xref: /openbmc/qemu/tcg/arm/tcg-target.c.inc (revision 197a1372)
1/*
2 * Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2008 Andrzej Zaborowski
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to deal
8 * in the Software without restriction, including without limitation the rights
9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10 * copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22 * THE SOFTWARE.
23 */
24
25#include "elf.h"
26#include "../tcg-ldst.c.inc"
27#include "../tcg-pool.c.inc"
28
29int arm_arch = __ARM_ARCH;
30
31#ifndef use_idiv_instructions
32bool use_idiv_instructions;
33#endif
34#ifndef use_neon_instructions
35bool use_neon_instructions;
36#endif
37
38#ifdef CONFIG_DEBUG_TCG
39static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
40    "%r0",  "%r1",  "%r2",  "%r3",  "%r4",  "%r5",  "%r6",  "%r7",
41    "%r8",  "%r9",  "%r10", "%r11", "%r12", "%sp",  "%r14", "%pc",
42    "%q0",  "%q1",  "%q2",  "%q3",  "%q4",  "%q5",  "%q6",  "%q7",
43    "%q8",  "%q9",  "%q10", "%q11", "%q12", "%q13", "%q14", "%q15",
44};
45#endif
46
47static const int tcg_target_reg_alloc_order[] = {
48    TCG_REG_R4,
49    TCG_REG_R5,
50    TCG_REG_R6,
51    TCG_REG_R7,
52    TCG_REG_R8,
53    TCG_REG_R9,
54    TCG_REG_R10,
55    TCG_REG_R11,
56    TCG_REG_R13,
57    TCG_REG_R0,
58    TCG_REG_R1,
59    TCG_REG_R2,
60    TCG_REG_R3,
61    TCG_REG_R12,
62    TCG_REG_R14,
63
64    TCG_REG_Q0,
65    TCG_REG_Q1,
66    TCG_REG_Q2,
67    TCG_REG_Q3,
68    /* Q4 - Q7 are call-saved, and skipped. */
69    TCG_REG_Q8,
70    TCG_REG_Q9,
71    TCG_REG_Q10,
72    TCG_REG_Q11,
73    TCG_REG_Q12,
74    TCG_REG_Q13,
75    TCG_REG_Q14,
76    TCG_REG_Q15,
77};
78
79static const int tcg_target_call_iarg_regs[4] = {
80    TCG_REG_R0, TCG_REG_R1, TCG_REG_R2, TCG_REG_R3
81};
82
83static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
84{
85    tcg_debug_assert(kind == TCG_CALL_RET_NORMAL);
86    tcg_debug_assert(slot >= 0 && slot <= 3);
87    return TCG_REG_R0 + slot;
88}
89
90#define TCG_REG_TMP  TCG_REG_R12
91#define TCG_VEC_TMP  TCG_REG_Q15
92#ifndef CONFIG_SOFTMMU
93#define TCG_REG_GUEST_BASE  TCG_REG_R11
94#endif
95
96typedef enum {
97    COND_EQ = 0x0,
98    COND_NE = 0x1,
99    COND_CS = 0x2,	/* Unsigned greater or equal */
100    COND_CC = 0x3,	/* Unsigned less than */
101    COND_MI = 0x4,	/* Negative */
102    COND_PL = 0x5,	/* Zero or greater */
103    COND_VS = 0x6,	/* Overflow */
104    COND_VC = 0x7,	/* No overflow */
105    COND_HI = 0x8,	/* Unsigned greater than */
106    COND_LS = 0x9,	/* Unsigned less or equal */
107    COND_GE = 0xa,
108    COND_LT = 0xb,
109    COND_GT = 0xc,
110    COND_LE = 0xd,
111    COND_AL = 0xe,
112} ARMCond;
113
114#define TO_CPSR (1 << 20)
115
116#define SHIFT_IMM_LSL(im)	(((im) << 7) | 0x00)
117#define SHIFT_IMM_LSR(im)	(((im) << 7) | 0x20)
118#define SHIFT_IMM_ASR(im)	(((im) << 7) | 0x40)
119#define SHIFT_IMM_ROR(im)	(((im) << 7) | 0x60)
120#define SHIFT_REG_LSL(rs)	(((rs) << 8) | 0x10)
121#define SHIFT_REG_LSR(rs)	(((rs) << 8) | 0x30)
122#define SHIFT_REG_ASR(rs)	(((rs) << 8) | 0x50)
123#define SHIFT_REG_ROR(rs)	(((rs) << 8) | 0x70)
124
125typedef enum {
126    ARITH_AND = 0x0 << 21,
127    ARITH_EOR = 0x1 << 21,
128    ARITH_SUB = 0x2 << 21,
129    ARITH_RSB = 0x3 << 21,
130    ARITH_ADD = 0x4 << 21,
131    ARITH_ADC = 0x5 << 21,
132    ARITH_SBC = 0x6 << 21,
133    ARITH_RSC = 0x7 << 21,
134    ARITH_TST = 0x8 << 21 | TO_CPSR,
135    ARITH_CMP = 0xa << 21 | TO_CPSR,
136    ARITH_CMN = 0xb << 21 | TO_CPSR,
137    ARITH_ORR = 0xc << 21,
138    ARITH_MOV = 0xd << 21,
139    ARITH_BIC = 0xe << 21,
140    ARITH_MVN = 0xf << 21,
141
142    INSN_B         = 0x0a000000,
143
144    INSN_CLZ       = 0x016f0f10,
145    INSN_RBIT      = 0x06ff0f30,
146
147    INSN_LDMIA     = 0x08b00000,
148    INSN_STMDB     = 0x09200000,
149
150    INSN_LDR_IMM   = 0x04100000,
151    INSN_LDR_REG   = 0x06100000,
152    INSN_STR_IMM   = 0x04000000,
153    INSN_STR_REG   = 0x06000000,
154
155    INSN_LDRH_IMM  = 0x005000b0,
156    INSN_LDRH_REG  = 0x001000b0,
157    INSN_LDRSH_IMM = 0x005000f0,
158    INSN_LDRSH_REG = 0x001000f0,
159    INSN_STRH_IMM  = 0x004000b0,
160    INSN_STRH_REG  = 0x000000b0,
161
162    INSN_LDRB_IMM  = 0x04500000,
163    INSN_LDRB_REG  = 0x06500000,
164    INSN_LDRSB_IMM = 0x005000d0,
165    INSN_LDRSB_REG = 0x001000d0,
166    INSN_STRB_IMM  = 0x04400000,
167    INSN_STRB_REG  = 0x06400000,
168
169    INSN_LDRD_IMM  = 0x004000d0,
170    INSN_LDRD_REG  = 0x000000d0,
171    INSN_STRD_IMM  = 0x004000f0,
172    INSN_STRD_REG  = 0x000000f0,
173
174    INSN_DMB_ISH   = 0xf57ff05b,
175    INSN_DMB_MCR   = 0xee070fba,
176
177    /* Architected nop introduced in v6k.  */
178    /* ??? This is an MSR (imm) 0,0,0 insn.  Anyone know if this
179       also Just So Happened to do nothing on pre-v6k so that we
180       don't need to conditionalize it?  */
181    INSN_NOP_v6k   = 0xe320f000,
182    /* Otherwise the assembler uses mov r0,r0 */
183    INSN_NOP_v4    = (COND_AL << 28) | ARITH_MOV,
184
185    INSN_VADD      = 0xf2000800,
186    INSN_VAND      = 0xf2000110,
187    INSN_VBIC      = 0xf2100110,
188    INSN_VEOR      = 0xf3000110,
189    INSN_VORN      = 0xf2300110,
190    INSN_VORR      = 0xf2200110,
191    INSN_VSUB      = 0xf3000800,
192    INSN_VMUL      = 0xf2000910,
193    INSN_VQADD     = 0xf2000010,
194    INSN_VQADD_U   = 0xf3000010,
195    INSN_VQSUB     = 0xf2000210,
196    INSN_VQSUB_U   = 0xf3000210,
197    INSN_VMAX      = 0xf2000600,
198    INSN_VMAX_U    = 0xf3000600,
199    INSN_VMIN      = 0xf2000610,
200    INSN_VMIN_U    = 0xf3000610,
201
202    INSN_VABS      = 0xf3b10300,
203    INSN_VMVN      = 0xf3b00580,
204    INSN_VNEG      = 0xf3b10380,
205
206    INSN_VCEQ0     = 0xf3b10100,
207    INSN_VCGT0     = 0xf3b10000,
208    INSN_VCGE0     = 0xf3b10080,
209    INSN_VCLE0     = 0xf3b10180,
210    INSN_VCLT0     = 0xf3b10200,
211
212    INSN_VCEQ      = 0xf3000810,
213    INSN_VCGE      = 0xf2000310,
214    INSN_VCGT      = 0xf2000300,
215    INSN_VCGE_U    = 0xf3000310,
216    INSN_VCGT_U    = 0xf3000300,
217
218    INSN_VSHLI     = 0xf2800510,  /* VSHL (immediate) */
219    INSN_VSARI     = 0xf2800010,  /* VSHR.S */
220    INSN_VSHRI     = 0xf3800010,  /* VSHR.U */
221    INSN_VSLI      = 0xf3800510,
222    INSN_VSHL_S    = 0xf2000400,  /* VSHL.S (register) */
223    INSN_VSHL_U    = 0xf3000400,  /* VSHL.U (register) */
224
225    INSN_VBSL      = 0xf3100110,
226    INSN_VBIT      = 0xf3200110,
227    INSN_VBIF      = 0xf3300110,
228
229    INSN_VTST      = 0xf2000810,
230
231    INSN_VDUP_G    = 0xee800b10,  /* VDUP (ARM core register) */
232    INSN_VDUP_S    = 0xf3b00c00,  /* VDUP (scalar) */
233    INSN_VLDR_D    = 0xed100b00,  /* VLDR.64 */
234    INSN_VLD1      = 0xf4200000,  /* VLD1 (multiple single elements) */
235    INSN_VLD1R     = 0xf4a00c00,  /* VLD1 (single element to all lanes) */
236    INSN_VST1      = 0xf4000000,  /* VST1 (multiple single elements) */
237    INSN_VMOVI     = 0xf2800010,  /* VMOV (immediate) */
238} ARMInsn;
239
240#define INSN_NOP   (use_armv7_instructions ? INSN_NOP_v6k : INSN_NOP_v4)
241
242static const uint8_t tcg_cond_to_arm_cond[] = {
243    [TCG_COND_EQ] = COND_EQ,
244    [TCG_COND_NE] = COND_NE,
245    [TCG_COND_LT] = COND_LT,
246    [TCG_COND_GE] = COND_GE,
247    [TCG_COND_LE] = COND_LE,
248    [TCG_COND_GT] = COND_GT,
249    /* unsigned */
250    [TCG_COND_LTU] = COND_CC,
251    [TCG_COND_GEU] = COND_CS,
252    [TCG_COND_LEU] = COND_LS,
253    [TCG_COND_GTU] = COND_HI,
254};
255
256static int encode_imm(uint32_t imm);
257
258/* TCG private relocation type: add with pc+imm8 */
259#define R_ARM_PC8  11
260
261/* TCG private relocation type: vldr with imm8 << 2 */
262#define R_ARM_PC11 12
263
264static bool reloc_pc24(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
265{
266    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
267    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) >> 2;
268
269    if (offset == sextract32(offset, 0, 24)) {
270        *src_rw = deposit32(*src_rw, 0, 24, offset);
271        return true;
272    }
273    return false;
274}
275
276static bool reloc_pc13(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
277{
278    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
279    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
280
281    if (offset >= -0xfff && offset <= 0xfff) {
282        tcg_insn_unit insn = *src_rw;
283        bool u = (offset >= 0);
284        if (!u) {
285            offset = -offset;
286        }
287        insn = deposit32(insn, 23, 1, u);
288        insn = deposit32(insn, 0, 12, offset);
289        *src_rw = insn;
290        return true;
291    }
292    return false;
293}
294
295static bool reloc_pc11(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
296{
297    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
298    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) / 4;
299
300    if (offset >= -0xff && offset <= 0xff) {
301        tcg_insn_unit insn = *src_rw;
302        bool u = (offset >= 0);
303        if (!u) {
304            offset = -offset;
305        }
306        insn = deposit32(insn, 23, 1, u);
307        insn = deposit32(insn, 0, 8, offset);
308        *src_rw = insn;
309        return true;
310    }
311    return false;
312}
313
314static bool reloc_pc8(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
315{
316    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
317    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
318    int imm12 = encode_imm(offset);
319
320    if (imm12 >= 0) {
321        *src_rw = deposit32(*src_rw, 0, 12, imm12);
322        return true;
323    }
324    return false;
325}
326
327static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
328                        intptr_t value, intptr_t addend)
329{
330    tcg_debug_assert(addend == 0);
331    switch (type) {
332    case R_ARM_PC24:
333        return reloc_pc24(code_ptr, (const tcg_insn_unit *)value);
334    case R_ARM_PC13:
335        return reloc_pc13(code_ptr, (const tcg_insn_unit *)value);
336    case R_ARM_PC11:
337        return reloc_pc11(code_ptr, (const tcg_insn_unit *)value);
338    case R_ARM_PC8:
339        return reloc_pc8(code_ptr, (const tcg_insn_unit *)value);
340    default:
341        g_assert_not_reached();
342    }
343}
344
345#define TCG_CT_CONST_ARM  0x100
346#define TCG_CT_CONST_INV  0x200
347#define TCG_CT_CONST_NEG  0x400
348#define TCG_CT_CONST_ZERO 0x800
349#define TCG_CT_CONST_ORRI 0x1000
350#define TCG_CT_CONST_ANDI 0x2000
351
352#define ALL_GENERAL_REGS  0xffffu
353#define ALL_VECTOR_REGS   0xffff0000u
354
355/*
356 * r0-r2 will be overwritten when reading the tlb entry (softmmu only)
357 * and r0-r1 doing the byte swapping, so don't use these.
358 * r3 is removed for softmmu to avoid clashes with helper arguments.
359 */
360#ifdef CONFIG_SOFTMMU
361#define ALL_QLOAD_REGS \
362    (ALL_GENERAL_REGS & ~((1 << TCG_REG_R0) | (1 << TCG_REG_R1) | \
363                          (1 << TCG_REG_R2) | (1 << TCG_REG_R3) | \
364                          (1 << TCG_REG_R14)))
365#define ALL_QSTORE_REGS \
366    (ALL_GENERAL_REGS & ~((1 << TCG_REG_R0) | (1 << TCG_REG_R1) | \
367                          (1 << TCG_REG_R2) | (1 << TCG_REG_R14) | \
368                          ((TARGET_LONG_BITS == 64) << TCG_REG_R3)))
369#else
370#define ALL_QLOAD_REGS   ALL_GENERAL_REGS
371#define ALL_QSTORE_REGS \
372    (ALL_GENERAL_REGS & ~((1 << TCG_REG_R0) | (1 << TCG_REG_R1)))
373#endif
374
375/*
376 * ARM immediates for ALU instructions are made of an unsigned 8-bit
377 * right-rotated by an even amount between 0 and 30.
378 *
379 * Return < 0 if @imm cannot be encoded, else the entire imm12 field.
380 */
381static int encode_imm(uint32_t imm)
382{
383    uint32_t rot, imm8;
384
385    /* Simple case, no rotation required. */
386    if ((imm & ~0xff) == 0) {
387        return imm;
388    }
389
390    /* Next, try a simple even shift.  */
391    rot = ctz32(imm) & ~1;
392    imm8 = imm >> rot;
393    rot = 32 - rot;
394    if ((imm8 & ~0xff) == 0) {
395        goto found;
396    }
397
398    /*
399     * Finally, try harder with rotations.
400     * The ctz test above will have taken care of rotates >= 8.
