xref: /openbmc/qemu/tcg/arm/tcg-target.c.inc (revision 5242ef88)
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};
82static const int tcg_target_call_oarg_regs[2] = {
83    TCG_REG_R0, TCG_REG_R1
84};
85
86#define TCG_REG_TMP  TCG_REG_R12
87#define TCG_VEC_TMP  TCG_REG_Q15
88#ifndef CONFIG_SOFTMMU
89#define TCG_REG_GUEST_BASE  TCG_REG_R11
90#endif
91
92typedef enum {
93    COND_EQ = 0x0,
94    COND_NE = 0x1,
95    COND_CS = 0x2,	/* Unsigned greater or equal */
96    COND_CC = 0x3,	/* Unsigned less than */
97    COND_MI = 0x4,	/* Negative */
98    COND_PL = 0x5,	/* Zero or greater */
99    COND_VS = 0x6,	/* Overflow */
100    COND_VC = 0x7,	/* No overflow */
101    COND_HI = 0x8,	/* Unsigned greater than */
102    COND_LS = 0x9,	/* Unsigned less or equal */
103    COND_GE = 0xa,
104    COND_LT = 0xb,
105    COND_GT = 0xc,
106    COND_LE = 0xd,
107    COND_AL = 0xe,
108} ARMCond;
109
110#define TO_CPSR (1 << 20)
111
112#define SHIFT_IMM_LSL(im)	(((im) << 7) | 0x00)
113#define SHIFT_IMM_LSR(im)	(((im) << 7) | 0x20)
114#define SHIFT_IMM_ASR(im)	(((im) << 7) | 0x40)
115#define SHIFT_IMM_ROR(im)	(((im) << 7) | 0x60)
116#define SHIFT_REG_LSL(rs)	(((rs) << 8) | 0x10)
117#define SHIFT_REG_LSR(rs)	(((rs) << 8) | 0x30)
118#define SHIFT_REG_ASR(rs)	(((rs) << 8) | 0x50)
119#define SHIFT_REG_ROR(rs)	(((rs) << 8) | 0x70)
120
121typedef enum {
122    ARITH_AND = 0x0 << 21,
123    ARITH_EOR = 0x1 << 21,
124    ARITH_SUB = 0x2 << 21,
125    ARITH_RSB = 0x3 << 21,
126    ARITH_ADD = 0x4 << 21,
127    ARITH_ADC = 0x5 << 21,
128    ARITH_SBC = 0x6 << 21,
129    ARITH_RSC = 0x7 << 21,
130    ARITH_TST = 0x8 << 21 | TO_CPSR,
131    ARITH_CMP = 0xa << 21 | TO_CPSR,
132    ARITH_CMN = 0xb << 21 | TO_CPSR,
133    ARITH_ORR = 0xc << 21,
134    ARITH_MOV = 0xd << 21,
135    ARITH_BIC = 0xe << 21,
136    ARITH_MVN = 0xf << 21,
137
138    INSN_CLZ       = 0x016f0f10,
139    INSN_RBIT      = 0x06ff0f30,
140
141    INSN_LDMIA     = 0x08b00000,
142    INSN_STMDB     = 0x09200000,
143
144    INSN_LDR_IMM   = 0x04100000,
145    INSN_LDR_REG   = 0x06100000,
146    INSN_STR_IMM   = 0x04000000,
147    INSN_STR_REG   = 0x06000000,
148
149    INSN_LDRH_IMM  = 0x005000b0,
150    INSN_LDRH_REG  = 0x001000b0,
151    INSN_LDRSH_IMM = 0x005000f0,
152    INSN_LDRSH_REG = 0x001000f0,
153    INSN_STRH_IMM  = 0x004000b0,
154    INSN_STRH_REG  = 0x000000b0,
155
156    INSN_LDRB_IMM  = 0x04500000,
157    INSN_LDRB_REG  = 0x06500000,
158    INSN_LDRSB_IMM = 0x005000d0,
159    INSN_LDRSB_REG = 0x001000d0,
160    INSN_STRB_IMM  = 0x04400000,
161    INSN_STRB_REG  = 0x06400000,
162
163    INSN_LDRD_IMM  = 0x004000d0,
164    INSN_LDRD_REG  = 0x000000d0,
165    INSN_STRD_IMM  = 0x004000f0,
166    INSN_STRD_REG  = 0x000000f0,
167
168    INSN_DMB_ISH   = 0xf57ff05b,
169    INSN_DMB_MCR   = 0xee070fba,
170
171    /* Architected nop introduced in v6k.  */
172    /* ??? This is an MSR (imm) 0,0,0 insn.  Anyone know if this
173       also Just So Happened to do nothing on pre-v6k so that we
174       don't need to conditionalize it?  */
175    INSN_NOP_v6k   = 0xe320f000,
176    /* Otherwise the assembler uses mov r0,r0 */
177    INSN_NOP_v4    = (COND_AL << 28) | ARITH_MOV,
178
179    INSN_VADD      = 0xf2000800,
180    INSN_VAND      = 0xf2000110,
181    INSN_VBIC      = 0xf2100110,
182    INSN_VEOR      = 0xf3000110,
183    INSN_VORN      = 0xf2300110,
184    INSN_VORR      = 0xf2200110,
185    INSN_VSUB      = 0xf3000800,
186    INSN_VMUL      = 0xf2000910,
187    INSN_VQADD     = 0xf2000010,
188    INSN_VQADD_U   = 0xf3000010,
189    INSN_VQSUB     = 0xf2000210,
190    INSN_VQSUB_U   = 0xf3000210,
191    INSN_VMAX      = 0xf2000600,
192    INSN_VMAX_U    = 0xf3000600,
193    INSN_VMIN      = 0xf2000610,
194    INSN_VMIN_U    = 0xf3000610,
195
196    INSN_VABS      = 0xf3b10300,
197    INSN_VMVN      = 0xf3b00580,
198    INSN_VNEG      = 0xf3b10380,
199
200    INSN_VCEQ0     = 0xf3b10100,
201    INSN_VCGT0     = 0xf3b10000,
202    INSN_VCGE0     = 0xf3b10080,
203    INSN_VCLE0     = 0xf3b10180,
204    INSN_VCLT0     = 0xf3b10200,
205
206    INSN_VCEQ      = 0xf3000810,
207    INSN_VCGE      = 0xf2000310,
208    INSN_VCGT      = 0xf2000300,
209    INSN_VCGE_U    = 0xf3000310,
210    INSN_VCGT_U    = 0xf3000300,
211
212    INSN_VSHLI     = 0xf2800510,  /* VSHL (immediate) */
213    INSN_VSARI     = 0xf2800010,  /* VSHR.S */
214    INSN_VSHRI     = 0xf3800010,  /* VSHR.U */
215    INSN_VSLI      = 0xf3800510,
216    INSN_VSHL_S    = 0xf2000400,  /* VSHL.S (register) */
217    INSN_VSHL_U    = 0xf3000400,  /* VSHL.U (register) */
218
219    INSN_VBSL      = 0xf3100110,
220    INSN_VBIT      = 0xf3200110,
221    INSN_VBIF      = 0xf3300110,
222
223    INSN_VTST      = 0xf2000810,
224
225    INSN_VDUP_G    = 0xee800b10,  /* VDUP (ARM core register) */
226    INSN_VDUP_S    = 0xf3b00c00,  /* VDUP (scalar) */
227    INSN_VLDR_D    = 0xed100b00,  /* VLDR.64 */
228    INSN_VLD1      = 0xf4200000,  /* VLD1 (multiple single elements) */
229    INSN_VLD1R     = 0xf4a00c00,  /* VLD1 (single element to all lanes) */
230    INSN_VST1      = 0xf4000000,  /* VST1 (multiple single elements) */
231    INSN_VMOVI     = 0xf2800010,  /* VMOV (immediate) */
232} ARMInsn;
233
234#define INSN_NOP   (use_armv7_instructions ? INSN_NOP_v6k : INSN_NOP_v4)
235
236static const uint8_t tcg_cond_to_arm_cond[] = {
237    [TCG_COND_EQ] = COND_EQ,
238    [TCG_COND_NE] = COND_NE,
239    [TCG_COND_LT] = COND_LT,
240    [TCG_COND_GE] = COND_GE,
241    [TCG_COND_LE] = COND_LE,
242    [TCG_COND_GT] = COND_GT,
243    /* unsigned */
244    [TCG_COND_LTU] = COND_CC,
245    [TCG_COND_GEU] = COND_CS,
246    [TCG_COND_LEU] = COND_LS,
247    [TCG_COND_GTU] = COND_HI,
248};
249
250static int encode_imm(uint32_t imm);
251
252/* TCG private relocation type: add with pc+imm8 */
253#define R_ARM_PC8  11
254
255/* TCG private relocation type: vldr with imm8 << 2 */
256#define R_ARM_PC11 12
257
258static bool reloc_pc24(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
259{
260    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
261    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) >> 2;
262
263    if (offset == sextract32(offset, 0, 24)) {
264        *src_rw = deposit32(*src_rw, 0, 24, offset);
265        return true;
266    }
267    return false;
268}
269
270static bool reloc_pc13(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
271{
272    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
273    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
274
275    if (offset >= -0xfff && offset <= 0xfff) {
276        tcg_insn_unit insn = *src_rw;
277        bool u = (offset >= 0);
278        if (!u) {
279            offset = -offset;
280        }
281        insn = deposit32(insn, 23, 1, u);
282        insn = deposit32(insn, 0, 12, offset);
283        *src_rw = insn;
284        return true;
285    }
286    return false;
287}
288
289static bool reloc_pc11(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
290{
291    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
292    ptrdiff_t offset = (tcg_ptr_byte_diff(target, src_rx) - 8) / 4;
293
294    if (offset >= -0xff && offset <= 0xff) {
295        tcg_insn_unit insn = *src_rw;
296        bool u = (offset >= 0);
297        if (!u) {
298            offset = -offset;
299        }
300        insn = deposit32(insn, 23, 1, u);
301        insn = deposit32(insn, 0, 8, offset);
302        *src_rw = insn;
303        return true;
304    }
305    return false;
306}
307
308static bool reloc_pc8(tcg_insn_unit *src_rw, const tcg_insn_unit *target)
309{
310    const tcg_insn_unit *src_rx = tcg_splitwx_to_rx(src_rw);
311    ptrdiff_t offset = tcg_ptr_byte_diff(target, src_rx) - 8;
312    int imm12 = encode_imm(offset);
313
314    if (imm12 >= 0) {
315        *src_rw = deposit32(*src_rw, 0, 12, imm12);
316        return true;
317    }
318    return false;
319}
320
321static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
322                        intptr_t value, intptr_t addend)
323{
324    tcg_debug_assert(addend == 0);
325    switch (type) {
326    case R_ARM_PC24:
327        return reloc_pc24(code_ptr, (const tcg_insn_unit *)value);
328    case R_ARM_PC13:
329        return reloc_pc13(code_ptr, (const tcg_insn_unit *)value);
330    case R_ARM_PC11:
331        return reloc_pc11(code_ptr, (const tcg_insn_unit *)value);
332    case R_ARM_PC8:
333        return reloc_pc8(code_ptr, (const tcg_insn_unit *)value);
334    default:
335        g_assert_not_reached();
336    }
337}
338
339#define TCG_CT_CONST_ARM  0x100
340#define TCG_CT_CONST_INV  0x200
341#define TCG_CT_CONST_NEG  0x400
342#define TCG_CT_CONST_ZERO 0x800
343#define TCG_CT_CONST_ORRI 0x1000
344#define TCG_CT_CONST_ANDI 0x2000
345
346#define ALL_GENERAL_REGS  0xffffu
347#define ALL_VECTOR_REGS   0xffff0000u
348
349/*
350 * r0-r2 will be overwritten when reading the tlb entry (softmmu only)
351 * and r0-r1 doing the byte swapping, so don't use these.
352 * r3 is removed for softmmu to avoid clashes with helper arguments.
353 */
354#ifdef CONFIG_SOFTMMU
355#define ALL_QLOAD_REGS \
356    (ALL_GENERAL_REGS & ~((1 << TCG_REG_R0) | (1 << TCG_REG_R1) | \
357                          (1 << TCG_REG_R2) | (1 << TCG_REG_R3) | \
358                          (1 << TCG_REG_R14)))
359#define ALL_QSTORE_REGS \
360    (ALL_GENERAL_REGS & ~((1 << TCG_REG_R0) | (1 << TCG_REG_R1) | \
361                          (1 << TCG_REG_R2) | (1 << TCG_REG_R14) | \
362                          ((TARGET_LONG_BITS == 64) << TCG_REG_R3)))
363#else
364#define ALL_QLOAD_REGS   ALL_GENERAL_REGS
365#define ALL_QSTORE_REGS \
366    (ALL_GENERAL_REGS & ~((1 << TCG_REG_R0) | (1 << TCG_REG_R1)))
367#endif
368
369/*
370 * ARM immediates for ALU instructions are made of an unsigned 8-bit
371 * right-rotated by an even amount between 0 and 30.
372 *
373 * Return < 0 if @imm cannot be encoded, else the entire imm12 field.
374 */
375static int encode_imm(uint32_t imm)
376{
377    uint32_t rot, imm8;
378
379    /* Simple case, no rotation required. */
380    if ((imm & ~0xff) == 0) {
381        return imm;
382    }
383
384    /* Next, try a simple even shift.  */
385    rot = ctz32(imm) & ~1;
386    imm8 = imm >> rot;
387    rot = 32 - rot;
388    if ((imm8 & ~0xff) == 0) {
389        goto found;
390    }
391
392    /*
393     * Finally, try harder with rotations.
394     * The ctz test above will have taken care of rotates >= 8.
395     */
396    for (rot = 2; rot < 8; rot += 2) {
397        imm8 = rol32(imm, rot);
398        if ((imm8 & ~0xff) == 0) {
399            goto found;
400        }
401    }
402    /* Fail: imm cannot be encoded. */
403    return -1;
404
405 found:
406    /* Note that rot is even, and we discard bit 0 by shifting by 7. */
407    return rot << 7 | imm8;
408}
409
410static int encode_imm_nofail(uint32_t imm)
411{
412    int ret = encode_imm(imm);
413    tcg_debug_assert(ret >= 0);
414    return ret;
415}
416
417static bool check_fit_imm(uint32_t imm)
418{
419    return encode_imm(imm) >= 0;
420}
421
422/* Return true if v16 is a valid 16-bit shifted immediate.  */
423static bool is_shimm16(uint16_t v16, int *cmode, int *imm8)
424{
425    if (v16 == (v16 & 0xff)) {
426        *cmode = 0x8;
427        *imm8 = v16 & 0xff;
428        return true;
429    } else if (v16 == (v16 & 0xff00)) {
430        *cmode = 0xa;
431        *imm8 = v16 >> 8;
432        return true;
433    }
434    return false;
435}
436
437/* Return true if v32 is a valid 32-bit shifted immediate.  */
438static bool is_shimm32(uint32_t v32, int *cmode, int *imm8)
439{
440    if (v32 == (v32 & 0xff)) {
441        *cmode = 0x0;
442        *imm8 = v32 & 0xff;
443        return true;
444    } else if (v32 == (v32 & 0xff00)) {
445        *cmode = 0x2;
446        *imm8 = (v32 >> 8) & 0xff;
447        return true;
448    } else if (v32 == (v32 & 0xff0000)) {
449        *cmode = 0x4;
450        *imm8 = (v32 >> 16) & 0xff;
451        return true;
452    } else if (v32 == (v32 & 0xff000000)) {
453        *cmode = 0x6;
454        *imm8 = v32 >> 24;
455        return true;
456    }
457    return false;
458}
459
460/* Return true if v32 is a valid 32-bit shifting ones immediate.  */
461static bool is_soimm32(uint32_t v32, int *cmode, int *imm8)
462{
463    if ((v32 & 0xffff00ff) == 0xff) {
464        *cmode = 0xc;
465        *imm8 = (v32 >> 8) & 0xff;
466        return true;
467    } else if ((v32 & 0xff00ffff) == 0xffff) {
468        *cmode = 0xd;
469        *imm8 = (v32 >> 16) & 0xff;
470        return true;
471    }
472    return false;
473}
474
475/*
476 * Return non-zero if v32 can be formed by MOVI+ORR.
477 * Place the parameters for MOVI in (cmode, imm8).
