcg_out_mb(s, args[0]);
2115        break;
2116
2117    case INDEX_op_mov_i32:  /* Always emitted via tcg_out_mov.  */
2118    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
2119    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
2120    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
2121    case INDEX_op_ext8s_i32:  /* Always emitted via tcg_reg_alloc_op.  */
2122    case INDEX_op_ext8u_i32:
2123    case INDEX_op_ext16s_i32:
2124    case INDEX_op_ext16u_i32:
2125    default:
2126        g_assert_not_reached();
2127    }
2128}
2129
2130static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
2131{
2132    switch (op) {
2133    case INDEX_op_goto_ptr:
2134        return C_O0_I1(r);
2135
2136    case INDEX_op_ld8u_i32:
2137    case INDEX_op_ld8s_i32:
2138    case INDEX_op_ld16u_i32:
2139    case INDEX_op_ld16s_i32:
2140    case INDEX_op_ld_i32:
2141    case INDEX_op_neg_i32:
2142    case INDEX_op_not_i32:
2143    case INDEX_op_bswap16_i32:
2144    case INDEX_op_bswap32_i32:
2145    case INDEX_op_ext8s_i32:
2146    case INDEX_op_ext16s_i32:
2147    case INDEX_op_ext16u_i32:
2148    case INDEX_op_extract_i32:
2149    case INDEX_op_sextract_i32:
2150        return C_O1_I1(r, r);
2151
2152    case INDEX_op_st8_i32:
2153    case INDEX_op_st16_i32:
2154    case INDEX_op_st_i32:
2155        return C_O0_I2(r, r);
2156
2157    case INDEX_op_add_i32:
2158    case INDEX_op_sub_i32:
2159    case INDEX_op_setcond_i32:
2160        return C_O1_I2(r, r, rIN);
2161
2162    case INDEX_op_and_i32:
2163    case INDEX_op_andc_i32:
2164    case INDEX_op_clz_i32:
2165    case INDEX_op_ctz_i32:
2166        return C_O1_I2(r, r, rIK);
2167
2168    case INDEX_op_mul_i32:
2169    case INDEX_op_div_i32:
2170    case INDEX_op_divu_i32:
2171        return C_O1_I2(r, r, r);
2172
2173    case INDEX_op_mulu2_i32:
2174    case INDEX_op_muls2_i32:
2175        return C_O2_I2(r, r, r, r);
2176
2177    case INDEX_op_or_i32:
2178    case INDEX_op_xor_i32:
2179        return C_O1_I2(r, r, rI);
2180
2181    case INDEX_op_shl_i32:
2182    case INDEX_op_shr_i32:
2183    case INDEX_op_sar_i32:
2184    case INDEX_op_rotl_i32:
2185    case INDEX_op_rotr_i32:
2186        return C_O1_I2(r, r, ri);
2187
2188    case INDEX_op_brcond_i32:
2189        return C_O0_I2(r, rIN);
2190    case INDEX_op_deposit_i32:
2191        return C_O1_I2(r, 0, rZ);
2192    case INDEX_op_extract2_i32:
2193        return C_O1_I2(r, rZ, rZ);
2194    case INDEX_op_movcond_i32:
2195        return C_O1_I4(r, r, rIN, rIK, 0);
2196    case INDEX_op_add2_i32:
2197        return C_O2_I4(r, r, r, r, rIN, rIK);
2198    case INDEX_op_sub2_i32:
2199        return C_O2_I4(r, r, rI, rI, rIN, rIK);
2200    case INDEX_op_brcond2_i32:
2201        return C_O0_I4(r, r, rI, rI);
2202    case INDEX_op_setcond2_i32:
2203        return C_O1_I4(r, r, r, rI, rI);
2204
2205    case INDEX_op_qemu_ld_i32:
2206        return TARGET_LONG_BITS == 32 ? C_O1_I1(r, l) : C_O1_I2(r, l, l);
2207    case INDEX_op_qemu_ld_i64:
2208        return TARGET_LONG_BITS == 32 ? C_O2_I1(e, p, l) : C_O2_I2(e, p, l, l);
2209    case INDEX_op_qemu_st_i32:
2210        return TARGET_LONG_BITS == 32 ? C_O0_I2(s, s) : C_O0_I3(s, s, s);
2211    case INDEX_op_qemu_st_i64:
2212        return TARGET_LONG_BITS == 32 ? C_O0_I3(S, p, s) : C_O0_I4(S, p, s, s);
2213
2214    case INDEX_op_st_vec:
2215        return C_O0_I2(w, r);
2216    case INDEX_op_ld_vec:
2217    case INDEX_op_dupm_vec:
2218        return C_O1_I1(w, r);
2219    case INDEX_op_dup_vec:
2220        return C_O1_I1(w, wr);
2221    case INDEX_op_abs_vec:
2222    case INDEX_op_neg_vec:
2223    case INDEX_op_not_vec:
2224    case INDEX_op_shli_vec:
2225    case INDEX_op_shri_vec:
2226    case INDEX_op_sari_vec:
2227        return C_O1_I1(w, w);
2228    case INDEX_op_dup2_vec:
2229    case INDEX_op_add_vec:
2230    case INDEX_op_mul_vec:
2231    case INDEX_op_smax_vec:
2232    case INDEX_op_smin_vec:
2233    case INDEX_op_ssadd_vec:
2234    case INDEX_op_sssub_vec:
2235    case INDEX_op_sub_vec:
2236    case INDEX_op_umax_vec:
2237    case INDEX_op_umin_vec:
