xref: /openbmc/qemu/tcg/arm/tcg-target.c.inc (revision f6e33708)
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#define TCG_REG_GUEST_BASE  TCG_REG_R11
93
94typedef enum {
95    COND_EQ = 0x0,
96    COND_NE = 0x1,
97    COND_CS = 0x2,	/* Unsigned greater or equal */
98    COND_CC = 0x3,	/* Unsigned less than */
99    COND_MI = 0x4,	/* Negative */
100    COND_PL = 0x5,	/* Zero or greater */
101    COND_VS = 0x6,	/* Overflow */
102    COND_VC = 0x7,	/* No overflow */
103    COND_HI = 0x8,	/* Unsigned greater than */
104    COND_LS = 0x9,	/* Unsigned less or equal */
105    COND_GE = 0xa,
106    COND_LT = 0xb,
107    COND_GT = 0xc,
108    COND_LE = 0xd,
109    COND_AL = 0xe,
110} ARMCond;
111
112#define TO_CPSR (1 << 20)
113
114#define SHIFT_IMM_LSL(im)	(((im) << 7) | 0x00)
115#define SHIFT_IMM_LSR(im)	(((im) << 7) | 0x20)
116#define SHIFT_IMM_ASR(im)	(((im) << 7) | 0x40)
117#define SHIFT_IMM_ROR(im)	(((im) << 7) | 0x60)
118#define SHIFT_REG_LSL(rs)	(((rs) << 8) | 0x10)
119#define SHIFT_REG_LSR(rs)	(((rs) << 8) | 0x30)
120#define SHIFT_REG_ASR(rs)	(((rs) << 8) | 0x50)
121#define SHIFT_REG_ROR(rs)	(((rs) << 8) | 0x70)
122
123typedef enum {
124    ARITH_AND = 0x0 << 21,
125    ARITH_EOR = 0x1 << 21,
126    ARITH_SUB = 0x2 << 21,
127    ARITH_RSB = 0x3 << 21,
128    ARITH_ADD = 0x4 << 21,
129    ARITH_ADC = 0x5 << 21,
130    ARITH_SBC = 0x6 << 21,
131    ARITH_RSC = 0x7 << 21,
132    ARITH_TST = 0x8 << 21 | TO_CPSR,
133    ARITH_CMP = 0xa << 21 | TO_CPSR,
134    ARITH_CMN = 0xb << 21 | TO_CPSR,
135    ARITH_ORR = 0xc << 21,
136    ARITH_MOV = 0xd << 21,
137    ARITH_BIC = 0xe << 21,
138    ARITH_MVN = 0xf << 21,
139
140    INSN_B         = 0x0a000000,
141
142    INSN_CLZ       = 0x016f0f10,
143    INSN_RBIT      = 0x06ff0f30,
144
145    INSN_LDMIA     = 0x08b00000,
146    INSN_STMDB     = 0x09200000,
147
148    INSN_LDR_IMM   = 0x04100000,
149    INSN_LDR_REG   = 0x06100000,
150    INSN_STR_IMM   = 0x04000000,
151    INSN_STR_REG   = 0x06000000,
152
153    INSN_LDRH_IMM  = 0x005000b0,
154    INSN_LDRH_REG  = 0x001000b0,
155    INSN_LDRSH_IMM = 0x005000f0,
156    INSN_LDRSH_REG = 0x001000f0,
157    INSN_STRH_IMM  = 0x004000b0,
158    INSN_STRH_REG  = 0x000000b0,
159
160    INSN_LDRB_IMM  = 0x04500000,
161    INSN_LDRB_REG  = 0x06500000,
162    INSN_LDRSB_IMM = 0x005000d0,
163    INSN_LDRSB_REG = 0x001000d0,
164    INSN_STRB_IMM  = 0x04400000,
165    INSN_STRB_REG  = 0x06400000,
166
167    INSN_LDRD_IMM  = 0x004000d0,
168    INSN_LDRD_REG  = 0x000000d0,
169    INSN_STRD_IMM  = 0x004000f0,
170    INSN_STRD_REG  = 0x000000f0,
171
172    INSN_DMB_ISH   = 0xf57ff05b,
173    INSN_DMB_MCR   = 0xee070fba,
174
175    /* Architected nop introduced in v6k.  */
176    /* ??? This is an MSR (imm) 0,0,0 insn.  Anyone know if this
177       also Just So Happened to do nothing on pre-v6k so that we
178       don't need to conditionalize it?  */
179    INSN_NOP_v6k   = 0xe320f000,
180    /* Otherwise the assembler uses mov r0,r0 */
181    INSN_NOP_v4    = (COND_AL << 28) | ARITH_MOV,
182
183    INSN_VADD      = 0xf2000800,
184    INSN_VAND      = 0xf2000110,
185    INSN_VBIC      = 0xf2100110,
186    INSN_VEOR      = 0xf3000110,
187    INSN_VORN      = 0xf2300110,
188    INSN_VORR      = 0xf2200110,
189    INSN_VSUB      = 0xf3000800,
190    INSN_VMUL      = 0xf2000910,
191    INSN_VQADD     = 0xf2000010,
192    INSN_VQADD_U   = 0xf3000010,
193    INSN_VQSUB     = 0xf2000210,
194    INSN_VQSUB_U   = 0xf3000210,
195    INSN_VMAX      = 0xf2000600,
196    INSN_VMAX_U    = 0xf3000600,
197    INSN_VMIN      = 0xf2000610,
198    INSN_VMIN_U    = 0xf3000610,
199
200    INSN_VABS      = 0xf3b10300,
201    INSN_VMVN      = 0xf3b00580,
202    INSN_VNEG      = 0xf3b10380,
203
204    INSN_VCEQ0     = 0xf3b10100,
205    INSN_VCGT0     = 0xf3b10000,
206    INSN_VCGE0     = 0xf3b10080,
207    INSN_VCLE0     = 0xf3b10180,
208    INSN_VCLT0     = 0xf3b10200,
209
210    INSN_VCEQ      = 0xf3000810,
211    INSN_VCGE      = 0xf2000310,
212    INSN_VCGT      = 0xf2000300,
213    INSN_VCGE_U    = 0xf3000310,
214    INSN_VCGT_U    = 0xf3000300,
215
216    INSN_VSHLI     = 0xf2800510,  /* VSHL (immediate) */
217    INSN_VSARI     = 0xf2800010,  /* VSHR.S */
218    INSN_VSHRI     = 0xf3800010,  /* VSHR.U */
219    INSN_VSLI      = 0xf3800510,
220    INSN_VSHL_S    = 0xf2000400,  /* VSHL.S (register) */
221    INSN_VSHL_U    = 0xf3000400,  /* VSHL.U (register) */
222
223    INSN_VBSL      = 0xf3100110,
224    INSN_VBIT      = 0xf3200110,
225    INSN_VBIF      = 0xf3300110,
226
227    INSN_VTST      = 0xf2000810,
228
229    INSN_VDUP_G    = 0xee800b10,  /* VDUP (ARM core register) */
230    INSN_VDUP_S    = 0xf3b00c00,  /* VDUP (scalar) */
231    INSN_VLDR_D    = 0xed100b00,  /* VLDR.64 */
232    INSN_VLD1      = 0xf4200000,  /* VLD1 (multiple single elements) */
233    INSN_VLD1R     = 0xf4a00c00,  /* VLD1 (single element to all lanes) */
234    INSN_VST1      = 0xf4000000,  /* VST1 (multiple single elements) */
235    INSN_VMOVI     = 0xf2800010,  /* VMOV (immediate) */
236} ARMInsn;
237
238#define INSN_NOP   (use_armv7_instructions ? INSN_NOP_v6k : INSN_NOP_v4)
239
240static const uint8_t tcg_cond_to_arm_cond[] = {
241    [TCG_COND_EQ] = COND_EQ,
242    [TCG_COND_NE] = COND_NE,
243    [TCG_COND_LT] = COND_LT,
244    [TCG_COND_GE] = COND_GE,
245    [TCG_COND_LE] = COND_LE,
246    [TCG_COND_GT] = COND_GT,
247    /* unsigned */
248    [TCG_COND_LTU] = COND_CC,
249    [TCG_COND_GEU] = COND_CS,
250    [TCG_COND_LEU] = COND_LS,
251    [TCG_COND_GTU] = COND_HI,
252};
253
254static int encode_imm(uint32_t imm);
255
256/* TCG private relocation type: add with pc+imm8 */
257#define R_ARM_PC8  11
258
259/* TCG private relocation type: vldr with imm8 << 2 */
260#define R_ARM_PC11 12
261
262static bool reloc_pc24(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
263{
264    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
265    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) >> 2;
266
267    if (offset == sextract32(offset, 0, 24)) {
268        *src_rw = deposit32(*src_rw, 0, 24, offset);
269        return true;
270    }
271    return false;
272}
273
274static bool reloc_pc13(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
275{
276    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
277    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
278
279    if (offset >= -0xfff && offset <= 0xfff) {
280        tcg_insn_unit insn = *src_rw;
281        bool u = (offset >= 0);
282        if (!u) {
283            offset = -offset;
284        }
285        insn = deposit32(insn, 23, 1, u);
286        insn = deposit32(insn, 0, 12, offset);
287        *src_rw = insn;
288        return true;
289    }
290    return false;
291}
292
293static bool reloc_pc11(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
294{
295    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
296    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) / 4;
297
298    if (offset >= -0xff && offset <= 0xff) {
299        tcg_insn_unit insn = *src_rw;
300        bool u = (offset >= 0);
301        if (!u) {
302            offset = -offset;
303        }
304        insn = deposit32(insn, 23, 1, u);
305        insn = deposit32(insn, 0, 8, offset);
306        *src_rw = insn;
307        return true;
308    }
309    return false;
310}
311
312static bool reloc_pc8(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
313{
314    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
315    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
316    int imm12 = encode_imm(offset);
317
318    if (imm12 >= 0) {
319        *src_rw = deposit32(*src_rw, 0, 12, imm12);
320        return true;
321    }
322    return false;
323}
324
325static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
326                        intptr_t value, intptr_t addend)
327{
328    tcg_debug_assert(addend == 0);
329    switch (type) {
330    case R_ARM_PC24:
331        return reloc_pc24(code_ptr, (const tcg_insn_unit *)value);
332    case R_ARM_PC13:
333        return reloc_pc13(code_ptr, (const tcg_insn_unit *)value);
334    case R_ARM_PC11:
335        return reloc_pc11(code_ptr, (const tcg_insn_unit *)value);
336    case R_ARM_PC8:
337        return reloc_pc8(code_ptr, (const tcg_insn_unit *)value);
338    default:
339        g_assert_not_reached();
340    }
341}
342
343#define TCG_CT_CONST_ARM  0x100
344#define TCG_CT_CONST_INV  0x200
345#define TCG_CT_CONST_NEG  0x400
346#define TCG_CT_CONST_ZERO 0x800
347#define TCG_CT_CONST_ORRI 0x1000
348#define TCG_CT_CONST_ANDI 0x2000
349
350#define ALL_GENERAL_REGS  0xffffu
351#define ALL_VECTOR_REGS   0xffff0000u
352
353/*
354 * r0-r3 will be overwritten when reading the tlb entry (system-mode only);
355 * r14 will be overwritten by the BLNE branching to the slow path.
