xref: /openbmc/qemu/tcg/i386/tcg-target.c.inc (revision 5e0d6590)
1/*
2 * Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2008 Fabrice Bellard
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 "../tcg-ldst.c.inc"
26#include "../tcg-pool.c.inc"
27
28#ifdef CONFIG_DEBUG_TCG
29static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
30#if TCG_TARGET_REG_BITS == 64
31    "%rax", "%rcx", "%rdx", "%rbx", "%rsp", "%rbp", "%rsi", "%rdi",
32#else
33    "%eax", "%ecx", "%edx", "%ebx", "%esp", "%ebp", "%esi", "%edi",
34#endif
35    "%r8",  "%r9",  "%r10", "%r11", "%r12", "%r13", "%r14", "%r15",
36    "%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5", "%xmm6", "%xmm7",
37#if TCG_TARGET_REG_BITS == 64
38    "%xmm8", "%xmm9", "%xmm10", "%xmm11",
39    "%xmm12", "%xmm13", "%xmm14", "%xmm15",
40#endif
41};
42#endif
43
44static const int tcg_target_reg_alloc_order[] = {
45#if TCG_TARGET_REG_BITS == 64
46    TCG_REG_RBP,
47    TCG_REG_RBX,
48    TCG_REG_R12,
49    TCG_REG_R13,
50    TCG_REG_R14,
51    TCG_REG_R15,
52    TCG_REG_R10,
53    TCG_REG_R11,
54    TCG_REG_R9,
55    TCG_REG_R8,
56    TCG_REG_RCX,
57    TCG_REG_RDX,
58    TCG_REG_RSI,
59    TCG_REG_RDI,
60    TCG_REG_RAX,
61#else
62    TCG_REG_EBX,
63    TCG_REG_ESI,
64    TCG_REG_EDI,
65    TCG_REG_EBP,
66    TCG_REG_ECX,
67    TCG_REG_EDX,
68    TCG_REG_EAX,
69#endif
70    TCG_REG_XMM0,
71    TCG_REG_XMM1,
72    TCG_REG_XMM2,
73    TCG_REG_XMM3,
74    TCG_REG_XMM4,
75    TCG_REG_XMM5,
76#ifndef _WIN64
77    /* The Win64 ABI has xmm6-xmm15 as caller-saves, and we do not save
78       any of them.  Therefore only allow xmm0-xmm5 to be allocated.  */
79    TCG_REG_XMM6,
80    TCG_REG_XMM7,
81#if TCG_TARGET_REG_BITS == 64
82    TCG_REG_XMM8,
83    TCG_REG_XMM9,
84    TCG_REG_XMM10,
85    TCG_REG_XMM11,
86    TCG_REG_XMM12,
87    TCG_REG_XMM13,
88    TCG_REG_XMM14,
89    TCG_REG_XMM15,
90#endif
91#endif
92};
93
94#define TCG_TMP_VEC  TCG_REG_XMM5
95
96static const int tcg_target_call_iarg_regs[] = {
97#if TCG_TARGET_REG_BITS == 64
98#if defined(_WIN64)
99    TCG_REG_RCX,
100    TCG_REG_RDX,
101#else
102    TCG_REG_RDI,
103    TCG_REG_RSI,
104    TCG_REG_RDX,
105    TCG_REG_RCX,
106#endif
107    TCG_REG_R8,
108    TCG_REG_R9,
109#else
110    /* 32 bit mode uses stack based calling convention (GCC default). */
111#endif
112};
113
114static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
115{
116    switch (kind) {
117    case TCG_CALL_RET_NORMAL:
118        tcg_debug_assert(slot >= 0 && slot <= 1);
119        return slot ? TCG_REG_EDX : TCG_REG_EAX;
120#ifdef _WIN64
121    case TCG_CALL_RET_BY_VEC:
122        tcg_debug_assert(slot == 0);
123        return TCG_REG_XMM0;
124#endif
125    default:
126        g_assert_not_reached();
127    }
128}
129
130/* Constants we accept.  */
131#define TCG_CT_CONST_S32 0x100
132#define TCG_CT_CONST_U32 0x200
133#define TCG_CT_CONST_I32 0x400
134#define TCG_CT_CONST_WSZ 0x800
135
136/* Registers used with L constraint, which are the first argument
137   registers on x86_64, and two random call clobbered registers on
138   i386. */
139#if TCG_TARGET_REG_BITS == 64
140# define TCG_REG_L0 tcg_target_call_iarg_regs[0]
141# define TCG_REG_L1 tcg_target_call_iarg_regs[1]
142#else
143# define TCG_REG_L0 TCG_REG_EAX
144# define TCG_REG_L1 TCG_REG_EDX
145#endif
146
147#define ALL_BYTEH_REGS         0x0000000fu