401     */
402    for (rot = 2; rot < 8; rot += 2) {
403        imm8 = rol32(imm, rot);
404        if ((imm8 & ~0xff) == 0) {
405            goto found;
406        }
407    }
408    /* Fail: imm cannot be encoded. */
409    return -1;
410
411 found:
412    /* Note that rot is even, and we discard bit 0 by shifting by 7. */
413    return rot << 7 | imm8;
414}
415
416static int encode_imm_nofail(uint32_t imm)
417{
418    int ret = encode_imm(imm);
419    tcg_debug_assert(ret >= 0);
420    return ret;
421}
422
423static bool check_fit_imm(uint32_t imm)
424{
425    return encode_imm(imm) >= 0;
426}
427
428/* Return true if v16 is a valid 16-bit shifted immediate.  */
429static bool is_shimm16(uint16_t v16, int *cmode, int *imm8)
430{
431    if (v16 == (v16 & 0xff)) {
432        *cmode = 0x8;
433        *imm8 = v16 & 0xff;
434        return true;
435    } else if (v16 == (v16 & 0xff00)) {
436        *cmode = 0xa;
437        *imm8 = v16 >> 8;
438        return true;
439    }
440    return false;
441}
442
443/* Return true if v32 is a valid 32-bit shifted immediate.  */
444static bool is_shimm32(uint32_t v32, int *cmode, int *imm8)
445{
446    if (v32 == (v32 & 0xff)) {
447        *cmode = 0x0;
448        *imm8 = v32 & 0xff;
449        return true;
450    } else if (v32 == (v32 & 0xff00)) {
451        *cmode = 0x2;
452        *imm8 = (v32 >> 8) & 0xff;
453        return true;
454    } else if (v32 == (v32 & 0xff0000)) {
455        *cmode = 0x4;
456        *imm8 = (v32 >> 16) & 0xff;
457        return true;
458    } else if (v32 == (v32 & 0xff000000)) {
459        *cmode = 0x6;
460        *imm8 = v32 >> 24;
461        return true;
462    }
463    return false;
464}
465
466/* Return true if v32 is a valid 32-bit shifting ones immediate.  */
467static bool is_soimm32(uint32_t v32, int *cmode, int *imm8)
468{
469    if ((v32 & 0xffff00ff) == 0xff) {
470        *cmode = 0xc;
471        *imm8 = (v32 >> 8) & 0xff;
472        return true;
473    } else if ((v32 & 0xff00ffff) == 0xffff) {
474        *cmode = 0xd;
475        *imm8 = (v32 >> 16) & 0xff;
476        return true;
477    }
478    return false;
479}
480
481/*
482 * Return non-zero if v32 can be formed by MOVI+ORR.
483 * Place the parameters for MOVI in (cmode, imm8).
484 * Return the cmode for ORR; the imm8 can be had via extraction from v32.
485 */
486static int is_shimm32_pair(uint32_t v32, int *cmode, int *imm8)
487{
488    int i;
489
490    for (i = 6; i > 0; i -= 2) {
491        /* Mask out one byte we can add with ORR.  */
492        uint32_t tmp = v32 & ~(0xffu << (i * 4));
493        if (is_shimm32(tmp, cmode, imm8) ||
494            is_soimm32(tmp, cmode, imm8)) {
495            break;
496        }
497    }
498    return i;
499}
500
501/* Return true if V is a valid 16-bit or 32-bit shifted immediate.  */
502static bool is_shimm1632(uint32_t v32, int *cmode, int *imm8)
503{
504    if (v32 == deposit32(v32, 16, 16, v32)) {
505        return is_shimm16(v32, cmode, imm8);
506    } else {
507        return is_shimm32(v32, cmode, imm8);
508    }
509}
510
511/* Test if a constant matches the constraint.
512 * TODO: define constraints for:
513 *
514 * ldr/str offset:   between -0xfff and 0xfff
515 * ldrh/strh offset: between -0xff and 0xff
516 * mov operand2:     values represented with x << (2 * y), x < 0x100
517 * add, sub, eor...: ditto
518 */
519static bool tcg_target_const_match(int64_t val, TCGType type, int ct)
520{
521    if (ct & TCG_CT_CONST) {
522        return 1;
523    } else if ((ct & TCG_CT_CONST_ARM) && check_fit_imm(val)) {
524        return 1;
525    } else if ((ct & TCG_CT_CONST_INV) && check_fit_imm(~val)) {
526        return 1;
527    } else if ((ct & TCG_CT_CONST_NEG) && check_fit_imm(-val)) {
528        return 1;
529    } else if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
530        return 1;
531    }
532
533    switch (ct & (TCG_CT_CONST_ORRI | TCG_CT_CONST_ANDI)) {
534    case 0:
535        break;
536    case TCG_CT_CONST_ANDI:
537        val = ~val;
538        /* fallthru */
539    case TCG_CT_CONST_ORRI:
540        if (val == deposit64(val, 32, 32, val)) {
541            int cmode, imm8;
542            return is_shimm1632(val, &cmode, &imm8);
543        }
544        break;
545    default:
546        /* Both bits should not be set for the same insn.  */
547        g_assert_not_reached();
548    }
549
550    return 0;
551}
552
553static void tcg_out_b_imm(TCGContext *s, ARMCond cond, int32_t offset)
554{
555    tcg_out32(s, (cond << 28) | INSN_B |
556                    (((offset - 8) >> 2) & 0x00ffffff));
557}
558
559static void tcg_out_bl_imm(TCGContext *s, ARMCond cond, int32_t offset)
560{
561    tcg_out32(s, (cond << 28) | 0x0b000000 |
562                    (((offset - 8) >> 2) & 0x00ffffff));
563}
564
565static void tcg_out_blx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
566{
567    tcg_out32(s, (cond << 28) | 0x012fff30 | rn);
568}
569
570static void tcg_out_blx_imm(TCGContext *s, int32_t offset)
571{
572    tcg_out32(s, 0xfa000000 | ((offset & 2) << 23) |
573                (((offset - 8) >> 2) & 0x00ffffff));
574}
575
576static void tcg_out_dat_reg(TCGContext *s, ARMCond cond, ARMInsn opc,
577                            TCGReg rd, TCGReg rn, TCGReg rm, int shift)
578{
579    tcg_out32(s, (cond << 28) | (0 << 25) | opc |
580                    (rn << 16) | (rd << 12) | shift | rm);
581}
582
583static void tcg_out_mov_reg(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rm)
584{
585    /* Simple reg-reg move, optimising out the 'do nothing' case */
586    if (rd != rm) {
587        tcg_out_dat_reg(s, cond, ARITH_MOV, rd, 0, rm, SHIFT_IMM_LSL(0));
588    }
589}
590
591static void tcg_out_bx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
592{
593    tcg_out32(s, (cond << 28) | 0x012fff10 | rn);
594}
595
596static void tcg_out_b_reg(TCGContext *s, ARMCond cond, TCGReg rn)
597{
598    /*
599     * Unless the C portion of QEMU is compiled as thumb, we don't need
600     * true BX semantics; merely a branch to an address held in a register.
601     */
602    tcg_out_bx_reg(s, cond, rn);
603}
604
605static void tcg_out_dat_imm(TCGContext *s, ARMCond cond, ARMInsn opc,
606                            TCGReg rd, TCGReg rn, int im)
607{
608    tcg_out32(s, (cond << 28) | (1 << 25) | opc |
609                    (rn << 16) | (rd << 12) | im);
610}
611
612static void tcg_out_ldstm(TCGContext *s, ARMCond cond, ARMInsn opc,
613                          TCGReg rn, uint16_t mask)
614{
615    tcg_out32(s, (cond << 28) | opc | (rn << 16) | mask);
616}
617
618/* Note that this routine is used for both LDR and LDRH formats, so we do
619   not wish to include an immediate shift at this point.  */
620static void tcg_out_memop_r(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
621                            TCGReg rn, TCGReg rm, bool u, bool p, bool w)
622{
623    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24)
624              | (w << 21) | (rn << 16) | (rt << 12) | rm);
625}
626
627static void tcg_out_memop_8(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
628                            TCGReg rn, int imm8, bool p, bool w)
629{
630    bool u = 1;
631    if (imm8 < 0) {
632        imm8 = -imm8;
633        u = 0;
634    }
635    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
636              (rn << 16) | (rt << 12) | ((imm8 & 0xf0) << 4) | (imm8 & 0xf));
637}
638
639static void tcg_out_memop_12(TCGContext *s, ARMCond cond, ARMInsn opc,
640                             TCGReg rt, TCGReg rn, int imm12, bool p, bool w)
641{
642    bool u = 1;
643    if (imm12 < 0) {
644        imm12 = -imm12;
645        u = 0;
646    }
647    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
648              (rn << 16) | (rt << 12) | imm12);
649}
650
651static void tcg_out_ld32_12(TCGContext *s, ARMCond cond, TCGReg rt,
652                            TCGReg rn, int imm12)
653{
654    tcg_out_memop_12(s, cond, INSN_LDR_IMM, rt, rn, imm12, 1, 0);
655}
656
657static void tcg_out_st32_12(TCGContext *s, ARMCond cond, TCGReg rt,
658                            TCGReg rn, int imm12)
659{
660    tcg_out_memop_12(s, cond, INSN_STR_IMM, rt, rn, imm12, 1, 0);
661}
662
663static void tcg_out_ld32_r(TCGContext *s, ARMCond cond, TCGReg rt,
664                           TCGReg rn, TCGReg rm)
665{
666    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 0);
667}
668
669static void tcg_out_st32_r(TCGContext *s, ARMCond cond, TCGReg rt,
670                           TCGReg rn, TCGReg rm)
671{
672    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 0);
673}
674
675static void tcg_out_ldrd_8(TCGContext *s, ARMCond cond, TCGReg rt,
676                           TCGReg rn, int imm8)
677{
678    tcg_out_memop_8(s, cond, INSN_LDRD_IMM, rt, rn, imm8, 1, 0);
679}
680
681static void tcg_out_ldrd_r(TCGContext *s, ARMCond cond, TCGReg rt,
682                           TCGReg rn, TCGReg rm)
683{
684    tcg_out_memop_r(s, cond, INSN_LDRD_REG, rt, rn, rm, 1, 1, 0);
685}
686
687static void __attribute__((unused))
688tcg_out_ldrd_rwb(TCGContext *s, ARMCond cond, TCGReg rt, TCGReg rn, TCGReg rm)
689{
690    tcg_out_memop_r(s, cond, INSN_LDRD_REG, rt, rn, rm, 1, 1, 1);
691}
692
693static void tcg_out_strd_8(TCGContext *s, ARMCond cond, TCGReg rt,
694                           TCGReg rn, int imm8)
695{
696    tcg_out_memop_8(s, cond, INSN_STRD_IMM, rt, rn, imm8, 1, 0);
697}
698
699static void tcg_out_strd_r(TCGContext *s, ARMCond cond, TCGReg rt,
700                           TCGReg rn, TCGReg rm)
701{
702    tcg_out_memop_r(s, cond, INSN_STRD_REG, rt, rn, rm, 1, 1, 0);
703}
704
705/* Register pre-increment with base writeback.  */
706static void tcg_out_ld32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
707                             TCGReg rn, TCGReg rm)
708{
709    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 1);
710}
711
712static void tcg_out_st32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
713                             TCGReg rn, TCGReg rm)
714{
715    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 1);
716}
717
718static void tcg_out_ld16u_8(TCGContext *s, ARMCond cond, TCGReg rt,
719                            TCGReg rn, int imm8)
720{
721    tcg_out_memop_8(s, cond, INSN_LDRH_IMM, rt, rn, imm8, 1, 0);
722}
723
724static void tcg_out_st16_8(TCGContext *s, ARMCond cond, TCGReg rt,
725                           TCGReg rn, int imm8)
726{
727    tcg_out_memop_8(s, cond, INSN_STRH_IMM, rt, rn, imm8, 1, 0);
728}
729
730static void tcg_out_ld16u_r(TCGContext *s, ARMCond cond, TCGReg rt,
731                            TCGReg rn, TCGReg rm)
732{
733    tcg_out_memop_r(s, cond, INSN_LDRH_REG, rt, rn, rm, 1, 1, 0);
734}
735
736static void tcg_out_st16_r(TCGContext *s, ARMCond cond, TCGReg rt,
737                           TCGReg rn, TCGReg rm)
738{
739    tcg_out_memop_r(s, cond, INSN_STRH_REG, rt, rn, rm, 1, 1, 0);
740}
741
742static void tcg_out_ld16s_8(TCGContext *s, ARMCond cond, TCGReg rt,
743                            TCGReg rn, int imm8)
744{
745    tcg_out_memop_8(s, cond, INSN_LDRSH_IMM, rt, rn, imm8, 1, 0);
746}
747
748static void tcg_out_ld16s_r(TCGContext *s, ARMCond cond, TCGReg rt,
749                            TCGReg rn, TCGReg rm)
750{
751    tcg_out_memop_r(s, cond, INSN_LDRSH_REG, rt, rn, rm, 1, 1, 0);
752}
753
754static void tcg_out_ld8_12(TCGContext *s, ARMCond cond, TCGReg rt,
755                           TCGReg rn, int imm12)
756{
757    tcg_out_memop_12(s, cond, INSN_LDRB_IMM, rt, rn, imm12, 1, 0);
758}
759
760static void tcg_out_st8_12(TCGContext *s, ARMCond cond, TCGReg rt,
761                           TCGReg rn, int imm12)
762{
763    tcg_out_memop_12(s, cond, INSN_STRB_IMM, rt, rn, imm12, 1, 0);
764}
765
766static void tcg_out_ld8_r(TCGContext *s, ARMCond cond, TCGReg rt,
767                          TCGReg rn, TCGReg rm)
768{
769    tcg_out_memop_r(s, cond, INSN_LDRB_REG, rt, rn, rm, 1, 1, 0);
770}
771
772static void tcg_out_st8_r(TCGContext *s, ARMCond cond, TCGReg rt,
773                          TCGReg rn, TCGReg rm)
774{
775    tcg_out_memop_r(s, cond, INSN_STRB_REG, rt, rn, rm, 1, 1, 0);
776}
777
778static void tcg_out_ld8s_8(TCGContext *s, ARMCond cond, TCGReg rt,
779                           TCGReg rn, int imm8)
780{
781    tcg_out_memop_8(s, cond, INSN_LDRSB_IMM, rt, rn, imm8, 1, 0);
782}
783
784static void tcg_out_ld8s_r(TCGContext *s, ARMCond cond, TCGReg rt,
785                           TCGReg rn, TCGReg rm)
786{
787    tcg_out_memop_r(s, cond, INSN_LDRSB_REG, rt, rn, rm, 1, 1, 0);
788}
789
790static void tcg_out_movi_pool(TCGContext *s, ARMCond cond,
791                              TCGReg rd, uint32_t arg)
792{
793    new_pool_label(s, arg, R_ARM_PC13, s->code_ptr, 0);
794    tcg_out_ld32_12(s, cond, rd, TCG_REG_PC, 0);
795}
796
797static void tcg_out_movi32(TCGContext *s, ARMCond cond,
798                           TCGReg rd, uint32_t arg)
799{
800    int imm12, diff, opc, sh1, sh2;
801    uint32_t tt0, tt1, tt2;
802
803    /* Check a single MOV/MVN before anything else.  */
804    imm12 = encode_imm(arg);
805    if (imm12 >= 0) {
806        tcg_out_dat_imm(s, cond, ARITH_MOV, rd, 0, imm12);
807        return;
808    }
809    imm12 = encode_imm(~arg);
810    if (imm12 >= 0) {
811        tcg_out_dat_imm(s, cond, ARITH_MVN, rd, 0, imm12);
812        return;
813    }
814
815    /* Check for a pc-relative address.  This will usually be the TB,
816       or within the TB, which is immediately before the code block.  */
817    diff = tcg_pcrel_diff(s, (void *)arg) - 8;
818    if (diff >= 0) {
819        imm12 = encode_imm(diff);
820        if (imm12 >= 0) {
821            tcg_out_dat_imm(s, cond, ARITH_ADD, rd, TCG_REG_PC, imm12);
822            return;
823        }
824    } else {
825        imm12 = encode_imm(-diff);
826        if (imm12 >= 0) {
827            tcg_out_dat_imm(s, cond, ARITH_SUB, rd, TCG_REG_PC, imm12);
828            return;
829        }
830    }
831
832    /* Use movw + movt.  */
833    if (use_armv7_instructions) {
834        /* movw */
835        tcg_out32(s, (cond << 28) | 0x03000000 | (rd << 12)
836                  | ((arg << 4) & 0x000f0000) | (arg & 0xfff));
837        if (arg & 0xffff0000) {
838            /* movt */
839            tcg_out32(s, (cond << 28) | 0x03400000 | (rd << 12)
840                      | ((arg >> 12) & 0x000f0000) | ((arg >> 16) & 0xfff));
841        }
842        return;
843    }
844
845    /* Look for sequences of two insns.  If we have lots of 1's, we can
846       shorten the sequence by beginning with mvn and then clearing
847       higher bits with eor.  */
848    tt0 = arg;
849    opc = ARITH_MOV;
850    if (ctpop32(arg) > 16) {
851        tt0 = ~arg;
852        opc = ARITH_MVN;
853    }
854    sh1 = ctz32(tt0) & ~1;
855    tt1 = tt0 & ~(0xff << sh1);
856    sh2 = ctz32(tt1) & ~1;
857    tt2 = tt1 & ~(0xff << sh2);
858    if (tt2 == 0) {
859        int rot;
860
861        rot = ((32 - sh1) << 7) & 0xf00;
862        tcg_out_dat_imm(s, cond, opc, rd,  0, ((tt0 >> sh1) & 0xff) | rot);
863        rot = ((32 - sh2) << 7) & 0xf00;
864        tcg_out_dat_imm(s, cond, ARITH_EOR, rd, rd,
865                        ((tt0 >> sh2) & 0xff) | rot);
866        return;
867    }
868
869    /* Otherwise, drop it into the constant pool.  */
870    tcg_out_movi_pool(s, cond, rd, arg);
871}
872
873/*
874 * Emit either the reg,imm or reg,reg form of a data-processing insn.
