xref: /openbmc/qemu/tcg/arm/tcg-target.c.inc (revision fa361eefac24dcaa1d6dfbc433fce0652fdd8ba8)
1/*
2 * Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2008 Andrzej Zaborowski
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to deal
8 * in the Software without restriction, including without limitation the rights
9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10 * copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22 * THE SOFTWARE.
23 */
24
25#include "elf.h"
26
27int arm_arch = __ARM_ARCH;
28
29#ifndef use_idiv_instructions
30bool use_idiv_instructions;
31#endif
32#ifndef use_neon_instructions
33bool use_neon_instructions;
34#endif
35
36/* Used for function call generation. */
37#define TCG_TARGET_STACK_ALIGN          8
38#define TCG_TARGET_CALL_STACK_OFFSET    0
39#define TCG_TARGET_CALL_ARG_I32         TCG_CALL_ARG_NORMAL
40#define TCG_TARGET_CALL_ARG_I64         TCG_CALL_ARG_EVEN
41#define TCG_TARGET_CALL_ARG_I128        TCG_CALL_ARG_EVEN
42#define TCG_TARGET_CALL_RET_I128        TCG_CALL_RET_BY_REF
43
44#ifdef CONFIG_DEBUG_TCG
45static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
46    "%r0",  "%r1",  "%r2",  "%r3",  "%r4",  "%r5",  "%r6",  "%r7",
47    "%r8",  "%r9",  "%r10", "%r11", "%r12", "%sp",  "%r14", "%pc",
48    "%q0",  "%q1",  "%q2",  "%q3",  "%q4",  "%q5",  "%q6",  "%q7",
49    "%q8",  "%q9",  "%q10", "%q11", "%q12", "%q13", "%q14", "%q15",
50};
51#endif
52
53static const int tcg_target_reg_alloc_order[] = {
54    TCG_REG_R4,
55    TCG_REG_R5,
56    TCG_REG_R6,
57    TCG_REG_R7,
58    TCG_REG_R8,
59    TCG_REG_R9,
60    TCG_REG_R10,
61    TCG_REG_R11,
62    TCG_REG_R13,
63    TCG_REG_R0,
64    TCG_REG_R1,
65    TCG_REG_R2,
66    TCG_REG_R3,
67    TCG_REG_R12,
68    TCG_REG_R14,
69
70    TCG_REG_Q0,
71    TCG_REG_Q1,
72    TCG_REG_Q2,
73    TCG_REG_Q3,
74    /* Q4 - Q7 are call-saved, and skipped. */
75    TCG_REG_Q8,
76    TCG_REG_Q9,
77    TCG_REG_Q10,
78    TCG_REG_Q11,
79    TCG_REG_Q12,
80    TCG_REG_Q13,
81    TCG_REG_Q14,
82    TCG_REG_Q15,
83};
84
85static const int tcg_target_call_iarg_regs[4] = {
86    TCG_REG_R0, TCG_REG_R1, TCG_REG_R2, TCG_REG_R3
87};
88
89static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
90{
91    tcg_debug_assert(kind == TCG_CALL_RET_NORMAL);
92    tcg_debug_assert(slot >= 0 && slot <= 3);
93    return TCG_REG_R0 + slot;
94}
95
96#define TCG_REG_TMP  TCG_REG_R12
97#define TCG_VEC_TMP  TCG_REG_Q15
98#define TCG_REG_GUEST_BASE  TCG_REG_R11
99
100typedef enum {
101    COND_EQ = 0x0,
102    COND_NE = 0x1,
103    COND_CS = 0x2,	/* Unsigned greater or equal */
104    COND_CC = 0x3,	/* Unsigned less than */
105    COND_MI = 0x4,	/* Negative */
106    COND_PL = 0x5,	/* Zero or greater */
107    COND_VS = 0x6,	/* Overflow */
108    COND_VC = 0x7,	/* No overflow */
109    COND_HI = 0x8,	/* Unsigned greater than */
110    COND_LS = 0x9,	/* Unsigned less or equal */
111    COND_GE = 0xa,
112    COND_LT = 0xb,
113    COND_GT = 0xc,
114    COND_LE = 0xd,
115    COND_AL = 0xe,
116} ARMCond;
117
118#define TO_CPSR (1 << 20)
119
120#define SHIFT_IMM_LSL(im)	(((im) << 7) | 0x00)
121#define SHIFT_IMM_LSR(im)	(((im) << 7) | 0x20)
122#define SHIFT_IMM_ASR(im)	(((im) << 7) | 0x40)
123#define SHIFT_IMM_ROR(im)	(((im) << 7) | 0x60)
124#define SHIFT_REG_LSL(rs)	(((rs) << 8) | 0x10)
125#define SHIFT_REG_LSR(rs)	(((rs) << 8) | 0x30)
126#define SHIFT_REG_ASR(rs)	(((rs) << 8) | 0x50)
127#define SHIFT_REG_ROR(rs)	(((rs) << 8) | 0x70)
128
129typedef enum {
130    ARITH_AND = 0x0 << 21,
131    ARITH_EOR = 0x1 << 21,
132    ARITH_SUB = 0x2 << 21,
133    ARITH_RSB = 0x3 << 21,
134    ARITH_ADD = 0x4 << 21,
135    ARITH_ADC = 0x5 << 21,
136    ARITH_SBC = 0x6 << 21,
137    ARITH_RSC = 0x7 << 21,
138    ARITH_TST = 0x8 << 21 | TO_CPSR,
139    ARITH_CMP = 0xa << 21 | TO_CPSR,
140    ARITH_CMN = 0xb << 21 | TO_CPSR,
141    ARITH_ORR = 0xc << 21,
142    ARITH_MOV = 0xd << 21,
143    ARITH_BIC = 0xe << 21,
144    ARITH_MVN = 0xf << 21,
145
146    INSN_B         = 0x0a000000,
147
148    INSN_CLZ       = 0x016f0f10,
149    INSN_RBIT      = 0x06ff0f30,
150
151    INSN_LDMIA     = 0x08b00000,
152    INSN_STMDB     = 0x09200000,
153
154    INSN_LDR_IMM   = 0x04100000,
155    INSN_LDR_REG   = 0x06100000,
156    INSN_STR_IMM   = 0x04000000,
157    INSN_STR_REG   = 0x06000000,
158
159    INSN_LDRH_IMM  = 0x005000b0,
160    INSN_LDRH_REG  = 0x001000b0,
161    INSN_LDRSH_IMM = 0x005000f0,
162    INSN_LDRSH_REG = 0x001000f0,
163    INSN_STRH_IMM  = 0x004000b0,
164    INSN_STRH_REG  = 0x000000b0,
165
166    INSN_LDRB_IMM  = 0x04500000,
167    INSN_LDRB_REG  = 0x06500000,
168    INSN_LDRSB_IMM = 0x005000d0,
169    INSN_LDRSB_REG = 0x001000d0,
170    INSN_STRB_IMM  = 0x04400000,
171    INSN_STRB_REG  = 0x06400000,
172
173    INSN_LDRD_IMM  = 0x004000d0,
174    INSN_LDRD_REG  = 0x000000d0,
175    INSN_STRD_IMM  = 0x004000f0,
176    INSN_STRD_REG  = 0x000000f0,
177
178    INSN_DMB_ISH   = 0xf57ff05b,
179    INSN_DMB_MCR   = 0xee070fba,
180
181    /* Architected nop introduced in v6k.  */
182    /* ??? This is an MSR (imm) 0,0,0 insn.  Anyone know if this
183       also Just So Happened to do nothing on pre-v6k so that we
184       don't need to conditionalize it?  */
185    INSN_NOP_v6k   = 0xe320f000,
186    /* Otherwise the assembler uses mov r0,r0 */
187    INSN_NOP_v4    = (COND_AL << 28) | ARITH_MOV,
188
189    INSN_VADD      = 0xf2000800,
190    INSN_VAND      = 0xf2000110,
191    INSN_VBIC      = 0xf2100110,
192    INSN_VEOR      = 0xf3000110,
193    INSN_VORN      = 0xf2300110,
194    INSN_VORR      = 0xf2200110,
195    INSN_VSUB      = 0xf3000800,
196    INSN_VMUL      = 0xf2000910,
197    INSN_VQADD     = 0xf2000010,
198    INSN_VQADD_U   = 0xf3000010,
199    INSN_VQSUB     = 0xf2000210,
200    INSN_VQSUB_U   = 0xf3000210,
201    INSN_VMAX      = 0xf2000600,
202    INSN_VMAX_U    = 0xf3000600,
203    INSN_VMIN      = 0xf2000610,
204    INSN_VMIN_U    = 0xf3000610,
205
206    INSN_VABS      = 0xf3b10300,
207    INSN_VMVN      = 0xf3b00580,
208    INSN_VNEG      = 0xf3b10380,
209
210    INSN_VCEQ0     = 0xf3b10100,
211    INSN_VCGT0     = 0xf3b10000,
212    INSN_VCGE0     = 0xf3b10080,
213    INSN_VCLE0     = 0xf3b10180,
214    INSN_VCLT0     = 0xf3b10200,
215
216    INSN_VCEQ      = 0xf3000810,
217    INSN_VCGE      = 0xf2000310,
218    INSN_VCGT      = 0xf2000300,
219    INSN_VCGE_U    = 0xf3000310,
220    INSN_VCGT_U    = 0xf3000300,
221
222    INSN_VSHLI     = 0xf2800510,  /* VSHL (immediate) */
223    INSN_VSARI     = 0xf2800010,  /* VSHR.S */
224    INSN_VSHRI     = 0xf3800010,  /* VSHR.U */
225    INSN_VSLI      = 0xf3800510,
226    INSN_VSHL_S    = 0xf2000400,  /* VSHL.S (register) */
227    INSN_VSHL_U    = 0xf3000400,  /* VSHL.U (register) */
228
229    INSN_VBSL      = 0xf3100110,
230    INSN_VBIT      = 0xf3200110,
231    INSN_VBIF      = 0xf3300110,
232
233    INSN_VTST      = 0xf2000810,
234
235    INSN_VDUP_G    = 0xee800b10,  /* VDUP (ARM core register) */
236    INSN_VDUP_S    = 0xf3b00c00,  /* VDUP (scalar) */
237    INSN_VLDR_D    = 0xed100b00,  /* VLDR.64 */
238    INSN_VLD1      = 0xf4200000,  /* VLD1 (multiple single elements) */
239    INSN_VLD1R     = 0xf4a00c00,  /* VLD1 (single element to all lanes) */
240    INSN_VST1      = 0xf4000000,  /* VST1 (multiple single elements) */
241    INSN_VMOVI     = 0xf2800010,  /* VMOV (immediate) */
242} ARMInsn;
243
244#define INSN_NOP   (use_armv7_instructions ? INSN_NOP_v6k : INSN_NOP_v4)
245
246static const uint8_t tcg_cond_to_arm_cond[] = {
247    [TCG_COND_EQ] = COND_EQ,
248    [TCG_COND_NE] = COND_NE,
249    [TCG_COND_LT] = COND_LT,
250    [TCG_COND_GE] = COND_GE,
251    [TCG_COND_LE] = COND_LE,
252    [TCG_COND_GT] = COND_GT,
253    /* unsigned */
254    [TCG_COND_LTU] = COND_CC,
255    [TCG_COND_GEU] = COND_CS,
256    [TCG_COND_LEU] = COND_LS,
257    [TCG_COND_GTU] = COND_HI,
258};
259
260static int encode_imm(uint32_t imm);
261
262/* TCG private relocation type: add with pc+imm8 */
263#define R_ARM_PC8  11
264
265/* TCG private relocation type: vldr with imm8 << 2 */
266#define R_ARM_PC11 12
267
268static bool reloc_pc24(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
269{
270    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
271    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) >> 2;
272
273    if (offset == sextract32(offset, 0, 24)) {
274        *src_rw = deposit32(*src_rw, 0, 24, offset);
275        return true;
276    }
277    return false;
278}
279
280static bool reloc_pc13(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
281{
282    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
283    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
284
285    if (offset >= -0xfff && offset <= 0xfff) {
286        tcg_insn_unit insn = *src_rw;
287        bool u = (offset >= 0);
288        if (!u) {
289            offset = -offset;
290        }
291        insn = deposit32(insn, 23, 1, u);
292        insn = deposit32(insn, 0, 12, offset);
293        *src_rw = insn;
294        return true;
295    }
296    return false;
297}
298
299static bool reloc_pc11(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
300{
301    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
302    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) / 4;
303
304    if (offset >= -0xff && offset <= 0xff) {
305        tcg_insn_unit insn = *src_rw;
306        bool u = (offset >= 0);
307        if (!u) {
308            offset = -offset;
309        }
310        insn = deposit32(insn, 23, 1, u);
311        insn = deposit32(insn, 0, 8, offset);
312        *src_rw = insn;
313        return true;
314    }
315    return false;
316}
317
318static bool reloc_pc8(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
319{
320    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
321    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
322    int imm12 = encode_imm(offset);
323
324    if (imm12 >= 0) {
325        *src_rw = deposit32(*src_rw, 0, 12, imm12);
326        return true;
327    }
328    return false;
329}
330
331static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
332                        intptr_t value, intptr_t addend)
333{
334    tcg_debug_assert(addend == 0);
335    switch (type) {
336    case R_ARM_PC24:
337        return reloc_pc24(code_ptr, (const tcg_insn_unit *)value);
338    case R_ARM_PC13:
339        return reloc_pc13(code_ptr, (const tcg_insn_unit *)value);
340    case R_ARM_PC11:
341        return reloc_pc11(code_ptr, (const tcg_insn_unit *)value);
342    case R_ARM_PC8:
343        return reloc_pc8(code_ptr, (const tcg_insn_unit *)value);
344    default:
345        g_assert_not_reached();
346    }
347}
348
349#define TCG_CT_CONST_ARM  0x100
350#define TCG_CT_CONST_INV  0x200
351#define TCG_CT_CONST_NEG  0x400
352#define TCG_CT_CONST_ZERO 0x800
353#define TCG_CT_CONST_ORRI 0x1000
354#define TCG_CT_CONST_ANDI 0x2000
355
356#define ALL_GENERAL_REGS  0xffffu
357#define ALL_VECTOR_REGS   0xffff0000u
358
359/*
360 * r0-r3 will be overwritten when reading the tlb entry (system-mode only);
361 * r14 will be overwritten by the BLNE branching to the slow path.
362 */
363#define ALL_QLDST_REGS \
364    (ALL_GENERAL_REGS & ~((tcg_use_softmmu ? 0xf : 0) | (1 << TCG_REG_R14)))
365
366/*
367 * ARM immediates for ALU instructions are made of an unsigned 8-bit
368 * right-rotated by an even amount between 0 and 30.
369 *
370 * Return < 0 if @imm cannot be encoded, else the entire imm12 field.
371 */
372static int encode_imm(uint32_t imm)
373{
374    uint32_t rot, imm8;
375
376    /* Simple case, no rotation required. */
377    if ((imm & ~0xff) == 0) {
378        return imm;
379    }
380
381    /* Next, try a simple even shift.  */
382    rot = ctz32(imm) & ~1;
383    imm8 = imm >> rot;
384    rot = 32 - rot;
385    if ((imm8 & ~0xff) == 0) {
386        goto found;
387    }
388
389    /*
390     * Finally, try harder with rotations.
391     * The ctz test above will have taken care of rotates >= 8.