148#if TCG_TARGET_REG_BITS == 64
149# define ALL_GENERAL_REGS      0x0000ffffu
150# define ALL_VECTOR_REGS       0xffff0000u
151# define ALL_BYTEL_REGS        ALL_GENERAL_REGS
152#else
153# define ALL_GENERAL_REGS      0x000000ffu
154# define ALL_VECTOR_REGS       0x00ff0000u
155# define ALL_BYTEL_REGS        ALL_BYTEH_REGS
156#endif
157#ifdef CONFIG_SOFTMMU
158# define SOFTMMU_RESERVE_REGS  ((1 << TCG_REG_L0) | (1 << TCG_REG_L1))
159#else
160# define SOFTMMU_RESERVE_REGS  0
161#endif
162
163/* For 64-bit, we always know that CMOV is available.  */
164#if TCG_TARGET_REG_BITS == 64
165# define have_cmov      true
166#else
167# define have_cmov      (cpuinfo & CPUINFO_CMOV)
168#endif
169#define have_bmi2       (cpuinfo & CPUINFO_BMI2)
170#define have_lzcnt      (cpuinfo & CPUINFO_LZCNT)
171
172static const tcg_insn_unit *tb_ret_addr;
173
174static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
175                        intptr_t value, intptr_t addend)
176{
177    value += addend;
178    switch(type) {
179    case R_386_PC32:
180        value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
181        if (value != (int32_t)value) {
182            return false;
183        }
184        /* FALLTHRU */
185    case R_386_32:
186        tcg_patch32(code_ptr, value);
187        break;
188    case R_386_PC8:
189        value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
190        if (value != (int8_t)value) {
191            return false;
192        }
193        tcg_patch8(code_ptr, value);
194        break;
195    default:
196        g_assert_not_reached();
197    }
198    return true;
199}
200
201/* test if a constant matches the constraint */
202static bool tcg_target_const_match(int64_t val, TCGType type, int ct)
203{
204    if (ct & TCG_CT_CONST) {
205        return 1;
206    }
207    if (type == TCG_TYPE_I32) {
208        if (ct & (TCG_CT_CONST_S32 | TCG_CT_CONST_U32 | TCG_CT_CONST_I32)) {
209            return 1;
210        }
211    } else {
212        if ((ct & TCG_CT_CONST_S32) && val == (int32_t)val) {
213            return 1;
214        }
215        if ((ct & TCG_CT_CONST_U32) && val == (uint32_t)val) {
216            return 1;
217        }
218        if ((ct & TCG_CT_CONST_I32) && ~val == (int32_t)~val) {
219            return 1;
220        }
221    }
222    if ((ct & TCG_CT_CONST_WSZ) && val == (type == TCG_TYPE_I32 ? 32 : 64)) {
223        return 1;
224    }
225    return 0;
226}
227
228# define LOWREGMASK(x)	((x) & 7)
229
230#define P_EXT		0x100		/* 0x0f opcode prefix */
231#define P_EXT38         0x200           /* 0x0f 0x38 opcode prefix */
232#define P_DATA16        0x400           /* 0x66 opcode prefix */
233#define P_VEXW          0x1000          /* Set VEX.W = 1 */
234#if TCG_TARGET_REG_BITS == 64
235# define P_REXW         P_VEXW          /* Set REX.W = 1; match VEXW */
236# define P_REXB_R       0x2000          /* REG field as byte register */
237# define P_REXB_RM      0x4000          /* R/M field as byte register */
238# define P_GS           0x8000          /* gs segment override */
239#else
240# define P_REXW		0
241# define P_REXB_R	0
242# define P_REXB_RM	0
243# define P_GS           0
244#endif
245#define P_EXT3A         0x10000         /* 0x0f 0x3a opcode prefix */
246#define P_SIMDF3        0x20000         /* 0xf3 opcode prefix */