478 * Return the cmode for ORR; the imm8 can be had via extraction from v32.
479 */
480static int is_shimm32_pair(uint32_t v32, int *cmode, int *imm8)
481{
482    int i;
483
484    for (i = 6; i > 0; i -= 2) {
485        /* Mask out one byte we can add with ORR.  */
486        uint32_t tmp = v32 & ~(0xffu << (i * 4));
487        if (is_shimm32(tmp, cmode, imm8) ||
488            is_soimm32(tmp, cmode, imm8)) {
489            break;
490        }
491    }
492    return i;
493}
494
495/* Return true if V is a valid 16-bit or 32-bit shifted immediate.  */
496static bool is_shimm1632(uint32_t v32, int *cmode, int *imm8)
497{
498    if (v32 == deposit32(v32, 16, 16, v32)) {
499        return is_shimm16(v32, cmode, imm8);
500    } else {
501        return is_shimm32(v32, cmode, imm8);
502    }
503}
504
505/* Test if a constant matches the constraint.
506 * TODO: define constraints for:
507 *
508 * ldr/str offset:   between -0xfff and 0xfff
509 * ldrh/strh offset: between -0xff and 0xff
510 * mov operand2:     values represented with x << (2 * y), x < 0x100
511 * add, sub, eor...: ditto
512 */
513static bool tcg_target_const_match(int64_t val, TCGType type, int ct)
514{
515    if (ct & TCG_CT_CONST) {
516        return 1;
517    } else if ((ct & TCG_CT_CONST_ARM) && check_fit_imm(val)) {
518        return 1;
519    } else if ((ct & TCG_CT_CONST_INV) && check_fit_imm(~val)) {
520        return 1;
521    } else if ((ct & TCG_CT_CONST_NEG) && check_fit_imm(-val)) {
522        return 1;
523    } else if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
524        return 1;
525    }
526
527    switch (ct & (TCG_CT_CONST_ORRI | TCG_CT_CONST_ANDI)) {
528    case 0:
529        break;
530    case TCG_CT_CONST_ANDI:
531        val = ~val;
532        /* fallthru */
533    case TCG_CT_CONST_ORRI:
534        if (val == deposit64(val, 32, 32, val)) {
535            int cmode, imm8;
536            return is_shimm1632(val, &cmode, &imm8);
537        }
538        break;
539    default:
540        /* Both bits should not be set for the same insn.  */
541        g_assert_not_reached();
542    }
543
544    return 0;
545}
546
547static void tcg_out_b_imm(TCGContext *s, ARMCond cond, int32_t offset)
548{
549    tcg_out32(s, (cond << 28) | 0x0a000000 |
550                    (((offset - 8) >> 2) & 0x00ffffff));
551}
552
553static void tcg_out_bl_imm(TCGContext *s, ARMCond cond, int32_t offset)
554{
555    tcg_out32(s, (cond << 28) | 0x0b000000 |
556                    (((offset - 8) >> 2) & 0x00ffffff));
557}
558
559static void tcg_out_blx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
560{
561    tcg_out32(s, (cond << 28) | 0x012fff30 | rn);
562}
563
564static void tcg_out_blx_imm(TCGContext *s, int32_t offset)
565{
566    tcg_out32(s, 0xfa000000 | ((offset & 2) << 23) |
567                (((offset - 8) >> 2) & 0x00ffffff));
568}
569
570static void tcg_out_dat_reg(TCGContext *s, ARMCond cond, ARMInsn opc,
571                            TCGReg rd, TCGReg rn, TCGReg rm, int shift)
572{
573    tcg_out32(s, (cond << 28) | (0 << 25) | opc |
574                    (rn << 16) | (rd << 12) | shift | rm);
575}
576
577static void tcg_out_mov_reg(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rm)
578{
579    /* Simple reg-reg move, optimising out the 'do nothing' case */
580    if (rd != rm) {
581        tcg_out_dat_reg(s, cond, ARITH_MOV, rd, 0, rm, SHIFT_IMM_LSL(0));
582    }
583}
584
585static void tcg_out_bx_reg(TCGContext *s, ARMCond cond, TCGReg rn)
586{
587    tcg_out32(s, (cond << 28) | 0x012fff10 | rn);
588}
589
590static void tcg_out_b_reg(TCGContext *s, ARMCond cond, TCGReg rn)
591{
592    /*
593     * Unless the C portion of QEMU is compiled as thumb, we don't need
594     * true BX semantics; merely a branch to an address held in a register.
595     */
596    tcg_out_bx_reg(s, cond, rn);
597}
598
599static void tcg_out_dat_imm(TCGContext *s, ARMCond cond, ARMInsn opc,
600                            TCGReg rd, TCGReg rn, int im)
601{
602    tcg_out32(s, (cond << 28) | (1 << 25) | opc |
603                    (rn << 16) | (rd << 12) | im);
604}
605
606static void tcg_out_ldstm(TCGContext *s, ARMCond cond, ARMInsn opc,
607                          TCGReg rn, uint16_t mask)
608{
609    tcg_out32(s, (cond << 28) | opc | (rn << 16) | mask);
610}
611
612/* Note that this routine is used for both LDR and LDRH formats, so we do
613   not wish to include an immediate shift at this point.  */
614static void tcg_out_memop_r(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
615                            TCGReg rn, TCGReg rm, bool u, bool p, bool w)
616{
617    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24)
618              | (w << 21) | (rn << 16) | (rt << 12) | rm);
619}
620
621static void tcg_out_memop_8(TCGContext *s, ARMCond cond, ARMInsn opc, TCGReg rt,
622                            TCGReg rn, int imm8, bool p, bool w)
623{
624    bool u = 1;
625    if (imm8 < 0) {
626        imm8 = -imm8;
627        u = 0;
628    }
629    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
630              (rn << 16) | (rt << 12) | ((imm8 & 0xf0) << 4) | (imm8 & 0xf));
631}
632
633static void tcg_out_memop_12(TCGContext *s, ARMCond cond, ARMInsn opc,
634                             TCGReg rt, TCGReg rn, int imm12, bool p, bool w)
635{
636    bool u = 1;
637    if (imm12 < 0) {
638        imm12 = -imm12;
639        u = 0;
640    }
641    tcg_out32(s, (cond << 28) | opc | (u << 23) | (p << 24) | (w << 21) |
642              (rn << 16) | (rt << 12) | imm12);
643}
644
645static void tcg_out_ld32_12(TCGContext *s, ARMCond cond, TCGReg rt,
646                            TCGReg rn, int imm12)
647{
648    tcg_out_memop_12(s, cond, INSN_LDR_IMM, rt, rn, imm12, 1, 0);
649}
650
651static void tcg_out_st32_12(TCGContext *s, ARMCond cond, TCGReg rt,
652                            TCGReg rn, int imm12)
653{
654    tcg_out_memop_12(s, cond, INSN_STR_IMM, rt, rn, imm12, 1, 0);
655}
656
657static void tcg_out_ld32_r(TCGContext *s, ARMCond cond, TCGReg rt,
658                           TCGReg rn, TCGReg rm)
659{
660    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 0);
661}
662
663static void tcg_out_st32_r(TCGContext *s, ARMCond cond, TCGReg rt,
664                           TCGReg rn, TCGReg rm)
665{
666    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 0);
667}
668
669static void tcg_out_ldrd_8(TCGContext *s, ARMCond cond, TCGReg rt,
670                           TCGReg rn, int imm8)
671{
672    tcg_out_memop_8(s, cond, INSN_LDRD_IMM, rt, rn, imm8, 1, 0);
673}
674
675static void tcg_out_ldrd_r(TCGContext *s, ARMCond cond, TCGReg rt,
676                           TCGReg rn, TCGReg rm)
677{
678    tcg_out_memop_r(s, cond, INSN_LDRD_REG, rt, rn, rm, 1, 1, 0);
679}
680
681static void __attribute__((unused))
682tcg_out_ldrd_rwb(TCGContext *s, ARMCond cond, TCGReg rt, TCGReg rn, TCGReg rm)
683{
684    tcg_out_memop_r(s, cond, INSN_LDRD_REG, rt, rn, rm, 1, 1, 1);
685}
686
687static void tcg_out_strd_8(TCGContext *s, ARMCond cond, TCGReg rt,
688                           TCGReg rn, int imm8)
689{
690    tcg_out_memop_8(s, cond, INSN_STRD_IMM, rt, rn, imm8, 1, 0);
691}
692
693static void tcg_out_strd_r(TCGContext *s, ARMCond cond, TCGReg rt,
694                           TCGReg rn, TCGReg rm)
695{
696    tcg_out_memop_r(s, cond, INSN_STRD_REG, rt, rn, rm, 1, 1, 0);
697}
698
699/* Register pre-increment with base writeback.  */
700static void tcg_out_ld32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
701                             TCGReg rn, TCGReg rm)
702{
703    tcg_out_memop_r(s, cond, INSN_LDR_REG, rt, rn, rm, 1, 1, 1);
704}
705
706static void tcg_out_st32_rwb(TCGContext *s, ARMCond cond, TCGReg rt,
707                             TCGReg rn, TCGReg rm)
708{
709    tcg_out_memop_r(s, cond, INSN_STR_REG, rt, rn, rm, 1, 1, 1);
710}
711
712static void tcg_out_ld16u_8(TCGContext *s, ARMCond cond, TCGReg rt,
713                            TCGReg rn, int imm8)
714{
715    tcg_out_memop_8(s, cond, INSN_LDRH_IMM, rt, rn, imm8, 1, 0);
716}
717
718static void tcg_out_st16_8(TCGContext *s, ARMCond cond, TCGReg rt,
719                           TCGReg rn, int imm8)
720{
721    tcg_out_memop_8(s, cond, INSN_STRH_IMM, rt, rn, imm8, 1, 0);
722}
723
724static void tcg_out_ld16u_r(TCGContext *s, ARMCond cond, TCGReg rt,
725                            TCGReg rn, TCGReg rm)
726{
727    tcg_out_memop_r(s, cond, INSN_LDRH_REG, rt, rn, rm, 1, 1, 0);
728}
729
730static void tcg_out_st16_r(TCGContext *s, ARMCond cond, TCGReg rt,
731                           TCGReg rn, TCGReg rm)
732{
733    tcg_out_memop_r(s, cond, INSN_STRH_REG, rt, rn, rm, 1, 1, 0);
734}
735
736static void tcg_out_ld16s_8(TCGContext *s, ARMCond cond, TCGReg rt,
737                            TCGReg rn, int imm8)
738{
739    tcg_out_memop_8(s, cond, INSN_LDRSH_IMM, rt, rn, imm8, 1, 0);
740}
741
742static void tcg_out_ld16s_r(TCGContext *s, ARMCond cond, TCGReg rt,
743                            TCGReg rn, TCGReg rm)
744{
745    tcg_out_memop_r(s, cond, INSN_LDRSH_REG, rt, rn, rm, 1, 1, 0);
746}
747
748static void tcg_out_ld8_12(TCGContext *s, ARMCond cond, TCGReg rt,
749                           TCGReg rn, int imm12)
750{
751    tcg_out_memop_12(s, cond, INSN_LDRB_IMM, rt, rn, imm12, 1, 0);
752}
753
754static void tcg_out_st8_12(TCGContext *s, ARMCond cond, TCGReg rt,
755                           TCGReg rn, int imm12)
756{
757    tcg_out_memop_12(s, cond, INSN_STRB_IMM, rt, rn, imm12, 1, 0);
758}
759
760static void tcg_out_ld8_r(TCGContext *s, ARMCond cond, TCGReg rt,
761                          TCGReg rn, TCGReg rm)
762{
763    tcg_out_memop_r(s, cond, INSN_LDRB_REG, rt, rn, rm, 1, 1, 0);
764}
765
766static void tcg_out_st8_r(TCGContext *s, ARMCond cond, TCGReg rt,
767                          TCGReg rn, TCGReg rm)
768{
769    tcg_out_memop_r(s, cond, INSN_STRB_REG, rt, rn, rm, 1, 1, 0);
770}
771
772static void tcg_out_ld8s_8(TCGContext *s, ARMCond cond, TCGReg rt,
773                           TCGReg rn, int imm8)
774{
775    tcg_out_memop_8(s, cond, INSN_LDRSB_IMM, rt, rn, imm8, 1, 0);
776}
777
778static void tcg_out_ld8s_r(TCGContext *s, ARMCond cond, TCGReg rt,
779                           TCGReg rn, TCGReg rm)
780{
781    tcg_out_memop_r(s, cond, INSN_LDRSB_REG, rt, rn, rm, 1, 1, 0);
782}
783
784static void tcg_out_movi_pool(TCGContext *s, ARMCond cond,
785                              TCGReg rd, uint32_t arg)
786{
787    new_pool_label(s, arg, R_ARM_PC13, s->code_ptr, 0);
788    tcg_out_ld32_12(s, cond, rd, TCG_REG_PC, 0);
789}
790
791static void tcg_out_movi32(TCGContext *s, ARMCond cond,
792                           TCGReg rd, uint32_t arg)
793{
794    int imm12, diff, opc, sh1, sh2;
795    uint32_t tt0, tt1, tt2;
796
797    /* Check a single MOV/MVN before anything else.  */
798    imm12 = encode_imm(arg);
799    if (imm12 >= 0) {
800        tcg_out_dat_imm(s, cond, ARITH_MOV, rd, 0, imm12);
801        return;
802    }
803    imm12 = encode_imm(~arg);
804    if (imm12 >= 0) {
805        tcg_out_dat_imm(s, cond, ARITH_MVN, rd, 0, imm12);
806        return;
807    }
808
809    /* Check for a pc-relative address.  This will usually be the TB,
810       or within the TB, which is immediately before the code block.  */
811    diff = tcg_pcrel_diff(s, (void *)arg) - 8;
812    if (diff >= 0) {
813        imm12 = encode_imm(diff);
814        if (imm12 >= 0) {
815            tcg_out_dat_imm(s, cond, ARITH_ADD, rd, TCG_REG_PC, imm12);
816            return;
817        }
818    } else {
819        imm12 = encode_imm(-diff);
820        if (imm12 >= 0) {
821            tcg_out_dat_imm(s, cond, ARITH_SUB, rd, TCG_REG_PC, imm12);
822            return;
823        }
824    }
825
826    /* Use movw + movt.  */
827    if (use_armv7_instructions) {
828        /* movw */
829        tcg_out32(s, (cond << 28) | 0x03000000 | (rd << 12)
830                  | ((arg << 4) & 0x000f0000) | (arg & 0xfff));
831        if (arg & 0xffff0000) {
832            /* movt */
833            tcg_out32(s, (cond << 28) | 0x03400000 | (rd << 12)
834                      | ((arg >> 12) & 0x000f0000) | ((arg >> 16) & 0xfff));
835        }
836        return;
837    }
838
839    /* Look for sequences of two insns.  If we have lots of 1's, we can
840       shorten the sequence by beginning with mvn and then clearing
841       higher bits with eor.  */
842    tt0 = arg;
843    opc = ARITH_MOV;
844    if (ctpop32(arg) > 16) {
845        tt0 = ~arg;
846        opc = ARITH_MVN;
847    }
848    sh1 = ctz32(tt0) & ~1;
849    tt1 = tt0 & ~(0xff << sh1);
850    sh2 = ctz32(tt1) & ~1;
851    tt2 = tt1 & ~(0xff << sh2);
852    if (tt2 == 0) {
853        int rot;
854
855        rot = ((32 - sh1) << 7) & 0xf00;
856        tcg_out_dat_imm(s, cond, opc, rd,  0, ((tt0 >> sh1) & 0xff) | rot);
857        rot = ((32 - sh2) << 7) & 0xf00;
858        tcg_out_dat_imm(s, cond, ARITH_EOR, rd, rd,
859                        ((tt0 >> sh2) & 0xff) | rot);
860        return;
861    }
862
863    /* Otherwise, drop it into the constant pool.  */
864    tcg_out_movi_pool(s, cond, rd, arg);
865}
866
867/*
868 * Emit either the reg,imm or reg,reg form of a data-processing insn.
869 * rhs must satisfy the "rI" constraint.
870 */
871static void tcg_out_dat_rI(TCGContext *s, ARMCond cond, ARMInsn opc,
872                           TCGReg dst, TCGReg lhs, TCGArg rhs, int rhs_is_const)
873{
874    if (rhs_is_const) {
875        tcg_out_dat_imm(s, cond, opc, dst, lhs, encode_imm_nofail(rhs));
876    } else {
877        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
878    }
879}
880
881/*
882 * Emit either the reg,imm or reg,reg form of a data-processing insn.