2238    case INDEX_op_usadd_vec:
2239    case INDEX_op_ussub_vec:
2240    case INDEX_op_xor_vec:
2241    case INDEX_op_arm_sshl_vec:
2242    case INDEX_op_arm_ushl_vec:
2243        return C_O1_I2(w, w, w);
2244    case INDEX_op_arm_sli_vec:
2245        return C_O1_I2(w, 0, w);
2246    case INDEX_op_or_vec:
2247    case INDEX_op_andc_vec:
2248        return C_O1_I2(w, w, wO);
2249    case INDEX_op_and_vec:
2250    case INDEX_op_orc_vec:
2251        return C_O1_I2(w, w, wV);
2252    case INDEX_op_cmp_vec:
2253        return C_O1_I2(w, w, wZ);
2254    case INDEX_op_bitsel_vec:
2255        return C_O1_I3(w, w, w, w);
2256    default:
2257        g_assert_not_reached();
2258    }
2259}
2260
2261static void tcg_target_init(TCGContext *s)
2262{
2263    /*
2264     * Only probe for the platform and capabilities if we haven't already
2265     * determined maximum values at compile time.
2266     */
2267#if !defined(use_idiv_instructions) || !defined(use_neon_instructions)
2268    {
2269        unsigned long hwcap = qemu_getauxval(AT_HWCAP);
2270#ifndef use_idiv_instructions
2271        use_idiv_instructions = (hwcap & HWCAP_ARM_IDIVA) != 0;
2272#endif
2273#ifndef use_neon_instructions
2274        use_neon_instructions = (hwcap & HWCAP_ARM_NEON) != 0;
2275#endif
2276    }
2277#endif
2278
2279    if (__ARM_ARCH < 7) {
2280        const char *pl = (const char *)qemu_getauxval(AT_PLATFORM);
2281        if (pl != NULL && pl[0] == 'v' && pl[1] >= '4' && pl[1] <= '9') {
2282            arm_arch = pl[1] - '0';
2283        }
2284
2285        if (arm_arch < 6) {
2286            error_report("TCG: ARMv%d is unsupported; exiting", arm_arch);
2287            exit(EXIT_FAILURE);
2288        }
2289    }
2290
2291    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
2292
2293    tcg_target_call_clobber_regs = 0;
2294    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R0);
2295    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R1);
2296    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R2);
2297    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R3);
2298    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R12);
2299    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R14);
2300
2301    if (use_neon_instructions) {
2302        tcg_target_available_regs[TCG_TYPE_V64]  = ALL_VECTOR_REGS;
2303        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
2304
2305        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q0);
2306        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q1);
2307        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q2);
2308        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q3);
2309        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q8);
2310        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q9);
2311        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q10);
2312        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q11);
2313        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q12);
2314        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q13);
2315        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q14);
2316        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q15);
2317    }
2318
2319    s->reserved_regs = 0;
2320    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
2321    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP);
2322    tcg_regset_set_reg(s->reserved_regs, TCG_REG_PC);
2323    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP);
2324}
2325
2326static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
2327                       TCGReg arg1, intptr_t arg2)
2328{
2329    switch (type) {
2330    case TCG_TYPE_I32:
2331        tcg_out_ld32u(s, COND_AL, arg, arg1, arg2);
2332        return;
2333    case TCG_TYPE_V64:
2334        /* regs 1; size 8; align 8 */
2335        tcg_out_vldst(s, INSN_VLD1 | 0x7d0, arg, arg1, arg2);
2336        return;
2337    case TCG_TYPE_V128:
2338        /*
2339         * We have only 8-byte alignment for the stack per the ABI.