356 */
357#define ALL_QLDST_REGS \
358    (ALL_GENERAL_REGS & ~((tcg_use_softmmu ? 0xf : 0) | (1 << TCG_REG_R14)))
359
360/*
361 * ARM immediates for ALU instructions are made of an unsigned 8-bit
362 * right-rotated by an even amount between 0 and 30.
363 *
364 * Return < 0 if @imm cannot be encoded, else the entire imm12 field.
365 */
366static int encode_imm(uint32_t imm)
367{
368    uint32_t rot, imm8;
369
370    /* Simple case, no rotation required. */
371    if ((imm & ~0xff) == 0) {
372        return imm;
373    }
374
375    /* Next, try a simple even shift.  */
376    rot = ctz32(imm) & ~1;
377    imm8 = imm >> rot;
378    rot = 32 - rot;
379    if ((imm8 & ~0xff) == 0) {
380        goto found;
381    }
382
383    /*
384     * Finally, try harder with rotations.
385     * The ctz test above will have taken care of rotates >= 8.
386     */
387    for (rot = 2; rot < 8; rot += 2) {
388        imm8 = rol32(imm, rot);
389        if ((imm8 & ~0xff) == 0) {
390            goto found;
391        }
392    }
393    /* Fail: imm cannot be encoded. */
394    return -1;
395
396 found:
397    /* Note that rot is even, and we discard bit 0 by shifting by 7. */
398    return rot << 7 | imm8;
399}
400
401static int encode_imm_nofail(uint32_t imm)
402{
403    int ret = encode_imm(imm);
404    tcg_debug_assert(ret >= 0);
405    return ret;
406}
407
408static bool check_fit_imm(uint32_t imm)
409{
410    return encode_imm(imm) >= 0;
411}
412
413/* Return true if v16 is a valid 16-bit shifted immediate.  */
414static bool is_shimm16(uint16_t v16, int *cmode, int *imm8)
415{
416    if (v16 == (v16 & 0xff)) {
417        *cmode = 0x8;
418        *imm8 = v16 & 0xff;
419        return true;
420    } else if (v16 == (v16 & 0xff00)) {
421        *cmode = 0xa;
422        *imm8 = v16 >> 8;
423        return true;
424    }
425    return false;
426}
427
428/* Return true if v32 is a valid 32-bit shifted immediate.  */
429static bool is_shimm32(uint32_t v32, int *cmode, int *imm8)
430{
431    if (v32 == (v32 & 0xff)) {
432        *cmode = 0x0;
433        *imm8 = v32 & 0xff;
434        return true;
435    } else if (v32 == (v32 & 0xff00)) {
436        *cmode = 0x2;
437        *imm8 = (v32 >> 8) & 0xff;
438        return true;
439    } else if (v32 == (v32 & 0xff0000)) {
440        *cmode = 0x4;
441        *imm8 = (v32 >> 16) & 0xff;
442        return true;
443    } else if (v32 == (v32 & 0xff000000)) {
444        *cmode = 0x6;
445        *imm8 = v32 >> 24;
446        return true;
447    }
448    return false;
449}
450
451/* Return true if v32 is a valid 32-bit shifting ones immediate.  */
452static bool is_soimm32(uint32_t v32, int *cmode, int *imm8)
453{
454    if ((v32 & 0xffff00ff) == 0xff) {
455        *cmode = 0xc;
456        *imm8 = (v32 >> 8) & 0xff;
457        return true;
458    } else if ((v32 & 0xff00ffff) == 0xffff) {
459        *cmode = 0xd;
460        *imm8 = (v32 >> 16) & 0xff;
461        return true;
462    }
463    return false;
464}
465
466/*
467 * Return non-zero if v32 can be formed by MOVI+ORR.
468 * Place the parameters for MOVI in (cmode, imm8).
469 * Return the cmode for ORR; the imm8 can be had via extraction from v32.
470 */
471static int is_shimm32_pair(uint32_t v32, int *cmode, int *imm8)
472{
473    int i;
474
475    for (i = 6; i > 0; i -= 2) {
476        /* Mask out one byte we can add with ORR.  */
477        uint32_t tmp = v32 & ~(0xffu << (i * 4));
478        if (is_shimm32(tmp, cmode, imm8) ||
479            is_soimm32(tmp, cmode, imm8)) {
480            break;
481        }
482    }
483    return i;
484}
485
486/* Return true if V is a valid 16-bit or 32-bit shifted immediate.  */
487static bool is_shimm1632(uint32_t v32, int *cmode, int *imm8)
488{
489    if (v32 == deposit32(v32, 16, 16, v32)) {
490        return is_shimm16(v32, cmode, imm8);
491    } else {
492        return is_shimm32(v32, cmode, imm8);
493    }
494}
495
496/* Test if a constant matches the constraint.
497 * TODO: define constraints for:
498 *
499 * ldr/str offset:   between -0xfff and 0xfff
500 * ldrh/strh offset: between -0xff and 0xff
501 * mov operand2:     values represented with x << (2 * y), x < 0x100
502 * add, sub, eor...: ditto
503 */
504static bool tcg_target_const_match(int64_t val, int ct,
505                                   TCGType type, TCGCond cond, int vece)
506{
507    if (ct & TCG_CT_CONST) {
508        return 1;
509    } else if ((ct & TCG_CT_CONST_ARM) && check_fit_imm(val)) {
510        return 1;
511    } else if ((ct & TCG_CT_CONST_INV) && check_fit_imm(~val)) {
512        return 1;
513    } else if ((ct & TCG_CT_CONST_NEG) && check_fit_imm(-val)) {
514        return 1;
515    } else if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
516        return 1;
517    }
518
519    switch (ct & (TCG_CT_CONST_ORRI | TCG_CT_CONST_ANDI)) {
520    case 0:
521        break;
522    case TCG_CT_CONST_ANDI:
523        val = ~val;
524        /* fallthru */
525    case TCG_CT_CONST_ORRI:
526        if (val == deposit64(val, 32, 32, val)) {
527            int cmode, imm8;
528            return is_shimm1632(val, &cmode, &imm8);
529        }
530        break;
531    default:
532        /* Both bits should not be set for the same insn.  */
533        g_assert_not_reached();
534    }
535
536    return 0;
537}
538
539static void tcg_out_b_imm(TCGContext *s, ARMCond cond, int32_t offset)
540{
541    tcg_out32(s, (cond << 28) | INSN_B |
542                    (((offset - 8) >> 2) & 0x00ffffff));
543}
544
545static void tcg_out_bl_imm(TCGContext *s, ARMCond cond, int32_t offset)
546{
547    tcg_out32(s, (cond << 28) | 0x0b000000 |
548                    (((offset - 8) >> 2) & 0x00ffffff));
549}
550
551static void tcg_out_blx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
552{
553    tcg_out32(s, (cond << 28) | 0x012fff30 | rn);
554}
555
556static void tcg_out_blx_imm(TCGContext *s, int32_t offset)
557{
558    tcg_out32(s, 0xfa000000 | ((offset & 2) << 23) |
559                (((offset - 8) >> 2) & 0x00ffffff));
560}
561
562static void tcg_out_dat_reg(TCGContext *s, ARMCond cond, ARMInsn opc,
563                            TCGReg rd, TCGReg rn, TCGReg rm, int shift)
564{
565    tcg_out32(s, (cond << 28) | (0 << 25) | opc |
566                    (rn << 16) | (rd << 12) | shift | rm);
567}
568
569static void tcg_out_mov_reg(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rm)
570{
571    /* Simple reg-reg move, optimising out the 'do nothing' case */
572    if (rd != rm) {
573        tcg_out_dat_reg(s, cond, ARITH_MOV, rd, 0, rm, SHIFT_IMM_LSL(0));
574    }
575}
576
577static void tcg_out_bx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
578{
579    tcg_out32(s, (cond << 28) | 0x012fff10 | rn);
580}
581
582static void tcg_out_b_reg(TCGContext *s, ARMCond cond, TCGReg rn)
583{
584    /*
585     * Unless the C portion of QEMU is compiled as thumb, we don't need
586     * true BX semantics; merely a branch to an address held in a register.