883 * rhs must satisfy the "rIK" constraint.
884 */
885static void tcg_out_dat_rIK(TCGContext *s, ARMCond cond, ARMInsn opc,
886                            ARMInsn opinv, TCGReg dst, TCGReg lhs, TCGArg rhs,
887                            bool rhs_is_const)
888{
889    if (rhs_is_const) {
890        int imm12 = encode_imm(rhs);
891        if (imm12 < 0) {
892            imm12 = encode_imm_nofail(~rhs);
893            opc = opinv;
894        }
895        tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
896    } else {
897        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
898    }
899}
900
901static void tcg_out_dat_rIN(TCGContext *s, ARMCond cond, ARMInsn opc,
902                            ARMInsn opneg, TCGReg dst, TCGReg lhs, TCGArg rhs,
903                            bool rhs_is_const)
904{
905    /* Emit either the reg,imm or reg,reg form of a data-processing insn.
906     * rhs must satisfy the "rIN" constraint.
907     */
908    if (rhs_is_const) {
909        int imm12 = encode_imm(rhs);
910        if (imm12 < 0) {
911            imm12 = encode_imm_nofail(-rhs);
912            opc = opneg;
913        }
914        tcg_out_dat_imm(s, cond, opc, dst, lhs, imm12);
915    } else {
916        tcg_out_dat_reg(s, cond, opc, dst, lhs, rhs, SHIFT_IMM_LSL(0));
917    }
918}
919
920static void tcg_out_mul32(TCGContext *s, ARMCond cond, TCGReg rd,
921                          TCGReg rn, TCGReg rm)
922{
923    /* mul */
924    tcg_out32(s, (cond << 28) | 0x90 | (rd << 16) | (rm << 8) | rn);
925}
926
927static void tcg_out_umull32(TCGContext *s, ARMCond cond, TCGReg rd0,
928                            TCGReg rd1, TCGReg rn, TCGReg rm)
929{
930    /* umull */
931    tcg_out32(s, (cond << 28) | 0x00800090 |
932              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
933}
934
935static void tcg_out_smull32(TCGContext *s, ARMCond cond, TCGReg rd0,
936                            TCGReg rd1, TCGReg rn, TCGReg rm)
937{
938    /* smull */
939    tcg_out32(s, (cond << 28) | 0x00c00090 |
940              (rd1 << 16) | (rd0 << 12) | (rm << 8) | rn);
941}
942
943static void tcg_out_sdiv(TCGContext *s, ARMCond cond,
944                         TCGReg rd, TCGReg rn, TCGReg rm)
945{
946    tcg_out32(s, 0x0710f010 | (cond << 28) | (rd << 16) | rn | (rm << 8));
947}
948
949static void tcg_out_udiv(TCGContext *s, ARMCond cond,
950                         TCGReg rd, TCGReg rn, TCGReg rm)
951{
952    tcg_out32(s, 0x0730f010 | (cond << 28) | (rd << 16) | rn | (rm << 8));
953}
954
955static void tcg_out_ext8s(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rn)
956{
957    /* sxtb */
958    tcg_out32(s, 0x06af0070 | (cond << 28) | (rd << 12) | rn);
959}
960
961static void __attribute__((unused))
962tcg_out_ext8u(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rn)
963{
964    tcg_out_dat_imm(s, cond, ARITH_AND, rd, rn, 0xff);
965}
966
967static void tcg_out_ext16s(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rn)
968{
969    /* sxth */
970    tcg_out32(s, 0x06bf0070 | (cond << 28) | (rd << 12) | rn);
971}
972
973static void tcg_out_ext16u(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rn)
974{
975    /* uxth */
976    tcg_out32(s, 0x06ff0070 | (cond << 28) | (rd << 12) | rn);
977}
978
979static void tcg_out_bswap16(TCGContext *s, ARMCond cond,
980                            TCGReg rd, TCGReg rn, int flags)
981{
982    if (flags & TCG_BSWAP_OS) {
983        /* revsh */
984        tcg_out32(s, 0x06ff0fb0 | (cond << 28) | (rd << 12) | rn);
985        return;
986    }
987
988    /* rev16 */
989    tcg_out32(s, 0x06bf0fb0 | (cond << 28) | (rd << 12) | rn);
990    if ((flags & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
991        /* uxth */
992        tcg_out32(s, 0x06ff0070 | (cond << 28) | (rd << 12) | rd);
993    }
994}
995
996static void tcg_out_bswap32(TCGContext *s, ARMCond cond, TCGReg rd, TCGReg rn)
997{
998    /* rev */
999    tcg_out32(s, 0x06bf0f30 | (cond << 28) | (rd << 12) | rn);
1000}
1001
1002static void tcg_out_deposit(TCGContext *s, ARMCond cond, TCGReg rd,
1003                            TCGArg a1, int ofs, int len, bool const_a1)
1004{
1005    if (const_a1) {
1006        /* bfi becomes bfc with rn == 15.  */
1007        a1 = 15;
1008    }
1009    /* bfi/bfc */
1010    tcg_out32(s, 0x07c00010 | (cond << 28) | (rd << 12) | a1
1011              | (ofs << 7) | ((ofs + len - 1) << 16));
1012}
1013
1014static void tcg_out_extract(TCGContext *s, ARMCond cond, TCGReg rd,
1015                            TCGReg rn, int ofs, int len)
1016{
1017    /* ubfx */
1018    tcg_out32(s, 0x07e00050 | (cond << 28) | (rd << 12) | rn
1019              | (ofs << 7) | ((len - 1) << 16));
1020}
1021
1022static void tcg_out_sextract(TCGContext *s, ARMCond cond, TCGReg rd,
1023                             TCGReg rn, int ofs, int len)
1024{
1025    /* sbfx */
1026    tcg_out32(s, 0x07a00050 | (cond << 28) | (rd << 12) | rn
1027              | (ofs << 7) | ((len - 1) << 16));
1028}
1029
1030static void tcg_out_ld32u(TCGContext *s, ARMCond cond,
1031                          TCGReg rd, TCGReg rn, int32_t offset)
1032{
1033    if (offset > 0xfff || offset < -0xfff) {
1034        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1035        tcg_out_ld32_r(s, cond, rd, rn, TCG_REG_TMP);
1036    } else
1037        tcg_out_ld32_12(s, cond, rd, rn, offset);
1038}
1039
1040static void tcg_out_st32(TCGContext *s, ARMCond cond,
1041                         TCGReg rd, TCGReg rn, int32_t offset)
1042{
1043    if (offset > 0xfff || offset < -0xfff) {
1044        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1045        tcg_out_st32_r(s, cond, rd, rn, TCG_REG_TMP);
1046    } else
1047        tcg_out_st32_12(s, cond, rd, rn, offset);
1048}
1049
1050static void tcg_out_ld16u(TCGContext *s, ARMCond cond,
1051                          TCGReg rd, TCGReg rn, int32_t offset)
1052{
1053    if (offset > 0xff || offset < -0xff) {
1054        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1055        tcg_out_ld16u_r(s, cond, rd, rn, TCG_REG_TMP);
1056    } else
1057        tcg_out_ld16u_8(s, cond, rd, rn, offset);
1058}
1059
1060static void tcg_out_ld16s(TCGContext *s, ARMCond cond,
1061                          TCGReg rd, TCGReg rn, int32_t offset)
1062{
1063    if (offset > 0xff || offset < -0xff) {
1064        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1065        tcg_out_ld16s_r(s, cond, rd, rn, TCG_REG_TMP);
1066    } else
1067        tcg_out_ld16s_8(s, cond, rd, rn, offset);
1068}
1069
1070static void tcg_out_st16(TCGContext *s, ARMCond cond,
1071                         TCGReg rd, TCGReg rn, int32_t offset)
1072{
1073    if (offset > 0xff || offset < -0xff) {
1074        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1075        tcg_out_st16_r(s, cond, rd, rn, TCG_REG_TMP);
1076    } else
1077        tcg_out_st16_8(s, cond, rd, rn, offset);
1078}
1079
1080static void tcg_out_ld8u(TCGContext *s, ARMCond cond,
1081                         TCGReg rd, TCGReg rn, int32_t offset)
1082{
1083    if (offset > 0xfff || offset < -0xfff) {
1084        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1085        tcg_out_ld8_r(s, cond, rd, rn, TCG_REG_TMP);
1086    } else
1087        tcg_out_ld8_12(s, cond, rd, rn, offset);
1088}
1089
1090static void tcg_out_ld8s(TCGContext *s, ARMCond cond,
1091                         TCGReg rd, TCGReg rn, int32_t offset)
1092{
1093    if (offset > 0xff || offset < -0xff) {
1094        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1095        tcg_out_ld8s_r(s, cond, rd, rn, TCG_REG_TMP);
1096    } else
1097        tcg_out_ld8s_8(s, cond, rd, rn, offset);
1098}
1099
1100static void tcg_out_st8(TCGContext *s, ARMCond cond,
1101                        TCGReg rd, TCGReg rn, int32_t offset)
1102{
1103    if (offset > 0xfff || offset < -0xfff) {
1104        tcg_out_movi32(s, cond, TCG_REG_TMP, offset);
1105        tcg_out_st8_r(s, cond, rd, rn, TCG_REG_TMP);
1106    } else
1107        tcg_out_st8_12(s, cond, rd, rn, offset);
1108}
1109
1110/*
1111 * The _goto case is normally between TBs within the same code buffer, and
1112 * with the code buffer limited to 16MB we wouldn't need the long case.
1113 * But we also use it for the tail-call to the qemu_ld/st helpers, which does.
1114 */
1115static void tcg_out_goto(TCGContext *s, ARMCond cond, const tcg_insn_unit *addr)
1116{
1117    intptr_t addri = (intptr_t)addr;
1118    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1119    bool arm_mode = !(addri & 1);
1120
1121    if (arm_mode && disp - 8 < 0x01fffffd && disp - 8 > -0x01fffffd) {
1122        tcg_out_b_imm(s, cond, disp);
1123        return;
1124    }
1125
1126    /* LDR is interworking from v5t. */
1127    tcg_out_movi_pool(s, cond, TCG_REG_PC, addri);
1128}
1129
1130/*
1131 * The call case is mostly used for helpers - so it's not unreasonable
1132 * for them to be beyond branch range.
1133 */
1134static void tcg_out_call(TCGContext *s, const tcg_insn_unit *addr)
1135{
1136    intptr_t addri = (intptr_t)addr;
1137    ptrdiff_t disp = tcg_pcrel_diff(s, addr);
1138    bool arm_mode = !(addri & 1);
1139
1140    if (disp - 8 < 0x02000000 && disp - 8 >= -0x02000000) {
1141        if (arm_mode) {
1142            tcg_out_bl_imm(s, COND_AL, disp);
1143        } else {
1144            tcg_out_blx_imm(s, disp);
1145        }
1146        return;
1147    }
1148
1149    tcg_out_movi32(s, COND_AL, TCG_REG_TMP, addri);
1150    tcg_out_blx_reg(s, COND_AL, TCG_REG_TMP);
1151}
1152
1153static void tcg_out_goto_label(TCGContext *s, ARMCond cond, TCGLabel *l)
1154{
1155    if (l->has_value) {
1156        tcg_out_goto(s, cond, l->u.value_ptr);
1157    } else {
1158        tcg_out_reloc(s, s->code_ptr, R_ARM_PC24, l, 0);
1159        tcg_out_b_imm(s, cond, 0);
1160    }
1161}
1162
1163static void tcg_out_mb(TCGContext *s, TCGArg a0)
1164{
1165    if (use_armv7_instructions) {
1166        tcg_out32(s, INSN_DMB_ISH);
1167    } else {
1168        tcg_out32(s, INSN_DMB_MCR);
1169    }
1170}
1171
1172static TCGCond tcg_out_cmp2(TCGContext *s, const TCGArg *args,
1173                            const int *const_args)
1174{
1175    TCGReg al = args[0];
1176    TCGReg ah = args[1];
1177    TCGArg bl = args[2];
1178    TCGArg bh = args[3];
1179    TCGCond cond = args[4];
1180    int const_bl = const_args[2];
1181    int const_bh = const_args[3];
1182
1183    switch (cond) {
1184    case TCG_COND_EQ:
1185    case TCG_COND_NE:
1186    case TCG_COND_LTU:
1187    case TCG_COND_LEU:
1188    case TCG_COND_GTU:
1189    case TCG_COND_GEU:
1190        /* We perform a conditional comparision.  If the high half is
1191           equal, then overwrite the flags with the comparison of the
1192           low half.  The resulting flags cover the whole.  */
1193        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, ah, bh, const_bh);
1194        tcg_out_dat_rI(s, COND_EQ, ARITH_CMP, 0, al, bl, const_bl);
1195        return cond;
1196
1197    case TCG_COND_LT:
1198    case TCG_COND_GE:
1199        /* We perform a double-word subtraction and examine the result.
1200           We do not actually need the result of the subtract, so the
1201           low part "subtract" is a compare.  For the high half we have
1202           no choice but to compute into a temporary.  */
1203        tcg_out_dat_rI(s, COND_AL, ARITH_CMP, 0, al, bl, const_bl);
1204        tcg_out_dat_rI(s, COND_AL, ARITH_SBC | TO_CPSR,
1205                       TCG_REG_TMP, ah, bh, const_bh);
1206        return cond;
1207
1208    case TCG_COND_LE:
1209    case TCG_COND_GT:
1210        /* Similar, but with swapped arguments, via reversed subtract.  */
1211        tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR,
1212                       TCG_REG_TMP, al, bl, const_bl);
1213        tcg_out_dat_rI(s, COND_AL, ARITH_RSC | TO_CPSR,
1214                       TCG_REG_TMP, ah, bh, const_bh);
1215        return tcg_swap_cond(cond);
1216
1217    default:
1218        g_assert_not_reached();
1219    }
1220}
1221
1222/*
1223 * Note that TCGReg references Q-registers.
1224 * Q-regno = 2 * D-regno, so shift left by 1 whlie inserting.