247#define P_SIMDF2        0x40000         /* 0xf2 opcode prefix */
248#define P_VEXL          0x80000         /* Set VEX.L = 1 */
249#define P_EVEX          0x100000        /* Requires EVEX encoding */
250
251#define OPC_ARITH_EvIz	(0x81)
252#define OPC_ARITH_EvIb	(0x83)
253#define OPC_ARITH_GvEv	(0x03)		/* ... plus (ARITH_FOO << 3) */
254#define OPC_ANDN        (0xf2 | P_EXT38)
255#define OPC_ADD_GvEv	(OPC_ARITH_GvEv | (ARITH_ADD << 3))
256#define OPC_AND_GvEv    (OPC_ARITH_GvEv | (ARITH_AND << 3))
257#define OPC_BLENDPS     (0x0c | P_EXT3A | P_DATA16)
258#define OPC_BSF         (0xbc | P_EXT)
259#define OPC_BSR         (0xbd | P_EXT)
260#define OPC_BSWAP	(0xc8 | P_EXT)
261#define OPC_CALL_Jz	(0xe8)
262#define OPC_CMOVCC      (0x40 | P_EXT)  /* ... plus condition code */
263#define OPC_CMP_GvEv	(OPC_ARITH_GvEv | (ARITH_CMP << 3))
264#define OPC_DEC_r32	(0x48)
265#define OPC_IMUL_GvEv	(0xaf | P_EXT)
266#define OPC_IMUL_GvEvIb	(0x6b)
267#define OPC_IMUL_GvEvIz	(0x69)
268#define OPC_INC_r32	(0x40)
269#define OPC_JCC_long	(0x80 | P_EXT)	/* ... plus condition code */
270#define OPC_JCC_short	(0x70)		/* ... plus condition code */
271#define OPC_JMP_long	(0xe9)
272#define OPC_JMP_short	(0xeb)
273#define OPC_LEA         (0x8d)
274#define OPC_LZCNT       (0xbd | P_EXT | P_SIMDF3)
275#define OPC_MOVB_EvGv	(0x88)		/* stores, more or less */
276#define OPC_MOVL_EvGv	(0x89)		/* stores, more or less */
277#define OPC_MOVL_GvEv	(0x8b)		/* loads, more or less */
278#define OPC_MOVB_EvIz   (0xc6)
279#define OPC_MOVL_EvIz	(0xc7)
280#define OPC_MOVL_Iv     (0xb8)
281#define OPC_MOVBE_GyMy  (0xf0 | P_EXT38)
282#define OPC_MOVBE_MyGy  (0xf1 | P_EXT38)
283#define OPC_MOVD_VyEy   (0x6e | P_EXT | P_DATA16)
284#define OPC_MOVD_EyVy   (0x7e | P_EXT | P_DATA16)
285#define OPC_MOVDDUP     (0x12 | P_EXT | P_SIMDF2)
286#define OPC_MOVDQA_VxWx (0x6f | P_EXT | P_DATA16)
287#define OPC_MOVDQA_WxVx (0x7f | P_EXT | P_DATA16)
288#define OPC_MOVDQU_VxWx (0x6f | P_EXT | P_SIMDF3)
289#define OPC_MOVDQU_WxVx (0x7f | P_EXT | P_SIMDF3)
290#define OPC_MOVQ_VqWq   (0x7e | P_EXT | P_SIMDF3)
291#define OPC_MOVQ_WqVq   (0xd6 | P_EXT | P_DATA16)
292#define OPC_MOVSBL	(0xbe | P_EXT)
293#define OPC_MOVSWL	(0xbf | P_EXT)
294#define OPC_MOVSLQ	(0x63 | P_REXW)
295#define OPC_MOVZBL	(0xb6 | P_EXT)
296#define OPC_MOVZWL	(0xb7 | P_EXT)
297#define OPC_PABSB       (0x1c | P_EXT38 | P_DATA16)
298#define OPC_PABSW       (0x1d | P_EXT38 | P_DATA16)
299#define OPC_PABSD       (0x1e | P_EXT38 | P_DATA16)
300#define OPC_VPABSQ      (0x1f | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
301#define OPC_PACKSSDW    (0x6b | P_EXT | P_DATA16)
302#define OPC_PACKSSWB    (0x63 | P_EXT | P_DATA16)
303#define OPC_PACKUSDW    (0x2b | P_EXT38 | P_DATA16)
304#define OPC_PACKUSWB    (0x67 | P_EXT | P_DATA16)
305#define OPC_PADDB       (0xfc | P_EXT | P_DATA16)
306#define OPC_PADDW       (0xfd | P_EXT | P_DATA16)
307#define OPC_PADDD       (0xfe | P_EXT | P_DATA16)
308#define OPC_PADDQ       (0xd4 | P_EXT | P_DATA16)
309#define OPC_PADDSB      (0xec | P_EXT | P_DATA16)
310#define OPC_PADDSW      (0xed | P_EXT | P_DATA16)
311#define OPC_PADDUB      (0xdc | P_EXT | P_DATA16)
312#define OPC_PADDUW      (0xdd | P_EXT | P_DATA16)
313#define OPC_PAND        (0xdb | P_EXT | P_DATA16)
314#define OPC_PANDN       (0xdf | P_EXT | P_DATA16)