392     */
393    for (rot = 2; rot < 8; rot += 2) {
394        imm8 = rol32(imm, rot);
395        if ((imm8 & ~0xff) == 0) {
396            goto found;
397        }
398    }
399    /* Fail: imm cannot be encoded. */
400    return -1;
401
402 found:
403    /* Note that rot is even, and we discard bit 0 by shifting by 7. */
404    return rot << 7 | imm8;
405}
406
407static int encode_imm_nofail(uint32_t imm)
408{
409    int ret = encode_imm(imm);
410    tcg_debug_assert(ret >= 0);
411    return ret;
412}
413
414static bool check_fit_imm(uint32_t imm)
415{
416    return encode_imm(imm) >= 0;
417}
418
419/* Return true if v16 is a valid 16-bit shifted immediate.  */
420static bool is_shimm16(uint16_t v16, int *cmode, int *imm8)
421{
422    if (v16 == (v16 & 0xff)) {
423        *cmode = 0x8;
424        *imm8 = v16 & 0xff;
425        return true;
426    } else if (v16 == (v16 & 0xff00)) {
427        *cmode = 0xa;
428        *imm8 = v16 >> 8;
429        return true;
430    }
431    return false;
432}
433
434/* Return true if v32 is a valid 32-bit shifted immediate.  */
435static bool is_shimm32(uint32_t v32, int *cmode, int *imm8)
436{
437    if (v32 == (v32 & 0xff)) {
438        *cmode = 0x0;
439        *imm8 = v32 & 0xff;
440        return true;
441    } else if (v32 == (v32 & 0xff00)) {
442        *cmode = 0x2;
443        *imm8 = (v32 >> 8) & 0xff;
444        return true;
445    } else if (v32 == (v32 & 0xff0000)) {
446        *cmode = 0x4;
447        *imm8 = (v32 >> 16) & 0xff;
448        return true;
449    } else if (v32 == (v32 & 0xff000000)) {
450        *cmode = 0x6;
451        *imm8 = v32 >> 24;
452        return true;
453    }
454    return false;
455}
456
457/* Return true if v32 is a valid 32-bit shifting ones immediate.  */
458static bool is_soimm32(uint32_t v32, int *cmode, int *imm8)
459{
460    if ((v32 & 0xffff00ff) == 0xff) {
461        *cmode = 0xc;
462        *imm8 = (v32 >> 8) & 0xff;
463        return true;
464    } else if ((v32 & 0xff00ffff) == 0xffff) {
465        *cmode = 0xd;
466        *imm8 = (v32 >> 16) & 0xff;
467        return true;
468    }
469    return false;
470}
471
472/*
473 * Return non-zero if v32 can be formed by MOVI+ORR.
474 * Place the parameters for MOVI in (cmode, imm8).
475 * Return the cmode for ORR; the imm8 can be had via extraction from v32.
476 */
477static int is_shimm32_pair(uint32_t v32, int *cmode, int *imm8)
478{
479    int i;
480
481    for (i = 6; i > 0; i -= 2) {
482        /* Mask out one byte we can add with ORR.  */
483        uint32_t tmp = v32 & ~(0xffu << (i * 4));
484        if (is_shimm32(tmp, cmode, imm8) ||
485            is_soimm32(tmp, cmode, imm8)) {
486            break;
487        }
488    }
489    return i;
490}
491
492/* Return true if V is a valid 16-bit or 32-bit shifted immediate.  */
493static bool is_shimm1632(uint32_t v32, int *cmode, int *imm8)
494{
495    if (v32 == deposit32(v32, 16, 16, v32)) {
496        return is_shimm16(v32, cmode, imm8);
497    } else {
498        return is_shimm32(v32, cmode, imm8);
499    }
500}
501
502/* Test if a constant matches the constraint.
503 * TODO: define constraints for:
504 *
505 * ldr/str offset:   between -0xfff and 0xfff
506 * ldrh/strh offset: between -0xff and 0xff
507 * mov operand2:     values represented with x << (2 * y), x < 0x100
508 * add, sub, eor...: ditto
509 */
510static bool tcg_target_const_match(int64_t val, int ct,
511                                   TCGType type, TCGCond cond, int vece)
512{
513    if (ct & TCG_CT_CONST) {
514        return 1;
515    } else if ((ct & TCG_CT_CONST_ARM) && check_fit_imm(val)) {
516        return 1;
517    } else if ((ct & TCG_CT_CONST_INV) && check_fit_imm(~val)) {
518        return 1;
519    } else if ((ct & TCG_CT_CONST_NEG) && check_fit_imm(-val)) {
520        return 1;
521    } else if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
522        return 1;
523    }
524
525    switch (ct & (TCG_CT_CONST_ORRI | TCG_CT_CONST_ANDI)) {
526    case 0:
527        break;
528    case TCG_CT_CONST_ANDI:
529        val = ~val;
530        /* fallthru */
531    case TCG_CT_CONST_ORRI:
532        if (val == deposit64(val, 32, 32, val)) {
533            int cmode, imm8;
534            return is_shimm1632(val, &cmode, &imm8);
535        }
536        break;
537    default:
538        /* Both bits should not be set for the same insn.  */
539        g_assert_not_reached();
540    }
541
542    return 0;
543}
544
545static void tcg_out_b_imm(TCGContext *s, ARMCond cond, int32_t offset)
546{
547    tcg_out32(s, (cond << 28) | INSN_B |
548                    (((offset - 8) >> 2) & 0x00ffffff));
549}
550
551static void tcg_out_bl_imm(TCGContext *s, ARMCond cond, int32_t offset)
552{
553    tcg_out32(s, (cond << 28) | 0x0b000000 |
554                    (((offset - 8) >> 2) & 0x00ffffff));
555}
556
557static void tcg_out_blx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
558{
559    tcg_out32(s, (cond << 28) | 0x012fff30 | rn);
560}
561
562static void tcg_out_blx_imm(TCGContext *s, int32_t offset)
563{
564    tcg_out32(s, 0xfa000000 | ((offset & 2) << 23) |
565                (((offset - 8) >> 2) & 0x00ffffff));
566}
567
568static void tcg_out_dat_reg(TCGContext *s, ARMCond cond, ARMInsn opc,
569                            TCGReg rd, TCGReg rn, TCGReg rm, int shift)
570{
571    tcg_out32(s, (cond << 28) | (0 << 25) | opc |
572                    (rn << 16) | (rd << 12) | shift | rm);
573}
574
575static void tcg_out_mov_reg(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rm)
576{
577    /* Simple reg-reg move, optimising out the 'do nothing' case */
578    if (rd != rm) {
579        tcg_out_dat_reg(s, cond, ARITH_MOV, rd, 0, rm, SHIFT_IMM_LSL(0));
580    }
581}
582
583static void tcg_out_bx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
584{
585    tcg_out32(s, (cond << 28) | 0x012fff10 | rn);
586}
587
588static void tcg_out_b_reg(TCGContext *s, ARMCond cond, TCGReg rn)
589{
590    /*
591     * Unless the C portion of QEMU is compiled as thumb, we don't need
592     * true BX semantics; merely a branch to an address held in a register.
593     */
594    tcg_out_bx_reg(s, cond, rn);
595}
596
597static void tcg_out_dat_imm(TCGContext *s, ARMCond cond, ARMInsn opc,
598                            TCGReg rd, TCGReg rn, int im)
599{
600    tcg_out32(s, (cond << 28) | (1 << 25) | opc |
601                    (rn << 16) | (rd << 12) | im);
602}
603
604static void tcg_out_ldstm(TCGContext *s, ARMCond cond, ARMInsn opc,
605                          TCGReg rn, uint16_t mask)
606{
607    tcg_out32(s, (cond << 28) | opc | (rn << 16) | mask);
608}
609
610/* Note that this routine is used for both LDR and LDRH formats, so we do
611   not wish to include an immediate shift at this point.  */
612static void tcg_out_memop_r(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
613                            TCGReg rn, TCGReg rm, bool u, bool p, bool w)
614{
615    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24)
616              | (w << 21) | (rn << 16) | (rt << 12) | rm);
617}
618
619static void tcg_out_memop_8(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
620                            TCGReg rn, int imm8, bool p, bool w)
621{
622    bool u = 1;
623    if (imm8 < 0) {
624        imm8 = -imm8;
625        u = 0;
626    }
627    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
628              (rn << 16) | (rt << 12) | ((imm8 & 0xf0) << 4) | (imm8 & 0xf));
629}
630
631static void tcg_out_memop_12(TCGContext *s, ARMCond cond, ARMInsn opc,
632                             TCGReg rt, TCGReg rn, int imm12, bool p, bool w)
633{
634    bool u = 1;
635    if (imm12 < 0) {
636        imm12 = -imm12;
637        u = 0;
638    }
639    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
640              (rn << 16) | (rt << 12) | imm12);
641}
642
643static void tcg_out_ld32_12(TCGContext *s, ARMCond cond, TCGReg rt,
644                            TCGReg rn, int imm12)
645{
646    tcg_out_memop_12(s, cond, INSN_LDR_IMM, rt, rn, imm12, 1, 0);
647}
648
649static void tcg_out_st32_12(TCGContext *s, ARMCond cond, TCGReg rt,
650                            TCGReg rn, int imm12)
651{
652    tcg_out_memop_12(s, cond, INSN_STR_IMM, rt, rn, imm12, 1, 0);
653}
654
655static void tcg_out_ld32_r(TCGContext *s, ARMCond cond, TCGReg rt,
656                           TCGReg rn, TCGReg rm)
657{
658    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 0);
659}
660
661static void tcg_out_st32_r(TCGContext *s, ARMCond cond, TCGReg rt,
662                           TCGReg rn, TCGReg rm)
663{
664    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 0);
665}
666
667static void tcg_out_ldrd_8(TCGContext *s, ARMCond cond, TCGReg rt,
668                           TCGReg rn, int imm8)
669{
670    tcg_out_memop_8(s, cond, INSN_LDRD_IMM, rt, rn, imm8, 1, 0);
671}
672
673static void tcg_out_ldrd_r(TCGContext *s, ARMCond cond, TCGReg rt,
674                           TCGReg rn, TCGReg rm)
675{
676    tcg_out_memop_r(s, cond, INSN_LDRD_REG, rt, rn, rm, 1, 1, 0);
677}
678
679static void tcg_out_strd_8(TCGContext *s, ARMCond cond, TCGReg rt,
680                           TCGReg rn, int imm8)
681{
682    tcg_out_memop_8(s, cond, INSN_STRD_IMM, rt, rn, imm8, 1, 0);
683}
684
685static void tcg_out_strd_r(TCGContext *s, ARMCond cond, TCGReg rt,
686                           TCGReg rn, TCGReg rm)
687{
688    tcg_out_memop_r(s, cond, INSN_STRD_REG, rt, rn, rm, 1, 1, 0);
689}
690
691/* Register pre-increment with base writeback.  */
692static void tcg_out_ld32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
693                             TCGReg rn, TCGReg rm)
694{
695    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 1);
696}
697
698static void tcg_out_st32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
699                             TCGReg rn, TCGReg rm)
700{
701    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 1);
702}
703
704static void tcg_out_ld16u_8(TCGContext *s, ARMCond cond, TCGReg rt,
705                            TCGReg rn, int imm8)
706{
707    tcg_out_memop_8(s, cond, INSN_LDRH_IMM, rt, rn, imm8, 1, 0);
708}
709
710static void tcg_out_st16_8(TCGContext *s, ARMCond cond, TCGReg rt,
711                           TCGReg rn, int imm8)
712{
713    tcg_out_memop_8(s, cond, INSN_STRH_IMM, rt, rn, imm8, 1, 0);
714}
715
716static void tcg_out_ld16u_r(TCGContext *s, ARMCond cond, TCGReg rt,
717                            TCGReg rn, TCGReg rm)
718{
719    tcg_out_memop_r(s, cond, INSN_LDRH_REG, rt, rn, rm, 1, 1, 0);
720}
721
722static void tcg_out_st16_r(TCGContext *s, ARMCond cond, TCGReg rt,
723                           TCGReg rn, TCGReg rm)
724{
725    tcg_out_memop_r(s, cond, INSN_STRH_REG, rt, rn, rm, 1, 1, 0);
726}
727
728static void tcg_out_ld16s_8(TCGContext *s, ARMCond cond, TCGReg rt,
729                            TCGReg rn, int imm8)
730{
731    tcg_out_memop_8(s, cond, INSN_LDRSH_IMM, rt, rn, imm8, 1, 0);
732}
733
734static void tcg_out_ld16s_r(TCGContext *s, ARMCond cond, TCGReg rt,
735                            TCGReg rn, TCGReg rm)
736{
737    tcg_out_memop_r(s, cond, INSN_LDRSH_REG, rt, rn, rm, 1, 1, 0);
738}
739
740static void tcg_out_ld8_12(TCGContext *s, ARMCond cond, TCGReg rt,
741                           TCGReg rn, int imm12)
742{
743    tcg_out_memop_12(s, cond, INSN_LDRB_IMM, rt, rn, imm12, 1, 0);
744}
745
746static void tcg_out_st8_12(TCGContext *s, ARMCond cond, TCGReg rt,
747                           TCGReg rn, int imm12)
748{
749    tcg_out_memop_12(s, cond, INSN_STRB_IMM, rt, rn, imm12, 1, 0);
750}
751
752static void tcg_out_ld8_r(TCGContext *s, ARMCond cond, TCGReg rt,
753                          TCGReg rn, TCGReg rm)
754{
755    tcg_out_memop_r(s, cond, INSN_LDRB_REG, rt, rn, rm, 1, 1, 0);
756}
757
758static void tcg_out_st8_r(TCGContext *s, ARMCond cond, TCGReg rt,
759                          TCGReg rn, TCGReg rm)
760{
761    tcg_out_memop_r(s, cond, INSN_STRB_REG, rt, rn, rm, 1, 1, 0);
762}
763
764static void tcg_out_ld8s_8(TCGContext *s, ARMCond cond, TCGReg rt,
765                           TCGReg rn, int imm8)
766{
767    tcg_out_memop_8(s, cond, INSN_LDRSB_IMM, rt, rn, imm8, 1, 0);
768}
769
770static void tcg_out_ld8s_r(TCGContext *s, ARMCond cond, TCGReg rt,
771                           TCGReg rn, TCGReg rm)
772{
773    tcg_out_memop_r(s, cond, INSN_LDRSB_REG, rt, rn, rm, 1, 1, 0);
774}
775
776static void tcg_out_movi_pool(TCGContext *s, ARMCond cond,
777                              TCGReg rd, uint32_t arg)
778{
779    new_pool_label(s, arg, R_ARM_PC13, s->code_ptr, 0);
780    tcg_out_ld32_12(s, cond, rd, TCG_REG_PC, 0);
781}
782
783static void tcg_out_movi32(TCGContext *s, ARMCond cond,
784                           TCGReg rd, uint32_t arg)
785{
786    int imm12, diff, opc, sh1, sh2;
787    uint32_t tt0, tt1, tt2;
788
789    /* Check a single MOV/MVN before anything else.  */
790    imm12 = encode_imm(arg);
791    if (imm12 >= 0) {
792        tcg_out_dat_imm(s, cond, ARITH_MOV, rd, 0, imm12);
793        return;
794    }
795    imm12 = encode_imm(~arg);
796    if (imm12 >= 0) {
797        tcg_out_dat_imm(s, cond, ARITH_MVN, rd, 0, imm12);
798        return;
799    }
800
801    /* Check for a pc-relative address.  This will usually be the TB,
802       or within the TB, which is immediately before the code block.  */
803    diff = tcg_pcrel_diff(s, (void *)arg) - 8;
804    if (diff >= 0) {
805        imm12 = encode_imm(diff);
806        if (imm12 >= 0) {
807            tcg_out_dat_imm(s, cond, ARITH_ADD, rd, TCG_REG_PC, imm12);
808            return;
809        }
810    } else {
811        imm12 = encode_imm(-diff);
812        if (imm12 >= 0) {
813            tcg_out_dat_imm(s, cond, ARITH_SUB, rd, TCG_REG_PC, imm12);
814            return;
815        }
816    }
817
818    /* Use movw + movt.  */
819    if (use_armv7_instructions) {
820        /* movw */
821        tcg_out32(s, (cond << 28) | 0x03000000 | (rd << 12)
822                  | ((arg << 4) & 0x000f0000) | (arg & 0xfff));
823        if (arg & 0xffff0000) {
824            /* movt */
825            tcg_out32(s, (cond << 28) | 0x03400000 | (rd << 12)
826                      | ((arg >> 12) & 0x000f0000) | ((arg >> 16) & 0xfff));
827        }
828        return;
829    }
830
831    /* Look for sequences of two insns.  If we have lots of 1's, we can
832       shorten the sequence by beginning with mvn and then clearing
833       higher bits with eor.  */
834    tt0 = arg;
835    opc = ARITH_MOV;
836    if (ctpop32(arg) > 16) {
837        tt0 = ~arg;
838        opc = ARITH_MVN;
839    }
840    sh1 = ctz32(tt0) & ~1;
841    tt1 = tt0 & ~(0xff << sh1);
842    sh2 = ctz32(tt1) & ~1;
843    tt2 = tt1 & ~(0xff << sh2);
844    if (tt2 == 0) {
845        int rot;
846
847        rot = ((32 - sh1) << 7) & 0xf00;
848        tcg_out_dat_imm(s, cond, opc, rd,  0, ((tt0 >> sh1) & 0xff) | rot);
849        rot = ((32 - sh2) << 7) & 0xf00;
850        tcg_out_dat_imm(s, cond, ARITH_EOR, rd, rd,
851                        ((tt0 >> sh2) & 0xff) | rot);
852        return;
853    }
854
855    /* Otherwise, drop it into the constant pool.  */
856    tcg_out_movi_pool(s, cond, rd, arg);
857}
858
859/*
860 * Emit either the reg,imm or reg,reg form of a data-processing insn.