587     */
588    tcg_out_bx_reg(s, cond, rn);
589}
590
591static void tcg_out_dat_imm(TCGContext *s, ARMCond cond, ARMInsn opc,
592                            TCGReg rd, TCGReg rn, int im)
593{
594    tcg_out32(s, (cond << 28) | (1 << 25) | opc |
595                    (rn << 16) | (rd << 12) | im);
596}
597
598static void tcg_out_ldstm(TCGContext *s, ARMCond cond, ARMInsn opc,
599                          TCGReg rn, uint16_t mask)
600{
601    tcg_out32(s, (cond << 28) | opc | (rn << 16) | mask);
602}
603
604/* Note that this routine is used for both LDR and LDRH formats, so we do
605   not wish to include an immediate shift at this point.  */
606static void tcg_out_memop_r(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
607                            TCGReg rn, TCGReg rm, bool u, bool p, bool w)
608{
609    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24)
610              | (w << 21) | (rn << 16) | (rt << 12) | rm);
611}
612
613static void tcg_out_memop_8(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
614                            TCGReg rn, int imm8, bool p, bool w)
615{
616    bool u = 1;
617    if (imm8 < 0) {
618        imm8 = -imm8;
619        u = 0;
620    }
621    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
622              (rn << 16) | (rt << 12) | ((imm8 & 0xf0) << 4) | (imm8 & 0xf));
623}
624
625static void tcg_out_memop_12(TCGContext *s, ARMCond cond, ARMInsn opc,
626                             TCGReg rt, TCGReg rn, int imm12, bool p, bool w)
627{
628    bool u = 1;
629    if (imm12 < 0) {
630        imm12 = -imm12;
631        u = 0;
632    }
633    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
634              (rn << 16) | (rt << 12) | imm12);
635}
636
637static void tcg_out_ld32_12(TCGContext *s, ARMCond cond, TCGReg rt,
638                            TCGReg rn, int imm12)
639{
640    tcg_out_memop_12(s, cond, INSN_LDR_IMM, rt, rn, imm12, 1, 0);
641}
642
643static void tcg_out_st32_12(TCGContext *s, ARMCond cond, TCGReg rt,
644                            TCGReg rn, int imm12)
645{
646    tcg_out_memop_12(s, cond, INSN_STR_IMM, rt, rn, imm12, 1, 0);
647}
648
649static void tcg_out_ld32_r(TCGContext *s, ARMCond cond, TCGReg rt,
650                           TCGReg rn, TCGReg rm)
651{
652    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 0);
653}
654
655static void tcg_out_st32_r(TCGContext *s, ARMCond cond, TCGReg rt,
656                           TCGReg rn, TCGReg rm)
657{
658    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 0);
659}
660
661static void tcg_out_ldrd_8(TCGContext *s, ARMCond cond, TCGReg rt,
662                           TCGReg rn, int imm8)
663{
664    tcg_out_memop_8(s, cond, INSN_LDRD_IMM, rt, rn, imm8, 1, 0);
665}
666
667static void tcg_out_ldrd_r(TCGContext *s, ARMCond cond, TCGReg rt,
668                           TCGReg rn, TCGReg rm)
669{
670    tcg_out_memop_r(s, cond, INSN_LDRD_REG, rt, rn, rm, 1, 1, 0);
671}
672
673static void __attribute__((unused))
674tcg_out_ldrd_rwb(TCGContext *s, ARMCond cond, TCGReg rt, TCGReg rn, TCGReg rm)
675{
676    tcg_out_memop_r(s, cond, INSN_LDRD_REG, rt, rn, rm, 1, 1, 1);
677}
678
679static void __attribute__((unused))
680tcg_out_strd_8(TCGContext *s, ARMCond cond, TCGReg rt, TCGReg rn, int imm8)
681{
682    tcg_out_memop_8(s, cond, INSN_STRD_IMM, rt, rn, imm8, 1, 0);
683}
684
685static void tcg_out_strd_r(TCGContext *s, ARMCond cond, TCGReg rt,
686                           TCGReg rn, TCGReg rm)
687{
688    tcg_out_memop_r(s, cond, INSN_STRD_REG, rt, rn, rm, 1, 1, 0);
689}
690
691/* Register pre-increment with base writeback.  */
692static void tcg_out_ld32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
693                             TCGReg rn, TCGReg rm)
694{
695    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 1);
696}
697
698static void tcg_out_st32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
699                             TCGReg rn, TCGReg rm)
700{
701    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 1);
702}
703
704static void tcg_out_ld16u_8(TCGContext *s, ARMCond cond, TCGReg rt,
705                            TCGReg rn, int imm8)
706{
707    tcg_out_memop_8(s, cond, INSN_LDRH_IMM, rt, rn, imm8, 1, 0);
708}
709
710static void tcg_out_st16_8(TCGContext *s, ARMCond cond, TCGReg rt,
711                           TCGReg rn, int imm8)
712{
713    tcg_out_memop_8(s, cond, INSN_STRH_IMM, rt, rn, imm8, 1, 0);
714}
715
716static void tcg_out_ld16u_r(TCGContext *s, ARMCond cond, TCGReg rt,
717                            TCGReg rn, TCGReg rm)
718{
719    tcg_out_memop_r(s, cond, INSN_LDRH_REG, rt, rn, rm, 1, 1, 0);
720}
721
722static void tcg_out_st16_r(TCGContext *s, ARMCond cond, TCGReg rt,
723                           TCGReg rn, TCGReg rm)
724{
725    tcg_out_memop_r(s, cond, INSN_STRH_REG, rt, rn, rm, 1, 1, 0);
726}
727
728static void tcg_out_ld16s_8(TCGContext *s, ARMCond cond, TCGReg rt,
729                            TCGReg rn, int imm8)
730{
731    tcg_out_memop_8(s, cond, INSN_LDRSH_IMM, rt, rn, imm8, 1, 0);
732}
733
734static void tcg_out_ld16s_r(TCGContext *s, ARMCond cond, TCGReg rt,
735                            TCGReg rn, TCGReg rm)
736{
737    tcg_out_memop_r(s, cond, INSN_LDRSH_REG, rt, rn, rm, 1, 1, 0);
738}
739
740static void tcg_out_ld8_12(TCGContext *s, ARMCond cond, TCGReg rt,
741                           TCGReg rn, int imm12)
742{
743    tcg_out_memop_12(s, cond, INSN_LDRB_IMM, rt, rn, imm12, 1, 0);
744}
745
746static void tcg_out_st8_12(TCGContext *s, ARMCond cond, TCGReg rt,
747                           TCGReg rn, int imm12)
748{
749    tcg_out_memop_12(s, cond, INSN_STRB_IMM, rt, rn, imm12, 1, 0);
750}
751
752static void tcg_out_ld8_r(TCGContext *s, ARMCond cond, TCGReg rt,
753                          TCGReg rn, TCGReg rm)
754{
755    tcg_out_memop_r(s, cond, INSN_LDRB_REG, rt, rn, rm, 1, 1, 0);
756}
757
758static void tcg_out_st8_r(TCGContext *s, ARMCond cond, TCGReg rt,
759                          TCGReg rn, TCGReg rm)
760{
761    tcg_out_memop_r(s, cond, INSN_STRB_REG, rt, rn, rm, 1, 1, 0);
762}
763
764static void tcg_out_ld8s_8(TCGContext *s, ARMCond cond, TCGReg rt,
765                           TCGReg rn, int imm8)
766{
767    tcg_out_memop_8(s, cond, INSN_LDRSB_IMM, rt, rn, imm8, 1, 0);
768}
769
770static void tcg_out_ld8s_r(TCGContext *s, ARMCond cond, TCGReg rt,
771                           TCGReg rn, TCGReg rm)
772{
773    tcg_out_memop_r(s, cond, INSN_LDRSB_REG, rt, rn, rm, 1, 1, 0);
774}
775
776static void tcg_out_movi_pool(TCGContext *s, ARMCond cond,
777                              TCGReg rd, uint32_t arg)
778{
779    new_pool_label(s, arg, R_ARM_PC13, s->code_ptr, 0);
780    tcg_out_ld32_12(s, cond, rd, TCG_REG_PC, 0);
781}
782
783static void tcg_out_movi32(TCGContext *s, ARMCond cond,
784                           TCGReg rd, uint32_t arg)
785{
786    int imm12, diff, opc, sh1, sh2;
787    uint32_t tt0, tt1, tt2;
788
789    /* Check a single MOV/MVN before anything else.  */
790    imm12 = encode_imm(arg);
791    if (imm12 >= 0) {
792        tcg_out_dat_imm(s, cond, ARITH_MOV, rd, 0, imm12);
793        return;
794    }
795    imm12 = encode_imm(~arg);
796    if (imm12 >= 0) {
797        tcg_out_dat_imm(s, cond, ARITH_MVN, rd, 0, imm12);
798        return;
799    }
800
801    /* Check for a pc-relative address.  This will usually be the TB,
802       or within the TB, which is immediately before the code block.  */
803    diff = tcg_pcrel_diff(s, (void *)arg) - 8;
804    if (diff >= 0) {
805        imm12 = encode_imm(diff);
806        if (imm12 >= 0) {
807            tcg_out_dat_imm(s, cond, ARITH_ADD, rd, TCG_REG_PC, imm12);
808            return;
809        }
810    } else {
811        imm12 = encode_imm(-diff);
812        if (imm12 >= 0) {
813            tcg_out_dat_imm(s, cond, ARITH_SUB, rd, TCG_REG_PC, imm12);
814            return;
815        }
816    }
817
818    /* Use movw + movt.  */
819    if (use_armv7_instructions) {
820        /* movw */
821        tcg_out32(s, (cond << 28) | 0x03000000 | (rd << 12)
822                  | ((arg << 4) & 0x000f0000) | (arg & 0xfff));
823        if (arg & 0xffff0000) {
824            /* movt */
825            tcg_out32(s, (cond << 28) | 0x03400000 | (rd << 12)
826                      | ((arg >> 12) & 0x000f0000) | ((arg >> 16) & 0xfff));
827        }
828        return;
829    }
830
831    /* Look for sequences of two insns.  If we have lots of 1's, we can
832       shorten the sequence by beginning with mvn and then clearing
833       higher bits with eor.  */
834    tt0 = arg;
835    opc = ARITH_MOV;
836    if (ctpop32(arg) > 16) {
837        tt0 = ~arg;
838        opc = ARITH_MVN;
839    }
840    sh1 = ctz32(tt0) & ~1;
841    tt1 = tt0 & ~(0xff << sh1);
842    sh2 = ctz32(tt1) & ~1;
843    tt2 = tt1 & ~(0xff << sh2);
844    if (tt2 == 0) {
845        int rot;
846
847        rot = ((32 - sh1) << 7) & 0xf00;
848        tcg_out_dat_imm(s, cond, opc, rd,  0, ((tt0 >> sh1) & 0xff) | rot);
849        rot = ((32 - sh2) << 7) & 0xf00;
850        tcg_out_dat_imm(s, cond, ARITH_EOR, rd, rd,
851                        ((tt0 >> sh2) & 0xff) | rot);
852        return;
853    }
854
855    /* Otherwise, drop it into the constant pool.  */
856    tcg_out_movi_pool(s, cond, rd, arg);
857}
858
859/*
860 * Emit either the reg,imm or reg,reg form of a data-processing insn.
861 * rhs must satisfy the "rI" constraint.
862 */
863static void tcg_out_dat_rI(TCGContext *s, ARMCond cond, ARMInsn opc,
864                           TCGReg dst, TCGReg lhs, TCGArg rhs, int rhs_is_const)
865{
866    if (rhs_is_const) {
867        tcg_out_dat_imm(s, cond, opc, dst, lhs, encode_imm_nofail(rhs));
868    } else {
869        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
870    }
871}
872
873/*
874 * Emit either the reg,imm or reg,reg form of a data-processing insn.
875 * rhs must satisfy the "rIK" constraint.
876 */
877static void tcg_out_dat_rIK(TCGContext *s, ARMCond cond, ARMInsn opc,
878                            ARMInsn opinv, TCGReg dst, TCGReg lhs, TCGArg rhs,
879                            bool rhs_is_const)
880{
881    if (rhs_is_const) {
882        int imm12 = encode_imm(rhs);
883        if (imm12 < 0) {
884            imm12 = encode_imm_nofail(~rhs);
885            opc = opinv;
886        }
887        tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
888    } else {
889        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
890    }
891}
892
893static void tcg_out_dat_rIN(TCGContext *s, ARMCond cond, ARMInsn opc,
894                            ARMInsn opneg, TCGReg dst, TCGReg lhs, TCGArg rhs,
895                            bool rhs_is_const)
896{
897    /* Emit either the reg,imm or reg,reg form of a data-processing insn.
898     * rhs must satisfy the "rIN" constraint.