315#define OPC_PBLENDW     (0x0e | P_EXT3A | P_DATA16)
316#define OPC_PCMPEQB     (0x74 | P_EXT | P_DATA16)
317#define OPC_PCMPEQW     (0x75 | P_EXT | P_DATA16)
318#define OPC_PCMPEQD     (0x76 | P_EXT | P_DATA16)
319#define OPC_PCMPEQQ     (0x29 | P_EXT38 | P_DATA16)
320#define OPC_PCMPGTB     (0x64 | P_EXT | P_DATA16)
321#define OPC_PCMPGTW     (0x65 | P_EXT | P_DATA16)
322#define OPC_PCMPGTD     (0x66 | P_EXT | P_DATA16)
323#define OPC_PCMPGTQ     (0x37 | P_EXT38 | P_DATA16)
324#define OPC_PEXTRD      (0x16 | P_EXT3A | P_DATA16)
325#define OPC_PINSRD      (0x22 | P_EXT3A | P_DATA16)
326#define OPC_PMAXSB      (0x3c | P_EXT38 | P_DATA16)
327#define OPC_PMAXSW      (0xee | P_EXT | P_DATA16)
328#define OPC_PMAXSD      (0x3d | P_EXT38 | P_DATA16)
329#define OPC_VPMAXSQ     (0x3d | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
330#define OPC_PMAXUB      (0xde | P_EXT | P_DATA16)
331#define OPC_PMAXUW      (0x3e | P_EXT38 | P_DATA16)
332#define OPC_PMAXUD      (0x3f | P_EXT38 | P_DATA16)
333#define OPC_VPMAXUQ     (0x3f | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
334#define OPC_PMINSB      (0x38 | P_EXT38 | P_DATA16)
335#define OPC_PMINSW      (0xea | P_EXT | P_DATA16)
336#define OPC_PMINSD      (0x39 | P_EXT38 | P_DATA16)
337#define OPC_VPMINSQ     (0x39 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
338#define OPC_PMINUB      (0xda | P_EXT | P_DATA16)
339#define OPC_PMINUW      (0x3a | P_EXT38 | P_DATA16)
340#define OPC_PMINUD      (0x3b | P_EXT38 | P_DATA16)
341#define OPC_VPMINUQ     (0x3b | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
342#define OPC_PMOVSXBW    (0x20 | P_EXT38 | P_DATA16)
343#define OPC_PMOVSXWD    (0x23 | P_EXT38 | P_DATA16)
344#define OPC_PMOVSXDQ    (0x25 | P_EXT38 | P_DATA16)
345#define OPC_PMOVZXBW    (0x30 | P_EXT38 | P_DATA16)
346#define OPC_PMOVZXWD    (0x33 | P_EXT38 | P_DATA16)
347#define OPC_PMOVZXDQ    (0x35 | P_EXT38 | P_DATA16)
348#define OPC_PMULLW      (0xd5 | P_EXT | P_DATA16)
349#define OPC_PMULLD      (0x40 | P_EXT38 | P_DATA16)
350#define OPC_VPMULLQ     (0x40 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
351#define OPC_POR         (0xeb | P_EXT | P_DATA16)
352#define OPC_PSHUFB      (0x00 | P_EXT38 | P_DATA16)
353#define OPC_PSHUFD      (0x70 | P_EXT | P_DATA16)
354#define OPC_PSHUFLW     (0x70 | P_EXT | P_SIMDF2)
355#define OPC_PSHUFHW     (0x70 | P_EXT | P_SIMDF3)
356#define OPC_PSHIFTW_Ib  (0x71 | P_EXT | P_DATA16) /* /2 /6 /4 */
357#define OPC_PSHIFTD_Ib  (0x72 | P_EXT | P_DATA16) /* /1 /2 /6 /4 */
358#define OPC_PSHIFTQ_Ib  (0x73 | P_EXT | P_DATA16) /* /2 /6 /4 */
359#define OPC_PSLLW       (0xf1 | P_EXT | P_DATA16)
360#define OPC_PSLLD       (0xf2 | P_EXT | P_DATA16)
361#define OPC_PSLLQ       (0xf3 | P_EXT | P_DATA16)
362#define OPC_PSRAW       (0xe1 | P_EXT | P_DATA16)
363#define OPC_PSRAD       (0xe2 | P_EXT | P_DATA16)
364#define OPC_VPSRAQ      (0xe2 | P_EXT | P_DATA16 | P_VEXW | P_EVEX)
365#define OPC_PSRLW       (0xd1 | P_EXT | P_DATA16)
366#define OPC_PSRLD       (0xd2 | P_EXT | P_DATA16)
367#define OPC_PSRLQ       (0xd3 | P_EXT | P_DATA16)
368#define OPC_PSUBB       (0xf8 | P_EXT | P_DATA16)
369#define OPC_PSUBW       (0xf9 | P_EXT | P_DATA16)
370#define OPC_PSUBD       (0xfa | P_EXT | P_DATA16)
371#define OPC_PSUBQ       (0xfb | P_EXT | P_DATA16)
372#define OPC_PSUBSB      (0xe8 | P_EXT | P_DATA16)
