xref: /openbmc/qemu/tcg/arm/tcg-target.c.inc (revision 07d5d502f2b4a8eedda3c6bdfcab31dc36d1d1d5)
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 tcg_out_sextract(TCGContext *s, ARMCond cond, TCGReg rd,
1024                             TCGReg rn, int ofs, int len)
1025{
1026    if (use_armv7_instructions) {
1027        /* sbfx */
1028        tcg_out32(s, 0x07a00050 | (cond << 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 << 28) | (rd << 12) | (ofs << 7) | rn);
1038        break;
1039    case 16:
1040        /* sxth */
1041        tcg_out32(s, 0x06bf0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
1042        break;
1043    default:
1044        g_assert_not_reached();
1045    }
1046}
1047
1048
1049static void tcg_out_ld32u(TCGContext *s, ARMCond cond,
1050                          TCGReg rd, TCGReg rn, int32_t offset)
1051{
1052    if (offset > 0xfff || offset < -0xfff) {
1053        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1054        tcg_out_ld32_r(s, cond, rd, rn, TCG_REG_TMP);
1055    } else
1056        tcg_out_ld32_12(s, cond, rd, rn, offset);
1057}
1058
1059static void tcg_out_st32(TCGContext *s, ARMCond cond,
1060                         TCGReg rd, TCGReg rn, int32_t offset)
1061{
1062    if (offset > 0xfff || offset < -0xfff) {
1063        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1064        tcg_out_st32_r(s, cond, rd, rn, TCG_REG_TMP);
1065    } else
1066        tcg_out_st32_12(s, cond, rd, rn, offset);
1067}
1068
1069static void tcg_out_ld16u(TCGContext *s, ARMCond cond,
1070                          TCGReg rd, TCGReg rn, int32_t offset)
1071{
1072    if (offset > 0xff || offset < -0xff) {
1073        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1074        tcg_out_ld16u_r(s, cond, rd, rn, TCG_REG_TMP);
1075    } else
1076        tcg_out_ld16u_8(s, cond, rd, rn, offset);
1077}
1078
1079static void tcg_out_ld16s(TCGContext *s, ARMCond cond,
1080                          TCGReg rd, TCGReg rn, int32_t offset)
1081{
1082    if (offset > 0xff || offset < -0xff) {
1083        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1084        tcg_out_ld16s_r(s, cond, rd, rn, TCG_REG_TMP);
1085    } else
1086        tcg_out_ld16s_8(s, cond, rd, rn, offset);
1087}
1088
1089static void tcg_out_st16(TCGContext *s, ARMCond cond,
1090                         TCGReg rd, TCGReg rn, int32_t offset)
1091{
1092    if (offset > 0xff || offset < -0xff) {
1093        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1094        tcg_out_st16_r(s, cond, rd, rn, TCG_REG_TMP);
1095    } else
1096        tcg_out_st16_8(s, cond, rd, rn, offset);
1097}
1098
1099static void tcg_out_ld8u(TCGContext *s, ARMCond cond,
1100                         TCGReg rd, TCGReg rn, int32_t offset)
1101{
1102    if (offset > 0xfff || offset < -0xfff) {
1103        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1104        tcg_out_ld8_r(s, cond, rd, rn, TCG_REG_TMP);
1105    } else
1106        tcg_out_ld8_12(s, cond, rd, rn, offset);
1107}
1108
1109static void tcg_out_ld8s(TCGContext *s, ARMCond cond,
1110                         TCGReg rd, TCGReg rn, int32_t offset)
1111{
1112    if (offset > 0xff || offset < -0xff) {
1113        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1114        tcg_out_ld8s_r(s, cond, rd, rn, TCG_REG_TMP);
1115    } else
1116        tcg_out_ld8s_8(s, cond, rd, rn, offset);
1117}
1118
1119static void tcg_out_st8(TCGContext *s, ARMCond cond,
1120                        TCGReg rd, TCGReg rn, int32_t offset)
1121{
1122    if (offset > 0xfff || offset < -0xfff) {
1123        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1124        tcg_out_st8_r(s, cond, rd, rn, TCG_REG_TMP);
1125    } else
1126        tcg_out_st8_12(s, cond, rd, rn, offset);
1127}
1128
1129/*
1130 * The _goto case is normally between TBs within the same code buffer, and
1131 * with the code buffer limited to 16MB we wouldn't need the long case.
1132 * But we also use it for the tail-call to the qemu_ld/st helpers, which does.
1133 */
1134static void tcg_out_goto(TCGContext *s, ARMCond cond, const tcg_insn_unit *addr)
1135{
1136    intptr_t addri = (intptr_t)addr;
1137    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1138    bool arm_mode = !(addri & 1);
1139
1140    if (arm_mode && disp - 8 < 0x01fffffd && disp - 8 > -0x01fffffd) {
1141        tcg_out_b_imm(s, cond, disp);
1142        return;
1143    }
1144
1145    /* LDR is interworking from v5t. */
1146    tcg_out_movi_pool(s, cond, TCG_REG_PC, addri);
1147}
1148
1149/*
1150 * The call case is mostly used for helpers - so it's not unreasonable
1151 * for them to be beyond branch range.
1152 */
1153static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *addr)
1154{
1155    intptr_t addri = (intptr_t)addr;
1156    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1157    bool arm_mode = !(addri & 1);
1158
1159    if (disp - 8 < 0x02000000 && disp - 8 >= -0x02000000) {
1160        if (arm_mode) {
1161            tcg_out_bl_imm(s, COND_AL, disp);
1162        } else {
1163            tcg_out_blx_imm(s, disp);
1164        }
1165        return;
1166    }
1167
1168    tcg_out_movi32(s, COND_AL, TCG_REG_TMP, addri);
1169    tcg_out_blx_reg(s, COND_AL, TCG_REG_TMP);
1170}
1171
1172static void tcg_out_call(TCGContext *s, const tcg_insn_unit *addr,
1173                         const TCGHelperInfo *info)
1174{
1175    tcg_out_call_int(s, addr);
1176}
1177
1178static void tcg_out_goto_label(TCGContext *s, ARMCond cond, TCGLabel *l)
1179{
1180    if (l->has_value) {
1181        tcg_out_goto(s, cond, l->u.value_ptr);
1182    } else {
1183        tcg_out_reloc(s, s->code_ptr, R_ARM_PC24, l, 0);
1184        tcg_out_b_imm(s, cond, 0);
1185    }
1186}
1187
1188static void tcg_out_mb(TCGContext *s, TCGArg a0)
1189{
1190    if (use_armv7_instructions) {
1191        tcg_out32(s, INSN_DMB_ISH);
1192    } else {
1193        tcg_out32(s, INSN_DMB_MCR);
1194    }
1195}
1196
1197static TCGCond tgen_cmp(TCGContext *s, TCGCond cond, TCGReg a, TCGReg b)
1198{
1199    if (is_tst_cond(cond)) {
1200        tcg_out_dat_reg(s, COND_AL, ARITH_TST, 0, a, b, SHIFT_IMM_LSL(0));
1201        return tcg_tst_eqne_cond(cond);
1202    }
1203    tcg_out_dat_reg(s, COND_AL, ARITH_CMP, 0, a, b, SHIFT_IMM_LSL(0));
1204    return cond;
1205}
1206
1207static TCGCond tgen_cmpi(TCGContext *s, TCGCond cond, TCGReg a, TCGArg b)
1208{
1209    int imm12;
1210
1211    if (!is_tst_cond(cond)) {
1212        tcg_out_dat_IN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0, a, b);
1213        return cond;
1214    }
1215
1216    /*
1217     * The compare constraints allow rIN, but TST does not support N.
1218     * Be prepared to load the constant into a scratch register.
1219     */
1220    imm12 = encode_imm(b);
1221    if (imm12 >= 0) {
1222        tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, a, imm12);
1223    } else {
1224        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, b);
1225        tcg_out_dat_reg(s, COND_AL, ARITH_TST, 0,
1226                        a, TCG_REG_TMP, SHIFT_IMM_LSL(0));
1227    }
1228    return tcg_tst_eqne_cond(cond);
1229}
1230
1231static TCGCond tcg_out_cmp(TCGContext *s, TCGCond cond, TCGReg a,
1232                           TCGArg b, int b_const)
1233{
1234    if (b_const) {
1235        return tgen_cmpi(s, cond, a, b);
1236    } else {
1237        return tgen_cmp(s, cond, a, b);
1238    }
1239}
1240
1241static TCGCond tcg_out_cmp2(TCGContext *s, TCGCond cond, TCGReg al, TCGReg ah,
1242                            TCGArg bl, bool const_bl, TCGArg bh, bool const_bh)
1243{
1244    switch (cond) {
1245    case TCG_COND_EQ:
1246    case TCG_COND_NE:
1247    case TCG_COND_LTU:
1248    case TCG_COND_LEU:
1249    case TCG_COND_GTU:
1250    case TCG_COND_GEU:
1251        /*
1252         * We perform a conditional comparison.  If the high half is
1253         * equal, then overwrite the flags with the comparison of the
1254         * low half.  The resulting flags cover the whole.
1255         */
1256        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, ah, bh, const_bh);
1257        tcg_out_dat_rI(s, COND_EQ, ARITH_CMP, 0, al, bl, const_bl);
1258        return cond;
1259
1260    case TCG_COND_TSTEQ:
1261    case TCG_COND_TSTNE:
1262        /* Similar, but with TST instead of CMP. */
1263        tcg_out_dat_rI(s, COND_AL, ARITH_TST, 0, ah, bh, const_bh);
1264        tcg_out_dat_rI(s, COND_EQ, ARITH_TST, 0, al, bl, const_bl);
1265        return tcg_tst_eqne_cond(cond);
1266
1267    case TCG_COND_LT:
1268    case TCG_COND_GE:
1269        /* We perform a double-word subtraction and examine the result.
1270           We do not actually need the result of the subtract, so the
1271           low part "subtract" is a compare.  For the high half we have
1272           no choice but to compute into a temporary.  */
1273        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, al, bl, const_bl);
1274        tcg_out_dat_rI(s, COND_AL, ARITH_SBC | TO_CPSR,
1275                       TCG_REG_TMP, ah, bh, const_bh);
1276        return cond;
1277
1278    case TCG_COND_LE:
1279    case TCG_COND_GT:
1280        /* Similar, but with swapped arguments, via reversed subtract.  */
1281        tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR,
1282                       TCG_REG_TMP, al, bl, const_bl);
1283        tcg_out_dat_rI(s, COND_AL, ARITH_RSC | TO_CPSR,
1284                       TCG_REG_TMP, ah, bh, const_bh);
1285        return tcg_swap_cond(cond);
1286
1287    default:
1288        g_assert_not_reached();
1289    }
1290}
1291
1292/*
1293 * Note that TCGReg references Q-registers.
1294 * Q-regno = 2 * D-regno, so shift left by 1 while inserting.