899     */
900    if (rhs_is_const) {
901        int imm12 = encode_imm(rhs);
902        if (imm12 < 0) {
903            imm12 = encode_imm_nofail(-rhs);
904            opc = opneg;
905        }
906        tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
907    } else {
908        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
909    }
910}
911
912static void tcg_out_mul32(TCGContext *s, ARMCond cond, TCGReg rd,
913                          TCGReg rn, TCGReg rm)
914{
915    /* mul */
916    tcg_out32(s, (cond << 28) | 0x90 | (rd << 16) | (rm << 8) | rn);
917}
918
919static void tcg_out_umull32(TCGContext *s, ARMCond cond, TCGReg rd0,
920                            TCGReg rd1, TCGReg rn, TCGReg rm)
921{
922    /* umull */
923    tcg_out32(s, (cond << 28) | 0x00800090 |
924              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
925}
926
927static void tcg_out_smull32(TCGContext *s, ARMCond cond, TCGReg rd0,
928                            TCGReg rd1, TCGReg rn, TCGReg rm)
929{
930    /* smull */
931    tcg_out32(s, (cond << 28) | 0x00c00090 |
932              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
933}
934
935static void tcg_out_sdiv(TCGContext *s, ARMCond cond,
936                         TCGReg rd, TCGReg rn, TCGReg rm)
937{
938    tcg_out32(s, 0x0710f010 | (cond << 28) | (rd << 16) | rn | (rm << 8));
939}
940
941static void tcg_out_udiv(TCGContext *s, ARMCond cond,
942                         TCGReg rd, TCGReg rn, TCGReg rm)
943{
944    tcg_out32(s, 0x0730f010 | (cond << 28) | (rd << 16) | rn | (rm << 8));
945}
946
947static void tcg_out_ext8s(TCGContext *s, TCGType t, TCGReg rd, TCGReg rn)
948{
949    /* sxtb */
950    tcg_out32(s, 0x06af0070 | (COND_AL << 28) | (rd << 12) | rn);
951}
952
953static void tcg_out_ext8u(TCGContext *s, TCGReg rd, TCGReg rn)
954{
955    tcg_out_dat_imm(s, COND_AL, ARITH_AND, rd, rn, 0xff);
956}
957
958static void tcg_out_ext16s(TCGContext *s, TCGType t, TCGReg rd, TCGReg rn)
959{
960    /* sxth */
961    tcg_out32(s, 0x06bf0070 | (COND_AL << 28) | (rd << 12) | rn);
962}
963
964static void tcg_out_ext16u(TCGContext *s, TCGReg rd, TCGReg rn)
965{
966    /* uxth */
967    tcg_out32(s, 0x06ff0070 | (COND_AL << 28) | (rd << 12) | rn);
968}
969
970static void tcg_out_ext32s(TCGContext *s, TCGReg rd, TCGReg rn)
971{
972    g_assert_not_reached();
973}
974
975static void tcg_out_ext32u(TCGContext *s, TCGReg rd, TCGReg rn)
976{
977    g_assert_not_reached();
978}
979
980static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
981{
982    g_assert_not_reached();
983}
984
985static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
986{
987    g_assert_not_reached();
988}
989
990static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg rd, TCGReg rn)
991{
992    g_assert_not_reached();
993}
994
995static void tcg_out_bswap16(TCGContext *s, ARMCond cond,
996                            TCGReg rd, TCGReg rn, int flags)
997{
998    if (flags & TCG_BSWAP_OS) {
999        /* revsh */
1000        tcg_out32(s, 0x06ff0fb0 | (cond << 28) | (rd << 12) | rn);
1001        return;
1002    }
1003
1004    /* rev16 */
1005    tcg_out32(s, 0x06bf0fb0 | (cond << 28) | (rd << 12) | rn);
1006    if ((flags & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
1007        /* uxth */
1008        tcg_out32(s, 0x06ff0070 | (cond << 28) | (rd << 12) | rd);
1009    }
1010}
1011
1012static void tcg_out_bswap32(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rn)
1013{
1014    /* rev */
1015    tcg_out32(s, 0x06bf0f30 | (cond << 28) | (rd << 12) | rn);
1016}
1017
1018static void tcg_out_deposit(TCGContext *s, ARMCond cond, TCGReg rd,
1019                            TCGArg a1, int ofs, int len, bool const_a1)
1020{
1021    if (const_a1) {
1022        /* bfi becomes bfc with rn == 15.  */
1023        a1 = 15;
1024    }
1025    /* bfi/bfc */
1026    tcg_out32(s, 0x07c00010 | (cond << 28) | (rd << 12) | a1
1027              | (ofs << 7) | ((ofs + len - 1) << 16));
1028}
1029
1030static void tcg_out_extract(TCGContext *s, ARMCond cond, TCGReg rd,
1031                            TCGReg rn, int ofs, int len)
1032{
1033    /* ubfx */
1034    tcg_out32(s, 0x07e00050 | (cond << 28) | (rd << 12) | rn
1035              | (ofs << 7) | ((len - 1) << 16));
1036}
1037
1038static void tcg_out_sextract(TCGContext *s, ARMCond cond, TCGReg rd,
1039                             TCGReg rn, int ofs, int len)
1040{
1041    /* sbfx */
1042    tcg_out32(s, 0x07a00050 | (cond << 28) | (rd << 12) | rn
1043              | (ofs << 7) | ((len - 1) << 16));
1044}
1045
1046static void tcg_out_ld32u(TCGContext *s, ARMCond cond,
1047                          TCGReg rd, TCGReg rn, int32_t offset)
1048{
1049    if (offset > 0xfff || offset < -0xfff) {
1050        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1051        tcg_out_ld32_r(s, cond, rd, rn, TCG_REG_TMP);
1052    } else
1053        tcg_out_ld32_12(s, cond, rd, rn, offset);
1054}
1055
1056static void tcg_out_st32(TCGContext *s, ARMCond cond,
1057                         TCGReg rd, TCGReg rn, int32_t offset)
1058{
1059    if (offset > 0xfff || offset < -0xfff) {
1060        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1061        tcg_out_st32_r(s, cond, rd, rn, TCG_REG_TMP);
1062    } else
1063        tcg_out_st32_12(s, cond, rd, rn, offset);
1064}
1065
1066static void tcg_out_ld16u(TCGContext *s, ARMCond cond,
1067                          TCGReg rd, TCGReg rn, int32_t offset)
1068{
1069    if (offset > 0xff || offset < -0xff) {
1070        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1071        tcg_out_ld16u_r(s, cond, rd, rn, TCG_REG_TMP);
1072    } else
1073        tcg_out_ld16u_8(s, cond, rd, rn, offset);
1074}
1075
1076static void tcg_out_ld16s(TCGContext *s, ARMCond cond,
1077                          TCGReg rd, TCGReg rn, int32_t offset)
1078{
1079    if (offset > 0xff || offset < -0xff) {
1080        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1081        tcg_out_ld16s_r(s, cond, rd, rn, TCG_REG_TMP);
1082    } else
1083        tcg_out_ld16s_8(s, cond, rd, rn, offset);
1084}
1085
1086static void tcg_out_st16(TCGContext *s, ARMCond cond,
1087                         TCGReg rd, TCGReg rn, int32_t offset)
1088{
1089    if (offset > 0xff || offset < -0xff) {
1090        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1091        tcg_out_st16_r(s, cond, rd, rn, TCG_REG_TMP);
1092    } else
1093        tcg_out_st16_8(s, cond, rd, rn, offset);
1094}
1095
1096static void tcg_out_ld8u(TCGContext *s, ARMCond cond,
1097                         TCGReg rd, TCGReg rn, int32_t offset)
1098{
1099    if (offset > 0xfff || offset < -0xfff) {
1100        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1101        tcg_out_ld8_r(s, cond, rd, rn, TCG_REG_TMP);
1102    } else
1103        tcg_out_ld8_12(s, cond, rd, rn, offset);
1104}
1105
1106static void tcg_out_ld8s(TCGContext *s, ARMCond cond,
1107                         TCGReg rd, TCGReg rn, int32_t offset)
1108{
1109    if (offset > 0xff || offset < -0xff) {
1110        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1111        tcg_out_ld8s_r(s, cond, rd, rn, TCG_REG_TMP);
1112    } else
1113        tcg_out_ld8s_8(s, cond, rd, rn, offset);
1114}
1115
1116static void tcg_out_st8(TCGContext *s, ARMCond cond,
1117                        TCGReg rd, TCGReg rn, int32_t offset)
1118{
1119    if (offset > 0xfff || offset < -0xfff) {
1120        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1121        tcg_out_st8_r(s, cond, rd, rn, TCG_REG_TMP);
1122    } else
1123        tcg_out_st8_12(s, cond, rd, rn, offset);
1124}
1125
1126/*
1127 * The _goto case is normally between TBs within the same code buffer, and
1128 * with the code buffer limited to 16MB we wouldn't need the long case.
1129 * But we also use it for the tail-call to the qemu_ld/st helpers, which does.
1130 */
1131static void tcg_out_goto(TCGContext *s, ARMCond cond, const tcg_insn_unit *addr)
1132{
1133    intptr_t addri = (intptr_t)addr;
1134    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1135    bool arm_mode = !(addri & 1);
1136
1137    if (arm_mode && disp - 8 < 0x01fffffd && disp - 8 > -0x01fffffd) {
1138        tcg_out_b_imm(s, cond, disp);
1139        return;
1140    }
1141
1142    /* LDR is interworking from v5t. */
1143    tcg_out_movi_pool(s, cond, TCG_REG_PC, addri);
1144}
1145
1146/*
1147 * The call case is mostly used for helpers - so it's not unreasonable
1148 * for them to be beyond branch range.
1149 */
1150static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *addr)
1151{
1152    intptr_t addri = (intptr_t)addr;
1153    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1154    bool arm_mode = !(addri & 1);
1155
1156    if (disp - 8 < 0x02000000 && disp - 8 >= -0x02000000) {
1157        if (arm_mode) {
1158            tcg_out_bl_imm(s, COND_AL, disp);
1159        } else {
1160            tcg_out_blx_imm(s, disp);
1161        }
1162        return;
1163    }
1164
1165    tcg_out_movi32(s, COND_AL, TCG_REG_TMP, addri);
1166    tcg_out_blx_reg(s, COND_AL, TCG_REG_TMP);
1167}
1168
1169static void tcg_out_call(TCGContext *s, const tcg_insn_unit *addr,
1170                         const TCGHelperInfo *info)
1171{
1172    tcg_out_call_int(s, addr);
1173}
1174
1175static void tcg_out_goto_label(TCGContext *s, ARMCond cond, TCGLabel *l)
1176{
1177    if (l->has_value) {
1178        tcg_out_goto(s, cond, l->u.value_ptr);
1179    } else {
1180        tcg_out_reloc(s, s->code_ptr, R_ARM_PC24, l, 0);
1181        tcg_out_b_imm(s, cond, 0);
1182    }
1183}
1184
1185static void tcg_out_mb(TCGContext *s, TCGArg a0)
1186{
1187    if (use_armv7_instructions) {
1188        tcg_out32(s, INSN_DMB_ISH);
1189    } else {
1190        tcg_out32(s, INSN_DMB_MCR);
1191    }
1192}
1193
1194static TCGCond tcg_out_cmp(TCGContext *s, TCGCond cond, TCGReg a,
1195                           TCGArg b, int b_const)
1196{
1197    if (!is_tst_cond(cond)) {
1198        tcg_out_dat_rIN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0, a, b, b_const);
1199        return cond;
1200    }
1201
1202    cond = tcg_tst_eqne_cond(cond);
1203    if (b_const) {
1204        int imm12 = encode_imm(b);
1205
1206        /*
1207         * The compare constraints allow rIN, but TST does not support N.
1208         * Be prepared to load the constant into a scratch register.
1209         */
1210        if (imm12 >= 0) {
1211            tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, a, imm12);
1212            return cond;
1213        }
1214        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, b);
1215        b = TCG_REG_TMP;
1216    }
1217    tcg_out_dat_reg(s, COND_AL, ARITH_TST, 0, a, b, SHIFT_IMM_LSL(0));
1218    return cond;
1219}
1220
1221static TCGCond tcg_out_cmp2(TCGContext *s, const TCGArg *args,
1222                            const int *const_args)
1223{
1224    TCGReg al = args[0];
1225    TCGReg ah = args[1];
1226    TCGArg bl = args[2];
1227    TCGArg bh = args[3];
1228    TCGCond cond = args[4];
1229    int const_bl = const_args[2];
1230    int const_bh = const_args[3];
1231
1232    switch (cond) {
1233    case TCG_COND_EQ:
1234    case TCG_COND_NE:
1235    case TCG_COND_LTU:
1236    case TCG_COND_LEU:
1237    case TCG_COND_GTU:
1238    case TCG_COND_GEU:
1239        /*
1240         * We perform a conditional comparison.  If the high half is
1241         * equal, then overwrite the flags with the comparison of the
1242         * low half.  The resulting flags cover the whole.