861 * rhs must satisfy the "rI" constraint.
862 */
863static void tcg_out_dat_rI(TCGContext *s, ARMCond cond, ARMInsn opc,
864                           TCGReg dst, TCGReg lhs, TCGArg rhs, int rhs_is_const)
865{
866    if (rhs_is_const) {
867        tcg_out_dat_imm(s, cond, opc, dst, lhs, encode_imm_nofail(rhs));
868    } else {
869        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
870    }
871}
872
873/*
874 * Emit either the reg,imm or reg,reg form of a data-processing insn.
875 * rhs must satisfy the "rIK" constraint.
876 */
877static void tcg_out_dat_IK(TCGContext *s, ARMCond cond, ARMInsn opc,
878                            ARMInsn opinv, TCGReg dst, TCGReg lhs, TCGArg rhs)
879{
880    int imm12 = encode_imm(rhs);
881    if (imm12 < 0) {
882        imm12 = encode_imm_nofail(~rhs);
883        opc = opinv;
884    }
885    tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
886}
887
888static void tcg_out_dat_rIK(TCGContext *s, ARMCond cond, ARMInsn opc,
889                            ARMInsn opinv, TCGReg dst, TCGReg lhs, TCGArg rhs,
890                            bool rhs_is_const)
891{
892    if (rhs_is_const) {
893        tcg_out_dat_IK(s, cond, opc, opinv, dst, lhs, rhs);
894    } else {
895        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
896    }
897}
898
899static void tcg_out_dat_IN(TCGContext *s, ARMCond cond, ARMInsn opc,
900                           ARMInsn opneg, TCGReg dst, TCGReg lhs, TCGArg rhs)
901{
902    int imm12 = encode_imm(rhs);
903    if (imm12 < 0) {
904        imm12 = encode_imm_nofail(-rhs);
905        opc = opneg;
906    }
907    tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
908}
909
910static void tcg_out_dat_rIN(TCGContext *s, ARMCond cond, ARMInsn opc,
911                            ARMInsn opneg, TCGReg dst, TCGReg lhs, TCGArg rhs,
912                            bool rhs_is_const)
913{
914    /* Emit either the reg,imm or reg,reg form of a data-processing insn.
915     * rhs must satisfy the "rIN" constraint.
916     */
917    if (rhs_is_const) {
918        tcg_out_dat_IN(s, cond, opc, opneg, dst, lhs, rhs);
919    } else {
920        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
921    }
922}
923
924static void tcg_out_ext8s(TCGContext *s, TCGType t, TCGReg rd, TCGReg rn)
925{
926    /* sxtb */
927    tcg_out32(s, 0x06af0070 | (COND_AL << 28) | (rd << 12) | rn);
928}
929
930static void tcg_out_ext8u(TCGContext *s, TCGReg rd, TCGReg rn)
931{
932    tcg_out_dat_imm(s, COND_AL, ARITH_AND, rd, rn, 0xff);
933}
934
935static void tcg_out_ext16s(TCGContext *s, TCGType t, TCGReg rd, TCGReg rn)
936{
937    /* sxth */
938    tcg_out32(s, 0x06bf0070 | (COND_AL << 28) | (rd << 12) | rn);
939}
940
941static void tcg_out_ext16u(TCGContext *s, TCGReg rd, TCGReg rn)
942{
943    /* uxth */
944    tcg_out32(s, 0x06ff0070 | (COND_AL << 28) | (rd << 12) | rn);
945}
946
947static void tcg_out_ext32s(TCGContext *s, TCGReg rd, TCGReg rn)
948{
949    g_assert_not_reached();
950}
951
952static void tcg_out_ext32u(TCGContext *s, TCGReg rd, TCGReg rn)
953{
954    g_assert_not_reached();
955}
956
957static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
958{
959    g_assert_not_reached();
960}
961
962static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg rd, TCGReg rn)
963{
964    g_assert_not_reached();
965}
966
967static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg rd, TCGReg rn)
968{
969    g_assert_not_reached();
970}
971
972static void tcg_out_deposit(TCGContext *s, ARMCond cond, TCGReg rd,
973                            TCGArg a1, int ofs, int len, bool const_a1)
974{
975    if (const_a1) {
976        /* bfi becomes bfc with rn == 15.  */
977        a1 = 15;
978    }
979    /* bfi/bfc */
980    tcg_out32(s, 0x07c00010 | (cond << 28) | (rd << 12) | a1
981              | (ofs << 7) | ((ofs + len - 1) << 16));
982}
983
984static void tgen_extract(TCGContext *s, TCGType type, TCGReg rd, TCGReg rn,
985                         unsigned ofs, unsigned len)
986{
987    /* According to gcc, AND can be faster. */
988    if (ofs == 0 && len <= 8) {
989        tcg_out_dat_imm(s, COND_AL, ARITH_AND, rd, rn,
990                        encode_imm_nofail((1 << len) - 1));
991        return;
992    }
993
994    if (use_armv7_instructions) {
995        /* ubfx */
996        tcg_out32(s, 0x07e00050 | (COND_AL << 28) | (rd << 12) | rn
997                  | (ofs << 7) | ((len - 1) << 16));
998        return;
999    }
1000
1001    assert(ofs % 8 == 0);
1002    switch (len) {
1003    case 8:
1004        /* uxtb */
1005        tcg_out32(s, 0x06ef0070 | (COND_AL << 28) |
1006                  (rd << 12) | (ofs << 7) | rn);
1007        break;
1008    case 16:
1009        /* uxth */
1010        tcg_out32(s, 0x06ff0070 | (COND_AL << 28) |
1011                  (rd << 12) | (ofs << 7) | rn);
1012        break;
1013    default:
1014        g_assert_not_reached();
1015    }
1016}
1017
1018static const TCGOutOpExtract outop_extract = {
1019    .base.static_constraint = C_O1_I1(r, r),
1020    .out_rr = tgen_extract,
1021};
1022
1023static void tgen_sextract(TCGContext *s, TCGType type, TCGReg rd, TCGReg rn,
1024                          unsigned ofs, unsigned len)
1025{
1026    if (use_armv7_instructions) {
1027        /* sbfx */
1028        tcg_out32(s, 0x07a00050 | (COND_AL << 28) | (rd << 12) | rn
1029                  | (ofs << 7) | ((len - 1) << 16));
1030        return;
1031    }
1032
1033    assert(ofs % 8 == 0);
1034    switch (len) {
1035    case 8:
1036        /* sxtb */
1037        tcg_out32(s, 0x06af0070 | (COND_AL << 28) |
1038                  (rd << 12) | (ofs << 7) | rn);
1039        break;
1040    case 16:
1041        /* sxth */
1042        tcg_out32(s, 0x06bf0070 | (COND_AL << 28) |
1043                  (rd << 12) | (ofs << 7) | rn);
1044        break;
1045    default:
1046        g_assert_not_reached();
1047    }
1048}
1049
1050static const TCGOutOpExtract outop_sextract = {
1051    .base.static_constraint = C_O1_I1(r, r),
1052    .out_rr = tgen_sextract,
1053};
1054
1055
1056static void tcg_out_ld32u(TCGContext *s, ARMCond cond,
1057                          TCGReg rd, TCGReg rn, int32_t offset)
1058{
1059    if (offset > 0xfff || offset < -0xfff) {
1060        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1061        tcg_out_ld32_r(s, cond, rd, rn, TCG_REG_TMP);
1062    } else
1063        tcg_out_ld32_12(s, cond, rd, rn, offset);
1064}
1065
1066static void tcg_out_st32(TCGContext *s, ARMCond cond,
1067                         TCGReg rd, TCGReg rn, int32_t offset)
1068{
1069    if (offset > 0xfff || offset < -0xfff) {
1070        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1071        tcg_out_st32_r(s, cond, rd, rn, TCG_REG_TMP);
1072    } else
1073        tcg_out_st32_12(s, cond, rd, rn, offset);
1074}
1075
1076static void tcg_out_ld16u(TCGContext *s, ARMCond cond,
1077                          TCGReg rd, TCGReg rn, int32_t offset)
1078{
1079    if (offset > 0xff || offset < -0xff) {
1080        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1081        tcg_out_ld16u_r(s, cond, rd, rn, TCG_REG_TMP);
1082    } else
1083        tcg_out_ld16u_8(s, cond, rd, rn, offset);
1084}
1085
1086static void tcg_out_ld16s(TCGContext *s, ARMCond cond,
1087                          TCGReg rd, TCGReg rn, int32_t offset)
1088{
1089    if (offset > 0xff || offset < -0xff) {
1090        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1091        tcg_out_ld16s_r(s, cond, rd, rn, TCG_REG_TMP);
1092    } else
1093        tcg_out_ld16s_8(s, cond, rd, rn, offset);
1094}
1095
1096static void tcg_out_st16(TCGContext *s, ARMCond cond,
1097                         TCGReg rd, TCGReg rn, int32_t offset)
1098{
1099    if (offset > 0xff || offset < -0xff) {
1100        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1101        tcg_out_st16_r(s, cond, rd, rn, TCG_REG_TMP);
1102    } else
1103        tcg_out_st16_8(s, cond, rd, rn, offset);
1104}
1105
1106static void tcg_out_ld8u(TCGContext *s, ARMCond cond,
1107                         TCGReg rd, TCGReg rn, int32_t offset)
1108{
1109    if (offset > 0xfff || offset < -0xfff) {
1110        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1111        tcg_out_ld8_r(s, cond, rd, rn, TCG_REG_TMP);
1112    } else
1113        tcg_out_ld8_12(s, cond, rd, rn, offset);
1114}
1115
1116static void tcg_out_ld8s(TCGContext *s, ARMCond cond,
1117                         TCGReg rd, TCGReg rn, int32_t offset)
1118{
1119    if (offset > 0xff || offset < -0xff) {
1120        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1121        tcg_out_ld8s_r(s, cond, rd, rn, TCG_REG_TMP);
1122    } else
1123        tcg_out_ld8s_8(s, cond, rd, rn, offset);
1124}
1125
1126static void tcg_out_st8(TCGContext *s, ARMCond cond,
1127                        TCGReg rd, TCGReg rn, int32_t offset)
1128{
1129    if (offset > 0xfff || offset < -0xfff) {
1130        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1131        tcg_out_st8_r(s, cond, rd, rn, TCG_REG_TMP);
1132    } else
1133        tcg_out_st8_12(s, cond, rd, rn, offset);
1134}
1135
1136/*
1137 * The _goto case is normally between TBs within the same code buffer, and
1138 * with the code buffer limited to 16MB we wouldn't need the long case.
1139 * But we also use it for the tail-call to the qemu_ld/st helpers, which does.
1140 */
1141static void tcg_out_goto(TCGContext *s, ARMCond cond, const tcg_insn_unit *addr)
1142{
1143    intptr_t addri = (intptr_t)addr;
1144    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1145    bool arm_mode = !(addri & 1);
1146
1147    if (arm_mode && disp - 8 < 0x01fffffd && disp - 8 > -0x01fffffd) {
1148        tcg_out_b_imm(s, cond, disp);
1149        return;
1150    }
1151
1152    /* LDR is interworking from v5t. */
1153    tcg_out_movi_pool(s, cond, TCG_REG_PC, addri);
1154}
1155
1156/*
1157 * The call case is mostly used for helpers - so it's not unreasonable
1158 * for them to be beyond branch range.
1159 */
1160static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *addr)
1161{
1162    intptr_t addri = (intptr_t)addr;
1163    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1164    bool arm_mode = !(addri & 1);
1165
1166    if (disp - 8 < 0x02000000 && disp - 8 >= -0x02000000) {
1167        if (arm_mode) {
1168            tcg_out_bl_imm(s, COND_AL, disp);
1169        } else {
1170            tcg_out_blx_imm(s, disp);
1171        }
1172        return;
1173    }
1174
1175    tcg_out_movi32(s, COND_AL, TCG_REG_TMP, addri);
1176    tcg_out_blx_reg(s, COND_AL, TCG_REG_TMP);
1177}
1178
1179static void tcg_out_call(TCGContext *s, const tcg_insn_unit *addr,
1180                         const TCGHelperInfo *info)
1181{
1182    tcg_out_call_int(s, addr);
1183}
1184
1185static void tcg_out_goto_label(TCGContext *s, ARMCond cond, TCGLabel *l)
1186{
1187    if (l->has_value) {
1188        tcg_out_goto(s, cond, l->u.value_ptr);
1189    } else {
1190        tcg_out_reloc(s, s->code_ptr, R_ARM_PC24, l, 0);
1191        tcg_out_b_imm(s, cond, 0);
1192    }
1193}
1194
1195static void tcg_out_mb(TCGContext *s, TCGArg a0)
1196{
1197    if (use_armv7_instructions) {
1198        tcg_out32(s, INSN_DMB_ISH);
1199    } else {
1200        tcg_out32(s, INSN_DMB_MCR);
1201    }
1202}
1203
1204static TCGCond tgen_cmp(TCGContext *s, TCGCond cond, TCGReg a, TCGReg b)
1205{
1206    if (is_tst_cond(cond)) {
1207        tcg_out_dat_reg(s, COND_AL, ARITH_TST, 0, a, b, SHIFT_IMM_LSL(0));
1208        return tcg_tst_eqne_cond(cond);
1209    }
1210    tcg_out_dat_reg(s, COND_AL, ARITH_CMP, 0, a, b, SHIFT_IMM_LSL(0));
1211    return cond;
1212}
1213
1214static TCGCond tgen_cmpi(TCGContext *s, TCGCond cond, TCGReg a, TCGArg b)
1215{
1216    int imm12;
1217
1218    if (!is_tst_cond(cond)) {
1219        tcg_out_dat_IN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0, a, b);
1220        return cond;
1221    }
1222
1223    /*
1224     * The compare constraints allow rIN, but TST does not support N.
1225     * Be prepared to load the constant into a scratch register.
1226     */
1227    imm12 = encode_imm(b);
1228    if (imm12 >= 0) {
1229        tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, a, imm12);
1230    } else {
1231        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, b);
1232        tcg_out_dat_reg(s, COND_AL, ARITH_TST, 0,
1233                        a, TCG_REG_TMP, SHIFT_IMM_LSL(0));
1234    }
1235    return tcg_tst_eqne_cond(cond);
1236}
1237
1238static TCGCond tcg_out_cmp(TCGContext *s, TCGCond cond, TCGReg a,
1239                           TCGArg b, int b_const)
1240{
1241    if (b_const) {
1242        return tgen_cmpi(s, cond, a, b);
1243    } else {
1244        return tgen_cmp(s, cond, a, b);
1245    }
1246}
1247
1248static TCGCond tcg_out_cmp2(TCGContext *s, TCGCond cond, TCGReg al, TCGReg ah,
1249                            TCGArg bl, bool const_bl, TCGArg bh, bool const_bh)
1250{
1251    switch (cond) {
1252    case TCG_COND_EQ:
1253    case TCG_COND_NE:
1254    case TCG_COND_LTU:
1255    case TCG_COND_LEU:
1256    case TCG_COND_GTU:
1257    case TCG_COND_GEU:
1258        /*
1259         * We perform a conditional comparison.  If the high half is
1260         * equal, then overwrite the flags with the comparison of the
1261         * low half.  The resulting flags cover the whole.