373#define OPC_PSUBSW      (0xe9 | P_EXT | P_DATA16)
374#define OPC_PSUBUB      (0xd8 | P_EXT | P_DATA16)
375#define OPC_PSUBUW      (0xd9 | P_EXT | P_DATA16)
376#define OPC_PUNPCKLBW   (0x60 | P_EXT | P_DATA16)
377#define OPC_PUNPCKLWD   (0x61 | P_EXT | P_DATA16)
378#define OPC_PUNPCKLDQ   (0x62 | P_EXT | P_DATA16)
379#define OPC_PUNPCKLQDQ  (0x6c | P_EXT | P_DATA16)
380#define OPC_PUNPCKHBW   (0x68 | P_EXT | P_DATA16)
381#define OPC_PUNPCKHWD   (0x69 | P_EXT | P_DATA16)
382#define OPC_PUNPCKHDQ   (0x6a | P_EXT | P_DATA16)
383#define OPC_PUNPCKHQDQ  (0x6d | P_EXT | P_DATA16)
384#define OPC_PXOR        (0xef | P_EXT | P_DATA16)
385#define OPC_POP_r32	(0x58)
386#define OPC_POPCNT      (0xb8 | P_EXT | P_SIMDF3)
387#define OPC_PUSH_r32	(0x50)
388#define OPC_PUSH_Iv	(0x68)
389#define OPC_PUSH_Ib	(0x6a)
390#define OPC_RET		(0xc3)
391#define OPC_SETCC	(0x90 | P_EXT | P_REXB_RM) /* ... plus cc */
392#define OPC_SHIFT_1	(0xd1)
393#define OPC_SHIFT_Ib	(0xc1)
394#define OPC_SHIFT_cl	(0xd3)
395#define OPC_SARX        (0xf7 | P_EXT38 | P_SIMDF3)
396#define OPC_SHUFPS      (0xc6 | P_EXT)
397#define OPC_SHLX        (0xf7 | P_EXT38 | P_DATA16)
398#define OPC_SHRX        (0xf7 | P_EXT38 | P_SIMDF2)
399#define OPC_SHRD_Ib     (0xac | P_EXT)
400#define OPC_TESTL	(0x85)
401#define OPC_TZCNT       (0xbc | P_EXT | P_SIMDF3)
402#define OPC_UD2         (0x0b | P_EXT)
403#define OPC_VPBLENDD    (0x02 | P_EXT3A | P_DATA16)
404#define OPC_VPBLENDVB   (0x4c | P_EXT3A | P_DATA16)
405#define OPC_VPINSRB     (0x20 | P_EXT3A | P_DATA16)
406#define OPC_VPINSRW     (0xc4 | P_EXT | P_DATA16)
407#define OPC_VBROADCASTSS (0x18 | P_EXT38 | P_DATA16)
408#define OPC_VBROADCASTSD (0x19 | P_EXT38 | P_DATA16)
409#define OPC_VPBROADCASTB (0x78 | P_EXT38 | P_DATA16)
410#define OPC_VPBROADCASTW (0x79 | P_EXT38 | P_DATA16)
411#define OPC_VPBROADCASTD (0x58 | P_EXT38 | P_DATA16)
412#define OPC_VPBROADCASTQ (0x59 | P_EXT38 | P_DATA16)
413#define OPC_VPERMQ      (0x00 | P_EXT3A | P_DATA16 | P_VEXW)
414#define OPC_VPERM2I128  (0x46 | P_EXT3A | P_DATA16 | P_VEXL)
415#define OPC_VPROLVD     (0x15 | P_EXT38 | P_DATA16 | P_EVEX)
416#define OPC_VPROLVQ     (0x15 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
417#define OPC_VPRORVD     (0x14 | P_EXT38 | P_DATA16 | P_EVEX)
418#define OPC_VPRORVQ     (0x14 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
419#define OPC_VPSHLDW     (0x70 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
420#define OPC_VPSHLDD     (0x71 | P_EXT3A | P_DATA16 | P_EVEX)
421#define OPC_VPSHLDQ     (0x71 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
422#define OPC_VPSHLDVW    (0x70 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
423#define OPC_VPSHLDVD    (0x71 | P_EXT38 | P_DATA16 | P_EVEX)
424#define OPC_VPSHLDVQ    (0x71 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
425#define OPC_VPSHRDVW    (0x72 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
426#define OPC_VPSHRDVD    (0x73 | P_EXT38 | P_DATA16 | P_EVEX)
427#define OPC_VPSHRDVQ    (0x73 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
428#define OPC_VPSLLVW     (0x12 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
429#define OPC_VPSLLVD     (0x47 | P_EXT38 | P_DATA16)
430#define OPC_VPSLLVQ     (0x47 | P_EXT38 | P_DATA16 | P_VEXW)
431#define OPC_VPSRAVW     (0x11 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