1295 */
1296static uint32_t encode_vd(TCGReg rd)
1297{
1298    tcg_debug_assert(rd >= TCG_REG_Q0);
1299    return (extract32(rd, 3, 1) << 22) | (extract32(rd, 0, 3) << 13);
1300}
1301
1302static uint32_t encode_vn(TCGReg rn)
1303{
1304    tcg_debug_assert(rn >= TCG_REG_Q0);
1305    return (extract32(rn, 3, 1) << 7) | (extract32(rn, 0, 3) << 17);
1306}
1307
1308static uint32_t encode_vm(TCGReg rm)
1309{
1310    tcg_debug_assert(rm >= TCG_REG_Q0);
1311    return (extract32(rm, 3, 1) << 5) | (extract32(rm, 0, 3) << 1);
1312}
1313
1314static void tcg_out_vreg2(TCGContext *s, ARMInsn insn, int q, int vece,
1315                          TCGReg d, TCGReg m)
1316{
1317    tcg_out32(s, insn | (vece << 18) | (q << 6) |
1318              encode_vd(d) | encode_vm(m));
1319}
1320
1321static void tcg_out_vreg3(TCGContext *s, ARMInsn insn, int q, int vece,
1322                          TCGReg d, TCGReg n, TCGReg m)
1323{
1324    tcg_out32(s, insn | (vece << 20) | (q << 6) |
1325              encode_vd(d) | encode_vn(n) | encode_vm(m));
1326}
1327
1328static void tcg_out_vmovi(TCGContext *s, TCGReg rd,
1329                          int q, int op, int cmode, uint8_t imm8)
1330{
1331    tcg_out32(s, INSN_VMOVI | encode_vd(rd) | (q << 6) | (op << 5)
1332              | (cmode << 8) | extract32(imm8, 0, 4)
1333              | (extract32(imm8, 4, 3) << 16)
1334              | (extract32(imm8, 7, 1) << 24));
1335}
1336
1337static void tcg_out_vshifti(TCGContext *s, ARMInsn insn, int q,
1338                            TCGReg rd, TCGReg rm, int l_imm6)
1339{
1340    tcg_out32(s, insn | (q << 6) | encode_vd(rd) | encode_vm(rm) |
1341              (extract32(l_imm6, 6, 1) << 7) |
1342              (extract32(l_imm6, 0, 6) << 16));
1343}
1344
1345static void tcg_out_vldst(TCGContext *s, ARMInsn insn,
1346                          TCGReg rd, TCGReg rn, int offset)
1347{
1348    if (offset != 0) {
1349        if (check_fit_imm(offset) || check_fit_imm(-offset)) {
1350            tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
1351                            TCG_REG_TMP, rn, offset, true);
1352        } else {
1353            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP, offset);
1354            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1355                            TCG_REG_TMP, TCG_REG_TMP, rn, 0);
1356        }
1357        rn = TCG_REG_TMP;
1358    }
1359    tcg_out32(s, insn | (rn << 16) | encode_vd(rd) | 0xf);
1360}
1361
1362typedef struct {
1363    ARMCond cond;
1364    TCGReg base;
1365    int index;
1366    bool index_scratch;
1367    TCGAtomAlign aa;
1368} HostAddress;
1369
1370bool tcg_target_has_memory_bswap(MemOp memop)
1371{
1372    return false;
1373}
1374
1375static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
1376{
1377    /* We arrive at the slow path via "BLNE", so R14 contains l->raddr. */
1378    return TCG_REG_R14;
1379}
1380
1381static const TCGLdstHelperParam ldst_helper_param = {
1382    .ra_gen = ldst_ra_gen,
1383    .ntmp = 1,
1384    .tmp = { TCG_REG_TMP },
1385};
1386
1387static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1388{
1389    MemOp opc = get_memop(lb->oi);
1390
1391    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1392        return false;
1393    }
1394
1395    tcg_out_ld_helper_args(s, lb, &ldst_helper_param);
1396    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE]);
1397    tcg_out_ld_helper_ret(s, lb, false, &ldst_helper_param);
1398
1399    tcg_out_goto(s, COND_AL, lb->raddr);
1400    return true;
1401}
1402
1403static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1404{
1405    MemOp opc = get_memop(lb->oi);
1406
1407    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1408        return false;
1409    }
1410
1411    tcg_out_st_helper_args(s, lb, &ldst_helper_param);
1412
1413    /* Tail-call to the helper, which will return to the fast path.  */
1414    tcg_out_goto(s, COND_AL, qemu_st_helpers[opc & MO_SIZE]);
1415    return true;
1416}
1417
1418/* We expect to use an 9-bit sign-magnitude negative offset from ENV.  */
1419#define MIN_TLB_MASK_TABLE_OFS  -256
1420
1421static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
1422                                           TCGReg addr, MemOpIdx oi, bool is_ld)
1423{
1424    TCGLabelQemuLdst *ldst = NULL;
1425    MemOp opc = get_memop(oi);
1426    unsigned a_mask;
1427
1428    if (tcg_use_softmmu) {
1429        *h = (HostAddress){
1430            .cond = COND_AL,
1431            .base = addr,
1432            .index = TCG_REG_R1,
1433            .index_scratch = true,
1434        };
1435    } else {
1436        *h = (HostAddress){
1437            .cond = COND_AL,
1438            .base = addr,
1439            .index = guest_base ? TCG_REG_GUEST_BASE : -1,
1440            .index_scratch = false,
1441        };
1442    }
1443
1444    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, false);
1445    a_mask = (1 << h->aa.align) - 1;
1446
1447    if (tcg_use_softmmu) {
1448        int mem_index = get_mmuidx(oi);
1449        int cmp_off = is_ld ? offsetof(CPUTLBEntry, addr_read)
1450                            : offsetof(CPUTLBEntry, addr_write);
1451        int fast_off = tlb_mask_table_ofs(s, mem_index);
1452        unsigned s_mask = (1 << (opc & MO_SIZE)) - 1;
1453        TCGReg t_addr;
1454
1455        ldst = new_ldst_label(s);
1456        ldst->is_ld = is_ld;
1457        ldst->oi = oi;
1458        ldst->addr_reg = addr;
1459
1460        /* Load cpu->neg.tlb.f[mmu_idx].{mask,table} into {r0,r1}.  */
1461        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, mask) != 0);
1462        QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, table) != 4);
1463        tcg_out_ldrd_8(s, COND_AL, TCG_REG_R0, TCG_AREG0, fast_off);
1464
1465        /* Extract the tlb index from the address into R0.  */
1466        tcg_out_dat_reg(s, COND_AL, ARITH_AND, TCG_REG_R0, TCG_REG_R0, addr,
1467                        SHIFT_IMM_LSR(s->page_bits - CPU_TLB_ENTRY_BITS));
1468
1469        /*
1470         * Add the tlb_table pointer, creating the CPUTLBEntry address in R1.
1471         * Load the tlb comparator into R2 and the fast path addend into R1.
1472         */
1473        if (cmp_off == 0) {
1474            tcg_out_ld32_rwb(s, COND_AL, TCG_REG_R2, TCG_REG_R1, TCG_REG_R0);
1475        } else {
1476            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1477                            TCG_REG_R1, TCG_REG_R1, TCG_REG_R0, 0);
1478            tcg_out_ld32_12(s, COND_AL, TCG_REG_R2, TCG_REG_R1, cmp_off);
1479        }
1480
1481        /* Load the tlb addend.  */
1482        tcg_out_ld32_12(s, COND_AL, TCG_REG_R1, TCG_REG_R1,
1483                        offsetof(CPUTLBEntry, addend));
1484
1485        /*
1486         * Check alignment, check comparators.
1487         * Do this in 2-4 insns.  Use MOVW for v7, if possible,
1488         * to reduce the number of sequential conditional instructions.
1489         * Almost all guests have at least 4k pages, which means that we need
1490         * to clear at least 9 bits even for an 8-byte memory, which means it
1491         * isn't worth checking for an immediate operand for BIC.
1492         *
1493         * For unaligned accesses, test the page of the last unit of alignment.
1494         * This leaves the least significant alignment bits unchanged, and of
1495         * course must be zero.
1496         */
1497        t_addr = addr;
1498        if (a_mask < s_mask) {
1499            t_addr = TCG_REG_R0;
1500            tcg_out_dat_imm(s, COND_AL, ARITH_ADD, t_addr,
1501                            addr, s_mask - a_mask);
1502        }
1503        if (use_armv7_instructions && s->page_bits <= 16) {
1504            tcg_out_movi32(s, COND_AL, TCG_REG_TMP, ~(s->page_mask | a_mask));
1505            tcg_out_dat_reg(s, COND_AL, ARITH_BIC, TCG_REG_TMP,
1506                            t_addr, TCG_REG_TMP, 0);
1507            tcg_out_dat_reg(s, COND_AL, ARITH_CMP, 0,
1508                            TCG_REG_R2, TCG_REG_TMP, 0);
1509        } else {
1510            if (a_mask) {
1511                tcg_debug_assert(a_mask <= 0xff);
1512                tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addr, a_mask);
1513            }
1514            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0, t_addr,
1515                            SHIFT_IMM_LSR(s->page_bits));
1516            tcg_out_dat_reg(s, (a_mask ? COND_EQ : COND_AL), ARITH_CMP,
1517                            0, TCG_REG_R2, TCG_REG_TMP,
1518                            SHIFT_IMM_LSL(s->page_bits));
1519        }
1520    } else if (a_mask) {
1521        ldst = new_ldst_label(s);
1522        ldst->is_ld = is_ld;
1523        ldst->oi = oi;
1524        ldst->addr_reg = addr;
1525
1526        /* We are expecting alignment to max out at 7 */
1527        tcg_debug_assert(a_mask <= 0xff);
1528        /* tst addr, #mask */
1529        tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addr, a_mask);
1530    }
1531
1532    return ldst;
1533}
1534
1535static void tcg_out_qemu_ld_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1536                                   TCGReg datahi, HostAddress h)
1537{
1538    TCGReg base;
1539
1540    /* Byte swapping is left to middle-end expansion. */
1541    tcg_debug_assert((opc & MO_BSWAP) == 0);
1542
1543    switch (opc & MO_SSIZE) {
1544    case MO_UB:
1545        if (h.index < 0) {
1546            tcg_out_ld8_12(s, h.cond, datalo, h.base, 0);
1547        } else {
1548            tcg_out_ld8_r(s, h.cond, datalo, h.base, h.index);
1549        }
1550        break;
1551    case MO_SB:
1552        if (h.index < 0) {
1553            tcg_out_ld8s_8(s, h.cond, datalo, h.base, 0);
1554        } else {
1555            tcg_out_ld8s_r(s, h.cond, datalo, h.base, h.index);
1556        }
1557        break;
1558    case MO_UW:
1559        if (h.index < 0) {
1560            tcg_out_ld16u_8(s, h.cond, datalo, h.base, 0);
1561        } else {
1562            tcg_out_ld16u_r(s, h.cond, datalo, h.base, h.index);
1563        }
1564        break;
1565    case MO_SW:
1566        if (h.index < 0) {
1567            tcg_out_ld16s_8(s, h.cond, datalo, h.base, 0);
1568        } else {
1569            tcg_out_ld16s_r(s, h.cond, datalo, h.base, h.index);
1570        }
1571        break;
1572    case MO_UL:
1573        if (h.index < 0) {
1574            tcg_out_ld32_12(s, h.cond, datalo, h.base, 0);
1575        } else {
1576            tcg_out_ld32_r(s, h.cond, datalo, h.base, h.index);
1577        }
1578        break;
1579    case MO_UQ:
1580        /* We used pair allocation for datalo, so already should be aligned. */
1581        tcg_debug_assert((datalo & 1) == 0);
1582        tcg_debug_assert(datahi == datalo + 1);
1583        /* LDRD requires alignment; double-check that. */
1584        if (memop_alignment_bits(opc) >= MO_64) {
1585            if (h.index < 0) {
1586                tcg_out_ldrd_8(s, h.cond, datalo, h.base, 0);
1587                break;
1588            }
1589            /*
1590             * Rm (the second address op) must not overlap Rt or Rt + 1.