875 * rhs must satisfy the "rI" constraint.
876 */
877static void tcg_out_dat_rI(TCGContext *s, ARMCond cond, ARMInsn opc,
878                           TCGReg dst, TCGReg lhs, TCGArg rhs, int rhs_is_const)
879{
880    if (rhs_is_const) {
881        tcg_out_dat_imm(s, cond, opc, dst, lhs, encode_imm_nofail(rhs));
882    } else {
883        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
884    }
885}
886
887/*
888 * Emit either the reg,imm or reg,reg form of a data-processing insn.
889 * rhs must satisfy the "rIK" constraint.
890 */
891static void tcg_out_dat_rIK(TCGContext *s, ARMCond cond, ARMInsn opc,
892                            ARMInsn opinv, TCGReg dst, TCGReg lhs, TCGArg rhs,
893                            bool rhs_is_const)
894{
895    if (rhs_is_const) {
896        int imm12 = encode_imm(rhs);
897        if (imm12 < 0) {
898            imm12 = encode_imm_nofail(~rhs);
899            opc = opinv;
900        }
901        tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
902    } else {
903        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
904    }
905}
906
907static void tcg_out_dat_rIN(TCGContext *s, ARMCond cond, ARMInsn opc,
908                            ARMInsn opneg, TCGReg dst, TCGReg lhs, TCGArg rhs,
909                            bool rhs_is_const)
910{
911    /* Emit either the reg,imm or reg,reg form of a data-processing insn.
912     * rhs must satisfy the "rIN" constraint.
913     */
914    if (rhs_is_const) {
915        int imm12 = encode_imm(rhs);
916        if (imm12 < 0) {
917            imm12 = encode_imm_nofail(-rhs);
918            opc = opneg;
919        }
920        tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
921    } else {
922        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
923    }
924}
925
926static void tcg_out_mul32(TCGContext *s, ARMCond cond, TCGReg rd,
927                          TCGReg rn, TCGReg rm)
928{
929    /* mul */
930    tcg_out32(s, (cond << 28) | 0x90 | (rd << 16) | (rm << 8) | rn);
931}
932
933static void tcg_out_umull32(TCGContext *s, ARMCond cond, TCGReg rd0,
934                            TCGReg rd1, TCGReg rn, TCGReg rm)
935{
936    /* umull */
937    tcg_out32(s, (cond << 28) | 0x00800090 |
938              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
939}
940
941static void tcg_out_smull32(TCGContext *s, ARMCond cond, TCGReg rd0,
942                            TCGReg rd1, TCGReg rn, TCGReg rm)
943{
944    /* smull */
945    tcg_out32(s, (cond << 28) | 0x00c00090 |
946              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
947}
948
949static void tcg_out_sdiv(TCGContext *s, ARMCond cond,
950                         TCGReg rd, TCGReg rn, TCGReg rm)
951{
952    tcg_out32(s, 0x0710f010 | (cond << 28) | (rd << 16) | rn | (rm << 8));
953}
954
955static void tcg_out_udiv(TCGContext *s, ARMCond cond,
956                         TCGReg rd, TCGReg rn, TCGReg rm)
957{
958    tcg_out32(s, 0x0730f010 | (cond << 28) | (rd << 16) | rn | (rm << 8));
959}
960
961static void tcg_out_ext8s(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rn)
962{
963    /* sxtb */
964    tcg_out32(s, 0x06af0070 | (cond << 28) | (rd << 12) | rn);
965}
966
967static void __attribute__((unused))
968tcg_out_ext8u(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rn)
969{
970    tcg_out_dat_imm(s, cond, ARITH_AND, rd, rn, 0xff);
971}
972
973static void tcg_out_ext16s(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rn)
974{
975    /* sxth */
976    tcg_out32(s, 0x06bf0070 | (cond << 28) | (rd << 12) | rn);
977}
978
979static void tcg_out_ext16u(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rn)
980{
981    /* uxth */
982    tcg_out32(s, 0x06ff0070 | (cond << 28) | (rd << 12) | rn);
983}
984
985static void tcg_out_bswap16(TCGContext *s, ARMCond cond,
986                            TCGReg rd, TCGReg rn, int flags)
987{
988    if (flags & TCG_BSWAP_OS) {
989        /* revsh */
990        tcg_out32(s, 0x06ff0fb0 | (cond << 28) | (rd << 12) | rn);
991        return;
992    }
993
994    /* rev16 */
995    tcg_out32(s, 0x06bf0fb0 | (cond << 28) | (rd << 12) | rn);
996    if ((flags & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
997        /* uxth */
998        tcg_out32(s, 0x06ff0070 | (cond << 28) | (rd << 12) | rd);
999    }
1000}
1001
1002static void tcg_out_bswap32(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rn)
1003{
1004    /* rev */
1005    tcg_out32(s, 0x06bf0f30 | (cond << 28) | (rd << 12) | rn);
1006}
1007
1008static void tcg_out_deposit(TCGContext *s, ARMCond cond, TCGReg rd,
1009                            TCGArg a1, int ofs, int len, bool const_a1)
1010{
1011    if (const_a1) {
1012        /* bfi becomes bfc with rn == 15.  */
1013        a1 = 15;
1014    }
1015    /* bfi/bfc */
1016    tcg_out32(s, 0x07c00010 | (cond << 28) | (rd << 12) | a1
1017              | (ofs << 7) | ((ofs + len - 1) << 16));
1018}
1019
1020static void tcg_out_extract(TCGContext *s, ARMCond cond, TCGReg rd,
1021                            TCGReg rn, int ofs, int len)
1022{
1023    /* ubfx */
1024    tcg_out32(s, 0x07e00050 | (cond << 28) | (rd << 12) | rn
1025              | (ofs << 7) | ((len - 1) << 16));
1026}
1027
1028static void tcg_out_sextract(TCGContext *s, ARMCond cond, TCGReg rd,
1029                             TCGReg rn, int ofs, int len)
1030{
1031    /* sbfx */
1032    tcg_out32(s, 0x07a00050 | (cond << 28) | (rd << 12) | rn
1033              | (ofs << 7) | ((len - 1) << 16));
1034}
1035
1036static void tcg_out_ld32u(TCGContext *s, ARMCond cond,
1037                          TCGReg rd, TCGReg rn, int32_t offset)
1038{
1039    if (offset > 0xfff || offset < -0xfff) {
1040        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1041        tcg_out_ld32_r(s, cond, rd, rn, TCG_REG_TMP);
1042    } else
1043        tcg_out_ld32_12(s, cond, rd, rn, offset);
1044}
1045
1046static void tcg_out_st32(TCGContext *s, ARMCond cond,
1047                         TCGReg rd, TCGReg rn, int32_t offset)
1048{
1049    if (offset > 0xfff || offset < -0xfff) {
1050        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1051        tcg_out_st32_r(s, cond, rd, rn, TCG_REG_TMP);
1052    } else
1053        tcg_out_st32_12(s, cond, rd, rn, offset);
1054}
1055
1056static void tcg_out_ld16u(TCGContext *s, ARMCond cond,
1057                          TCGReg rd, TCGReg rn, int32_t offset)
1058{
1059    if (offset > 0xff || offset < -0xff) {
1060        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1061        tcg_out_ld16u_r(s, cond, rd, rn, TCG_REG_TMP);
1062    } else
1063        tcg_out_ld16u_8(s, cond, rd, rn, offset);
1064}
1065
1066static void tcg_out_ld16s(TCGContext *s, ARMCond cond,
1067                          TCGReg rd, TCGReg rn, int32_t offset)
1068{
1069    if (offset > 0xff || offset < -0xff) {
1070        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1071        tcg_out_ld16s_r(s, cond, rd, rn, TCG_REG_TMP);
1072    } else
1073        tcg_out_ld16s_8(s, cond, rd, rn, offset);
1074}
1075
1076static void tcg_out_st16(TCGContext *s, ARMCond cond,
1077                         TCGReg rd, TCGReg rn, int32_t offset)
1078{
1079    if (offset > 0xff || offset < -0xff) {
1080        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1081        tcg_out_st16_r(s, cond, rd, rn, TCG_REG_TMP);
1082    } else
1083        tcg_out_st16_8(s, cond, rd, rn, offset);
1084}
1085
1086static void tcg_out_ld8u(TCGContext *s, ARMCond cond,
1087                         TCGReg rd, TCGReg rn, int32_t offset)
1088{
1089    if (offset > 0xfff || offset < -0xfff) {
1090        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1091        tcg_out_ld8_r(s, cond, rd, rn, TCG_REG_TMP);
1092    } else
1093        tcg_out_ld8_12(s, cond, rd, rn, offset);
1094}
1095
1096static void tcg_out_ld8s(TCGContext *s, ARMCond cond,
1097                         TCGReg rd, TCGReg rn, int32_t offset)
1098{
1099    if (offset > 0xff || offset < -0xff) {
1100        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1101        tcg_out_ld8s_r(s, cond, rd, rn, TCG_REG_TMP);
1102    } else
1103        tcg_out_ld8s_8(s, cond, rd, rn, offset);
1104}
1105
1106static void tcg_out_st8(TCGContext *s, ARMCond cond,
1107                        TCGReg rd, TCGReg rn, int32_t offset)
1108{
1109    if (offset > 0xfff || offset < -0xfff) {
1110        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1111        tcg_out_st8_r(s, cond, rd, rn, TCG_REG_TMP);
1112    } else
1113        tcg_out_st8_12(s, cond, rd, rn, offset);
1114}
1115
1116/*
1117 * The _goto case is normally between TBs within the same code buffer, and
1118 * with the code buffer limited to 16MB we wouldn't need the long case.
1119 * But we also use it for the tail-call to the qemu_ld/st helpers, which does.
1120 */
1121static void tcg_out_goto(TCGContext *s, ARMCond cond, const tcg_insn_unit *addr)
1122{
1123    intptr_t addri = (intptr_t)addr;
1124    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1125    bool arm_mode = !(addri & 1);
1126
1127    if (arm_mode && disp - 8 < 0x01fffffd && disp - 8 > -0x01fffffd) {
1128        tcg_out_b_imm(s, cond, disp);
1129        return;
1130    }
1131
1132    /* LDR is interworking from v5t. */
1133    tcg_out_movi_pool(s, cond, TCG_REG_PC, addri);
1134}
1135
1136/*
1137 * The call case is mostly used for helpers - so it's not unreasonable
1138 * for them to be beyond branch range.
1139 */
1140static void tcg_out_call_int(TCGContext *s, const tcg_insn_unit *addr)
1141{
1142    intptr_t addri = (intptr_t)addr;
1143    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1144    bool arm_mode = !(addri & 1);
1145
1146    if (disp - 8 < 0x02000000 && disp - 8 >= -0x02000000) {
1147        if (arm_mode) {
1148            tcg_out_bl_imm(s, COND_AL, disp);
1149        } else {
1150            tcg_out_blx_imm(s, disp);
1151        }
1152        return;
1153    }
1154
1155    tcg_out_movi32(s, COND_AL, TCG_REG_TMP, addri);
1156    tcg_out_blx_reg(s, COND_AL, TCG_REG_TMP);
1157}
1158
1159static void tcg_out_call(TCGContext *s, const tcg_insn_unit *addr,
1160                         const TCGHelperInfo *info)
1161{
1162    tcg_out_call_int(s, addr);
1163}
1164
1165static void tcg_out_goto_label(TCGContext *s, ARMCond cond, TCGLabel *l)
1166{
1167    if (l->has_value) {
1168        tcg_out_goto(s, cond, l->u.value_ptr);
1169    } else {
1170        tcg_out_reloc(s, s->code_ptr, R_ARM_PC24, l, 0);
1171        tcg_out_b_imm(s, cond, 0);
1172    }
1173}
1174
1175static void tcg_out_mb(TCGContext *s, TCGArg a0)
1176{
1177    if (use_armv7_instructions) {
1178        tcg_out32(s, INSN_DMB_ISH);
1179    } else {
1180        tcg_out32(s, INSN_DMB_MCR);
1181    }
1182}
1183
1184static TCGCond tcg_out_cmp2(TCGContext *s, const TCGArg *args,
1185                            const int *const_args)
1186{
1187    TCGReg al = args[0];
1188    TCGReg ah = args[1];
1189    TCGArg bl = args[2];
1190    TCGArg bh = args[3];
1191    TCGCond cond = args[4];
1192    int const_bl = const_args[2];
1193    int const_bh = const_args[3];
1194
1195    switch (cond) {
1196    case TCG_COND_EQ:
1197    case TCG_COND_NE:
1198    case TCG_COND_LTU:
1199    case TCG_COND_LEU:
1200    case TCG_COND_GTU:
1201    case TCG_COND_GEU:
1202        /* We perform a conditional comparision.  If the high half is
1203           equal, then overwrite the flags with the comparison of the
1204           low half.  The resulting flags cover the whole.  */
1205        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, ah, bh, const_bh);
1206        tcg_out_dat_rI(s, COND_EQ, ARITH_CMP, 0, al, bl, const_bl);
1207        return cond;
1208
1209    case TCG_COND_LT:
1210    case TCG_COND_GE:
1211        /* We perform a double-word subtraction and examine the result.
1212           We do not actually need the result of the subtract, so the
1213           low part "subtract" is a compare.  For the high half we have
1214           no choice but to compute into a temporary.  */
1215        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, al, bl, const_bl);
1216        tcg_out_dat_rI(s, COND_AL, ARITH_SBC | TO_CPSR,
1217                       TCG_REG_TMP, ah, bh, const_bh);
1218        return cond;
1219
1220    case TCG_COND_LE:
1221    case TCG_COND_GT:
1222        /* Similar, but with swapped arguments, via reversed subtract.  */
1223        tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR,
1224                       TCG_REG_TMP, al, bl, const_bl);
1225        tcg_out_dat_rI(s, COND_AL, ARITH_RSC | TO_CPSR,
1226                       TCG_REG_TMP, ah, bh, const_bh);
1227        return tcg_swap_cond(cond);
1228
1229    default:
1230        g_assert_not_reached();
1231    }
1232}
1233
1234/*
1235 * Note that TCGReg references Q-registers.
1236 * Q-regno = 2 * D-regno, so shift left by 1 whlie inserting.