1243         */
1244        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, ah, bh, const_bh);
1245        tcg_out_dat_rI(s, COND_EQ, ARITH_CMP, 0, al, bl, const_bl);
1246        return cond;
1247
1248    case TCG_COND_TSTEQ:
1249    case TCG_COND_TSTNE:
1250        /* Similar, but with TST instead of CMP. */
1251        tcg_out_dat_rI(s, COND_AL, ARITH_TST, 0, ah, bh, const_bh);
1252        tcg_out_dat_rI(s, COND_EQ, ARITH_TST, 0, al, bl, const_bl);
1253        return tcg_tst_eqne_cond(cond);
1254
1255    case TCG_COND_LT:
1256    case TCG_COND_GE:
1257        /* We perform a double-word subtraction and examine the result.
1258           We do not actually need the result of the subtract, so the
1259           low part "subtract" is a compare.  For the high half we have
1260           no choice but to compute into a temporary.  */
1261        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, al, bl, const_bl);
1262        tcg_out_dat_rI(s, COND_AL, ARITH_SBC | TO_CPSR,
1263                       TCG_REG_TMP, ah, bh, const_bh);
1264        return cond;
1265
1266    case TCG_COND_LE:
1267    case TCG_COND_GT:
1268        /* Similar, but with swapped arguments, via reversed subtract.  */
1269        tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR,
1270                       TCG_REG_TMP, al, bl, const_bl);
1271        tcg_out_dat_rI(s, COND_AL, ARITH_RSC | TO_CPSR,
1272                       TCG_REG_TMP, ah, bh, const_bh);
1273        return tcg_swap_cond(cond);
1274
1275    default:
1276        g_assert_not_reached();
1277    }
1278}
1279
1280/*
1281 * Note that TCGReg references Q-registers.
1282 * Q-regno = 2 * D-regno, so shift left by 1 while inserting.
1283 */
1284static uint32_t encode_vd(TCGReg rd)
1285{
1286    tcg_debug_assert(rd >= TCG_REG_Q0);
1287    return (extract32(rd, 3, 1) << 22) | (extract32(rd, 0, 3) << 13);
1288}
1289
1290static uint32_t encode_vn(TCGReg rn)
1291{
1292    tcg_debug_assert(rn >= TCG_REG_Q0);
1293    return (extract32(rn, 3, 1) << 7) | (extract32(rn, 0, 3) << 17);
1294}
1295
1296static uint32_t encode_vm(TCGReg rm)
1297{
1298    tcg_debug_assert(rm >= TCG_REG_Q0);
1299    return (extract32(rm, 3, 1) << 5) | (extract32(rm, 0, 3) << 1);
1300}
1301
1302static void tcg_out_vreg2(TCGContext *s, ARMInsn insn, int q, int vece,
1303                          TCGReg d, TCGReg m)
1304{
1305    tcg_out32(s, insn | (vece << 18) | (q << 6) |
1306              encode_vd(d) | encode_vm(m));
1307}
1308
1309static void tcg_out_vreg3(TCGContext *s, ARMInsn insn, int q, int vece,
1310                          TCGReg d, TCGReg n, TCGReg m)
1311{
1312    tcg_out32(s, insn | (vece << 20) | (q << 6) |
1313              encode_vd(d) | encode_vn(n) | encode_vm(m));
1314}
1315
1316static void tcg_out_vmovi(TCGContext *s, TCGReg rd,
1317                          int q, int op, int cmode, uint8_t imm8)
1318{
1319    tcg_out32(s, INSN_VMOVI | encode_vd(rd) | (q << 6) | (op << 5)
1320              | (cmode << 8) | extract32(imm8, 0, 4)
1321              | (extract32(imm8, 4, 3) << 16)
1322              | (extract32(imm8, 7, 1) << 24));
1323}
1324
1325static void tcg_out_vshifti(TCGContext *s, ARMInsn insn, int q,
1326                            TCGReg rd, TCGReg rm, int l_imm6)
1327{
1328    tcg_out32(s, insn | (q << 6) | encode_vd(rd) | encode_vm(rm) |
1329              (extract32(l_imm6, 6, 1) << 7) |
1330              (extract32(l_imm6, 0, 6) << 16));
1331}
1332
1333static void tcg_out_vldst(TCGContext *s, ARMInsn insn,
1334                          TCGReg rd, TCGReg rn, int offset)
1335{
1336    if (offset != 0) {
1337        if (check_fit_imm(offset) || check_fit_imm(-offset)) {
1338            tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
1339                            TCG_REG_TMP, rn, offset, true);
1340        } else {
1341            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP, offset);
1342            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1343                            TCG_REG_TMP, TCG_REG_TMP, rn, 0);
1344        }
1345        rn = TCG_REG_TMP;
1346    }
1347    tcg_out32(s, insn | (rn << 16) | encode_vd(rd) | 0xf);
1348}
1349
1350typedef struct {
1351    ARMCond cond;
1352    TCGReg base;
1353    int index;
1354    bool index_scratch;
1355    TCGAtomAlign aa;
1356} HostAddress;
1357
1358bool tcg_target_has_memory_bswap(MemOp memop)
1359{
1360    return false;
1361}
1362
1363static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
1364{
1365    /* We arrive at the slow path via "BLNE", so R14 contains l->raddr. */
1366    return TCG_REG_R14;
1367}
1368
1369static const TCGLdstHelperParam ldst_helper_param = {
1370    .ra_gen = ldst_ra_gen,
1371    .ntmp = 1,
1372    .tmp = { TCG_REG_TMP },
1373};
1374
1375static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1376{
1377    MemOp opc = get_memop(lb->oi);
1378
1379    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1380        return false;
1381    }
1382
1383    tcg_out_ld_helper_args(s, lb, &ldst_helper_param);
1384    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE]);
1385    tcg_out_ld_helper_ret(s, lb, false, &ldst_helper_param);
1386
1387    tcg_out_goto(s, COND_AL, lb->raddr);
1388    return true;
1389}
1390
1391static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1392{
1393    MemOp opc = get_memop(lb->oi);
1394
1395    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1396        return false;
1397    }
1398
1399    tcg_out_st_helper_args(s, lb, &ldst_helper_param);
1400
1401    /* Tail-call to the helper, which will return to the fast path.  */
1402    tcg_out_goto(s, COND_AL, qemu_st_helpers[opc & MO_SIZE]);
1403    return true;
1404}
1405
1406/* We expect to use an 9-bit sign-magnitude negative offset from ENV.  */
1407#define MIN_TLB_MASK_TABLE_OFS  -256
1408
1409static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
1410                                           TCGReg addrlo, TCGReg addrhi,
1411                                           MemOpIdx oi, bool is_ld)
1412{
1413    TCGLabelQemuLdst *ldst = NULL;
1414    MemOp opc = get_memop(oi);
1415    unsigned a_mask;
1416
1417    if (tcg_use_softmmu) {
1418        *h = (HostAddress){
1419            .cond = COND_AL,
1420            .base = addrlo,
1421            .index = TCG_REG_R1,
1422            .index_scratch = true,
1423        };
1424    } else {
1425        *h = (HostAddress){
1426            .cond = COND_AL,
1427            .base = addrlo,
1428            .index = guest_base ? TCG_REG_GUEST_BASE : -1,
1429            .index_scratch = false,
1430        };
1431    }
1432
1433    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, false);
1434    a_mask = (1 << h->aa.align) - 1;
1435
1436    if (tcg_use_softmmu) {
1437        int mem_index = get_mmuidx(oi);
1438        int cmp_off = is_ld ? offsetof(CPUTLBEntry, addr_read)
1439                            : offsetof(CPUTLBEntry, addr_write);
1440        int fast_off = tlb_mask_table_ofs(s, mem_index);
1441        unsigned s_mask = (1 << (opc & MO_SIZE)) - 1;
1442        TCGReg t_addr;
1443
1444        ldst = new_ldst_label(s);
1445        ldst->is_ld = is_ld;
1446        ldst->oi = oi;
1447        ldst->addrlo_reg = addrlo;
1448        ldst->addrhi_reg = addrhi;
1449
1450        /* Load cpu->neg.tlb.f[mmu_idx].{mask,table} into {r0,r1}.  */
1451        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, mask) != 0);
1452        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, table) != 4);
1453        tcg_out_ldrd_8(s, COND_AL, TCG_REG_R0, TCG_AREG0, fast_off);
1454
1455        /* Extract the tlb index from the address into R0.  */
1456        tcg_out_dat_reg(s, COND_AL, ARITH_AND, TCG_REG_R0, TCG_REG_R0, addrlo,
1457                        SHIFT_IMM_LSR(s->page_bits - CPU_TLB_ENTRY_BITS));
1458
1459        /*
1460         * Add the tlb_table pointer, creating the CPUTLBEntry address in R1.
1461         * Load the tlb comparator into R2/R3 and the fast path addend into R1.
1462         */
1463        QEMU_BUILD_BUG_ON(HOST_BIG_ENDIAN);
1464        if (cmp_off == 0) {
1465            if (s->addr_type == TCG_TYPE_I32) {
1466                tcg_out_ld32_rwb(s, COND_AL, TCG_REG_R2,
1467                                 TCG_REG_R1, TCG_REG_R0);
1468            } else {
1469                tcg_out_ldrd_rwb(s, COND_AL, TCG_REG_R2,
1470                                 TCG_REG_R1, TCG_REG_R0);
1471            }
1472        } else {
1473            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1474                            TCG_REG_R1, TCG_REG_R1, TCG_REG_R0, 0);
1475            if (s->addr_type == TCG_TYPE_I32) {
1476                tcg_out_ld32_12(s, COND_AL, TCG_REG_R2, TCG_REG_R1, cmp_off);
1477            } else {
1478                tcg_out_ldrd_8(s, COND_AL, TCG_REG_R2, TCG_REG_R1, cmp_off);
1479            }
1480        }
1481
1482        /* Load the tlb addend.  */
1483        tcg_out_ld32_12(s, COND_AL, TCG_REG_R1, TCG_REG_R1,
1484                        offsetof(CPUTLBEntry, addend));
1485
1486        /*
1487         * Check alignment, check comparators.
1488         * Do this in 2-4 insns.  Use MOVW for v7, if possible,
1489         * to reduce the number of sequential conditional instructions.
1490         * Almost all guests have at least 4k pages, which means that we need
1491         * to clear at least 9 bits even for an 8-byte memory, which means it
1492         * isn't worth checking for an immediate operand for BIC.
1493         *
1494         * For unaligned accesses, test the page of the last unit of alignment.
1495         * This leaves the least significant alignment bits unchanged, and of
1496         * course must be zero.