1591             * Since datalo is aligned, we can simplify the test via alignment.
1592             * Flip the two address arguments if that works.
1593             */
1594            if ((h.index & ~1) != datalo) {
1595                tcg_out_ldrd_r(s, h.cond, datalo, h.base, h.index);
1596                break;
1597            }
1598            if ((h.base & ~1) != datalo) {
1599                tcg_out_ldrd_r(s, h.cond, datalo, h.index, h.base);
1600                break;
1601            }
1602        }
1603        if (h.index < 0) {
1604            base = h.base;
1605            if (datalo == h.base) {
1606                tcg_out_mov_reg(s, h.cond, TCG_REG_TMP, base);
1607                base = TCG_REG_TMP;
1608            }
1609        } else if (h.index_scratch) {
1610            tcg_out_ld32_rwb(s, h.cond, datalo, h.index, h.base);
1611            tcg_out_ld32_12(s, h.cond, datahi, h.index, 4);
1612            break;
1613        } else {
1614            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1615                            h.base, h.index, SHIFT_IMM_LSL(0));
1616            base = TCG_REG_TMP;
1617        }
1618        tcg_out_ld32_12(s, h.cond, datalo, base, 0);
1619        tcg_out_ld32_12(s, h.cond, datahi, base, 4);
1620        break;
1621    default:
1622        g_assert_not_reached();
1623    }
1624}
1625
1626static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
1627                            TCGReg addr, MemOpIdx oi, TCGType data_type)
1628{
1629    MemOp opc = get_memop(oi);
1630    TCGLabelQemuLdst *ldst;
1631    HostAddress h;
1632
1633    ldst = prepare_host_addr(s, &h, addr, oi, true);
1634    if (ldst) {
1635        ldst->type = data_type;
1636        ldst->datalo_reg = datalo;
1637        ldst->datahi_reg = datahi;
1638
1639        /*
1640         * This a conditional BL only to load a pointer within this
1641         * opcode into LR for the slow path.  We will not be using
1642         * the value for a tail call.
1643         */
1644        ldst->label_ptr[0] = s->code_ptr;
1645        tcg_out_bl_imm(s, COND_NE, 0);
1646
1647        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1648        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1649    } else {
1650        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, h);
1651    }
1652}
1653
1654static void tcg_out_qemu_st_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1655                                   TCGReg datahi, HostAddress h)
1656{
1657    /* Byte swapping is left to middle-end expansion. */
1658    tcg_debug_assert((opc & MO_BSWAP) == 0);
1659
1660    switch (opc & MO_SIZE) {
1661    case MO_8:
1662        if (h.index < 0) {
1663            tcg_out_st8_12(s, h.cond, datalo, h.base, 0);
1664        } else {
1665            tcg_out_st8_r(s, h.cond, datalo, h.base, h.index);
1666        }
1667        break;
1668    case MO_16:
1669        if (h.index < 0) {
1670            tcg_out_st16_8(s, h.cond, datalo, h.base, 0);
1671        } else {
1672            tcg_out_st16_r(s, h.cond, datalo, h.base, h.index);
1673        }
1674        break;
1675    case MO_32:
1676        if (h.index < 0) {
1677            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1678        } else {
1679            tcg_out_st32_r(s, h.cond, datalo, h.base, h.index);
1680        }
1681        break;
1682    case MO_64:
1683        /* We used pair allocation for datalo, so already should be aligned. */
1684        tcg_debug_assert((datalo & 1) == 0);
1685        tcg_debug_assert(datahi == datalo + 1);
1686        /* STRD requires alignment; double-check that. */
1687        if (memop_alignment_bits(opc) >= MO_64) {
1688            if (h.index < 0) {
1689                tcg_out_strd_8(s, h.cond, datalo, h.base, 0);
1690            } else {
1691                tcg_out_strd_r(s, h.cond, datalo, h.base, h.index);
1692            }
1693        } else if (h.index < 0) {
1694            tcg_out_st32_12(s, h.cond, datalo, h.base, 0);
1695            tcg_out_st32_12(s, h.cond, datahi, h.base, 4);
1696        } else if (h.index_scratch) {
1697            tcg_out_st32_rwb(s, h.cond, datalo, h.index, h.base);
1698            tcg_out_st32_12(s, h.cond, datahi, h.index, 4);
1699        } else {
1700            tcg_out_dat_reg(s, h.cond, ARITH_ADD, TCG_REG_TMP,
1701                            h.base, h.index, SHIFT_IMM_LSL(0));
1702            tcg_out_st32_12(s, h.cond, datalo, TCG_REG_TMP, 0);
1703            tcg_out_st32_12(s, h.cond, datahi, TCG_REG_TMP, 4);
1704        }
1705        break;
1706    default:
1707        g_assert_not_reached();
1708    }
1709}
1710
1711static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
1712                            TCGReg addr, MemOpIdx oi, TCGType data_type)
1713{
1714    MemOp opc = get_memop(oi);
1715    TCGLabelQemuLdst *ldst;
1716    HostAddress h;
1717
1718    ldst = prepare_host_addr(s, &h, addr, oi, false);
1719    if (ldst) {
1720        ldst->type = data_type;
1721        ldst->datalo_reg = datalo;
1722        ldst->datahi_reg = datahi;
1723
1724        h.cond = COND_EQ;
1725        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1726
1727        /* The conditional call is last, as we're going to return here. */
1728        ldst->label_ptr[0] = s->code_ptr;
1729        tcg_out_bl_imm(s, COND_NE, 0);
1730        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
1731    } else {
1732        tcg_out_qemu_st_direct(s, opc, datalo, datahi, h);
1733    }
1734}
1735
1736static void tcg_out_epilogue(TCGContext *s);
1737
1738static void tcg_out_exit_tb(TCGContext *s, uintptr_t arg)
1739{
1740    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, arg);
1741    tcg_out_epilogue(s);
1742}
1743
1744static void tcg_out_goto_tb(TCGContext *s, int which)
1745{
1746    uintptr_t i_addr;
1747    intptr_t i_disp;
1748
1749    /* Direct branch will be patched by tb_target_set_jmp_target. */
1750    set_jmp_insn_offset(s, which);
1751    tcg_out32(s, INSN_NOP);
1752
1753    /* When branch is out of range, fall through to indirect. */
1754    i_addr = get_jmp_target_addr(s, which);
1755    i_disp = tcg_pcrel_diff(s, (void *)i_addr) - 8;
1756    tcg_debug_assert(i_disp < 0);
1757    if (i_disp >= -0xfff) {
1758        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_PC, i_disp);
1759    } else {
1760        /*
1761         * The TB is close, but outside the 12 bits addressable by
1762         * the load.  We can extend this to 20 bits with a sub of a
1763         * shifted immediate from pc.
1764         */
1765        int h = -i_disp;
1766        int l = -(h & 0xfff);
1767
1768        h = encode_imm_nofail(h + l);
1769        tcg_out_dat_imm(s, COND_AL, ARITH_SUB, TCG_REG_R0, TCG_REG_PC, h);
1770        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_R0, l);
1771    }
1772    set_jmp_reset_offset(s, which);
1773}
1774
1775void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1776                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1777{
1778    uintptr_t addr = tb->jmp_target_addr[n];
1779    ptrdiff_t offset = addr - (jmp_rx + 8);
1780    tcg_insn_unit insn;
1781
1782    /* Either directly branch, or fall through to indirect branch. */
1783    if (offset == sextract64(offset, 0, 26)) {
1784        /* B <addr> */
1785        insn = deposit32((COND_AL << 28) | INSN_B, 0, 24, offset >> 2);
1786    } else {
1787        insn = INSN_NOP;
1788    }
1789
1790    qatomic_set((uint32_t *)jmp_rw, insn);
1791    flush_idcache_range(jmp_rx, jmp_rw, 4);
1792}
1793
1794
1795static void tgen_add(TCGContext *s, TCGType type,
1796                     TCGReg a0, TCGReg a1, TCGReg a2)
1797{
1798    tcg_out_dat_reg(s, COND_AL, ARITH_ADD, a0, a1, a2, SHIFT_IMM_LSL(0));
1799}
1800
1801static void tgen_addi(TCGContext *s, TCGType type,
1802                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1803{
1804    tcg_out_dat_IN(s, COND_AL, ARITH_ADD, ARITH_SUB, a0, a1, a2);
1805}
1806
1807static const TCGOutOpBinary outop_add = {
1808    .base.static_constraint = C_O1_I2(r, r, rIN),
1809    .out_rrr = tgen_add,
1810    .out_rri = tgen_addi,
1811};
1812
1813static void tgen_and(TCGContext *s, TCGType type,
1814                     TCGReg a0, TCGReg a1, TCGReg a2)
1815{
1816    tcg_out_dat_reg(s, COND_AL, ARITH_AND, a0, a1, a2, SHIFT_IMM_LSL(0));
1817}
1818
1819static void tgen_andi(TCGContext *s, TCGType type,
1820                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1821{
1822    tcg_out_dat_IK(s, COND_AL, ARITH_AND, ARITH_BIC, a0, a1, a2);
1823}
1824
1825static const TCGOutOpBinary outop_and = {
1826    .base.static_constraint = C_O1_I2(r, r, rIK),
1827    .out_rrr = tgen_and,
1828    .out_rri = tgen_andi,
1829};
1830
1831static void tgen_andc(TCGContext *s, TCGType type,
1832                      TCGReg a0, TCGReg a1, TCGReg a2)
1833{
1834    tcg_out_dat_reg(s, COND_AL, ARITH_BIC, a0, a1, a2, SHIFT_IMM_LSL(0));
1835}
1836
1837static const TCGOutOpBinary outop_andc = {
1838    .base.static_constraint = C_O1_I2(r, r, r),
1839    .out_rrr = tgen_andc,
1840};
1841
1842static void tgen_clz(TCGContext *s, TCGType type,
1843                     TCGReg a0, TCGReg a1, TCGReg a2)
1844{
1845    tcg_out_dat_imm(s, COND_AL, ARITH_CMP, 0, a1, 0);
1846    tcg_out_dat_reg(s, COND_NE, INSN_CLZ, a0, 0, a1, 0);
1847    tcg_out_mov_reg(s, COND_EQ, a0, a2);
1848}
1849
1850static void tgen_clzi(TCGContext *s, TCGType type,
1851                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1852{
1853    if (a2 == 32) {
1854        tcg_out_dat_reg(s, COND_AL, INSN_CLZ, a0, 0, a1, 0);
1855    } else {
1856        tcg_out_dat_imm(s, COND_AL, ARITH_CMP, 0, a1, 0);
1857        tcg_out_dat_reg(s, COND_NE, INSN_CLZ, a0, 0, a1, 0);
1858        tcg_out_movi32(s, COND_EQ, a0, a2);
1859    }
1860}
1861
1862static const TCGOutOpBinary outop_clz = {
1863    .base.static_constraint = C_O1_I2(r, r, rIK),
1864    .out_rrr = tgen_clz,
1865    .out_rri = tgen_clzi,
1866};
1867
1868static const TCGOutOpUnary outop_ctpop = {
1869    .base.static_constraint = C_NotImplemented,
1870};
1871
1872static void tgen_ctz(TCGContext *s, TCGType type,
1873                     TCGReg a0, TCGReg a1, TCGReg a2)
1874{
1875    tcg_out_dat_reg(s, COND_AL, INSN_RBIT, TCG_REG_TMP, 0, a1, 0);