1237 */
1238static uint32_t encode_vd(TCGReg rd)
1239{
1240    tcg_debug_assert(rd >= TCG_REG_Q0);
1241    return (extract32(rd, 3, 1) << 22) | (extract32(rd, 0, 3) << 13);
1242}
1243
1244static uint32_t encode_vn(TCGReg rn)
1245{
1246    tcg_debug_assert(rn >= TCG_REG_Q0);
1247    return (extract32(rn, 3, 1) << 7) | (extract32(rn, 0, 3) << 17);
1248}
1249
1250static uint32_t encode_vm(TCGReg rm)
1251{
1252    tcg_debug_assert(rm >= TCG_REG_Q0);
1253    return (extract32(rm, 3, 1) << 5) | (extract32(rm, 0, 3) << 1);
1254}
1255
1256static void tcg_out_vreg2(TCGContext *s, ARMInsn insn, int q, int vece,
1257                          TCGReg d, TCGReg m)
1258{
1259    tcg_out32(s, insn | (vece << 18) | (q << 6) |
1260              encode_vd(d) | encode_vm(m));
1261}
1262
1263static void tcg_out_vreg3(TCGContext *s, ARMInsn insn, int q, int vece,
1264                          TCGReg d, TCGReg n, TCGReg m)
1265{
1266    tcg_out32(s, insn | (vece << 20) | (q << 6) |
1267              encode_vd(d) | encode_vn(n) | encode_vm(m));
1268}
1269
1270static void tcg_out_vmovi(TCGContext *s, TCGReg rd,
1271                          int q, int op, int cmode, uint8_t imm8)
1272{
1273    tcg_out32(s, INSN_VMOVI | encode_vd(rd) | (q << 6) | (op << 5)
1274              | (cmode << 8) | extract32(imm8, 0, 4)
1275              | (extract32(imm8, 4, 3) << 16)
1276              | (extract32(imm8, 7, 1) << 24));
1277}
1278
1279static void tcg_out_vshifti(TCGContext *s, ARMInsn insn, int q,
1280                            TCGReg rd, TCGReg rm, int l_imm6)
1281{
1282    tcg_out32(s, insn | (q << 6) | encode_vd(rd) | encode_vm(rm) |
1283              (extract32(l_imm6, 6, 1) << 7) |
1284              (extract32(l_imm6, 0, 6) << 16));
1285}
1286
1287static void tcg_out_vldst(TCGContext *s, ARMInsn insn,
1288                          TCGReg rd, TCGReg rn, int offset)
1289{
1290    if (offset != 0) {
1291        if (check_fit_imm(offset) || check_fit_imm(-offset)) {
1292            tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
1293                            TCG_REG_TMP, rn, offset, true);
1294        } else {
1295            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP, offset);
1296            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1297                            TCG_REG_TMP, TCG_REG_TMP, rn, 0);
1298        }
1299        rn = TCG_REG_TMP;
1300    }
1301    tcg_out32(s, insn | (rn << 16) | encode_vd(rd) | 0xf);
1302}
1303
1304#ifdef CONFIG_SOFTMMU
1305/* helper signature: helper_ret_ld_mmu(CPUState *env, target_ulong addr,
1306 *                                     int mmu_idx, uintptr_t ra)
1307 */
1308static void * const qemu_ld_helpers[MO_SSIZE + 1] = {
1309    [MO_UB]   = helper_ret_ldub_mmu,
1310    [MO_SB]   = helper_ret_ldsb_mmu,
1311#if HOST_BIG_ENDIAN
1312    [MO_UW] = helper_be_lduw_mmu,
1313    [MO_UL] = helper_be_ldul_mmu,
1314    [MO_UQ] = helper_be_ldq_mmu,
1315    [MO_SW] = helper_be_ldsw_mmu,
1316    [MO_SL] = helper_be_ldul_mmu,
1317#else
1318    [MO_UW] = helper_le_lduw_mmu,
1319    [MO_UL] = helper_le_ldul_mmu,
1320    [MO_UQ] = helper_le_ldq_mmu,
1321    [MO_SW] = helper_le_ldsw_mmu,
1322    [MO_SL] = helper_le_ldul_mmu,
1323#endif
1324};
1325
1326/* helper signature: helper_ret_st_mmu(CPUState *env, target_ulong addr,
1327 *                                     uintxx_t val, int mmu_idx, uintptr_t ra)
1328 */
1329static void * const qemu_st_helpers[MO_SIZE + 1] = {
1330    [MO_8]   = helper_ret_stb_mmu,
1331#if HOST_BIG_ENDIAN
1332    [MO_16] = helper_be_stw_mmu,
1333    [MO_32] = helper_be_stl_mmu,
1334    [MO_64] = helper_be_stq_mmu,
1335#else
1336    [MO_16] = helper_le_stw_mmu,
1337    [MO_32] = helper_le_stl_mmu,
1338    [MO_64] = helper_le_stq_mmu,
1339#endif
1340};
1341
1342/* Helper routines for marshalling helper function arguments into
1343 * the correct registers and stack.
1344 * argreg is where we want to put this argument, arg is the argument itself.
1345 * Return value is the updated argreg ready for the next call.
1346 * Note that argreg 0..3 is real registers, 4+ on stack.
1347 *
1348 * We provide routines for arguments which are: immediate, 32 bit
1349 * value in register, 16 and 8 bit values in register (which must be zero
1350 * extended before use) and 64 bit value in a lo:hi register pair.
1351 */
1352#define DEFINE_TCG_OUT_ARG(NAME, ARGTYPE, MOV_ARG, EXT_ARG)                \
1353static TCGReg NAME(TCGContext *s, TCGReg argreg, ARGTYPE arg)              \
1354{                                                                          \
1355    if (argreg < 4) {                                                      \
1356        MOV_ARG(s, COND_AL, argreg, arg);                                  \
1357    } else {                                                               \
1358        int ofs = (argreg - 4) * 4;                                        \
1359        EXT_ARG;                                                           \
1360        tcg_debug_assert(ofs + 4 <= TCG_STATIC_CALL_ARGS_SIZE);            \
1361        tcg_out_st32_12(s, COND_AL, arg, TCG_REG_CALL_STACK, ofs);         \
1362    }                                                                      \
1363    return argreg + 1;                                                     \
1364}
1365
1366DEFINE_TCG_OUT_ARG(tcg_out_arg_imm32, uint32_t, tcg_out_movi32,
1367    (tcg_out_movi32(s, COND_AL, TCG_REG_TMP, arg), arg = TCG_REG_TMP))
1368DEFINE_TCG_OUT_ARG(tcg_out_arg_reg8, TCGReg, tcg_out_ext8u,
1369    (tcg_out_ext8u(s, COND_AL, TCG_REG_TMP, arg), arg = TCG_REG_TMP))
1370DEFINE_TCG_OUT_ARG(tcg_out_arg_reg16, TCGReg, tcg_out_ext16u,
1371    (tcg_out_ext16u(s, COND_AL, TCG_REG_TMP, arg), arg = TCG_REG_TMP))
1372DEFINE_TCG_OUT_ARG(tcg_out_arg_reg32, TCGReg, tcg_out_mov_reg, )
1373
1374static TCGReg tcg_out_arg_reg64(TCGContext *s, TCGReg argreg,
1375                                TCGReg arglo, TCGReg arghi)
1376{
1377    /* 64 bit arguments must go in even/odd register pairs
1378     * and in 8-aligned stack slots.
1379     */
1380    if (argreg & 1) {
1381        argreg++;
1382    }
1383    if (argreg >= 4 && (arglo & 1) == 0 && arghi == arglo + 1) {
1384        tcg_out_strd_8(s, COND_AL, arglo,
1385                       TCG_REG_CALL_STACK, (argreg - 4) * 4);
1386        return argreg + 2;
1387    } else {
1388        argreg = tcg_out_arg_reg32(s, argreg, arglo);
1389        argreg = tcg_out_arg_reg32(s, argreg, arghi);
1390        return argreg;
1391    }
1392}
1393
1394#define TLB_SHIFT	(CPU_TLB_ENTRY_BITS + CPU_TLB_BITS)
1395
1396/* We expect to use an 9-bit sign-magnitude negative offset from ENV.  */
1397QEMU_BUILD_BUG_ON(TLB_MASK_TABLE_OFS(0) > 0);
1398QEMU_BUILD_BUG_ON(TLB_MASK_TABLE_OFS(0) < -256);
1399
1400/* These offsets are built into the LDRD below.  */
1401QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, mask) != 0);
1402QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, table) != 4);
1403
1404/* Load and compare a TLB entry, leaving the flags set.  Returns the register
1405   containing the addend of the tlb entry.  Clobbers R0, R1, R2, TMP.  */
1406
1407static TCGReg tcg_out_tlb_read(TCGContext *s, TCGReg addrlo, TCGReg addrhi,
1408                               MemOp opc, int mem_index, bool is_load)
1409{
1410    int cmp_off = (is_load ? offsetof(CPUTLBEntry, addr_read)
1411                   : offsetof(CPUTLBEntry, addr_write));
1412    int fast_off = TLB_MASK_TABLE_OFS(mem_index);
1413    unsigned s_mask = (1 << (opc & MO_SIZE)) - 1;
1414    unsigned a_mask = (1 << get_alignment_bits(opc)) - 1;
1415    TCGReg t_addr;
1416
1417    /* Load env_tlb(env)->f[mmu_idx].{mask,table} into {r0,r1}.  */
1418    tcg_out_ldrd_8(s, COND_AL, TCG_REG_R0, TCG_AREG0, fast_off);
1419
1420    /* Extract the tlb index from the address into R0.  */
1421    tcg_out_dat_reg(s, COND_AL, ARITH_AND, TCG_REG_R0, TCG_REG_R0, addrlo,
1422                    SHIFT_IMM_LSR(TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS));
1423
1424    /*
1425     * Add the tlb_table pointer, creating the CPUTLBEntry address in R1.
1426     * Load the tlb comparator into R2/R3 and the fast path addend into R1.
1427     */
1428    if (cmp_off == 0) {
1429        if (TARGET_LONG_BITS == 64) {
1430            tcg_out_ldrd_rwb(s, COND_AL, TCG_REG_R2, TCG_REG_R1, TCG_REG_R0);
1431        } else {
1432            tcg_out_ld32_rwb(s, COND_AL, TCG_REG_R2, TCG_REG_R1, TCG_REG_R0);
1433        }
1434    } else {
1435        tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1436                        TCG_REG_R1, TCG_REG_R1, TCG_REG_R0, 0);
1437        if (TARGET_LONG_BITS == 64) {
1438            tcg_out_ldrd_8(s, COND_AL, TCG_REG_R2, TCG_REG_R1, cmp_off);
1439        } else {
1440            tcg_out_ld32_12(s, COND_AL, TCG_REG_R2, TCG_REG_R1, cmp_off);
1441        }
1442    }
1443
1444    /* Load the tlb addend.  */
1445    tcg_out_ld32_12(s, COND_AL, TCG_REG_R1, TCG_REG_R1,
1446                    offsetof(CPUTLBEntry, addend));
1447
1448    /*
1449     * Check alignment, check comparators.
1450     * Do this in 2-4 insns.  Use MOVW for v7, if possible,
1451     * to reduce the number of sequential conditional instructions.
1452     * Almost all guests have at least 4k pages, which means that we need
1453     * to clear at least 9 bits even for an 8-byte memory, which means it
1454     * isn't worth checking for an immediate operand for BIC.
1455     *
1456     * For unaligned accesses, test the page of the last unit of alignment.
1457     * This leaves the least significant alignment bits unchanged, and of
1458     * course must be zero.
1459     */
1460    t_addr = addrlo;
1461    if (a_mask < s_mask) {
1462        t_addr = TCG_REG_R0;
1463        tcg_out_dat_imm(s, COND_AL, ARITH_ADD, t_addr,
1464                        addrlo, s_mask - a_mask);
1465    }
1466    if (use_armv7_instructions && TARGET_PAGE_BITS <= 16) {
1467        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, ~(TARGET_PAGE_MASK | a_mask));
1468        tcg_out_dat_reg(s, COND_AL, ARITH_BIC, TCG_REG_TMP,
1469                        t_addr, TCG_REG_TMP, 0);
1470        tcg_out_dat_reg(s, COND_AL, ARITH_CMP, 0, TCG_REG_R2, TCG_REG_TMP, 0);
1471    } else {
1472        if (a_mask) {
1473            tcg_debug_assert(a_mask <= 0xff);
1474            tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addrlo, a_mask);
1475        }
1476        tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0, t_addr,
1477                        SHIFT_IMM_LSR(TARGET_PAGE_BITS));
1478        tcg_out_dat_reg(s, (a_mask ? COND_EQ : COND_AL), ARITH_CMP,
1479                        0, TCG_REG_R2, TCG_REG_TMP,
1480                        SHIFT_IMM_LSL(TARGET_PAGE_BITS));
1481    }
1482
1483    if (TARGET_LONG_BITS == 64) {
1484        tcg_out_dat_reg(s, COND_EQ, ARITH_CMP, 0, TCG_REG_R3, addrhi, 0);
1485    }
1486
1487    return TCG_REG_R1;
1488}
1489
1490/* Record the context of a call to the out of line helper code for the slow
1491   path for a load or store, so that we can later generate the correct
1492   helper code.  */
1493static void add_qemu_ldst_label(TCGContext *s, bool is_ld, MemOpIdx oi,
1494                                TCGReg datalo, TCGReg datahi, TCGReg addrlo,
1495                                TCGReg addrhi, tcg_insn_unit *raddr,
1496                                tcg_insn_unit *label_ptr)
1497{
1498    TCGLabelQemuLdst *label = new_ldst_label(s);
1499
1500    label->is_ld = is_ld;
1501    label->oi = oi;
1502    label->datalo_reg = datalo;
1503    label->datahi_reg = datahi;
1504    label->addrlo_reg = addrlo;
1505    label->addrhi_reg = addrhi;
1506    label->raddr = tcg_splitwx_to_rx(raddr);
1507    label->label_ptr[0] = label_ptr;
1508}
1509
1510static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1511{
1512    TCGReg argreg, datalo, datahi;
1513    MemOpIdx oi = lb->oi;
1514    MemOp opc = get_memop(oi);
1515
1516    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1517        return false;
1518    }
1519
1520    argreg = tcg_out_arg_reg32(s, TCG_REG_R0, TCG_AREG0);
1521    if (TARGET_LONG_BITS == 64) {
1522        argreg = tcg_out_arg_reg64(s, argreg, lb->addrlo_reg, lb->addrhi_reg);
1523    } else {
1524        argreg = tcg_out_arg_reg32(s, argreg, lb->addrlo_reg);
1525    }
1526    argreg = tcg_out_arg_imm32(s, argreg, oi);
1527    argreg = tcg_out_arg_reg32(s, argreg, TCG_REG_R14);
1528
1529    /* Use the canonical unsigned helpers and minimize icache usage. */
1530    tcg_out_call_int(s, qemu_ld_helpers[opc & MO_SIZE]);
1531
1532    datalo = lb->datalo_reg;
1533    datahi = lb->datahi_reg;
1534    switch (opc & MO_SSIZE) {
1535    case MO_SB:
1536        tcg_out_ext8s(s, COND_AL, datalo, TCG_REG_R0);
1537        break;
1538    case MO_SW:
1539        tcg_out_ext16s(s, COND_AL, datalo, TCG_REG_R0);
1540        break;
1541    default:
1542        tcg_out_mov_reg(s, COND_AL, datalo, TCG_REG_R0);
1543        break;
1544    case MO_UQ:
1545        if (datalo != TCG_REG_R1) {
1546            tcg_out_mov_reg(s, COND_AL, datalo, TCG_REG_R0);
1547            tcg_out_mov_reg(s, COND_AL, datahi, TCG_REG_R1);
1548        } else if (datahi != TCG_REG_R0) {
1549            tcg_out_mov_reg(s, COND_AL, datahi, TCG_REG_R1);
1550            tcg_out_mov_reg(s, COND_AL, datalo, TCG_REG_R0);
1551        } else {
1552            tcg_out_mov_reg(s, COND_AL, TCG_REG_TMP, TCG_REG_R0);
1553            tcg_out_mov_reg(s, COND_AL, datahi, TCG_REG_R1);
1554            tcg_out_mov_reg(s, COND_AL, datalo, TCG_REG_TMP);
1555        }
1556        break;
1557    }
1558
1559    tcg_out_goto(s, COND_AL, lb->raddr);
1560    return true;
1561}
1562
1563static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1564{
1565    TCGReg argreg, datalo, datahi;
1566    MemOpIdx oi = lb->oi;
1567    MemOp opc = get_memop(oi);
1568
1569    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1570        return false;
1571    }
1572
1573    argreg = TCG_REG_R0;
1574    argreg = tcg_out_arg_reg32(s, argreg, TCG_AREG0);
1575    if (TARGET_LONG_BITS == 64) {
1576        argreg = tcg_out_arg_reg64(s, argreg, lb->addrlo_reg, lb->addrhi_reg);
1577    } else {
1578        argreg = tcg_out_arg_reg32(s, argreg, lb->addrlo_reg);
1579    }
1580
1581    datalo = lb->datalo_reg;
1582    datahi = lb->datahi_reg;
1583    switch (opc & MO_SIZE) {
1584    case MO_8:
1585        argreg = tcg_out_arg_reg8(s, argreg, datalo);
1586        break;
1587    case MO_16:
1588        argreg = tcg_out_arg_reg16(s, argreg, datalo);
1589        break;
1590    case MO_32:
1591    default:
1592        argreg = tcg_out_arg_reg32(s, argreg, datalo);
1593        break;
1594    case MO_64:
1595        argreg = tcg_out_arg_reg64(s, argreg, datalo, datahi);
1596        break;
1597    }
1598
1599    argreg = tcg_out_arg_imm32(s, argreg, oi);
1600    argreg = tcg_out_arg_reg32(s, argreg, TCG_REG_R14);
1601
1602    /* Tail-call to the helper, which will return to the fast path.  */
1603    tcg_out_goto(s, COND_AL, qemu_st_helpers[opc & MO_SIZE]);
1604    return true;
1605}
1606#else
1607
1608static void tcg_out_test_alignment(TCGContext *s, bool is_ld, TCGReg addrlo,
1609                                   TCGReg addrhi, unsigned a_bits)
1610{
1611    unsigned a_mask = (1 << a_bits) - 1;
1612    TCGLabelQemuLdst *label = new_ldst_label(s);
1613
1614    label->is_ld = is_ld;
1615    label->addrlo_reg = addrlo;
1616    label->addrhi_reg = addrhi;
1617
1618    /* We are expecting a_bits to max out at 7, and can easily support 8. */
1619    tcg_debug_assert(a_mask <= 0xff);
1620    /* tst addr, #mask */
1621    tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addrlo, a_mask);
1622
1623    /* blne slow_path */
1624    label->label_ptr[0] = s->code_ptr;
1625    tcg_out_bl_imm(s, COND_NE, 0);
1626
1627    label->raddr = tcg_splitwx_to_rx(s->code_ptr);
1628}
1629
1630static bool tcg_out_fail_alignment(TCGContext *s, TCGLabelQemuLdst *l)
1631{
1632    if (!reloc_pc24(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1633        return false;
1634    }
1635
1636    if (TARGET_LONG_BITS == 64) {
1637        /* 64-bit target address is aligned into R2:R3. */
1638        if (l->addrhi_reg != TCG_REG_R2) {
1639            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R2, l->addrlo_reg);
1640            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R3, l->addrhi_reg);
1641        } else if (l->addrlo_reg != TCG_REG_R3) {
1642            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R3, l->addrhi_reg);
1643            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R2, l->addrlo_reg);
1644        } else {
1645            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R1, TCG_REG_R2);
1646            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R2, TCG_REG_R3);
1647            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R3, TCG_REG_R1);
1648        }
1649    } else {
1650        tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R1, l->addrlo_reg);
1651    }
1652    tcg_out_mov(s, TCG_TYPE_PTR, TCG_REG_R0, TCG_AREG0);
1653
1654    /*
1655     * Tail call to the helper, with the return address back inline,
1656     * just for the clarity of the debugging traceback -- the helper
1657     * cannot return.  We have used BLNE to arrive here, so LR is
1658     * already set.