1225 */
1226static uint32_t encode_vd(TCGReg rd)
1227{
1228    tcg_debug_assert(rd >= TCG_REG_Q0);
1229    return (extract32(rd, 3, 1) << 22) | (extract32(rd, 0, 3) << 13);
1230}
1231
1232static uint32_t encode_vn(TCGReg rn)
1233{
1234    tcg_debug_assert(rn >= TCG_REG_Q0);
1235    return (extract32(rn, 3, 1) << 7) | (extract32(rn, 0, 3) << 17);
1236}
1237
1238static uint32_t encode_vm(TCGReg rm)
1239{
1240    tcg_debug_assert(rm >= TCG_REG_Q0);
1241    return (extract32(rm, 3, 1) << 5) | (extract32(rm, 0, 3) << 1);
1242}
1243
1244static void tcg_out_vreg2(TCGContext *s, ARMInsn insn, int q, int vece,
1245                          TCGReg d, TCGReg m)
1246{
1247    tcg_out32(s, insn | (vece << 18) | (q << 6) |
1248              encode_vd(d) | encode_vm(m));
1249}
1250
1251static void tcg_out_vreg3(TCGContext *s, ARMInsn insn, int q, int vece,
1252                          TCGReg d, TCGReg n, TCGReg m)
1253{
1254    tcg_out32(s, insn | (vece << 20) | (q << 6) |
1255              encode_vd(d) | encode_vn(n) | encode_vm(m));
1256}
1257
1258static void tcg_out_vmovi(TCGContext *s, TCGReg rd,
1259                          int q, int op, int cmode, uint8_t imm8)
1260{
1261    tcg_out32(s, INSN_VMOVI | encode_vd(rd) | (q << 6) | (op << 5)
1262              | (cmode << 8) | extract32(imm8, 0, 4)
1263              | (extract32(imm8, 4, 3) << 16)
1264              | (extract32(imm8, 7, 1) << 24));
1265}
1266
1267static void tcg_out_vshifti(TCGContext *s, ARMInsn insn, int q,
1268                            TCGReg rd, TCGReg rm, int l_imm6)
1269{
1270    tcg_out32(s, insn | (q << 6) | encode_vd(rd) | encode_vm(rm) |
1271              (extract32(l_imm6, 6, 1) << 7) |
1272              (extract32(l_imm6, 0, 6) << 16));
1273}
1274
1275static void tcg_out_vldst(TCGContext *s, ARMInsn insn,
1276                          TCGReg rd, TCGReg rn, int offset)
1277{
1278    if (offset != 0) {
1279        if (check_fit_imm(offset) || check_fit_imm(-offset)) {
1280            tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
1281                            TCG_REG_TMP, rn, offset, true);
1282        } else {
1283            tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_TMP, offset);
1284            tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1285                            TCG_REG_TMP, TCG_REG_TMP, rn, 0);
1286        }
1287        rn = TCG_REG_TMP;
1288    }
1289    tcg_out32(s, insn | (rn << 16) | encode_vd(rd) | 0xf);
1290}
1291
1292#ifdef CONFIG_SOFTMMU
1293/* helper signature: helper_ret_ld_mmu(CPUState *env, target_ulong addr,
1294 *                                     int mmu_idx, uintptr_t ra)
1295 */
1296static void * const qemu_ld_helpers[MO_SSIZE + 1] = {
1297    [MO_UB]   = helper_ret_ldub_mmu,
1298    [MO_SB]   = helper_ret_ldsb_mmu,
1299#ifdef HOST_WORDS_BIGENDIAN
1300    [MO_UW] = helper_be_lduw_mmu,
1301    [MO_UL] = helper_be_ldul_mmu,
1302    [MO_UQ] = helper_be_ldq_mmu,
1303    [MO_SW] = helper_be_ldsw_mmu,
1304    [MO_SL] = helper_be_ldul_mmu,
1305#else
1306    [MO_UW] = helper_le_lduw_mmu,
1307    [MO_UL] = helper_le_ldul_mmu,
1308    [MO_UQ] = helper_le_ldq_mmu,
1309    [MO_SW] = helper_le_ldsw_mmu,
1310    [MO_SL] = helper_le_ldul_mmu,
1311#endif
1312};
1313
1314/* helper signature: helper_ret_st_mmu(CPUState *env, target_ulong addr,
1315 *                                     uintxx_t val, int mmu_idx, uintptr_t ra)
1316 */
1317static void * const qemu_st_helpers[MO_SIZE + 1] = {
1318    [MO_8]   = helper_ret_stb_mmu,
1319#ifdef HOST_WORDS_BIGENDIAN
1320    [MO_16] = helper_be_stw_mmu,
1321    [MO_32] = helper_be_stl_mmu,
1322    [MO_64] = helper_be_stq_mmu,
1323#else
1324    [MO_16] = helper_le_stw_mmu,
1325    [MO_32] = helper_le_stl_mmu,
1326    [MO_64] = helper_le_stq_mmu,
1327#endif
1328};
1329
1330/* Helper routines for marshalling helper function arguments into
1331 * the correct registers and stack.
1332 * argreg is where we want to put this argument, arg is the argument itself.
1333 * Return value is the updated argreg ready for the next call.
1334 * Note that argreg 0..3 is real registers, 4+ on stack.
1335 *
1336 * We provide routines for arguments which are: immediate, 32 bit
1337 * value in register, 16 and 8 bit values in register (which must be zero
1338 * extended before use) and 64 bit value in a lo:hi register pair.
1339 */
1340#define DEFINE_TCG_OUT_ARG(NAME, ARGTYPE, MOV_ARG, EXT_ARG)                \
1341static TCGReg NAME(TCGContext *s, TCGReg argreg, ARGTYPE arg)              \
1342{                                                                          \
1343    if (argreg < 4) {                                                      \
1344        MOV_ARG(s, COND_AL, argreg, arg);                                  \
1345    } else {                                                               \
1346        int ofs = (argreg - 4) * 4;                                        \
1347        EXT_ARG;                                                           \
1348        tcg_debug_assert(ofs + 4 <= TCG_STATIC_CALL_ARGS_SIZE);            \
1349        tcg_out_st32_12(s, COND_AL, arg, TCG_REG_CALL_STACK, ofs);         \
1350    }                                                                      \
1351    return argreg + 1;                                                     \
1352}
1353
1354DEFINE_TCG_OUT_ARG(tcg_out_arg_imm32, uint32_t, tcg_out_movi32,
1355    (tcg_out_movi32(s, COND_AL, TCG_REG_TMP, arg), arg = TCG_REG_TMP))
1356DEFINE_TCG_OUT_ARG(tcg_out_arg_reg8, TCGReg, tcg_out_ext8u,
1357    (tcg_out_ext8u(s, COND_AL, TCG_REG_TMP, arg), arg = TCG_REG_TMP))
1358DEFINE_TCG_OUT_ARG(tcg_out_arg_reg16, TCGReg, tcg_out_ext16u,
1359    (tcg_out_ext16u(s, COND_AL, TCG_REG_TMP, arg), arg = TCG_REG_TMP))
1360DEFINE_TCG_OUT_ARG(tcg_out_arg_reg32, TCGReg, tcg_out_mov_reg, )
1361
1362static TCGReg tcg_out_arg_reg64(TCGContext *s, TCGReg argreg,
1363                                TCGReg arglo, TCGReg arghi)
1364{
1365    /* 64 bit arguments must go in even/odd register pairs
1366     * and in 8-aligned stack slots.
1367     */
1368    if (argreg & 1) {
1369        argreg++;
1370    }
1371    if (argreg >= 4 && (arglo & 1) == 0 && arghi == arglo + 1) {
1372        tcg_out_strd_8(s, COND_AL, arglo,
1373                       TCG_REG_CALL_STACK, (argreg - 4) * 4);
1374        return argreg + 2;
1375    } else {
1376        argreg = tcg_out_arg_reg32(s, argreg, arglo);
1377        argreg = tcg_out_arg_reg32(s, argreg, arghi);
1378        return argreg;
1379    }
1380}
1381
1382#define TLB_SHIFT	(CPU_TLB_ENTRY_BITS + CPU_TLB_BITS)
1383
1384/* We expect to use an 9-bit sign-magnitude negative offset from ENV.  */
1385QEMU_BUILD_BUG_ON(TLB_MASK_TABLE_OFS(0) > 0);
1386QEMU_BUILD_BUG_ON(TLB_MASK_TABLE_OFS(0) < -256);
1387
1388/* These offsets are built into the LDRD below.  */
1389QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, mask) != 0);
1390QEMU_BUILD_BUG_ON(offsetof(CPUTLBDescFast, table) != 4);
1391
1392/* Load and compare a TLB entry, leaving the flags set.  Returns the register
1393   containing the addend of the tlb entry.  Clobbers R0, R1, R2, TMP.  */
1394
1395static TCGReg tcg_out_tlb_read(TCGContext *s, TCGReg addrlo, TCGReg addrhi,
1396                               MemOp opc, int mem_index, bool is_load)
1397{
1398    int cmp_off = (is_load ? offsetof(CPUTLBEntry, addr_read)
1399                   : offsetof(CPUTLBEntry, addr_write));
1400    int fast_off = TLB_MASK_TABLE_OFS(mem_index);
1401    unsigned s_mask = (1 << (opc & MO_SIZE)) - 1;
1402    unsigned a_mask = (1 << get_alignment_bits(opc)) - 1;
1403    TCGReg t_addr;
1404
1405    /* Load env_tlb(env)->f[mmu_idx].{mask,table} into {r0,r1}.  */
1406    tcg_out_ldrd_8(s, COND_AL, TCG_REG_R0, TCG_AREG0, fast_off);
1407
1408    /* Extract the tlb index from the address into R0.  */
1409    tcg_out_dat_reg(s, COND_AL, ARITH_AND, TCG_REG_R0, TCG_REG_R0, addrlo,
1410                    SHIFT_IMM_LSR(TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS));
1411
1412    /*
1413     * Add the tlb_table pointer, creating the CPUTLBEntry address in R1.
1414     * Load the tlb comparator into R2/R3 and the fast path addend into R1.
1415     */
1416    if (cmp_off == 0) {
1417        if (TARGET_LONG_BITS == 64) {
1418            tcg_out_ldrd_rwb(s, COND_AL, TCG_REG_R2, TCG_REG_R1, TCG_REG_R0);
1419        } else {
1420            tcg_out_ld32_rwb(s, COND_AL, TCG_REG_R2, TCG_REG_R1, TCG_REG_R0);
1421        }
1422    } else {
1423        tcg_out_dat_reg(s, COND_AL, ARITH_ADD,
1424                        TCG_REG_R1, TCG_REG_R1, TCG_REG_R0, 0);
1425        if (TARGET_LONG_BITS == 64) {
1426            tcg_out_ldrd_8(s, COND_AL, TCG_REG_R2, TCG_REG_R1, cmp_off);
1427        } else {
1428            tcg_out_ld32_12(s, COND_AL, TCG_REG_R2, TCG_REG_R1, cmp_off);
1429        }
1430    }
1431
1432    /* Load the tlb addend.  */
1433    tcg_out_ld32_12(s, COND_AL, TCG_REG_R1, TCG_REG_R1,
1434                    offsetof(CPUTLBEntry, addend));
1435
1436    /*
1437     * Check alignment, check comparators.
1438     * Do this in 2-4 insns.  Use MOVW for v7, if possible,
1439     * to reduce the number of sequential conditional instructions.
1440     * Almost all guests have at least 4k pages, which means that we need
1441     * to clear at least 9 bits even for an 8-byte memory, which means it
1442     * isn't worth checking for an immediate operand for BIC.
1443     *
1444     * For unaligned accesses, test the page of the last unit of alignment.
1445     * This leaves the least significant alignment bits unchanged, and of
1446     * course must be zero.
1447     */
1448    t_addr = addrlo;
1449    if (a_mask < s_mask) {
1450        t_addr = TCG_REG_R0;
1451        tcg_out_dat_imm(s, COND_AL, ARITH_ADD, t_addr,
1452                        addrlo, s_mask - a_mask);
1453    }
1454    if (use_armv7_instructions && TARGET_PAGE_BITS <= 16) {
1455        tcg_out_movi32(s, COND_AL, TCG_REG_TMP, ~(TARGET_PAGE_MASK | a_mask));
1456        tcg_out_dat_reg(s, COND_AL, ARITH_BIC, TCG_REG_TMP,
1457                        t_addr, TCG_REG_TMP, 0);
1458        tcg_out_dat_reg(s, COND_AL, ARITH_CMP, 0, TCG_REG_R2, TCG_REG_TMP, 0);
1459    } else {
1460        if (a_mask) {
1461            tcg_debug_assert(a_mask <= 0xff);
1462            tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addrlo, a_mask);
1463        }
1464        tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0, t_addr,
1465                        SHIFT_IMM_LSR(TARGET_PAGE_BITS));
1466        tcg_out_dat_reg(s, (a_mask ? COND_EQ : COND_AL), ARITH_CMP,
1467                        0, TCG_REG_R2, TCG_REG_TMP,
1468                        SHIFT_IMM_LSL(TARGET_PAGE_BITS));
1469    }
1470
1471    if (TARGET_LONG_BITS == 64) {
1472        tcg_out_dat_reg(s, COND_EQ, ARITH_CMP, 0, TCG_REG_R3, addrhi, 0);
1473    }
1474
1475    return TCG_REG_R1;
1476}
1477
1478/* Record the context of a call to the out of line helper code for the slow
1479   path for a load or store, so that we can later generate the correct
1480   helper code.  */
1481static void add_qemu_ldst_label(TCGContext *s, bool is_ld, MemOpIdx oi,
1482                                TCGReg datalo, TCGReg datahi, TCGReg addrlo,
1483                                TCGReg addrhi, tcg_insn_unit *raddr,
1484                                tcg_insn_unit *label_ptr)
1485{
1486    TCGLabelQemuLdst *label = new_ldst_label(s);
1487
1488    label->is_ld = is_ld;
1489    label->oi = oi;
1490    label->datalo_reg = datalo;
1491    label->datahi_reg = datahi;
1492    label->addrlo_reg = addrlo;
1493    label->addrhi_reg = addrhi;
1494    label->raddr = tcg_splitwx_to_rx(raddr);
1495    label->label_ptr[0] = label_ptr;
1496}
1497
1498static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1499{
1500    TCGReg argreg, datalo, datahi;
1501    MemOpIdx oi = lb->oi;
1502    MemOp opc = get_memop(oi);
1503
1504    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1505        return false;
1506    }
1507
1508    argreg = tcg_out_arg_reg32(s, TCG_REG_R0, TCG_AREG0);
1509    if (TARGET_LONG_BITS == 64) {
1510        argreg = tcg_out_arg_reg64(s, argreg, lb->addrlo_reg, lb->addrhi_reg);
1511    } else {
1512        argreg = tcg_out_arg_reg32(s, argreg, lb->addrlo_reg);
1513    }
1514    argreg = tcg_out_arg_imm32(s, argreg, oi);
1515    argreg = tcg_out_arg_reg32(s, argreg, TCG_REG_R14);
1516
1517    /* Use the canonical unsigned helpers and minimize icache usage. */
1518    tcg_out_call(s, qemu_ld_helpers[opc & MO_SIZE]);
1519
1520    datalo = lb->datalo_reg;
1521    datahi = lb->datahi_reg;
1522    switch (opc & MO_SSIZE) {
1523    case MO_SB:
1524        tcg_out_ext8s(s, COND_AL, datalo, TCG_REG_R0);
1525        break;
1526    case MO_SW:
1527        tcg_out_ext16s(s, COND_AL, datalo, TCG_REG_R0);
1528        break;
1529    default:
1530        tcg_out_mov_reg(s, COND_AL, datalo, TCG_REG_R0);
1531        break;
1532    case MO_UQ:
1533        if (datalo != TCG_REG_R1) {
1534            tcg_out_mov_reg(s, COND_AL, datalo, TCG_REG_R0);
1535            tcg_out_mov_reg(s, COND_AL, datahi, TCG_REG_R1);
1536        } else if (datahi != TCG_REG_R0) {
1537            tcg_out_mov_reg(s, COND_AL, datahi, TCG_REG_R1);
1538            tcg_out_mov_reg(s, COND_AL, datalo, TCG_REG_R0);
1539        } else {
1540            tcg_out_mov_reg(s, COND_AL, TCG_REG_TMP, TCG_REG_R0);
1541            tcg_out_mov_reg(s, COND_AL, datahi, TCG_REG_R1);
1542            tcg_out_mov_reg(s, COND_AL, datalo, TCG_REG_TMP);
1543        }
1544        break;
1545    }
1546
1547    tcg_out_goto(s, COND_AL, lb->raddr);
1548    return true;
1549}
1550
1551static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *lb)
1552{
1553    TCGReg argreg, datalo, datahi;
1554    MemOpIdx oi = lb->oi;
1555    MemOp opc = get_memop(oi);
1556
1557    if (!reloc_pc24(lb->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1558        return false;
1559    }
1560
1561    argreg = TCG_REG_R0;
1562    argreg = tcg_out_arg_reg32(s, argreg, TCG_AREG0);
1563    if (TARGET_LONG_BITS == 64) {
1564        argreg = tcg_out_arg_reg64(s, argreg, lb->addrlo_reg, lb->addrhi_reg);
1565    } else {
1566        argreg = tcg_out_arg_reg32(s, argreg, lb->addrlo_reg);
1567    }
1568
1569    datalo = lb->datalo_reg;
1570    datahi = lb->datahi_reg;
1571    switch (opc & MO_SIZE) {
1572    case MO_8:
1573        argreg = tcg_out_arg_reg8(s, argreg, datalo);
1574        break;
1575    case MO_16:
1576        argreg = tcg_out_arg_reg16(s, argreg, datalo);
1577        break;
1578    case MO_32:
1579    default:
1580        argreg = tcg_out_arg_reg32(s, argreg, datalo);
1581        break;
1582    case MO_64:
1583        argreg = tcg_out_arg_reg64(s, argreg, datalo, datahi);
1584        break;
1585    }
1586
1587    argreg = tcg_out_arg_imm32(s, argreg, oi);
1588    argreg = tcg_out_arg_reg32(s, argreg, TCG_REG_R14);
1589
1590    /* Tail-call to the helper, which will return to the fast path.  */
1591    tcg_out_goto(s, COND_AL, qemu_st_helpers[opc & MO_SIZE]);
1592    return true;
1593}
1594#else
1595
1596static void tcg_out_test_alignment(TCGContext *s, bool is_ld, TCGReg addrlo,
1597                                   TCGReg addrhi, unsigned a_bits)
1598{
1599    unsigned a_mask = (1 << a_bits) - 1;
1600    TCGLabelQemuLdst *label = new_ldst_label(s);
1601
1602    label->is_ld = is_ld;
1603    label->addrlo_reg = addrlo;
1604    label->addrhi_reg = addrhi;
1605
1606    /* We are expecting a_bits to max out at 7, and can easily support 8. */
1607    tcg_debug_assert(a_mask <= 0xff);
1608    /* tst addr, #mask */
1609    tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addrlo, a_mask);
1610
1611    /* blne slow_path */
1612    label->label_ptr[0] = s->code_ptr;
1613    tcg_out_bl_imm(s, COND_NE, 0);
1614
1615    label->raddr = tcg_splitwx_to_rx(s->code_ptr);
1616}
1617
1618static bool tcg_out_fail_alignment(TCGContext *s, TCGLabelQemuLdst *l)
1619{
1620    if (!reloc_pc24(l->label_ptr[0], tcg_splitwx_to_rx(s->code_ptr))) {
1621        return false;
1622    }
1623
1624    if (TARGET_LONG_BITS == 64) {
1625        /* 64-bit target address is aligned into R2:R3. */
1626        if (l->addrhi_reg != TCG_REG_R2) {
1627            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R2, l->addrlo_reg);
1628            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R3, l->addrhi_reg);
1629        } else if (l->addrlo_reg != TCG_REG_R3) {
1630            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R3, l->addrhi_reg);
1631            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R2, l->addrlo_reg);
1632        } else {
1633            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R1, TCG_REG_R2);
1634            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R2, TCG_REG_R3);
1635            tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R3, TCG_REG_R1);
1636        }
1637    } else {
1638        tcg_out_mov(s, TCG_TYPE_I32, TCG_REG_R1, l->addrlo_reg);
1639    }
1640    tcg_out_mov(s, TCG_TYPE_PTR, TCG_REG_R0, TCG_AREG0);
1641
1642    /*
1643     * Tail call to the helper, with the return address back inline,
1644     * just for the clarity of the debugging traceback -- the helper
1645     * cannot return.  We have used BLNE to arrive here, so LR is
1646     * already set.