1497         */
1498        t_addr = addrlo;
1499        if (a_mask < s_mask) {
1500            t_addr = TCG_REG_R0;
1501            tcg_out_dat_imm(s, COND_AL, ARITH_ADD, t_addr,
1502                            addrlo, s_mask - a_mask);
1503        }
1504        if (use_armv7_instructions && s->page_bits <= 16) {
1505            tcg_out_movi32(s, COND_AL, TCG_REG_TMP, ~(s->page_mask | a_mask));
1506            tcg_out_dat_reg(s, COND_AL, ARITH_BIC, TCG_REG_TMP,
1507                            t_addr, TCG_REG_TMP, 0);
1508            tcg_out_dat_reg(s, COND_AL, ARITH_CMP, 0,
1509                            TCG_REG_R2, TCG_REG_TMP, 0);
1510        } else {
1511            if (a_mask) {
1512                tcg_debug_assert(a_mask <= 0xff);
1513                tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addrlo, a_mask);
1514            }
1515            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0, t_addr,
1516                            SHIFT_IMM_LSR(s->page_bits));
1517            tcg_out_dat_reg(s, (a_mask ? COND_EQ : COND_AL), ARITH_CMP,
1518                            0, TCG_REG_R2, TCG_REG_TMP,
1519                            SHIFT_IMM_LSL(s->page_bits));
1520        }
1521
1522        if (s->addr_type != TCG_TYPE_I32) {
1523            tcg_out_dat_reg(s, COND_EQ, ARITH_CMP, 0, TCG_REG_R3, addrhi, 0);
1524        }
1525    } else if (a_mask) {
1526        ldst = new_ldst_label(s);
1527        ldst->is_ld = is_ld;
1528        ldst->oi = oi;
1529        ldst->addrlo_reg = addrlo;
1530        ldst->addrhi_reg = addrhi;
1531
1532        /* We are expecting alignment to max out at 7 */
1533        tcg_debug_assert(a_mask <= 0xff);
1534        /* tst addr, #mask */
1535        tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addrlo, a_mask);
1536    }
1537
1538    return ldst;
1539}
1540
1541static void tcg_out_qemu_ld_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1542                                   TCGReg datahi, HostAddress h)
1543{
1544    TCGReg base;
1545
1546    /* Byte swapping is left to middle-end expansion. */
1547    tcg_debug_assert((opc & MO_BSWAP) == 0);
1548
1549    switch (opc & MO_SSIZE) {
1550    case MO_UB:
1551        if (h.index < 0) {
1552            tcg_out_ld8_12(s, h.cond, datalo, h.base, 0);
1553        } else {
1554            tcg_out_ld8_r(s, h.cond, datalo, h.base, h.index);
1555        }
1556        break;
1557    case MO_SB:
1558        if (h.index < 0) {
1559            tcg_out_ld8s_8(s, h.cond, datalo, h.base, 0);
1560        } else {
1561            tcg_out_ld8s_r(s, h.cond, datalo, h.base, h.index);
1562        }
1563        break;
1564    case MO_UW:
1565        if (h.index < 0) {
1566            tcg_out_ld16u_8(s, h.cond, datalo, h.base, 0);
1567        } else {
1568            tcg_out_ld16u_r(s, h.cond, datalo, h.base, h.index);
1569        }
1570        break;
1571    case MO_SW:
1572        if (h.index < 0) {
1573            tcg_out_ld16s_8(s, h.cond, datalo, h.base, 0);
1574        } else {
1575            tcg_out_ld16s_r(s, h.cond, datalo, h.base, h.index);
1576        }
1577        break;
1578    case MO_UL:
1579        if (h.index < 0) {
1580            tcg_out_ld32_12(s, h.cond, datalo, h.base, 0);
1581        } else {
1582            tcg_out_ld32_r(s, h.cond, datalo, h.base, h.index);
1583        }
1584        break;
1585    case MO_UQ:
1586        /* We used pair allocation for datalo, so already should be aligned. */
1587        tcg_debug_assert((datalo & 1) == 0);
1588        tcg_debug_assert(datahi == datalo + 1);
1589        /* LDRD requires alignment; double-check that. */
1590        if (get_alignment_bits(opc) >= MO_64) {
1591            if (h.index < 0) {
1592                tcg_out_ldrd_8(s, h.cond, datalo, h.base, 0);
1593                break;
1594            }
1595            /*
1596             * Rm (the second address op) must not overlap Rt or Rt + 1.
1597             * Since datalo is aligned, we can simplify the test via alignment.
1598             * Flip the two address arguments if that works.
1599             */
1600            if ((h.index & ~1) != datalo) {
1601                tcg_out_ldrd_r(s, h.cond, datalo, h.base, h.index);
1602                break;
1603            }
1604            if ((h.base & ~1) != datalo) {
1605                tcg_out_ldrd_r(s, h.cond, datalo, h.index, h.base);
1606                break;
1607            }
1608        }
1609        if (h.index < 0) {
1610            base = h.base;
1611            if (datalo == h.base) {
1612                tcg_out_mov_reg(s, h.cond, TCG_REG_TMP, base);
1613                base = TCG_REG_TMP;
1614            }
1615        } else if (h.index_scratch) {
1616            tcg_out_ld32_rwb(s, h.cond, datalo, h.index, h.base);
1617            tcg_out_ld32_12(s, h.cond, datahi, h.index, 4);
1618            break;
1619        } else {
1620            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1621                            h.base, h.index, SHIFT_IMM_LSL(0));
1622            base = TCG_REG_TMP;
1623        }
1624        tcg_out_ld32_12(s, h.cond, datalo, base, 0);
1625        tcg_out_ld32_12(s, h.cond, datahi, base, 4);
1626        break;
1627    default:
1628        g_assert_not_reached();
1629    }
1630}
1631
1632static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
1633                            TCGReg addrlo, TCGReg addrhi,
1634                            MemOpIdx oi, TCGType data_type)
1635{
1636    MemOp opc = get_memop(oi);
1637    TCGLabelQemuLdst *ldst;
1638    HostAddress h;
1639
1640    ldst = prepare_host_addr(s, &h, addrlo, addrhi, oi, true);
1641    if (ldst) {
1642        ldst->type = data_type;
1643        ldst->datalo_reg = datalo;
1644        ldst->datahi_reg = datahi;
1645
1646        /*
1647         * This a conditional BL only to load a pointer within this
1648         * opcode into LR for the slow path.  We will not be using
1649         * the value for a tail call.
1650         */
1651        ldst->label_ptr[0] = s->code_ptr;
1652        tcg_out_bl_imm(s, COND_NE, 0);
1653
1654        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1655        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1656    } else {
1657        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1658    }
1659}
1660
1661static void tcg_out_qemu_st_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1662                                   TCGReg datahi, HostAddress h)
1663{
1664    /* Byte swapping is left to middle-end expansion. */
1665    tcg_debug_assert((opc & MO_BSWAP) == 0);
1666
1667    switch (opc & MO_SIZE) {
1668    case MO_8:
1669        if (h.index < 0) {
1670            tcg_out_st8_12(s, h.cond, datalo, h.base, 0);
1671        } else {
1672            tcg_out_st8_r(s, h.cond, datalo, h.base, h.index);
1673        }
1674        break;
1675    case MO_16:
1676        if (h.index < 0) {
1677            tcg_out_st16_8(s, h.cond, datalo, h.base, 0);
1678        } else {
1679            tcg_out_st16_r(s, h.cond, datalo, h.base, h.index);
1680        }
1681        break;
1682    case MO_32:
1683        if (h.index < 0) {
1684            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1685        } else {
1686            tcg_out_st32_r(s, h.cond, datalo, h.base, h.index);
1687        }
1688        break;
1689    case MO_64:
1690        /* We used pair allocation for datalo, so already should be aligned. */
1691        tcg_debug_assert((datalo & 1) == 0);
1692        tcg_debug_assert(datahi == datalo + 1);
1693        /* STRD requires alignment; double-check that. */
1694        if (get_alignment_bits(opc) >= MO_64) {
1695            if (h.index < 0) {
1696                tcg_out_strd_8(s, h.cond, datalo, h.base, 0);
1697            } else {
1698                tcg_out_strd_r(s, h.cond, datalo, h.base, h.index);
1699            }
1700        } else if (h.index < 0) {
1701            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1702            tcg_out_st32_12(s, h.cond, datahi, h.base, 4);
1703        } else if (h.index_scratch) {
1704            tcg_out_st32_rwb(s, h.cond, datalo, h.index, h.base);
1705            tcg_out_st32_12(s, h.cond, datahi, h.index, 4);
1706        } else {
1707            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1708                            h.base, h.index, SHIFT_IMM_LSL(0));
1709            tcg_out_st32_12(s, h.cond, datalo, TCG_REG_TMP, 0);
1710            tcg_out_st32_12(s, h.cond, datahi, TCG_REG_TMP, 4);
1711        }
1712        break;
1713    default:
1714        g_assert_not_reached();
1715    }
1716}
1717
1718static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
1719                            TCGReg addrlo, TCGReg addrhi,
1720                            MemOpIdx oi, TCGType data_type)
1721{
1722    MemOp opc = get_memop(oi);
1723    TCGLabelQemuLdst *ldst;
1724    HostAddress h;
1725
1726    ldst = prepare_host_addr(s, &h, addrlo, addrhi, oi, false);
1727    if (ldst) {
1728        ldst->type = data_type;
1729        ldst->datalo_reg = datalo;
1730        ldst->datahi_reg = datahi;
1731
1732        h.cond = COND_EQ;
1733        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1734
1735        /* The conditional call is last, as we're going to return here. */
1736        ldst->label_ptr[0] = s->code_ptr;
1737        tcg_out_bl_imm(s, COND_NE, 0);
1738        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1739    } else {
1740        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1741    }
1742}
1743
1744static void tcg_out_epilogue(TCGContext *s);
1745
1746static void tcg_out_exit_tb(TCGContext *s, uintptr_t arg)
1747{
1748    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, arg);
1749    tcg_out_epilogue(s);
1750}
1751
1752static void tcg_out_goto_tb(TCGContext *s, int which)
1753{
1754    uintptr_t i_addr;
1755    intptr_t i_disp;
1756
1757    /* Direct branch will be patched by tb_target_set_jmp_target. */
1758    set_jmp_insn_offset(s, which);
1759    tcg_out32(s, INSN_NOP);
1760
1761    /* When branch is out of range, fall through to indirect. */
1762    i_addr = get_jmp_target_addr(s, which);
1763    i_disp = tcg_pcrel_diff(s, (void *)i_addr) - 8;
1764    tcg_debug_assert(i_disp < 0);
1765    if (i_disp >= -0xfff) {
1766        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_PC, i_disp);
1767    } else {
1768        /*
1769         * The TB is close, but outside the 12 bits addressable by
1770         * the load.  We can extend this to 20 bits with a sub of a
1771         * shifted immediate from pc.