1262         */
1263        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, ah, bh, const_bh);
1264        tcg_out_dat_rI(s, COND_EQ, ARITH_CMP, 0, al, bl, const_bl);
1265        return cond;
1266
1267    case TCG_COND_TSTEQ:
1268    case TCG_COND_TSTNE:
1269        /* Similar, but with TST instead of CMP. */
1270        tcg_out_dat_rI(s, COND_AL, ARITH_TST, 0, ah, bh, const_bh);
1271        tcg_out_dat_rI(s, COND_EQ, ARITH_TST, 0, al, bl, const_bl);
1272        return tcg_tst_eqne_cond(cond);
1273
1274    case TCG_COND_LT:
1275    case TCG_COND_GE:
1276        /* We perform a double-word subtraction and examine the result.
1277           We do not actually need the result of the subtract, so the
1278           low part "subtract" is a compare.  For the high half we have
1279           no choice but to compute into a temporary.  */
1280        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, al, bl, const_bl);
1281        tcg_out_dat_rI(s, COND_AL, ARITH_SBC | TO_CPSR,
1282                       TCG_REG_TMP, ah, bh, const_bh);
1283        return cond;
1284
1285    case TCG_COND_LE:
1286    case TCG_COND_GT:
1287        /* Similar, but with swapped arguments, via reversed subtract.  */
1288        tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR,
1289                       TCG_REG_TMP, al, bl, const_bl);
1290        tcg_out_dat_rI(s, COND_AL, ARITH_RSC | TO_CPSR,
1291                       TCG_REG_TMP, ah, bh, const_bh);
1292        return tcg_swap_cond(cond);
1293
1294    default:
1295        g_assert_not_reached();
1296    }
1297}
1298
1299/*
1300 * Note that TCGReg references Q-registers.
1301 * Q-regno = 2 * D-regno, so shift left by 1 while inserting.
1302 */
1303static uint32_t encode_vd(TCGReg rd)
1304{
1305    tcg_debug_assert(rd >= TCG_REG_Q0);
1306    return (extract32(rd, 3, 1) << 22) | (extract32(rd, 0, 3) << 13);
1307}
1308
1309static uint32_t encode_vn(TCGReg rn)
1310{
1311    tcg_debug_assert(rn >= TCG_REG_Q0);
1312    return (extract32(rn, 3, 1) << 7) | (extract32(rn, 0, 3) << 17);
1313}
1314
1315static uint32_t encode_vm(TCGReg rm)
1316{
1317    tcg_debug_assert(rm >= TCG_REG_Q0);
1318    return (extract32(rm, 3, 1) << 5) | (extract32(rm, 0, 3) << 1);
1319}
1320
1321static void tcg_out_vreg2(TCGContext *s, ARMInsn insn, int q, int vece,
1322                          TCGReg d, TCGReg m)
1323{
1324    tcg_out32(s, insn | (vece << 18) | (q << 6) |
1325              encode_vd(d) | encode_vm(m));
1326}
1327
1328static void tcg_out_vreg3(TCGContext *s, ARMInsn insn, int q, int vece,
1329                          TCGReg d, TCGReg n, TCGReg m)
1330{
1331    tcg_out32(s, insn | (vece << 20) | (q << 6) |
1332              encode_vd(d) | encode_vn(n) | encode_vm(m));
1333}
1334
1335static void tcg_out_vmovi(TCGContext *s, TCGReg rd,
1336                          int q, int op, int cmode, uint8_t imm8)
1337{
1338    tcg_out32(s, INSN_VMOVI | encode_vd(rd) | (q << 6) | (op << 5)
1339              | (cmode << 8) | extract32(imm8, 0, 4)
1340              | (extract32(imm8, 4, 3) << 16)
1341              | (extract32(imm8, 7, 1) << 24));
1342}
1343
1344static void tcg_out_vshifti(TCGContext *s, ARMInsn insn, int q,
1345                            TCGReg rd, TCGReg rm, int l_imm6)
1346{
1347    tcg_out32(s, insn | (q << 6) | encode_vd(rd) | encode_vm(rm) |
1348              (extract32(l_imm6, 6, 1) << 7) |
1349              (extract32(l_imm6, 0, 6) << 16));
1350}
1351
1352static void tcg_out_vldst(TCGContext *s, ARMInsn insn,
1353                          TCGReg rd, TCGReg rn, int offset)
1354{
1355    if (offset != 0) {
1356        if (check_fit_imm(offset) || check_fit_imm(-offset)) {
1357            tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
1358                            TCG_REG_TMP, rn, offset, true);
1359        } else {
1360            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP, offset);
1361            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1362                            TCG_REG_TMP, TCG_REG_TMP, rn, 0);
1363        }
1364        rn = TCG_REG_TMP;
1365    }
1366    tcg_out32(s, insn | (rn << 16) | encode_vd(rd) | 0xf);
1367}
1368
1369typedef struct {
1370    ARMCond cond;
1371    TCGReg base;
1372    int index;
1373    bool index_scratch;
1374    TCGAtomAlign aa;
1375} HostAddress;
1376
1377bool tcg_target_has_memory_bswap(MemOp memop)
1378{
1379    return false;
1380}
1381
1382static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
1383{
1384    /* We arrive at the slow path via "BLNE", so R14 contains l->raddr. */
1385    return TCG_REG_R14;
1386}
1387
1388static const TCGLdstHelperParam ldst_helper_param = {
1389    .ra_gen = ldst_ra_gen,
1390    .ntmp = 1,
1391    .tmp = { TCG_REG_TMP },
1392};
1393
1394static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1395{
1396    MemOp opc = get_memop(lb->oi);
1397
1398    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1399        return false;
1400    }
1401
1402    tcg_out_ld_helper_args(s, lb, &ldst_helper_param);
1403    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE]);
1404    tcg_out_ld_helper_ret(s, lb, false, &ldst_helper_param);
1405
1406    tcg_out_goto(s, COND_AL, lb->raddr);
1407    return true;
1408}
1409
1410static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1411{
1412    MemOp opc = get_memop(lb->oi);
1413
1414    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1415        return false;
1416    }
1417
1418    tcg_out_st_helper_args(s, lb, &ldst_helper_param);
1419
1420    /* Tail-call to the helper, which will return to the fast path.  */
1421    tcg_out_goto(s, COND_AL, qemu_st_helpers[opc & MO_SIZE]);
1422    return true;
1423}
1424
1425/* We expect to use an 9-bit sign-magnitude negative offset from ENV.  */
1426#define MIN_TLB_MASK_TABLE_OFS  -256
1427
1428static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
1429                                           TCGReg addr, MemOpIdx oi, bool is_ld)
1430{
1431    TCGLabelQemuLdst *ldst = NULL;
1432    MemOp opc = get_memop(oi);
1433    unsigned a_mask;
1434
1435    if (tcg_use_softmmu) {
1436        *h = (HostAddress){
1437            .cond = COND_AL,
1438            .base = addr,
1439            .index = TCG_REG_R1,
1440            .index_scratch = true,
1441        };
1442    } else {
1443        *h = (HostAddress){
1444            .cond = COND_AL,
1445            .base = addr,
1446            .index = guest_base ? TCG_REG_GUEST_BASE : -1,
1447            .index_scratch = false,
1448        };
1449    }
1450
1451    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, false);
1452    a_mask = (1 << h->aa.align) - 1;
1453
1454    if (tcg_use_softmmu) {
1455        int mem_index = get_mmuidx(oi);
1456        int cmp_off = is_ld ? offsetof(CPUTLBEntry, addr_read)
1457                            : offsetof(CPUTLBEntry, addr_write);
1458        int fast_off = tlb_mask_table_ofs(s, mem_index);
1459        unsigned s_mask = (1 << (opc & MO_SIZE)) - 1;
1460        TCGReg t_addr;
1461
1462        ldst = new_ldst_label(s);
1463        ldst->is_ld = is_ld;
1464        ldst->oi = oi;
1465        ldst->addr_reg = addr;
1466
1467        /* Load cpu->neg.tlb.f[mmu_idx].{mask,table} into {r0,r1}.  */
1468        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, mask) != 0);
1469        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, table) != 4);
1470        tcg_out_ldrd_8(s, COND_AL, TCG_REG_R0, TCG_AREG0, fast_off);
1471
1472        /* Extract the tlb index from the address into R0.  */
1473        tcg_out_dat_reg(s, COND_AL, ARITH_AND, TCG_REG_R0, TCG_REG_R0, addr,
1474                        SHIFT_IMM_LSR(s->page_bits - CPU_TLB_ENTRY_BITS));
1475
1476        /*
1477         * Add the tlb_table pointer, creating the CPUTLBEntry address in R1.
1478         * Load the tlb comparator into R2 and the fast path addend into R1.
1479         */
1480        if (cmp_off == 0) {
1481            tcg_out_ld32_rwb(s, COND_AL, TCG_REG_R2, TCG_REG_R1, TCG_REG_R0);
1482        } else {
1483            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1484                            TCG_REG_R1, TCG_REG_R1, TCG_REG_R0, 0);
1485            tcg_out_ld32_12(s, COND_AL, TCG_REG_R2, TCG_REG_R1, cmp_off);
1486        }
1487
1488        /* Load the tlb addend.  */
1489        tcg_out_ld32_12(s, COND_AL, TCG_REG_R1, TCG_REG_R1,
1490                        offsetof(CPUTLBEntry, addend));
1491
1492        /*
1493         * Check alignment, check comparators.
1494         * Do this in 2-4 insns.  Use MOVW for v7, if possible,
1495         * to reduce the number of sequential conditional instructions.
1496         * Almost all guests have at least 4k pages, which means that we need
1497         * to clear at least 9 bits even for an 8-byte memory, which means it
1498         * isn't worth checking for an immediate operand for BIC.
1499         *
1500         * For unaligned accesses, test the page of the last unit of alignment.
1501         * This leaves the least significant alignment bits unchanged, and of
1502         * course must be zero.
1503         */
1504        t_addr = addr;
1505        if (a_mask < s_mask) {
1506            t_addr = TCG_REG_R0;
1507            tcg_out_dat_imm(s, COND_AL, ARITH_ADD, t_addr,
1508                            addr, s_mask - a_mask);
1509        }
1510        if (use_armv7_instructions && s->page_bits <= 16) {
1511            tcg_out_movi32(s, COND_AL, TCG_REG_TMP, ~(s->page_mask | a_mask));
1512            tcg_out_dat_reg(s, COND_AL, ARITH_BIC, TCG_REG_TMP,
1513                            t_addr, TCG_REG_TMP, 0);
1514            tcg_out_dat_reg(s, COND_AL, ARITH_CMP, 0,
1515                            TCG_REG_R2, TCG_REG_TMP, 0);
1516        } else {
1517            if (a_mask) {
1518                tcg_debug_assert(a_mask <= 0xff);
1519                tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addr, a_mask);
1520            }
1521            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0, t_addr,
1522                            SHIFT_IMM_LSR(s->page_bits));
1523            tcg_out_dat_reg(s, (a_mask ? COND_EQ : COND_AL), ARITH_CMP,
1524                            0, TCG_REG_R2, TCG_REG_TMP,
1525                            SHIFT_IMM_LSL(s->page_bits));
1526        }
1527    } else if (a_mask) {
1528        ldst = new_ldst_label(s);
1529        ldst->is_ld = is_ld;
1530        ldst->oi = oi;
1531        ldst->addr_reg = addr;
1532
1533        /* We are expecting alignment to max out at 7 */
1534        tcg_debug_assert(a_mask <= 0xff);
1535        /* tst addr, #mask */
1536        tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addr, a_mask);
1537    }
1538
1539    return ldst;
1540}
1541
1542static void tcg_out_qemu_ld_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1543                                   TCGReg datahi, HostAddress h)
1544{
1545    TCGReg base;
1546
1547    /* Byte swapping is left to middle-end expansion. */
1548    tcg_debug_assert((opc & MO_BSWAP) == 0);
1549
1550    switch (opc & MO_SSIZE) {
1551    case MO_UB:
1552        if (h.index < 0) {
1553            tcg_out_ld8_12(s, h.cond, datalo, h.base, 0);
1554        } else {
1555            tcg_out_ld8_r(s, h.cond, datalo, h.base, h.index);
1556        }
1557        break;
1558    case MO_SB:
1559        if (h.index < 0) {
1560            tcg_out_ld8s_8(s, h.cond, datalo, h.base, 0);
1561        } else {
1562            tcg_out_ld8s_r(s, h.cond, datalo, h.base, h.index);
1563        }
1564        break;
1565    case MO_UW:
1566        if (h.index < 0) {
1567            tcg_out_ld16u_8(s, h.cond, datalo, h.base, 0);
1568        } else {
1569            tcg_out_ld16u_r(s, h.cond, datalo, h.base, h.index);
1570        }
1571        break;
1572    case MO_SW:
1573        if (h.index < 0) {
1574            tcg_out_ld16s_8(s, h.cond, datalo, h.base, 0);
1575        } else {
1576            tcg_out_ld16s_r(s, h.cond, datalo, h.base, h.index);
1577        }
1578        break;
1579    case MO_UL:
1580        if (h.index < 0) {
1581            tcg_out_ld32_12(s, h.cond, datalo, h.base, 0);
1582        } else {
1583            tcg_out_ld32_r(s, h.cond, datalo, h.base, h.index);
1584        }
1585        break;
1586    case MO_UQ:
1587        /* We used pair allocation for datalo, so already should be aligned. */
1588        tcg_debug_assert((datalo & 1) == 0);
1589        tcg_debug_assert(datahi == datalo + 1);
1590        /* LDRD requires alignment; double-check that. */
1591        if (memop_alignment_bits(opc) >= MO_64) {
1592            if (h.index < 0) {
1593                tcg_out_ldrd_8(s, h.cond, datalo, h.base, 0);
1594                break;
1595            }
1596            /*
1597             * Rm (the second address op) must not overlap Rt or Rt + 1.
1598             * Since datalo is aligned, we can simplify the test via alignment.
1599             * Flip the two address arguments if that works.
1600             */
1601            if ((h.index & ~1) != datalo) {
1602                tcg_out_ldrd_r(s, h.cond, datalo, h.base, h.index);
1603                break;
1604            }
1605            if ((h.base & ~1) != datalo) {
1606                tcg_out_ldrd_r(s, h.cond, datalo, h.index, h.base);
1607                break;
1608            }
1609        }
1610        if (h.index < 0) {
1611            base = h.base;
1612            if (datalo == h.base) {
1613                tcg_out_mov_reg(s, h.cond, TCG_REG_TMP, base);
1614                base = TCG_REG_TMP;
1615            }
1616        } else if (h.index_scratch) {
1617            tcg_out_ld32_rwb(s, h.cond, datalo, h.index, h.base);
1618            tcg_out_ld32_12(s, h.cond, datahi, h.index, 4);
1619            break;
1620        } else {
1621            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1622                            h.base, h.index, SHIFT_IMM_LSL(0));
1623            base = TCG_REG_TMP;
1624        }
1625        tcg_out_ld32_12(s, h.cond, datalo, base, 0);
1626        tcg_out_ld32_12(s, h.cond, datahi, base, 4);
1627        break;
1628    default:
1629        g_assert_not_reached();
1630    }
1631}
1632
1633static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
1634                            TCGReg addr, MemOpIdx oi, TCGType data_type)
1635{
1636    MemOp opc = get_memop(oi);
1637    TCGLabelQemuLdst *ldst;
1638    HostAddress h;
1639
1640    ldst = prepare_host_addr(s, &h, addr, oi, true);
1641    if (ldst) {
1642        ldst->type = data_type;
1643        ldst->datalo_reg = datalo;
1644        ldst->datahi_reg = datahi;
1645
1646        /*
1647         * This a conditional BL only to load a pointer within this
1648         * opcode into LR for the slow path.  We will not be using
1649         * the value for a tail call.