2340         * Rather than dynamically re-align the stack, it's easier
2341         * to simply not request alignment beyond that.  So:
2342         * regs 2; size 8; align 8
2343         */
2344        tcg_out_vldst(s, INSN_VLD1 | 0xad0, arg, arg1, arg2);
2345        return;
2346    default:
2347        g_assert_not_reached();
2348    }
2349}
2350
2351static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
2352                       TCGReg arg1, intptr_t arg2)
2353{
2354    switch (type) {
2355    case TCG_TYPE_I32:
2356        tcg_out_st32(s, COND_AL, arg, arg1, arg2);
2357        return;
2358    case TCG_TYPE_V64:
2359        /* regs 1; size 8; align 8 */
2360        tcg_out_vldst(s, INSN_VST1 | 0x7d0, arg, arg1, arg2);
2361        return;
2362    case TCG_TYPE_V128:
2363        /* See tcg_out_ld re alignment: regs 2; size 8; align 8 */
2364        tcg_out_vldst(s, INSN_VST1 | 0xad0, arg, arg1, arg2);
2365        return;
2366    default:
2367        g_assert_not_reached();
2368    }
2369}
2370
2371static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
2372                        TCGReg base, intptr_t ofs)
2373{
2374    return false;
2375}
2376
2377static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
2378{
2379    if (ret == arg) {
2380        return true;
2381    }
2382    switch (type) {
2383    case TCG_TYPE_I32:
2384        if (ret < TCG_REG_Q0 && arg < TCG_REG_Q0) {
2385            tcg_out_mov_reg(s, COND_AL, ret, arg);
2386            return true;
2387        }
2388        return false;
2389
2390    case TCG_TYPE_V64:
2391    case TCG_TYPE_V128:
2392        /* "VMOV D,N" is an alias for "VORR D,N,N". */
2393        tcg_out_vreg3(s, INSN_VORR, type - TCG_TYPE_V64, 0, ret, arg, arg);
2394        return true;
2395
2396    default:
2397        g_assert_not_reached();
2398    }
2399}
2400
2401static void tcg_out_movi(TCGContext *s, TCGType type,
2402                         TCGReg ret, tcg_target_long arg)
2403{
2404    tcg_debug_assert(type == TCG_TYPE_I32);
2405    tcg_debug_assert(ret < TCG_REG_Q0);
2406    tcg_out_movi32(s, COND_AL, ret, arg);
2407}
2408
2409static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
2410{
2411    return false;
2412}
2413
2414static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
2415                             tcg_target_long imm)
2416{
2417    int enc, opc = ARITH_ADD;
2418
2419    /* All of the easiest immediates to encode are positive. */
2420    if (imm < 0) {
2421        imm = -imm;
2422        opc = ARITH_SUB;
2423    }
2424    enc = encode_imm(imm);
2425    if (enc >= 0) {
2426        tcg_out_dat_imm(s, COND_AL, opc, rd, rs, enc);
2427    } else {
2428        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, imm);
2429        tcg_out_dat_reg(s, COND_AL, opc, rd, rs,
2430                        TCG_REG_TMP, SHIFT_IMM_LSL(0));
2431    }
2432}
2433
2434/* Type is always V128, with I64 elements.  */
2435static void tcg_out_dup2_vec(TCGContext *s, TCGReg rd, TCGReg rl, TCGReg rh)
2436{
2437    /* Move high element into place first. */
2438    /* VMOV Dd+1, Ds */
2439    tcg_out_vreg3(s, INSN_VORR | (1 << 12), 0, 0, rd, rh, rh);
2440    /* Move low element into place; tcg_out_mov will check for nop. */
2441    tcg_out_mov(s, TCG_TYPE_V64, rd, rl);
2442}
2443
2444static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
2445                            TCGReg rd, TCGReg rs)
2446{
2447    int q = type - TCG_TYPE_V64;
2448
2449    if (vece == MO_64) {
2450        if (type == TCG_TYPE_V128) {
2451            tcg_out_dup2_vec(s, rd, rs, rs);
2452        } else {
2453            tcg_out_mov(s, TCG_TYPE_V64, rd, rs);
2454        }
2455    } else if (rs < TCG_REG_Q0) {
2456        int b = (vece == MO_8);
2457        int e = (vece == MO_16);
2458        tcg_out32(s, INSN_VDUP_G | (b << 22) | (q << 21) | (e << 5) |
2459                  encode_vn(rd) | (rs << 12));
2460    } else {
2461        int imm4 = 1 << vece;
2462        tcg_out32(s, INSN_VDUP_S | (imm4 << 16) | (q << 6) |
2463                  encode_vd(rd) | encode_vm(rs));
2464    }
2465    return true;
2466}
2467
2468static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
2469                             TCGReg rd, TCGReg base, intptr_t offset)
2470{
2471    if (vece == MO_64) {
2472        tcg_out_ld(s, TCG_TYPE_V64, rd, base, offset);
2473        if (type == TCG_TYPE_V128) {
2474            tcg_out_dup2_vec(s, rd, rd, rd);
2475        }
2476    } else {
2477        int q = type - TCG_TYPE_V64;
2478        tcg_out_vldst(s, INSN_VLD1R | (vece << 6) | (q << 5),
2479                      rd, base, offset);
2480    }
2481    return true;
2482}
2483
2484static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
2485                             TCGReg rd, int64_t v64)
2486{
2487    int q = type - TCG_TYPE_V64;
2488    int cmode, imm8, i;
2489
2490    /* Test all bytes equal first.  */
2491    if (vece == MO_8) {
2492        tcg_out_vmovi(s, rd, q, 0, 0xe, v64);
2493        return;
2494    }
2495
2496    /*
2497     * Test all bytes 0x00 or 0xff second.  This can match cases that
2498     * might otherwise take 2 or 3 insns for MO_16 or MO_32 below.
