xref: /openbmc/qemu/tcg/arm/tcg-target.c.inc (revision 3521f9cadc29c7d68b73b325ddb46a7acebf6212)
1/*
2 * Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2008 Andrzej Zaborowski
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to deal
8 * in the Software without restriction, including without limitation the rights
9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10 * copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22 * THE SOFTWARE.
23 */
24
25#include "elf.h"
26
27int arm_arch = __ARM_ARCH;
28
29#ifndef use_idiv_instructions
30bool use_idiv_instructions;
31#endif
32#ifndef use_neon_instructions
33bool use_neon_instructions;
34#endif
35
36/* Used for function call generation. */
37#define TCG_TARGET_STACK_ALIGN          8
38#define TCG_TARGET_CALL_STACK_OFFSET    0
39#define TCG_TARGET_CALL_ARG_I32         TCG_CALL_ARG_NORMAL
40#define TCG_TARGET_CALL_ARG_I64         TCG_CALL_ARG_EVEN
41#define TCG_TARGET_CALL_ARG_I128        TCG_CALL_ARG_EVEN
42#define TCG_TARGET_CALL_RET_I128        TCG_CALL_RET_BY_REF
43
44#ifdef CONFIG_DEBUG_TCG
45static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
46    "%r0",  "%r1",  "%r2",  "%r3",  "%r4",  "%r5",  "%r6",  "%r7",
47    "%r8",  "%r9",  "%r10", "%r11", "%r12", "%sp",  "%r14", "%pc",
48    "%q0",  "%q1",  "%q2",  "%q3",  "%q4",  "%q5",  "%q6",  "%q7",
49    "%q8",  "%q9",  "%q10", "%q11", "%q12", "%q13", "%q14", "%q15",
50};
51#endif
52
53static const int tcg_target_reg_alloc_order[] = {
54    TCG_REG_R4,
55    TCG_REG_R5,
56    TCG_REG_R6,
57    TCG_REG_R7,
58    TCG_REG_R8,
59    TCG_REG_R9,
60    TCG_REG_R10,
61    TCG_REG_R11,
62    TCG_REG_R13,
63    TCG_REG_R0,
64    TCG_REG_R1,
65    TCG_REG_R2,
66    TCG_REG_R3,
67    TCG_REG_R12,
68    TCG_REG_R14,
69
70    TCG_REG_Q0,
71    TCG_REG_Q1,
72    TCG_REG_Q2,
73    TCG_REG_Q3,
74    /* Q4 - Q7 are call-saved, and skipped. */
75    TCG_REG_Q8,
76    TCG_REG_Q9,
77    TCG_REG_Q10,
78    TCG_REG_Q11,
79    TCG_REG_Q12,
80    TCG_REG_Q13,
81    TCG_REG_Q14,
82    TCG_REG_Q15,
83};
84
85static const int tcg_target_call_iarg_regs[4] = {
86    TCG_REG_R0, TCG_REG_R1, TCG_REG_R2, TCG_REG_R3
87};
88
89static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
90{
91    tcg_debug_assert(kind == TCG_CALL_RET_NORMAL);
92    tcg_debug_assert(slot >= 0 && slot <= 3);
93    return TCG_REG_R0 + slot;
94}
95
96#define TCG_REG_TMP  TCG_REG_R12
97#define TCG_VEC_TMP  TCG_REG_Q15
98#define TCG_REG_GUEST_BASE  TCG_REG_R11
99
100typedef enum {
101    COND_EQ = 0x0,
102    COND_NE = 0x1,
103    COND_CS = 0x2,	/* Unsigned greater or equal */
104    COND_CC = 0x3,	/* Unsigned less than */
105    COND_MI = 0x4,	/* Negative */
106    COND_PL = 0x5,	/* Zero or greater */
107    COND_VS = 0x6,	/* Overflow */
108    COND_VC = 0x7,	/* No overflow */
109    COND_HI = 0x8,	/* Unsigned greater than */
110    COND_LS = 0x9,	/* Unsigned less or equal */
111    COND_GE = 0xa,
112    COND_LT = 0xb,
113    COND_GT = 0xc,
114    COND_LE = 0xd,
115    COND_AL = 0xe,
116} ARMCond;
117
118#define TO_CPSR (1 << 20)
119
120#define SHIFT_IMM_LSL(im)	(((im) << 7) | 0x00)
121#define SHIFT_IMM_LSR(im)	(((im) << 7) | 0x20)
122#define SHIFT_IMM_ASR(im)	(((im) << 7) | 0x40)
123#define SHIFT_IMM_ROR(im)	(((im) << 7) | 0x60)
124#define SHIFT_REG_LSL(rs)	(((rs) << 8) | 0x10)
125#define SHIFT_REG_LSR(rs)	(((rs) << 8) | 0x30)
126#define SHIFT_REG_ASR(rs)	(((rs) << 8) | 0x50)
127#define SHIFT_REG_ROR(rs)	(((rs) << 8) | 0x70)
128
129typedef enum {
130    ARITH_AND = 0x0 << 21,
131    ARITH_EOR = 0x1 << 21,
132    ARITH_SUB = 0x2 << 21,
133    ARITH_RSB = 0x3 << 21,
134    ARITH_ADD = 0x4 << 21,
135    ARITH_ADC = 0x5 << 21,
136    ARITH_SBC = 0x6 << 21,
137    ARITH_RSC = 0x7 << 21,
138    ARITH_TST = 0x8 << 21 | TO_CPSR,
139    ARITH_CMP = 0xa << 21 | TO_CPSR,
140    ARITH_CMN = 0xb << 21 | TO_CPSR,
141    ARITH_ORR = 0xc << 21,
142    ARITH_MOV = 0xd << 21,
143    ARITH_BIC = 0xe << 21,
144    ARITH_MVN = 0xf << 21,
145
146    INSN_B         = 0x0a000000,
147
148    INSN_CLZ       = 0x016f0f10,
149    INSN_RBIT      = 0x06ff0f30,
150
151    INSN_LDMIA     = 0x08b00000,
152    INSN_STMDB     = 0x09200000,
153
154    INSN_LDR_IMM   = 0x04100000,
155    INSN_LDR_REG   = 0x06100000,
156    INSN_STR_IMM   = 0x04000000,
157    INSN_STR_REG   = 0x06000000,
158
159    INSN_LDRH_IMM  = 0x005000b0,
160    INSN_LDRH_REG  = 0x001000b0,
161    INSN_LDRSH_IMM = 0x005000f0,
162    INSN_LDRSH_REG = 0x001000f0,
163    INSN_STRH_IMM  = 0x004000b0,
164    INSN_STRH_REG  = 0x000000b0,
165
166    INSN_LDRB_IMM  = 0x04500000,
167    INSN_LDRB_REG  = 0x06500000,
168    INSN_LDRSB_IMM = 0x005000d0,
169    INSN_LDRSB_REG = 0x001000d0,
170    INSN_STRB_IMM  = 0x04400000,
171    INSN_STRB_REG  = 0x06400000,
172
173    INSN_LDRD_IMM  = 0x004000d0,
174    INSN_LDRD_REG  = 0x000000d0,
175    INSN_STRD_IMM  = 0x004000f0,
176    INSN_STRD_REG  = 0x000000f0,
177
178    INSN_DMB_ISH   = 0xf57ff05b,
179    INSN_DMB_MCR   = 0xee070fba,
180
181    /* Architected nop introduced in v6k.  */
182    /* ??? This is an MSR (imm) 0,0,0 insn.  Anyone know if this
183       also Just So Happened to do nothing on pre-v6k so that we
184       don't need to conditionalize it?  */
185    INSN_NOP_v6k   = 0xe320f000,
186    /* Otherwise the assembler uses mov r0,r0 */
187    INSN_NOP_v4    = (COND_AL << 28) | ARITH_MOV,
188
189    INSN_VADD      = 0xf2000800,
190    INSN_VAND      = 0xf2000110,
191    INSN_VBIC      = 0xf2100110,
192    INSN_VEOR      = 0xf3000110,
193    INSN_VORN      = 0xf2300110,
194    INSN_VORR      = 0xf2200110,
195    INSN_VSUB      = 0xf3000800,
196    INSN_VMUL      = 0xf2000910,
197    INSN_VQADD     = 0xf2000010,
198    INSN_VQADD_U   = 0xf3000010,
199    INSN_VQSUB     = 0xf2000210,
200    INSN_VQSUB_U   = 0xf3000210,
201    INSN_VMAX      = 0xf2000600,
202    INSN_VMAX_U    = 0xf3000600,
203    INSN_VMIN      = 0xf2000610,
204    INSN_VMIN_U    = 0xf3000610,
205
206    INSN_VABS      = 0xf3b10300,
207    INSN_VMVN      = 0xf3b00580,
208    INSN_VNEG      = 0xf3b10380,
209
210    INSN_VCEQ0     = 0xf3b10100,
211    INSN_VCGT0     = 0xf3b10000,
212    INSN_VCGE0     = 0xf3b10080,
213    INSN_VCLE0     = 0xf3b10180,
214    INSN_VCLT0     = 0xf3b10200,
215
216    INSN_VCEQ      = 0xf3000810,
217    INSN_VCGE      = 0xf2000310,
218    INSN_VCGT      = 0xf2000300,
219    INSN_VCGE_U    = 0xf3000310,
220    INSN_VCGT_U    = 0xf3000300,
221
222    INSN_VSHLI     = 0xf2800510,  /* VSHL (immediate) */
223    INSN_VSARI     = 0xf2800010,  /* VSHR.S */
224    INSN_VSHRI     = 0xf3800010,  /* VSHR.U */
225    INSN_VSLI      = 0xf3800510,
226    INSN_VSHL_S    = 0xf2000400,  /* VSHL.S (register) */
227    INSN_VSHL_U    = 0xf3000400,  /* VSHL.U (register) */
228
229    INSN_VBSL      = 0xf3100110,
230    INSN_VBIT      = 0xf3200110,
231    INSN_VBIF      = 0xf3300110,
232
233    INSN_VTST      = 0xf2000810,
234
235    INSN_VDUP_G    = 0xee800b10,  /* VDUP (ARM core register) */
236    INSN_VDUP_S    = 0xf3b00c00,  /* VDUP (scalar) */
237    INSN_VLDR_D    = 0xed100b00,  /* VLDR.64 */
238    INSN_VLD1      = 0xf4200000,  /* VLD1 (multiple single elements) */
239    INSN_VLD1R     = 0xf4a00c00,  /* VLD1 (single element to all lanes) */
240    INSN_VST1      = 0xf4000000,  /* VST1 (multiple single elements) */
241    INSN_VMOVI     = 0xf2800010,  /* VMOV (immediate) */
242} ARMInsn;
243
244#define INSN_NOP   (use_armv7_instructions ? INSN_NOP_v6k : INSN_NOP_v4)
245
246static const uint8_t tcg_cond_to_arm_cond[] = {
247    [TCG_COND_EQ] = COND_EQ,
248    [TCG_COND_NE] = COND_NE,
249    [TCG_COND_LT] = COND_LT,
250    [TCG_COND_GE] = COND_GE,
251    [TCG_COND_LE] = COND_LE,
252    [TCG_COND_GT] = COND_GT,
253    /* unsigned */
254    [TCG_COND_LTU] = COND_CC,
255    [TCG_COND_GEU] = COND_CS,
256    [TCG_COND_LEU] = COND_LS,
257    [TCG_COND_GTU] = COND_HI,
258};
259
260static int encode_imm(uint32_t imm);
261
262/* TCG private relocation type: add with pc+imm8 */
263#define R_ARM_PC8  11
264
265/* TCG private relocation type: vldr with imm8 << 2 */
266#define R_ARM_PC11 12
267
268static bool reloc_pc24(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
269{
270    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
271    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) >> 2;
272
273    if (offset == sextract32(offset, 0, 24)) {
274        *src_rw = deposit32(*src_rw, 0, 24, offset);
275        return true;
276    }
277    return false;
278}
279
280static bool reloc_pc13(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
281{
282    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
283    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
284
285    if (offset >= -0xfff && offset <= 0xfff) {
286        tcg_insn_unit insn = *src_rw;
287        bool u = (offset >= 0);
288        if (!u) {
289            offset = -offset;
290        }
291        insn = deposit32(insn, 23, 1, u);
292        insn = deposit32(insn, 0, 12, offset);
293        *src_rw = insn;
294        return true;
295    }
296    return false;
297}
298
299static bool reloc_pc11(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
300{
301    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
302    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) / 4;
303
304    if (offset >= -0xff && offset <= 0xff) {
305        tcg_insn_unit insn = *src_rw;
306        bool u = (offset >= 0);
307        if (!u) {
308            offset = -offset;
309        }
310        insn = deposit32(insn, 23, 1, u);
311        insn = deposit32(insn, 0, 8, offset);
312        *src_rw = insn;
313        return true;
314    }
315    return false;
316}
317
318static bool reloc_pc8(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
319{
320    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
321    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
322    int imm12 = encode_imm(offset);
323
324    if (imm12 >= 0) {
325        *src_rw = deposit32(*src_rw, 0, 12, imm12);
326        return true;
327    }
328    return false;
329}
330
331static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
332                        intptr_t value, intptr_t addend)
333{
334    tcg_debug_assert(addend == 0);
335    switch (type) {
336    case R_ARM_PC24:
337        return reloc_pc24(code_ptr, (const tcg_insn_unit *)value);
338    case R_ARM_PC13:
339        return reloc_pc13(code_ptr, (const tcg_insn_unit *)value);
340    case R_ARM_PC11:
341        return reloc_pc11(code_ptr, (const tcg_insn_unit *)value);
342    case R_ARM_PC8:
343        return reloc_pc8(code_ptr, (const tcg_insn_unit *)value);
344    default:
345        g_assert_not_reached();
346    }
347}
348
349#define TCG_CT_CONST_ARM  0x100
350#define TCG_CT_CONST_INV  0x200
351#define TCG_CT_CONST_NEG  0x400
352#define TCG_CT_CONST_ZERO 0x800
353#define TCG_CT_CONST_ORRI 0x1000
354#define TCG_CT_CONST_ANDI 0x2000
355
356#define ALL_GENERAL_REGS  0xffffu
357#define ALL_VECTOR_REGS   0xffff0000u
358
359/*
360 * r0-r3 will be overwritten when reading the tlb entry (system-mode only);
361 * r14 will be overwritten by the BLNE branching to the slow path.