1659     */
1660    tcg_out_goto(s, COND_AL, (const void *)
1661                 (l->is_ld ? helper_unaligned_ld : helper_unaligned_st));
1662    return true;
1663}
1664
1665static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
1666{
1667    return tcg_out_fail_alignment(s, l);
1668}
1669
1670static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
1671{
1672    return tcg_out_fail_alignment(s, l);
1673}
1674#endif /* SOFTMMU */
1675
1676static void tcg_out_qemu_ld_index(TCGContext *s, MemOp opc,
1677                                  TCGReg datalo, TCGReg datahi,
1678                                  TCGReg addrlo, TCGReg addend,
1679                                  bool scratch_addend)
1680{
1681    /* Byte swapping is left to middle-end expansion. */
1682    tcg_debug_assert((opc & MO_BSWAP) == 0);
1683
1684    switch (opc & MO_SSIZE) {
1685    case MO_UB:
1686        tcg_out_ld8_r(s, COND_AL, datalo, addrlo, addend);
1687        break;
1688    case MO_SB:
1689        tcg_out_ld8s_r(s, COND_AL, datalo, addrlo, addend);
1690        break;
1691    case MO_UW:
1692        tcg_out_ld16u_r(s, COND_AL, datalo, addrlo, addend);
1693        break;
1694    case MO_SW:
1695        tcg_out_ld16s_r(s, COND_AL, datalo, addrlo, addend);
1696        break;
1697    case MO_UL:
1698        tcg_out_ld32_r(s, COND_AL, datalo, addrlo, addend);
1699        break;
1700    case MO_UQ:
1701        /* We used pair allocation for datalo, so already should be aligned. */
1702        tcg_debug_assert((datalo & 1) == 0);
1703        tcg_debug_assert(datahi == datalo + 1);
1704        /* LDRD requires alignment; double-check that. */
1705        if (get_alignment_bits(opc) >= MO_64) {
1706            /*
1707             * Rm (the second address op) must not overlap Rt or Rt + 1.
1708             * Since datalo is aligned, we can simplify the test via alignment.
1709             * Flip the two address arguments if that works.
1710             */
1711            if ((addend & ~1) != datalo) {
1712                tcg_out_ldrd_r(s, COND_AL, datalo, addrlo, addend);
1713                break;
1714            }
1715            if ((addrlo & ~1) != datalo) {
1716                tcg_out_ldrd_r(s, COND_AL, datalo, addend, addrlo);
1717                break;
1718            }
1719        }
1720        if (scratch_addend) {
1721            tcg_out_ld32_rwb(s, COND_AL, datalo, addend, addrlo);
1722            tcg_out_ld32_12(s, COND_AL, datahi, addend, 4);
1723        } else {
1724            tcg_out_dat_reg(s, COND_AL, ARITH_ADD, TCG_REG_TMP,
1725                            addend, addrlo, SHIFT_IMM_LSL(0));
1726            tcg_out_ld32_12(s, COND_AL, datalo, TCG_REG_TMP, 0);
1727            tcg_out_ld32_12(s, COND_AL, datahi, TCG_REG_TMP, 4);
1728        }
1729        break;
1730    default:
1731        g_assert_not_reached();
1732    }
1733}
1734
1735#ifndef CONFIG_SOFTMMU
1736static void tcg_out_qemu_ld_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1737                                   TCGReg datahi, TCGReg addrlo)
1738{
1739    /* Byte swapping is left to middle-end expansion. */
1740    tcg_debug_assert((opc & MO_BSWAP) == 0);
1741
1742    switch (opc & MO_SSIZE) {
1743    case MO_UB:
1744        tcg_out_ld8_12(s, COND_AL, datalo, addrlo, 0);
1745        break;
1746    case MO_SB:
1747        tcg_out_ld8s_8(s, COND_AL, datalo, addrlo, 0);
1748        break;
1749    case MO_UW:
1750        tcg_out_ld16u_8(s, COND_AL, datalo, addrlo, 0);
1751        break;
1752    case MO_SW:
1753        tcg_out_ld16s_8(s, COND_AL, datalo, addrlo, 0);
1754        break;
1755    case MO_UL:
1756        tcg_out_ld32_12(s, COND_AL, datalo, addrlo, 0);
1757        break;
1758    case MO_UQ:
1759        /* We used pair allocation for datalo, so already should be aligned. */
1760        tcg_debug_assert((datalo & 1) == 0);
1761        tcg_debug_assert(datahi == datalo + 1);
1762        /* LDRD requires alignment; double-check that. */
1763        if (get_alignment_bits(opc) >= MO_64) {
1764            tcg_out_ldrd_8(s, COND_AL, datalo, addrlo, 0);
1765        } else if (datalo == addrlo) {
1766            tcg_out_ld32_12(s, COND_AL, datahi, addrlo, 4);
1767            tcg_out_ld32_12(s, COND_AL, datalo, addrlo, 0);
1768        } else {
1769            tcg_out_ld32_12(s, COND_AL, datalo, addrlo, 0);
1770            tcg_out_ld32_12(s, COND_AL, datahi, addrlo, 4);
1771        }
1772        break;
1773    default:
1774        g_assert_not_reached();
1775    }
1776}
1777#endif
1778
1779static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, bool is64)
1780{
1781    TCGReg addrlo, datalo, datahi, addrhi __attribute__((unused));
1782    MemOpIdx oi;
1783    MemOp opc;
1784#ifdef CONFIG_SOFTMMU
1785    int mem_index;
1786    TCGReg addend;
1787    tcg_insn_unit *label_ptr;
1788#else
1789    unsigned a_bits;
1790#endif
1791
1792    datalo = *args++;
1793    datahi = (is64 ? *args++ : 0);
1794    addrlo = *args++;
1795    addrhi = (TARGET_LONG_BITS == 64 ? *args++ : 0);
1796    oi = *args++;
1797    opc = get_memop(oi);
1798
1799#ifdef CONFIG_SOFTMMU
1800    mem_index = get_mmuidx(oi);
1801    addend = tcg_out_tlb_read(s, addrlo, addrhi, opc, mem_index, 1);
1802
1803    /* This a conditional BL only to load a pointer within this opcode into LR
1804       for the slow path.  We will not be using the value for a tail call.  */
1805    label_ptr = s->code_ptr;
1806    tcg_out_bl_imm(s, COND_NE, 0);
1807
1808    tcg_out_qemu_ld_index(s, opc, datalo, datahi, addrlo, addend, true);
1809
1810    add_qemu_ldst_label(s, true, oi, datalo, datahi, addrlo, addrhi,
1811                        s->code_ptr, label_ptr);
1812#else /* !CONFIG_SOFTMMU */
1813    a_bits = get_alignment_bits(opc);
1814    if (a_bits) {
1815        tcg_out_test_alignment(s, true, addrlo, addrhi, a_bits);
1816    }
1817    if (guest_base) {
1818        tcg_out_qemu_ld_index(s, opc, datalo, datahi,
1819                              addrlo, TCG_REG_GUEST_BASE, false);
1820    } else {
1821        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, addrlo);
1822    }
1823#endif
1824}
1825
1826static void tcg_out_qemu_st_index(TCGContext *s, ARMCond cond, MemOp opc,
1827                                  TCGReg datalo, TCGReg datahi,
1828                                  TCGReg addrlo, TCGReg addend,
1829                                  bool scratch_addend)
1830{
1831    /* Byte swapping is left to middle-end expansion. */
1832    tcg_debug_assert((opc & MO_BSWAP) == 0);
1833
1834    switch (opc & MO_SIZE) {
1835    case MO_8:
1836        tcg_out_st8_r(s, cond, datalo, addrlo, addend);
1837        break;
1838    case MO_16:
1839        tcg_out_st16_r(s, cond, datalo, addrlo, addend);
1840        break;
1841    case MO_32:
1842        tcg_out_st32_r(s, cond, datalo, addrlo, addend);
1843        break;
1844    case MO_64:
1845        /* We used pair allocation for datalo, so already should be aligned. */
1846        tcg_debug_assert((datalo & 1) == 0);
1847        tcg_debug_assert(datahi == datalo + 1);
1848        /* STRD requires alignment; double-check that. */
1849        if (get_alignment_bits(opc) >= MO_64) {
1850            tcg_out_strd_r(s, cond, datalo, addrlo, addend);
1851        } else if (scratch_addend) {
1852            tcg_out_st32_rwb(s, cond, datalo, addend, addrlo);
1853            tcg_out_st32_12(s, cond, datahi, addend, 4);
1854        } else {
1855            tcg_out_dat_reg(s, cond, ARITH_ADD, TCG_REG_TMP,
1856                            addend, addrlo, SHIFT_IMM_LSL(0));
1857            tcg_out_st32_12(s, cond, datalo, TCG_REG_TMP, 0);
1858            tcg_out_st32_12(s, cond, datahi, TCG_REG_TMP, 4);
1859        }
1860        break;
1861    default:
1862        g_assert_not_reached();
1863    }
1864}
1865
1866#ifndef CONFIG_SOFTMMU
1867static void tcg_out_qemu_st_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1868                                   TCGReg datahi, TCGReg addrlo)
1869{
1870    /* Byte swapping is left to middle-end expansion. */
1871    tcg_debug_assert((opc & MO_BSWAP) == 0);
1872
1873    switch (opc & MO_SIZE) {
1874    case MO_8:
1875        tcg_out_st8_12(s, COND_AL, datalo, addrlo, 0);
1876        break;
1877    case MO_16:
1878        tcg_out_st16_8(s, COND_AL, datalo, addrlo, 0);
1879        break;
1880    case MO_32:
1881        tcg_out_st32_12(s, COND_AL, datalo, addrlo, 0);
1882        break;
1883    case MO_64:
1884        /* We used pair allocation for datalo, so already should be aligned. */
1885        tcg_debug_assert((datalo & 1) == 0);
1886        tcg_debug_assert(datahi == datalo + 1);
1887        /* STRD requires alignment; double-check that. */
1888        if (get_alignment_bits(opc) >= MO_64) {
1889            tcg_out_strd_8(s, COND_AL, datalo, addrlo, 0);
1890        } else {
1891            tcg_out_st32_12(s, COND_AL, datalo, addrlo, 0);
1892            tcg_out_st32_12(s, COND_AL, datahi, addrlo, 4);
1893        }
1894        break;
1895    default:
1896        g_assert_not_reached();
1897    }
1898}
1899#endif
1900
1901static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is64)
1902{
1903    TCGReg addrlo, datalo, datahi, addrhi __attribute__((unused));
1904    MemOpIdx oi;
1905    MemOp opc;
1906#ifdef CONFIG_SOFTMMU
1907    int mem_index;
1908    TCGReg addend;
1909    tcg_insn_unit *label_ptr;
1910#else
1911    unsigned a_bits;
1912#endif
1913
1914    datalo = *args++;
1915    datahi = (is64 ? *args++ : 0);
1916    addrlo = *args++;
1917    addrhi = (TARGET_LONG_BITS == 64 ? *args++ : 0);
1918    oi = *args++;
1919    opc = get_memop(oi);
1920
1921#ifdef CONFIG_SOFTMMU
1922    mem_index = get_mmuidx(oi);
1923    addend = tcg_out_tlb_read(s, addrlo, addrhi, opc, mem_index, 0);
1924
1925    tcg_out_qemu_st_index(s, COND_EQ, opc, datalo, datahi,
1926                          addrlo, addend, true);
1927
1928    /* The conditional call must come last, as we're going to return here.  */
1929    label_ptr = s->code_ptr;
1930    tcg_out_bl_imm(s, COND_NE, 0);
1931
1932    add_qemu_ldst_label(s, false, oi, datalo, datahi, addrlo, addrhi,
1933                        s->code_ptr, label_ptr);
1934#else /* !CONFIG_SOFTMMU */
1935    a_bits = get_alignment_bits(opc);
1936    if (a_bits) {
1937        tcg_out_test_alignment(s, false, addrlo, addrhi, a_bits);
1938    }
1939    if (guest_base) {
1940        tcg_out_qemu_st_index(s, COND_AL, opc, datalo, datahi,
1941                              addrlo, TCG_REG_GUEST_BASE, false);
1942    } else {
1943        tcg_out_qemu_st_direct(s, opc, datalo, datahi, addrlo);
1944    }
1945#endif
1946}
1947
1948static void tcg_out_epilogue(TCGContext *s);
1949
1950static void tcg_out_exit_tb(TCGContext *s, uintptr_t arg)
1951{
1952    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, arg);
1953    tcg_out_epilogue(s);
1954}
1955
1956static void tcg_out_goto_tb(TCGContext *s, int which)
1957{
1958    uintptr_t i_addr;
1959    intptr_t i_disp;
1960
1961    /* Direct branch will be patched by tb_target_set_jmp_target. */
1962    set_jmp_insn_offset(s, which);
1963    tcg_out32(s, INSN_NOP);
1964
1965    /* When branch is out of range, fall through to indirect. */
1966    i_addr = get_jmp_target_addr(s, which);
1967    i_disp = tcg_pcrel_diff(s, (void *)i_addr) - 8;
1968    tcg_debug_assert(i_disp < 0);
1969    if (i_disp >= -0xfff) {
1970        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_PC, i_disp);
1971    } else {
1972        /*
1973         * The TB is close, but outside the 12 bits addressable by
1974         * the load.  We can extend this to 20 bits with a sub of a
1975         * shifted immediate from pc.