432#define OPC_VPSRAVD     (0x46 | P_EXT38 | P_DATA16)
433#define OPC_VPSRAVQ     (0x46 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
434#define OPC_VPSRLVW     (0x10 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
435#define OPC_VPSRLVD     (0x45 | P_EXT38 | P_DATA16)
436#define OPC_VPSRLVQ     (0x45 | P_EXT38 | P_DATA16 | P_VEXW)
437#define OPC_VPTERNLOGQ  (0x25 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
438#define OPC_VZEROUPPER  (0x77 | P_EXT)
439#define OPC_XCHG_ax_r32	(0x90)
440#define OPC_XCHG_EvGv   (0x87)
441
442#define OPC_GRP3_Eb     (0xf6)
443#define OPC_GRP3_Ev     (0xf7)
444#define OPC_GRP5        (0xff)
445#define OPC_GRP14       (0x73 | P_EXT | P_DATA16)
446
447/* Group 1 opcode extensions for 0x80-0x83.
448   These are also used as modifiers for OPC_ARITH.  */
449#define ARITH_ADD 0
450#define ARITH_OR  1
451#define ARITH_ADC 2
452#define ARITH_SBB 3
453#define ARITH_AND 4
454#define ARITH_SUB 5
455#define ARITH_XOR 6
456#define ARITH_CMP 7
457
458/* Group 2 opcode extensions for 0xc0, 0xc1, 0xd0-0xd3.  */
459#define SHIFT_ROL 0
460#define SHIFT_ROR 1
461#define SHIFT_SHL 4
462#define SHIFT_SHR 5
463#define SHIFT_SAR 7
464
465/* Group 3 opcode extensions for 0xf6, 0xf7.  To be used with OPC_GRP3.  */
466#define EXT3_TESTi 0
467#define EXT3_NOT   2
468#define EXT3_NEG   3
469#define EXT3_MUL   4
470#define EXT3_IMUL  5
471#define EXT3_DIV   6
472#define EXT3_IDIV  7
473
474/* Group 5 opcode extensions for 0xff.  To be used with OPC_GRP5.  */
475#define EXT5_INC_Ev	0
476#define EXT5_DEC_Ev	1
477#define EXT5_CALLN_Ev	2
478#define EXT5_JMPN_Ev	4
479
480/* Condition codes to be added to OPC_JCC_{long,short}.  */
481#define JCC_JMP (-1)
482#define JCC_JO  0x0
483#define JCC_JNO 0x1
484#define JCC_JB  0x2
485#define JCC_JAE 0x3
486#define JCC_JE  0x4
487#define JCC_JNE 0x5
488#define JCC_JBE 0x6
489#define JCC_JA  0x7
490#define JCC_JS  0x8
491#define JCC_JNS 0x9
492#define JCC_JP  0xa
493#define JCC_JNP 0xb
494#define JCC_JL  0xc
495#define JCC_JGE 0xd
496#define JCC_JLE 0xe
497#define JCC_JG  0xf
498
499static const uint8_t tcg_cond_to_jcc[] = {
500    [TCG_COND_EQ] = JCC_JE,
501    [TCG_COND_NE] = JCC_JNE,
502    [TCG_COND_LT] = JCC_JL,
503    [TCG_COND_GE] = JCC_JGE,
504    [TCG_COND_LE] = JCC_JLE,
505    [TCG_COND_GT] = JCC_JG,
506    [TCG_COND_LTU] = JCC_JB,
507    [TCG_COND_GEU] = JCC_JAE,
508    [TCG_COND_LEU] = JCC_JBE,
509    [TCG_COND_GTU] = JCC_JA,
510};
511
512#if TCG_TARGET_REG_BITS == 64
513static void tcg_out_opc(TCGContext *s, int opc, int r, int rm, int x)
514{
515    int rex;
516
517    if (opc & P_GS) {
518        tcg_out8(s, 0x65);
519    }
520    if (opc & P_DATA16) {
521        /* We should never be asking for both 16 and 64-bit operation.  */
522        tcg_debug_assert((opc & P_REXW) == 0);
523        tcg_out8(s, 0x66);
524    }
525    if (opc & P_SIMDF3) {
526        tcg_out8(s, 0xf3);
527    } else if (opc & P_SIMDF2) {
528        tcg_out8(s, 0xf2);
529    }
530
531    rex = 0;
532    rex |= (opc & P_REXW) ? 0x8 : 0x0;  /* REX.W */
533    rex |= (r & 8) >> 1;                /* REX.R */
534    rex |= (x & 8) >> 2;                /* REX.X */
535    rex |= (rm & 8) >> 3;               /* REX.B */
536
537    /* P_REXB_{R,RM} indicates that the given register is the low byte.