1772         */
1773        int h = -i_disp;
1774        int l = h & 0xfff;
1775
1776        h = encode_imm_nofail(h - l);
1777        tcg_out_dat_imm(s, COND_AL, ARITH_SUB, TCG_REG_R0, TCG_REG_PC, h);
1778        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_R0, l);
1779    }
1780    set_jmp_reset_offset(s, which);
1781}
1782
1783void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1784                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1785{
1786    uintptr_t addr = tb->jmp_target_addr[n];
1787    ptrdiff_t offset = addr - (jmp_rx + 8);
1788    tcg_insn_unit insn;
1789
1790    /* Either directly branch, or fall through to indirect branch. */
1791    if (offset == sextract64(offset, 0, 26)) {
1792        /* B <addr> */
1793        insn = deposit32((COND_AL << 28) | INSN_B, 0, 24, offset >> 2);
1794    } else {
1795        insn = INSN_NOP;
1796    }
1797
1798    qatomic_set((uint32_t *)jmp_rw, insn);
1799    flush_idcache_range(jmp_rx, jmp_rw, 4);
1800}
1801
1802static void tcg_out_op(TCGContext *s, TCGOpcode opc,
1803                       const TCGArg args[TCG_MAX_OP_ARGS],
1804                       const int const_args[TCG_MAX_OP_ARGS])
1805{
1806    TCGArg a0, a1, a2, a3, a4, a5;
1807    int c;
1808
1809    switch (opc) {
1810    case INDEX_op_goto_ptr:
1811        tcg_out_b_reg(s, COND_AL, args[0]);
1812        break;
1813    case INDEX_op_br:
1814        tcg_out_goto_label(s, COND_AL, arg_label(args[0]));
1815        break;
1816
1817    case INDEX_op_ld8u_i32:
1818        tcg_out_ld8u(s, COND_AL, args[0], args[1], args[2]);
1819        break;
1820    case INDEX_op_ld8s_i32:
1821        tcg_out_ld8s(s, COND_AL, args[0], args[1], args[2]);
1822        break;
1823    case INDEX_op_ld16u_i32:
1824        tcg_out_ld16u(s, COND_AL, args[0], args[1], args[2]);
1825        break;
1826    case INDEX_op_ld16s_i32:
1827        tcg_out_ld16s(s, COND_AL, args[0], args[1], args[2]);
1828        break;
1829    case INDEX_op_ld_i32:
1830        tcg_out_ld32u(s, COND_AL, args[0], args[1], args[2]);
1831        break;
1832    case INDEX_op_st8_i32:
1833        tcg_out_st8(s, COND_AL, args[0], args[1], args[2]);
1834        break;
1835    case INDEX_op_st16_i32:
1836        tcg_out_st16(s, COND_AL, args[0], args[1], args[2]);
1837        break;
1838    case INDEX_op_st_i32:
1839        tcg_out_st32(s, COND_AL, args[0], args[1], args[2]);
1840        break;
1841
1842    case INDEX_op_movcond_i32:
1843        /* Constraints mean that v2 is always in the same register as dest,
1844         * so we only need to do "if condition passed, move v1 to dest".
1845         */
1846        c = tcg_out_cmp(s, args[5], args[1], args[2], const_args[2]);
1847        tcg_out_dat_rIK(s, tcg_cond_to_arm_cond[c], ARITH_MOV,
1848                        ARITH_MVN, args[0], 0, args[3], const_args[3]);
1849        break;
1850    case INDEX_op_add_i32:
1851        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
1852                        args[0], args[1], args[2], const_args[2]);
1853        break;
1854    case INDEX_op_sub_i32:
1855        if (const_args[1]) {
1856            if (const_args[2]) {
1857                tcg_out_movi32(s, COND_AL, args[0], args[1] - args[2]);
1858            } else {
1859                tcg_out_dat_rI(s, COND_AL, ARITH_RSB,
1860                               args[0], args[2], args[1], 1);
1861            }
1862        } else {
1863            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB, ARITH_ADD,
1864                            args[0], args[1], args[2], const_args[2]);
1865        }
1866        break;
1867    case INDEX_op_and_i32:
1868        tcg_out_dat_rIK(s, COND_AL, ARITH_AND, ARITH_BIC,
1869                        args[0], args[1], args[2], const_args[2]);
1870        break;
1871    case INDEX_op_andc_i32:
1872        tcg_out_dat_rIK(s, COND_AL, ARITH_BIC, ARITH_AND,
1873                        args[0], args[1], args[2], const_args[2]);
1874        break;
1875    case INDEX_op_or_i32:
1876        c = ARITH_ORR;
1877        goto gen_arith;
1878    case INDEX_op_xor_i32:
1879        c = ARITH_EOR;
1880        /* Fall through.  */
1881    gen_arith:
1882        tcg_out_dat_rI(s, COND_AL, c, args[0], args[1], args[2], const_args[2]);
1883        break;
1884    case INDEX_op_add2_i32:
1885        a0 = args[0], a1 = args[1], a2 = args[2];
1886        a3 = args[3], a4 = args[4], a5 = args[5];
1887        if (a0 == a3 || (a0 == a5 && !const_args[5])) {
1888            a0 = TCG_REG_TMP;
1889        }
1890        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD | TO_CPSR, ARITH_SUB | TO_CPSR,
1891                        a0, a2, a4, const_args[4]);
1892        tcg_out_dat_rIK(s, COND_AL, ARITH_ADC, ARITH_SBC,
1893                        a1, a3, a5, const_args[5]);
1894        tcg_out_mov_reg(s, COND_AL, args[0], a0);
1895        break;
1896    case INDEX_op_sub2_i32:
1897        a0 = args[0], a1 = args[1], a2 = args[2];
1898        a3 = args[3], a4 = args[4], a5 = args[5];
1899        if ((a0 == a3 && !const_args[3]) || (a0 == a5 && !const_args[5])) {
1900            a0 = TCG_REG_TMP;
1901        }
1902        if (const_args[2]) {
1903            if (const_args[4]) {
1904                tcg_out_movi32(s, COND_AL, a0, a4);
1905                a4 = a0;
1906            }
1907            tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR, a0, a4, a2, 1);
1908        } else {
1909            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB | TO_CPSR,
1910                            ARITH_ADD | TO_CPSR, a0, a2, a4, const_args[4]);
1911        }
1912        if (const_args[3]) {
1913            if (const_args[5]) {
1914                tcg_out_movi32(s, COND_AL, a1, a5);
1915                a5 = a1;
1916            }
1917            tcg_out_dat_rI(s, COND_AL, ARITH_RSC, a1, a5, a3, 1);
1918        } else {
1919            tcg_out_dat_rIK(s, COND_AL, ARITH_SBC, ARITH_ADC,
1920                            a1, a3, a5, const_args[5]);
1921        }
1922        tcg_out_mov_reg(s, COND_AL, args[0], a0);
1923        break;
1924    case INDEX_op_neg_i32:
1925        tcg_out_dat_imm(s, COND_AL, ARITH_RSB, args[0], args[1], 0);
1926        break;
1927    case INDEX_op_not_i32:
1928        tcg_out_dat_reg(s, COND_AL,
1929                        ARITH_MVN, args[0], 0, args[1], SHIFT_IMM_LSL(0));
1930        break;
1931    case INDEX_op_mul_i32:
1932        tcg_out_mul32(s, COND_AL, args[0], args[1], args[2]);
1933        break;
1934    case INDEX_op_mulu2_i32:
1935        tcg_out_umull32(s, COND_AL, args[0], args[1], args[2], args[3]);
1936        break;
1937    case INDEX_op_muls2_i32:
1938        tcg_out_smull32(s, COND_AL, args[0], args[1], args[2], args[3]);
1939        break;
1940    /* XXX: Perhaps args[2] & 0x1f is wrong */
1941    case INDEX_op_shl_i32:
1942        c = const_args[2] ?
1943                SHIFT_IMM_LSL(args[2] & 0x1f) : SHIFT_REG_LSL(args[2]);
1944        goto gen_shift32;
1945    case INDEX_op_shr_i32:
1946        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_LSR(args[2] & 0x1f) :
1947                SHIFT_IMM_LSL(0) : SHIFT_REG_LSR(args[2]);
1948        goto gen_shift32;
1949    case INDEX_op_sar_i32:
1950        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ASR(args[2] & 0x1f) :
1951                SHIFT_IMM_LSL(0) : SHIFT_REG_ASR(args[2]);
1952        goto gen_shift32;
1953    case INDEX_op_rotr_i32:
1954        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ROR(args[2] & 0x1f) :
1955                SHIFT_IMM_LSL(0) : SHIFT_REG_ROR(args[2]);
1956        /* Fall through.  */
1957    gen_shift32:
1958        tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1], c);
1959        break;
1960
1961    case INDEX_op_rotl_i32:
1962        if (const_args[2]) {
1963            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
1964                            ((0x20 - args[2]) & 0x1f) ?