1976         */
1977        int h = -i_disp;
1978        int l = h & 0xfff;
1979
1980        h = encode_imm_nofail(h - l);
1981        tcg_out_dat_imm(s, COND_AL, ARITH_SUB, TCG_REG_R0, TCG_REG_PC, h);
1982        tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, TCG_REG_R0, l);
1983    }
1984    set_jmp_reset_offset(s, which);
1985}
1986
1987void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
1988                              uintptr_t jmp_rx, uintptr_t jmp_rw)
1989{
1990    uintptr_t addr = tb->jmp_target_addr[n];
1991    ptrdiff_t offset = addr - (jmp_rx + 8);
1992    tcg_insn_unit insn;
1993
1994    /* Either directly branch, or fall through to indirect branch. */
1995    if (offset == sextract64(offset, 0, 26)) {
1996        /* B <addr> */
1997        insn = deposit32((COND_AL << 28) | INSN_B, 0, 24, offset >> 2);
1998    } else {
1999        insn = INSN_NOP;
2000    }
2001
2002    qatomic_set((uint32_t *)jmp_rw, insn);
2003    flush_idcache_range(jmp_rx, jmp_rw, 4);
2004}
2005
2006static void tcg_out_op(TCGContext *s, TCGOpcode opc,
2007                       const TCGArg args[TCG_MAX_OP_ARGS],
2008                       const int const_args[TCG_MAX_OP_ARGS])
2009{
2010    TCGArg a0, a1, a2, a3, a4, a5;
2011    int c;
2012
2013    switch (opc) {
2014    case INDEX_op_goto_ptr:
2015        tcg_out_b_reg(s, COND_AL, args[0]);
2016        break;
2017    case INDEX_op_br:
2018        tcg_out_goto_label(s, COND_AL, arg_label(args[0]));
2019        break;
2020
2021    case INDEX_op_ld8u_i32:
2022        tcg_out_ld8u(s, COND_AL, args[0], args[1], args[2]);
2023        break;
2024    case INDEX_op_ld8s_i32:
2025        tcg_out_ld8s(s, COND_AL, args[0], args[1], args[2]);
2026        break;
2027    case INDEX_op_ld16u_i32:
2028        tcg_out_ld16u(s, COND_AL, args[0], args[1], args[2]);
2029        break;
2030    case INDEX_op_ld16s_i32:
2031        tcg_out_ld16s(s, COND_AL, args[0], args[1], args[2]);
2032        break;
2033    case INDEX_op_ld_i32:
2034        tcg_out_ld32u(s, COND_AL, args[0], args[1], args[2]);
2035        break;
2036    case INDEX_op_st8_i32:
2037        tcg_out_st8(s, COND_AL, args[0], args[1], args[2]);
2038        break;
2039    case INDEX_op_st16_i32:
2040        tcg_out_st16(s, COND_AL, args[0], args[1], args[2]);
2041        break;
2042    case INDEX_op_st_i32:
2043        tcg_out_st32(s, COND_AL, args[0], args[1], args[2]);
2044        break;
2045
2046    case INDEX_op_movcond_i32:
2047        /* Constraints mean that v2 is always in the same register as dest,
2048         * so we only need to do "if condition passed, move v1 to dest".
2049         */
2050        tcg_out_dat_rIN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0,
2051                        args[1], args[2], const_args[2]);
2052        tcg_out_dat_rIK(s, tcg_cond_to_arm_cond[args[5]], ARITH_MOV,
2053                        ARITH_MVN, args[0], 0, args[3], const_args[3]);
2054        break;
2055    case INDEX_op_add_i32:
2056        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
2057                        args[0], args[1], args[2], const_args[2]);
2058        break;
2059    case INDEX_op_sub_i32:
2060        if (const_args[1]) {
2061            if (const_args[2]) {
2062                tcg_out_movi32(s, COND_AL, args[0], args[1] - args[2]);
2063            } else {
2064                tcg_out_dat_rI(s, COND_AL, ARITH_RSB,
2065                               args[0], args[2], args[1], 1);
2066            }
2067        } else {
2068            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB, ARITH_ADD,
2069                            args[0], args[1], args[2], const_args[2]);
2070        }
2071        break;
2072    case INDEX_op_and_i32:
2073        tcg_out_dat_rIK(s, COND_AL, ARITH_AND, ARITH_BIC,
2074                        args[0], args[1], args[2], const_args[2]);
2075        break;
2076    case INDEX_op_andc_i32:
2077        tcg_out_dat_rIK(s, COND_AL, ARITH_BIC, ARITH_AND,
2078                        args[0], args[1], args[2], const_args[2]);
2079        break;
2080    case INDEX_op_or_i32:
2081        c = ARITH_ORR;
2082        goto gen_arith;
2083    case INDEX_op_xor_i32:
2084        c = ARITH_EOR;
2085        /* Fall through.  */
2086    gen_arith:
2087        tcg_out_dat_rI(s, COND_AL, c, args[0], args[1], args[2], const_args[2]);
2088        break;
2089    case INDEX_op_add2_i32:
2090        a0 = args[0], a1 = args[1], a2 = args[2];
2091        a3 = args[3], a4 = args[4], a5 = args[5];
2092        if (a0 == a3 || (a0 == a5 && !const_args[5])) {
2093            a0 = TCG_REG_TMP;
2094        }
2095        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD | TO_CPSR, ARITH_SUB | TO_CPSR,
2096                        a0, a2, a4, const_args[4]);
2097        tcg_out_dat_rIK(s, COND_AL, ARITH_ADC, ARITH_SBC,
2098                        a1, a3, a5, const_args[5]);
2099        tcg_out_mov_reg(s, COND_AL, args[0], a0);
2100        break;
2101    case INDEX_op_sub2_i32:
2102        a0 = args[0], a1 = args[1], a2 = args[2];
2103        a3 = args[3], a4 = args[4], a5 = args[5];
2104        if ((a0 == a3 && !const_args[3]) || (a0 == a5 && !const_args[5])) {
2105            a0 = TCG_REG_TMP;
2106        }
2107        if (const_args[2]) {
2108            if (const_args[4]) {
2109                tcg_out_movi32(s, COND_AL, a0, a4);
2110                a4 = a0;
2111            }
2112            tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR, a0, a4, a2, 1);
2113        } else {
2114            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB | TO_CPSR,
2115                            ARITH_ADD | TO_CPSR, a0, a2, a4, const_args[4]);
2116        }
2117        if (const_args[3]) {
2118            if (const_args[5]) {
2119                tcg_out_movi32(s, COND_AL, a1, a5);
2120                a5 = a1;
2121            }
2122            tcg_out_dat_rI(s, COND_AL, ARITH_RSC, a1, a5, a3, 1);
2123        } else {
2124            tcg_out_dat_rIK(s, COND_AL, ARITH_SBC, ARITH_ADC,
2125                            a1, a3, a5, const_args[5]);
2126        }
2127        tcg_out_mov_reg(s, COND_AL, args[0], a0);
2128        break;
2129    case INDEX_op_neg_i32:
2130        tcg_out_dat_imm(s, COND_AL, ARITH_RSB, args[0], args[1], 0);
2131        break;
2132    case INDEX_op_not_i32:
2133        tcg_out_dat_reg(s, COND_AL,
2134                        ARITH_MVN, args[0], 0, args[1], SHIFT_IMM_LSL(0));
2135        break;
2136    case INDEX_op_mul_i32:
2137        tcg_out_mul32(s, COND_AL, args[0], args[1], args[2]);
2138        break;
2139    case INDEX_op_mulu2_i32:
2140        tcg_out_umull32(s, COND_AL, args[0], args[1], args[2], args[3]);
2141        break;
2142    case INDEX_op_muls2_i32:
2143        tcg_out_smull32(s, COND_AL, args[0], args[1], args[2], args[3]);
2144        break;
2145    /* XXX: Perhaps args[2] & 0x1f is wrong */
2146    case INDEX_op_shl_i32:
2147        c = const_args[2] ?
2148                SHIFT_IMM_LSL(args[2] & 0x1f) : SHIFT_REG_LSL(args[2]);
2149        goto gen_shift32;
2150    case INDEX_op_shr_i32:
2151        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_LSR(args[2] & 0x1f) :
2152                SHIFT_IMM_LSL(0) : SHIFT_REG_LSR(args[2]);
2153        goto gen_shift32;
2154    case INDEX_op_sar_i32:
2155        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ASR(args[2] & 0x1f) :
2156                SHIFT_IMM_LSL(0) : SHIFT_REG_ASR(args[2]);
2157        goto gen_shift32;
2158    case INDEX_op_rotr_i32:
2159        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ROR(args[2] & 0x1f) :
2160                SHIFT_IMM_LSL(0) : SHIFT_REG_ROR(args[2]);
2161        /* Fall through.  */
2162    gen_shift32:
2163        tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1], c);
2164        break;
2165
2166    case INDEX_op_rotl_i32:
2167        if (const_args[2]) {
2168            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
2169                            ((0x20 - args[2]) & 0x1f) ?
2170                            SHIFT_IMM_ROR((0x20 - args[2]) & 0x1f) :
2171                            SHIFT_IMM_LSL(0));
2172        } else {
2173            tcg_out_dat_imm(s, COND_AL, ARITH_RSB, TCG_REG_TMP, args[2], 0x20);
2174            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
2175                            SHIFT_REG_ROR(TCG_REG_TMP));
2176        }
2177        break;
2178
2179    case INDEX_op_ctz_i32:
2180        tcg_out_dat_reg(s, COND_AL, INSN_RBIT, TCG_REG_TMP, 0, args[1], 0);
2181        a1 = TCG_REG_TMP;
2182        goto do_clz;
2183
2184    case INDEX_op_clz_i32:
2185        a1 = args[1];
2186    do_clz:
2187        a0 = args[0];
2188        a2 = args[2];
2189        c = const_args[2];
2190        if (c && a2 == 32) {
2191            tcg_out_dat_reg(s, COND_AL, INSN_CLZ, a0, 0, a1, 0);
2192            break;
2193        }
2194        tcg_out_dat_imm(s, COND_AL, ARITH_CMP, 0, a1, 0);
2195        tcg_out_dat_reg(s, COND_NE, INSN_CLZ, a0, 0, a1, 0);
2196        if (c || a0 != a2) {
2197            tcg_out_dat_rIK(s, COND_EQ, ARITH_MOV, ARITH_MVN, a0, 0, a2, c);
2198        }
2199        break;
2200
2201    case INDEX_op_brcond_i32:
2202        tcg_out_dat_rIN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0,
2203                       args[0], args[1], const_args[1]);
2204        tcg_out_goto_label(s, tcg_cond_to_arm_cond[args[2]],
2205                           arg_label(args[3]));
2206        break;
2207    case INDEX_op_setcond_i32:
2208        tcg_out_dat_rIN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0,
2209                        args[1], args[2], const_args[2]);
2210        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[args[3]],
2211                        ARITH_MOV, args[0], 0, 1);
2212        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(args[3])],
2213                        ARITH_MOV, args[0], 0, 0);
2214        break;
2215
2216    case INDEX_op_brcond2_i32:
2217        c = tcg_out_cmp2(s, args, const_args);
2218        tcg_out_goto_label(s, tcg_cond_to_arm_cond[c], arg_label(args[5]));
2219        break;
2220    case INDEX_op_setcond2_i32:
2221        c = tcg_out_cmp2(s, args + 1, const_args + 1);
2222        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c], ARITH_MOV, args[0], 0, 1);
2223        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2224                        ARITH_MOV, args[0], 0, 0);
2225        break;
2226
2227    case INDEX_op_qemu_ld_i32:
2228        tcg_out_qemu_ld(s, args, 0);
2229        break;
2230    case INDEX_op_qemu_ld_i64:
2231        tcg_out_qemu_ld(s, args, 1);
2232        break;
2233    case INDEX_op_qemu_st_i32:
2234        tcg_out_qemu_st(s, args, 0);
2235        break;
2236    case INDEX_op_qemu_st_i64:
2237        tcg_out_qemu_st(s, args, 1);
2238        break;
2239
2240    case INDEX_op_bswap16_i32:
2241        tcg_out_bswap16(s, COND_AL, args[0], args[1], args[2]);
2242        break;
2243    case INDEX_op_bswap32_i32:
2244        tcg_out_bswap32(s, COND_AL, args[0], args[1]);
2245        break;
2246
2247    case INDEX_op_ext8s_i32:
2248        tcg_out_ext8s(s, COND_AL, args[0], args[1]);
2249        break;
2250    case INDEX_op_ext16s_i32:
2251        tcg_out_ext16s(s, COND_AL, args[0], args[1]);
2252        break;
2253    case INDEX_op_ext16u_i32:
2254        tcg_out_ext16u(s, COND_AL, args[0], args[1]);
2255        break;
2256
2257    case INDEX_op_deposit_i32:
2258        tcg_out_deposit(s, COND_AL, args[0], args[2],
2259                        args[3], args[4], const_args[2]);
2260        break;
2261    case INDEX_op_extract_i32:
2262        tcg_out_extract(s, COND_AL, args[0], args[1], args[2], args[3]);
2263        break;
2264    case INDEX_op_sextract_i32:
2265        tcg_out_sextract(s, COND_AL, args[0], args[1], args[2], args[3]);
2266        break;
2267    case INDEX_op_extract2_i32:
2268        /* ??? These optimization vs zero should be generic.  */
2269        /* ??? But we can't substitute 2 for 1 in the opcode stream yet.  */
2270        if (const_args[1]) {
2271            if (const_args[2]) {
2272                tcg_out_movi(s, TCG_TYPE_REG, args[0], 0);
2273            } else {
2274                tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2275                                args[2], SHIFT_IMM_LSL(32 - args[3]));
2276            }
2277        } else if (const_args[2]) {
2278            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2279                            args[1], SHIFT_IMM_LSR(args[3]));
2280        } else {
2281            /* We can do extract2 in 2 insns, vs the 3 required otherwise.  */
2282            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0,
2283                            args[2], SHIFT_IMM_LSL(32 - args[3]));
2284            tcg_out_dat_reg(s, COND_AL, ARITH_ORR, args[0], TCG_REG_TMP,
2285                            args[1], SHIFT_IMM_LSR(args[3]));
2286        }
2287        break;
2288
2289    case INDEX_op_div_i32:
2290        tcg_out_sdiv(s, COND_AL, args[0], args[1], args[2]);
2291        break;
2292    case INDEX_op_divu_i32:
2293        tcg_out_udiv(s, COND_AL, args[0], args[1], args[2]);
2294        break;
2295
2296    case INDEX_op_mb:
2297        tcg_out_mb(s, args[0]);
2298        break;
2299
2300    case INDEX_op_mov_i32:  /* Always emitted via tcg_out_mov.  */
2301    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
2302    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
2303    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
2304    default:
2305        tcg_abort();
2306    }
2307}
2308
2309static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
2310{
2311    switch (op) {
2312    case INDEX_op_goto_ptr:
2313        return C_O0_I1(r);
2314
2315    case INDEX_op_ld8u_i32:
2316    case INDEX_op_ld8s_i32:
2317    case INDEX_op_ld16u_i32:
2318    case INDEX_op_ld16s_i32:
2319    case INDEX_op_ld_i32:
2320    case INDEX_op_neg_i32:
2321    case INDEX_op_not_i32:
2322    case INDEX_op_bswap16_i32:
2323    case INDEX_op_bswap32_i32:
2324    case INDEX_op_ext8s_i32:
2325    case INDEX_op_ext16s_i32:
2326    case INDEX_op_ext16u_i32:
2327    case INDEX_op_extract_i32:
2328    case INDEX_op_sextract_i32:
2329        return C_O1_I1(r, r);
2330
2331    case INDEX_op_st8_i32:
2332    case INDEX_op_st16_i32:
2333    case INDEX_op_st_i32:
2334        return C_O0_I2(r, r);
2335
2336    case INDEX_op_add_i32:
2337    case INDEX_op_sub_i32:
2338    case INDEX_op_setcond_i32:
2339        return C_O1_I2(r, r, rIN);
2340
2341    case INDEX_op_and_i32:
2342    case INDEX_op_andc_i32:
2343    case INDEX_op_clz_i32:
2344    case INDEX_op_ctz_i32:
2345        return C_O1_I2(r, r, rIK);
2346
2347    case INDEX_op_mul_i32:
2348    case INDEX_op_div_i32:
2349    case INDEX_op_divu_i32:
2350        return C_O1_I2(r, r, r);
2351
2352    case INDEX_op_mulu2_i32:
2353    case INDEX_op_muls2_i32:
2354        return C_O2_I2(r, r, r, r);
2355
2356    case INDEX_op_or_i32:
2357    case INDEX_op_xor_i32:
2358        return C_O1_I2(r, r, rI);
2359
2360    case INDEX_op_shl_i32:
2361    case INDEX_op_shr_i32:
2362    case INDEX_op_sar_i32:
2363    case INDEX_op_rotl_i32:
2364    case INDEX_op_rotr_i32:
2365        return C_O1_I2(r, r, ri);
2366
2367    case INDEX_op_brcond_i32:
2368        return C_O0_I2(r, rIN);
2369    case INDEX_op_deposit_i32:
2370        return C_O1_I2(r, 0, rZ);
2371    case INDEX_op_extract2_i32:
2372        return C_O1_I2(r, rZ, rZ);
2373    case INDEX_op_movcond_i32:
2374        return C_O1_I4(r, r, rIN, rIK, 0);
2375    case INDEX_op_add2_i32:
2376        return C_O2_I4(r, r, r, r, rIN, rIK);
2377    case INDEX_op_sub2_i32:
2378        return C_O2_I4(r, r, rI, rI, rIN, rIK);
2379    case INDEX_op_brcond2_i32:
2380        return C_O0_I4(r, r, rI, rI);
2381    case INDEX_op_setcond2_i32:
2382        return C_O1_I4(r, r, r, rI, rI);
2383
2384    case INDEX_op_qemu_ld_i32:
2385        return TARGET_LONG_BITS == 32 ? C_O1_I1(r, l) : C_O1_I2(r, l, l);
2386    case INDEX_op_qemu_ld_i64:
2387        return TARGET_LONG_BITS == 32 ? C_O2_I1(e, p, l) : C_O2_I2(e, p, l, l);
2388    case INDEX_op_qemu_st_i32:
2389        return TARGET_LONG_BITS == 32 ? C_O0_I2(s, s) : C_O0_I3(s, s, s);
2390    case INDEX_op_qemu_st_i64:
2391        return TARGET_LONG_BITS == 32 ? C_O0_I3(S, p, s) : C_O0_I4(S, p, s, s);
2392
2393    case INDEX_op_st_vec:
2394        return C_O0_I2(w, r);
2395    case INDEX_op_ld_vec:
2396    case INDEX_op_dupm_vec:
2397        return C_O1_I1(w, r);
2398    case INDEX_op_dup_vec:
2399        return C_O1_I1(w, wr);
2400    case INDEX_op_abs_vec:
2401    case INDEX_op_neg_vec:
2402    case INDEX_op_not_vec:
2403    case INDEX_op_shli_vec:
2404    case INDEX_op_shri_vec:
2405    case INDEX_op_sari_vec:
2406        return C_O1_I1(w, w);
2407    case INDEX_op_dup2_vec:
2408    case INDEX_op_add_vec:
2409    case INDEX_op_mul_vec:
2410    case INDEX_op_smax_vec:
2411    case INDEX_op_smin_vec:
2412    case INDEX_op_ssadd_vec:
2413    case INDEX_op_sssub_vec:
2414    case INDEX_op_sub_vec:
2415    case INDEX_op_umax_vec:
2416    case INDEX_op_umin_vec:
2417    case INDEX_op_usadd_vec:
2418    case INDEX_op_ussub_vec:
2419    case INDEX_op_xor_vec:
2420    case INDEX_op_arm_sshl_vec:
2421    case INDEX_op_arm_ushl_vec:
2422        return C_O1_I2(w, w, w);
2423    case INDEX_op_arm_sli_vec:
2424        return C_O1_I2(w, 0, w);
2425    case INDEX_op_or_vec:
2426    case INDEX_op_andc_vec:
2427        return C_O1_I2(w, w, wO);
2428    case INDEX_op_and_vec:
2429    case INDEX_op_orc_vec:
2430        return C_O1_I2(w, w, wV);
2431    case INDEX_op_cmp_vec:
2432        return C_O1_I2(w, w, wZ);
2433    case INDEX_op_bitsel_vec:
2434        return C_O1_I3(w, w, w, w);
2435    default:
2436        g_assert_not_reached();
2437    }
2438}
2439
2440static void tcg_target_init(TCGContext *s)
2441{
2442    /*
2443     * Only probe for the platform and capabilities if we haven't already
2444     * determined maximum values at compile time.