1650         */
1651        ldst->label_ptr[0] = s->code_ptr;
1652        tcg_out_bl_imm(s, COND_NE, 0);
1653
1654        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1655        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1656    } else {
1657        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1658    }
1659}
1660
1661static void tcg_out_qemu_st_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1662                                   TCGReg datahi, HostAddress h)
1663{
1664    /* Byte swapping is left to middle-end expansion. */
1665    tcg_debug_assert((opc & MO_BSWAP) == 0);
1666
1667    switch (opc & MO_SIZE) {
1668    case MO_8:
1669        if (h.index < 0) {
1670            tcg_out_st8_12(s, h.cond, datalo, h.base, 0);
1671        } else {
1672            tcg_out_st8_r(s, h.cond, datalo, h.base, h.index);
1673        }
1674        break;
1675    case MO_16:
1676        if (h.index < 0) {
1677            tcg_out_st16_8(s, h.cond, datalo, h.base, 0);
1678        } else {
1679            tcg_out_st16_r(s, h.cond, datalo, h.base, h.index);
1680        }
1681        break;
1682    case MO_32:
1683        if (h.index < 0) {
1684            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1685        } else {
1686            tcg_out_st32_r(s, h.cond, datalo, h.base, h.index);
1687        }
1688        break;
1689    case MO_64:
1690        /* We used pair allocation for datalo, so already should be aligned. */
1691        tcg_debug_assert((datalo & 1) == 0);
1692        tcg_debug_assert(datahi == datalo + 1);
1693        /* STRD requires alignment; double-check that. */
1694        if (memop_alignment_bits(opc) >= MO_64) {
1695            if (h.index < 0) {
1696                tcg_out_strd_8(s, h.cond, datalo, h.base, 0);
1697            } else {
1698                tcg_out_strd_r(s, h.cond, datalo, h.base, h.index);
1699            }
1700        } else if (h.index < 0) {
1701            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1702            tcg_out_st32_12(s, h.cond, datahi, h.base, 4);
1703        } else if (h.index_scratch) {
1704            tcg_out_st32_rwb(s, h.cond, datalo, h.index, h.base);
1705            tcg_out_st32_12(s, h.cond, datahi, h.index, 4);
1706        } else {
1707            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1708                            h.base, h.index, SHIFT_IMM_LSL(0));
1709            tcg_out_st32_12(s, h.cond, datalo, TCG_REG_TMP, 0);
1710            tcg_out_st32_12(s, h.cond, datahi, TCG_REG_TMP, 4);
1711        }
1712        break;
1713    default:
1714        g_assert_not_reached();
1715    }
1716}
1717
1718static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
1719                            TCGReg addr, MemOpIdx oi, TCGType data_type)
1720{
1721    MemOp opc = get_memop(oi);
1722    TCGLabelQemuLdst *ldst;
1723    HostAddress h;
1724
1725    ldst = prepare_host_addr(s, &h, addr, oi, false);
1726    if (ldst) {
1727        ldst->type = data_type;
1728        ldst->datalo_reg = datalo;
1729        ldst->datahi_reg = datahi;
1730
1731        h.cond = COND_EQ;
1732        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1733
1734        /* The conditional call is last, as we're going to return here. */
1735        ldst->label_ptr[0] = s->code_ptr;
1736        tcg_out_bl_imm(s, COND_NE, 0);
1737        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1738    } else {
1739        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1740    }
1741}
1742
1743static void tcg_out_epilogue(TCGContext *s);
1744
1745static void tcg_out_exit_tb(TCGContext *s, uintptr_t arg)
1746{
1747    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, arg);
1748    tcg_out_epilogue(s);
1749}
1750
1751static void tcg_out_goto_tb(TCGContext *s, int which)
1752{
1753    uintptr_t i_addr;
1754    intptr_t i_disp;
1755
1756    /* Direct branch will be patched by tb_target_set_jmp_target. */
1757    set_jmp_insn_offset(s, which);
1758    tcg_out32(s, INSN_NOP);
1759
1760    /* When branch is out of range, fall through to indirect. */
1761    i_addr = get_jmp_target_addr(s, which);
1762    i_disp = tcg_pcrel_diff(s, (void *)i_addr) - 8;
1763    tcg_debug_assert(i_disp < 0);
1764    if (i_disp >= -0xfff) {
1765        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_PC, i_disp);
1766    } else {
1767        /*
1768         * The TB is close, but outside the 12 bits addressable by
1769         * the load.  We can extend this to 20 bits with a sub of a
1770         * shifted immediate from pc.
1771         */
1772        int h = -i_disp;
1773        int l = -(h & 0xfff);
1774
1775        h = encode_imm_nofail(h + l);
1776        tcg_out_dat_imm(s, COND_AL, ARITH_SUB, TCG_REG_R0, TCG_REG_PC, h);
1777        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_R0, l);
1778    }
1779    set_jmp_reset_offset(s, which);
1780}
1781
1782void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1783                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1784{
1785    uintptr_t addr = tb->jmp_target_addr[n];
1786    ptrdiff_t offset = addr - (jmp_rx + 8);
1787    tcg_insn_unit insn;
1788
1789    /* Either directly branch, or fall through to indirect branch. */
1790    if (offset == sextract64(offset, 0, 26)) {
1791        /* B <addr> */
1792        insn = deposit32((COND_AL << 28) | INSN_B, 0, 24, offset >> 2);
1793    } else {
1794        insn = INSN_NOP;
1795    }
1796
1797    qatomic_set((uint32_t *)jmp_rw, insn);
1798    flush_idcache_range(jmp_rx, jmp_rw, 4);
1799}
1800
1801
1802static void tgen_add(TCGContext *s, TCGType type,
1803                     TCGReg a0, TCGReg a1, TCGReg a2)
1804{
1805    tcg_out_dat_reg(s, COND_AL, ARITH_ADD, a0, a1, a2, SHIFT_IMM_LSL(0));
1806}
1807
1808static void tgen_addi(TCGContext *s, TCGType type,
1809                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1810{
1811    tcg_out_dat_IN(s, COND_AL, ARITH_ADD, ARITH_SUB, a0, a1, a2);
1812}
1813
1814static const TCGOutOpBinary outop_add = {
1815    .base.static_constraint = C_O1_I2(r, r, rIN),
1816    .out_rrr = tgen_add,
1817    .out_rri = tgen_addi,
1818};
1819
1820static void tgen_and(TCGContext *s, TCGType type,
1821                     TCGReg a0, TCGReg a1, TCGReg a2)
1822{
1823    tcg_out_dat_reg(s, COND_AL, ARITH_AND, a0, a1, a2, SHIFT_IMM_LSL(0));
1824}
1825
1826static void tgen_andi(TCGContext *s, TCGType type,
1827                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1828{
1829    tcg_out_dat_IK(s, COND_AL, ARITH_AND, ARITH_BIC, a0, a1, a2);
1830}
1831
1832static const TCGOutOpBinary outop_and = {
1833    .base.static_constraint = C_O1_I2(r, r, rIK),
1834    .out_rrr = tgen_and,
1835    .out_rri = tgen_andi,
1836};
1837
1838static void tgen_andc(TCGContext *s, TCGType type,
1839                      TCGReg a0, TCGReg a1, TCGReg a2)
1840{
1841    tcg_out_dat_reg(s, COND_AL, ARITH_BIC, a0, a1, a2, SHIFT_IMM_LSL(0));
1842}
1843
1844static const TCGOutOpBinary outop_andc = {
1845    .base.static_constraint = C_O1_I2(r, r, r),
1846    .out_rrr = tgen_andc,
1847};
1848
1849static void tgen_clz(TCGContext *s, TCGType type,
1850                     TCGReg a0, TCGReg a1, TCGReg a2)
1851{
1852    tcg_out_dat_imm(s, COND_AL, ARITH_CMP, 0, a1, 0);
1853    tcg_out_dat_reg(s, COND_NE, INSN_CLZ, a0, 0, a1, 0);
1854    tcg_out_mov_reg(s, COND_EQ, a0, a2);
1855}
1856
1857static void tgen_clzi(TCGContext *s, TCGType type,
1858                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1859{
1860    if (a2 == 32) {
1861        tcg_out_dat_reg(s, COND_AL, INSN_CLZ, a0, 0, a1, 0);
1862    } else {
1863        tcg_out_dat_imm(s, COND_AL, ARITH_CMP, 0, a1, 0);
1864        tcg_out_dat_reg(s, COND_NE, INSN_CLZ, a0, 0, a1, 0);
1865        tcg_out_movi32(s, COND_EQ, a0, a2);
1866    }
1867}
1868
1869static const TCGOutOpBinary outop_clz = {
1870    .base.static_constraint = C_O1_I2(r, r, rIK),
1871    .out_rrr = tgen_clz,
1872    .out_rri = tgen_clzi,
1873};
1874
1875static const TCGOutOpUnary outop_ctpop = {
1876    .base.static_constraint = C_NotImplemented,
1877};
1878
1879static void tgen_ctz(TCGContext *s, TCGType type,
1880                     TCGReg a0, TCGReg a1, TCGReg a2)
1881{
1882    tcg_out_dat_reg(s, COND_AL, INSN_RBIT, TCG_REG_TMP, 0, a1, 0);
1883    tgen_clz(s, TCG_TYPE_I32, a0, TCG_REG_TMP, a2);
1884}
1885
1886static void tgen_ctzi(TCGContext *s, TCGType type,
1887                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1888{
1889    tcg_out_dat_reg(s, COND_AL, INSN_RBIT, TCG_REG_TMP, 0, a1, 0);
1890    tgen_clzi(s, TCG_TYPE_I32, a0, TCG_REG_TMP, a2);
1891}
1892
1893static TCGConstraintSetIndex cset_ctz(TCGType type, unsigned flags)
1894{
1895    return use_armv7_instructions ? C_O1_I2(r, r, rIK) : C_NotImplemented;
1896}
1897
1898static const TCGOutOpBinary outop_ctz = {
1899    .base.static_constraint = C_Dynamic,
1900    .base.dynamic_constraint = cset_ctz,
1901    .out_rrr = tgen_ctz,
1902    .out_rri = tgen_ctzi,
1903};
1904
1905static TCGConstraintSetIndex cset_idiv(TCGType type, unsigned flags)
1906{
1907    return use_idiv_instructions ? C_O1_I2(r, r, r) : C_NotImplemented;
1908}
1909
1910static void tgen_divs(TCGContext *s, TCGType type,
1911                      TCGReg a0, TCGReg a1, TCGReg a2)
1912{
1913    /* sdiv */
1914    tcg_out32(s, 0x0710f010 | (COND_AL << 28) | (a0 << 16) | a1 | (a2 << 8));
1915}
1916
1917static const TCGOutOpBinary outop_divs = {
1918    .base.static_constraint = C_Dynamic,
1919    .base.dynamic_constraint = cset_idiv,
1920    .out_rrr = tgen_divs,
1921};
1922
1923static const TCGOutOpDivRem outop_divs2 = {
1924    .base.static_constraint = C_NotImplemented,
1925};
1926
1927static void tgen_divu(TCGContext *s, TCGType type,
1928                      TCGReg a0, TCGReg a1, TCGReg a2)
1929{
1930    /* udiv */
1931    tcg_out32(s, 0x0730f010 | (COND_AL << 28) | (a0 << 16) | a1 | (a2 << 8));
1932}
1933
1934static const TCGOutOpBinary outop_divu = {
1935    .base.static_constraint = C_Dynamic,
1936    .base.dynamic_constraint = cset_idiv,
1937    .out_rrr = tgen_divu,
1938};
1939
1940static const TCGOutOpDivRem outop_divu2 = {
1941    .base.static_constraint = C_NotImplemented,
1942};
1943
1944static const TCGOutOpBinary outop_eqv = {
1945    .base.static_constraint = C_NotImplemented,
1946};
1947
1948static void tgen_mul(TCGContext *s, TCGType type,
1949                     TCGReg a0, TCGReg a1, TCGReg a2)
1950{
1951    /* mul */
1952    tcg_out32(s, (COND_AL << 28) | 0x90 | (a0 << 16) | (a1 << 8) | a2);
1953}
1954
1955static const TCGOutOpBinary outop_mul = {
1956    .base.static_constraint = C_O1_I2(r, r, r),
1957    .out_rrr = tgen_mul,
1958};
1959
1960static void tgen_muls2(TCGContext *s, TCGType type,
1961                       TCGReg rd0, TCGReg rd1, TCGReg rn, TCGReg rm)
1962{
1963    /* smull */
1964    tcg_out32(s, (COND_AL << 28) | 0x00c00090 |
1965              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
1966}
1967
1968static const TCGOutOpMul2 outop_muls2 = {
1969    .base.static_constraint = C_O2_I2(r, r, r, r),
1970    .out_rrrr = tgen_muls2,
1971};
1972
1973static const TCGOutOpBinary outop_mulsh = {
1974    .base.static_constraint = C_NotImplemented,
1975};
1976
1977static void tgen_mulu2(TCGContext *s, TCGType type,
1978                       TCGReg rd0, TCGReg rd1, TCGReg rn, TCGReg rm)
1979{
1980    /* umull */
1981    tcg_out32(s, (COND_AL << 28) | 0x00800090 |
1982              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
1983}
1984
1985static const TCGOutOpMul2 outop_mulu2 = {
1986    .base.static_constraint = C_O2_I2(r, r, r, r),
1987    .out_rrrr = tgen_mulu2,
1988};
1989
1990static const TCGOutOpBinary outop_muluh = {
1991    .base.static_constraint = C_NotImplemented,
1992};
1993
1994static const TCGOutOpBinary outop_nand = {
1995    .base.static_constraint = C_NotImplemented,
1996};
1997
1998static const TCGOutOpBinary outop_nor = {
1999    .base.static_constraint = C_NotImplemented,
2000};
2001
2002static void tgen_or(TCGContext *s, TCGType type,
2003                     TCGReg a0, TCGReg a1, TCGReg a2)
2004{
2005    tcg_out_dat_reg(s, COND_AL, ARITH_ORR, a0, a1, a2, SHIFT_IMM_LSL(0));
2006}
2007
2008static void tgen_ori(TCGContext *s, TCGType type,
2009                     TCGReg a0, TCGReg a1, tcg_target_long a2)
2010{
2011    tcg_out_dat_imm(s, COND_AL, ARITH_ORR, a0, a1, encode_imm_nofail(a2));
2012}
2013
2014static const TCGOutOpBinary outop_or = {
2015    .base.static_constraint = C_O1_I2(r, r, rI),
2016    .out_rrr = tgen_or,
2017    .out_rri = tgen_ori,
2018};
2019
2020static const TCGOutOpBinary outop_orc = {
2021    .base.static_constraint = C_NotImplemented,
2022};
2023
2024static const TCGOutOpBinary outop_rems = {
2025    .base.static_constraint = C_NotImplemented,
2026};
2027
2028static const TCGOutOpBinary outop_remu = {
2029    .base.static_constraint = C_NotImplemented,
2030};
2031
2032static const TCGOutOpBinary outop_rotl = {
2033    .base.static_constraint = C_NotImplemented,
2034};
2035
2036static void tgen_rotr(TCGContext *s, TCGType type,
2037                      TCGReg a0, TCGReg a1, TCGReg a2)
2038{
2039    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1, SHIFT_REG_ROR(a2));
2040}
2041
2042static void tgen_rotri(TCGContext *s, TCGType type,
2043                       TCGReg a0, TCGReg a1, tcg_target_long a2)
2044{
2045    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1, SHIFT_IMM_ROR(a2 & 0x1f));
2046}
2047
2048static const TCGOutOpBinary outop_rotr = {
2049    .base.static_constraint = C_O1_I2(r, r, ri),
2050    .out_rrr = tgen_rotr,
2051    .out_rri = tgen_rotri,
2052};
2053
2054static void tgen_sar(TCGContext *s, TCGType type,
2055                     TCGReg a0, TCGReg a1, TCGReg a2)
2056{
2057    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1, SHIFT_REG_ASR(a2));
2058}
2059
2060static void tgen_sari(TCGContext *s, TCGType type,
2061                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2062{
2063    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1,
2064                    SHIFT_IMM_ASR(a2 & 0x1f));
2065}
2066
2067static const TCGOutOpBinary outop_sar = {
2068    .base.static_constraint = C_O1_I2(r, r, ri),
2069    .out_rrr = tgen_sar,
2070    .out_rri = tgen_sari,
2071};
2072
2073static void tgen_shl(TCGContext *s, TCGType type,
2074                     TCGReg a0, TCGReg a1, TCGReg a2)
2075{
2076    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1, SHIFT_REG_LSL(a2));
2077}
2078
2079static void tgen_shli(TCGContext *s, TCGType type,
2080                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2081{
2082    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1,
2083                    SHIFT_IMM_LSL(a2 & 0x1f));
2084}
2085
2086static const TCGOutOpBinary outop_shl = {
2087    .base.static_constraint = C_O1_I2(r, r, ri),
2088    .out_rrr = tgen_shl,
2089    .out_rri = tgen_shli,
2090};
2091
2092static void tgen_shr(TCGContext *s, TCGType type,
2093                     TCGReg a0, TCGReg a1, TCGReg a2)
2094{
2095    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1, SHIFT_REG_LSR(a2));
2096}
2097
2098static void tgen_shri(TCGContext *s, TCGType type,
2099                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2100{