2499     */
2500    for (i = imm8 = 0; i < 8; i++) {
2501        uint8_t byte = v64 >> (i * 8);
2502        if (byte == 0xff) {
2503            imm8 |= 1 << i;
2504        } else if (byte != 0) {
2505            goto fail_bytes;
2506        }
2507    }
2508    tcg_out_vmovi(s, rd, q, 1, 0xe, imm8);
2509    return;
2510 fail_bytes:
2511
2512    /*
2513     * Tests for various replications.  For each element width, if we
2514     * cannot find an expansion there's no point checking a larger
2515     * width because we already know by replication it cannot match.
2516     */
2517    if (vece == MO_16) {
2518        uint16_t v16 = v64;
2519
2520        if (is_shimm16(v16, &cmode, &imm8)) {
2521            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2522            return;
2523        }
2524        if (is_shimm16(~v16, &cmode, &imm8)) {
2525            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2526            return;
2527        }
2528
2529        /*
2530         * Otherwise, all remaining constants can be loaded in two insns:
2531         * rd = v16 & 0xff, rd |= v16 & 0xff00.
2532         */
2533        tcg_out_vmovi(s, rd, q, 0, 0x8, v16 & 0xff);
2534        tcg_out_vmovi(s, rd, q, 0, 0xb, v16 >> 8);   /* VORRI */
2535        return;
2536    }
2537
2538    if (vece == MO_32) {
2539        uint32_t v32 = v64;
2540
2541        if (is_shimm32(v32, &cmode, &imm8) ||
2542            is_soimm32(v32, &cmode, &imm8)) {
2543            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2544            return;
2545        }
2546        if (is_shimm32(~v32, &cmode, &imm8) ||
2547            is_soimm32(~v32, &cmode, &imm8)) {
2548            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2549            return;
2550        }
2551
2552        /*
2553         * Restrict the set of constants to those we can load with
2554         * two instructions.  Others we load from the pool.
2555         */
2556        i = is_shimm32_pair(v32, &cmode, &imm8);
2557        if (i) {
2558            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2559            tcg_out_vmovi(s, rd, q, 0, i | 1, extract32(v32, i * 4, 8));
2560            return;
2561        }
2562        i = is_shimm32_pair(~v32, &cmode, &imm8);
2563        if (i) {
2564            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2565            tcg_out_vmovi(s, rd, q, 1, i | 1, extract32(~v32, i * 4, 8));
2566            return;
2567        }
2568    }
2569
2570    /*
2571     * As a last resort, load from the constant pool.