1647     */
1648    tcg_out_goto(s, COND_AL, (const void *)
1649                 (l->is_ld ? helper_unaligned_ld : helper_unaligned_st));
1650    return true;
1651}
1652
1653static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
1654{
1655    return tcg_out_fail_alignment(s, l);
1656}
1657
1658static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
1659{
1660    return tcg_out_fail_alignment(s, l);
1661}
1662#endif /* SOFTMMU */
1663
1664static void tcg_out_qemu_ld_index(TCGContext *s, MemOp opc,
1665                                  TCGReg datalo, TCGReg datahi,
1666                                  TCGReg addrlo, TCGReg addend,
1667                                  bool scratch_addend)
1668{
1669    /* Byte swapping is left to middle-end expansion. */
1670    tcg_debug_assert((opc & MO_BSWAP) == 0);
1671
1672    switch (opc & MO_SSIZE) {
1673    case MO_UB:
1674        tcg_out_ld8_r(s, COND_AL, datalo, addrlo, addend);
1675        break;
1676    case MO_SB:
1677        tcg_out_ld8s_r(s, COND_AL, datalo, addrlo, addend);
1678        break;
1679    case MO_UW:
1680        tcg_out_ld16u_r(s, COND_AL, datalo, addrlo, addend);
1681        break;
1682    case MO_SW:
1683        tcg_out_ld16s_r(s, COND_AL, datalo, addrlo, addend);
1684        break;
1685    case MO_UL:
1686        tcg_out_ld32_r(s, COND_AL, datalo, addrlo, addend);
1687        break;
1688    case MO_UQ:
1689        /* LDRD requires alignment; double-check that. */
1690        if (get_alignment_bits(opc) >= MO_64
1691            && (datalo & 1) == 0 && datahi == datalo + 1) {
1692            /*
1693             * Rm (the second address op) must not overlap Rt or Rt + 1.
1694             * Since datalo is aligned, we can simplify the test via alignment.
1695             * Flip the two address arguments if that works.
1696             */
1697            if ((addend & ~1) != datalo) {
1698                tcg_out_ldrd_r(s, COND_AL, datalo, addrlo, addend);
1699                break;
1700            }
1701            if ((addrlo & ~1) != datalo) {
1702                tcg_out_ldrd_r(s, COND_AL, datalo, addend, addrlo);
1703                break;
1704            }
1705        }
1706        if (scratch_addend) {
1707            tcg_out_ld32_rwb(s, COND_AL, datalo, addend, addrlo);
1708            tcg_out_ld32_12(s, COND_AL, datahi, addend, 4);
1709        } else {
1710            tcg_out_dat_reg(s, COND_AL, ARITH_ADD, TCG_REG_TMP,
1711                            addend, addrlo, SHIFT_IMM_LSL(0));
1712            tcg_out_ld32_12(s, COND_AL, datalo, TCG_REG_TMP, 0);
1713            tcg_out_ld32_12(s, COND_AL, datahi, TCG_REG_TMP, 4);
1714        }
1715        break;
1716    default:
1717        g_assert_not_reached();
1718    }
1719}
1720
1721#ifndef CONFIG_SOFTMMU
1722static void tcg_out_qemu_ld_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1723                                   TCGReg datahi, TCGReg addrlo)
1724{
1725    /* Byte swapping is left to middle-end expansion. */
1726    tcg_debug_assert((opc & MO_BSWAP) == 0);
1727
1728    switch (opc & MO_SSIZE) {
1729    case MO_UB:
1730        tcg_out_ld8_12(s, COND_AL, datalo, addrlo, 0);
1731        break;
1732    case MO_SB:
1733        tcg_out_ld8s_8(s, COND_AL, datalo, addrlo, 0);
1734        break;
1735    case MO_UW:
1736        tcg_out_ld16u_8(s, COND_AL, datalo, addrlo, 0);
1737        break;
1738    case MO_SW:
1739        tcg_out_ld16s_8(s, COND_AL, datalo, addrlo, 0);
1740        break;
1741    case MO_UL:
1742        tcg_out_ld32_12(s, COND_AL, datalo, addrlo, 0);
1743        break;
1744    case MO_UQ:
1745        /* LDRD requires alignment; double-check that. */
1746        if (get_alignment_bits(opc) >= MO_64
1747            && (datalo & 1) == 0 && datahi == datalo + 1) {
1748            tcg_out_ldrd_8(s, COND_AL, datalo, addrlo, 0);
1749        } else if (datalo == addrlo) {
1750            tcg_out_ld32_12(s, COND_AL, datahi, addrlo, 4);
1751            tcg_out_ld32_12(s, COND_AL, datalo, addrlo, 0);
1752        } else {
1753            tcg_out_ld32_12(s, COND_AL, datalo, addrlo, 0);
1754            tcg_out_ld32_12(s, COND_AL, datahi, addrlo, 4);
1755        }
1756        break;
1757    default:
1758        g_assert_not_reached();
1759    }
1760}
1761#endif
1762
1763static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, bool is64)
1764{
1765    TCGReg addrlo, datalo, datahi, addrhi __attribute__((unused));
1766    MemOpIdx oi;
1767    MemOp opc;
1768#ifdef CONFIG_SOFTMMU
1769    int mem_index;
1770    TCGReg addend;
1771    tcg_insn_unit *label_ptr;
1772#else
1773    unsigned a_bits;
1774#endif
1775
1776    datalo = *args++;
1777    datahi = (is64 ? *args++ : 0);
1778    addrlo = *args++;
1779    addrhi = (TARGET_LONG_BITS == 64 ? *args++ : 0);
1780    oi = *args++;
1781    opc = get_memop(oi);
1782
1783#ifdef CONFIG_SOFTMMU
1784    mem_index = get_mmuidx(oi);
1785    addend = tcg_out_tlb_read(s, addrlo, addrhi, opc, mem_index, 1);
1786
1787    /* This a conditional BL only to load a pointer within this opcode into LR
1788       for the slow path.  We will not be using the value for a tail call.  */
1789    label_ptr = s->code_ptr;
1790    tcg_out_bl_imm(s, COND_NE, 0);
1791
1792    tcg_out_qemu_ld_index(s, opc, datalo, datahi, addrlo, addend, true);
1793
1794    add_qemu_ldst_label(s, true, oi, datalo, datahi, addrlo, addrhi,
1795                        s->code_ptr, label_ptr);
1796#else /* !CONFIG_SOFTMMU */
1797    a_bits = get_alignment_bits(opc);
1798    if (a_bits) {
1799        tcg_out_test_alignment(s, true, addrlo, addrhi, a_bits);
1800    }
1801    if (guest_base) {
1802        tcg_out_qemu_ld_index(s, opc, datalo, datahi,
1803                              addrlo, TCG_REG_GUEST_BASE, false);
1804    } else {
1805        tcg_out_qemu_ld_direct(s, opc, datalo, datahi, addrlo);
1806    }
1807#endif
1808}
1809
1810static void tcg_out_qemu_st_index(TCGContext *s, ARMCond cond, MemOp opc,
1811                                  TCGReg datalo, TCGReg datahi,
1812                                  TCGReg addrlo, TCGReg addend,
1813                                  bool scratch_addend)
1814{
1815    /* Byte swapping is left to middle-end expansion. */
1816    tcg_debug_assert((opc & MO_BSWAP) == 0);
1817
1818    switch (opc & MO_SIZE) {
1819    case MO_8:
1820        tcg_out_st8_r(s, cond, datalo, addrlo, addend);
1821        break;
1822    case MO_16:
1823        tcg_out_st16_r(s, cond, datalo, addrlo, addend);
1824        break;
1825    case MO_32:
1826        tcg_out_st32_r(s, cond, datalo, addrlo, addend);
1827        break;
1828    case MO_64:
1829        /* STRD requires alignment; double-check that. */
1830        if (get_alignment_bits(opc) >= MO_64
1831            && (datalo & 1) == 0 && datahi == datalo + 1) {
1832            tcg_out_strd_r(s, cond, datalo, addrlo, addend);
1833        } else if (scratch_addend) {
1834            tcg_out_st32_rwb(s, cond, datalo, addend, addrlo);
1835            tcg_out_st32_12(s, cond, datahi, addend, 4);
1836        } else {
1837            tcg_out_dat_reg(s, cond, ARITH_ADD, TCG_REG_TMP,
1838                            addend, addrlo, SHIFT_IMM_LSL(0));
1839            tcg_out_st32_12(s, cond, datalo, TCG_REG_TMP, 0);
1840            tcg_out_st32_12(s, cond, datahi, TCG_REG_TMP, 4);
1841        }
1842        break;
1843    default:
1844        g_assert_not_reached();
1845    }
1846}
1847
1848#ifndef CONFIG_SOFTMMU
1849static void tcg_out_qemu_st_direct(TCGContext *s, MemOp opc, TCGReg datalo,
1850                                   TCGReg datahi, TCGReg addrlo)
1851{
1852    /* Byte swapping is left to middle-end expansion. */
1853    tcg_debug_assert((opc & MO_BSWAP) == 0);
1854
1855    switch (opc & MO_SIZE) {
1856    case MO_8:
1857        tcg_out_st8_12(s, COND_AL, datalo, addrlo, 0);
1858        break;
1859    case MO_16:
1860        tcg_out_st16_8(s, COND_AL, datalo, addrlo, 0);
1861        break;
1862    case MO_32:
1863        tcg_out_st32_12(s, COND_AL, datalo, addrlo, 0);
1864        break;
1865    case MO_64:
1866        /* STRD requires alignment; double-check that. */
1867        if (get_alignment_bits(opc) >= MO_64
1868            && (datalo & 1) == 0 && datahi == datalo + 1) {
1869            tcg_out_strd_8(s, COND_AL, datalo, addrlo, 0);
1870        } else {
1871            tcg_out_st32_12(s, COND_AL, datalo, addrlo, 0);
1872            tcg_out_st32_12(s, COND_AL, datahi, addrlo, 4);
1873        }
1874        break;
1875    default:
1876        g_assert_not_reached();
1877    }
1878}
1879#endif
1880
1881static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is64)
1882{
1883    TCGReg addrlo, datalo, datahi, addrhi __attribute__((unused));
1884    MemOpIdx oi;
1885    MemOp opc;
1886#ifdef CONFIG_SOFTMMU
1887    int mem_index;
1888    TCGReg addend;
1889    tcg_insn_unit *label_ptr;
1890#else
1891    unsigned a_bits;
1892#endif
1893
1894    datalo = *args++;
1895    datahi = (is64 ? *args++ : 0);
1896    addrlo = *args++;
1897    addrhi = (TARGET_LONG_BITS == 64 ? *args++ : 0);
1898    oi = *args++;
1899    opc = get_memop(oi);
1900
1901#ifdef CONFIG_SOFTMMU
1902    mem_index = get_mmuidx(oi);
1903    addend = tcg_out_tlb_read(s, addrlo, addrhi, opc, mem_index, 0);
1904
1905    tcg_out_qemu_st_index(s, COND_EQ, opc, datalo, datahi,
1906                          addrlo, addend, true);
1907
1908    /* The conditional call must come last, as we're going to return here.  */
1909    label_ptr = s->code_ptr;
1910    tcg_out_bl_imm(s, COND_NE, 0);
1911
1912    add_qemu_ldst_label(s, false, oi, datalo, datahi, addrlo, addrhi,
1913                        s->code_ptr, label_ptr);
1914#else /* !CONFIG_SOFTMMU */
1915    a_bits = get_alignment_bits(opc);
1916    if (a_bits) {
1917        tcg_out_test_alignment(s, false, addrlo, addrhi, a_bits);
1918    }
1919    if (guest_base) {
1920        tcg_out_qemu_st_index(s, COND_AL, opc, datalo, datahi,
1921                              addrlo, TCG_REG_GUEST_BASE, false);
1922    } else {
1923        tcg_out_qemu_st_direct(s, opc, datalo, datahi, addrlo);
1924    }
1925#endif
1926}
1927
1928static void tcg_out_epilogue(TCGContext *s);
1929
1930static void tcg_out_op(TCGContext *s, TCGOpcode opc,
1931                       const TCGArg args[TCG_MAX_OP_ARGS],
1932                       const int const_args[TCG_MAX_OP_ARGS])
1933{
1934    TCGArg a0, a1, a2, a3, a4, a5;
1935    int c;
1936
1937    switch (opc) {
1938    case INDEX_op_exit_tb:
1939        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, args[0]);
1940        tcg_out_epilogue(s);
1941        break;
1942    case INDEX_op_goto_tb:
1943        {
1944            /* Indirect jump method */
1945            intptr_t ptr, dif, dil;
1946            TCGReg base = TCG_REG_PC;
1947
1948            tcg_debug_assert(s->tb_jmp_insn_offset == 0);
1949            ptr = (intptr_t)tcg_splitwx_to_rx(s->tb_jmp_target_addr + args[0]);
1950            dif = tcg_pcrel_diff(s, (void *)ptr) - 8;
1951            dil = sextract32(dif, 0, 12);
1952            if (dif != dil) {
1953                /* The TB is close, but outside the 12 bits addressable by
1954                   the load.  We can extend this to 20 bits with a sub of a
1955                   shifted immediate from pc.  In the vastly unlikely event
1956                   the code requires more than 1MB, we'll use 2 insns and
1957                   be no worse off.  */
1958                base = TCG_REG_R0;
1959                tcg_out_movi32(s, COND_AL, base, ptr - dil);
1960            }
1961            tcg_out_ld32_12(s, COND_AL, TCG_REG_PC, base, dil);
1962            set_jmp_reset_offset(s, args[0]);
1963        }
1964        break;
1965    case INDEX_op_goto_ptr:
1966        tcg_out_b_reg(s, COND_AL, args[0]);
1967        break;
1968    case INDEX_op_br:
1969        tcg_out_goto_label(s, COND_AL, arg_label(args[0]));
1970        break;
1971
1972    case INDEX_op_ld8u_i32:
1973        tcg_out_ld8u(s, COND_AL, args[0], args[1], args[2]);
1974        break;
1975    case INDEX_op_ld8s_i32:
1976        tcg_out_ld8s(s, COND_AL, args[0], args[1], args[2]);
1977        break;
1978    case INDEX_op_ld16u_i32:
1979        tcg_out_ld16u(s, COND_AL, args[0], args[1], args[2]);
1980        break;
1981    case INDEX_op_ld16s_i32:
1982        tcg_out_ld16s(s, COND_AL, args[0], args[1], args[2]);
1983        break;
1984    case INDEX_op_ld_i32:
1985        tcg_out_ld32u(s, COND_AL, args[0], args[1], args[2]);
1986        break;
1987    case INDEX_op_st8_i32:
1988        tcg_out_st8(s, COND_AL, args[0], args[1], args[2]);
1989        break;
1990    case INDEX_op_st16_i32:
1991        tcg_out_st16(s, COND_AL, args[0], args[1], args[2]);
1992        break;
1993    case INDEX_op_st_i32:
1994        tcg_out_st32(s, COND_AL, args[0], args[1], args[2]);
1995        break;
1996
1997    case INDEX_op_movcond_i32:
1998        /* Constraints mean that v2 is always in the same register as dest,
1999         * so we only need to do "if condition passed, move v1 to dest".