362 */
363#define ALL_QLDST_REGS \
364    (ALL_GENERAL_REGS & ~((tcg_use_softmmu ? 0xf : 0) | (1 << TCG_REG_R14)))
365
366/*
367 * ARM immediates for ALU instructions are made of an unsigned 8-bit
368 * right-rotated by an even amount between 0 and 30.
369 *
370 * Return < 0 if @imm cannot be encoded, else the entire imm12 field.
371 */
372static int encode_imm(uint32_t imm)
373{
374    uint32_t rot, imm8;
375
376    /* Simple case, no rotation required. */
377    if ((imm & ~0xff) == 0) {
378        return imm;
379    }
380
381    /* Next, try a simple even shift.  */
382    rot = ctz32(imm) & ~1;
383    imm8 = imm >> rot;
384    rot = 32 - rot;
385    if ((imm8 & ~0xff) == 0) {
386        goto found;
387    }
388
389    /*
390     * Finally, try harder with rotations.
391     * The ctz test above will have taken care of rotates >= 8.
392     */
393    for (rot = 2; rot < 8; rot += 2) {
394        imm8 = rol32(imm, rot);
395        if ((imm8 & ~0xff) == 0) {
396            goto found;
397        }
398    }
399    /* Fail: imm cannot be encoded. */
400    return -1;
401
402 found:
403    /* Note that rot is even, and we discard bit 0 by shifting by 7. */
404    return rot << 7 | imm8;
405}
406
407static int encode_imm_nofail(uint32_t imm)
408{
409    int ret = encode_imm(imm);
410    tcg_debug_assert(ret >= 0);
411    return ret;
412}
413
414static bool check_fit_imm(uint32_t imm)
415{
416    return encode_imm(imm) >= 0;
417}
418
419/* Return true if v16 is a valid 16-bit shifted immediate.  */
420static bool is_shimm16(uint16_t v16, int *cmode, int *imm8)
421{
422    if (v16 == (v16 & 0xff)) {
423        *cmode = 0x8;
424        *imm8 = v16 & 0xff;
425        return true;
426    } else if (v16 == (v16 & 0xff00)) {
427        *cmode = 0xa;
428        *imm8 = v16 >> 8;
429        return true;
430    }
431    return false;
432}
433
434/* Return true if v32 is a valid 32-bit shifted immediate.  */
435static bool is_shimm32(uint32_t v32, int *cmode, int *imm8)
436{
437    if (v32 == (v32 & 0xff)) {
438        *cmode = 0x0;
439        *imm8 = v32 & 0xff;
440        return true;
441    } else if (v32 == (v32 & 0xff00)) {
442        *cmode = 0x2;
443        *imm8 = (v32 >> 8) & 0xff;
444        return true;
445    } else if (v32 == (v32 & 0xff0000)) {
446        *cmode = 0x4;
447        *imm8 = (v32 >> 16) & 0xff;
448        return true;
449    } else if (v32 == (v32 & 0xff000000)) {
450        *cmode = 0x6;
451        *imm8 = v32 >> 24;
452        return true;
453    }
454    return false;
455}
456
457/* Return true if v32 is a valid 32-bit shifting ones immediate.  */
458static bool is_soimm32(uint32_t v32, int *cmode, int *imm8)
459{
460    if ((v32 & 0xffff00ff) == 0xff) {
461        *cmode = 0xc;
462        *imm8 = (v32 >> 8) & 0xff;
463        return true;
464    } else if ((v32 & 0xff00ffff) == 0xffff) {
465        *cmode = 0xd;
466        *imm8 = (v32 >> 16) & 0xff;
467        return true;
468    }
469    return false;
470}
471
472/*
473 * Return non-zero if v32 can be formed by MOVI+ORR.
474 * Place the parameters for MOVI in (cmode, imm8).
475 * Return the cmode for ORR; the imm8 can be had via extraction from v32.
476 */
477static int is_shimm32_pair(uint32_t v32, int *cmode, int *imm8)
478{
479    int i;
480
481    for (i = 6; i > 0; i -= 2) {
482        /* Mask out one byte we can add with ORR.  */
483        uint32_t tmp = v32 & ~(0xffu << (i * 4));
484        if (is_shimm32(tmp, cmode, imm8) ||
485            is_soimm32(tmp, cmode, imm8)) {
486            break;
487        }
488    }
489    return i;
490}
491
492/* Return true if V is a valid 16-bit or 32-bit shifted immediate.  */
493static bool is_shimm1632(uint32_t v32, int *cmode, int *imm8)
494{
495    if (v32 == deposit32(v32, 16, 16, v32)) {
496        return is_shimm16(v32, cmode, imm8);
497    } else {
498        return is_shimm32(v32, cmode, imm8);
499    }
500}
501
502/* Test if a constant matches the constraint.
503 * TODO: define constraints for:
504 *
505 * ldr/str offset:   between -0xfff and 0xfff
506 * ldrh/strh offset: between -0xff and 0xff
507 * mov operand2:     values represented with x << (2 * y), x < 0x100
508 * add, sub, eor...: ditto
509 */
510static bool tcg_target_const_match(int64_t val, int ct,
511                                   TCGType type, TCGCond cond, int vece)
512{
513    if (ct & TCG_CT_CONST) {
514        return 1;
515    } else if ((ct & TCG_CT_CONST_ARM) && check_fit_imm(val)) {
516        return 1;
517    } else if ((ct & TCG_CT_CONST_INV) && check_fit_imm(~val)) {
518        return 1;
519    } else if ((ct & TCG_CT_CONST_NEG) && check_fit_imm(-val)) {
520        return 1;
521    } else if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
522        return 1;
523    }
524
525    switch (ct & (TCG_CT_CONST_ORRI | TCG_CT_CONST_ANDI)) {
526    case 0:
527        break;
528    case TCG_CT_CONST_ANDI:
529        val = ~val;
530        /* fallthru */
531    case TCG_CT_CONST_ORRI:
532        if (val == deposit64(val, 32, 32, val)) {
533            int cmode, imm8;
534            return is_shimm1632(val, &cmode, &imm8);
535        }
536        break;
537    default:
538        /* Both bits should not be set for the same insn.  */
539        g_assert_not_reached();
540    }
541
542    return 0;
543}
544
545static void tcg_out_b_imm(TCGContext *s, ARMCond cond, int32_t offset)
546{
547    tcg_out32(s, (cond << 28) | INSN_B |
548                    (((offset - 8) >> 2) & 0x00ffffff));
549}
550
551static void tcg_out_bl_imm(TCGContext *s, ARMCond cond, int32_t offset)
552{
553    tcg_out32(s, (cond << 28) | 0x0b000000 |
554                    (((offset - 8) >> 2) & 0x00ffffff));
555}
556
557static void tcg_out_blx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
558{
559    tcg_out32(s, (cond << 28) | 0x012fff30 | rn);
560}
561
562static void tcg_out_blx_imm(TCGContext *s, int32_t offset)
563{
564    tcg_out32(s, 0xfa000000 | ((offset & 2) << 23) |
565                (((offset - 8) >> 2) & 0x00ffffff));
566}
567
568static void tcg_out_dat_reg(TCGContext *s, ARMCond cond, ARMInsn opc,
569                            TCGReg rd, TCGReg rn, TCGReg rm, int shift)
570{
571    tcg_out32(s, (cond << 28) | (0 << 25) | opc |
572                    (rn << 16) | (rd << 12) | shift | rm);
573}
574
575static void tcg_out_mov_reg(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rm)
576{
577    /* Simple reg-reg move, optimising out the 'do nothing' case */
578    if (rd != rm) {
579        tcg_out_dat_reg(s, cond, ARITH_MOV, rd, 0, rm, SHIFT_IMM_LSL(0));
580    }
581}
582
583static void tcg_out_bx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
584{
585    tcg_out32(s, (cond << 28) | 0x012fff10 | rn);
586}
587
588static void tcg_out_b_reg(TCGContext *s, ARMCond cond, TCGReg rn)
589{
590    /*
591     * Unless the C portion of QEMU is compiled as thumb, we don't need
592     * true BX semantics; merely a branch to an address held in a register.
593     */
594    tcg_out_bx_reg(s, cond, rn);
595}
596
597static void tcg_out_dat_imm(TCGContext *s, ARMCond cond, ARMInsn opc,
598                            TCGReg rd, TCGReg rn, int im)
599{
600    tcg_out32(s, (cond << 28) | (1 << 25) | opc |
601                    (rn << 16) | (rd << 12) | im);
602}
603
604static void tcg_out_ldstm(TCGContext *s, ARMCond cond, ARMInsn opc,
605                          TCGReg rn, uint16_t mask)
606{
607    tcg_out32(s, (cond << 28) | opc | (rn << 16) | mask);
608}
609
610/* Note that this routine is used for both LDR and LDRH formats, so we do
611   not wish to include an immediate shift at this point.  */
612static void tcg_out_memop_r(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
613                            TCGReg rn, TCGReg rm, bool u, bool p, bool w)
614{
615    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24)
616              | (w << 21) | (rn << 16) | (rt << 12) | rm);
617}
618
619static void tcg_out_memop_8(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
620                            TCGReg rn, int imm8, bool p, bool w)
621{
622    bool u = 1;
623    if (imm8 < 0) {
624        imm8 = -imm8;
625        u = 0;
626    }
627    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
628              (rn << 16) | (rt << 12) | ((imm8 & 0xf0) << 4) | (imm8 & 0xf));
629}
630
631static void tcg_out_memop_12(TCGContext *s, ARMCond cond, ARMInsn opc,
632                             TCGReg rt, TCGReg rn, int imm12, bool p, bool w)
633{
634    bool u = 1;
635    if (imm12 < 0) {
636        imm12 = -imm12;
637        u = 0;
638    }
639    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
640              (rn << 16) | (rt << 12) | imm12);
641}
642
643static void tcg_out_ld32_12(TCGContext *s, ARMCond cond, TCGReg rt,
644                            TCGReg rn, int imm12)
645{
646    tcg_out_memop_12(s, cond, INSN_LDR_IMM, rt, rn, imm12, 1, 0);
647}
648
649static void tcg_out_st32_12(TCGContext *s, ARMCond cond, TCGReg rt,
650                            TCGReg rn, int imm12)
651{
652    tcg_out_memop_12(s, cond, INSN_STR_IMM, rt, rn, imm12, 1, 0);
653}
654
655static void tcg_out_ld32_r(TCGContext *s, ARMCond cond, TCGReg rt,
656                           TCGReg rn, TCGReg rm)
657{
658    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 0);
659}
660
661static void tcg_out_st32_r(TCGContext *s, ARMCond cond, TCGReg rt,
662                           TCGReg rn, TCGReg rm)
663{
664    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 0);
665}
666
667static void tcg_out_ldrd_8(TCGContext *s, ARMCond cond, TCGReg rt,
668                           TCGReg rn, int imm8)
669{
670    tcg_out_memop_8(s, cond, INSN_LDRD_IMM, rt, rn, imm8, 1, 0);
671}
672
673static void tcg_out_ldrd_r(TCGContext *s, ARMCond cond, TCGReg rt,
674                           TCGReg rn, TCGReg rm)
675{
676    tcg_out_memop_r(s, cond, INSN_LDRD_REG, rt, rn, rm, 1, 1, 0);
677}
678
679static void tcg_out_strd_8(TCGContext *s, ARMCond cond, TCGReg rt,
680                           TCGReg rn, int imm8)
681{
682    tcg_out_memop_8(s, cond, INSN_STRD_IMM, rt, rn, imm8, 1, 0);
683}
684
685static void tcg_out_strd_r(TCGContext *s, ARMCond cond, TCGReg rt,
686                           TCGReg rn, TCGReg rm)
687{
688    tcg_out_memop_r(s, cond, INSN_STRD_REG, rt, rn, rm, 1, 1, 0);
689}
690
691/* Register pre-increment with base writeback.  */
692static void tcg_out_ld32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
693                             TCGReg rn, TCGReg rm)
694{
695    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 1);
696}
697
698static void tcg_out_st32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
699                             TCGReg rn, TCGReg rm)
700{
701    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 1);
702}
703
704static void tcg_out_ld16u_8(TCGContext *s, ARMCond cond, TCGReg rt,
705                            TCGReg rn, int imm8)
706{
707    tcg_out_memop_8(s, cond, INSN_LDRH_IMM, rt, rn, imm8, 1, 0);
708}
709
710static void tcg_out_st16_8(TCGContext *s, ARMCond cond, TCGReg rt,
711                           TCGReg rn, int imm8)
712{
713    tcg_out_memop_8(s, cond, INSN_STRH_IMM, rt, rn, imm8, 1, 0);
714}
715
716static void tcg_out_ld16u_r(TCGContext *s, ARMCond cond, TCGReg rt,
717                            TCGReg rn, TCGReg rm)
718{
719    tcg_out_memop_r(s, cond, INSN_LDRH_REG, rt, rn, rm, 1, 1, 0);
720}
721
722static void tcg_out_st16_r(TCGContext *s, ARMCond cond, TCGReg rt,
723                           TCGReg rn, TCGReg rm)
724{
725    tcg_out_memop_r(s, cond, INSN_STRH_REG, rt, rn, rm, 1, 1, 0);
726}
727
728static void tcg_out_ld16s_8(TCGContext *s, ARMCond cond, TCGReg rt,
729                            TCGReg rn, int imm8)
730{
731    tcg_out_memop_8(s, cond, INSN_LDRSH_IMM, rt, rn, imm8, 1, 0);
732}
733
734static void tcg_out_ld16s_r(TCGContext *s, ARMCond cond, TCGReg rt,
735                            TCGReg rn, TCGReg rm)
736{
737    tcg_out_memop_r(s, cond, INSN_LDRSH_REG, rt, rn, rm, 1, 1, 0);
738}
739
740static void tcg_out_ld8_12(TCGContext *s, ARMCond cond, TCGReg rt,
741                           TCGReg rn, int imm12)
742{
743    tcg_out_memop_12(s, cond, INSN_LDRB_IMM, rt, rn, imm12, 1, 0);
744}
745
746static void tcg_out_st8_12(TCGContext *s, ARMCond cond, TCGReg rt,
747                           TCGReg rn, int imm12)
748{
749    tcg_out_memop_12(s, cond, INSN_STRB_IMM, rt, rn, imm12, 1, 0);
750}
751
752static void tcg_out_ld8_r(TCGContext *s, ARMCond cond, TCGReg rt,
753                          TCGReg rn, TCGReg rm)
754{
755    tcg_out_memop_r(s, cond, INSN_LDRB_REG, rt, rn, rm, 1, 1, 0);
756}
757