2572     */
2573    if (!q || vece == MO_64) {
2574        new_pool_l2(s, R_ARM_PC11, s->code_ptr, 0, v64, v64 >> 32);
2575        /* VLDR Dd, [pc + offset] */
2576        tcg_out32(s, INSN_VLDR_D | encode_vd(rd) | (0xf << 16));
2577        if (q) {
2578            tcg_out_dup2_vec(s, rd, rd, rd);
2579        }
2580    } else {
2581        new_pool_label(s, (uint32_t)v64, R_ARM_PC8, s->code_ptr, 0);
2582        /* add tmp, pc, offset */
2583        tcg_out_dat_imm(s, COND_AL, ARITH_ADD, TCG_REG_TMP, TCG_REG_PC, 0);
2584        tcg_out_dupm_vec(s, type, MO_32, rd, TCG_REG_TMP, 0);
2585    }
2586}
2587
2588static const ARMInsn vec_cmp_insn[16] = {
2589    [TCG_COND_EQ] = INSN_VCEQ,
2590    [TCG_COND_GT] = INSN_VCGT,
2591    [TCG_COND_GE] = INSN_VCGE,
2592    [TCG_COND_GTU] = INSN_VCGT_U,
2593    [TCG_COND_GEU] = INSN_VCGE_U,
2594};
2595
2596static const ARMInsn vec_cmp0_insn[16] = {
2597    [TCG_COND_EQ] = INSN_VCEQ0,
2598    [TCG_COND_GT] = INSN_VCGT0,
2599    [TCG_COND_GE] = INSN_VCGE0,
2600    [TCG_COND_LT] = INSN_VCLT0,
2601    [TCG_COND_LE] = INSN_VCLE0,
2602};
2603
2604static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2605                           unsigned vecl, unsigned vece,
2606                           const TCGArg args[TCG_MAX_OP_ARGS],
2607                           const int const_args[TCG_MAX_OP_ARGS])
2608{
2609    TCGType type = vecl + TCG_TYPE_V64;
2610    unsigned q = vecl;
2611    TCGArg a0, a1, a2, a3;
2612    int cmode, imm8;
2613
2614    a0 = args[0];
2615    a1 = args[1];
2616    a2 = args[2];
2617
2618    switch (opc) {
2619    case INDEX_op_ld_vec:
2620        tcg_out_ld(s, type, a0, a1, a2);
2621        return;
2622    case INDEX_op_st_vec:
2623        tcg_out_st(s, type, a0, a1, a2);
2624        return;
2625    case INDEX_op_dupm_vec:
2626        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
2627        return;
2628    case INDEX_op_dup2_vec:
2629        tcg_out_dup2_vec(s, a0, a1, a2);
2630        return;
2631    case INDEX_op_abs_vec:
2632        tcg_out_vreg2(s, INSN_VABS, q, vece, a0, a1);
2633        return;
2634    case INDEX_op_neg_vec:
2635        tcg_out_vreg2(s, INSN_VNEG, q, vece, a0, a1);
2636        return;
2637    case INDEX_op_not_vec:
2638        tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a1);
2639        return;
2640    case INDEX_op_add_vec:
2641        tcg_out_vreg3(s, INSN_VADD, q, vece, a0, a1, a2);
2642        return;
2643    case INDEX_op_mul_vec:
2644        tcg_out_vreg3(s, INSN_VMUL, q, vece, a0, a1, a2);
2645        return;
2646    case INDEX_op_smax_vec:
2647        tcg_out_vreg3(s, INSN_VMAX, q, vece, a0, a1, a2);
2648        return;
2649    case INDEX_op_smin_vec:
2650        tcg_out_vreg3(s, INSN_VMIN, q, vece, a0, a1, a2);
2651        return;
2652    case INDEX_op_sub_vec:
2653        tcg_out_vreg3(s, INSN_VSUB, q, vece, a0, a1, a2);
2654        return;
2655    case INDEX_op_ssadd_vec:
2656        tcg_out_vreg3(s, INSN_VQADD, q, vece, a0, a1, a2);
2657        return;
2658    case INDEX_op_sssub_vec:
2659        tcg_out_vreg3(s, INSN_VQSUB, q, vece, a0, a1, a2);
2660        return;
2661    case INDEX_op_umax_vec:
2662        tcg_out_vreg3(s, INSN_VMAX_U, q, vece, a0, a1, a2);
2663        return;
2664    case INDEX_op_umin_vec:
2665        tcg_out_vreg3(s, INSN_VMIN_U, q, vece, a0, a1, a2);
2666        return;
2667    case INDEX_op_usadd_vec:
2668        tcg_out_vreg3(s, INSN_VQADD_U, q, vece, a0, a1, a2);
2669        return;
2670    case INDEX_op_ussub_vec:
2671        tcg_out_vreg3(s, INSN_VQSUB_U, q, vece, a0, a1, a2);
2672        return;
2673    case INDEX_op_xor_vec:
2674        tcg_out_vreg3(s, INSN_VEOR, q, 0, a0, a1, a2);
2675        return;
2676    case INDEX_op_arm_sshl_vec:
2677        /*
2678         * Note that Vm is the data and Vn is the shift count,
2679         * therefore the arguments appear reversed.