2445     */
2446#if !defined(use_idiv_instructions) || !defined(use_neon_instructions)
2447    {
2448        unsigned long hwcap = qemu_getauxval(AT_HWCAP);
2449#ifndef use_idiv_instructions
2450        use_idiv_instructions = (hwcap & HWCAP_ARM_IDIVA) != 0;
2451#endif
2452#ifndef use_neon_instructions
2453        use_neon_instructions = (hwcap & HWCAP_ARM_NEON) != 0;
2454#endif
2455    }
2456#endif
2457
2458    if (__ARM_ARCH < 7) {
2459        const char *pl = (const char *)qemu_getauxval(AT_PLATFORM);
2460        if (pl != NULL && pl[0] == 'v' && pl[1] >= '4' && pl[1] <= '9') {
2461            arm_arch = pl[1] - '0';
2462        }
2463
2464        if (arm_arch < 6) {
2465            error_report("TCG: ARMv%d is unsupported; exiting", arm_arch);
2466            exit(EXIT_FAILURE);
2467        }
2468    }
2469
2470    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
2471
2472    tcg_target_call_clobber_regs = 0;
2473    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R0);
2474    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R1);
2475    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R2);
2476    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R3);
2477    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R12);
2478    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R14);
2479
2480    if (use_neon_instructions) {
2481        tcg_target_available_regs[TCG_TYPE_V64]  = ALL_VECTOR_REGS;
2482        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
2483
2484        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q0);
2485        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q1);
2486        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q2);
2487        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q3);
2488        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q8);
2489        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q9);
2490        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q10);
2491        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q11);
2492        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q12);
2493        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q13);
2494        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q14);
2495        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q15);
2496    }
2497
2498    s->reserved_regs = 0;
2499    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
2500    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP);
2501    tcg_regset_set_reg(s->reserved_regs, TCG_REG_PC);
2502    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP);
2503}
2504
2505static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
2506                       TCGReg arg1, intptr_t arg2)
2507{
2508    switch (type) {
2509    case TCG_TYPE_I32:
2510        tcg_out_ld32u(s, COND_AL, arg, arg1, arg2);
2511        return;
2512    case TCG_TYPE_V64:
2513        /* regs 1; size 8; align 8 */
2514        tcg_out_vldst(s, INSN_VLD1 | 0x7d0, arg, arg1, arg2);
2515        return;
2516    case TCG_TYPE_V128:
2517        /*
2518         * We have only 8-byte alignment for the stack per the ABI.
2519         * Rather than dynamically re-align the stack, it's easier
2520         * to simply not request alignment beyond that.  So:
2521         * regs 2; size 8; align 8
2522         */
2523        tcg_out_vldst(s, INSN_VLD1 | 0xad0, arg, arg1, arg2);
2524        return;
2525    default:
2526        g_assert_not_reached();
2527    }
2528}
2529
2530static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
2531                       TCGReg arg1, intptr_t arg2)
2532{
2533    switch (type) {
2534    case TCG_TYPE_I32:
2535        tcg_out_st32(s, COND_AL, arg, arg1, arg2);
2536        return;
2537    case TCG_TYPE_V64:
2538        /* regs 1; size 8; align 8 */
2539        tcg_out_vldst(s, INSN_VST1 | 0x7d0, arg, arg1, arg2);
2540        return;
2541    case TCG_TYPE_V128:
2542        /* See tcg_out_ld re alignment: regs 2; size 8; align 8 */
2543        tcg_out_vldst(s, INSN_VST1 | 0xad0, arg, arg1, arg2);
2544        return;
2545    default:
2546        g_assert_not_reached();
2547    }
2548}
2549
2550static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
2551                        TCGReg base, intptr_t ofs)
2552{
2553    return false;
2554}
2555
2556static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
2557{
2558    if (ret == arg) {
2559        return true;
2560    }
2561    switch (type) {
2562    case TCG_TYPE_I32:
2563        if (ret < TCG_REG_Q0 && arg < TCG_REG_Q0) {
2564            tcg_out_mov_reg(s, COND_AL, ret, arg);
2565            return true;
2566        }
2567        return false;
2568
2569    case TCG_TYPE_V64:
2570    case TCG_TYPE_V128:
2571        /* "VMOV D,N" is an alias for "VORR D,N,N". */
2572        tcg_out_vreg3(s, INSN_VORR, type - TCG_TYPE_V64, 0, ret, arg, arg);
2573        return true;
2574
2575    default:
2576        g_assert_not_reached();
2577    }
2578}
2579
2580static void tcg_out_movi(TCGContext *s, TCGType type,
2581                         TCGReg ret, tcg_target_long arg)
2582{
2583    tcg_debug_assert(type == TCG_TYPE_I32);
2584    tcg_debug_assert(ret < TCG_REG_Q0);
2585    tcg_out_movi32(s, COND_AL, ret, arg);
2586}
2587
2588static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
2589                             tcg_target_long imm)
2590{
2591    int enc, opc = ARITH_ADD;
2592
2593    /* All of the easiest immediates to encode are positive. */
2594    if (imm < 0) {
2595        imm = -imm;
2596        opc = ARITH_SUB;
2597    }
2598    enc = encode_imm(imm);
2599    if (enc >= 0) {
2600        tcg_out_dat_imm(s, COND_AL, opc, rd, rs, enc);
2601    } else {
2602        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, imm);
2603        tcg_out_dat_reg(s, COND_AL, opc, rd, rs,
2604                        TCG_REG_TMP, SHIFT_IMM_LSL(0));
2605    }
2606}
2607
2608/* Type is always V128, with I64 elements.  */
2609static void tcg_out_dup2_vec(TCGContext *s, TCGReg rd, TCGReg rl, TCGReg rh)
2610{
2611    /* Move high element into place first. */
2612    /* VMOV Dd+1, Ds */
2613    tcg_out_vreg3(s, INSN_VORR | (1 << 12), 0, 0, rd, rh, rh);
2614    /* Move low element into place; tcg_out_mov will check for nop. */
2615    tcg_out_mov(s, TCG_TYPE_V64, rd, rl);
2616}
2617
2618static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
2619                            TCGReg rd, TCGReg rs)
2620{
2621    int q = type - TCG_TYPE_V64;
2622
2623    if (vece == MO_64) {
2624        if (type == TCG_TYPE_V128) {
2625            tcg_out_dup2_vec(s, rd, rs, rs);
2626        } else {
2627            tcg_out_mov(s, TCG_TYPE_V64, rd, rs);
2628        }
2629    } else if (rs < TCG_REG_Q0) {
2630        int b = (vece == MO_8);
2631        int e = (vece == MO_16);
2632        tcg_out32(s, INSN_VDUP_G | (b << 22) | (q << 21) | (e << 5) |
2633                  encode_vn(rd) | (rs << 12));
2634    } else {
2635        int imm4 = 1 << vece;
2636        tcg_out32(s, INSN_VDUP_S | (imm4 << 16) | (q << 6) |
2637                  encode_vd(rd) | encode_vm(rs));
2638    }
2639    return true;
2640}
2641
2642static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
2643                             TCGReg rd, TCGReg base, intptr_t offset)
2644{
2645    if (vece == MO_64) {
2646        tcg_out_ld(s, TCG_TYPE_V64, rd, base, offset);
2647        if (type == TCG_TYPE_V128) {
2648            tcg_out_dup2_vec(s, rd, rd, rd);
2649        }
2650    } else {
2651        int q = type - TCG_TYPE_V64;
2652        tcg_out_vldst(s, INSN_VLD1R | (vece << 6) | (q << 5),
2653                      rd, base, offset);
2654    }
2655    return true;
2656}
2657
2658static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
2659                             TCGReg rd, int64_t v64)
2660{
2661    int q = type - TCG_TYPE_V64;
2662    int cmode, imm8, i;
2663
2664    /* Test all bytes equal first.  */
2665    if (vece == MO_8) {
2666        tcg_out_vmovi(s, rd, q, 0, 0xe, v64);
2667        return;
2668    }
2669
2670    /*
2671     * Test all bytes 0x00 or 0xff second.  This can match cases that
2672     * might otherwise take 2 or 3 insns for MO_16 or MO_32 below.
2673     */
2674    for (i = imm8 = 0; i < 8; i++) {
2675        uint8_t byte = v64 >> (i * 8);
2676        if (byte == 0xff) {
2677            imm8 |= 1 << i;
2678        } else if (byte != 0) {
2679            goto fail_bytes;
2680        }
2681    }
2682    tcg_out_vmovi(s, rd, q, 1, 0xe, imm8);
2683    return;
2684 fail_bytes:
2685
2686    /*
2687     * Tests for various replications.  For each element width, if we
2688     * cannot find an expansion there's no point checking a larger
2689     * width because we already know by replication it cannot match.
2690     */
2691    if (vece == MO_16) {
2692        uint16_t v16 = v64;
2693
2694        if (is_shimm16(v16, &cmode, &imm8)) {
2695            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2696            return;
2697        }
2698        if (is_shimm16(~v16, &cmode, &imm8)) {
2699            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2700            return;
2701        }
2702
2703        /*
2704         * Otherwise, all remaining constants can be loaded in two insns:
2705         * rd = v16 & 0xff, rd |= v16 & 0xff00.
2706         */
2707        tcg_out_vmovi(s, rd, q, 0, 0x8, v16 & 0xff);
2708        tcg_out_vmovi(s, rd, q, 0, 0xb, v16 >> 8);   /* VORRI */
2709        return;
2710    }
2711
2712    if (vece == MO_32) {
2713        uint32_t v32 = v64;
2714
2715        if (is_shimm32(v32, &cmode, &imm8) ||
2716            is_soimm32(v32, &cmode, &imm8)) {
2717            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2718            return;
2719        }
2720        if (is_shimm32(~v32, &cmode, &imm8) ||
2721            is_soimm32(~v32, &cmode, &imm8)) {
2722            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2723            return;
2724        }
2725
2726        /*
2727         * Restrict the set of constants to those we can load with
2728         * two instructions.  Others we load from the pool.
2729         */
2730        i = is_shimm32_pair(v32, &cmode, &imm8);
2731        if (i) {
2732            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2733            tcg_out_vmovi(s, rd, q, 0, i | 1, extract32(v32, i * 4, 8));
2734            return;
2735        }
2736        i = is_shimm32_pair(~v32, &cmode, &imm8);
2737        if (i) {
2738            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2739            tcg_out_vmovi(s, rd, q, 1, i | 1, extract32(~v32, i * 4, 8));
2740            return;
2741        }
2742    }
2743
2744    /*
2745     * As a last resort, load from the constant pool.