1965                            SHIFT_IMM_ROR((0x20 - args[2]) & 0x1f) :
1966                            SHIFT_IMM_LSL(0));
1967        } else {
1968            tcg_out_dat_imm(s, COND_AL, ARITH_RSB, TCG_REG_TMP, args[2], 0x20);
1969            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
1970                            SHIFT_REG_ROR(TCG_REG_TMP));
1971        }
1972        break;
1973
1974    case INDEX_op_ctz_i32:
1975        tcg_out_dat_reg(s, COND_AL, INSN_RBIT, TCG_REG_TMP, 0, args[1], 0);
1976        a1 = TCG_REG_TMP;
1977        goto do_clz;
1978
1979    case INDEX_op_clz_i32:
1980        a1 = args[1];
1981    do_clz:
1982        a0 = args[0];
1983        a2 = args[2];
1984        c = const_args[2];
1985        if (c && a2 == 32) {
1986            tcg_out_dat_reg(s, COND_AL, INSN_CLZ, a0, 0, a1, 0);
1987            break;
1988        }
1989        tcg_out_dat_imm(s, COND_AL, ARITH_CMP, 0, a1, 0);
1990        tcg_out_dat_reg(s, COND_NE, INSN_CLZ, a0, 0, a1, 0);
1991        if (c || a0 != a2) {
1992            tcg_out_dat_rIK(s, COND_EQ, ARITH_MOV, ARITH_MVN, a0, 0, a2, c);
1993        }
1994        break;
1995
1996    case INDEX_op_brcond_i32:
1997        c = tcg_out_cmp(s, args[2], args[0], args[1], const_args[1]);
1998        tcg_out_goto_label(s, tcg_cond_to_arm_cond[c], arg_label(args[3]));
1999        break;
2000    case INDEX_op_setcond_i32:
2001        c = tcg_out_cmp(s, args[3], args[1], args[2], const_args[2]);
2002        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c],
2003                        ARITH_MOV, args[0], 0, 1);
2004        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2005                        ARITH_MOV, args[0], 0, 0);
2006        break;
2007    case INDEX_op_negsetcond_i32:
2008        c = tcg_out_cmp(s, args[3], args[1], args[2], const_args[2]);
2009        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c],
2010                        ARITH_MVN, args[0], 0, 0);
2011        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2012                        ARITH_MOV, args[0], 0, 0);
2013        break;
2014
2015    case INDEX_op_brcond2_i32:
2016        c = tcg_out_cmp2(s, args, const_args);
2017        tcg_out_goto_label(s, tcg_cond_to_arm_cond[c], arg_label(args[5]));
2018        break;
2019    case INDEX_op_setcond2_i32:
2020        c = tcg_out_cmp2(s, args + 1, const_args + 1);
2021        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c], ARITH_MOV, args[0], 0, 1);
2022        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2023                        ARITH_MOV, args[0], 0, 0);
2024        break;
2025
2026    case INDEX_op_qemu_ld_a32_i32:
2027        tcg_out_qemu_ld(s, args[0], -1, args[1], -1, args[2], TCG_TYPE_I32);
2028        break;
2029    case INDEX_op_qemu_ld_a64_i32:
2030        tcg_out_qemu_ld(s, args[0], -1, args[1], args[2],
2031                        args[3], TCG_TYPE_I32);
2032        break;
2033    case INDEX_op_qemu_ld_a32_i64:
2034        tcg_out_qemu_ld(s, args[0], args[1], args[2], -1,
2035                        args[3], TCG_TYPE_I64);
2036        break;
2037    case INDEX_op_qemu_ld_a64_i64:
2038        tcg_out_qemu_ld(s, args[0], args[1], args[2], args[3],
2039                        args[4], TCG_TYPE_I64);
2040        break;
2041
2042    case INDEX_op_qemu_st_a32_i32:
2043        tcg_out_qemu_st(s, args[0], -1, args[1], -1, args[2], TCG_TYPE_I32);
2044        break;
2045    case INDEX_op_qemu_st_a64_i32:
2046        tcg_out_qemu_st(s, args[0], -1, args[1], args[2],
2047                        args[3], TCG_TYPE_I32);
2048        break;
2049    case INDEX_op_qemu_st_a32_i64:
2050        tcg_out_qemu_st(s, args[0], args[1], args[2], -1,
2051                        args[3], TCG_TYPE_I64);
2052        break;
2053    case INDEX_op_qemu_st_a64_i64:
2054        tcg_out_qemu_st(s, args[0], args[1], args[2], args[3],
2055                        args[4], TCG_TYPE_I64);
2056        break;
2057
2058    case INDEX_op_bswap16_i32:
2059        tcg_out_bswap16(s, COND_AL, args[0], args[1], args[2]);
2060        break;
2061    case INDEX_op_bswap32_i32:
2062        tcg_out_bswap32(s, COND_AL, args[0], args[1]);
2063        break;
2064
2065    case INDEX_op_deposit_i32:
2066        tcg_out_deposit(s, COND_AL, args[0], args[2],
2067                        args[3], args[4], const_args[2]);
2068        break;
2069    case INDEX_op_extract_i32:
2070        tcg_out_extract(s, COND_AL, args[0], args[1], args[2], args[3]);
2071        break;
2072    case INDEX_op_sextract_i32:
2073        tcg_out_sextract(s, COND_AL, args[0], args[1], args[2], args[3]);
2074        break;
2075    case INDEX_op_extract2_i32:
2076        /* ??? These optimization vs zero should be generic.  */
2077        /* ??? But we can't substitute 2 for 1 in the opcode stream yet.  */
2078        if (const_args[1]) {
2079            if (const_args[2]) {
2080                tcg_out_movi(s, TCG_TYPE_REG, args[0], 0);
2081            } else {
2082                tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2083                                args[2], SHIFT_IMM_LSL(32 - args[3]));
2084            }
2085        } else if (const_args[2]) {
2086            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2087                            args[1], SHIFT_IMM_LSR(args[3]));
2088        } else {
2089            /* We can do extract2 in 2 insns, vs the 3 required otherwise.  */
2090            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0,
2091                            args[2], SHIFT_IMM_LSL(32 - args[3]));
2092            tcg_out_dat_reg(s, COND_AL, ARITH_ORR, args[0], TCG_REG_TMP,
2093                            args[1], SHIFT_IMM_LSR(args[3]));
2094        }
2095        break;
2096
2097    case INDEX_op_div_i32:
2098        tcg_out_sdiv(s, COND_AL, args[0], args[1], args[2]);
2099        break;
2100    case INDEX_op_divu_i32:
2101        tcg_out_udiv(s, COND_AL, args[0], args[1], args[2]);
2102        break;
2103
2104    case INDEX_op_mb:
2105        tcg_out_mb(s, args[0]);
2106        break;
2107
2108    case INDEX_op_mov_i32:  /* Always emitted via tcg_out_mov.  */
2109    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
2110    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
2111    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
2112    case INDEX_op_ext8s_i32:  /* Always emitted via tcg_reg_alloc_op.  */
2113    case INDEX_op_ext8u_i32:
2114    case INDEX_op_ext16s_i32:
2115    case INDEX_op_ext16u_i32:
2116    default:
2117        g_assert_not_reached();
2118    }
2119}
2120
2121static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
2122{
2123    switch (op) {
2124    case INDEX_op_goto_ptr:
2125        return C_O0_I1(r);
2126
2127    case INDEX_op_ld8u_i32:
2128    case INDEX_op_ld8s_i32:
2129    case INDEX_op_ld16u_i32:
2130    case INDEX_op_ld16s_i32:
2131    case INDEX_op_ld_i32:
2132    case INDEX_op_neg_i32:
2133    case INDEX_op_not_i32:
2134    case INDEX_op_bswap16_i32:
2135    case INDEX_op_bswap32_i32:
2136    case INDEX_op_ext8s_i32:
2137    case INDEX_op_ext16s_i32:
2138    case INDEX_op_ext16u_i32:
2139    case INDEX_op_extract_i32:
2140    case INDEX_op_sextract_i32:
2141        return C_O1_I1(r, r);
2142
2143    case INDEX_op_st8_i32:
2144    case INDEX_op_st16_i32:
2145    case INDEX_op_st_i32:
2146        return C_O0_I2(r, r);
2147
2148    case INDEX_op_add_i32:
2149    case INDEX_op_sub_i32:
2150    case INDEX_op_setcond_i32:
2151    case INDEX_op_negsetcond_i32:
2152        return C_O1_I2(r, r, rIN);
2153
2154    case INDEX_op_and_i32:
2155    case INDEX_op_andc_i32:
2156    case INDEX_op_clz_i32:
2157    case INDEX_op_ctz_i32:
2158        return C_O1_I2(r, r, rIK);
2159
2160    case INDEX_op_mul_i32:
2161    case INDEX_op_div_i32:
2162    case INDEX_op_divu_i32:
2163        return C_O1_I2(r, r, r);
2164
2165    case INDEX_op_mulu2_i32:
2166    case INDEX_op_muls2_i32:
2167        return C_O2_I2(r, r, r, r);
2168
2169    case INDEX_op_or_i32:
2170    case INDEX_op_xor_i32:
2171        return C_O1_I2(r, r, rI);
2172
2173    case INDEX_op_shl_i32:
2174    case INDEX_op_shr_i32:
2175    case INDEX_op_sar_i32:
2176    case INDEX_op_rotl_i32:
2177    case INDEX_op_rotr_i32:
2178        return C_O1_I2(r, r, ri);
2179
2180    case INDEX_op_brcond_i32:
2181        return C_O0_I2(r, rIN);
2182    case INDEX_op_deposit_i32:
2183        return C_O1_I2(r, 0, rZ);
2184    case INDEX_op_extract2_i32:
2185        return C_O1_I2(r, rZ, rZ);
2186    case INDEX_op_movcond_i32:
2187        return C_O1_I4(r, r, rIN, rIK, 0);
2188    case INDEX_op_add2_i32:
2189        return C_O2_I4(r, r, r, r, rIN, rIK);
2190    case INDEX_op_sub2_i32:
2191        return C_O2_I4(r, r, rI, rI, rIN, rIK);
2192    case INDEX_op_brcond2_i32:
2193        return C_O0_I4(r, r, rI, rI);
2194    case INDEX_op_setcond2_i32:
2195        return C_O1_I4(r, r, r, rI, rI);
2196
2197    case INDEX_op_qemu_ld_a32_i32:
2198        return C_O1_I1(r, q);
2199    case INDEX_op_qemu_ld_a64_i32:
2200        return C_O1_I2(r, q, q);
2201    case INDEX_op_qemu_ld_a32_i64:
2202        return C_O2_I1(e, p, q);
2203    case INDEX_op_qemu_ld_a64_i64:
2204        return C_O2_I2(e, p, q, q);
2205    case INDEX_op_qemu_st_a32_i32:
2206        return C_O0_I2(q, q);
2207    case INDEX_op_qemu_st_a64_i32:
2208        return C_O0_I3(q, q, q);
2209    case INDEX_op_qemu_st_a32_i64:
2210        return C_O0_I3(Q, p, q);
2211    case INDEX_op_qemu_st_a64_i64:
2212        return C_O0_I4(Q, p, q, q);
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    if (!tcg_use_softmmu && guest_base) {
2961        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_GUEST_BASE, guest_base);
2962        tcg_regset_set_reg(s->reserved_regs, TCG_REG_GUEST_BASE);
2963    }
2964
2965    tcg_out_b_reg(s, COND_AL, tcg_target_call_iarg_regs[1]);
2966
2967    /*
2968     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
2969     * and fall through to the rest of the epilogue.
2970     */
2971    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
2972    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, 0);
2973    tcg_out_epilogue(s);
2974}
2975
2976static void tcg_out_epilogue(TCGContext *s)
2977{
2978    /* Release local stack frame.  */
2979    tcg_out_dat_rI(s, COND_AL, ARITH_ADD, TCG_REG_CALL_STACK,
2980                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
2981
2982    /* ldmia sp!, { r4 - r11, pc } */
2983    tcg_out_ldstm(s, COND_AL, INSN_LDMIA, TCG_REG_CALL_STACK,
2984                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
2985                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
2986                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_PC));
2987}
2988
2989static void tcg_out_tb_start(TCGContext *s)
2990{
2991    /* nothing to do */
2992}
2993
2994typedef struct {
2995    DebugFrameHeader h;
2996    uint8_t fde_def_cfa[4];
2997    uint8_t fde_reg_ofs[18];
2998} DebugFrame;
2999
3000#define ELF_HOST_MACHINE EM_ARM
3001
3002/* We're expecting a 2 byte uleb128 encoded value.  */
3003QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
3004
3005static const DebugFrame debug_frame = {
3006    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
3007    .h.cie.id = -1,
3008    .h.cie.version = 1,
3009    .h.cie.code_align = 1,
3010    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
3011    .h.cie.return_column = 14,
3012
3013    /* Total FDE size does not include the "len" member.  */
3014    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
3015
3016    .fde_def_cfa = {
3017        12, 13,                         /* DW_CFA_def_cfa sp, ... */
3018        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
3019        (FRAME_SIZE >> 7)
3020    },
3021    .fde_reg_ofs = {
3022        /* The following must match the stmdb in the prologue.  */
3023        0x8e, 1,                        /* DW_CFA_offset, lr, -4 */
3024        0x8b, 2,                        /* DW_CFA_offset, r11, -8 */
3025        0x8a, 3,                        /* DW_CFA_offset, r10, -12 */
3026        0x89, 4,                        /* DW_CFA_offset, r9, -16 */
3027        0x88, 5,                        /* DW_CFA_offset, r8, -20 */
3028        0x87, 6,                        /* DW_CFA_offset, r7, -24 */
3029        0x86, 7,                        /* DW_CFA_offset, r6, -28 */
3030        0x85, 8,                        /* DW_CFA_offset, r5, -32 */
3031        0x84, 9,                        /* DW_CFA_offset, r4, -36 */
3032    }
3033};
3034
3035void tcg_register_jit(const void *buf, size_t buf_size)
3036{
3037    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
3038}
3039