538       For %[abcd]l we need no REX prefix, but for %{si,di,bp,sp}l we do,
539       as otherwise the encoding indicates %[abcd]h.  Note that the values
540       that are ORed in merely indicate that the REX byte must be present;
541       those bits get discarded in output.  */
542    rex |= opc & (r >= 4 ? P_REXB_R : 0);
543    rex |= opc & (rm >= 4 ? P_REXB_RM : 0);
544
545    if (rex) {
546        tcg_out8(s, (uint8_t)(rex | 0x40));
547    }
548
549    if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
550        tcg_out8(s, 0x0f);
551        if (opc & P_EXT38) {
552            tcg_out8(s, 0x38);
553        } else if (opc & P_EXT3A) {
554            tcg_out8(s, 0x3a);
555        }
556    }
557
558    tcg_out8(s, opc);
559}
560#else
561static void tcg_out_opc(TCGContext *s, int opc)
562{
563    if (opc & P_DATA16) {
564        tcg_out8(s, 0x66);
565    }
566    if (opc & P_SIMDF3) {
567        tcg_out8(s, 0xf3);
568    } else if (opc & P_SIMDF2) {
569        tcg_out8(s, 0xf2);
570    }
571    if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
572        tcg_out8(s, 0x0f);
573        if (opc & P_EXT38) {
574            tcg_out8(s, 0x38);
575        } else if (opc & P_EXT3A) {
576            tcg_out8(s, 0x3a);
577        }
578    }
579    tcg_out8(s, opc);
580}
581/* Discard the register arguments to tcg_out_opc early, so as not to penalize
582   the 32-bit compilation paths.  This method works with all versions of gcc,
583   whereas relying on optimization may not be able to exclude them.  */
584#define tcg_out_opc(s, opc, r, rm, x)  (tcg_out_opc)(s, opc)
585#endif
586
587static void tcg_out_modrm(TCGContext *s, int opc, int r, int rm)
588{
589    tcg_out_opc(s, opc, r, rm, 0);
590    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
591}
592
593static void tcg_out_vex_opc(TCGContext *s, int opc, int r, int v,
594                            int rm, int index)
595{
596    int tmp;
597
598    if (opc & P_GS) {
599        tcg_out8(s, 0x65);
600    }
601    /* Use the two byte form if possible, which cannot encode
602       VEX.W, VEX.B, VEX.X, or an m-mmmm field other than P_EXT.  */
603    if ((opc & (P_EXT | P_EXT38 | P_EXT3A | P_VEXW)) == P_EXT
604        && ((rm | index) & 8) == 0) {
605        /* Two byte VEX prefix.  */
606        tcg_out8(s, 0xc5);
607
608        tmp = (r & 8 ? 0 : 0x80);              /* VEX.R */
609    } else {
610        /* Three byte VEX prefix.  */
611        tcg_out8(s, 0xc4);
612
613        /* VEX.m-mmmm */
614        if (opc & P_EXT3A) {
615            tmp = 3;
616        } else if (opc & P_EXT38) {
617            tmp = 2;
618        } else if (opc & P_EXT) {
619            tmp = 1;
620        } else {
621            g_assert_not_reached();
622        }
623        tmp |= (r & 8 ? 0 : 0x80);             /* VEX.R */
624        tmp |= (index & 8 ? 0 : 0x40);         /* VEX.X */
625        tmp |= (rm & 8 ? 0 : 0x20);            /* VEX.B */
626        tcg_out8(s, tmp);
627
628        tmp = (opc & P_VEXW ? 0x80 : 0);       /* VEX.W */
629    }
630
631    tmp |= (opc & P_VEXL ? 0x04 : 0);      /* VEX.L */
632    /* VEX.pp */
633    if (opc & P_DATA16) {
634        tmp |= 1;                          /* 0x66 */
635    } else if (opc & P_SIMDF3) {
636        tmp |= 2;                          /* 0xf3 */
637    } else if (opc & P_SIMDF2) {
638        tmp |= 3;                          /* 0xf2 */
639    }
640    tmp |= (~v & 15) << 3;                 /* VEX.vvvv */
641    tcg_out8(s, tmp);
642    tcg_out8(s, opc);
643}
644
645static void tcg_out_evex_opc(TCGContext *s, int opc, int r, int v,
646                             int rm, int index)
647{
648    /* The entire 4-byte evex prefix; with R' and V' set. */
649    uint32_t p = 0x08041062;
650    int mm, pp;
651
652    tcg_debug_assert(have_avx512vl);