758static void tcg_out_st8_r(TCGContext *s, ARMCond cond, TCGReg rt,
759                          TCGReg rn, TCGReg rm)
760{
761    tcg_out_memop_r(s, cond, INSN_STRB_REG, rt, rn, rm, 1, 1, 0);
762}
763
764static void tcg_out_ld8s_8(TCGContext *s, ARMCond cond, TCGReg rt,
765                           TCGReg rn, int imm8)
766{
767    tcg_out_memop_8(s, cond, INSN_LDRSB_IMM, rt, rn, imm8, 1, 0);
768}
769
770static void tcg_out_ld8s_r(TCGContext *s, ARMCond cond, TCGReg rt,
771                           TCGReg rn, TCGReg rm)
772{
773    tcg_out_memop_r(s, cond, INSN_LDRSB_REG, rt, rn, rm, 1, 1, 0);
774}
775
776static void tcg_out_movi_pool(TCGContext *s, ARMCond cond,
777                              TCGReg rd, uint32_t arg)
778{
779    new_pool_label(s, arg, R_ARM_PC13, s->code_ptr, 0);
780    tcg_out_ld32_12(s, cond, rd, TCG_REG_PC, 0);
781}
782
783static void tcg_out_movi32(TCGContext *s, ARMCond cond,
784                           TCGReg rd, uint32_t arg)
785{
786    int imm12, diff, opc, sh1, sh2;
787    uint32_t tt0, tt1, tt2;
788
789    /* Check a single MOV/MVN before anything else.  */
790    imm12 = encode_imm(arg);
791    if (imm12 >= 0) {
792        tcg_out_dat_imm(s, cond, ARITH_MOV, rd, 0, imm12);
793        return;
794    }
795    imm12 = encode_imm(~arg);
796    if (imm12 >= 0) {
797        tcg_out_dat_imm(s, cond, ARITH_MVN, rd, 0, imm12);
798        return;
799    }
800
801    /* Check for a pc-relative address.  This will usually be the TB,
802       or within the TB, which is immediately before the code block.  */
803    diff = tcg_pcrel_diff(s, (void *)arg) - 8;
804    if (diff >= 0) {
805        imm12 = encode_imm(diff);
806        if (imm12 >= 0) {
807            tcg_out_dat_imm(s, cond, ARITH_ADD, rd, TCG_REG_PC, imm12);
808            return;
809        }
810    } else {
811        imm12 = encode_imm(-diff);
812        if (imm12 >= 0) {
813            tcg_out_dat_imm(s, cond, ARITH_SUB, rd, TCG_REG_PC, imm12);
814            return;
815        }
816    }
817
818    /* Use movw + movt.  */
819    if (use_armv7_instructions) {
820        /* movw */
821        tcg_out32(s, (cond << 28) | 0x03000000 | (rd << 12)
822                  | ((arg << 4) & 0x000f0000) | (arg & 0xfff));
823        if (arg & 0xffff0000) {
824            /* movt */
825            tcg_out32(s, (cond << 28) | 0x03400000 | (rd << 12)
826                      | ((arg >> 12) & 0x000f0000) | ((arg >> 16) & 0xfff));
827        }
828        return;
829    }
830
831    /* Look for sequences of two insns.  If we have lots of 1's, we can
832       shorten the sequence by beginning with mvn and then clearing
833       higher bits with eor.  */
834    tt0 = arg;
835    opc = ARITH_MOV;
836    if (ctpop32(arg) > 16) {
837        tt0 = ~arg;
838        opc = ARITH_MVN;
839    }
840    sh1 = ctz32(tt0) & ~1;
841    tt1 = tt0 & ~(0xff << sh1);
842    sh2 = ctz32(tt1) & ~1;
843    tt2 = tt1 & ~(0xff << sh2);
844    if (tt2 == 0) {
845        int rot;
846
847        rot = ((32 - sh1) << 7) & 0xf00;
848        tcg_out_dat_imm(s, cond, opc, rd,  0, ((tt0 >> sh1) & 0xff) | rot);
849        rot = ((32 - sh2) << 7) & 0xf00;
850        tcg_out_dat_imm(s, cond, ARITH_EOR, rd, rd,
851                        ((tt0 >> sh2) & 0xff) | rot);
852        return;
853    }
854
855    /* Otherwise, drop it into the constant pool.  */
856    tcg_out_movi_pool(s, cond, rd, arg);
857}
858
859/*
860 * Emit either the reg,imm or reg,reg form of a data-processing insn.
861 * rhs must satisfy the "rI" constraint.
862 */
863static void tcg_out_dat_rI(TCGContext *s, ARMCond cond, ARMInsn opc,
864                           TCGReg dst, TCGReg lhs, TCGArg rhs, int rhs_is_const)
865{
866    if (rhs_is_const) {
867        tcg_out_dat_imm(s, cond, opc, dst, lhs, encode_imm_nofail(rhs));
868    } else {
869        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
870    }
871}
872
873/*
874 * Emit either the reg,imm or reg,reg form of a data-processing insn.
875 * rhs must satisfy the "rIK" constraint.
876 */
877static void tcg_out_dat_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    /* According to gcc, AND can be faster. */
1034    if (ofs == 0 && len <= 8) {
1035        tcg_out_dat_imm(s, cond, ARITH_AND, rd, rn,
1036                        encode_imm_nofail((1 << len) - 1));
1037        return;
1038    }
1039
1040    if (use_armv7_instructions) {
1041        /* ubfx */
1042        tcg_out32(s, 0x07e00050 | (cond << 28) | (rd << 12) | rn
1043                  | (ofs << 7) | ((len - 1) << 16));
1044        return;
1045    }
1046
1047    assert(ofs % 8 == 0);
1048    switch (len) {
1049    case 8:
1050        /* uxtb */
1051        tcg_out32(s, 0x06ef0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1052        break;
1053    case 16:
1054        /* uxth */
1055        tcg_out32(s, 0x06ff0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1056        break;
1057    default:
1058        g_assert_not_reached();
1059    }
1060}
1061
1062static void tcg_out_sextract(TCGContext *s, ARMCond cond, TCGReg rd,
1063                             TCGReg rn, int ofs, int len)
1064{
1065    if (use_armv7_instructions) {
1066        /* sbfx */
1067        tcg_out32(s, 0x07a00050 | (cond << 28) | (rd << 12) | rn
1068                  | (ofs << 7) | ((len - 1) << 16));
1069        return;
1070    }
1071
1072    assert(ofs % 8 == 0);
1073    switch (len) {
1074    case 8:
1075        /* sxtb */
1076        tcg_out32(s, 0x06af0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1077        break;
1078    case 16:
1079        /* sxth */
1080        tcg_out32(s, 0x06bf0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1081        break;
1082    default:
1083        g_assert_not_reached();
1084    }
1085}
1086
1087
1088static void tcg_out_ld32u(TCGContext *s, ARMCond cond,
1089                          TCGReg rd, TCGReg rn, int32_t offset)
1090{
1091    if (offset > 0xfff || offset < -0xfff) {
1092        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1093        tcg_out_ld32_r(s, cond, rd, rn, TCG_REG_TMP);
1094    } else
1095        tcg_out_ld32_12(s, cond, rd, rn, offset);
1096}
1097
1098static void tcg_out_st32(TCGContext *s, ARMCond cond,
1099                         TCGReg rd, TCGReg rn, int32_t offset)
1100{
1101    if (offset > 0xfff || offset < -0xfff) {
1102        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1103        tcg_out_st32_r(s, cond, rd, rn, TCG_REG_TMP);
1104    } else
1105        tcg_out_st32_12(s, cond, rd, rn, offset);
1106}
1107
1108static void tcg_out_ld16u(TCGContext *s, ARMCond cond,
1109                          TCGReg rd, TCGReg rn, int32_t offset)
1110{
1111    if (offset > 0xff || offset < -0xff) {
1112        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1113        tcg_out_ld16u_r(s, cond, rd, rn, TCG_REG_TMP);
1114    } else
1115        tcg_out_ld16u_8(s, cond, rd, rn, offset);
1116}
1117
1118static void tcg_out_ld16s(TCGContext *s, ARMCond cond,
1119                          TCGReg rd, TCGReg rn, int32_t offset)
1120{
1121    if (offset > 0xff || offset < -0xff) {
1122        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1123        tcg_out_ld16s_r(s, cond, rd, rn, TCG_REG_TMP);
1124    } else
1125        tcg_out_ld16s_8(s, cond, rd, rn, offset);
1126}
1127
1128static void tcg_out_st16(TCGContext *s, ARMCond cond,
1129                         TCGReg rd, TCGReg rn, int32_t offset)
1130{
1131    if (offset > 0xff || offset < -0xff) {
1132        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1133        tcg_out_st16_r(s, cond, rd, rn, TCG_REG_TMP);
1134    } else
1135        tcg_out_st16_8(s, cond, rd, rn, offset);
1136}
1137
1138static void tcg_out_ld8u(TCGContext *s, ARMCond cond,
1139                         TCGReg rd, TCGReg rn, int32_t offset)
1140{
1141    if (offset > 0xfff || offset < -0xfff) {
1142        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1143        tcg_out_ld8_r(s, cond, rd, rn, TCG_REG_TMP);
1144    } else
1145        tcg_out_ld8_12(s, cond, rd, rn, offset);
1146}
1147
1148static void tcg_out_ld8s(TCGContext *s, ARMCond cond,
1149                         TCGReg rd, TCGReg rn, int32_t offset)
1150{
1151    if (offset > 0xff || offset < -0xff) {
1152        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1153        tcg_out_ld8s_r(s, cond, rd, rn, TCG_REG_TMP);
1154    } else
1155        tcg_out_ld8s_8(s, cond, rd, rn, offset);
1156}
1157
1158static void tcg_out_st8(TCGContext *s, ARMCond cond,
1159                        TCGReg rd, TCGReg rn, int32_t offset)
1160{
1161    if (offset > 0xfff || offset < -0xfff) {
1162        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1163        tcg_out_st8_r(s, cond, rd, rn, TCG_REG_TMP);
1164    } else
1165        tcg_out_st8_12(s, cond, rd, rn, offset);
1166}
1167
1168/*
1169 * The _goto case is normally between TBs within the same code buffer, and
1170 * with the code buffer limited to 16MB we wouldn't need the long case.
1171 * But we also use it for the tail-call to the qemu_ld/st helpers, which does.
1172 */
1173static void tcg_out_goto(TCGContext *s, ARMCond cond, const tcg_insn_unit *addr)
1174{
1175    intptr_t addri = (intptr_t)addr;
1176    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1177    bool arm_mode = !(addri & 1);
1178
1179    if (arm_mode && disp - 8 < 0x01fffffd && disp - 8 > -0x01fffffd) {
1180        tcg_out_b_imm(s, cond, disp);
1181        return;
1182    }
1183
1184    /* LDR is interworking from v5t. */
1185    tcg_out_movi_pool(s, cond, TCG_REG_PC, addri);
1186}
1187
1188/*
1189 * The call case is mostly used for helpers - so it's not unreasonable
1190 * for them to be beyond branch range.
1191 */
1192static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *addr)
1193{
1194    intptr_t addri = (intptr_t)addr;
1195    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1196    bool arm_mode = !(addri & 1);
1197
1198    if (disp - 8 < 0x02000000 && disp - 8 >= -0x02000000) {
1199        if (arm_mode) {
1200            tcg_out_bl_imm(s, COND_AL, disp);
1201        } else {
1202            tcg_out_blx_imm(s, disp);
1203        }
1204        return;
1205    }
1206
1207    tcg_out_movi32(s, COND_AL, TCG_REG_TMP, addri);
1208    tcg_out_blx_reg(s, COND_AL, TCG_REG_TMP);
1209}
1210
1211static void tcg_out_call(TCGContext *s, const tcg_insn_unit *addr,
1212                         const TCGHelperInfo *info)
1213{
1214    tcg_out_call_int(s, addr);
1215}
1216
1217static void tcg_out_goto_label(TCGContext *s, ARMCond cond, TCGLabel *l)
1218{
1219    if (l->has_value) {
1220        tcg_out_goto(s, cond, l->u.value_ptr);
1221    } else {
1222        tcg_out_reloc(s, s->code_ptr, R_ARM_PC24, l, 0);
1223        tcg_out_b_imm(s, cond, 0);
1224    }
1225}
1226
1227static void tcg_out_mb(TCGContext *s, TCGArg a0)
1228{
1229    if (use_armv7_instructions) {
1230        tcg_out32(s, INSN_DMB_ISH);
1231    } else {
1232        tcg_out32(s, INSN_DMB_MCR);
1233    }
1234}
1235
1236static TCGCond tcg_out_cmp(TCGContext *s, TCGCond cond, TCGReg a,
1237                           TCGArg b, int b_const)
1238{
1239    if (!is_tst_cond(cond)) {
1240        tcg_out_dat_rIN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0, a, b, b_const);
1241        return cond;
1242    }
1243
1244    cond = tcg_tst_eqne_cond(cond);
1245    if (b_const) {
1246        int imm12 = encode_imm(b);
1247
1248        /*
1249         * The compare constraints allow rIN, but TST does not support N.
1250         * Be prepared to load the constant into a scratch register.
1251         */
1252        if (imm12 >= 0) {
1253            tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, a, imm12);
1254            return cond;
1255        }
1256        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, b);
1257        b = TCG_REG_TMP;
1258    }
1259    tcg_out_dat_reg(s, COND_AL, ARITH_TST, 0, a, b, SHIFT_IMM_LSL(0));
1260    return cond;
1261}
1262
1263static TCGCond tcg_out_cmp2(TCGContext *s, const TCGArg *args,
1264                            const int *const_args)
1265{
1266    TCGReg al = args[0];
1267    TCGReg ah = args[1];
1268    TCGArg bl = args[2];
1269    TCGArg bh = args[3];
1270    TCGCond cond = args[4];
1271    int const_bl = const_args[2];
1272    int const_bh = const_args[3];
1273
1274    switch (cond) {
1275    case TCG_COND_EQ:
1276    case TCG_COND_NE:
1277    case TCG_COND_LTU:
1278    case TCG_COND_LEU:
1279    case TCG_COND_GTU:
1280    case TCG_COND_GEU:
1281        /*
1282         * We perform a conditional comparison.  If the high half is
1283         * equal, then overwrite the flags with the comparison of the
1284         * low half.  The resulting flags cover the whole.