1876    tgen_clz(s, TCG_TYPE_I32, a0, TCG_REG_TMP, a2);
1877}
1878
1879static void tgen_ctzi(TCGContext *s, TCGType type,
1880                      TCGReg a0, TCGReg a1, tcg_target_long a2)
1881{
1882    tcg_out_dat_reg(s, COND_AL, INSN_RBIT, TCG_REG_TMP, 0, a1, 0);
1883    tgen_clzi(s, TCG_TYPE_I32, a0, TCG_REG_TMP, a2);
1884}
1885
1886static TCGConstraintSetIndex cset_ctz(TCGType type, unsigned flags)
1887{
1888    return use_armv7_instructions ? C_O1_I2(r, r, rIK) : C_NotImplemented;
1889}
1890
1891static const TCGOutOpBinary outop_ctz = {
1892    .base.static_constraint = C_Dynamic,
1893    .base.dynamic_constraint = cset_ctz,
1894    .out_rrr = tgen_ctz,
1895    .out_rri = tgen_ctzi,
1896};
1897
1898static TCGConstraintSetIndex cset_idiv(TCGType type, unsigned flags)
1899{
1900    return use_idiv_instructions ? C_O1_I2(r, r, r) : C_NotImplemented;
1901}
1902
1903static void tgen_divs(TCGContext *s, TCGType type,
1904                      TCGReg a0, TCGReg a1, TCGReg a2)
1905{
1906    /* sdiv */
1907    tcg_out32(s, 0x0710f010 | (COND_AL << 28) | (a0 << 16) | a1 | (a2 << 8));
1908}
1909
1910static const TCGOutOpBinary outop_divs = {
1911    .base.static_constraint = C_Dynamic,
1912    .base.dynamic_constraint = cset_idiv,
1913    .out_rrr = tgen_divs,
1914};
1915
1916static const TCGOutOpDivRem outop_divs2 = {
1917    .base.static_constraint = C_NotImplemented,
1918};
1919
1920static void tgen_divu(TCGContext *s, TCGType type,
1921                      TCGReg a0, TCGReg a1, TCGReg a2)
1922{
1923    /* udiv */
1924    tcg_out32(s, 0x0730f010 | (COND_AL << 28) | (a0 << 16) | a1 | (a2 << 8));
1925}
1926
1927static const TCGOutOpBinary outop_divu = {
1928    .base.static_constraint = C_Dynamic,
1929    .base.dynamic_constraint = cset_idiv,
1930    .out_rrr = tgen_divu,
1931};
1932
1933static const TCGOutOpDivRem outop_divu2 = {
1934    .base.static_constraint = C_NotImplemented,
1935};
1936
1937static const TCGOutOpBinary outop_eqv = {
1938    .base.static_constraint = C_NotImplemented,
1939};
1940
1941static void tgen_mul(TCGContext *s, TCGType type,
1942                     TCGReg a0, TCGReg a1, TCGReg a2)
1943{
1944    /* mul */
1945    tcg_out32(s, (COND_AL << 28) | 0x90 | (a0 << 16) | (a1 << 8) | a2);
1946}
1947
1948static const TCGOutOpBinary outop_mul = {
1949    .base.static_constraint = C_O1_I2(r, r, r),
1950    .out_rrr = tgen_mul,
1951};
1952
1953static void tgen_muls2(TCGContext *s, TCGType type,
1954                       TCGReg rd0, TCGReg rd1, TCGReg rn, TCGReg rm)
1955{
1956    /* smull */
1957    tcg_out32(s, (COND_AL << 28) | 0x00c00090 |
1958              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
1959}
1960
1961static const TCGOutOpMul2 outop_muls2 = {
1962    .base.static_constraint = C_O2_I2(r, r, r, r),
1963    .out_rrrr = tgen_muls2,
1964};
1965
1966static const TCGOutOpBinary outop_mulsh = {
1967    .base.static_constraint = C_NotImplemented,
1968};
1969
1970static void tgen_mulu2(TCGContext *s, TCGType type,
1971                       TCGReg rd0, TCGReg rd1, TCGReg rn, TCGReg rm)
1972{
1973    /* umull */
1974    tcg_out32(s, (COND_AL << 28) | 0x00800090 |
1975              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
1976}
1977
1978static const TCGOutOpMul2 outop_mulu2 = {
1979    .base.static_constraint = C_O2_I2(r, r, r, r),
1980    .out_rrrr = tgen_mulu2,
1981};
1982
1983static const TCGOutOpBinary outop_muluh = {
1984    .base.static_constraint = C_NotImplemented,
1985};
1986
1987static const TCGOutOpBinary outop_nand = {
1988    .base.static_constraint = C_NotImplemented,
1989};
1990
1991static const TCGOutOpBinary outop_nor = {
1992    .base.static_constraint = C_NotImplemented,
1993};
1994
1995static void tgen_or(TCGContext *s, TCGType type,
1996                     TCGReg a0, TCGReg a1, TCGReg a2)
1997{
1998    tcg_out_dat_reg(s, COND_AL, ARITH_ORR, a0, a1, a2, SHIFT_IMM_LSL(0));
1999}
2000
2001static void tgen_ori(TCGContext *s, TCGType type,
2002                     TCGReg a0, TCGReg a1, tcg_target_long a2)
2003{
2004    tcg_out_dat_imm(s, COND_AL, ARITH_ORR, a0, a1, encode_imm_nofail(a2));
2005}
2006
2007static const TCGOutOpBinary outop_or = {
2008    .base.static_constraint = C_O1_I2(r, r, rI),
2009    .out_rrr = tgen_or,
2010    .out_rri = tgen_ori,
2011};
2012
2013static const TCGOutOpBinary outop_orc = {
2014    .base.static_constraint = C_NotImplemented,
2015};
2016
2017static const TCGOutOpBinary outop_rems = {
2018    .base.static_constraint = C_NotImplemented,
2019};
2020
2021static const TCGOutOpBinary outop_remu = {
2022    .base.static_constraint = C_NotImplemented,
2023};
2024
2025static const TCGOutOpBinary outop_rotl = {
2026    .base.static_constraint = C_NotImplemented,
2027};
2028
2029static void tgen_rotr(TCGContext *s, TCGType type,
2030                      TCGReg a0, TCGReg a1, TCGReg a2)
2031{
2032    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1, SHIFT_REG_ROR(a2));
2033}
2034
2035static void tgen_rotri(TCGContext *s, TCGType type,
2036                       TCGReg a0, TCGReg a1, tcg_target_long a2)
2037{
2038    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1, SHIFT_IMM_ROR(a2 & 0x1f));
2039}
2040
2041static const TCGOutOpBinary outop_rotr = {
2042    .base.static_constraint = C_O1_I2(r, r, ri),
2043    .out_rrr = tgen_rotr,
2044    .out_rri = tgen_rotri,
2045};
2046
2047static void tgen_sar(TCGContext *s, TCGType type,
2048                     TCGReg a0, TCGReg a1, TCGReg a2)
2049{
2050    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1, SHIFT_REG_ASR(a2));
2051}
2052
2053static void tgen_sari(TCGContext *s, TCGType type,
2054                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2055{
2056    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1,
2057                    SHIFT_IMM_ASR(a2 & 0x1f));
2058}
2059
2060static const TCGOutOpBinary outop_sar = {
2061    .base.static_constraint = C_O1_I2(r, r, ri),
2062    .out_rrr = tgen_sar,
2063    .out_rri = tgen_sari,
2064};
2065
2066static void tgen_shl(TCGContext *s, TCGType type,
2067                     TCGReg a0, TCGReg a1, TCGReg a2)
2068{
2069    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1, SHIFT_REG_LSL(a2));
2070}
2071
2072static void tgen_shli(TCGContext *s, TCGType type,
2073                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2074{
2075    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1,
2076                    SHIFT_IMM_LSL(a2 & 0x1f));
2077}
2078
2079static const TCGOutOpBinary outop_shl = {
2080    .base.static_constraint = C_O1_I2(r, r, ri),
2081    .out_rrr = tgen_shl,
2082    .out_rri = tgen_shli,
2083};
2084
2085static void tgen_shr(TCGContext *s, TCGType type,
2086                     TCGReg a0, TCGReg a1, TCGReg a2)
2087{
2088    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1, SHIFT_REG_LSR(a2));
2089}
2090
2091static void tgen_shri(TCGContext *s, TCGType type,
2092                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2093{
2094    tcg_out_dat_reg(s, COND_AL, ARITH_MOV, a0, 0, a1,
2095                    SHIFT_IMM_LSR(a2 & 0x1f));
2096}
2097
2098static const TCGOutOpBinary outop_shr = {
2099    .base.static_constraint = C_O1_I2(r, r, ri),
2100    .out_rrr = tgen_shr,
2101    .out_rri = tgen_shri,
2102};
2103
2104static void tgen_sub(TCGContext *s, TCGType type,
2105                     TCGReg a0, TCGReg a1, TCGReg a2)
2106{
2107    tcg_out_dat_reg(s, COND_AL, ARITH_SUB, a0, a1, a2, SHIFT_IMM_LSL(0));
2108}
2109
2110static void tgen_subfi(TCGContext *s, TCGType type,
2111                       TCGReg a0, tcg_target_long a1, TCGReg a2)
2112{
2113    tcg_out_dat_imm(s, COND_AL, ARITH_RSB, a0, a2, encode_imm_nofail(a1));
2114}
2115
2116static const TCGOutOpSubtract outop_sub = {
2117    .base.static_constraint = C_O1_I2(r, rI, r),
2118    .out_rrr = tgen_sub,
2119    .out_rir = tgen_subfi,
2120};
2121
2122static void tgen_xor(TCGContext *s, TCGType type,
2123                     TCGReg a0, TCGReg a1, TCGReg a2)
2124{
2125    tcg_out_dat_reg(s, COND_AL, ARITH_EOR, a0, a1, a2, SHIFT_IMM_LSL(0));
2126}
2127
2128static void tgen_xori(TCGContext *s, TCGType type,
2129                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2130{
2131    tcg_out_dat_imm(s, COND_AL, ARITH_EOR, a0, a1, encode_imm_nofail(a2));
2132}
2133
2134static const TCGOutOpBinary outop_xor = {
2135    .base.static_constraint = C_O1_I2(r, r, rI),
2136    .out_rrr = tgen_xor,
2137    .out_rri = tgen_xori,
2138};
2139
2140static void tgen_bswap16(TCGContext *s, TCGType type,
2141                         TCGReg rd, TCGReg rn, unsigned flags)
2142{
2143    if (flags & TCG_BSWAP_OS) {
2144        /* revsh */
2145        tcg_out32(s, 0x06ff0fb0 | (COND_AL << 28) | (rd << 12) | rn);
2146        return;
2147    }
2148
2149    /* rev16 */
2150    tcg_out32(s, 0x06bf0fb0 | (COND_AL << 28) | (rd << 12) | rn);
2151    if ((flags & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
2152        tcg_out_ext16u(s, rd, rd);
2153    }
2154}
2155
2156static const TCGOutOpBswap outop_bswap16 = {
2157    .base.static_constraint = C_O1_I1(r, r),
2158    .out_rr = tgen_bswap16,
2159};
2160
2161static void tgen_bswap32(TCGContext *s, TCGType type,
2162                         TCGReg rd, TCGReg rn, unsigned flags)
2163{
2164    /* rev */
2165    tcg_out32(s, 0x06bf0f30 | (COND_AL << 28) | (rd << 12) | rn);
2166}
2167
2168static const TCGOutOpBswap outop_bswap32 = {
2169    .base.static_constraint = C_O1_I1(r, r),
2170    .out_rr = tgen_bswap32,
2171};
2172
2173static const TCGOutOpUnary outop_bswap64 = {
2174    .base.static_constraint = C_NotImplemented,
2175};
2176
2177static void tgen_neg(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
2178{
2179    tgen_subfi(s, type, a0, 0, a1);
2180}
2181
2182static const TCGOutOpUnary outop_neg = {
2183    .base.static_constraint = C_O1_I1(r, r),
2184    .out_rr = tgen_neg,
2185};
2186
2187static void tgen_not(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
2188{
2189    tcg_out_dat_reg(s, COND_AL, ARITH_MVN, a0, 0, a1, SHIFT_IMM_LSL(0));
2190}
2191
2192static const TCGOutOpUnary outop_not = {
2193    .base.static_constraint = C_O1_I1(r, r),
2194    .out_rr = tgen_not,
2195};
2196
2197static void tgen_brcond(TCGContext *s, TCGType type, TCGCond cond,