2101    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1,
2102                    SHIFT_IMM_LSR(a2 & 0x1f));
2103}
2104
2105static const TCGOutOpBinary outop_shr = {
2106    .base.static_constraint = C_O1_I2(r, r, ri),
2107    .out_rrr = tgen_shr,
2108    .out_rri = tgen_shri,
2109};
2110
2111static void tgen_sub(TCGContext *s, TCGType type,
2112                     TCGReg a0, TCGReg a1, TCGReg a2)
2113{
2114    tcg_out_dat_reg(s, COND_AL, ARITH_SUB, a0, a1, a2, SHIFT_IMM_LSL(0));
2115}
2116
2117static void tgen_subfi(TCGContext *s, TCGType type,
2118                       TCGReg a0, tcg_target_long a1, TCGReg a2)
2119{
2120    tcg_out_dat_imm(s, COND_AL, ARITH_RSB, a0, a2, encode_imm_nofail(a1));
2121}
2122
2123static const TCGOutOpSubtract outop_sub = {
2124    .base.static_constraint = C_O1_I2(r, rI, r),
2125    .out_rrr = tgen_sub,
2126    .out_rir = tgen_subfi,
2127};
2128
2129static void tgen_xor(TCGContext *s, TCGType type,
2130                     TCGReg a0, TCGReg a1, TCGReg a2)
2131{
2132    tcg_out_dat_reg(s, COND_AL, ARITH_EOR, a0, a1, a2, SHIFT_IMM_LSL(0));
2133}
2134
2135static void tgen_xori(TCGContext *s, TCGType type,
2136                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2137{
2138    tcg_out_dat_imm(s, COND_AL, ARITH_EOR, a0, a1, encode_imm_nofail(a2));
2139}
2140
2141static const TCGOutOpBinary outop_xor = {
2142    .base.static_constraint = C_O1_I2(r, r, rI),
2143    .out_rrr = tgen_xor,
2144    .out_rri = tgen_xori,
2145};
2146
2147static void tgen_bswap16(TCGContext *s, TCGType type,
2148                         TCGReg rd, TCGReg rn, unsigned flags)
2149{
2150    if (flags & TCG_BSWAP_OS) {
2151        /* revsh */
2152        tcg_out32(s, 0x06ff0fb0 | (COND_AL << 28) | (rd << 12) | rn);
2153        return;
2154    }
2155
2156    /* rev16 */
2157    tcg_out32(s, 0x06bf0fb0 | (COND_AL << 28) | (rd << 12) | rn);
2158    if ((flags & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
2159        tcg_out_ext16u(s, rd, rd);
2160    }
2161}
2162
2163static const TCGOutOpBswap outop_bswap16 = {
2164    .base.static_constraint = C_O1_I1(r, r),
2165    .out_rr = tgen_bswap16,
2166};
2167
2168static void tgen_bswap32(TCGContext *s, TCGType type,
2169                         TCGReg rd, TCGReg rn, unsigned flags)
2170{
2171    /* rev */
2172    tcg_out32(s, 0x06bf0f30 | (COND_AL << 28) | (rd << 12) | rn);
2173}
2174
2175static const TCGOutOpBswap outop_bswap32 = {
2176    .base.static_constraint = C_O1_I1(r, r),
2177    .out_rr = tgen_bswap32,
2178};
2179
2180static const TCGOutOpUnary outop_bswap64 = {
2181    .base.static_constraint = C_NotImplemented,
2182};
2183
2184static void tgen_neg(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
2185{
2186    tgen_subfi(s, type, a0, 0, a1);
2187}
2188
2189static const TCGOutOpUnary outop_neg = {
2190    .base.static_constraint = C_O1_I1(r, r),
2191    .out_rr = tgen_neg,
2192};
2193
2194static void tgen_not(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
2195{
2196    tcg_out_dat_reg(s, COND_AL, ARITH_MVN, a0, 0, a1, SHIFT_IMM_LSL(0));
2197}
2198
2199static const TCGOutOpUnary outop_not = {
2200    .base.static_constraint = C_O1_I1(r, r),
2201    .out_rr = tgen_not,
2202};
2203
2204static void tgen_brcond(TCGContext *s, TCGType type, TCGCond cond,
2205                        TCGReg a0, TCGReg a1, TCGLabel *l)
2206{
2207    cond = tgen_cmp(s, cond, a0, a1);
2208    tcg_out_goto_label(s, tcg_cond_to_arm_cond[cond], l);
2209}
2210
2211static void tgen_brcondi(TCGContext *s, TCGType type, TCGCond cond,
2212                         TCGReg a0, tcg_target_long a1, TCGLabel *l)
2213{
2214    cond = tgen_cmpi(s, cond, a0, a1);
2215    tcg_out_goto_label(s, tcg_cond_to_arm_cond[cond], l);
2216}
2217
2218static const TCGOutOpBrcond outop_brcond = {
2219    .base.static_constraint = C_O0_I2(r, rIN),
2220    .out_rr = tgen_brcond,
2221    .out_ri = tgen_brcondi,
2222};
2223
2224static void finish_setcond(TCGContext *s, TCGCond cond, TCGReg ret, bool neg)
2225{
2226    tcg_out_movi32(s, tcg_cond_to_arm_cond[tcg_invert_cond(cond)], ret, 0);
2227    tcg_out_movi32(s, tcg_cond_to_arm_cond[cond], ret, neg ? -1 : 1);
2228}
2229
2230static void tgen_setcond(TCGContext *s, TCGType type, TCGCond cond,
2231                         TCGReg a0, TCGReg a1, TCGReg a2)
2232{
2233    cond = tgen_cmp(s, cond, a1, a2);
2234    finish_setcond(s, cond, a0, false);
2235}
2236
2237static void tgen_setcondi(TCGContext *s, TCGType type, TCGCond cond,
2238                          TCGReg a0, TCGReg a1, tcg_target_long a2)
2239{
2240    cond = tgen_cmpi(s, cond, a1, a2);
2241    finish_setcond(s, cond, a0, false);
2242}
2243
2244static const TCGOutOpSetcond outop_setcond = {
2245    .base.static_constraint = C_O1_I2(r, r, rIN),
2246    .out_rrr = tgen_setcond,
2247    .out_rri = tgen_setcondi,
2248};
2249
2250static void tgen_negsetcond(TCGContext *s, TCGType type, TCGCond cond,
2251                            TCGReg a0, TCGReg a1, TCGReg a2)
2252{
2253    cond = tgen_cmp(s, cond, a1, a2);
2254    finish_setcond(s, cond, a0, true);
2255}
2256
2257static void tgen_negsetcondi(TCGContext *s, TCGType type, TCGCond cond,
2258                             TCGReg a0, TCGReg a1, tcg_target_long a2)
2259{
2260    cond = tgen_cmpi(s, cond, a1, a2);
2261    finish_setcond(s, cond, a0, true);
2262}
2263
2264static const TCGOutOpSetcond outop_negsetcond = {
2265    .base.static_constraint = C_O1_I2(r, r, rIN),
2266    .out_rrr = tgen_negsetcond,
2267    .out_rri = tgen_negsetcondi,
2268};
2269
2270static void tgen_movcond(TCGContext *s, TCGType type, TCGCond cond,
2271                         TCGReg ret, TCGReg c1, TCGArg c2, bool const_c2,
2272                         TCGArg vt, bool const_vt, TCGArg vf, bool consf_vf)
2273{
2274    cond = tcg_out_cmp(s, cond, c1, c2, const_c2);
2275    tcg_out_dat_rIK(s, tcg_cond_to_arm_cond[cond], ARITH_MOV, ARITH_MVN,
2276                    ret, 0, vt, const_vt);
2277}
2278
2279static const TCGOutOpMovcond outop_movcond = {
2280    .base.static_constraint = C_O1_I4(r, r, rIN, rIK, 0),
2281    .out = tgen_movcond,
2282};
2283
2284static void tgen_brcond2(TCGContext *s, TCGCond cond, TCGReg al, TCGReg ah,
2285                         TCGArg bl, bool const_bl, TCGArg bh, bool const_bh,
2286                         TCGLabel *l)
2287{
2288    cond = tcg_out_cmp2(s, cond, al, ah, bl, const_bl, bh, const_bh);
2289    tcg_out_goto_label(s, tcg_cond_to_arm_cond[cond], l);
2290}
2291
2292static const TCGOutOpBrcond2 outop_brcond2 = {
2293    .base.static_constraint = C_O0_I4(r, r, rI, rI),
2294    .out = tgen_brcond2,
2295};
2296
2297static void tgen_setcond2(TCGContext *s, TCGCond cond, TCGReg ret,
2298                          TCGReg al, TCGReg ah,
2299                          TCGArg bl, bool const_bl,
2300                          TCGArg bh, bool const_bh)
2301{
2302    cond = tcg_out_cmp2(s, cond, al, ah, bl, const_bl, bh, const_bh);
2303    finish_setcond(s, cond, ret, false);
2304}
2305
2306static const TCGOutOpSetcond2 outop_setcond2 = {
2307    .base.static_constraint = C_O1_I4(r, r, r, rI, rI),
2308    .out = tgen_setcond2,
2309};
2310
2311static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
2312                       const TCGArg args[TCG_MAX_OP_ARGS],
2313                       const int const_args[TCG_MAX_OP_ARGS])
2314{
2315    TCGArg a0, a1, a2, a3, a4, a5;
2316
2317    switch (opc) {
2318    case INDEX_op_goto_ptr:
2319        tcg_out_b_reg(s, COND_AL, args[0]);
2320        break;
2321    case INDEX_op_br:
2322        tcg_out_goto_label(s, COND_AL, arg_label(args[0]));
2323        break;
2324
2325    case INDEX_op_ld8u_i32:
2326        tcg_out_ld8u(s, COND_AL, args[0], args[1], args[2]);
2327        break;
2328    case INDEX_op_ld8s_i32:
2329        tcg_out_ld8s(s, COND_AL, args[0], args[1], args[2]);
2330        break;
2331    case INDEX_op_ld16u_i32:
2332        tcg_out_ld16u(s, COND_AL, args[0], args[1], args[2]);
2333        break;
2334    case INDEX_op_ld16s_i32:
2335        tcg_out_ld16s(s, COND_AL, args[0], args[1], args[2]);
2336        break;
2337    case INDEX_op_ld_i32:
2338        tcg_out_ld32u(s, COND_AL, args[0], args[1], args[2]);
2339        break;
2340    case INDEX_op_st8_i32:
2341        tcg_out_st8(s, COND_AL, args[0], args[1], args[2]);
2342        break;
2343    case INDEX_op_st16_i32:
2344        tcg_out_st16(s, COND_AL, args[0], args[1], args[2]);
2345        break;
2346    case INDEX_op_st_i32:
2347        tcg_out_st32(s, COND_AL, args[0], args[1], args[2]);
2348        break;
2349
2350    case INDEX_op_add2_i32:
2351        a0 = args[0], a1 = args[1], a2 = args[2];
2352        a3 = args[3], a4 = args[4], a5 = args[5];
2353        if (a0 == a3 || (a0 == a5 && !const_args[5])) {
2354            a0 = TCG_REG_TMP;
2355        }
2356        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD | TO_CPSR, ARITH_SUB | TO_CPSR,
2357                        a0, a2, a4, const_args[4]);
2358        tcg_out_dat_rIK(s, COND_AL, ARITH_ADC, ARITH_SBC,
2359                        a1, a3, a5, const_args[5]);
2360        tcg_out_mov_reg(s, COND_AL, args[0], a0);
2361        break;
2362    case INDEX_op_sub2_i32:
2363        a0 = args[0], a1 = args[1], a2 = args[2];
2364        a3 = args[3], a4 = args[4], a5 = args[5];
2365        if ((a0 == a3 && !const_args[3]) || (a0 == a5 && !const_args[5])) {
2366            a0 = TCG_REG_TMP;
2367        }
2368        if (const_args[2]) {
2369            if (const_args[4]) {
2370                tcg_out_movi32(s, COND_AL, a0, a4);
2371                a4 = a0;
2372            }
2373            tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR, a0, a4, a2, 1);
2374        } else {
2375            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB | TO_CPSR,
2376                            ARITH_ADD | TO_CPSR, a0, a2, a4, const_args[4]);
2377        }
2378        if (const_args[3]) {
2379            if (const_args[5]) {
2380                tcg_out_movi32(s, COND_AL, a1, a5);
2381                a5 = a1;
2382            }
2383            tcg_out_dat_rI(s, COND_AL, ARITH_RSC, a1, a5, a3, 1);
2384        } else {
2385            tcg_out_dat_rIK(s, COND_AL, ARITH_SBC, ARITH_ADC,
2386                            a1, a3, a5, const_args[5]);
2387        }
2388        tcg_out_mov_reg(s, COND_AL, args[0], a0);
2389        break;
2390
2391    case INDEX_op_qemu_ld_i32:
2392        tcg_out_qemu_ld(s, args[0], -1, args[1], args[2], TCG_TYPE_I32);
2393        break;
2394    case INDEX_op_qemu_ld_i64:
2395        tcg_out_qemu_ld(s, args[0], args[1], args[2], args[3], TCG_TYPE_I64);
2396        break;
2397
2398    case INDEX_op_qemu_st_i32:
2399        tcg_out_qemu_st(s, args[0], -1, args[1], args[2], TCG_TYPE_I32);
2400        break;
2401    case INDEX_op_qemu_st_i64:
2402        tcg_out_qemu_st(s, args[0], args[1], args[2], args[3], TCG_TYPE_I64);
2403        break;
2404
2405    case INDEX_op_deposit_i32:
2406        tcg_out_deposit(s, COND_AL, args[0], args[2],
2407                        args[3], args[4], const_args[2]);
2408        break;
2409    case INDEX_op_extract2_i32:
2410        /* ??? These optimization vs zero should be generic.  */
2411        /* ??? But we can't substitute 2 for 1 in the opcode stream yet.  */
2412        if (const_args[1]) {
2413            if (const_args[2]) {
2414                tcg_out_movi(s, TCG_TYPE_REG, args[0], 0);
2415            } else {
2416                tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2417                                args[2], SHIFT_IMM_LSL(32 - args[3]));
2418            }
2419        } else if (const_args[2]) {
2420            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2421                            args[1], SHIFT_IMM_LSR(args[3]));
2422        } else {
2423            /* We can do extract2 in 2 insns, vs the 3 required otherwise.  */
2424            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0,
2425                            args[2], SHIFT_IMM_LSL(32 - args[3]));
2426            tcg_out_dat_reg(s, COND_AL, ARITH_ORR, args[0], TCG_REG_TMP,
2427                            args[1], SHIFT_IMM_LSR(args[3]));
2428        }
2429        break;
2430
2431    case INDEX_op_mb:
2432        tcg_out_mb(s, args[0]);
2433        break;
2434
2435    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
2436    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
2437    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
2438    default:
2439        g_assert_not_reached();
2440    }
2441}
2442
2443static TCGConstraintSetIndex
2444tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
2445{
2446    switch (op) {
2447    case INDEX_op_goto_ptr:
2448        return C_O0_I1(r);
2449
2450    case INDEX_op_ld8u_i32:
2451    case INDEX_op_ld8s_i32:
2452    case INDEX_op_ld16u_i32:
2453    case INDEX_op_ld16s_i32:
2454    case INDEX_op_ld_i32:
2455        return C_O1_I1(r, r);
2456
2457    case INDEX_op_st8_i32:
2458    case INDEX_op_st16_i32:
2459    case INDEX_op_st_i32:
2460        return C_O0_I2(r, r);
2461
2462    case INDEX_op_deposit_i32:
2463        return C_O1_I2(r, 0, rZ);
2464    case INDEX_op_extract2_i32:
2465        return C_O1_I2(r, rZ, rZ);
2466    case INDEX_op_add2_i32:
2467        return C_O2_I4(r, r, r, r, rIN, rIK);
2468    case INDEX_op_sub2_i32:
2469        return C_O2_I4(r, r, rI, rI, rIN, rIK);
2470    case INDEX_op_qemu_ld_i32:
2471        return C_O1_I1(r, q);
2472    case INDEX_op_qemu_ld_i64:
2473        return C_O2_I1(e, p, q);
2474    case INDEX_op_qemu_st_i32:
2475        return C_O0_I2(q, q);
2476    case INDEX_op_qemu_st_i64:
2477        return C_O0_I3(Q, p, q);
2478
2479    case INDEX_op_st_vec:
2480        return C_O0_I2(w, r);
2481    case INDEX_op_ld_vec:
2482    case INDEX_op_dupm_vec:
2483        return C_O1_I1(w, r);
2484    case INDEX_op_dup_vec:
2485        return C_O1_I1(w, wr);
2486    case INDEX_op_abs_vec:
2487    case INDEX_op_neg_vec:
2488    case INDEX_op_not_vec:
2489    case INDEX_op_shli_vec:
2490    case INDEX_op_shri_vec:
2491    case INDEX_op_sari_vec:
2492        return C_O1_I1(w, w);
2493    case INDEX_op_dup2_vec:
2494    case INDEX_op_add_vec:
2495    case INDEX_op_mul_vec:
2496    case INDEX_op_smax_vec:
2497    case INDEX_op_smin_vec:
2498    case INDEX_op_ssadd_vec:
2499    case INDEX_op_sssub_vec:
2500    case INDEX_op_sub_vec:
2501    case INDEX_op_umax_vec:
2502    case INDEX_op_umin_vec:
2503    case INDEX_op_usadd_vec:
2504    case INDEX_op_ussub_vec:
2505    case INDEX_op_xor_vec:
2506    case INDEX_op_arm_sshl_vec:
2507    case INDEX_op_arm_ushl_vec:
2508        return C_O1_I2(w, w, w);
2509    case INDEX_op_arm_sli_vec:
2510        return C_O1_I2(w, 0, w);
2511    case INDEX_op_or_vec:
2512    case INDEX_op_andc_vec:
2513        return C_O1_I2(w, w, wO);
2514    case INDEX_op_and_vec:
2515    case INDEX_op_orc_vec:
2516        return C_O1_I2(w, w, wV);
2517    case INDEX_op_cmp_vec:
2518        return C_O1_I2(w, w, wZ);
2519    case INDEX_op_bitsel_vec:
2520        return C_O1_I3(w, w, w, w);
2521    default:
2522        return C_NotImplemented;
2523    }
2524}
2525
2526static void tcg_target_init(TCGContext *s)
2527{
2528    /*
2529     * Only probe for the platform and capabilities if we haven't already
2530     * determined maximum values at compile time.