2000         */
2001        tcg_out_dat_rIN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0,
2002                        args[1], args[2], const_args[2]);
2003        tcg_out_dat_rIK(s, tcg_cond_to_arm_cond[args[5]], ARITH_MOV,
2004                        ARITH_MVN, args[0], 0, args[3], const_args[3]);
2005        break;
2006    case INDEX_op_add_i32:
2007        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD, ARITH_SUB,
2008                        args[0], args[1], args[2], const_args[2]);
2009        break;
2010    case INDEX_op_sub_i32:
2011        if (const_args[1]) {
2012            if (const_args[2]) {
2013                tcg_out_movi32(s, COND_AL, args[0], args[1] - args[2]);
2014            } else {
2015                tcg_out_dat_rI(s, COND_AL, ARITH_RSB,
2016                               args[0], args[2], args[1], 1);
2017            }
2018        } else {
2019            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB, ARITH_ADD,
2020                            args[0], args[1], args[2], const_args[2]);
2021        }
2022        break;
2023    case INDEX_op_and_i32:
2024        tcg_out_dat_rIK(s, COND_AL, ARITH_AND, ARITH_BIC,
2025                        args[0], args[1], args[2], const_args[2]);
2026        break;
2027    case INDEX_op_andc_i32:
2028        tcg_out_dat_rIK(s, COND_AL, ARITH_BIC, ARITH_AND,
2029                        args[0], args[1], args[2], const_args[2]);
2030        break;
2031    case INDEX_op_or_i32:
2032        c = ARITH_ORR;
2033        goto gen_arith;
2034    case INDEX_op_xor_i32:
2035        c = ARITH_EOR;
2036        /* Fall through.  */
2037    gen_arith:
2038        tcg_out_dat_rI(s, COND_AL, c, args[0], args[1], args[2], const_args[2]);
2039        break;
2040    case INDEX_op_add2_i32:
2041        a0 = args[0], a1 = args[1], a2 = args[2];
2042        a3 = args[3], a4 = args[4], a5 = args[5];
2043        if (a0 == a3 || (a0 == a5 && !const_args[5])) {
2044            a0 = TCG_REG_TMP;
2045        }
2046        tcg_out_dat_rIN(s, COND_AL, ARITH_ADD | TO_CPSR, ARITH_SUB | TO_CPSR,
2047                        a0, a2, a4, const_args[4]);
2048        tcg_out_dat_rIK(s, COND_AL, ARITH_ADC, ARITH_SBC,
2049                        a1, a3, a5, const_args[5]);
2050        tcg_out_mov_reg(s, COND_AL, args[0], a0);
2051        break;
2052    case INDEX_op_sub2_i32:
2053        a0 = args[0], a1 = args[1], a2 = args[2];
2054        a3 = args[3], a4 = args[4], a5 = args[5];
2055        if ((a0 == a3 && !const_args[3]) || (a0 == a5 && !const_args[5])) {
2056            a0 = TCG_REG_TMP;
2057        }
2058        if (const_args[2]) {
2059            if (const_args[4]) {
2060                tcg_out_movi32(s, COND_AL, a0, a4);
2061                a4 = a0;
2062            }
2063            tcg_out_dat_rI(s, COND_AL, ARITH_RSB | TO_CPSR, a0, a4, a2, 1);
2064        } else {
2065            tcg_out_dat_rIN(s, COND_AL, ARITH_SUB | TO_CPSR,
2066                            ARITH_ADD | TO_CPSR, a0, a2, a4, const_args[4]);
2067        }
2068        if (const_args[3]) {
2069            if (const_args[5]) {
2070                tcg_out_movi32(s, COND_AL, a1, a5);
2071                a5 = a1;
2072            }
2073            tcg_out_dat_rI(s, COND_AL, ARITH_RSC, a1, a5, a3, 1);
2074        } else {
2075            tcg_out_dat_rIK(s, COND_AL, ARITH_SBC, ARITH_ADC,
2076                            a1, a3, a5, const_args[5]);
2077        }
2078        tcg_out_mov_reg(s, COND_AL, args[0], a0);
2079        break;
2080    case INDEX_op_neg_i32:
2081        tcg_out_dat_imm(s, COND_AL, ARITH_RSB, args[0], args[1], 0);
2082        break;
2083    case INDEX_op_not_i32:
2084        tcg_out_dat_reg(s, COND_AL,
2085                        ARITH_MVN, args[0], 0, args[1], SHIFT_IMM_LSL(0));
2086        break;
2087    case INDEX_op_mul_i32:
2088        tcg_out_mul32(s, COND_AL, args[0], args[1], args[2]);
2089        break;
2090    case INDEX_op_mulu2_i32:
2091        tcg_out_umull32(s, COND_AL, args[0], args[1], args[2], args[3]);
2092        break;
2093    case INDEX_op_muls2_i32:
2094        tcg_out_smull32(s, COND_AL, args[0], args[1], args[2], args[3]);
2095        break;
2096    /* XXX: Perhaps args[2] & 0x1f is wrong */
2097    case INDEX_op_shl_i32:
2098        c = const_args[2] ?
2099                SHIFT_IMM_LSL(args[2] & 0x1f) : SHIFT_REG_LSL(args[2]);
2100        goto gen_shift32;
2101    case INDEX_op_shr_i32:
2102        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_LSR(args[2] & 0x1f) :
2103                SHIFT_IMM_LSL(0) : SHIFT_REG_LSR(args[2]);
2104        goto gen_shift32;
2105    case INDEX_op_sar_i32:
2106        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ASR(args[2] & 0x1f) :
2107                SHIFT_IMM_LSL(0) : SHIFT_REG_ASR(args[2]);
2108        goto gen_shift32;
2109    case INDEX_op_rotr_i32:
2110        c = const_args[2] ? (args[2] & 0x1f) ? SHIFT_IMM_ROR(args[2] & 0x1f) :
2111                SHIFT_IMM_LSL(0) : SHIFT_REG_ROR(args[2]);
2112        /* Fall through.  */
2113    gen_shift32:
2114        tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1], c);
2115        break;
2116
2117    case INDEX_op_rotl_i32:
2118        if (const_args[2]) {
2119            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
2120                            ((0x20 - args[2]) & 0x1f) ?
2121                            SHIFT_IMM_ROR((0x20 - args[2]) & 0x1f) :
2122                            SHIFT_IMM_LSL(0));
2123        } else {
2124            tcg_out_dat_imm(s, COND_AL, ARITH_RSB, TCG_REG_TMP, args[2], 0x20);
2125            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0, args[1],
2126                            SHIFT_REG_ROR(TCG_REG_TMP));
2127        }
2128        break;
2129
2130    case INDEX_op_ctz_i32:
2131        tcg_out_dat_reg(s, COND_AL, INSN_RBIT, TCG_REG_TMP, 0, args[1], 0);
2132        a1 = TCG_REG_TMP;
2133        goto do_clz;
2134
2135    case INDEX_op_clz_i32:
2136        a1 = args[1];
2137    do_clz:
2138        a0 = args[0];
2139        a2 = args[2];
2140        c = const_args[2];
2141        if (c && a2 == 32) {
2142            tcg_out_dat_reg(s, COND_AL, INSN_CLZ, a0, 0, a1, 0);
2143            break;
2144        }
2145        tcg_out_dat_imm(s, COND_AL, ARITH_CMP, 0, a1, 0);
2146        tcg_out_dat_reg(s, COND_NE, INSN_CLZ, a0, 0, a1, 0);
2147        if (c || a0 != a2) {
2148            tcg_out_dat_rIK(s, COND_EQ, ARITH_MOV, ARITH_MVN, a0, 0, a2, c);
2149        }
2150        break;
2151
2152    case INDEX_op_brcond_i32:
2153        tcg_out_dat_rIN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0,
2154                       args[0], args[1], const_args[1]);
2155        tcg_out_goto_label(s, tcg_cond_to_arm_cond[args[2]],
2156                           arg_label(args[3]));
2157        break;
2158    case INDEX_op_setcond_i32:
2159        tcg_out_dat_rIN(s, COND_AL, ARITH_CMP, ARITH_CMN, 0,
2160                        args[1], args[2], const_args[2]);
2161        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[args[3]],
2162                        ARITH_MOV, args[0], 0, 1);
2163        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(args[3])],
2164                        ARITH_MOV, args[0], 0, 0);
2165        break;
2166
2167    case INDEX_op_brcond2_i32:
2168        c = tcg_out_cmp2(s, args, const_args);
2169        tcg_out_goto_label(s, tcg_cond_to_arm_cond[c], arg_label(args[5]));
2170        break;
2171    case INDEX_op_setcond2_i32:
2172        c = tcg_out_cmp2(s, args + 1, const_args + 1);
2173        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[c], ARITH_MOV, args[0], 0, 1);
2174        tcg_out_dat_imm(s, tcg_cond_to_arm_cond[tcg_invert_cond(c)],
2175                        ARITH_MOV, args[0], 0, 0);
2176        break;
2177
2178    case INDEX_op_qemu_ld_i32:
2179        tcg_out_qemu_ld(s, args, 0);
2180        break;
2181    case INDEX_op_qemu_ld_i64:
2182        tcg_out_qemu_ld(s, args, 1);
2183        break;
2184    case INDEX_op_qemu_st_i32:
2185        tcg_out_qemu_st(s, args, 0);
2186        break;
2187    case INDEX_op_qemu_st_i64:
2188        tcg_out_qemu_st(s, args, 1);
2189        break;
2190
2191    case INDEX_op_bswap16_i32:
2192        tcg_out_bswap16(s, COND_AL, args[0], args[1], args[2]);
2193        break;
2194    case INDEX_op_bswap32_i32:
2195        tcg_out_bswap32(s, COND_AL, args[0], args[1]);
2196        break;
2197
2198    case INDEX_op_ext8s_i32:
2199        tcg_out_ext8s(s, COND_AL, args[0], args[1]);
2200        break;
2201    case INDEX_op_ext16s_i32:
2202        tcg_out_ext16s(s, COND_AL, args[0], args[1]);
2203        break;
2204    case INDEX_op_ext16u_i32:
2205        tcg_out_ext16u(s, COND_AL, args[0], args[1]);
2206        break;
2207
2208    case INDEX_op_deposit_i32:
2209        tcg_out_deposit(s, COND_AL, args[0], args[2],
2210                        args[3], args[4], const_args[2]);
2211        break;
2212    case INDEX_op_extract_i32:
2213        tcg_out_extract(s, COND_AL, args[0], args[1], args[2], args[3]);
2214        break;
2215    case INDEX_op_sextract_i32:
2216        tcg_out_sextract(s, COND_AL, args[0], args[1], args[2], args[3]);
2217        break;
2218    case INDEX_op_extract2_i32:
2219        /* ??? These optimization vs zero should be generic.  */
2220        /* ??? But we can't substitute 2 for 1 in the opcode stream yet.  */
2221        if (const_args[1]) {
2222            if (const_args[2]) {
2223                tcg_out_movi(s, TCG_TYPE_REG, args[0], 0);
2224            } else {
2225                tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2226                                args[2], SHIFT_IMM_LSL(32 - args[3]));
2227            }
2228        } else if (const_args[2]) {
2229            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, args[0], 0,
2230                            args[1], SHIFT_IMM_LSR(args[3]));
2231        } else {
2232            /* We can do extract2 in 2 insns, vs the 3 required otherwise.  */
2233            tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, 0,
2234                            args[2], SHIFT_IMM_LSL(32 - args[3]));
2235            tcg_out_dat_reg(s, COND_AL, ARITH_ORR, args[0], TCG_REG_TMP,
2236                            args[1], SHIFT_IMM_LSR(args[3]));
2237        }
2238        break;
2239
2240    case INDEX_op_div_i32:
2241        tcg_out_sdiv(s, COND_AL, args[0], args[1], args[2]);
2242        break;
2243    case INDEX_op_divu_i32:
2244        tcg_out_udiv(s, COND_AL, args[0], args[1], args[2]);
2245        break;
2246
2247    case INDEX_op_mb:
2248        tcg_out_mb(s, args[0]);
2249        break;
2250
2251    case INDEX_op_mov_i32:  /* Always emitted via tcg_out_mov.  */
2252    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
2253    default:
2254        tcg_abort();
2255    }
2256}
2257
2258static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
2259{
2260    switch (op) {
2261    case INDEX_op_goto_ptr:
2262        return C_O0_I1(r);
2263
2264    case INDEX_op_ld8u_i32:
2265    case INDEX_op_ld8s_i32:
2266    case INDEX_op_ld16u_i32:
2267    case INDEX_op_ld16s_i32:
2268    case INDEX_op_ld_i32:
2269    case INDEX_op_neg_i32:
2270    case INDEX_op_not_i32:
2271    case INDEX_op_bswap16_i32:
2272    case INDEX_op_bswap32_i32:
2273    case INDEX_op_ext8s_i32:
2274    case INDEX_op_ext16s_i32:
2275    case INDEX_op_ext16u_i32:
2276    case INDEX_op_extract_i32:
2277    case INDEX_op_sextract_i32:
2278        return C_O1_I1(r, r);
2279
2280    case INDEX_op_st8_i32:
2281    case INDEX_op_st16_i32:
2282    case INDEX_op_st_i32:
2283        return C_O0_I2(r, r);
2284
2285    case INDEX_op_add_i32:
2286    case INDEX_op_sub_i32:
2287    case INDEX_op_setcond_i32:
2288        return C_O1_I2(r, r, rIN);
2289
2290    case INDEX_op_and_i32:
2291    case INDEX_op_andc_i32:
2292    case INDEX_op_clz_i32:
2293    case INDEX_op_ctz_i32:
2294        return C_O1_I2(r, r, rIK);
2295
2296    case INDEX_op_mul_i32:
2297    case INDEX_op_div_i32:
2298    case INDEX_op_divu_i32:
2299        return C_O1_I2(r, r, r);
2300
2301    case INDEX_op_mulu2_i32:
2302    case INDEX_op_muls2_i32:
2303        return C_O2_I2(r, r, r, r);
2304
2305    case INDEX_op_or_i32:
2306    case INDEX_op_xor_i32:
2307        return C_O1_I2(r, r, rI);
2308
2309    case INDEX_op_shl_i32:
2310    case INDEX_op_shr_i32:
2311    case INDEX_op_sar_i32:
2312    case INDEX_op_rotl_i32:
2313    case INDEX_op_rotr_i32:
2314        return C_O1_I2(r, r, ri);
2315
2316    case INDEX_op_brcond_i32:
2317        return C_O0_I2(r, rIN);
2318    case INDEX_op_deposit_i32:
2319        return C_O1_I2(r, 0, rZ);
2320    case INDEX_op_extract2_i32:
2321        return C_O1_I2(r, rZ, rZ);
2322    case INDEX_op_movcond_i32:
2323        return C_O1_I4(r, r, rIN, rIK, 0);
2324    case INDEX_op_add2_i32:
2325        return C_O2_I4(r, r, r, r, rIN, rIK);
2326    case INDEX_op_sub2_i32:
2327        return C_O2_I4(r, r, rI, rI, rIN, rIK);
2328    case INDEX_op_brcond2_i32:
2329        return C_O0_I4(r, r, rI, rI);
2330    case INDEX_op_setcond2_i32:
2331        return C_O1_I4(r, r, r, rI, rI);
2332
2333    case INDEX_op_qemu_ld_i32:
2334        return TARGET_LONG_BITS == 32 ? C_O1_I1(r, l) : C_O1_I2(r, l, l);
2335    case INDEX_op_qemu_ld_i64:
2336        return TARGET_LONG_BITS == 32 ? C_O2_I1(r, r, l) : C_O2_I2(r, r, l, l);
2337    case INDEX_op_qemu_st_i32:
2338        return TARGET_LONG_BITS == 32 ? C_O0_I2(s, s) : C_O0_I3(s, s, s);
2339    case INDEX_op_qemu_st_i64:
2340        return TARGET_LONG_BITS == 32 ? C_O0_I3(s, s, s) : C_O0_I4(s, s, s, s);
2341
2342    case INDEX_op_st_vec:
2343        return C_O0_I2(w, r);
2344    case INDEX_op_ld_vec:
2345    case INDEX_op_dupm_vec:
2346        return C_O1_I1(w, r);
2347    case INDEX_op_dup_vec:
2348        return C_O1_I1(w, wr);
2349    case INDEX_op_abs_vec:
2350    case INDEX_op_neg_vec:
2351    case INDEX_op_not_vec:
2352    case INDEX_op_shli_vec:
2353    case INDEX_op_shri_vec:
2354    case INDEX_op_sari_vec:
2355        return C_O1_I1(w, w);
2356    case INDEX_op_dup2_vec:
2357    case INDEX_op_add_vec:
2358    case INDEX_op_mul_vec:
2359    case INDEX_op_smax_vec:
2360    case INDEX_op_smin_vec:
2361    case INDEX_op_ssadd_vec:
2362    case INDEX_op_sssub_vec:
2363    case INDEX_op_sub_vec:
2364    case INDEX_op_umax_vec:
2365    case INDEX_op_umin_vec:
2366    case INDEX_op_usadd_vec:
2367    case INDEX_op_ussub_vec:
2368    case INDEX_op_xor_vec:
2369    case INDEX_op_arm_sshl_vec:
2370    case INDEX_op_arm_ushl_vec:
2371        return C_O1_I2(w, w, w);
2372    case INDEX_op_arm_sli_vec:
2373        return C_O1_I2(w, 0, w);
2374    case INDEX_op_or_vec:
2375    case INDEX_op_andc_vec:
2376        return C_O1_I2(w, w, wO);
2377    case INDEX_op_and_vec:
2378    case INDEX_op_orc_vec:
2379        return C_O1_I2(w, w, wV);
2380    case INDEX_op_cmp_vec:
2381        return C_O1_I2(w, w, wZ);
2382    case INDEX_op_bitsel_vec:
2383        return C_O1_I3(w, w, w, w);
2384    default:
2385        g_assert_not_reached();
2386    }
2387}
2388
2389static void tcg_target_init(TCGContext *s)
2390{
2391    /*
2392     * Only probe for the platform and capabilities if we haven't already
2393     * determined maximum values at compile time.