2680         */
2681        tcg_out_vreg3(s, INSN_VSHL_S, q, vece, a0, a2, a1);
2682        return;
2683    case INDEX_op_arm_ushl_vec:
2684        /* See above. */
2685        tcg_out_vreg3(s, INSN_VSHL_U, q, vece, a0, a2, a1);
2686        return;
2687    case INDEX_op_shli_vec:
2688        tcg_out_vshifti(s, INSN_VSHLI, q, a0, a1, a2 + (8 << vece));
2689        return;
2690    case INDEX_op_shri_vec:
2691        tcg_out_vshifti(s, INSN_VSHRI, q, a0, a1, (16 << vece) - a2);
2692        return;
2693    case INDEX_op_sari_vec:
2694        tcg_out_vshifti(s, INSN_VSARI, q, a0, a1, (16 << vece) - a2);
2695        return;
2696    case INDEX_op_arm_sli_vec:
2697        tcg_out_vshifti(s, INSN_VSLI, q, a0, a2, args[3] + (8 << vece));
2698        return;
2699
2700    case INDEX_op_andc_vec:
2701        if (!const_args[2]) {
2702            tcg_out_vreg3(s, INSN_VBIC, q, 0, a0, a1, a2);
2703            return;
2704        }
2705        a2 = ~a2;
2706        /* fall through */
2707    case INDEX_op_and_vec:
2708        if (const_args[2]) {
2709            is_shimm1632(~a2, &cmode, &imm8);
2710            if (a0 == a1) {
2711                tcg_out_vmovi(s, a0, q, 1, cmode | 1, imm8); /* VBICI */
2712                return;
2713            }
2714            tcg_out_vmovi(s, a0, q, 1, cmode, imm8); /* VMVNI */
2715            a2 = a0;
2716        }
2717        tcg_out_vreg3(s, INSN_VAND, q, 0, a0, a1, a2);
2718        return;
2719
2720    case INDEX_op_orc_vec:
2721        if (!const_args[2]) {
2722            tcg_out_vreg3(s, INSN_VORN, q, 0, a0, a1, a2);
2723            return;
2724        }
2725        a2 = ~a2;
2726        /* fall through */
2727    case INDEX_op_or_vec:
2728        if (const_args[2]) {
2729            is_shimm1632(a2, &cmode, &imm8);
2730            if (a0 == a1) {
2731                tcg_out_vmovi(s, a0, q, 0, cmode | 1, imm8); /* VORRI */
2732                return;
2733            }
2734            tcg_out_vmovi(s, a0, q, 0, cmode, imm8); /* VMOVI */
2735            a2 = a0;
2736        }
2737        tcg_out_vreg3(s, INSN_VORR, q, 0, a0, a1, a2);
2738        return;
2739
2740    case INDEX_op_cmp_vec:
2741        {
2742            TCGCond cond = args[3];
2743
2744            if (cond == TCG_COND_NE) {
2745                if (const_args[2]) {
2746                    tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a1);
2747                } else {
2748                    tcg_out_vreg3(s, INSN_VCEQ, q, vece, a0, a1, a2);
2749                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
2750                }
2751            } else {
2752                ARMInsn insn;
2753
2754                if (const_args[2]) {
2755                    insn = vec_cmp0_insn[cond];
2756                    if (insn) {
2757                        tcg_out_vreg2(s, insn, q, vece, a0, a1);
2758                        return;
2759                    }
2760                    tcg_out_dupi_vec(s, type, MO_8, TCG_VEC_TMP, 0);
2761                    a2 = TCG_VEC_TMP;
2762                }
2763                insn = vec_cmp_insn[cond];
2764                if (insn == 0) {
2765                    TCGArg t;
2766                    t = a1, a1 = a2, a2 = t;
2767                    cond = tcg_swap_cond(cond);
2768                    insn = vec_cmp_insn[cond];
2769                    tcg_debug_assert(insn != 0);
2770                }
2771                tcg_out_vreg3(s, insn, q, vece, a0, a1, a2);
2772            }
2773        }
2774        return;
2775
2776    case INDEX_op_bitsel_vec:
2777        a3 = args[3];
2778        if (a0 == a3) {
2779            tcg_out_vreg3(s, INSN_VBIT, q, 0, a0, a2, a1);
2780        } else if (a0 == a2) {
2781            tcg_out_vreg3(s, INSN_VBIF, q, 0, a0, a3, a1);
2782        } else {
2783            tcg_out_mov(s, type, a0, a1);
2784            tcg_out_vreg3(s, INSN_VBSL, q, 0, a0, a2, a3);
2785        }
2786        return;
2787
2788    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
2789    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
2790    default:
2791        g_assert_not_reached();
2792    }
2793}
2794
2795int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
2796{
2797    switch (opc) {
2798    case INDEX_op_add_vec:
2799    case INDEX_op_sub_vec:
2800    case INDEX_op_and_vec:
2801    case INDEX_op_andc_vec:
2802    case INDEX_op_or_vec:
2803    case INDEX_op_orc_vec:
2804    case INDEX_op_xor_vec:
2805    case INDEX_op_not_vec:
2806    case INDEX_op_shli_vec:
2807    case INDEX_op_shri_vec:
2808    case INDEX_op_sari_vec:
2809    case INDEX_op_ssadd_vec:
2810    case INDEX_op_sssub_vec:
2811    case INDEX_op_usadd_vec:
2812    case INDEX_op_ussub_vec:
2813    case INDEX_op_bitsel_vec:
2814        return 1;
2815    case INDEX_op_abs_vec:
2816    case INDEX_op_cmp_vec:
2817    case INDEX_op_mul_vec:
2818    case INDEX_op_neg_vec:
2819    case INDEX_op_smax_vec:
2820    case INDEX_op_smin_vec:
2821    case INDEX_op_umax_vec:
2822    case INDEX_op_umin_vec:
2823        return vece < MO_64;
2824    case INDEX_op_shlv_vec:
2825    case INDEX_op_shrv_vec:
2826    case INDEX_op_sarv_vec:
2827    case INDEX_op_rotli_vec:
2828    case INDEX_op_rotlv_vec:
2829    case INDEX_op_rotrv_vec:
2830        return -1;
2831    default:
2832        return 0;
2833    }
2834}
2835
2836void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
2837                       TCGArg a0, ...)
2838{
2839    va_list va;
2840    TCGv_vec v0, v1, v2, t1, t2, c1;
2841    TCGArg a2;
2842
2843    va_start(va, a0);
2844    v0 = temp_tcgv_vec(arg_temp(a0));
2845    v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
2846    a2 = va_arg(va, TCGArg);
2847    va_end(va);
2848
2849    switch (opc) {
2850    case INDEX_op_shlv_vec:
2851        /*
2852         * Merely propagate shlv_vec to arm_ushl_vec.
2853         * In this way we don't set TCG_TARGET_HAS_shv_vec
2854         * because everything is done via expansion.
2855         */
2856        v2 = temp_tcgv_vec(arg_temp(a2));
2857        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
2858                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
2859        break;
2860
2861    case INDEX_op_shrv_vec:
2862    case INDEX_op_sarv_vec:
2863        /* Right shifts are negative left shifts for NEON.  */
2864        v2 = temp_tcgv_vec(arg_temp(a2));
2865        t1 = tcg_temp_new_vec(type);
2866        tcg_gen_neg_vec(vece, t1, v2);
2867        if (opc == INDEX_op_shrv_vec) {
2868            opc = INDEX_op_arm_ushl_vec;
2869        } else {
2870            opc = INDEX_op_arm_sshl_vec;
2871        }
2872        vec_gen_3(opc, type, vece, tcgv_vec_arg(v0),
2873                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2874        tcg_temp_free_vec(t1);
2875        break;
2876
2877    case INDEX_op_rotli_vec:
2878        t1 = tcg_temp_new_vec(type);
2879        tcg_gen_shri_vec(vece, t1, v1, -a2 & ((8 << vece) - 1));
2880        vec_gen_4(INDEX_op_arm_sli_vec, type, vece,
2881                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(v1), a2);
2882        tcg_temp_free_vec(t1);
2883        break;
2884
2885    case INDEX_op_rotlv_vec:
2886        v2 = temp_tcgv_vec(arg_temp(a2));
2887        t1 = tcg_temp_new_vec(type);
2888        c1 = tcg_constant_vec(type, vece, 8 << vece);
2889        tcg_gen_sub_vec(vece, t1, v2, c1);
2890        /* Right shifts are negative left shifts for NEON.  */
2891        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
2892                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2893        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
2894                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
2895        tcg_gen_or_vec(vece, v0, v0, t1);
2896        tcg_temp_free_vec(t1);
2897        break;
2898
2899    case INDEX_op_rotrv_vec:
2900        v2 = temp_tcgv_vec(arg_temp(a2));
2901        t1 = tcg_temp_new_vec(type);
2902        t2 = tcg_temp_new_vec(type);
2903        c1 = tcg_constant_vec(type, vece, 8 << vece);
2904        tcg_gen_neg_vec(vece, t1, v2);
2905        tcg_gen_sub_vec(vece, t2, c1, v2);
2906        /* Right shifts are negative left shifts for NEON.  */
2907        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
2908                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2909        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t2),
2910                  tcgv_vec_arg(v1), tcgv_vec_arg(t2));
2911        tcg_gen_or_vec(vece, v0, t1, t2);
2912        tcg_temp_free_vec(t1);
2913        tcg_temp_free_vec(t2);
2914        break;
2915
2916    default:
2917        g_assert_not_reached();
2918    }
2919}
2920
2921static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
2922{
2923    int i;
2924    for (i = 0; i < count; ++i) {
2925        p[i] = INSN_NOP;
2926    }