1285         */
1286        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, ah, bh, const_bh);
1287        tcg_out_dat_rI(s, COND_EQ, ARITH_CMP, 0, al, bl, const_bl);
1288        return cond;
1289
1290    case TCG_COND_TSTEQ:
1291    case TCG_COND_TSTNE:
1292        /* Similar, but with TST instead of CMP. */
1293        tcg_out_dat_rI(s, COND_AL, ARITH_TST, 0, ah, bh, const_bh);
1294        tcg_out_dat_rI(s, COND_EQ, ARITH_TST, 0, al, bl, const_bl);
1295        return tcg_tst_eqne_cond(cond);
1296
1297    case TCG_COND_LT:
1298    case TCG_COND_GE:
1299        /* We perform a double-word subtraction and examine the result.
1300           We do not actually need the result of the subtract, so the
1301           low part "subtract" is a compare.  For the high half we have
1302           no choice but to compute into a temporary.  */
1303        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, al, bl, const_bl);
1304        tcg_out_dat_rI(s, COND_AL, ARITH_SBC | TO_CPSR,
1305                       TCG_REG_TMP, ah, bh, const_bh);
1306        return cond;
1307
1308    case TCG_COND_LE:
1309    case TCG_COND_GT:
1310        /* Similar, but with swapped arguments, via reversed subtract.  */
1311        tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR,
1312                       TCG_REG_TMP, al, bl, const_bl);
1313        tcg_out_dat_rI(s, COND_AL, ARITH_RSC | TO_CPSR,
1314                       TCG_REG_TMP, ah, bh, const_bh);
1315        return tcg_swap_cond(cond);
1316
1317    default:
1318        g_assert_not_reached();
1319    }
1320}
1321
1322/*
1323 * Note that TCGReg references Q-registers.
1324 * Q-regno = 2 * D-regno, so shift left by 1 while inserting.
1325 */
1326static uint32_t encode_vd(TCGReg rd)
1327{
1328    tcg_debug_assert(rd >= TCG_REG_Q0);
1329    return (extract32(rd, 3, 1) << 22) | (extract32(rd, 0, 3) << 13);
1330}
1331
1332static uint32_t encode_vn(TCGReg rn)
1333{
1334    tcg_debug_assert(rn >= TCG_REG_Q0);
1335    return (extract32(rn, 3, 1) << 7) | (extract32(rn, 0, 3) << 17);
1336}
1337
1338static uint32_t encode_vm(TCGReg rm)
1339{
1340    tcg_debug_assert(rm >= TCG_REG_Q0);
1341    return (extract32(rm, 3, 1) << 5) | (extract32(rm, 0, 3) << 1);
1342}
1343
1344static void tcg_out_vreg2(TCGContext *s, ARMInsn insn, int q, int vece,
1345                          TCGReg d, TCGReg m)
1346{
1347    tcg_out32(s, insn | (vece << 18) | (q << 6) |
1348              encode_vd(d) | encode_vm(m));
1349}
1350
1351static void tcg_out_vreg3(TCGContext *s, ARMInsn insn, int q, int vece,
1352                          TCGReg d, TCGReg n, TCGReg m)
1353{
1354    tcg_out32(s, insn | (vece << 20) | (q << 6) |
1355              encode_vd(d) | encode_vn(n) | encode_vm(m));
1356}
1357
1358static void tcg_out_vmovi(TCGContext *s, TCGReg rd,
1359                          int q, int op, int cmode, uint8_t imm8)
1360{
1361    tcg_out32(s, INSN_VMOVI | encode_vd(rd) | (q << 6) | (op << 5)
1362              | (cmode << 8) | extract32(imm8, 0, 4)
1363              | (extract32(imm8, 4, 3) << 16)
1364              | (extract32(imm8, 7, 1) << 24));
1365}
1366
1367static void tcg_out_vshifti(TCGContext *s, ARMInsn insn, int q,
1368                            TCGReg rd, TCGReg rm, int l_imm6)
1369{
1370    tcg_out32(s, insn | (q << 6) | encode_vd(rd) | encode_vm(rm) |
1371              (extract32(l_imm6, 6, 1) << 7) |
1372              (extract32(l_imm6, 0, 6) << 16));
1373}
1374
1375static void tcg_out_vldst(TCGContext *s, ARMInsn insn,
1376                          TCGReg rd, TCGReg rn, int offset)
1377{
1378    if (offset != 0) {
1379        if (check_fit_imm(offset) || check_fit_imm(-offset)) {
1380            tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
1381                            TCG_REG_TMP, rn, offset, true);
1382        } else {
1383            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP, offset);
1384            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1385                            TCG_REG_TMP, TCG_REG_TMP, rn, 0);
1386        }
1387        rn = TCG_REG_TMP;
1388    }
1389    tcg_out32(s, insn | (rn << 16) | encode_vd(rd) | 0xf);
1390}
1391
1392typedef struct {
1393    ARMCond cond;
1394    TCGReg base;
1395    int index;
1396    bool index_scratch;
1397    TCGAtomAlign aa;
1398} HostAddress;
1399
1400bool tcg_target_has_memory_bswap(MemOp memop)
1401{
1402    return false;
1403}
1404
1405static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
1406{
1407    /* We arrive at the slow path via "BLNE", so R14 contains l->raddr. */
1408    return TCG_REG_R14;
1409}
1410
1411static const TCGLdstHelperParam ldst_helper_param = {
1412    .ra_gen = ldst_ra_gen,
1413    .ntmp = 1,
1414    .tmp = { TCG_REG_TMP },
1415};
1416
1417static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1418{
1419    MemOp opc = get_memop(lb->oi);
1420
1421    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1422        return false;
1423    }
1424
1425    tcg_out_ld_helper_args(s, lb, &ldst_helper_param);
1426    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE]);
1427    tcg_out_ld_helper_ret(s, lb, false, &ldst_helper_param);
1428
1429    tcg_out_goto(s, COND_AL, lb->raddr);
1430    return true;
1431}
1432
1433static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1434{
1435    MemOp opc = get_memop(lb->oi);
1436
1437    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1438        return false;
1439    }
1440
1441    tcg_out_st_helper_args(s, lb, &ldst_helper_param);
1442
1443    /* Tail-call to the helper, which will return to the fast path.  */
1444    tcg_out_goto(s, COND_AL, qemu_st_helpers[opc & MO_SIZE]);
1445    return true;
1446}
1447
1448/* We expect to use an 9-bit sign-magnitude negative offset from ENV.  */
1449#define MIN_TLB_MASK_TABLE_OFS  -256
1450
1451static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
1452                                           TCGReg addr, MemOpIdx oi, bool is_ld)
1453{
1454    TCGLabelQemuLdst *ldst = NULL;
1455    MemOp opc = get_memop(oi);
1456    unsigned a_mask;
1457
1458    if (tcg_use_softmmu) {
1459        *h = (HostAddress){
1460            .cond = COND_AL,
1461            .base = addr,
1462            .index = TCG_REG_R1,
1463            .index_scratch = true,
1464        };
1465    } else {
1466        *h = (HostAddress){
1467            .cond = COND_AL,
1468            .base = addr,
1469            .index = guest_base ? TCG_REG_GUEST_BASE : -1,
1470            .index_scratch = false,
1471        };
1472    }
1473
1474    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, false);
1475    a_mask = (1 << h->aa.align) - 1;
1476
1477    if (tcg_use_softmmu) {
1478        int mem_index = get_mmuidx(oi);
1479        int cmp_off = is_ld ? offsetof(CPUTLBEntry, addr_read)
1480                            : offsetof(CPUTLBEntry, addr_write);
1481        int fast_off = tlb_mask_table_ofs(s, mem_index);
1482        unsigned s_mask = (1 << (opc & MO_SIZE)) - 1;
1483        TCGReg t_addr;
1484
1485        ldst = new_ldst_label(s);
1486        ldst->is_ld = is_ld;
1487        ldst->oi = oi;
1488        ldst->addr_reg = addr;
1489
1490        /* Load cpu->neg.tlb.f[mmu_idx].{mask,table} into {r0,r1}.  */
1491        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, mask) != 0);
1492        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, table) != 4);
1493        tcg_out_ldrd_8(s, COND_AL, TCG_REG_R0, TCG_AREG0, fast_off);
1494
1495        /* Extract the tlb index from the address into R0.  */
1496        tcg_out_dat_reg(s, COND_AL, ARITH_AND, TCG_REG_R0, TCG_REG_R0, addr,
1497                        SHIFT_IMM_LSR(s->page_bits - CPU_TLB_ENTRY_BITS));
1498
1499        /*
1500         * Add the tlb_table pointer, creating the CPUTLBEntry address in R1.
1501         * Load the tlb comparator into R2 and the fast path addend into R1.
1502         */
1503        if (cmp_off == 0) {
1504            tcg_out_ld32_rwb(s, COND_AL, TCG_REG_R2, TCG_REG_R1, TCG_REG_R0);
1505        } else {
1506            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1507                            TCG_REG_R1, TCG_REG_R1, TCG_REG_R0, 0);
1508            tcg_out_ld32_12(s, COND_AL, TCG_REG_R2, TCG_REG_R1, cmp_off);
1509        }
1510
1511        /* Load the tlb addend.  */
1512        tcg_out_ld32_12(s, COND_AL, TCG_REG_R1, TCG_REG_R1,
1513                        offsetof(CPUTLBEntry, addend));
1514
1515        /*
1516         * Check alignment, check comparators.
1517         * Do this in 2-4 insns.  Use MOVW for v7, if possible,
1518         * to reduce the number of sequential conditional instructions.
1519         * Almost all guests have at least 4k pages, which means that we need
1520         * to clear at least 9 bits even for an 8-byte memory, which means it
1521         * isn't worth checking for an immediate operand for BIC.
1522         *
1523         * For unaligned accesses, test the page of the last unit of alignment.
1524         * This leaves the least significant alignment bits unchanged, and of
1525         * course must be zero.
1526         */
1527        t_addr = addr;
1528        if (a_mask < s_mask) {
1529            t_addr = TCG_REG_R0;
1530            tcg_out_dat_imm(s, COND_AL, ARITH_ADD, t_addr,
1531                            addr, s_mask - a_mask);
1532        }
1533        if (use_armv7_instructions && s->page_bits <= 16) {
1534            tcg_out_movi32(s, COND_AL, TCG_REG_TMP, ~(s->page_mask | a_mask));
1535            tcg_out_dat_reg(s, COND_AL, ARITH_BIC, TCG_REG_TMP,
1536                            t_addr, TCG_REG_TMP, 0);
1537            tcg_out_dat_reg(s, COND_AL, ARITH_CMP, 0,
1538                            TCG_REG_R2, TCG_REG_TMP, 0);
1539        } else {
1540            if (a_mask) {
1541                tcg_debug_assert(a_mask <= 0xff);
1542                tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addr, a_mask);
1543            }
1544            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0, t_addr,
1545                            SHIFT_IMM_LSR(s->page_bits));
1546            tcg_out_dat_reg(s, (a_mask ? COND_EQ : COND_AL), ARITH_CMP,
1547                            0, TCG_REG_R2, TCG_REG_TMP,
1548                            SHIFT_IMM_LSL(s->page_bits));
1549        }
1550    } else if (a_mask) {
1551        ldst = new_ldst_label(s);
1552        ldst->is_ld = is_ld;
1553        ldst->oi = oi;
1554        ldst->addr_reg = addr;
1555
1556        /* We are expecting alignment to max out at 7 */
1557        tcg_debug_assert(a_mask <= 0xff);
1558        /* tst addr, #mask */
1559        tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addr, a_mask);
1560    }
1561
1562    return ldst;
1563}
1564
1565static void tcg_out_qemu_ld_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1566                                   TCGReg datahi, HostAddress h)
1567{
1568    TCGReg base;
1569
1570    /* Byte swapping is left to middle-end expansion. */
1571    tcg_debug_assert((opc & MO_BSWAP) == 0);
1572
1573    switch (opc & MO_SSIZE) {
1574    case MO_UB:
1575        if (h.index < 0) {
1576            tcg_out_ld8_12(s, h.cond, datalo, h.base, 0);
1577        } else {
1578            tcg_out_ld8_r(s, h.cond, datalo, h.base, h.index);
1579        }
1580        break;
1581    case MO_SB:
1582        if (h.index < 0) {
1583            tcg_out_ld8s_8(s, h.cond, datalo, h.base, 0);
1584        } else {
1585            tcg_out_ld8s_r(s, h.cond, datalo, h.base, h.index);
1586        }
1587        break;
1588    case MO_UW:
1589        if (h.index < 0) {
1590            tcg_out_ld16u_8(s, h.cond, datalo, h.base, 0);
1591        } else {
1592            tcg_out_ld16u_r(s, h.cond, datalo, h.base, h.index);
1593        }
1594        break;
1595    case MO_SW:
1596        if (h.index < 0) {
1597            tcg_out_ld16s_8(s, h.cond, datalo, h.base, 0);
1598        } else {
1599            tcg_out_ld16s_r(s, h.cond, datalo, h.base, h.index);
1600        }
1601        break;
1602    case MO_UL:
1603        if (h.index < 0) {
1604            tcg_out_ld32_12(s, h.cond, datalo, h.base, 0);
1605        } else {
1606            tcg_out_ld32_r(s, h.cond, datalo, h.base, h.index);
1607        }
1608        break;
1609    case MO_UQ:
1610        /* We used pair allocation for datalo, so already should be aligned. */
1611        tcg_debug_assert((datalo & 1) == 0);
1612        tcg_debug_assert(datahi == datalo + 1);
1613        /* LDRD requires alignment; double-check that. */
1614        if (memop_alignment_bits(opc) >= MO_64) {
1615            if (h.index < 0) {
1616                tcg_out_ldrd_8(s, h.cond, datalo, h.base, 0);
1617                break;
1618            }
1619            /*
1620             * Rm (the second address op) must not overlap Rt or Rt + 1.