2394     */
2395#if !defined(use_idiv_instructions) || !defined(use_neon_instructions)
2396    {
2397        unsigned long hwcap = qemu_getauxval(AT_HWCAP);
2398#ifndef use_idiv_instructions
2399        use_idiv_instructions = (hwcap & HWCAP_ARM_IDIVA) != 0;
2400#endif
2401#ifndef use_neon_instructions
2402        use_neon_instructions = (hwcap & HWCAP_ARM_NEON) != 0;
2403#endif
2404    }
2405#endif
2406
2407    if (__ARM_ARCH < 7) {
2408        const char *pl = (const char *)qemu_getauxval(AT_PLATFORM);
2409        if (pl != NULL && pl[0] == 'v' && pl[1] >= '4' && pl[1] <= '9') {
2410            arm_arch = pl[1] - '0';
2411        }
2412
2413        if (arm_arch < 6) {
2414            error_report("TCG: ARMv%d is unsupported; exiting", arm_arch);
2415            exit(EXIT_FAILURE);
2416        }
2417    }
2418
2419    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
2420
2421    tcg_target_call_clobber_regs = 0;
2422    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R0);
2423    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R1);
2424    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R2);
2425    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R3);
2426    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R12);
2427    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R14);
2428
2429    if (use_neon_instructions) {
2430        tcg_target_available_regs[TCG_TYPE_V64]  = ALL_VECTOR_REGS;
2431        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
2432
2433        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q0);
2434        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q1);
2435        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q2);
2436        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q3);
2437        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q8);
2438        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q9);
2439        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q10);
2440        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q11);
2441        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q12);
2442        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q13);
2443        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q14);
2444        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_Q15);
2445    }
2446
2447    s->reserved_regs = 0;
2448    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
2449    tcg_regset_set_reg(s->reserved_regs, TCG_REG_TMP);
2450    tcg_regset_set_reg(s->reserved_regs, TCG_REG_PC);
2451    tcg_regset_set_reg(s->reserved_regs, TCG_VEC_TMP);
2452}
2453
2454static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
2455                       TCGReg arg1, intptr_t arg2)
2456{
2457    switch (type) {
2458    case TCG_TYPE_I32:
2459        tcg_out_ld32u(s, COND_AL, arg, arg1, arg2);
2460        return;
2461    case TCG_TYPE_V64:
2462        /* regs 1; size 8; align 8 */
2463        tcg_out_vldst(s, INSN_VLD1 | 0x7d0, arg, arg1, arg2);
2464        return;
2465    case TCG_TYPE_V128:
2466        /*
2467         * We have only 8-byte alignment for the stack per the ABI.
2468         * Rather than dynamically re-align the stack, it's easier
2469         * to simply not request alignment beyond that.  So:
2470         * regs 2; size 8; align 8
2471         */
2472        tcg_out_vldst(s, INSN_VLD1 | 0xad0, arg, arg1, arg2);
2473        return;
2474    default:
2475        g_assert_not_reached();
2476    }
2477}
2478
2479static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
2480                       TCGReg arg1, intptr_t arg2)
2481{
2482    switch (type) {
2483    case TCG_TYPE_I32:
2484        tcg_out_st32(s, COND_AL, arg, arg1, arg2);
2485        return;
2486    case TCG_TYPE_V64:
2487        /* regs 1; size 8; align 8 */
2488        tcg_out_vldst(s, INSN_VST1 | 0x7d0, arg, arg1, arg2);
2489        return;
2490    case TCG_TYPE_V128:
2491        /* See tcg_out_ld re alignment: regs 2; size 8; align 8 */
2492        tcg_out_vldst(s, INSN_VST1 | 0xad0, arg, arg1, arg2);
2493        return;
2494    default:
2495        g_assert_not_reached();
2496    }
2497}
2498
2499static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
2500                        TCGReg base, intptr_t ofs)
2501{
2502    return false;
2503}
2504
2505static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
2506{
2507    if (ret == arg) {
2508        return true;
2509    }
2510    switch (type) {
2511    case TCG_TYPE_I32:
2512        if (ret < TCG_REG_Q0 && arg < TCG_REG_Q0) {
2513            tcg_out_mov_reg(s, COND_AL, ret, arg);
2514            return true;
2515        }
2516        return false;
2517
2518    case TCG_TYPE_V64:
2519    case TCG_TYPE_V128:
2520        /* "VMOV D,N" is an alias for "VORR D,N,N". */
2521        tcg_out_vreg3(s, INSN_VORR, type - TCG_TYPE_V64, 0, ret, arg, arg);
2522        return true;
2523
2524    default:
2525        g_assert_not_reached();
2526    }
2527}
2528
2529static void tcg_out_movi(TCGContext *s, TCGType type,
2530                         TCGReg ret, tcg_target_long arg)
2531{
2532    tcg_debug_assert(type == TCG_TYPE_I32);
2533    tcg_debug_assert(ret < TCG_REG_Q0);
2534    tcg_out_movi32(s, COND_AL, ret, arg);
2535}
2536
2537/* Type is always V128, with I64 elements.  */
2538static void tcg_out_dup2_vec(TCGContext *s, TCGReg rd, TCGReg rl, TCGReg rh)
2539{
2540    /* Move high element into place first. */
2541    /* VMOV Dd+1, Ds */
2542    tcg_out_vreg3(s, INSN_VORR | (1 << 12), 0, 0, rd, rh, rh);
2543    /* Move low element into place; tcg_out_mov will check for nop. */
2544    tcg_out_mov(s, TCG_TYPE_V64, rd, rl);
2545}
2546
2547static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
2548                            TCGReg rd, TCGReg rs)
2549{
2550    int q = type - TCG_TYPE_V64;
2551
2552    if (vece == MO_64) {
2553        if (type == TCG_TYPE_V128) {
2554            tcg_out_dup2_vec(s, rd, rs, rs);
2555        } else {
2556            tcg_out_mov(s, TCG_TYPE_V64, rd, rs);
2557        }
2558    } else if (rs < TCG_REG_Q0) {
2559        int b = (vece == MO_8);
2560        int e = (vece == MO_16);
2561        tcg_out32(s, INSN_VDUP_G | (b << 22) | (q << 21) | (e << 5) |
2562                  encode_vn(rd) | (rs << 12));
2563    } else {
2564        int imm4 = 1 << vece;
2565        tcg_out32(s, INSN_VDUP_S | (imm4 << 16) | (q << 6) |
2566                  encode_vd(rd) | encode_vm(rs));
2567    }
2568    return true;
2569}
2570
2571static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
2572                             TCGReg rd, TCGReg base, intptr_t offset)
2573{
2574    if (vece == MO_64) {
2575        tcg_out_ld(s, TCG_TYPE_V64, rd, base, offset);
2576        if (type == TCG_TYPE_V128) {
2577            tcg_out_dup2_vec(s, rd, rd, rd);
2578        }
2579    } else {
2580        int q = type - TCG_TYPE_V64;
2581        tcg_out_vldst(s, INSN_VLD1R | (vece << 6) | (q << 5),
2582                      rd, base, offset);
2583    }
2584    return true;
2585}
2586
2587static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
2588                             TCGReg rd, int64_t v64)
2589{
2590    int q = type - TCG_TYPE_V64;
2591    int cmode, imm8, i;
2592
2593    /* Test all bytes equal first.  */
2594    if (vece == MO_8) {
2595        tcg_out_vmovi(s, rd, q, 0, 0xe, v64);
2596        return;
2597    }
2598
2599    /*
2600     * Test all bytes 0x00 or 0xff second.  This can match cases that
2601     * might otherwise take 2 or 3 insns for MO_16 or MO_32 below.
2602     */
2603    for (i = imm8 = 0; i < 8; i++) {
2604        uint8_t byte = v64 >> (i * 8);
2605        if (byte == 0xff) {
2606            imm8 |= 1 << i;
2607        } else if (byte != 0) {
2608            goto fail_bytes;
2609        }
2610    }
2611    tcg_out_vmovi(s, rd, q, 1, 0xe, imm8);
2612    return;
2613 fail_bytes:
2614
2615    /*
2616     * Tests for various replications.  For each element width, if we
2617     * cannot find an expansion there's no point checking a larger
2618     * width because we already know by replication it cannot match.
2619     */
2620    if (vece == MO_16) {
2621        uint16_t v16 = v64;
2622
2623        if (is_shimm16(v16, &cmode, &imm8)) {
2624            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2625            return;
2626        }
2627        if (is_shimm16(~v16, &cmode, &imm8)) {
2628            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2629            return;
2630        }
2631
2632        /*
2633         * Otherwise, all remaining constants can be loaded in two insns:
2634         * rd = v16 & 0xff, rd |= v16 & 0xff00.
2635         */
2636        tcg_out_vmovi(s, rd, q, 0, 0x8, v16 & 0xff);
2637        tcg_out_vmovi(s, rd, q, 0, 0xb, v16 >> 8);   /* VORRI */
2638        return;
2639    }
2640
2641    if (vece == MO_32) {
2642        uint32_t v32 = v64;
2643
2644        if (is_shimm32(v32, &cmode, &imm8) ||
2645            is_soimm32(v32, &cmode, &imm8)) {
2646            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2647            return;
2648        }
2649        if (is_shimm32(~v32, &cmode, &imm8) ||
2650            is_soimm32(~v32, &cmode, &imm8)) {
2651            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2652            return;
2653        }
2654
2655        /*
2656         * Restrict the set of constants to those we can load with
2657         * two instructions.  Others we load from the pool.
2658         */
2659        i = is_shimm32_pair(v32, &cmode, &imm8);
2660        if (i) {
2661            tcg_out_vmovi(s, rd, q, 0, cmode, imm8);
2662            tcg_out_vmovi(s, rd, q, 0, i | 1, extract32(v32, i * 4, 8));
2663            return;
2664        }
2665        i = is_shimm32_pair(~v32, &cmode, &imm8);
2666        if (i) {
2667            tcg_out_vmovi(s, rd, q, 1, cmode, imm8);
2668            tcg_out_vmovi(s, rd, q, 1, i | 1, extract32(~v32, i * 4, 8));
2669            return;
2670        }
2671    }
2672
2673    /*
2674     * As a last resort, load from the constant pool.
2675     */
2676    if (!q || vece == MO_64) {
2677        new_pool_l2(s, R_ARM_PC11, s->code_ptr, 0, v64, v64 >> 32);
2678        /* VLDR Dd, [pc + offset] */
2679        tcg_out32(s, INSN_VLDR_D | encode_vd(rd) | (0xf << 16));
2680        if (q) {
2681            tcg_out_dup2_vec(s, rd, rd, rd);
2682        }
2683    } else {
2684        new_pool_label(s, (uint32_t)v64, R_ARM_PC8, s->code_ptr, 0);
2685        /* add tmp, pc, offset */
2686        tcg_out_dat_imm(s, COND_AL, ARITH_ADD, TCG_REG_TMP, TCG_REG_PC, 0);
2687        tcg_out_dupm_vec(s, type, MO_32, rd, TCG_REG_TMP, 0);
2688    }
2689}
2690
2691static const ARMInsn vec_cmp_insn[16] = {
2692    [TCG_COND_EQ] = INSN_VCEQ,
2693    [TCG_COND_GT] = INSN_VCGT,
2694    [TCG_COND_GE] = INSN_VCGE,
2695    [TCG_COND_GTU] = INSN_VCGT_U,
2696    [TCG_COND_GEU] = INSN_VCGE_U,
2697};
2698
2699static const ARMInsn vec_cmp0_insn[16] = {
2700    [TCG_COND_EQ] = INSN_VCEQ0,
2701    [TCG_COND_GT] = INSN_VCGT0,
2702    [TCG_COND_GE] = INSN_VCGE0,
2703    [TCG_COND_LT] = INSN_VCLT0,
2704    [TCG_COND_LE] = INSN_VCLE0,
2705};
2706
2707static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2708                           unsigned vecl, unsigned vece,
2709                           const TCGArg args[TCG_MAX_OP_ARGS],
2710                           const int const_args[TCG_MAX_OP_ARGS])
2711{
2712    TCGType type = vecl + TCG_TYPE_V64;
2713    unsigned q = vecl;
2714    TCGArg a0, a1, a2, a3;
2715    int cmode, imm8;
2716
2717    a0 = args[0];
2718    a1 = args[1];
2719    a2 = args[2];
2720
2721    switch (opc) {
2722    case INDEX_op_ld_vec:
2723        tcg_out_ld(s, type, a0, a1, a2);
2724        return;
2725    case INDEX_op_st_vec:
2726        tcg_out_st(s, type, a0, a1, a2);
2727        return;
2728    case INDEX_op_dupm_vec:
2729        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
2730        return;
2731    case INDEX_op_dup2_vec:
2732        tcg_out_dup2_vec(s, a0, a1, a2);
2733        return;
2734    case INDEX_op_abs_vec:
2735        tcg_out_vreg2(s, INSN_VABS, q, vece, a0, a1);
2736        return;
2737    case INDEX_op_neg_vec:
2738        tcg_out_vreg2(s, INSN_VNEG, q, vece, a0, a1);
2739        return;
2740    case INDEX_op_not_vec:
2741        tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a1);
2742        return;
2743    case INDEX_op_add_vec:
2744        tcg_out_vreg3(s, INSN_VADD, q, vece, a0, a1, a2);
2745        return;
2746    case INDEX_op_mul_vec:
2747        tcg_out_vreg3(s, INSN_VMUL, q, vece, a0, a1, a2);
2748        return;
2749    case INDEX_op_smax_vec:
2750        tcg_out_vreg3(s, INSN_VMAX, q, vece, a0, a1, a2);
2751        return;
2752    case INDEX_op_smin_vec:
2753        tcg_out_vreg3(s, INSN_VMIN, q, vece, a0, a1, a2);
2754        return;
2755    case INDEX_op_sub_vec:
2756        tcg_out_vreg3(s, INSN_VSUB, q, vece, a0, a1, a2);
2757        return;
2758    case INDEX_op_ssadd_vec:
2759        tcg_out_vreg3(s, INSN_VQADD, q, vece, a0, a1, a2);
2760        return;
2761    case INDEX_op_sssub_vec:
2762        tcg_out_vreg3(s, INSN_VQSUB, q, vece, a0, a1, a2);
2763        return;
2764    case INDEX_op_umax_vec:
2765        tcg_out_vreg3(s, INSN_VMAX_U, q, vece, a0, a1, a2);
2766        return;
2767    case INDEX_op_umin_vec:
2768        tcg_out_vreg3(s, INSN_VMIN_U, q, vece, a0, a1, a2);
2769        return;
2770    case INDEX_op_usadd_vec:
2771        tcg_out_vreg3(s, INSN_VQADD_U, q, vece, a0, a1, a2);
2772        return;
2773    case INDEX_op_ussub_vec:
2774        tcg_out_vreg3(s, INSN_VQSUB_U, q, vece, a0, a1, a2);
2775        return;
2776    case INDEX_op_xor_vec:
2777        tcg_out_vreg3(s, INSN_VEOR, q, 0, a0, a1, a2);
2778        return;
2779    case INDEX_op_arm_sshl_vec:
2780        /*
2781         * Note that Vm is the data and Vn is the shift count,
2782         * therefore the arguments appear reversed.