2746     */
2747    if (!q || vece == MO_64) {
2748        new_pool_l2(s, R_ARM_PC11, s->code_ptr, 0, v64, v64 >> 32);
2749        /* VLDR Dd, [pc + offset] */
2750        tcg_out32(s, INSN_VLDR_D | encode_vd(rd) | (0xf << 16));
2751        if (q) {
2752            tcg_out_dup2_vec(s, rd, rd, rd);
2753        }
2754    } else {
2755        new_pool_label(s, (uint32_t)v64, R_ARM_PC8, s->code_ptr, 0);
2756        /* add tmp, pc, offset */
2757        tcg_out_dat_imm(s, COND_AL, ARITH_ADD, TCG_REG_TMP, TCG_REG_PC, 0);
2758        tcg_out_dupm_vec(s, type, MO_32, rd, TCG_REG_TMP, 0);
2759    }
2760}
2761
2762static const ARMInsn vec_cmp_insn[16] = {
2763    [TCG_COND_EQ] = INSN_VCEQ,
2764    [TCG_COND_GT] = INSN_VCGT,
2765    [TCG_COND_GE] = INSN_VCGE,
2766    [TCG_COND_GTU] = INSN_VCGT_U,
2767    [TCG_COND_GEU] = INSN_VCGE_U,
2768};
2769
2770static const ARMInsn vec_cmp0_insn[16] = {
2771    [TCG_COND_EQ] = INSN_VCEQ0,
2772    [TCG_COND_GT] = INSN_VCGT0,
2773    [TCG_COND_GE] = INSN_VCGE0,
2774    [TCG_COND_LT] = INSN_VCLT0,
2775    [TCG_COND_LE] = INSN_VCLE0,
2776};
2777
2778static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2779                           unsigned vecl, unsigned vece,
2780                           const TCGArg args[TCG_MAX_OP_ARGS],
2781                           const int const_args[TCG_MAX_OP_ARGS])
2782{
2783    TCGType type = vecl + TCG_TYPE_V64;
2784    unsigned q = vecl;
2785    TCGArg a0, a1, a2, a3;
2786    int cmode, imm8;
2787
2788    a0 = args[0];
2789    a1 = args[1];
2790    a2 = args[2];
2791
2792    switch (opc) {
2793    case INDEX_op_ld_vec:
2794        tcg_out_ld(s, type, a0, a1, a2);
2795        return;
2796    case INDEX_op_st_vec:
2797        tcg_out_st(s, type, a0, a1, a2);
2798        return;
2799    case INDEX_op_dupm_vec:
2800        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
2801        return;
2802    case INDEX_op_dup2_vec:
2803        tcg_out_dup2_vec(s, a0, a1, a2);
2804        return;
2805    case INDEX_op_abs_vec:
2806        tcg_out_vreg2(s, INSN_VABS, q, vece, a0, a1);
2807        return;
2808    case INDEX_op_neg_vec:
2809        tcg_out_vreg2(s, INSN_VNEG, q, vece, a0, a1);
2810        return;
2811    case INDEX_op_not_vec:
2812        tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a1);
2813        return;
2814    case INDEX_op_add_vec:
2815        tcg_out_vreg3(s, INSN_VADD, q, vece, a0, a1, a2);
2816        return;
2817    case INDEX_op_mul_vec:
2818        tcg_out_vreg3(s, INSN_VMUL, q, vece, a0, a1, a2);
2819        return;
2820    case INDEX_op_smax_vec:
2821        tcg_out_vreg3(s, INSN_VMAX, q, vece, a0, a1, a2);
2822        return;
2823    case INDEX_op_smin_vec:
2824        tcg_out_vreg3(s, INSN_VMIN, q, vece, a0, a1, a2);
2825        return;
2826    case INDEX_op_sub_vec:
2827        tcg_out_vreg3(s, INSN_VSUB, q, vece, a0, a1, a2);
2828        return;
2829    case INDEX_op_ssadd_vec:
2830        tcg_out_vreg3(s, INSN_VQADD, q, vece, a0, a1, a2);
2831        return;
2832    case INDEX_op_sssub_vec:
2833        tcg_out_vreg3(s, INSN_VQSUB, q, vece, a0, a1, a2);
2834        return;
2835    case INDEX_op_umax_vec:
2836        tcg_out_vreg3(s, INSN_VMAX_U, q, vece, a0, a1, a2);
2837        return;
2838    case INDEX_op_umin_vec:
2839        tcg_out_vreg3(s, INSN_VMIN_U, q, vece, a0, a1, a2);
2840        return;
2841    case INDEX_op_usadd_vec:
2842        tcg_out_vreg3(s, INSN_VQADD_U, q, vece, a0, a1, a2);
2843        return;
2844    case INDEX_op_ussub_vec:
2845        tcg_out_vreg3(s, INSN_VQSUB_U, q, vece, a0, a1, a2);
2846        return;
2847    case INDEX_op_xor_vec:
2848        tcg_out_vreg3(s, INSN_VEOR, q, 0, a0, a1, a2);
2849        return;
2850    case INDEX_op_arm_sshl_vec:
2851        /*
2852         * Note that Vm is the data and Vn is the shift count,
2853         * therefore the arguments appear reversed.
2854         */
2855        tcg_out_vreg3(s, INSN_VSHL_S, q, vece, a0, a2, a1);
2856        return;
2857    case INDEX_op_arm_ushl_vec:
2858        /* See above. */
2859        tcg_out_vreg3(s, INSN_VSHL_U, q, vece, a0, a2, a1);
2860        return;
2861    case INDEX_op_shli_vec:
2862        tcg_out_vshifti(s, INSN_VSHLI, q, a0, a1, a2 + (8 << vece));
2863        return;
2864    case INDEX_op_shri_vec:
2865        tcg_out_vshifti(s, INSN_VSHRI, q, a0, a1, (16 << vece) - a2);
2866        return;
2867    case INDEX_op_sari_vec:
2868        tcg_out_vshifti(s, INSN_VSARI, q, a0, a1, (16 << vece) - a2);
2869        return;
2870    case INDEX_op_arm_sli_vec:
2871        tcg_out_vshifti(s, INSN_VSLI, q, a0, a2, args[3] + (8 << vece));
2872        return;
2873
2874    case INDEX_op_andc_vec:
2875        if (!const_args[2]) {
2876            tcg_out_vreg3(s, INSN_VBIC, q, 0, a0, a1, a2);
2877            return;
2878        }
2879        a2 = ~a2;
2880        /* fall through */
2881    case INDEX_op_and_vec:
2882        if (const_args[2]) {
2883            is_shimm1632(~a2, &cmode, &imm8);
2884            if (a0 == a1) {
2885                tcg_out_vmovi(s, a0, q, 1, cmode | 1, imm8); /* VBICI */
2886                return;
2887            }
2888            tcg_out_vmovi(s, a0, q, 1, cmode, imm8); /* VMVNI */
2889            a2 = a0;
2890        }
2891        tcg_out_vreg3(s, INSN_VAND, q, 0, a0, a1, a2);
2892        return;
2893
2894    case INDEX_op_orc_vec:
2895        if (!const_args[2]) {
2896            tcg_out_vreg3(s, INSN_VORN, q, 0, a0, a1, a2);
2897            return;
2898        }
2899        a2 = ~a2;
2900        /* fall through */
2901    case INDEX_op_or_vec:
2902        if (const_args[2]) {
2903            is_shimm1632(a2, &cmode, &imm8);
2904            if (a0 == a1) {
2905                tcg_out_vmovi(s, a0, q, 0, cmode | 1, imm8); /* VORRI */
2906                return;
2907            }
2908            tcg_out_vmovi(s, a0, q, 0, cmode, imm8); /* VMOVI */
2909            a2 = a0;
2910        }
2911        tcg_out_vreg3(s, INSN_VORR, q, 0, a0, a1, a2);
2912        return;
2913
2914    case INDEX_op_cmp_vec:
2915        {
2916            TCGCond cond = args[3];
2917
2918            if (cond == TCG_COND_NE) {
2919                if (const_args[2]) {
2920                    tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a1);
2921                } else {
2922                    tcg_out_vreg3(s, INSN_VCEQ, q, vece, a0, a1, a2);
2923                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
2924                }
2925            } else {
2926                ARMInsn insn;
2927
2928                if (const_args[2]) {
2929                    insn = vec_cmp0_insn[cond];
2930                    if (insn) {
2931                        tcg_out_vreg2(s, insn, q, vece, a0, a1);
2932                        return;
2933                    }
2934                    tcg_out_dupi_vec(s, type, MO_8, TCG_VEC_TMP, 0);
2935                    a2 = TCG_VEC_TMP;
2936                }
2937                insn = vec_cmp_insn[cond];
2938                if (insn == 0) {
2939                    TCGArg t;
2940                    t = a1, a1 = a2, a2 = t;
2941                    cond = tcg_swap_cond(cond);
2942                    insn = vec_cmp_insn[cond];
2943                    tcg_debug_assert(insn != 0);
2944                }
2945                tcg_out_vreg3(s, insn, q, vece, a0, a1, a2);
2946            }
2947        }
2948        return;
2949
2950    case INDEX_op_bitsel_vec:
2951        a3 = args[3];
2952        if (a0 == a3) {
2953            tcg_out_vreg3(s, INSN_VBIT, q, 0, a0, a2, a1);
2954        } else if (a0 == a2) {
2955            tcg_out_vreg3(s, INSN_VBIF, q, 0, a0, a3, a1);
2956        } else {
2957            tcg_out_mov(s, type, a0, a1);
2958            tcg_out_vreg3(s, INSN_VBSL, q, 0, a0, a2, a3);
2959        }
2960        return;
2961
2962    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
2963    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
2964    default:
2965        g_assert_not_reached();
2966    }
2967}
2968
2969int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
2970{
2971    switch (opc) {
2972    case INDEX_op_add_vec:
2973    case INDEX_op_sub_vec:
2974    case INDEX_op_and_vec:
2975    case INDEX_op_andc_vec:
2976    case INDEX_op_or_vec:
2977    case INDEX_op_orc_vec:
2978    case INDEX_op_xor_vec:
2979    case INDEX_op_not_vec:
2980    case INDEX_op_shli_vec:
2981    case INDEX_op_shri_vec:
2982    case INDEX_op_sari_vec:
2983    case INDEX_op_ssadd_vec:
2984    case INDEX_op_sssub_vec:
2985    case INDEX_op_usadd_vec:
2986    case INDEX_op_ussub_vec:
2987    case INDEX_op_bitsel_vec:
2988        return 1;
2989    case INDEX_op_abs_vec:
2990    case INDEX_op_cmp_vec:
2991    case INDEX_op_mul_vec:
2992    case INDEX_op_neg_vec:
2993    case INDEX_op_smax_vec:
2994    case INDEX_op_smin_vec:
2995    case INDEX_op_umax_vec:
2996    case INDEX_op_umin_vec:
2997        return vece < MO_64;
2998    case INDEX_op_shlv_vec:
2999    case INDEX_op_shrv_vec:
3000    case INDEX_op_sarv_vec:
3001    case INDEX_op_rotli_vec:
3002    case INDEX_op_rotlv_vec:
3003    case INDEX_op_rotrv_vec:
3004        return -1;
3005    default:
3006        return 0;
3007    }
3008}
3009
3010void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
3011                       TCGArg a0, ...)
3012{
3013    va_list va;
3014    TCGv_vec v0, v1, v2, t1, t2, c1;
3015    TCGArg a2;
3016
3017    va_start(va, a0);
3018    v0 = temp_tcgv_vec(arg_temp(a0));
3019    v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
3020    a2 = va_arg(va, TCGArg);
3021    va_end(va);
3022
3023    switch (opc) {
3024    case INDEX_op_shlv_vec:
3025        /*
3026         * Merely propagate shlv_vec to arm_ushl_vec.
3027         * In this way we don't set TCG_TARGET_HAS_shv_vec
3028         * because everything is done via expansion.
3029         */
3030        v2 = temp_tcgv_vec(arg_temp(a2));
3031        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
3032                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
3033        break;
3034
3035    case INDEX_op_shrv_vec:
3036    case INDEX_op_sarv_vec:
3037        /* Right shifts are negative left shifts for NEON.  */
3038        v2 = temp_tcgv_vec(arg_temp(a2));
3039        t1 = tcg_temp_new_vec(type);
3040        tcg_gen_neg_vec(vece, t1, v2);
3041        if (opc == INDEX_op_shrv_vec) {
3042            opc = INDEX_op_arm_ushl_vec;
3043        } else {
3044            opc = INDEX_op_arm_sshl_vec;
3045        }
3046        vec_gen_3(opc, type, vece, tcgv_vec_arg(v0),
3047                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
3048        tcg_temp_free_vec(t1);
3049        break;
3050
3051    case INDEX_op_rotli_vec:
3052        t1 = tcg_temp_new_vec(type);
3053        tcg_gen_shri_vec(vece, t1, v1, -a2 & ((8 << vece) - 1));
3054        vec_gen_4(INDEX_op_arm_sli_vec, type, vece,
3055                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(v1), a2);
3056        tcg_temp_free_vec(t1);
3057        break;
3058
3059    case INDEX_op_rotlv_vec:
3060        v2 = temp_tcgv_vec(arg_temp(a2));
3061        t1 = tcg_temp_new_vec(type);
3062        c1 = tcg_constant_vec(type, vece, 8 << vece);
3063        tcg_gen_sub_vec(vece, t1, v2, c1);
3064        /* Right shifts are negative left shifts for NEON.  */
3065        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
3066                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
3067        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
3068                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
3069        tcg_gen_or_vec(vece, v0, v0, t1);
3070        tcg_temp_free_vec(t1);
3071        break;
3072
3073    case INDEX_op_rotrv_vec:
3074        v2 = temp_tcgv_vec(arg_temp(a2));
3075        t1 = tcg_temp_new_vec(type);
3076        t2 = tcg_temp_new_vec(type);
3077        c1 = tcg_constant_vec(type, vece, 8 << vece);
3078        tcg_gen_neg_vec(vece, t1, v2);
3079        tcg_gen_sub_vec(vece, t2, c1, v2);
3080        /* Right shifts are negative left shifts for NEON.  */
3081        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
3082                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
3083        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t2),
3084                  tcgv_vec_arg(v1), tcgv_vec_arg(t2));
3085        tcg_gen_or_vec(vece, v0, t1, t2);
3086        tcg_temp_free_vec(t1);
3087        tcg_temp_free_vec(t2);
3088        break;
3089
3090    default:
3091        g_assert_not_reached();
3092    }
3093}
3094
3095static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
3096{
3097    int i;
3098    for (i = 0; i < count; ++i) {
3099        p[i] = INSN_NOP;
3100    }
3101}
3102
3103/* Compute frame size via macros, to share between tcg_target_qemu_prologue
3104   and tcg_register_jit.  */
3105
3106#define PUSH_SIZE  ((11 - 4 + 1 + 1) * sizeof(tcg_target_long))
3107
3108#define FRAME_SIZE \
3109    ((PUSH_SIZE \
3110      + TCG_STATIC_CALL_ARGS_SIZE \
3111      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
3112      + TCG_TARGET_STACK_ALIGN - 1) \
3113     & -TCG_TARGET_STACK_ALIGN)
3114
3115#define STACK_ADDEND  (FRAME_SIZE - PUSH_SIZE)
3116
3117static void tcg_target_qemu_prologue(TCGContext *s)
3118{
3119    /* Calling convention requires us to save r4-r11 and lr.  */
3120    /* stmdb sp!, { r4 - r11, lr } */
3121    tcg_out_ldstm(s, COND_AL, INSN_STMDB, TCG_REG_CALL_STACK,
3122                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
3123                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3124                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_R14));
3125
3126    /* Reserve callee argument and tcg temp space.  */
3127    tcg_out_dat_rI(s, COND_AL, ARITH_SUB, TCG_REG_CALL_STACK,
3128                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
3129    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
3130                  CPU_TEMP_BUF_NLONGS * sizeof(long));
3131
3132    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
3133
3134#ifndef CONFIG_SOFTMMU
3135    if (guest_base) {
3136        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_GUEST_BASE, guest_base);
3137        tcg_regset_set_reg(s->reserved_regs, TCG_REG_GUEST_BASE);
3138    }
3139#endif
3140
3141    tcg_out_b_reg(s, COND_AL, tcg_target_call_iarg_regs[1]);
3142
3143    /*
3144     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
3145     * and fall through to the rest of the epilogue.
3146     */
3147    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
3148    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, 0);
3149    tcg_out_epilogue(s);
3150}
3151
3152static void tcg_out_epilogue(TCGContext *s)
3153{
3154    /* Release local stack frame.  */
3155    tcg_out_dat_rI(s, COND_AL, ARITH_ADD, TCG_REG_CALL_STACK,
3156                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
3157
3158    /* ldmia sp!, { r4 - r11, pc } */
3159    tcg_out_ldstm(s, COND_AL, INSN_LDMIA, TCG_REG_CALL_STACK,
3160                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
3161                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3162                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_PC));
3163}
3164
3165typedef struct {
3166    DebugFrameHeader h;
3167    uint8_t fde_def_cfa[4];
3168    uint8_t fde_reg_ofs[18];
3169} DebugFrame;
3170
3171#define ELF_HOST_MACHINE EM_ARM
3172
3173/* We're expecting a 2 byte uleb128 encoded value.  */
3174QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
3175
3176static const DebugFrame debug_frame = {
3177    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
3178    .h.cie.id = -1,
3179    .h.cie.version = 1,
3180    .h.cie.code_align = 1,
3181    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
3182    .h.cie.return_column = 14,
3183
3184    /* Total FDE size does not include the "len" member.  */
3185    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
3186
3187    .fde_def_cfa = {
3188        12, 13,                         /* DW_CFA_def_cfa sp, ... */
3189        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
3190        (FRAME_SIZE >> 7)
3191    },
3192    .fde_reg_ofs = {
3193        /* The following must match the stmdb in the prologue.  */
3194        0x8e, 1,                        /* DW_CFA_offset, lr, -4 */
3195        0x8b, 2,                        /* DW_CFA_offset, r11, -8 */
3196        0x8a, 3,                        /* DW_CFA_offset, r10, -12 */
3197        0x89, 4,                        /* DW_CFA_offset, r9, -16 */
3198        0x88, 5,                        /* DW_CFA_offset, r8, -20 */
3199        0x87, 6,                        /* DW_CFA_offset, r7, -24 */
3200        0x86, 7,                        /* DW_CFA_offset, r6, -28 */
3201        0x85, 8,                        /* DW_CFA_offset, r5, -32 */
3202        0x84, 9,                        /* DW_CFA_offset, r4, -36 */
3203    }
3204};
3205
3206void tcg_register_jit(const void *buf, size_t buf_size)
3207{
3208    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
3209}
3210