2531     */
2532#if !defined(use_idiv_instructions) || !defined(use_neon_instructions)
2533    {
2534        unsigned long hwcap = qemu_getauxval(AT_HWCAP);
2535#ifndef use_idiv_instructions
2536        use_idiv_instructions = (hwcap & HWCAP_ARM_IDIVA) != 0;
2537#endif
2538#ifndef use_neon_instructions
2539        use_neon_instructions = (hwcap & HWCAP_ARM_NEON) != 0;
2540#endif
2541    }
2542#endif
2543
2544    if (__ARM_ARCH < 7) {
2545        const char *pl = (const char *)qemu_getauxval(AT_PLATFORM);
2546        if (pl != NULL && pl[0] == 'v' && pl[1] >= '4' && pl[1] <= '9') {
2547            arm_arch = pl[1] - '0';
2548        }
2549
2550        if (arm_arch < 6) {
2551            error_report("TCG: ARMv%d is unsupported; exiting", arm_arch);
2552            exit(EXIT_FAILURE);
2553        }
2554    }
2555
2556    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
2557
2558    tcg_target_call_clobber_regs = 0;
2559    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R0);
2560    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R1);
2561    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R2);
2562    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R3);
2563    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R12);
2564    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R14);
2565
2566    if (use_neon_instructions) {
2567        tcg_target_available_regs[TCG_TYPE_V64]  = ALL_VECTOR_REGS;
2568        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
2569
2570        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q0);
2571        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q1);
2572        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q2);
2573        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q3);
2574        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q8);
2575        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q9);
2576        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q10);
2577        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q11);
2578        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q12);
2579        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q13);
2580        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q14);
2581        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q15);
2582    }
2583
2584    s->reserved_regs = 0;
2585    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
2586    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP);
2587    tcg_regset_set_reg(s->reserved_regs, TCG_REG_PC);
2588    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP);
2589}
2590
2591static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
2592                       TCGReg arg1, intptr_t arg2)
2593{
2594    switch (type) {
2595    case TCG_TYPE_I32:
2596        tcg_out_ld32u(s, COND_AL, arg, arg1, arg2);
2597        return;
2598    case TCG_TYPE_V64:
2599        /* regs 1; size 8; align 8 */
2600        tcg_out_vldst(s, INSN_VLD1 | 0x7d0, arg, arg1, arg2);
2601        return;
2602    case TCG_TYPE_V128:
2603        /*
2604         * We have only 8-byte alignment for the stack per the ABI.
2605         * Rather than dynamically re-align the stack, it's easier
2606         * to simply not request alignment beyond that.  So:
2607         * regs 2; size 8; align 8
2608         */
2609        tcg_out_vldst(s, INSN_VLD1 | 0xad0, arg, arg1, arg2);
2610        return;
2611    default:
2612        g_assert_not_reached();
2613    }
2614}
2615
2616static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
2617                       TCGReg arg1, intptr_t arg2)
2618{
2619    switch (type) {
2620    case TCG_TYPE_I32:
2621        tcg_out_st32(s, COND_AL, arg, arg1, arg2);
2622        return;
2623    case TCG_TYPE_V64:
2624        /* regs 1; size 8; align 8 */
2625        tcg_out_vldst(s, INSN_VST1 | 0x7d0, arg, arg1, arg2);
2626        return;
2627    case TCG_TYPE_V128:
2628        /* See tcg_out_ld re alignment: regs 2; size 8; align 8 */
2629        tcg_out_vldst(s, INSN_VST1 | 0xad0, arg, arg1, arg2);
2630        return;
2631    default:
2632        g_assert_not_reached();
2633    }
2634}
2635
2636static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
2637                        TCGReg base, intptr_t ofs)
2638{
2639    return false;
2640}
2641
2642static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
2643{
2644    if (ret == arg) {
2645        return true;
2646    }
2647    switch (type) {
2648    case TCG_TYPE_I32:
2649        if (ret < TCG_REG_Q0 && arg < TCG_REG_Q0) {
2650            tcg_out_mov_reg(s, COND_AL, ret, arg);
2651            return true;
2652        }
2653        return false;
2654
2655    case TCG_TYPE_V64:
2656    case TCG_TYPE_V128:
2657        /* "VMOV D,N" is an alias for "VORR D,N,N". */
2658        tcg_out_vreg3(s, INSN_VORR, type - TCG_TYPE_V64, 0, ret, arg, arg);
2659        return true;
2660
2661    default:
2662        g_assert_not_reached();
2663    }
2664}
2665
2666static void tcg_out_movi(TCGContext *s, TCGType type,
2667                         TCGReg ret, tcg_target_long arg)
2668{
2669    tcg_debug_assert(type == TCG_TYPE_I32);
2670    tcg_debug_assert(ret < TCG_REG_Q0);
2671    tcg_out_movi32(s, COND_AL, ret, arg);
2672}
2673
2674static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
2675{
2676    return false;
2677}
2678
2679static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
2680                             tcg_target_long imm)
2681{
2682    int enc, opc = ARITH_ADD;
2683
2684    /* All of the easiest immediates to encode are positive. */
2685    if (imm < 0) {
2686        imm = -imm;
2687        opc = ARITH_SUB;
2688    }
2689    enc = encode_imm(imm);
2690    if (enc >= 0) {
2691        tcg_out_dat_imm(s, COND_AL, opc, rd, rs, enc);
2692    } else {
2693        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, imm);
2694        tcg_out_dat_reg(s, COND_AL, opc, rd, rs,
2695                        TCG_REG_TMP, SHIFT_IMM_LSL(0));
2696    }
2697}
2698
2699/* Type is always V128, with I64 elements.  */
2700static void tcg_out_dup2_vec(TCGContext *s, TCGReg rd, TCGReg rl, TCGReg rh)
2701{
2702    /* Move high element into place first. */
2703    /* VMOV Dd+1, Ds */
2704    tcg_out_vreg3(s, INSN_VORR | (1 << 12), 0, 0, rd, rh, rh);
2705    /* Move low element into place; tcg_out_mov will check for nop. */
2706    tcg_out_mov(s, TCG_TYPE_V64, rd, rl);
2707}
2708
2709static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
2710                            TCGReg rd, TCGReg rs)
2711{
2712    int q = type - TCG_TYPE_V64;
2713
2714    if (vece == MO_64) {
2715        if (type == TCG_TYPE_V128) {
2716            tcg_out_dup2_vec(s, rd, rs, rs);
2717        } else {
2718            tcg_out_mov(s, TCG_TYPE_V64, rd, rs);
2719        }
2720    } else if (rs < TCG_REG_Q0) {
2721        int b = (vece == MO_8);
2722        int e = (vece == MO_16);
2723        tcg_out32(s, INSN_VDUP_G | (b << 22) | (q << 21) | (e << 5) |
2724                  encode_vn(rd) | (rs << 12));
2725    } else {
2726        int imm4 = 1 << vece;
2727        tcg_out32(s, INSN_VDUP_S | (imm4 << 16) | (q << 6) |
2728                  encode_vd(rd) | encode_vm(rs));
2729    }
2730    return true;
2731}
2732
2733static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
2734                             TCGReg rd, TCGReg base, intptr_t offset)
2735{
2736    if (vece == MO_64) {
2737        tcg_out_ld(s, TCG_TYPE_V64, rd, base, offset);
2738        if (type == TCG_TYPE_V128) {
2739            tcg_out_dup2_vec(s, rd, rd, rd);
2740        }
2741    } else {
2742        int q = type - TCG_TYPE_V64;
2743        tcg_out_vldst(s, INSN_VLD1R | (vece << 6) | (q << 5),
2744                      rd, base, offset);
2745    }
2746    return true;
2747}
2748
2749static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
2750                             TCGReg rd, int64_t v64)
2751{
2752    int q = type - TCG_TYPE_V64;
2753    int cmode, imm8, i;
2754
2755    /* Test all bytes equal first.  */
2756    if (vece == MO_8) {
2757        tcg_out_vmovi(s, rd, q, 0, 0xe, v64);
2758        return;
2759    }
2760
2761    /*
2762     * Test all bytes 0x00 or 0xff second.  This can match cases that
2763     * might otherwise take 2 or 3 insns for MO_16 or MO_32 below.
2764     */
2765    for (i = imm8 = 0; i < 8; i++) {
2766        uint8_t byte = v64 >> (i * 8);
2767        if (byte == 0xff) {
2768            imm8 |= 1 << i;
2769        } else if (byte != 0) {
2770            goto fail_bytes;
2771        }
2772    }
2773    tcg_out_vmovi(s, rd, q, 1, 0xe, imm8);
2774    return;
2775 fail_bytes:
2776
2777    /*
2778     * Tests for various replications.  For each element width, if we
2779     * cannot find an expansion there's no point checking a larger
2780     * width because we already know by replication it cannot match.
2781     */
2782    if (vece == MO_16) {
2783        uint16_t v16 = v64;
2784
2785        if (is_shimm16(v16, &cmode, &imm8)) {
2786            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2787            return;
2788        }
2789        if (is_shimm16(~v16, &cmode, &imm8)) {
2790            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2791            return;
2792        }
2793
2794        /*
2795         * Otherwise, all remaining constants can be loaded in two insns:
2796         * rd = v16 & 0xff, rd |= v16 & 0xff00.
2797         */
2798        tcg_out_vmovi(s, rd, q, 0, 0x8, v16 & 0xff);
2799        tcg_out_vmovi(s, rd, q, 0, 0xb, v16 >> 8);   /* VORRI */
2800        return;
2801    }
2802
2803    if (vece == MO_32) {
2804        uint32_t v32 = v64;
2805
2806        if (is_shimm32(v32, &cmode, &imm8) ||
2807            is_soimm32(v32, &cmode, &imm8)) {
2808            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2809            return;
2810        }
2811        if (is_shimm32(~v32, &cmode, &imm8) ||
2812            is_soimm32(~v32, &cmode, &imm8)) {
2813            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2814            return;
2815        }
2816
2817        /*
2818         * Restrict the set of constants to those we can load with
2819         * two instructions.  Others we load from the pool.
2820         */
2821        i = is_shimm32_pair(v32, &cmode, &imm8);
2822        if (i) {
2823            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2824            tcg_out_vmovi(s, rd, q, 0, i | 1, extract32(v32, i * 4, 8));
2825            return;
2826        }
2827        i = is_shimm32_pair(~v32, &cmode, &imm8);
2828        if (i) {
2829            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2830            tcg_out_vmovi(s, rd, q, 1, i | 1, extract32(~v32, i * 4, 8));
2831            return;
2832        }
2833    }
2834
2835    /*
2836     * As a last resort, load from the constant pool.
2837     */
2838    if (!q || vece == MO_64) {
2839        new_pool_l2(s, R_ARM_PC11, s->code_ptr, 0, v64, v64 >> 32);
2840        /* VLDR Dd, [pc + offset] */
2841        tcg_out32(s, INSN_VLDR_D | encode_vd(rd) | (0xf << 16));
2842        if (q) {
2843            tcg_out_dup2_vec(s, rd, rd, rd);
2844        }
2845    } else {
2846        new_pool_label(s, (uint32_t)v64, R_ARM_PC8, s->code_ptr, 0);
2847        /* add tmp, pc, offset */
2848        tcg_out_dat_imm(s, COND_AL, ARITH_ADD, TCG_REG_TMP, TCG_REG_PC, 0);
2849        tcg_out_dupm_vec(s, type, MO_32, rd, TCG_REG_TMP, 0);
2850    }
2851}
2852
2853static const ARMInsn vec_cmp_insn[16] = {
2854    [TCG_COND_EQ] = INSN_VCEQ,
2855    [TCG_COND_GT] = INSN_VCGT,
2856    [TCG_COND_GE] = INSN_VCGE,
2857    [TCG_COND_GTU] = INSN_VCGT_U,
2858    [TCG_COND_GEU] = INSN_VCGE_U,
2859};
2860
2861static const ARMInsn vec_cmp0_insn[16] = {
2862    [TCG_COND_EQ] = INSN_VCEQ0,
2863    [TCG_COND_GT] = INSN_VCGT0,
2864    [TCG_COND_GE] = INSN_VCGE0,
2865    [TCG_COND_LT] = INSN_VCLT0,
2866    [TCG_COND_LE] = INSN_VCLE0,
2867};
2868
2869static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2870                           unsigned vecl, unsigned vece,
2871                           const TCGArg args[TCG_MAX_OP_ARGS],
2872                           const int const_args[TCG_MAX_OP_ARGS])
2873{
2874    TCGType type = vecl + TCG_TYPE_V64;
2875    unsigned q = vecl;
2876    TCGArg a0, a1, a2, a3;
2877    int cmode, imm8;
2878
2879    a0 = args[0];
2880    a1 = args[1];
2881    a2 = args[2];
2882
2883    switch (opc) {
2884    case INDEX_op_ld_vec:
2885        tcg_out_ld(s, type, a0, a1, a2);
2886        return;
2887    case INDEX_op_st_vec:
2888        tcg_out_st(s, type, a0, a1, a2);
2889        return;
2890    case INDEX_op_dupm_vec:
2891        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
2892        return;
2893    case INDEX_op_dup2_vec:
2894        tcg_out_dup2_vec(s, a0, a1, a2);
2895        return;
2896    case INDEX_op_abs_vec:
2897        tcg_out_vreg2(s, INSN_VABS, q, vece, a0, a1);
2898        return;
2899    case INDEX_op_neg_vec:
2900        tcg_out_vreg2(s, INSN_VNEG, q, vece, a0, a1);
2901        return;
2902    case INDEX_op_not_vec:
2903        tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a1);
2904        return;
2905    case INDEX_op_add_vec:
2906        tcg_out_vreg3(s, INSN_VADD, q, vece, a0, a1, a2);
2907        return;
2908    case INDEX_op_mul_vec:
2909        tcg_out_vreg3(s, INSN_VMUL, q, vece, a0, a1, a2);
2910        return;
2911    case INDEX_op_smax_vec:
2912        tcg_out_vreg3(s, INSN_VMAX, q, vece, a0, a1, a2);
2913        return;
2914    case INDEX_op_smin_vec:
2915        tcg_out_vreg3(s, INSN_VMIN, q, vece, a0, a1, a2);
2916        return;
2917    case INDEX_op_sub_vec:
2918        tcg_out_vreg3(s, INSN_VSUB, q, vece, a0, a1, a2);
2919        return;
2920    case INDEX_op_ssadd_vec:
2921        tcg_out_vreg3(s, INSN_VQADD, q, vece, a0, a1, a2);
2922        return;
2923    case INDEX_op_sssub_vec:
2924        tcg_out_vreg3(s, INSN_VQSUB, q, vece, a0, a1, a2);
2925        return;
2926    case INDEX_op_umax_vec:
2927        tcg_out_vreg3(s, INSN_VMAX_U, q, vece, a0, a1, a2);
2928        return;
2929    case INDEX_op_umin_vec:
2930        tcg_out_vreg3(s, INSN_VMIN_U, q, vece, a0, a1, a2);
2931        return;
2932    case INDEX_op_usadd_vec:
2933        tcg_out_vreg3(s, INSN_VQADD_U, q, vece, a0, a1, a2);
2934        return;
2935    case INDEX_op_ussub_vec:
2936        tcg_out_vreg3(s, INSN_VQSUB_U, q, vece, a0, a1, a2);
2937        return;
2938    case INDEX_op_xor_vec:
2939        tcg_out_vreg3(s, INSN_VEOR, q, 0, a0, a1, a2);
2940        return;
2941    case INDEX_op_arm_sshl_vec:
2942        /*
2943         * Note that Vm is the data and Vn is the shift count,
2944         * therefore the arguments appear reversed.