2927}
2928
2929/* Compute frame size via macros, to share between tcg_target_qemu_prologue
2930   and tcg_register_jit.  */
2931
2932#define PUSH_SIZE  ((11 - 4 + 1 + 1) * sizeof(tcg_target_long))
2933
2934#define FRAME_SIZE \
2935    ((PUSH_SIZE \
2936      + TCG_STATIC_CALL_ARGS_SIZE \
2937      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
2938      + TCG_TARGET_STACK_ALIGN - 1) \
2939     & -TCG_TARGET_STACK_ALIGN)
2940
2941#define STACK_ADDEND  (FRAME_SIZE - PUSH_SIZE)
2942
2943static void tcg_target_qemu_prologue(TCGContext *s)
2944{
2945    /* Calling convention requires us to save r4-r11 and lr.  */
2946    /* stmdb sp!, { r4 - r11, lr } */
2947    tcg_out_ldstm(s, COND_AL, INSN_STMDB, TCG_REG_CALL_STACK,
2948                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
2949                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
2950                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_R14));
2951
2952    /* Reserve callee argument and tcg temp space.  */
2953    tcg_out_dat_rI(s, COND_AL, ARITH_SUB, TCG_REG_CALL_STACK,
2954                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
2955    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
2956                  CPU_TEMP_BUF_NLONGS * sizeof(long));
2957
2958    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
2959
2960#ifndef CONFIG_SOFTMMU
2961    if (guest_base) {
2962        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_GUEST_BASE, guest_base);
2963        tcg_regset_set_reg(s->reserved_regs, TCG_REG_GUEST_BASE);
2964    }
2965#endif
2966
2967    tcg_out_b_reg(s, COND_AL, tcg_target_call_iarg_regs[1]);
2968
2969    /*
2970     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
2971     * and fall through to the rest of the epilogue.
2972     */
2973    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
2974    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, 0);
2975    tcg_out_epilogue(s);
2976}
2977
2978static void tcg_out_epilogue(TCGContext *s)
2979{
2980    /* Release local stack frame.  */
2981    tcg_out_dat_rI(s, COND_AL, ARITH_ADD, TCG_REG_CALL_STACK,
2982                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
2983
2984    /* ldmia sp!, { r4 - r11, pc } */
2985    tcg_out_ldstm(s, COND_AL, INSN_LDMIA, TCG_REG_CALL_STACK,
2986                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
2987                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
2988                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_PC));
2989}
2990
2991typedef struct {
2992    DebugFrameHeader h;
2993    uint8_t fde_def_cfa[4];
2994    uint8_t fde_reg_ofs[18];
2995} DebugFrame;
2996
2997#define ELF_HOST_MACHINE EM_ARM
2998
2999/* We're expecting a 2 byte uleb128 encoded value.  */
3000QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
3001
3002static const DebugFrame debug_frame = {
3003    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
3004    .h.cie.id = -1,
3005    .h.cie.version = 1,
3006    .h.cie.code_align = 1,
3007    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
3008    .h.cie.return_column = 14,
3009
3010    /* Total FDE size does not include the "len" member.  */
3011    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
3012
3013    .fde_def_cfa = {
3014        12, 13,                         /* DW_CFA_def_cfa sp, ... */
3015        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
3016        (FRAME_SIZE >> 7)
3017    },
3018    .fde_reg_ofs = {
3019        /* The following must match the stmdb in the prologue.  */
3020        0x8e, 1,                        /* DW_CFA_offset, lr, -4 */
3021        0x8b, 2,                        /* DW_CFA_offset, r11, -8 */
3022        0x8a, 3,                        /* DW_CFA_offset, r10, -12 */
3023        0x89, 4,                        /* DW_CFA_offset, r9, -16 */
3024        0x88, 5,                        /* DW_CFA_offset, r8, -20 */
3025        0x87, 6,                        /* DW_CFA_offset, r7, -24 */
3026        0x86, 7,                        /* DW_CFA_offset, r6, -28 */
3027        0x85, 8,                        /* DW_CFA_offset, r5, -32 */
3028        0x84, 9,                        /* DW_CFA_offset, r4, -36 */
3029    }
3030};
3031
3032void tcg_register_jit(const void *buf, size_t buf_size)
3033{
3034    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
3035}
3036