1621             * Since datalo is aligned, we can simplify the test via alignment.
1622             * Flip the two address arguments if that works.
1623             */
1624            if ((h.index & ~1) != datalo) {
1625                tcg_out_ldrd_r(s, h.cond, datalo, h.base, h.index);
1626                break;
1627            }
1628            if ((h.base & ~1) != datalo) {
1629                tcg_out_ldrd_r(s, h.cond, datalo, h.index, h.base);
1630                break;
1631            }
1632        }
1633        if (h.index < 0) {
1634            base = h.base;
1635            if (datalo == h.base) {
1636                tcg_out_mov_reg(s, h.cond, TCG_REG_TMP, base);
1637                base = TCG_REG_TMP;
1638            }
1639        } else if (h.index_scratch) {
1640            tcg_out_ld32_rwb(s, h.cond, datalo, h.index, h.base);
1641            tcg_out_ld32_12(s, h.cond, datahi, h.index, 4);
1642            break;
1643        } else {
1644            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1645                            h.base, h.index, SHIFT_IMM_LSL(0));
1646            base = TCG_REG_TMP;
1647        }
1648        tcg_out_ld32_12(s, h.cond, datalo, base, 0);
1649        tcg_out_ld32_12(s, h.cond, datahi, base, 4);
1650        break;
1651    default:
1652        g_assert_not_reached();
1653    }
1654}
1655
1656static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
1657                            TCGReg addr, MemOpIdx oi, TCGType data_type)
1658{
1659    MemOp opc = get_memop(oi);
1660    TCGLabelQemuLdst *ldst;
1661    HostAddress h;
1662
1663    ldst = prepare_host_addr(s, &h, addr, oi, true);
1664    if (ldst) {
1665        ldst->type = data_type;
1666        ldst->datalo_reg = datalo;
1667        ldst->datahi_reg = datahi;
1668
1669        /*
1670         * This a conditional BL only to load a pointer within this
1671         * opcode into LR for the slow path.  We will not be using
1672         * the value for a tail call.
1673         */
1674        ldst->label_ptr[0] = s->code_ptr;
1675        tcg_out_bl_imm(s, COND_NE, 0);
1676
1677        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1678        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1679    } else {
1680        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1681    }
1682}
1683
1684static void tcg_out_qemu_st_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1685                                   TCGReg datahi, HostAddress h)
1686{
1687    /* Byte swapping is left to middle-end expansion. */
1688    tcg_debug_assert((opc & MO_BSWAP) == 0);
1689
1690    switch (opc & MO_SIZE) {
1691    case MO_8:
1692        if (h.index < 0) {
1693            tcg_out_st8_12(s, h.cond, datalo, h.base, 0);
1694        } else {
1695            tcg_out_st8_r(s, h.cond, datalo, h.base, h.index);
1696        }
1697        break;
1698    case MO_16:
1699        if (h.index < 0) {
1700            tcg_out_st16_8(s, h.cond, datalo, h.base, 0);
1701        } else {
1702            tcg_out_st16_r(s, h.cond, datalo, h.base, h.index);
1703        }
1704        break;
1705    case MO_32:
1706        if (h.index < 0) {
1707            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1708        } else {
1709            tcg_out_st32_r(s, h.cond, datalo, h.base, h.index);
1710        }
1711        break;
1712    case MO_64:
1713        /* We used pair allocation for datalo, so already should be aligned. */
1714        tcg_debug_assert((datalo & 1) == 0);
1715        tcg_debug_assert(datahi == datalo + 1);
1716        /* STRD requires alignment; double-check that. */
1717        if (memop_alignment_bits(opc) >= MO_64) {
1718            if (h.index < 0) {
1719                tcg_out_strd_8(s, h.cond, datalo, h.base, 0);
1720            } else {
1721                tcg_out_strd_r(s, h.cond, datalo, h.base, h.index);
1722            }
1723        } else if (h.index < 0) {
1724            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1725            tcg_out_st32_12(s, h.cond, datahi, h.base, 4);
1726        } else if (h.index_scratch) {
1727            tcg_out_st32_rwb(s, h.cond, datalo, h.index, h.base);
1728            tcg_out_st32_12(s, h.cond, datahi, h.index, 4);
1729        } else {
1730            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1731                            h.base, h.index, SHIFT_IMM_LSL(0));
1732            tcg_out_st32_12(s, h.cond, datalo, TCG_REG_TMP, 0);
1733            tcg_out_st32_12(s, h.cond, datahi, TCG_REG_TMP, 4);
1734        }
1735        break;
1736    default:
1737        g_assert_not_reached();
1738    }
1739}
1740
1741static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
1742                            TCGReg addr, MemOpIdx oi, TCGType data_type)
1743{
1744    MemOp opc = get_memop(oi);
1745    TCGLabelQemuLdst *ldst;
1746    HostAddress h;
1747
1748    ldst = prepare_host_addr(s, &h, addr, oi, false);
1749    if (ldst) {
1750        ldst->type = data_type;
1751        ldst->datalo_reg = datalo;
1752        ldst->datahi_reg = datahi;
1753
1754        h.cond = COND_EQ;
1755        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1756
1757        /* The conditional call is last, as we're going to return here. */
1758        ldst->label_ptr[0] = s->code_ptr;
1759        tcg_out_bl_imm(s, COND_NE, 0);
1760        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1761    } else {
1762        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1763    }
1764}
1765
1766static void tcg_out_epilogue(TCGContext *s);
1767
1768static void tcg_out_exit_tb(TCGContext *s, uintptr_t arg)
1769{
1770    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, arg);
1771    tcg_out_epilogue(s);
1772}
1773
1774static void tcg_out_goto_tb(TCGContext *s, int which)
1775{
1776    uintptr_t i_addr;
1777    intptr_t i_disp;
1778
1779    /* Direct branch will be patched by tb_target_set_jmp_target. */
1780    set_jmp_insn_offset(s, which);
1781    tcg_out32(s, INSN_NOP);
1782
1783    /* When branch is out of range, fall through to indirect. */
1784    i_addr = get_jmp_target_addr(s, which);
1785    i_disp = tcg_pcrel_diff(s, (void *)i_addr) - 8;
1786    tcg_debug_assert(i_disp < 0);
1787    if (i_disp >= -0xfff) {
1788        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_PC, i_disp);
1789    } else {
1790        /*
1791         * The TB is close, but outside the 12 bits addressable by
1792         * the load.  We can extend this to 20 bits with a sub of a
1793         * shifted immediate from pc.
1794         */
1795        int h = -i_disp;
1796        int l = -(h & 0xfff);
1797
1798        h = encode_imm_nofail(h + l);
1799        tcg_out_dat_imm(s, COND_AL, ARITH_SUB, TCG_REG_R0, TCG_REG_PC, h);
1800        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_R0, l);
1801    }
1802    set_jmp_reset_offset(s, which);
1803}
1804
1805void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1806                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1807{
1808    uintptr_t addr = tb->jmp_target_addr[n];
1809    ptrdiff_t offset = addr - (jmp_rx + 8);
1810    tcg_insn_unit insn;
1811
1812    /* Either directly branch, or fall through to indirect branch. */
1813    if (offset == sextract64(offset, 0, 26)) {
1814        /* B <addr> */
1815        insn = deposit32((COND_AL << 28) | INSN_B, 0, 24, offset >> 2);
1816    } else {
1817        insn = INSN_NOP;
1818    }
1819
1820    qatomic_set((uint32_t *)jmp_rw, insn);
1821    flush_idcache_range(jmp_rx, jmp_rw, 4);
1822}
1823
1824static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
1825                       const TCGArg args[TCG_MAX_OP_ARGS],
1826                       const int const_args[TCG_MAX_OP_ARGS])
1827{
1828    TCGArg a0, a1, a2, a3, a4, a5;
1829    int c;
1830
1831    switch (opc) {
1832    case INDEX_op_goto_ptr:
1833        tcg_out_b_reg(s, COND_AL, args[0]);
1834        break;
1835    case INDEX_op_br:
1836        tcg_out_goto_label(s, COND_AL, arg_label(args[0]));
1837        break;
1838
1839    case INDEX_op_ld8u_i32:
1840        tcg_out_ld8u(s, COND_AL, args[0], args[1], args[2]);
1841        break;
1842    case INDEX_op_ld8s_i32:
1843        tcg_out_ld8s(s, COND_AL, args[0], args[1], args[2]);
1844        break;
1845    case INDEX_op_ld16u_i32:
1846        tcg_out_ld16u(s, COND_AL, args[0], args[1], args[2]);
1847        break;
1848    case INDEX_op_ld16s_i32:
1849        tcg_out_ld16s(s, COND_AL, args[0], args[1], args[2]);
1850        break;
1851    case INDEX_op_ld_i32:
1852        tcg_out_ld32u(s, COND_AL, args[0], args[1], args[2]);
1853        break;
1854    case INDEX_op_st8_i32:
1855        tcg_out_st8(s, COND_AL, args[0], args[1], args[2]);
1856        break;
1857    case INDEX_op_st16_i32:
1858        tcg_out_st16(s, COND_AL, args[0], args[1], args[2]);
1859        break;
1860    case INDEX_op_st_i32:
1861        tcg_out_st32(s, COND_AL, args[0], args[1], args[2]);
1862        break;
1863
1864    case INDEX_op_movcond_i32:
1865        /* Constraints mean that v2 is always in the same register as dest,
1866         * so we only need to do "if condition passed, move v1 to dest".
1867         */
1868        c = tcg_out_cmp(s, args[5], args[1], args[2], const_args[2]);
1869        tcg_out_dat_rIK(s, tcg_cond_to_arm_cond[c], ARITH_MOV,
1870                        ARITH_MVN, args[0], 0, args[3], const_args[3]);
1871        break;
1872    case INDEX_op_add_i32:
1873        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
1874                        args[0], args[1], args[2], const_args[2]);
1875        break;
1876    case INDEX_op_sub_i32:
1877        if (const_args[1]) {
1878            if (const_args[2]) {
1879                tcg_out_movi32(s, COND_AL, args[0], args[1] - args[2]);
1880            } else {
1881                tcg_out_dat_rI(s, COND_AL, ARITH_RSB,
1882                               args[0], args[2], args[1], 1);
1883            }
1884        } else {
1885            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB, ARITH_ADD,
1886                            args[0], args[1], args[2], const_args[2]);
1887        }
1888        break;
1889    case INDEX_op_and_i32:
1890        tcg_out_dat_rIK(s, COND_AL, ARITH_AND, ARITH_BIC,
1891                        args[0], args[1], args[2], const_args[2]);
1892        break;
1893    case INDEX_op_andc_i32:
1894        tcg_out_dat_rIK(s, COND_AL, ARITH_BIC, ARITH_AND,
1895                        args[0], args[1], args[2], const_args[2]);
1896        break;
1897    case INDEX_op_or_i32:
1898        c = ARITH_ORR;
1899        goto gen_arith;
1900    case INDEX_op_xor_i32:
1901        c = ARITH_EOR;
1902        /* Fall through.  */
1903    gen_arith:
1904        tcg_out_dat_rI(s, COND_AL, c, args[0], args[1], args[2], const_args[2]);
1905        break;
1906    case INDEX_op_add2_i32:
1907        a0 = args[0], a1 = args[1], a2 = args[2];
1908        a3 = args[3], a4 = args[4], a5 = args[5];
1909        if (a0 == a3 || (a0 == a5 && !const_args[5])) {
1910            a0 = TCG_REG_TMP;
1911        }
1912        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD | TO_CPSR, ARITH_SUB | TO_CPSR,
1913                        a0, a2, a4, const_args[4]);
1914        tcg_out_dat_rIK(s, COND_AL, ARITH_ADC, ARITH_SBC,
1915                        a1, a3, a5, const_args[5]);
1916        tcg_out_mov_reg(s, COND_AL, args[0], a0);
1917        break;
1918    case INDEX_op_sub2_i32:
1919        a0 = args[0], a1 = args[1], a2 = args[2];
1920        a3 = args[3], a4 = args[4], a5 = args[5];
1921        if ((a0 == a3 && !const_args[3]) || (a0 == a5 && !const_args[5])) {
1922            a0 = TCG_REG_TMP;
1923        }
1924        if (const_args[2]) {
1925            if (const_args[4]) {
1926                tcg_out_movi32(s, COND_AL, a0, a4);
1927                a4 = a0;
1928            }
1929            tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR, a0, a4, a2, 1);
1930        } else {
1931            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB | TO_CPSR,
1932                            ARITH_ADD | TO_CPSR, a0, a2, a4, const_args[4]);
1933        }
1934        if (const_args[3]) {
1935            if (const_args[5]) {
1936                tcg_out_movi32(s, COND_AL, a1, a5);
1937                a5 = a1;
1938            }
1939            tcg_out_dat_rI(s, COND_AL, ARITH_RSC, a1, a5, a3, 1);
1940        } else {
1941            tcg_out_dat_rIK(s, COND_AL, ARITH_SBC, ARITH_ADC,
1942                            a1, a3, a5, const_args[5]);
1943        }
1944        tcg_out_mov_reg(s, COND_AL, args[0], a0);
1945        break;
1946    case INDEX_op_neg_i32:
1947        tcg_out_dat_imm(s, COND_AL, ARITH_RSB, args[0], args[1], 0);
1948        break;
1949    case INDEX_op_not_i32:
1950        tcg_out_dat_reg(s, COND_AL,
1951                        ARITH_MVN, args[0], 0, args[1], SHIFT_IMM_LSL(0));
1952        break;
1953    case INDEX_op_mul_i32:
1954        tcg_out_mul32(s, COND_AL, args[0], args[1], args[2]);
1955        break;
1956    case INDEX_op_mulu2_i32:
1957        tcg_out_umull32(s, COND_AL, args[0], args[1], args[2], args[3]);
1958        break;
1959    case INDEX_op_muls2_i32:
1960        tcg_out_smull32(s, COND_AL, args[0], args[1], args[2], args[3]);
1961        break;
1962    /* XXX: Perhaps args[2] & 0x1f is wrong */
1963    case INDEX_op_shl_i32:
1964        c = const_args[2] ?