2945         */
2946        tcg_out_vreg3(s, INSN_VSHL_S, q, vece, a0, a2, a1);
2947        return;
2948    case INDEX_op_arm_ushl_vec:
2949        /* See above. */
2950        tcg_out_vreg3(s, INSN_VSHL_U, q, vece, a0, a2, a1);
2951        return;
2952    case INDEX_op_shli_vec:
2953        tcg_out_vshifti(s, INSN_VSHLI, q, a0, a1, a2 + (8 << vece));
2954        return;
2955    case INDEX_op_shri_vec:
2956        tcg_out_vshifti(s, INSN_VSHRI, q, a0, a1, (16 << vece) - a2);
2957        return;
2958    case INDEX_op_sari_vec:
2959        tcg_out_vshifti(s, INSN_VSARI, q, a0, a1, (16 << vece) - a2);
2960        return;
2961    case INDEX_op_arm_sli_vec:
2962        tcg_out_vshifti(s, INSN_VSLI, q, a0, a2, args[3] + (8 << vece));
2963        return;
2964
2965    case INDEX_op_andc_vec:
2966        if (!const_args[2]) {
2967            tcg_out_vreg3(s, INSN_VBIC, q, 0, a0, a1, a2);
2968            return;
2969        }
2970        a2 = ~a2;
2971        /* fall through */
2972    case INDEX_op_and_vec:
2973        if (const_args[2]) {
2974            is_shimm1632(~a2, &cmode, &imm8);
2975            if (a0 == a1) {
2976                tcg_out_vmovi(s, a0, q, 1, cmode | 1, imm8); /* VBICI */
2977                return;
2978            }
2979            tcg_out_vmovi(s, a0, q, 1, cmode, imm8); /* VMVNI */
2980            a2 = a0;
2981        }
2982        tcg_out_vreg3(s, INSN_VAND, q, 0, a0, a1, a2);
2983        return;
2984
2985    case INDEX_op_orc_vec:
2986        if (!const_args[2]) {
2987            tcg_out_vreg3(s, INSN_VORN, q, 0, a0, a1, a2);
2988            return;
2989        }
2990        a2 = ~a2;
2991        /* fall through */
2992    case INDEX_op_or_vec:
2993        if (const_args[2]) {
2994            is_shimm1632(a2, &cmode, &imm8);
2995            if (a0 == a1) {
2996                tcg_out_vmovi(s, a0, q, 0, cmode | 1, imm8); /* VORRI */
2997                return;
2998            }
2999            tcg_out_vmovi(s, a0, q, 0, cmode, imm8); /* VMOVI */
3000            a2 = a0;
3001        }
3002        tcg_out_vreg3(s, INSN_VORR, q, 0, a0, a1, a2);
3003        return;
3004
3005    case INDEX_op_cmp_vec:
3006        {
3007            TCGCond cond = args[3];
3008            ARMInsn insn;
3009
3010            switch (cond) {
3011            case TCG_COND_NE:
3012                if (const_args[2]) {
3013                    tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a1);
3014                } else {
3015                    tcg_out_vreg3(s, INSN_VCEQ, q, vece, a0, a1, a2);
3016                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
3017                }
3018                break;
3019
3020            case TCG_COND_TSTNE:
3021            case TCG_COND_TSTEQ:
3022                if (const_args[2]) {
3023                    /* (x & 0) == 0 */
3024                    tcg_out_dupi_vec(s, type, MO_8, a0,
3025                                     -(cond == TCG_COND_TSTEQ));
3026                    break;
3027                }
3028                tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a2);
3029                if (cond == TCG_COND_TSTEQ) {
3030                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
3031                }
3032                break;
3033
3034            default:
3035                if (const_args[2]) {
3036                    insn = vec_cmp0_insn[cond];
3037                    if (insn) {
3038                        tcg_out_vreg2(s, insn, q, vece, a0, a1);
3039                        return;
3040                    }
3041                    tcg_out_dupi_vec(s, type, MO_8, TCG_VEC_TMP, 0);
3042                    a2 = TCG_VEC_TMP;
3043                }
3044                insn = vec_cmp_insn[cond];
3045                if (insn == 0) {
3046                    TCGArg t;
3047                    t = a1, a1 = a2, a2 = t;
3048                    cond = tcg_swap_cond(cond);
3049                    insn = vec_cmp_insn[cond];
3050                    tcg_debug_assert(insn != 0);
3051                }
3052                tcg_out_vreg3(s, insn, q, vece, a0, a1, a2);
3053                break;
3054            }
3055        }
3056        return;
3057
3058    case INDEX_op_bitsel_vec:
3059        a3 = args[3];
3060        if (a0 == a3) {
3061            tcg_out_vreg3(s, INSN_VBIT, q, 0, a0, a2, a1);
3062        } else if (a0 == a2) {
3063            tcg_out_vreg3(s, INSN_VBIF, q, 0, a0, a3, a1);
3064        } else {
3065            tcg_out_mov(s, type, a0, a1);
3066            tcg_out_vreg3(s, INSN_VBSL, q, 0, a0, a2, a3);
3067        }
3068        return;
3069
3070    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
3071    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
3072    default:
3073        g_assert_not_reached();
3074    }
3075}
3076
3077int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
3078{
3079    switch (opc) {
3080    case INDEX_op_add_vec:
3081    case INDEX_op_sub_vec:
3082    case INDEX_op_and_vec:
3083    case INDEX_op_andc_vec:
3084    case INDEX_op_or_vec:
3085    case INDEX_op_orc_vec:
3086    case INDEX_op_xor_vec:
3087    case INDEX_op_not_vec:
3088    case INDEX_op_shli_vec:
3089    case INDEX_op_shri_vec:
3090    case INDEX_op_sari_vec:
3091    case INDEX_op_ssadd_vec:
3092    case INDEX_op_sssub_vec:
3093    case INDEX_op_usadd_vec:
3094    case INDEX_op_ussub_vec:
3095    case INDEX_op_bitsel_vec:
3096        return 1;
3097    case INDEX_op_abs_vec:
3098    case INDEX_op_cmp_vec:
3099    case INDEX_op_mul_vec:
3100    case INDEX_op_neg_vec:
3101    case INDEX_op_smax_vec:
3102    case INDEX_op_smin_vec:
3103    case INDEX_op_umax_vec:
3104    case INDEX_op_umin_vec:
3105        return vece < MO_64;
3106    case INDEX_op_shlv_vec:
3107    case INDEX_op_shrv_vec:
3108    case INDEX_op_sarv_vec:
3109    case INDEX_op_rotli_vec:
3110    case INDEX_op_rotlv_vec:
3111    case INDEX_op_rotrv_vec:
3112        return -1;
3113    default:
3114        return 0;
3115    }
3116}
3117
3118void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
3119                       TCGArg a0, ...)
3120{
3121    va_list va;
3122    TCGv_vec v0, v1, v2, t1, t2, c1;
3123    TCGArg a2;
3124
3125    va_start(va, a0);
3126    v0 = temp_tcgv_vec(arg_temp(a0));
3127    v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
3128    a2 = va_arg(va, TCGArg);
3129    va_end(va);
3130
3131    switch (opc) {
3132    case INDEX_op_shlv_vec:
3133        /*
3134         * Merely propagate shlv_vec to arm_ushl_vec.
3135         * In this way we don't set TCG_TARGET_HAS_shv_vec
3136         * because everything is done via expansion.
3137         */
3138        v2 = temp_tcgv_vec(arg_temp(a2));
3139        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
3140                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
3141        break;
3142
3143    case INDEX_op_shrv_vec:
3144    case INDEX_op_sarv_vec:
3145        /* Right shifts are negative left shifts for NEON.  */
3146        v2 = temp_tcgv_vec(arg_temp(a2));
3147        t1 = tcg_temp_new_vec(type);
3148        tcg_gen_neg_vec(vece, t1, v2);
3149        if (opc == INDEX_op_shrv_vec) {
3150            opc = INDEX_op_arm_ushl_vec;
3151        } else {
3152            opc = INDEX_op_arm_sshl_vec;
3153        }
3154        vec_gen_3(opc, type, vece, tcgv_vec_arg(v0),
3155                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
3156        tcg_temp_free_vec(t1);
3157        break;
3158
3159    case INDEX_op_rotli_vec:
3160        t1 = tcg_temp_new_vec(type);
3161        tcg_gen_shri_vec(vece, t1, v1, -a2 & ((8 << vece) - 1));
3162        vec_gen_4(INDEX_op_arm_sli_vec, type, vece,
3163                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(v1), a2);
3164        tcg_temp_free_vec(t1);
3165        break;
3166
3167    case INDEX_op_rotlv_vec:
3168        v2 = temp_tcgv_vec(arg_temp(a2));
3169        t1 = tcg_temp_new_vec(type);
3170        c1 = tcg_constant_vec(type, vece, 8 << vece);
3171        tcg_gen_sub_vec(vece, t1, v2, c1);
3172        /* Right shifts are negative left shifts for NEON.  */
3173        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
3174                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
3175        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
3176                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
3177        tcg_gen_or_vec(vece, v0, v0, t1);
3178        tcg_temp_free_vec(t1);
3179        break;
3180
3181    case INDEX_op_rotrv_vec:
3182        v2 = temp_tcgv_vec(arg_temp(a2));
3183        t1 = tcg_temp_new_vec(type);
3184        t2 = tcg_temp_new_vec(type);
3185        c1 = tcg_constant_vec(type, vece, 8 << vece);
3186        tcg_gen_neg_vec(vece, t1, v2);
3187        tcg_gen_sub_vec(vece, t2, c1, v2);
3188        /* Right shifts are negative left shifts for NEON.  */
3189        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
3190                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
3191        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t2),
3192                  tcgv_vec_arg(v1), tcgv_vec_arg(t2));
3193        tcg_gen_or_vec(vece, v0, t1, t2);
3194        tcg_temp_free_vec(t1);
3195        tcg_temp_free_vec(t2);
3196        break;
3197
3198    default:
3199        g_assert_not_reached();
3200    }
3201}
3202
3203static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
3204{
3205    int i;
3206    for (i = 0; i < count; ++i) {
3207        p[i] = INSN_NOP;
3208    }
3209}
3210
3211/* Compute frame size via macros, to share between tcg_target_qemu_prologue
3212   and tcg_register_jit.  */
3213
3214#define PUSH_SIZE  ((11 - 4 + 1 + 1) * sizeof(tcg_target_long))
3215
3216#define FRAME_SIZE \
3217    ((PUSH_SIZE \
3218      + TCG_STATIC_CALL_ARGS_SIZE \
3219      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
3220      + TCG_TARGET_STACK_ALIGN - 1) \
3221     & -TCG_TARGET_STACK_ALIGN)
3222
3223#define STACK_ADDEND  (FRAME_SIZE - PUSH_SIZE)
3224
3225static void tcg_target_qemu_prologue(TCGContext *s)
3226{
3227    /* Calling convention requires us to save r4-r11 and lr.  */
3228    /* stmdb sp!, { r4 - r11, lr } */
3229    tcg_out_ldstm(s, COND_AL, INSN_STMDB, TCG_REG_CALL_STACK,
3230                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
3231                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3232                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_R14));
3233
3234    /* Reserve callee argument and tcg temp space.  */
3235    tcg_out_dat_rI(s, COND_AL, ARITH_SUB, TCG_REG_CALL_STACK,
3236                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
3237    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
3238                  CPU_TEMP_BUF_NLONGS * sizeof(long));
3239
3240    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
3241
3242    if (!tcg_use_softmmu && guest_base) {
3243        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_GUEST_BASE, guest_base);
3244        tcg_regset_set_reg(s->reserved_regs, TCG_REG_GUEST_BASE);
3245    }
3246
3247    tcg_out_b_reg(s, COND_AL, tcg_target_call_iarg_regs[1]);
3248
3249    /*
3250     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
3251     * and fall through to the rest of the epilogue.
3252     */
3253    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
3254    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, 0);
3255    tcg_out_epilogue(s);
3256}
3257
3258static void tcg_out_epilogue(TCGContext *s)
3259{
3260    /* Release local stack frame.  */
3261    tcg_out_dat_rI(s, COND_AL, ARITH_ADD, TCG_REG_CALL_STACK,
3262                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
3263
3264    /* ldmia sp!, { r4 - r11, pc } */
3265    tcg_out_ldstm(s, COND_AL, INSN_LDMIA, TCG_REG_CALL_STACK,
3266                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
3267                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3268                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_PC));
3269}
3270
3271static void tcg_out_tb_start(TCGContext *s)
3272{
3273    /* nothing to do */
3274}
3275
3276typedef struct {
3277    DebugFrameHeader h;
3278    uint8_t fde_def_cfa[4];
3279    uint8_t fde_reg_ofs[18];
3280} DebugFrame;
3281
3282#define ELF_HOST_MACHINE EM_ARM
3283
3284/* We're expecting a 2 byte uleb128 encoded value.  */
3285QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
3286
3287static const DebugFrame debug_frame = {
3288    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
3289    .h.cie.id = -1,
3290    .h.cie.version = 1,
3291    .h.cie.code_align = 1,
3292    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
3293    .h.cie.return_column = 14,
3294
3295    /* Total FDE size does not include the "len" member.  */
3296    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
3297
3298    .fde_def_cfa = {
3299        12, 13,                         /* DW_CFA_def_cfa sp, ... */
3300        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
3301        (FRAME_SIZE >> 7)
3302    },
3303    .fde_reg_ofs = {
3304        /* The following must match the stmdb in the prologue.  */
3305        0x8e, 1,                        /* DW_CFA_offset, lr, -4 */
3306        0x8b, 2,                        /* DW_CFA_offset, r11, -8 */
3307        0x8a, 3,                        /* DW_CFA_offset, r10, -12 */
3308        0x89, 4,                        /* DW_CFA_offset, r9, -16 */
3309        0x88, 5,                        /* DW_CFA_offset, r8, -20 */
3310        0x87, 6,                        /* DW_CFA_offset, r7, -24 */
3311        0x86, 7,                        /* DW_CFA_offset, r6, -28 */
3312        0x85, 8,                        /* DW_CFA_offset, r5, -32 */
3313        0x84, 9,                        /* DW_CFA_offset, r4, -36 */
3314    }
3315};
3316
3317void tcg_register_jit(const void *buf, size_t buf_size)
3318{
3319    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
3320}
3321