2198                        TCGReg a0, TCGReg a1, TCGLabel *l)
2199{
2200    cond = tgen_cmp(s, cond, a0, a1);
2201    tcg_out_goto_label(s, tcg_cond_to_arm_cond[cond], l);
2202}
2203
2204static void tgen_brcondi(TCGContext *s, TCGType type, TCGCond cond,
2205                         TCGReg a0, tcg_target_long a1, TCGLabel *l)
2206{
2207    cond = tgen_cmpi(s, cond, a0, a1);
2208    tcg_out_goto_label(s, tcg_cond_to_arm_cond[cond], l);
2209}
2210
2211static const TCGOutOpBrcond outop_brcond = {
2212    .base.static_constraint = C_O0_I2(r, rIN),
2213    .out_rr = tgen_brcond,
2214    .out_ri = tgen_brcondi,
2215};
2216
2217static void finish_setcond(TCGContext *s, TCGCond cond, TCGReg ret, bool neg)
2218{
2219    tcg_out_movi32(s, tcg_cond_to_arm_cond[tcg_invert_cond(cond)], ret, 0);
2220    tcg_out_movi32(s, tcg_cond_to_arm_cond[cond], ret, neg ? -1 : 1);
2221}
2222
2223static void tgen_setcond(TCGContext *s, TCGType type, TCGCond cond,
2224                         TCGReg a0, TCGReg a1, TCGReg a2)
2225{
2226    cond = tgen_cmp(s, cond, a1, a2);
2227    finish_setcond(s, cond, a0, false);
2228}
2229
2230static void tgen_setcondi(TCGContext *s, TCGType type, TCGCond cond,
2231                          TCGReg a0, TCGReg a1, tcg_target_long a2)
2232{
2233    cond = tgen_cmpi(s, cond, a1, a2);
2234    finish_setcond(s, cond, a0, false);
2235}
2236
2237static const TCGOutOpSetcond outop_setcond = {
2238    .base.static_constraint = C_O1_I2(r, r, rIN),
2239    .out_rrr = tgen_setcond,
2240    .out_rri = tgen_setcondi,
2241};
2242
2243static void tgen_negsetcond(TCGContext *s, TCGType type, TCGCond cond,
2244                            TCGReg a0, TCGReg a1, TCGReg a2)
2245{
2246    cond = tgen_cmp(s, cond, a1, a2);
2247    finish_setcond(s, cond, a0, true);
2248}
2249
2250static void tgen_negsetcondi(TCGContext *s, TCGType type, TCGCond cond,
2251                             TCGReg a0, TCGReg a1, tcg_target_long a2)
2252{
2253    cond = tgen_cmpi(s, cond, a1, a2);
2254    finish_setcond(s, cond, a0, true);
2255}
2256
2257static const TCGOutOpSetcond outop_negsetcond = {
2258    .base.static_constraint = C_O1_I2(r, r, rIN),
2259    .out_rrr = tgen_negsetcond,
2260    .out_rri = tgen_negsetcondi,
2261};
2262
2263static void tgen_movcond(TCGContext *s, TCGType type, TCGCond cond,
2264                         TCGReg ret, TCGReg c1, TCGArg c2, bool const_c2,
2265                         TCGArg vt, bool const_vt, TCGArg vf, bool consf_vf)
2266{
2267    cond = tcg_out_cmp(s, cond, c1, c2, const_c2);
2268    tcg_out_dat_rIK(s, tcg_cond_to_arm_cond[cond], ARITH_MOV, ARITH_MVN,
2269                    ret, 0, vt, const_vt);
2270}
2271
2272static const TCGOutOpMovcond outop_movcond = {
2273    .base.static_constraint = C_O1_I4(r, r, rIN, rIK, 0),
2274    .out = tgen_movcond,
2275};
2276
2277static void tgen_brcond2(TCGContext *s, TCGCond cond, TCGReg al, TCGReg ah,
2278                         TCGArg bl, bool const_bl, TCGArg bh, bool const_bh,
2279                         TCGLabel *l)
2280{
2281    cond = tcg_out_cmp2(s, cond, al, ah, bl, const_bl, bh, const_bh);
2282    tcg_out_goto_label(s, tcg_cond_to_arm_cond[cond], l);
2283}
2284
2285static const TCGOutOpBrcond2 outop_brcond2 = {
2286    .base.static_constraint = C_O0_I4(r, r, rI, rI),
2287    .out = tgen_brcond2,
2288};
2289
2290static void tgen_setcond2(TCGContext *s, TCGCond cond, TCGReg ret,
2291                          TCGReg al, TCGReg ah,
2292                          TCGArg bl, bool const_bl,
2293                          TCGArg bh, bool const_bh)
2294{
2295    cond = tcg_out_cmp2(s, cond, al, ah, bl, const_bl, bh, const_bh);
2296    finish_setcond(s, cond, ret, false);
2297}
2298
2299static const TCGOutOpSetcond2 outop_setcond2 = {
2300    .base.static_constraint = C_O1_I4(r, r, r, rI, rI),
2301    .out = tgen_setcond2,
2302};
2303
2304static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
2305                       const TCGArg args[TCG_MAX_OP_ARGS],
2306                       const int const_args[TCG_MAX_OP_ARGS])
2307{
2308    TCGArg a0, a1, a2, a3, a4, a5;
2309
2310    switch (opc) {
2311    case INDEX_op_goto_ptr:
2312        tcg_out_b_reg(s, COND_AL, args[0]);
2313        break;
2314    case INDEX_op_br:
2315        tcg_out_goto_label(s, COND_AL, arg_label(args[0]));
2316        break;
2317
2318    case INDEX_op_ld8u_i32:
2319        tcg_out_ld8u(s, COND_AL, args[0], args[1], args[2]);
2320        break;
2321    case INDEX_op_ld8s_i32:
2322        tcg_out_ld8s(s, COND_AL, args[0], args[1], args[2]);
2323        break;
2324    case INDEX_op_ld16u_i32:
2325        tcg_out_ld16u(s, COND_AL, args[0], args[1], args[2]);
2326        break;
2327    case INDEX_op_ld16s_i32:
2328        tcg_out_ld16s(s, COND_AL, args[0], args[1], args[2]);
2329        break;
2330    case INDEX_op_ld_i32:
2331        tcg_out_ld32u(s, COND_AL, args[0], args[1], args[2]);
2332        break;
2333    case INDEX_op_st8_i32:
2334        tcg_out_st8(s, COND_AL, args[0], args[1], args[2]);
2335        break;
2336    case INDEX_op_st16_i32:
2337        tcg_out_st16(s, COND_AL, args[0], args[1], args[2]);
2338        break;
2339    case INDEX_op_st_i32:
2340        tcg_out_st32(s, COND_AL, args[0], args[1], args[2]);
2341        break;
2342
2343    case INDEX_op_add2_i32:
2344        a0 = args[0], a1 = args[1], a2 = args[2];
2345        a3 = args[3], a4 = args[4], a5 = args[5];
2346        if (a0 == a3 || (a0 == a5 && !const_args[5])) {
2347            a0 = TCG_REG_TMP;
2348        }
2349        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD | TO_CPSR, ARITH_SUB | TO_CPSR,
2350                        a0, a2, a4, const_args[4]);
2351        tcg_out_dat_rIK(s, COND_AL, ARITH_ADC, ARITH_SBC,
2352                        a1, a3, a5, const_args[5]);
2353        tcg_out_mov_reg(s, COND_AL, args[0], a0);
2354        break;
2355    case INDEX_op_sub2_i32:
2356        a0 = args[0], a1 = args[1], a2 = args[2];
2357        a3 = args[3], a4 = args[4], a5 = args[5];
2358        if ((a0 == a3 && !const_args[3]) || (a0 == a5 && !const_args[5])) {
2359            a0 = TCG_REG_TMP;
2360        }
2361        if (const_args[2]) {
2362            if (const_args[4]) {
2363                tcg_out_movi32(s, COND_AL, a0, a4);
2364                a4 = a0;
2365            }
2366            tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR, a0, a4, a2, 1);
2367        } else {
2368            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB | TO_CPSR,
2369                            ARITH_ADD | TO_CPSR, a0, a2, a4, const_args[4]);
2370        }
2371        if (const_args[3]) {
2372            if (const_args[5]) {
2373                tcg_out_movi32(s, COND_AL, a1, a5);
2374                a5 = a1;
2375            }
2376            tcg_out_dat_rI(s, COND_AL, ARITH_RSC, a1, a5, a3, 1);
2377        } else {
2378            tcg_out_dat_rIK(s, COND_AL, ARITH_SBC, ARITH_ADC,
2379                            a1, a3, a5, const_args[5]);
2380        }
2381        tcg_out_mov_reg(s, COND_AL, args[0], a0);
2382        break;
2383
2384    case INDEX_op_qemu_ld_i32:
2385        tcg_out_qemu_ld(s, args[0], -1, args[1], args[2], TCG_TYPE_I32);
2386        break;
2387    case INDEX_op_qemu_ld_i64:
2388        tcg_out_qemu_ld(s, args[0], args[1], args[2], args[3], TCG_TYPE_I64);
2389        break;
2390
2391    case INDEX_op_qemu_st_i32:
2392        tcg_out_qemu_st(s, args[0], -1, args[1], args[2], TCG_TYPE_I32);
2393        break;
2394    case INDEX_op_qemu_st_i64:
2395        tcg_out_qemu_st(s, args[0], args[1], args[2], args[3], TCG_TYPE_I64);
2396        break;
2397
2398    case INDEX_op_deposit_i32:
2399        tcg_out_deposit(s, COND_AL, args[0], args[2],
2400                        args[3], args[4], const_args[2]);
2401        break;
2402    case INDEX_op_sextract_i32:
2403        tcg_out_sextract(s, COND_AL, args[0], args[1], args[2], args[3]);
2404        break;
2405    case INDEX_op_extract2_i32:
2406        /* ??? These optimization vs zero should be generic.  */
2407        /* ??? But we can't substitute 2 for 1 in the opcode stream yet.  */
2408        if (const_args[1]) {
2409            if (const_args[2]) {
2410                tcg_out_movi(s, TCG_TYPE_REG, args[0], 0);
2411            } else {
2412                tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2413                                args[2], SHIFT_IMM_LSL(32 - args[3]));
2414            }
2415        } else if (const_args[2]) {
2416            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2417                            args[1], SHIFT_IMM_LSR(args[3]));
2418        } else {
2419            /* We can do extract2 in 2 insns, vs the 3 required otherwise.  */
2420            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0,
2421                            args[2], SHIFT_IMM_LSL(32 - args[3]));
2422            tcg_out_dat_reg(s, COND_AL, ARITH_ORR, args[0], TCG_REG_TMP,
2423                            args[1], SHIFT_IMM_LSR(args[3]));
2424        }
2425        break;
2426
2427    case INDEX_op_mb:
2428        tcg_out_mb(s, args[0]);
2429        break;
2430
2431    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
2432    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
2433    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
2434    default:
2435        g_assert_not_reached();
2436    }
2437}
2438
2439static TCGConstraintSetIndex
2440tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
2441{
2442    switch (op) {
2443    case INDEX_op_goto_ptr:
2444        return C_O0_I1(r);
2445
2446    case INDEX_op_ld8u_i32:
2447    case INDEX_op_ld8s_i32:
2448    case INDEX_op_ld16u_i32:
2449    case INDEX_op_ld16s_i32:
2450    case INDEX_op_ld_i32:
2451    case INDEX_op_sextract_i32:
2452        return C_O1_I1(r, r);
2453
2454    case INDEX_op_st8_i32:
2455    case INDEX_op_st16_i32:
2456    case INDEX_op_st_i32:
2457        return C_O0_I2(r, r);
2458
2459    case INDEX_op_deposit_i32:
2460        return C_O1_I2(r, 0, rZ);
2461    case INDEX_op_extract2_i32:
2462        return C_O1_I2(r, rZ, rZ);
2463    case INDEX_op_add2_i32:
2464        return C_O2_I4(r, r, r, r, rIN, rIK);
2465    case INDEX_op_sub2_i32:
2466        return C_O2_I4(r, r, rI, rI, rIN, rIK);
2467    case INDEX_op_qemu_ld_i32:
2468        return C_O1_I1(r, q);
2469    case INDEX_op_qemu_ld_i64:
2470        return C_O2_I1(e, p, q);
2471    case INDEX_op_qemu_st_i32:
2472        return C_O0_I2(q, q);
2473    case INDEX_op_qemu_st_i64:
2474        return C_O0_I3(Q, p, q);
2475
2476    case INDEX_op_st_vec:
2477        return C_O0_I2(w, r);
2478    case INDEX_op_ld_vec:
2479    case INDEX_op_dupm_vec:
2480        return C_O1_I1(w, r);
2481    case INDEX_op_dup_vec:
2482        return C_O1_I1(w, wr);
2483    case INDEX_op_abs_vec:
2484    case INDEX_op_neg_vec:
2485    case INDEX_op_not_vec:
2486    case INDEX_op_shli_vec:
2487    case INDEX_op_shri_vec:
2488    case INDEX_op_sari_vec:
2489        return C_O1_I1(w, w);
2490    case INDEX_op_dup2_vec:
2491    case INDEX_op_add_vec:
2492    case INDEX_op_mul_vec:
2493    case INDEX_op_smax_vec:
2494    case INDEX_op_smin_vec:
2495    case INDEX_op_ssadd_vec:
2496    case INDEX_op_sssub_vec:
2497    case INDEX_op_sub_vec:
2498    case INDEX_op_umax_vec:
2499    case INDEX_op_umin_vec:
2500    case INDEX_op_usadd_vec:
2501    case INDEX_op_ussub_vec:
2502    case INDEX_op_xor_vec:
2503    case INDEX_op_arm_sshl_vec:
2504    case INDEX_op_arm_ushl_vec:
2505        return C_O1_I2(w, w, w);
2506    case INDEX_op_arm_sli_vec:
2507        return C_O1_I2(w, 0, w);
2508    case INDEX_op_or_vec:
2509    case INDEX_op_andc_vec:
2510        return C_O1_I2(w, w, wO);
2511    case INDEX_op_and_vec:
2512    case INDEX_op_orc_vec:
2513        return C_O1_I2(w, w, wV);
2514    case INDEX_op_cmp_vec:
2515        return C_O1_I2(w, w, wZ);
2516    case INDEX_op_bitsel_vec:
2517        return C_O1_I3(w, w, w, w);
2518    default:
2519        return C_NotImplemented;
2520    }
2521}
2522
2523static void tcg_target_init(TCGContext *s)
2524{
2525    /*
2526     * Only probe for the platform and capabilities if we haven't already
2527     * determined maximum values at compile time.
2528     */
2529#if !defined(use_idiv_instructions) || !defined(use_neon_instructions)
2530    {
2531        unsigned long hwcap = qemu_getauxval(AT_HWCAP);
2532#ifndef use_idiv_instructions
2533        use_idiv_instructions = (hwcap & HWCAP_ARM_IDIVA) != 0;
2534#endif
2535#ifndef use_neon_instructions
2536        use_neon_instructions = (hwcap & HWCAP_ARM_NEON) != 0;
2537#endif
2538    }
2539#endif
2540
2541    if (__ARM_ARCH < 7) {
2542        const char *pl = (const char *)qemu_getauxval(AT_PLATFORM);
2543        if (pl != NULL && pl[0] == 'v' && pl[1] >= '4' && pl[1] <= '9') {
2544            arm_arch = pl[1] - '0';
2545        }
2546
2547        if (arm_arch < 6) {
2548            error_report("TCG: ARMv%d is unsupported; exiting", arm_arch);
2549            exit(EXIT_FAILURE);
2550        }
2551    }
2552
2553    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
2554
2555    tcg_target_call_clobber_regs = 0;
2556    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R0);
2557    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R1);
2558    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R2);
2559    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R3);
2560    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R12);
2561    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R14);
2562
2563    if (use_neon_instructions) {
2564        tcg_target_available_regs[TCG_TYPE_V64]  = ALL_VECTOR_REGS;
2565        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
2566
2567        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q0);
2568        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q1);
2569        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q2);
2570        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q3);
2571        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q8);
2572        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q9);
2573        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q10);
2574        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q11);
2575        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q12);
2576        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q13);
2577        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q14);
2578        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q15);
2579    }
2580
2581    s->reserved_regs = 0;
2582    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
2583    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP);
2584    tcg_regset_set_reg(s->reserved_regs, TCG_REG_PC);
2585    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP);
2586}
2587
2588static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
2589                       TCGReg arg1, intptr_t arg2)
2590{
2591    switch (type) {
2592    case TCG_TYPE_I32:
2593        tcg_out_ld32u(s, COND_AL, arg, arg1, arg2);
2594        return;
2595    case TCG_TYPE_V64:
2596        /* regs 1; size 8; align 8 */
2597        tcg_out_vldst(s, INSN_VLD1 | 0x7d0, arg, arg1, arg2);
2598        return;
2599    case TCG_TYPE_V128:
2600        /*
2601         * We have only 8-byte alignment for the stack per the ABI.
2602         * Rather than dynamically re-align the stack, it's easier
2603         * to simply not request alignment beyond that.  So:
2604         * regs 2; size 8; align 8
2605         */
2606        tcg_out_vldst(s, INSN_VLD1 | 0xad0, arg, arg1, arg2);
2607        return;
2608    default:
2609        g_assert_not_reached();
2610    }
2611}
2612
2613static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
2614                       TCGReg arg1, intptr_t arg2)
2615{
2616    switch (type) {
2617    case TCG_TYPE_I32:
2618        tcg_out_st32(s, COND_AL, arg, arg1, arg2);
2619        return;
2620    case TCG_TYPE_V64:
2621        /* regs 1; size 8; align 8 */
2622        tcg_out_vldst(s, INSN_VST1 | 0x7d0, arg, arg1, arg2);
2623        return;
2624    case TCG_TYPE_V128:
2625        /* See tcg_out_ld re alignment: regs 2; size 8; align 8 */
2626        tcg_out_vldst(s, INSN_VST1 | 0xad0, arg, arg1, arg2);
2627        return;
2628    default:
2629        g_assert_not_reached();
2630    }
2631}
2632
2633static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
2634                        TCGReg base, intptr_t ofs)
2635{
2636    return false;
2637}
2638
2639static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
2640{
2641    if (ret == arg) {
2642        return true;
2643    }
2644    switch (type) {
2645    case TCG_TYPE_I32:
2646        if (ret < TCG_REG_Q0 && arg < TCG_REG_Q0) {
2647            tcg_out_mov_reg(s, COND_AL, ret, arg);
2648            return true;
2649        }
2650        return false;
2651
2652    case TCG_TYPE_V64:
2653    case TCG_TYPE_V128:
2654        /* "VMOV D,N" is an alias for "VORR D,N,N". */
2655        tcg_out_vreg3(s, INSN_VORR, type - TCG_TYPE_V64, 0, ret, arg, arg);
2656        return true;
2657
2658    default:
2659        g_assert_not_reached();
2660    }
2661}
2662
2663static void tcg_out_movi(TCGContext *s, TCGType type,
2664                         TCGReg ret, tcg_target_long arg)
2665{
2666    tcg_debug_assert(type == TCG_TYPE_I32);
2667    tcg_debug_assert(ret < TCG_REG_Q0);
2668    tcg_out_movi32(s, COND_AL, ret, arg);
2669}
2670
2671static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
2672{
2673    return false;
2674}
2675
2676static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
2677                             tcg_target_long imm)
2678{
2679    int enc, opc = ARITH_ADD;
2680
2681    /* All of the easiest immediates to encode are positive. */
2682    if (imm < 0) {
2683        imm = -imm;
2684        opc = ARITH_SUB;
2685    }
2686    enc = encode_imm(imm);
2687    if (enc >= 0) {
2688        tcg_out_dat_imm(s, COND_AL, opc, rd, rs, enc);
2689    } else {
2690        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, imm);
2691        tcg_out_dat_reg(s, COND_AL, opc, rd, rs,
2692                        TCG_REG_TMP, SHIFT_IMM_LSL(0));
2693    }
2694}
2695
2696/* Type is always V128, with I64 elements.  */
2697static void tcg_out_dup2_vec(TCGContext *s, TCGReg rd, TCGReg rl, TCGReg rh)
2698{
2699    /* Move high element into place first. */
2700    /* VMOV Dd+1, Ds */
2701    tcg_out_vreg3(s, INSN_VORR | (1 << 12), 0, 0, rd, rh, rh);
2702    /* Move low element into place; tcg_out_mov will check for nop. */
2703    tcg_out_mov(s, TCG_TYPE_V64, rd, rl);
2704}
2705
2706static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
2707                            TCGReg rd, TCGReg rs)
2708{
2709    int q = type - TCG_TYPE_V64;
2710
2711    if (vece == MO_64) {
2712        if (type == TCG_TYPE_V128) {
2713            tcg_out_dup2_vec(s, rd, rs, rs);
2714        } else {
2715            tcg_out_mov(s, TCG_TYPE_V64, rd, rs);
2716        }
2717    } else if (rs < TCG_REG_Q0) {
2718        int b = (vece == MO_8);
2719        int e = (vece == MO_16);
2720        tcg_out32(s, INSN_VDUP_G | (b << 22) | (q << 21) | (e << 5) |
2721                  encode_vn(rd) | (rs << 12));
2722    } else {
2723        int imm4 = 1 << vece;
2724        tcg_out32(s, INSN_VDUP_S | (imm4 << 16) | (q << 6) |
2725                  encode_vd(rd) | encode_vm(rs));
2726    }
2727    return true;
2728}
2729
2730static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
2731                             TCGReg rd, TCGReg base, intptr_t offset)
2732{
2733    if (vece == MO_64) {
2734        tcg_out_ld(s, TCG_TYPE_V64, rd, base, offset);
2735        if (type == TCG_TYPE_V128) {
2736            tcg_out_dup2_vec(s, rd, rd, rd);
2737        }
2738    } else {
2739        int q = type - TCG_TYPE_V64;
2740        tcg_out_vldst(s, INSN_VLD1R | (vece << 6) | (q << 5),
2741                      rd, base, offset);
2742    }
2743    return true;
2744}
2745
2746static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
2747                             TCGReg rd, int64_t v64)
2748{
2749    int q = type - TCG_TYPE_V64;
2750    int cmode, imm8, i;
2751
2752    /* Test all bytes equal first.  */
2753    if (vece == MO_8) {
2754        tcg_out_vmovi(s, rd, q, 0, 0xe, v64);
2755        return;
2756    }
2757
2758    /*
2759     * Test all bytes 0x00 or 0xff second.  This can match cases that
2760     * might otherwise take 2 or 3 insns for MO_16 or MO_32 below.