1965                SHIFT_IMM_LSL(args[2] & 0x1f) : SHIFT_REG_LSL(args[2]);
1966        goto gen_shift32;
1967    case INDEX_op_shr_i32:
1968        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_LSR(args[2] & 0x1f) :
1969                SHIFT_IMM_LSL(0) : SHIFT_REG_LSR(args[2]);
1970        goto gen_shift32;
1971    case INDEX_op_sar_i32:
1972        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ASR(args[2] & 0x1f) :
1973                SHIFT_IMM_LSL(0) : SHIFT_REG_ASR(args[2]);
1974        goto gen_shift32;
1975    case INDEX_op_rotr_i32:
1976        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ROR(args[2] & 0x1f) :
1977                SHIFT_IMM_LSL(0) : SHIFT_REG_ROR(args[2]);
1978        /* Fall through.  */
1979    gen_shift32:
1980        tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1], c);
1981        break;
1982
1983    case INDEX_op_rotl_i32:
1984        if (const_args[2]) {
1985            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
1986                            ((0x20 - args[2]) & 0x1f) ?
1987                            SHIFT_IMM_ROR((0x20 - args[2]) & 0x1f) :
1988                            SHIFT_IMM_LSL(0));
1989        } else {
1990            tcg_out_dat_imm(s, COND_AL, ARITH_RSB, TCG_REG_TMP, args[2], 0x20);
1991            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
1992                            SHIFT_REG_ROR(TCG_REG_TMP));
1993        }
1994        break;
1995
1996    case INDEX_op_ctz_i32:
1997        tcg_out_dat_reg(s, COND_AL, INSN_RBIT, TCG_REG_TMP, 0, args[1], 0);
1998        a1 = TCG_REG_TMP;
1999        goto do_clz;
2000
2001    case INDEX_op_clz_i32:
2002        a1 = args[1];
2003    do_clz:
2004        a0 = args[0];
2005        a2 = args[2];
2006        c = const_args[2];
2007        if (c && a2 == 32) {
2008            tcg_out_dat_reg(s, COND_AL, INSN_CLZ, a0, 0, a1, 0);
2009            break;
2010        }
2011        tcg_out_dat_imm(s, COND_AL, ARITH_CMP, 0, a1, 0);
2012        tcg_out_dat_reg(s, COND_NE, INSN_CLZ, a0, 0, a1, 0);
2013        if (c || a0 != a2) {
2014            tcg_out_dat_rIK(s, COND_EQ, ARITH_MOV, ARITH_MVN, a0, 0, a2, c);
2015        }
2016        break;
2017
2018    case INDEX_op_brcond_i32:
2019        c = tcg_out_cmp(s, args[2], args[0], args[1], const_args[1]);
2020        tcg_out_goto_label(s, tcg_cond_to_arm_cond[c], arg_label(args[3]));
2021        break;
2022    case INDEX_op_setcond_i32:
2023        c = tcg_out_cmp(s, args[3], args[1], args[2], const_args[2]);
2024        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c],
2025                        ARITH_MOV, args[0], 0, 1);
2026        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2027                        ARITH_MOV, args[0], 0, 0);
2028        break;
2029    case INDEX_op_negsetcond_i32:
2030        c = tcg_out_cmp(s, args[3], args[1], args[2], const_args[2]);
2031        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c],
2032                        ARITH_MVN, args[0], 0, 0);
2033        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2034                        ARITH_MOV, args[0], 0, 0);
2035        break;
2036
2037    case INDEX_op_brcond2_i32:
2038        c = tcg_out_cmp2(s, args, const_args);
2039        tcg_out_goto_label(s, tcg_cond_to_arm_cond[c], arg_label(args[5]));
2040        break;
2041    case INDEX_op_setcond2_i32:
2042        c = tcg_out_cmp2(s, args + 1, const_args + 1);
2043        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c], ARITH_MOV, args[0], 0, 1);
2044        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2045                        ARITH_MOV, args[0], 0, 0);
2046        break;
2047
2048    case INDEX_op_qemu_ld_i32:
2049        tcg_out_qemu_ld(s, args[0], -1, args[1], args[2], TCG_TYPE_I32);
2050        break;
2051    case INDEX_op_qemu_ld_i64:
2052        tcg_out_qemu_ld(s, args[0], args[1], args[2], args[3], TCG_TYPE_I64);
2053        break;
2054
2055    case INDEX_op_qemu_st_i32:
2056        tcg_out_qemu_st(s, args[0], -1, args[1], args[2], TCG_TYPE_I32);
2057        break;
2058    case INDEX_op_qemu_st_i64:
2059        tcg_out_qemu_st(s, args[0], args[1], args[2], args[3], TCG_TYPE_I64);
2060        break;
2061
2062    case INDEX_op_bswap16_i32:
2063        tcg_out_bswap16(s, COND_AL, args[0], args[1], args[2]);
2064        break;
2065    case INDEX_op_bswap32_i32:
2066        tcg_out_bswap32(s, COND_AL, args[0], args[1]);
2067        break;
2068
2069    case INDEX_op_deposit_i32:
2070        tcg_out_deposit(s, COND_AL, args[0], args[2],
2071                        args[3], args[4], const_args[2]);
2072        break;
2073    case INDEX_op_extract_i32:
2074        tcg_out_extract(s, COND_AL, args[0], args[1], args[2], args[3]);
2075        break;
2076    case INDEX_op_sextract_i32:
2077        tcg_out_sextract(s, COND_AL, args[0], args[1], args[2], args[3]);
2078        break;
2079    case INDEX_op_extract2_i32:
2080        /* ??? These optimization vs zero should be generic.  */
2081        /* ??? But we can't substitute 2 for 1 in the opcode stream yet.  */
2082        if (const_args[1]) {
2083            if (const_args[2]) {
2084                tcg_out_movi(s, TCG_TYPE_REG, args[0], 0);
2085            } else {
2086                tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2087                                args[2], SHIFT_IMM_LSL(32 - args[3]));
2088            }
2089        } else if (const_args[2]) {
2090            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2091                            args[1], SHIFT_IMM_LSR(args[3]));
2092        } else {
2093            /* We can do extract2 in 2 insns, vs the 3 required otherwise.  */
2094            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0,
2095                            args[2], SHIFT_IMM_LSL(32 - args[3]));
2096            tcg_out_dat_reg(s, COND_AL, ARITH_ORR, args[0], TCG_REG_TMP,
2097                            args[1], SHIFT_IMM_LSR(args[3]));
2098        }
2099        break;
2100
2101    case INDEX_op_div_i32:
2102        tcg_out_sdiv(s, COND_AL, args[0], args[1], args[2]);
2103        break;
2104    case INDEX_op_divu_i32:
2105        tcg_out_udiv(s, COND_AL, args[0], args[1], args[2]);
2106        break;
2107
2108    case INDEX_op_mb:
2109        tcg_out_mb(s, args[0]);
2110        break;
2111
2112    case INDEX_op_mov_i32:  /* Always emitted via tcg_out_mov.  */
2113    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
2114    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
2115    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
2116    case INDEX_op_ext8s_i32:  /* Always emitted via tcg_reg_alloc_op.  */
2117    case INDEX_op_ext8u_i32:
2118    case INDEX_op_ext16s_i32:
2119    case INDEX_op_ext16u_i32:
2120    default:
2121        g_assert_not_reached();
2122    }
2123}
2124
2125static TCGConstraintSetIndex
2126tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
2127{
2128    switch (op) {
2129    case INDEX_op_goto_ptr:
2130        return C_O0_I1(r);
2131
2132    case INDEX_op_ld8u_i32:
2133    case INDEX_op_ld8s_i32:
2134    case INDEX_op_ld16u_i32:
2135    case INDEX_op_ld16s_i32:
2136    case INDEX_op_ld_i32:
2137    case INDEX_op_neg_i32:
2138    case INDEX_op_not_i32:
2139    case INDEX_op_bswap16_i32:
2140    case INDEX_op_bswap32_i32:
2141    case INDEX_op_ext8s_i32:
2142    case INDEX_op_ext16s_i32:
2143    case INDEX_op_ext16u_i32:
2144    case INDEX_op_extract_i32:
2145    case INDEX_op_sextract_i32:
2146        return C_O1_I1(r, r);
2147
2148    case INDEX_op_st8_i32:
2149    case INDEX_op_st16_i32:
2150    case INDEX_op_st_i32:
2151        return C_O0_I2(r, r);
2152
2153    case INDEX_op_add_i32:
2154    case INDEX_op_sub_i32:
2155    case INDEX_op_setcond_i32:
2156    case INDEX_op_negsetcond_i32:
2157        return C_O1_I2(r, r, rIN);
2158
2159    case INDEX_op_and_i32:
2160    case INDEX_op_andc_i32:
2161    case INDEX_op_clz_i32:
2162    case INDEX_op_ctz_i32:
2163        return C_O1_I2(r, r, rIK);
2164
2165    case INDEX_op_mul_i32:
2166    case INDEX_op_div_i32:
2167    case INDEX_op_divu_i32:
2168        return C_O1_I2(r, r, r);
2169
2170    case INDEX_op_mulu2_i32:
2171    case INDEX_op_muls2_i32:
2172        return C_O2_I2(r, r, r, r);
2173
2174    case INDEX_op_or_i32:
2175    case INDEX_op_xor_i32:
2176        return C_O1_I2(r, r, rI);
2177
2178    case INDEX_op_shl_i32:
2179    case INDEX_op_shr_i32:
2180    case INDEX_op_sar_i32:
2181    case INDEX_op_rotl_i32:
2182    case INDEX_op_rotr_i32:
2183        return C_O1_I2(r, r, ri);
2184
2185    case INDEX_op_brcond_i32:
2186        return C_O0_I2(r, rIN);
2187    case INDEX_op_deposit_i32:
2188        return C_O1_I2(r, 0, rZ);
2189    case INDEX_op_extract2_i32:
2190        return C_O1_I2(r, rZ, rZ);
2191    case INDEX_op_movcond_i32:
2192        return C_O1_I4(r, r, rIN, rIK, 0);
2193    case INDEX_op_add2_i32:
2194        return C_O2_I4(r, r, r, r, rIN, rIK);
2195    case INDEX_op_sub2_i32:
2196        return C_O2_I4(r, r, rI, rI, rIN, rIK);
2197    case INDEX_op_brcond2_i32:
2198        return C_O0_I4(r, r, rI, rI);
2199    case INDEX_op_setcond2_i32:
2200        return C_O1_I4(r, r, r, rI, rI);
2201
2202    case INDEX_op_qemu_ld_i32:
2203        return C_O1_I1(r, q);
2204    case INDEX_op_qemu_ld_i64:
2205        return C_O2_I1(e, p, q);
2206    case INDEX_op_qemu_st_i32:
2207        return C_O0_I2(q, q);
2208    case INDEX_op_qemu_st_i64:
2209        return C_O0_I3(Q, p, q);
2210
2211    case INDEX_op_st_vec:
2212        return C_O0_I2(w, r);
2213    case INDEX_op_ld_vec:
2214    case INDEX_op_dupm_vec:
2215        return C_O1_I1(w, r);
2216    case INDEX_op_dup_vec:
2217        return C_O1_I1(w, wr);
2218    case INDEX_op_abs_vec:
2219    case INDEX_op_neg_vec:
2220    case INDEX_op_not_vec:
2221    case INDEX_op_shli_vec:
2222    case INDEX_op_shri_vec:
2223    case INDEX_op_sari_vec:
2224        return C_O1_I1(w, w);
2225    case INDEX_op_dup2_vec:
2226    case INDEX_op_add_vec:
2227    case INDEX_op_mul_vec:
2228    case INDEX_op_smax_vec:
2229    case INDEX_op_smin_vec:
2230    case INDEX_op_ssadd_vec:
2231    case INDEX_op_sssub_vec:
2232    case INDEX_op_sub_vec:
2233    case INDEX_op_umax_vec:
2234    case INDEX_op_umin_vec:
2235    case INDEX_op_usadd_vec:
2236    case INDEX_op_ussub_vec:
2237    case INDEX_op_xor_vec:
2238    case INDEX_op_arm_sshl_vec:
2239    case INDEX_op_arm_ushl_vec:
2240        return C_O1_I2(w, w, w);
2241    case INDEX_op_arm_sli_vec:
2242        return C_O1_I2(w, 0, w);
2243    case INDEX_op_or_vec:
2244    case INDEX_op_andc_vec:
2245        return C_O1_I2(w, w, wO);
2246    case INDEX_op_and_vec:
2247    case INDEX_op_orc_vec:
2248        return C_O1_I2(w, w, wV);
2249    case INDEX_op_cmp_vec:
2250        return C_O1_I2(w, w, wZ);
2251    case INDEX_op_bitsel_vec:
2252        return C_O1_I3(w, w, w, w);
2253    default:
2254        return C_NotImplemented;
2255    }
2256}
2257
2258static void tcg_target_init(TCGContext *s)
2259{
2260    /*
2261     * Only probe for the platform and capabilities if we haven't already
2262     * determined maximum values at compile time.
2263     */
2264#if !defined(use_idiv_instructions) || !defined(use_neon_instructions)
2265    {
2266        unsigned long hwcap = qemu_getauxval(AT_HWCAP);
2267#ifndef use_idiv_instructions
2268        use_idiv_instructions = (hwcap & HWCAP_ARM_IDIVA) != 0;
2269#endif
2270#ifndef use_neon_instructions
2271        use_neon_instructions = (hwcap & HWCAP_ARM_NEON) != 0;
2272#endif
2273    }
2274#endif
2275
2276    if (__ARM_ARCH < 7) {
2277        const char *pl = (const char *)qemu_getauxval(AT_PLATFORM);
2278        if (pl != NULL && pl[0] == 'v' && pl[1] >= '4' && pl[1] <= '9') {
2279            arm_arch = pl[1] - '0';
2280        }
2281
2282        if (arm_arch < 6) {
2283            error_report("TCG: ARMv%d is unsupported; exiting", arm_arch);
2284            exit(EXIT_FAILURE);
2285        }
2286    }
2287
2288    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
2289
2290    tcg_target_call_clobber_regs = 0;
2291    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R0);
2292    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R1);
2293    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R2);
2294    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R3);
2295    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R12);
2296    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R14);
2297
2298    if (use_neon_instructions) {
2299        tcg_target_available_regs[TCG_TYPE_V64]  = ALL_VECTOR_REGS;
2300        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
2301
2302        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q0);
2303        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q1);
2304        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q2);
2305        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q3);
2306        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q8);
2307        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q9);
2308        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q10);
2309        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q11);
2310        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q12);
2311        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q13);
2312        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q14);
2313        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q15);
2314    }
2315
2316    s->reserved_regs = 0;
2317    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
2318    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP);
2319    tcg_regset_set_reg(s->reserved_regs, TCG_REG_PC);
2320    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP);
2321}
2322
2323static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
2324                       TCGReg arg1, intptr_t arg2)
2325{
2326    switch (type) {
2327    case TCG_TYPE_I32:
2328        tcg_out_ld32u(s, COND_AL, arg, arg1, arg2);
2329        return;
2330    case TCG_TYPE_V64:
2331        /* regs 1; size 8; align 8 */
2332        tcg_out_vldst(s, INSN_VLD1 | 0x7d0, arg, arg1, arg2);
2333        return;
2334    case TCG_TYPE_V128:
2335        /*
2336         * We have only 8-byte alignment for the stack per the ABI.