2783         */
2784        tcg_out_vreg3(s, INSN_VSHL_S, q, vece, a0, a2, a1);
2785        return;
2786    case INDEX_op_arm_ushl_vec:
2787        /* See above. */
2788        tcg_out_vreg3(s, INSN_VSHL_U, q, vece, a0, a2, a1);
2789        return;
2790    case INDEX_op_shli_vec:
2791        tcg_out_vshifti(s, INSN_VSHLI, q, a0, a1, a2 + (8 << vece));
2792        return;
2793    case INDEX_op_shri_vec:
2794        tcg_out_vshifti(s, INSN_VSHRI, q, a0, a1, (16 << vece) - a2);
2795        return;
2796    case INDEX_op_sari_vec:
2797        tcg_out_vshifti(s, INSN_VSARI, q, a0, a1, (16 << vece) - a2);
2798        return;
2799    case INDEX_op_arm_sli_vec:
2800        tcg_out_vshifti(s, INSN_VSLI, q, a0, a2, args[3] + (8 << vece));
2801        return;
2802
2803    case INDEX_op_andc_vec:
2804        if (!const_args[2]) {
2805            tcg_out_vreg3(s, INSN_VBIC, q, 0, a0, a1, a2);
2806            return;
2807        }
2808        a2 = ~a2;
2809        /* fall through */
2810    case INDEX_op_and_vec:
2811        if (const_args[2]) {
2812            is_shimm1632(~a2, &cmode, &imm8);
2813            if (a0 == a1) {
2814                tcg_out_vmovi(s, a0, q, 1, cmode | 1, imm8); /* VBICI */
2815                return;
2816            }
2817            tcg_out_vmovi(s, a0, q, 1, cmode, imm8); /* VMVNI */
2818            a2 = a0;
2819        }
2820        tcg_out_vreg3(s, INSN_VAND, q, 0, a0, a1, a2);
2821        return;
2822
2823    case INDEX_op_orc_vec:
2824        if (!const_args[2]) {
2825            tcg_out_vreg3(s, INSN_VORN, q, 0, a0, a1, a2);
2826            return;
2827        }
2828        a2 = ~a2;
2829        /* fall through */
2830    case INDEX_op_or_vec:
2831        if (const_args[2]) {
2832            is_shimm1632(a2, &cmode, &imm8);
2833            if (a0 == a1) {
2834                tcg_out_vmovi(s, a0, q, 0, cmode | 1, imm8); /* VORRI */
2835                return;
2836            }
2837            tcg_out_vmovi(s, a0, q, 0, cmode, imm8); /* VMOVI */
2838            a2 = a0;
2839        }
2840        tcg_out_vreg3(s, INSN_VORR, q, 0, a0, a1, a2);
2841        return;
2842
2843    case INDEX_op_cmp_vec:
2844        {
2845            TCGCond cond = args[3];
2846
2847            if (cond == TCG_COND_NE) {
2848                if (const_args[2]) {
2849                    tcg_out_vreg3(s, INSN_VTST, q, vece, a0, a1, a1);
2850                } else {
2851                    tcg_out_vreg3(s, INSN_VCEQ, q, vece, a0, a1, a2);
2852                    tcg_out_vreg2(s, INSN_VMVN, q, 0, a0, a0);
2853                }
2854            } else {
2855                ARMInsn insn;
2856
2857                if (const_args[2]) {
2858                    insn = vec_cmp0_insn[cond];
2859                    if (insn) {
2860                        tcg_out_vreg2(s, insn, q, vece, a0, a1);
2861                        return;
2862                    }
2863                    tcg_out_dupi_vec(s, type, MO_8, TCG_VEC_TMP, 0);
2864                    a2 = TCG_VEC_TMP;
2865                }
2866                insn = vec_cmp_insn[cond];
2867                if (insn == 0) {
2868                    TCGArg t;
2869                    t = a1, a1 = a2, a2 = t;
2870                    cond = tcg_swap_cond(cond);
2871                    insn = vec_cmp_insn[cond];
2872                    tcg_debug_assert(insn != 0);
2873                }
2874                tcg_out_vreg3(s, insn, q, vece, a0, a1, a2);
2875            }
2876        }
2877        return;
2878
2879    case INDEX_op_bitsel_vec:
2880        a3 = args[3];
2881        if (a0 == a3) {
2882            tcg_out_vreg3(s, INSN_VBIT, q, 0, a0, a2, a1);
2883        } else if (a0 == a2) {
2884            tcg_out_vreg3(s, INSN_VBIF, q, 0, a0, a3, a1);
2885        } else {
2886            tcg_out_mov(s, type, a0, a1);
2887            tcg_out_vreg3(s, INSN_VBSL, q, 0, a0, a2, a3);
2888        }
2889        return;
2890
2891    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
2892    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
2893    default:
2894        g_assert_not_reached();
2895    }
2896}
2897
2898int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
2899{
2900    switch (opc) {
2901    case INDEX_op_add_vec:
2902    case INDEX_op_sub_vec:
2903    case INDEX_op_and_vec:
2904    case INDEX_op_andc_vec:
2905    case INDEX_op_or_vec:
2906    case INDEX_op_orc_vec:
2907    case INDEX_op_xor_vec:
2908    case INDEX_op_not_vec:
2909    case INDEX_op_shli_vec:
2910    case INDEX_op_shri_vec:
2911    case INDEX_op_sari_vec:
2912    case INDEX_op_ssadd_vec:
2913    case INDEX_op_sssub_vec:
2914    case INDEX_op_usadd_vec:
2915    case INDEX_op_ussub_vec:
2916    case INDEX_op_bitsel_vec:
2917        return 1;
2918    case INDEX_op_abs_vec:
2919    case INDEX_op_cmp_vec:
2920    case INDEX_op_mul_vec:
2921    case INDEX_op_neg_vec:
2922    case INDEX_op_smax_vec:
2923    case INDEX_op_smin_vec:
2924    case INDEX_op_umax_vec:
2925    case INDEX_op_umin_vec:
2926        return vece < MO_64;
2927    case INDEX_op_shlv_vec:
2928    case INDEX_op_shrv_vec:
2929    case INDEX_op_sarv_vec:
2930    case INDEX_op_rotli_vec:
2931    case INDEX_op_rotlv_vec:
2932    case INDEX_op_rotrv_vec:
2933        return -1;
2934    default:
2935        return 0;
2936    }
2937}
2938
2939void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
2940                       TCGArg a0, ...)
2941{
2942    va_list va;
2943    TCGv_vec v0, v1, v2, t1, t2, c1;
2944    TCGArg a2;
2945
2946    va_start(va, a0);
2947    v0 = temp_tcgv_vec(arg_temp(a0));
2948    v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
2949    a2 = va_arg(va, TCGArg);
2950    va_end(va);
2951
2952    switch (opc) {
2953    case INDEX_op_shlv_vec:
2954        /*
2955         * Merely propagate shlv_vec to arm_ushl_vec.
2956         * In this way we don't set TCG_TARGET_HAS_shv_vec
2957         * because everything is done via expansion.
2958         */
2959        v2 = temp_tcgv_vec(arg_temp(a2));
2960        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
2961                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
2962        break;
2963
2964    case INDEX_op_shrv_vec:
2965    case INDEX_op_sarv_vec:
2966        /* Right shifts are negative left shifts for NEON.  */
2967        v2 = temp_tcgv_vec(arg_temp(a2));
2968        t1 = tcg_temp_new_vec(type);
2969        tcg_gen_neg_vec(vece, t1, v2);
2970        if (opc == INDEX_op_shrv_vec) {
2971            opc = INDEX_op_arm_ushl_vec;
2972        } else {
2973            opc = INDEX_op_arm_sshl_vec;
2974        }
2975        vec_gen_3(opc, type, vece, tcgv_vec_arg(v0),
2976                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2977        tcg_temp_free_vec(t1);
2978        break;
2979
2980    case INDEX_op_rotli_vec:
2981        t1 = tcg_temp_new_vec(type);
2982        tcg_gen_shri_vec(vece, t1, v1, -a2 & ((8 << vece) - 1));
2983        vec_gen_4(INDEX_op_arm_sli_vec, type, vece,
2984                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(v1), a2);
2985        tcg_temp_free_vec(t1);
2986        break;
2987
2988    case INDEX_op_rotlv_vec:
2989        v2 = temp_tcgv_vec(arg_temp(a2));
2990        t1 = tcg_temp_new_vec(type);
2991        c1 = tcg_constant_vec(type, vece, 8 << vece);
2992        tcg_gen_sub_vec(vece, t1, v2, c1);
2993        /* Right shifts are negative left shifts for NEON.  */
2994        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
2995                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
2996        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(v0),
2997                  tcgv_vec_arg(v1), tcgv_vec_arg(v2));
2998        tcg_gen_or_vec(vece, v0, v0, t1);
2999        tcg_temp_free_vec(t1);
3000        break;
3001
3002    case INDEX_op_rotrv_vec:
3003        v2 = temp_tcgv_vec(arg_temp(a2));
3004        t1 = tcg_temp_new_vec(type);
3005        t2 = tcg_temp_new_vec(type);
3006        c1 = tcg_constant_vec(type, vece, 8 << vece);
3007        tcg_gen_neg_vec(vece, t1, v2);
3008        tcg_gen_sub_vec(vece, t2, c1, v2);
3009        /* Right shifts are negative left shifts for NEON.  */
3010        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t1),
3011                  tcgv_vec_arg(v1), tcgv_vec_arg(t1));
3012        vec_gen_3(INDEX_op_arm_ushl_vec, type, vece, tcgv_vec_arg(t2),
3013                  tcgv_vec_arg(v1), tcgv_vec_arg(t2));
3014        tcg_gen_or_vec(vece, v0, t1, t2);
3015        tcg_temp_free_vec(t1);
3016        tcg_temp_free_vec(t2);
3017        break;
3018
3019    default:
3020        g_assert_not_reached();
3021    }
3022}
3023
3024static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
3025{
3026    int i;
3027    for (i = 0; i < count; ++i) {
3028        p[i] = INSN_NOP;
3029    }
3030}
3031
3032/* Compute frame size via macros, to share between tcg_target_qemu_prologue
3033   and tcg_register_jit.  */
3034
3035#define PUSH_SIZE  ((11 - 4 + 1 + 1) * sizeof(tcg_target_long))
3036
3037#define FRAME_SIZE \
3038    ((PUSH_SIZE \
3039      + TCG_STATIC_CALL_ARGS_SIZE \
3040      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
3041      + TCG_TARGET_STACK_ALIGN - 1) \
3042     & -TCG_TARGET_STACK_ALIGN)
3043
3044#define STACK_ADDEND  (FRAME_SIZE - PUSH_SIZE)
3045
3046static void tcg_target_qemu_prologue(TCGContext *s)
3047{
3048    /* Calling convention requires us to save r4-r11 and lr.  */
3049    /* stmdb sp!, { r4 - r11, lr } */
3050    tcg_out_ldstm(s, COND_AL, INSN_STMDB, TCG_REG_CALL_STACK,
3051                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
3052                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3053                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_R14));
3054
3055    /* Reserve callee argument and tcg temp space.  */
3056    tcg_out_dat_rI(s, COND_AL, ARITH_SUB, TCG_REG_CALL_STACK,
3057                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
3058    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
3059                  CPU_TEMP_BUF_NLONGS * sizeof(long));
3060
3061    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
3062
3063#ifndef CONFIG_SOFTMMU
3064    if (guest_base) {
3065        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_GUEST_BASE, guest_base);
3066        tcg_regset_set_reg(s->reserved_regs, TCG_REG_GUEST_BASE);
3067    }
3068#endif
3069
3070    tcg_out_b_reg(s, COND_AL, tcg_target_call_iarg_regs[1]);
3071
3072    /*
3073     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
3074     * and fall through to the rest of the epilogue.
3075     */
3076    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
3077    tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_R0, 0);
3078    tcg_out_epilogue(s);
3079}
3080
3081static void tcg_out_epilogue(TCGContext *s)
3082{
3083    /* Release local stack frame.  */
3084    tcg_out_dat_rI(s, COND_AL, ARITH_ADD, TCG_REG_CALL_STACK,
3085                   TCG_REG_CALL_STACK, STACK_ADDEND, 1);
3086
3087    /* ldmia sp!, { r4 - r11, pc } */
3088    tcg_out_ldstm(s, COND_AL, INSN_LDMIA, TCG_REG_CALL_STACK,
3089                  (1 << TCG_REG_R4) | (1 << TCG_REG_R5) | (1 << TCG_REG_R6) |
3090                  (1 << TCG_REG_R7) | (1 << TCG_REG_R8) | (1 << TCG_REG_R9) |
3091                  (1 << TCG_REG_R10) | (1 << TCG_REG_R11) | (1 << TCG_REG_PC));
3092}
3093
3094typedef struct {
3095    DebugFrameHeader h;
3096    uint8_t fde_def_cfa[4];
3097    uint8_t fde_reg_ofs[18];
3098} DebugFrame;
3099
3100#define ELF_HOST_MACHINE EM_ARM
3101
3102/* We're expecting a 2 byte uleb128 encoded value.  */
3103QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
3104
3105static const DebugFrame debug_frame = {
3106    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
3107    .h.cie.id = -1,
3108    .h.cie.version = 1,
3109    .h.cie.code_align = 1,
3110    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
3111    .h.cie.return_column = 14,
3112
3113    /* Total FDE size does not include the "len" member.  */
3114    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
3115
3116    .fde_def_cfa = {
3117        12, 13,                         /* DW_CFA_def_cfa sp, ... */
3118        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
3119        (FRAME_SIZE >> 7)
3120    },
3121    .fde_reg_ofs = {
3122        /* The following must match the stmdb in the prologue.  */
3123        0x8e, 1,                        /* DW_CFA_offset, lr, -4 */
3124        0x8b, 2,                        /* DW_CFA_offset, r11, -8 */
3125        0x8a, 3,                        /* DW_CFA_offset, r10, -12 */
3126        0x89, 4,                        /* DW_CFA_offset, r9, -16 */
3127        0x88, 5,                        /* DW_CFA_offset, r8, -20 */
3128        0x87, 6,                        /* DW_CFA_offset, r7, -24 */
3129        0x86, 7,                        /* DW_CFA_offset, r6, -28 */
3130        0x85, 8,                        /* DW_CFA_offset, r5, -32 */
3131        0x84, 9,                        /* DW_CFA_offset, r4, -36 */
3132    }
3133};
3134
3135void tcg_register_jit(const void *buf, size_t buf_size)
3136{
3137    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
3138}
3139