2761     */
2762    for (i = imm8 = 0; i < 8; i++) {
2763        uint8_t byte = v64 >> (i * 8);
2764        if (byte == 0xff) {
2765            imm8 |= 1 << i;
2766        } else if (byte != 0) {
2767            goto fail_bytes;
2768        }
2769    }
2770    tcg_out_vmovi(s, rd, q, 1, 0xe, imm8);
2771    return;
2772 fail_bytes:
2773
2774    /*
2775     * Tests for various replications.  For each element width, if we
2776     * cannot find an expansion there's no point checking a larger
2777     * width because we already know by replication it cannot match.
2778     */
2779    if (vece == MO_16) {
2780        uint16_t v16 = v64;
2781
2782        if (is_shimm16(v16, &cmode, &imm8)) {
2783            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2784            return;
2785        }
2786        if (is_shimm16(~v16, &cmode, &imm8)) {
2787            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2788            return;
2789        }
2790
2791        /*
2792         * Otherwise, all remaining constants can be loaded in two insns:
2793         * rd = v16 & 0xff, rd |= v16 & 0xff00.
2794         */
2795        tcg_out_vmovi(s, rd, q, 0, 0x8, v16 & 0xff);
2796        tcg_out_vmovi(s, rd, q, 0, 0xb, v16 >> 8);   /* VORRI */
2797        return;
2798    }
2799
2800    if (vece == MO_32) {
2801        uint32_t v32 = v64;
2802
2803        if (is_shimm32(v32, &cmode, &imm8) ||
2804            is_soimm32(v32, &cmode, &imm8)) {
2805            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2806            return;
2807        }
2808        if (is_shimm32(~v32, &cmode, &imm8) ||
2809            is_soimm32(~v32, &cmode, &imm8)) {
2810            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2811            return;
2812        }
2813
2814        /*
2815         * Restrict the set of constants to those we can load with
2816         * two instructions.  Others we load from the pool.
2817         */
2818        i = is_shimm32_pair(v32, &cmode, &imm8);
2819        if (i) {
2820            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2821            tcg_out_vmovi(s, rd, q, 0, i | 1, extract32(v32, i * 4, 8));
2822            return;
2823        }
2824        i = is_shimm32_pair(~v32, &cmode, &imm8);
2825        if (i) {
2826            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2827            tcg_out_vmovi(s, rd, q, 1, i | 1, extract32(~v32, i * 4, 8));
2828            return;
2829        }
2830    }
2831
2832    /*
2833     * As a last resort, load from the constant pool.
2834     */
2835    if (!q || vece == MO_64) {
2836        new_pool_l2(s, R_ARM_PC11, s->code_ptr, 0, v64, v64 >> 32);
2837        /* VLDR Dd, [pc + offset] */
2838        tcg_out32(s, INSN_VLDR_D | encode_vd(rd) | (0xf << 16));
2839        if (q) {
2840            tcg_out_dup2_vec(s, rd, rd, rd);
2841        }
2842    } else {
2843        new_pool_label(s, (uint32_t)v64, R_ARM_PC8, s->code_ptr, 0);
2844        /* add tmp, pc, offset */
2845        tcg_out_dat_imm(s, COND_AL, ARITH_ADD, TCG_REG_TMP, TCG_REG_PC, 0);
2846        tcg_out_dupm_vec(s, type, MO_32, rd, TCG_REG_TMP, 0);
2847    }
2848}
2849
2850static const ARMInsn vec_cmp_insn[16] = {
2851    [TCG_COND_EQ] = INSN_VCEQ,
2852    [TCG_COND_GT] = INSN_VCGT,
2853    [TCG_COND_GE] = INSN_VCGE,
2854    [TCG_COND_GTU] = INSN_VCGT_U,
2855    [TCG_COND_GEU] = INSN_VCGE_U,
2856};
2857
2858static const ARMInsn vec_cmp0_insn[16] = {
2859    [TCG_COND_EQ] = INSN_VCEQ0,
2860    [TCG_COND_GT] = INSN_VCGT0,
2861    [TCG_COND_GE] = INSN_VCGE0,
2862    [TCG_COND_LT] = INSN_VCLT0,
2863    [TCG_COND_LE] = INSN_VCLE0,
2864};
2865
2866static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2867                           unsigned vecl, unsigned vece,
2868                           const TCGArg args[TCG_MAX_OP_ARGS],
2869                           const int const_args[TCG_MAX_OP_ARGS])
2870{
2871    TCGType type = vecl + TCG_TYPE_V64;
2872    unsigned q = vecl;
2873    TCGArg a0, a1, a2, a3;
2874    int cmode, imm8;
2875
2876    a0 = args[0];
2877    a1 = args[1];
2878    a2 = args[2];
2879
2880    switch (opc) {
2881    case INDEX_op_ld_vec:
2882        tcg_out_ld(s, type, a0, a1, a2);
2883        return;
2884    case INDEX_op_st_vec:
2885        tcg_out_st(s, type, a0, a1, a2);
2886        return;
2887    case INDEX_op_dupm_vec:
2888        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
2889        return;
2890    case INDEX_op_dup2_vec:
2891        tcg_out_dup2_vec(s, a0, a1, a2);
2892        return;
2893    case INDEX_op_abs_vec:
2894        tcg_out_vreg2(s, INSN_VABS, q, vece, a0, a1);
2895        return;
2896    case INDEX_op_neg_vec:
2897        tcg_out_vreg2(s, INSN_VNEG, q, vece, a0, a1);
2898        return;
2899    case INDEX_op_not_vec:
2900        tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a1);
2901        return;
2902    case INDEX_op_add_vec:
2903        tcg_out_vreg3(s, INSN_VADD, q, vece, a0, a1, a2);
2904        return;
2905    case INDEX_op_mul_vec:
2906        tcg_out_vreg3(s, INSN_VMUL, q, vece, a0, a1, a2);
2907        return;
2908    case INDEX_op_smax_vec:
2909        tcg_out_vreg3(s, INSN_VMAX, q, vece, a0, a1, a2);
2910        return;
2911    case INDEX_op_smin_vec:
2912        tcg_out_vreg3(s, INSN_VMIN, q, vece, a0, a1, a2);
2913        return;
2914    case INDEX_op_sub_vec:
2915        tcg_out_vreg3(s, INSN_VSUB, q, vece, a0, a1, a2);
2916        return;
2917    case INDEX_op_ssadd_vec:
2918        tcg_out_vreg3(s, INSN_VQADD, q, vece, a0, a1, a2);
2919        return;
2920    case INDEX_op_sssub_vec:
2921        tcg_out_vreg3(s, INSN_VQSUB, q, vece, a0, a1, a2);
2922        return;
2923    case INDEX_op_umax_vec:
2924        tcg_out_vreg3(s, INSN_VMAX_U, q, vece, a0, a1, a2);
2925        return;
2926    case INDEX_op_umin_vec:
2927        tcg_out_vreg3(s, INSN_VMIN_U, q, vece, a0, a1, a2);
2928        return;
2929    case INDEX_op_usadd_vec:
2930        tcg_out_vreg3(s, INSN_VQADD_U, q, vece, a0, a1, a2);
2931        return;
2932    case INDEX_op_ussub_vec:
2933        tcg_out_vreg3(s, INSN_VQSUB_U, q, vece, a0, a1, a2);
2934        return;
2935    case INDEX_op_xor_vec:
2936        tcg_out_vreg3(s, INSN_VEOR, q, 0, a0, a1, a2);
2937        return;
2938    case INDEX_op_arm_sshl_vec:
2939        /*
2940         * Note that Vm is the data and Vn is the shift count,
2941         * therefore the arguments appear reversed.