2337         * Rather than dynamically re-align the stack, it's easier
2338         * to simply not request alignment beyond that.  So:
2339         * regs 2; size 8; align 8
2340         */
2341        tcg_out_vldst(s, INSN_VLD1 | 0xad0, arg, arg1, arg2);
2342        return;
2343    default:
2344        g_assert_not_reached();
2345    }
2346}
2347
2348static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
2349                       TCGReg arg1, intptr_t arg2)
2350{
2351    switch (type) {
2352    case TCG_TYPE_I32:
2353        tcg_out_st32(s, COND_AL, arg, arg1, arg2);
2354        return;
2355    case TCG_TYPE_V64:
2356        /* regs 1; size 8; align 8 */
2357        tcg_out_vldst(s, INSN_VST1 | 0x7d0, arg, arg1, arg2);
2358        return;
2359    case TCG_TYPE_V128:
2360        /* See tcg_out_ld re alignment: regs 2; size 8; align 8 */
2361        tcg_out_vldst(s, INSN_VST1 | 0xad0, arg, arg1, arg2);
2362        return;
2363    default:
2364        g_assert_not_reached();
2365    }
2366}
2367
2368static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
2369                        TCGReg base, intptr_t ofs)
2370{
2371    return false;
2372}
2373
2374static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
2375{
2376    if (ret == arg) {
2377        return true;
2378    }
2379    switch (type) {
2380    case TCG_TYPE_I32:
2381        if (ret < TCG_REG_Q0 && arg < TCG_REG_Q0) {
2382            tcg_out_mov_reg(s, COND_AL, ret, arg);
2383            return true;
2384        }
2385        return false;
2386
2387    case TCG_TYPE_V64:
2388    case TCG_TYPE_V128:
2389        /* "VMOV D,N" is an alias for "VORR D,N,N". */
2390        tcg_out_vreg3(s, INSN_VORR, type - TCG_TYPE_V64, 0, ret, arg, arg);
2391        return true;
2392
2393    default:
2394        g_assert_not_reached();
2395    }
2396}
2397
2398static void tcg_out_movi(TCGContext *s, TCGType type,
2399                         TCGReg ret, tcg_target_long arg)
2400{
2401    tcg_debug_assert(type == TCG_TYPE_I32);
2402    tcg_debug_assert(ret < TCG_REG_Q0);
2403    tcg_out_movi32(s, COND_AL, ret, arg);
2404}
2405
2406static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
2407{
2408    return false;
2409}
2410
2411static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
2412                             tcg_target_long imm)
2413{
2414    int enc, opc = ARITH_ADD;
2415
2416    /* All of the easiest immediates to encode are positive. */
2417    if (imm < 0) {
2418        imm = -imm;
2419        opc = ARITH_SUB;
2420    }
2421    enc = encode_imm(imm);
2422    if (enc >= 0) {
2423        tcg_out_dat_imm(s, COND_AL, opc, rd, rs, enc);
2424    } else {
2425        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, imm);
2426        tcg_out_dat_reg(s, COND_AL, opc, rd, rs,
2427                        TCG_REG_TMP, SHIFT_IMM_LSL(0));
2428    }
2429}
2430
2431/* Type is always V128, with I64 elements.  */
2432static void tcg_out_dup2_vec(TCGContext *s, TCGReg rd, TCGReg rl, TCGReg rh)
2433{
2434    /* Move high element into place first. */
2435    /* VMOV Dd+1, Ds */
2436    tcg_out_vreg3(s, INSN_VORR | (1 << 12), 0, 0, rd, rh, rh);
2437    /* Move low element into place; tcg_out_mov will check for nop. */
2438    tcg_out_mov(s, TCG_TYPE_V64, rd, rl);
2439}
2440
2441static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
2442                            TCGReg rd, TCGReg rs)
2443{
2444    int q = type - TCG_TYPE_V64;
2445
2446    if (vece == MO_64) {
2447        if (type == TCG_TYPE_V128) {
2448            tcg_out_dup2_vec(s, rd, rs, rs);
2449        } else {
2450            tcg_out_mov(s, TCG_TYPE_V64, rd, rs);
2451        }
2452    } else if (rs < TCG_REG_Q0) {
2453        int b = (vece == MO_8);
2454        int e = (vece == MO_16);
2455        tcg_out32(s, INSN_VDUP_G | (b << 22) | (q << 21) | (e << 5) |
2456                  encode_vn(rd) | (rs << 12));
2457    } else {
2458        int imm4 = 1 << vece;
2459        tcg_out32(s, INSN_VDUP_S | (imm4 << 16) | (q << 6) |
2460                  encode_vd(rd) | encode_vm(rs));
2461    }
2462    return true;
2463}
2464
2465static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
2466                             TCGReg rd, TCGReg base, intptr_t offset)
2467{
2468    if (vece == MO_64) {
2469        tcg_out_ld(s, TCG_TYPE_V64, rd, base, offset);
2470        if (type == TCG_TYPE_V128) {
2471            tcg_out_dup2_vec(s, rd, rd, rd);
2472        }
2473    } else {
2474        int q = type - TCG_TYPE_V64;
2475        tcg_out_vldst(s, INSN_VLD1R | (vece << 6) | (q << 5),
2476                      rd, base, offset);
2477    }
2478    return true;
2479}
2480
2481static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
2482                             TCGReg rd, int64_t v64)
2483{
2484    int q = type - TCG_TYPE_V64;
2485    int cmode, imm8, i;
2486
2487    /* Test all bytes equal first.  */
2488    if (vece == MO_8) {
2489        tcg_out_vmovi(s, rd, q, 0, 0xe, v64);
2490        return;
2491    }
2492
2493    /*
2494     * Test all bytes 0x00 or 0xff second.  This can match cases that
2495     * might otherwise take 2 or 3 insns for MO_16 or MO_32 below.
2496     */
2497    for (i = imm8 = 0; i < 8; i++) {
2498        uint8_t byte = v64 >> (i * 8);
2499        if (byte == 0xff) {
2500            imm8 |= 1 << i;
2501        } else if (byte != 0) {
2502            goto fail_bytes;
2503        }
2504    }
2505    tcg_out_vmovi(s, rd, q, 1, 0xe, imm8);
2506    return;
2507 fail_bytes:
2508
2509    /*
2510     * Tests for various replications.  For each element width, if we
2511     * cannot find an expansion there's no point checking a larger
2512     * width because we already know by replication it cannot match.
2513     */
2514    if (vece == MO_16) {
2515        uint16_t v16 = v64;
2516
2517        if (is_shimm16(v16, &cmode, &imm8)) {
2518            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2519            return;
2520        }
2521        if (is_shimm16(~v16, &cmode, &imm8)) {
2522            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2523            return;
2524        }
2525
2526        /*
2527         * Otherwise, all remaining constants can be loaded in two insns:
2528         * rd = v16 & 0xff, rd |= v16 & 0xff00.
2529         */
2530        tcg_out_vmovi(s, rd, q, 0, 0x8, v16 & 0xff);
2531        tcg_out_vmovi(s, rd, q, 0, 0xb, v16 >> 8);   /* VORRI */
2532        return;
2533    }
2534
2535    if (vece == MO_32) {
2536        uint32_t v32 = v64;
2537
2538        if (is_shimm32(v32, &cmode, &imm8) ||
2539            is_soimm32(v32, &cmode, &imm8)) {
2540            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2541            return;
2542        }
2543        if (is_shimm32(~v32, &cmode, &imm8) ||
2544            is_soimm32(~v32, &cmode, &imm8)) {
2545            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2546            return;
2547        }
2548
2549        /*
2550         * Restrict the set of constants to those we can load with
2551         * two instructions.  Others we load from the pool.
2552         */
2553        i = is_shimm32_pair(v32, &cmode, &imm8);
2554        if (i) {
2555            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2556            tcg_out_vmovi(s, rd, q, 0, i | 1, extract32(v32, i * 4, 8));
2557            return;
2558        }
2559        i = is_shimm32_pair(~v32, &cmode, &imm8);
2560        if (i) {
2561            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2562            tcg_out_vmovi(s, rd, q, 1, i | 1, extract32(~v32, i * 4, 8));
2563            return;
2564        }
2565    }
2566
2567    /*
2568     * As a last resort, load from the constant pool.
2569     */
2570    if (!q || vece == MO_64) {
2571        new_pool_l2(s, R_ARM_PC11, s->code_ptr, 0, v64, v64 >> 32);
2572        /* VLDR Dd, [pc + offset] */
2573        tcg_out32(s, INSN_VLDR_D | encode_vd(rd) | (0xf << 16));
2574        if (q) {
2575            tcg_out_dup2_vec(s, rd, rd, rd);
2576        }
2577    } else {
2578        new_pool_label(s, (uint32_t)v64, R_ARM_PC8, s->code_ptr, 0);
2579        /* add tmp, pc, offset */
2580        tcg_out_dat_imm(s, COND_AL, ARITH_ADD, TCG_REG_TMP, TCG_REG_PC, 0);
2581        tcg_out_dupm_vec(s, type, MO_32, rd, TCG_REG_TMP, 0);
2582    }
2583}
2584
2585static const ARMInsn vec_cmp_insn[16] = {
2586    [TCG_COND_EQ] = INSN_VCEQ,
2587    [TCG_COND_GT] = INSN_VCGT,
2588    [TCG_COND_GE] = INSN_VCGE,
2589    [TCG_COND_GTU] = INSN_VCGT_U,
2590    [TCG_COND_GEU] = INSN_VCGE_U,
2591};
2592
2593static const ARMInsn vec_cmp0_insn[16] = {
2594    [TCG_COND_EQ] = INSN_VCEQ0,
2595    [TCG_COND_GT] = INSN_VCGT0,
2596    [TCG_COND_GE] = INSN_VCGE0,
2597    [TCG_COND_LT] = INSN_VCLT0,
2598    [TCG_COND_LE] = INSN_VCLE0,
2599};
2600
2601static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2602                           unsigned vecl, unsigned vece,
2603                           const TCGArg args[TCG_MAX_OP_ARGS],
2604                           const int const_args[TCG_MAX_OP_ARGS])
2605{
2606    TCGType type = vecl + TCG_TYPE_V64;
2607    unsigned q = vecl;
2608    TCGArg a0, a1, a2, a3;
2609    int cmode, imm8;
2610
2611    a0 = args[0];
2612    a1 = args[1];
2613    a2 = args[2];
2614
2615    switch (opc) {
2616    case INDEX_op_ld_vec:
2617        tcg_out_ld(s, type, a0, a1, a2);
2618        return;
2619    case INDEX_op_st_vec:
2620        tcg_out_st(s, type, a0, a1, a2);
2621        return;
2622    case INDEX_op_dupm_vec:
2623        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
2624        return;
2625    case INDEX_op_dup2_vec:
2626        tcg_out_dup2_vec(s, a0, a1, a2);
2627        return;
2628    case INDEX_op_abs_vec:
2629        tcg_out_vreg2(s, INSN_VABS, q, vece, a0, a1);
2630        return;
2631    case INDEX_op_neg_vec:
2632        tcg_out_vreg2(s, INSN_VNEG, q, vece, a0, a1);
2633        return;
2634    case INDEX_op_not_vec:
2635        tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a1);
2636        return;
2637    case INDEX_op_add_vec:
2638        tcg_out_vreg3(s, INSN_VADD, q, vece, a0, a1, a2);
2639        return;
2640    case INDEX_op_mul_vec:
2641        tcg_out_vreg3(s, INSN_VMUL, q, vece, a0, a1, a2);
2642        return;
2643    case INDEX_op_smax_vec:
2644        tcg_out_vreg3(s, INSN_VMAX, q, vece, a0, a1, a2);
2645        return;
2646    case INDEX_op_smin_vec:
2647        tcg_out_vreg3(s, INSN_VMIN, q, vece, a0, a1, a2);
2648        return;
2649    case INDEX_op_sub_vec:
2650        tcg_out_vreg3(s, INSN_VSUB, q, vece, a0, a1, a2);
2651        return;
2652    case INDEX_op_ssadd_vec:
2653        tcg_out_vreg3(s, INSN_VQADD, q, vece, a0, a1, a2);
2654        return;
2655    case INDEX_op_sssub_vec:
2656        tcg_out_vreg3(s, INSN_VQSUB, q, vece, a0, a1, a2);
2657        return;
2658    case INDEX_op_umax_vec:
2659        tcg_out_vreg3(s, INSN_VMAX_U, q, vece, a0, a1, a2);
2660        return;
2661    case INDEX_op_umin_vec:
2662        tcg_out_vreg3(s, INSN_VMIN_U, q, vece, a0, a1, a2);
2663        return;
2664    case INDEX_op_usadd_vec:
2665        tcg_out_vreg3(s, INSN_VQADD_U, q, vece, a0, a1, a2);
2666        return;
2667    case INDEX_op_ussub_vec:
2668        tcg_out_vreg3(s, INSN_VQSUB_U, q, vece, a0, a1, a2);
2669        return;
2670    case INDEX_op_xor_vec:
2671        tcg_out_vreg3(s, INSN_VEOR, q, 0, a0, a1, a2);
2672        return;
2673    case INDEX_op_arm_sshl_vec:
2674        /*
2675         * Note that Vm is the data and Vn is the shift count,
2676         * therefore the arguments appear reversed.
2677         */
2678        tcg_out_vreg3(s, INSN_VSHL_S, q, vece, a0, a2, a1);
2679        return;
2680    case INDEX_op_arm_ushl_vec:
2681        /* See above. */
2682        tcg_out_vreg3(s, INSN_VSHL_U, q, vece, a0, a2, a1);
2683        return;
2684    case INDEX_op_shli_vec:
2685        tcg_out_vshifti(s, INSN_VSHLI, q, a0, a1, a2 + (8 << vece));
2686        return;
2687    case INDEX_op_shri_vec:
2688        tcg_out_vshifti(s, INSN_VSHRI, q, a0, a1, (16 << vece) - a2);
2689        return;
2690    case INDEX_op_sari_vec:
2691        tcg_out_vshifti(s, INSN_VSARI, q, a0, a1, (16 << vece) - a2);
2692        return;
2693    case INDEX_op_arm_sli_vec:
2694        tcg_out_vshifti(s, INSN_VSLI, q, a0, a2, args[3] + (8 << vece));
2695        return;
2696
2697    case INDEX_op_andc_vec:
2698        if (!const_args[2]) {
2699            tcg_out_vreg3(s, INSN_VBIC, q, 0, a0, a1, a2);
2700            return;
2701        }
2702        a2 = ~a2;
2703        /* fall through */
2704    case INDEX_op_and_vec:
2705        if (const_args[2]) {
2706            is_shimm1632(~a2, &cmode, &imm8);
2707            if (a0 == a1) {
2708                tcg_out_vmovi(s, a0, q, 1, cmode | 1, imm8); /* VBICI */
2709                return;
2710            }
2711            tcg_out_vmovi(s, a0, q, 1, cmode, imm8); /* VMVNI */
2712            a2 = a0;
2713        }
2714        tcg_out_vreg3(s, INSN_VAND, q, 0, a0, a1, a2);
2715        return;
2716
2717    case INDEX_op_orc_vec:
2718        if (!const_args[2]) {
2719            tcg_out_vreg3(s, INSN_VORN, q, 0, a0, a1, a2);
2720            return;
2721        }
2722        a2 = ~a2;
2723        /* fall through */
2724    case INDEX_op_or_vec:
2725        if (const_args[2]) {
2726            is_shimm1632(a2, &cmode, &imm8);
2727            if (a0 == a1) {
2728                tcg_out_vmovi(s, a0, q, 0, cmode | 1, imm8); /* VORRI */
2729                return;
2730            }
2731            tcg_out_vmovi(s, a0, q, 0, cmode, imm8); /* VMOVI */
2732            a2 = a0;
2733        }
2734        tcg_out_vreg3(s, INSN_VORR, q, 0, a0, a1, a2);
2735        return;
2736
2737    case INDEX_op_cmp_vec:
2738        {
2739            TCGCond cond = args[3];
2740            ARMInsn insn;
2741
2742            switch (cond) {
2743            case TCG_COND_NE:
2744                if (const_args[2]) {
2745                    tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a1);
2746                } else {
2747                    tcg_out_vreg3(s, INSN_VCEQ, q, vece, a0, a1, a2);
2748                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
2749                }
2750                break;
2751
2752            case TCG_COND_TSTNE:
2753            case TCG_COND_TSTEQ:
2754                if (const_args[2]) {
2755                    /* (x & 0) == 0 */
2756                    tcg_out_dupi_vec(s, type, MO_8, a0,
2757                                     -(cond == TCG_COND_TSTEQ));
2758                    break;
2759                }
2760                tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a2);
2761                if (cond == TCG_COND_TSTEQ) {
2762                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
2763                }
2764                break;
2765
2766            default:
2767                if (const_args[2]) {
2768                    insn = vec_cmp0_insn[cond];
2769                    if (insn) {
2770                        tcg_out_vreg2(s, insn, q, vece, a0, a1);
2771                        return;
2772                    }
2773                    tcg_out_dupi_vec(s, type, MO_8, TCG_VEC_TMP, 0);
2774                    a2 = TCG_VEC_TMP;
2775                }
2776                insn = vec_cmp_insn[cond];
2777                if (insn == 0) {
2778                    TCGArg t;
2779                    t = a1, a1 = a2, a2 = t;
2780                    cond = tcg_swap_cond(cond);
2781                    insn = vec_cmp_insn[cond];
2782                    tcg_debug_assert(insn != 0);
2783                }
2784                tcg_out_vreg3(s, insn, q, vece, a0, a1, a2);
2785                break;
2786            }
2787        }
2788        return;
2789
2790    case INDEX_op_bitsel_vec:
2791        a3 = args[3];
2792        if (a0 == a3) {
2793            tcg_out_vreg3(s, INSN_VBIT, q, 0, a0, a2, a1);
2794        } else if (a0 == a2) {
2795            tcg_out_vreg3(s, INSN_VBIF, q, 0, a0, a3, a1);
2796        } else {
2797            tcg_out_mov(s, type, a0, a1);
2798            tcg_out_vreg3(s, INSN_VBSL, q, 0, a0, a2, a3);
2799        }
2800        return;
2801
2802    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
2803    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
2804    default:
2805        g_assert_not_reached();
2806    }
2807}
2808
2809int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
2810{
2811    switch (opc) {
2812    case INDEX_op_add_vec:
2813    case INDEX_op_sub_vec:
2814    case INDEX_op_and_vec:
2815    case INDEX_op_andc_vec:
2816    case INDEX_op_or_vec:
2817    case INDEX_op_orc_vec:
2818    case INDEX_op_xor_vec:
2819    case INDEX_op_not_vec:
2820    case INDEX_op_shli_vec:
2821    case INDEX_op_shri_vec:
2822    case INDEX_op_sari_vec:
2823    case INDEX_op_ssadd_vec:
2824    case INDEX_op_sssub_vec:
2825    case INDEX_op_usadd_vec:
2826    case INDEX_op_ussub_vec:
2827    case INDEX_op_bitsel_vec:
2828        return 1;
2829    case INDEX_op_abs_vec:
2830    case INDEX_op_cmp_vec:
2831    case INDEX_op_mul_vec:
2832    case INDEX_op_neg_vec:
2833    case INDEX_op_smax_vec:
2834    case INDEX_op_smin_vec:
2835    case INDEX_op_umax_vec:
2836    case INDEX_op_umin_vec:
2837        return vece < MO_64;
2838    case INDEX_op_shlv_vec:
2839    case INDEX_op_shrv_vec:
2840    case INDEX_op_sarv_vec:
2841    case INDEX_op_rotli_vec:
2842    case INDEX_op_rotlv_vec:
2843    case INDEX_op_rotrv_vec:
2844        return -1;
2845    default:
2846        return 0;
2847    }
2848}
2849
2850void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
2851                       TCGArg a0, ...)
2852{
2853    va_list va;
2854    TCGv_vec v0, v1, v2, t1, t2, c1;
2855    TCGArg a2;
2856
2857    va_start(va, a0);
2858    v0 = temp_tcgv_vec(arg_temp(a0));
2859    v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
2860    a2 = va_arg(va, TCGArg);
2861    va_end(va);
2862
2863    switch (opc) {
2864    case INDEX_op_shlv_vec:
2865        /*
2866         * Merely propagate shlv_vec to arm_ushl_vec.
2867         * In this way we don't set TCG_TARGET_HAS_shv_vec
2868         * because everything is done via expansion.
2869         */
2870        v2 = temp_tcgv_vec(arg_temp(a2));
2871        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
2872                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
2873        break;
2874
2875    case INDEX_op_shrv_vec:
2876    case INDEX_op_sarv_vec:
2877        /* Right shifts are negative left shifts for NEON.  */
2878        v2 = temp_tcgv_vec(arg_temp(a2));
2879        t1 = tcg_temp_new_vec(type);
2880        tcg_gen_neg_vec(vece, t1, v2);
2881        if (opc == INDEX_op_shrv_vec) {
2882            opc = INDEX_op_arm_ushl_vec;
2883        } else {
2884            opc = INDEX_op_arm_sshl_vec;
2885        }
2886        vec_gen_3(opc, type, vece, tcgv_vec_arg(v0),
2887                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2888        tcg_temp_free_vec(t1);
2889        break;
2890
2891    case INDEX_op_rotli_vec:
2892        t1 = tcg_temp_new_vec(type);
2893        tcg_gen_shri_vec(vece, t1, v1, -a2 & ((8 << vece) - 1));
2894        vec_gen_4(INDEX_op_arm_sli_vec, type, vece,
2895                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(v1), a2);
2896        tcg_temp_free_vec(t1);
2897        break;
2898
2899    case INDEX_op_rotlv_vec:
2900        v2 = temp_tcgv_vec(arg_temp(a2));
2901        t1 = tcg_temp_new_vec(type);
2902        c1 = tcg_constant_vec(type, vece, 8 << vece);
2903        tcg_gen_sub_vec(vece, t1, v2, c1);
2904        /* Right shifts are negative left shifts for NEON.  */
2905        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
2906                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2907        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
2908                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
2909        tcg_gen_or_vec(vece, v0, v0, t1);
2910        tcg_temp_free_vec(t1);
2911        break;
2912
2913    case INDEX_op_rotrv_vec:
2914        v2 = temp_tcgv_vec(arg_temp(a2));
2915        t1 = tcg_temp_new_vec(type);
2916        t2 = tcg_temp_new_vec(type);
2917        c1 = tcg_constant_vec(type, vece, 8 << vece);
2918        tcg_gen_neg_vec(vece, t1, v2);
2919        tcg_gen_sub_vec(vece, t2, c1, v2);
2920        /* Right shifts are negative left shifts for NEON.  */
2921        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
2922                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2923        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t2),
2924                  tcgv_vec_arg(v1), tcgv_vec_arg(t2));
2925        tcg_gen_or_vec(vece, v0, t1, t2);
2926        tcg_temp_free_vec(t1);
2927        tcg_temp_free_vec(t2);
2928        break;
2929
2930    default:
2931        g_assert_not_reached();
2932    }
2933}
2934
2935static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
2936{
2937    int i;
2938    for (i = 0; i < count; ++i) {
2939        p[i] = INSN_NOP;
2940    }
2941}
2942
2943/* Compute frame size via macros, to share between tcg_target_qemu_prologue
2944   and tcg_register_jit.  */
2945
2946#define PUSH_SIZE  ((11 - 4 + 1 + 1) * sizeof(tcg_target_long))
2947
2948#define FRAME_SIZE \
2949    ((PUSH_SIZE \
2950      + TCG_STATIC_CALL_ARGS_SIZE \
2951      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
2952      + TCG_TARGET_STACK_ALIGN - 1) \
2953     & -TCG_TARGET_STACK_ALIGN)
2954
2955#define STACK_ADDEND  (FRAME_SIZE - PUSH_SIZE)
2956
2957static void tcg_target_qemu_prologue(TCGContext *s)
2958{
2959    /* Calling convention requires us to save r4-r11 and lr.  */
2960    /* stmdb sp!, { r4 - r11, lr } */
2961    tcg_out_ldstm(s, COND_AL, INSN_STMDB, TCG_REG_CALL_STACK,
2962                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
2963                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
2964                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_R14));
2965
2966    /* Reserve callee argument and tcg temp space.  */
2967    tcg_out_dat_rI(s, COND_AL, ARITH_SUB, TCG_REG_CALL_STACK,
2968                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
2969    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
2970                  CPU_TEMP_BUF_NLONGS * sizeof(long));
2971
2972    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
2973
2974    if (!tcg_use_softmmu && guest_base) {
2975        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_GUEST_BASE, guest_base);
2976        tcg_regset_set_reg(s->reserved_regs, TCG_REG_GUEST_BASE);
2977    }
2978
2979    tcg_out_b_reg(s, COND_AL, tcg_target_call_iarg_regs[1]);
2980
2981    /*
2982     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
2983     * and fall through to the rest of the epilogue.
2984     */
2985    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
2986    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, 0);
2987    tcg_out_epilogue(s);
2988}
2989
2990static void tcg_out_epilogue(TCGContext *s)
2991{
2992    /* Release local stack frame.  */
2993    tcg_out_dat_rI(s, COND_AL, ARITH_ADD, TCG_REG_CALL_STACK,
2994                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
2995
2996    /* ldmia sp!, { r4 - r11, pc } */
2997    tcg_out_ldstm(s, COND_AL, INSN_LDMIA, TCG_REG_CALL_STACK,
2998                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
2999                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3000                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_PC));
3001}
3002
3003static void tcg_out_tb_start(TCGContext *s)
3004{
3005    /* nothing to do */
3006}
3007
3008typedef struct {
3009    DebugFrameHeader h;
3010    uint8_t fde_def_cfa[4];
3011    uint8_t fde_reg_ofs[18];
3012} DebugFrame;
3013
3014#define ELF_HOST_MACHINE EM_ARM
3015
3016/* We're expecting a 2 byte uleb128 encoded value.  */
3017QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
3018
3019static const DebugFrame debug_frame = {
3020    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
3021    .h.cie.id = -1,
3022    .h.cie.version = 1,
3023    .h.cie.code_align = 1,
3024    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
3025    .h.cie.return_column = 14,
3026
3027    /* Total FDE size does not include the "len" member.  */
3028    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
3029
3030    .fde_def_cfa = {
3031        12, 13,                         /* DW_CFA_def_cfa sp, ... */
3032        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
3033        (FRAME_SIZE >> 7)
3034    },
3035    .fde_reg_ofs = {
3036        /* The following must match the stmdb in the prologue.  */
3037        0x8e, 1,                        /* DW_CFA_offset, lr, -4 */
3038        0x8b, 2,                        /* DW_CFA_offset, r11, -8 */
3039        0x8a, 3,                        /* DW_CFA_offset, r10, -12 */
3040        0x89, 4,                        /* DW_CFA_offset, r9, -16 */
3041        0x88, 5,                        /* DW_CFA_offset, r8, -20 */
3042        0x87, 6,                        /* DW_CFA_offset, r7, -24 */
3043        0x86, 7,                        /* DW_CFA_offset, r6, -28 */
3044        0x85, 8,                        /* DW_CFA_offset, r5, -32 */
3045        0x84, 9,                        /* DW_CFA_offset, r4, -36 */
3046    }
3047};
3048
3049void tcg_register_jit(const void *buf, size_t buf_size)
3050{
3051    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
3052}
3053