2942         */
2943        tcg_out_vreg3(s, INSN_VSHL_S, q, vece, a0, a2, a1);
2944        return;
2945    case INDEX_op_arm_ushl_vec:
2946        /* See above. */
2947        tcg_out_vreg3(s, INSN_VSHL_U, q, vece, a0, a2, a1);
2948        return;
2949    case INDEX_op_shli_vec:
2950        tcg_out_vshifti(s, INSN_VSHLI, q, a0, a1, a2 + (8 << vece));
2951        return;
2952    case INDEX_op_shri_vec:
2953        tcg_out_vshifti(s, INSN_VSHRI, q, a0, a1, (16 << vece) - a2);
2954        return;
2955    case INDEX_op_sari_vec:
2956        tcg_out_vshifti(s, INSN_VSARI, q, a0, a1, (16 << vece) - a2);
2957        return;
2958    case INDEX_op_arm_sli_vec:
2959        tcg_out_vshifti(s, INSN_VSLI, q, a0, a2, args[3] + (8 << vece));
2960        return;
2961
2962    case INDEX_op_andc_vec:
2963        if (!const_args[2]) {
2964            tcg_out_vreg3(s, INSN_VBIC, q, 0, a0, a1, a2);
2965            return;
2966        }
2967        a2 = ~a2;
2968        /* fall through */
2969    case INDEX_op_and_vec:
2970        if (const_args[2]) {
2971            is_shimm1632(~a2, &cmode, &imm8);
2972            if (a0 == a1) {
2973                tcg_out_vmovi(s, a0, q, 1, cmode | 1, imm8); /* VBICI */
2974                return;
2975            }
2976            tcg_out_vmovi(s, a0, q, 1, cmode, imm8); /* VMVNI */
2977            a2 = a0;
2978        }
2979        tcg_out_vreg3(s, INSN_VAND, q, 0, a0, a1, a2);
2980        return;
2981
2982    case INDEX_op_orc_vec:
2983        if (!const_args[2]) {
2984            tcg_out_vreg3(s, INSN_VORN, q, 0, a0, a1, a2);
2985            return;
2986        }
2987        a2 = ~a2;
2988        /* fall through */
2989    case INDEX_op_or_vec:
2990        if (const_args[2]) {
2991            is_shimm1632(a2, &cmode, &imm8);
2992            if (a0 == a1) {
2993                tcg_out_vmovi(s, a0, q, 0, cmode | 1, imm8); /* VORRI */
2994                return;
2995            }
2996            tcg_out_vmovi(s, a0, q, 0, cmode, imm8); /* VMOVI */
2997            a2 = a0;
2998        }
2999        tcg_out_vreg3(s, INSN_VORR, q, 0, a0, a1, a2);
3000        return;
3001
3002    case INDEX_op_cmp_vec:
3003        {
3004            TCGCond cond = args[3];
3005            ARMInsn insn;
3006
3007            switch (cond) {
3008            case TCG_COND_NE:
3009                if (const_args[2]) {
3010                    tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a1);
3011                } else {
3012                    tcg_out_vreg3(s, INSN_VCEQ, q, vece, a0, a1, a2);
3013                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
3014                }
3015                break;
3016
3017            case TCG_COND_TSTNE:
3018            case TCG_COND_TSTEQ:
3019                if (const_args[2]) {
3020                    /* (x & 0) == 0 */
3021                    tcg_out_dupi_vec(s, type, MO_8, a0,
3022                                     -(cond == TCG_COND_TSTEQ));
3023                    break;
3024                }
3025                tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a2);
3026                if (cond == TCG_COND_TSTEQ) {
3027                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
3028                }
3029                break;
3030
3031            default:
3032                if (const_args[2]) {
3033                    insn = vec_cmp0_insn[cond];
3034                    if (insn) {
3035                        tcg_out_vreg2(s, insn, q, vece, a0, a1);
3036                        return;
3037                    }
3038                    tcg_out_dupi_vec(s, type, MO_8, TCG_VEC_TMP, 0);
3039                    a2 = TCG_VEC_TMP;
3040                }
3041                insn = vec_cmp_insn[cond];
3042                if (insn == 0) {
3043                    TCGArg t;
3044                    t = a1, a1 = a2, a2 = t;
3045                    cond = tcg_swap_cond(cond);
3046                    insn = vec_cmp_insn[cond];
3047                    tcg_debug_assert(insn != 0);
3048                }
3049                tcg_out_vreg3(s, insn, q, vece, a0, a1, a2);
3050                break;
3051            }
3052        }
3053        return;
3054
3055    case INDEX_op_bitsel_vec:
3056        a3 = args[3];
3057        if (a0 == a3) {
3058            tcg_out_vreg3(s, INSN_VBIT, q, 0, a0, a2, a1);
3059        } else if (a0 == a2) {
3060            tcg_out_vreg3(s, INSN_VBIF, q, 0, a0, a3, a1);
3061        } else {
3062            tcg_out_mov(s, type, a0, a1);
3063            tcg_out_vreg3(s, INSN_VBSL, q, 0, a0, a2, a3);
3064        }
3065        return;
3066
3067    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
3068    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
3069    default:
3070        g_assert_not_reached();
3071    }
3072}
3073
3074int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
3075{
3076    switch (opc) {
3077    case INDEX_op_add_vec:
3078    case INDEX_op_sub_vec:
3079    case INDEX_op_and_vec:
3080    case INDEX_op_andc_vec:
3081    case INDEX_op_or_vec:
3082    case INDEX_op_orc_vec:
3083    case INDEX_op_xor_vec:
3084    case INDEX_op_not_vec:
3085    case INDEX_op_shli_vec:
3086    case INDEX_op_shri_vec:
3087    case INDEX_op_sari_vec:
3088    case INDEX_op_ssadd_vec:
3089    case INDEX_op_sssub_vec:
3090    case INDEX_op_usadd_vec:
3091    case INDEX_op_ussub_vec:
3092    case INDEX_op_bitsel_vec:
3093        return 1;
3094    case INDEX_op_abs_vec:
3095    case INDEX_op_cmp_vec:
3096    case INDEX_op_mul_vec:
3097    case INDEX_op_neg_vec:
3098    case INDEX_op_smax_vec:
3099    case INDEX_op_smin_vec:
3100    case INDEX_op_umax_vec:
3101    case INDEX_op_umin_vec:
3102        return vece < MO_64;
3103    case INDEX_op_shlv_vec:
3104    case INDEX_op_shrv_vec:
3105    case INDEX_op_sarv_vec:
3106    case INDEX_op_rotli_vec:
3107    case INDEX_op_rotlv_vec:
3108    case INDEX_op_rotrv_vec:
3109        return -1;
3110    default:
3111        return 0;
3112    }
3113}
3114
3115void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
3116                       TCGArg a0, ...)
3117{
3118    va_list va;
3119    TCGv_vec v0, v1, v2, t1, t2, c1;
3120    TCGArg a2;
3121
3122    va_start(va, a0);
3123    v0 = temp_tcgv_vec(arg_temp(a0));
3124    v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
3125    a2 = va_arg(va, TCGArg);
3126    va_end(va);
3127
3128    switch (opc) {
3129    case INDEX_op_shlv_vec:
3130        /*
3131         * Merely propagate shlv_vec to arm_ushl_vec.
3132         * In this way we don't set TCG_TARGET_HAS_shv_vec
3133         * because everything is done via expansion.
3134         */
3135        v2 = temp_tcgv_vec(arg_temp(a2));
3136        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
3137                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
3138        break;
3139
3140    case INDEX_op_shrv_vec:
3141    case INDEX_op_sarv_vec:
3142        /* Right shifts are negative left shifts for NEON.  */
3143        v2 = temp_tcgv_vec(arg_temp(a2));
3144        t1 = tcg_temp_new_vec(type);
3145        tcg_gen_neg_vec(vece, t1, v2);
3146        if (opc == INDEX_op_shrv_vec) {
3147            opc = INDEX_op_arm_ushl_vec;
3148        } else {
3149            opc = INDEX_op_arm_sshl_vec;
3150        }
3151        vec_gen_3(opc, type, vece, tcgv_vec_arg(v0),
3152                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
3153        tcg_temp_free_vec(t1);
3154        break;
3155
3156    case INDEX_op_rotli_vec:
3157        t1 = tcg_temp_new_vec(type);
3158        tcg_gen_shri_vec(vece, t1, v1, -a2 & ((8 << vece) - 1));
3159        vec_gen_4(INDEX_op_arm_sli_vec, type, vece,
3160                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(v1), a2);
3161        tcg_temp_free_vec(t1);
3162        break;
3163
3164    case INDEX_op_rotlv_vec:
3165        v2 = temp_tcgv_vec(arg_temp(a2));
3166        t1 = tcg_temp_new_vec(type);
3167        c1 = tcg_constant_vec(type, vece, 8 << vece);
3168        tcg_gen_sub_vec(vece, t1, v2, c1);
3169        /* Right shifts are negative left shifts for NEON.  */
3170        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
3171                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
3172        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
3173                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
3174        tcg_gen_or_vec(vece, v0, v0, t1);
3175        tcg_temp_free_vec(t1);
3176        break;
3177
3178    case INDEX_op_rotrv_vec:
3179        v2 = temp_tcgv_vec(arg_temp(a2));
3180        t1 = tcg_temp_new_vec(type);
3181        t2 = tcg_temp_new_vec(type);
3182        c1 = tcg_constant_vec(type, vece, 8 << vece);
3183        tcg_gen_neg_vec(vece, t1, v2);
3184        tcg_gen_sub_vec(vece, t2, c1, v2);
3185        /* Right shifts are negative left shifts for NEON.  */
3186        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
3187                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
3188        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t2),
3189                  tcgv_vec_arg(v1), tcgv_vec_arg(t2));
3190        tcg_gen_or_vec(vece, v0, t1, t2);
3191        tcg_temp_free_vec(t1);
3192        tcg_temp_free_vec(t2);
3193        break;
3194
3195    default:
3196        g_assert_not_reached();
3197    }
3198}
3199
3200static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
3201{
3202    int i;
3203    for (i = 0; i < count; ++i) {
3204        p[i] = INSN_NOP;
3205    }
3206}
3207
3208/* Compute frame size via macros, to share between tcg_target_qemu_prologue
3209   and tcg_register_jit.  */
3210
3211#define PUSH_SIZE  ((11 - 4 + 1 + 1) * sizeof(tcg_target_long))
3212
3213#define FRAME_SIZE \
3214    ((PUSH_SIZE \
3215      + TCG_STATIC_CALL_ARGS_SIZE \
3216      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
3217      + TCG_TARGET_STACK_ALIGN - 1) \
3218     & -TCG_TARGET_STACK_ALIGN)
3219
3220#define STACK_ADDEND  (FRAME_SIZE - PUSH_SIZE)
3221
3222static void tcg_target_qemu_prologue(TCGContext *s)
3223{
3224    /* Calling convention requires us to save r4-r11 and lr.  */
3225    /* stmdb sp!, { r4 - r11, lr } */
3226    tcg_out_ldstm(s, COND_AL, INSN_STMDB, TCG_REG_CALL_STACK,
3227                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
3228                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3229                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_R14));
3230
3231    /* Reserve callee argument and tcg temp space.  */
3232    tcg_out_dat_rI(s, COND_AL, ARITH_SUB, TCG_REG_CALL_STACK,
3233                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
3234    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
3235                  CPU_TEMP_BUF_NLONGS * sizeof(long));
3236
3237    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
3238
3239    if (!tcg_use_softmmu && guest_base) {
3240        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_GUEST_BASE, guest_base);
3241        tcg_regset_set_reg(s->reserved_regs, TCG_REG_GUEST_BASE);
3242    }
3243
3244    tcg_out_b_reg(s, COND_AL, tcg_target_call_iarg_regs[1]);
3245
3246    /*
3247     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
3248     * and fall through to the rest of the epilogue.
3249     */
3250    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
3251    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, 0);
3252    tcg_out_epilogue(s);
3253}
3254
3255static void tcg_out_epilogue(TCGContext *s)
3256{
3257    /* Release local stack frame.  */
3258    tcg_out_dat_rI(s, COND_AL, ARITH_ADD, TCG_REG_CALL_STACK,
3259                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
3260
3261    /* ldmia sp!, { r4 - r11, pc } */
3262    tcg_out_ldstm(s, COND_AL, INSN_LDMIA, TCG_REG_CALL_STACK,
3263                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
3264                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3265                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_PC));
3266}
3267
3268static void tcg_out_tb_start(TCGContext *s)
3269{
3270    /* nothing to do */
3271}
3272
3273typedef struct {
3274    DebugFrameHeader h;
3275    uint8_t fde_def_cfa[4];
3276    uint8_t fde_reg_ofs[18];
3277} DebugFrame;
3278
3279#define ELF_HOST_MACHINE EM_ARM
3280
3281/* We're expecting a 2 byte uleb128 encoded value.  */
3282QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
3283
3284static const DebugFrame debug_frame = {
3285    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
3286    .h.cie.id = -1,
3287    .h.cie.version = 1,
3288    .h.cie.code_align = 1,
3289    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
3290    .h.cie.return_column = 14,
3291
3292    /* Total FDE size does not include the "len" member.  */
3293    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
3294
3295    .fde_def_cfa = {
3296        12, 13,                         /* DW_CFA_def_cfa sp, ... */
3297        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
3298        (FRAME_SIZE >> 7)
3299    },
3300    .fde_reg_ofs = {
3301        /* The following must match the stmdb in the prologue.  */
3302        0x8e, 1,                        /* DW_CFA_offset, lr, -4 */
3303        0x8b, 2,                        /* DW_CFA_offset, r11, -8 */
3304        0x8a, 3,                        /* DW_CFA_offset, r10, -12 */
3305        0x89, 4,                        /* DW_CFA_offset, r9, -16 */
3306        0x88, 5,                        /* DW_CFA_offset, r8, -20 */
3307        0x87, 6,                        /* DW_CFA_offset, r7, -24 */
3308        0x86, 7,                        /* DW_CFA_offset, r6, -28 */
3309        0x85, 8,                        /* DW_CFA_offset, r5, -32 */
3310        0x84, 9,                        /* DW_CFA_offset, r4, -36 */
3311    }
3312};
3313
3314void tcg_register_jit(const void *buf, size_t buf_size)
3315{
3316    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
3317}
3318