653
654    /* EVEX.mm */
655    if (opc & P_EXT3A) {
656        mm = 3;
657    } else if (opc & P_EXT38) {
658        mm = 2;
659    } else if (opc & P_EXT) {
660        mm = 1;
661    } else {
662        g_assert_not_reached();
663    }
664
665    /* EVEX.pp */
666    if (opc & P_DATA16) {
667        pp = 1;                          /* 0x66 */
668    } else if (opc & P_SIMDF3) {
669        pp = 2;                          /* 0xf3 */
670    } else if (opc & P_SIMDF2) {
671        pp = 3;                          /* 0xf2 */
672    } else {
673        pp = 0;
674    }
675
676    p = deposit32(p, 8, 2, mm);
677    p = deposit32(p, 13, 1, (rm & 8) == 0);             /* EVEX.RXB.B */
678    p = deposit32(p, 14, 1, (index & 8) == 0);          /* EVEX.RXB.X */
679    p = deposit32(p, 15, 1, (r & 8) == 0);              /* EVEX.RXB.R */
680    p = deposit32(p, 16, 2, pp);
681    p = deposit32(p, 19, 4, ~v);
682    p = deposit32(p, 23, 1, (opc & P_VEXW) != 0);
683    p = deposit32(p, 29, 2, (opc & P_VEXL) != 0);
684
685    tcg_out32(s, p);
686    tcg_out8(s, opc);
687}
688
689static void tcg_out_vex_modrm(TCGContext *s, int opc, int r, int v, int rm)
690{
691    if (opc & P_EVEX) {
692        tcg_out_evex_opc(s, opc, r, v, rm, 0);
693    } else {
694        tcg_out_vex_opc(s, opc, r, v, rm, 0);
695    }
696    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
697}
698
699/* Output an opcode with a full "rm + (index<<shift) + offset" address mode.
700   We handle either RM and INDEX missing with a negative value.  In 64-bit
701   mode for absolute addresses, ~RM is the size of the immediate operand
702   that will follow the instruction.  */
703
704static void tcg_out_sib_offset(TCGContext *s, int r, int rm, int index,
705                               int shift, intptr_t offset)
706{
707    int mod, len;
708
709    if (index < 0 && rm < 0) {
710        if (TCG_TARGET_REG_BITS == 64) {
711            /* Try for a rip-relative addressing mode.  This has replaced
712               the 32-bit-mode absolute addressing encoding.  */
713            intptr_t pc = (intptr_t)s->code_ptr + 5 + ~rm;
714            intptr_t disp = offset - pc;
715            if (disp == (int32_t)disp) {
716                tcg_out8(s, (LOWREGMASK(r) << 3) | 5);
717                tcg_out32(s, disp);
718                return;
719            }
720
721            /* Try for an absolute address encoding.  This requires the
722               use of the MODRM+SIB encoding and is therefore larger than
723               rip-relative addressing.  */
724            if (offset == (int32_t)offset) {
725                tcg_out8(s, (LOWREGMASK(r) << 3) | 4);
726                tcg_out8(s, (4 << 3) | 5);
727                tcg_out32(s, offset);
728                return;
729            }
730
731            /* ??? The memory isn't directly addressable.  */
732            g_assert_not_reached();
733        } else {
734            /* Absolute address.  */
735            tcg_out8(s, (r << 3) | 5);
736            tcg_out32(s, offset);
737            return;
738        }
739    }
740
741    /* Find the length of the immediate addend.  Note that the encoding
742       that would be used for (%ebp) indicates absolute addressing.  */
743    if (rm < 0) {
744        mod = 0, len = 4, rm = 5;
745    } else if (offset == 0 && LOWREGMASK(rm) != TCG_REG_EBP) {
746        mod = 0, len = 0;
747    } else if (offset == (int8_t)offset) {
748        mod = 0x40, len = 1;
749    } else {
750        mod = 0x80, len = 4;
751    }
752
753    /* Use a single byte MODRM format if possible.  Note that the encoding
754       that would be used for %esp is the escape to the two byte form.  */
755    if (index < 0 && LOWREGMASK(rm) != TCG_REG_ESP) {
756        /* Single byte MODRM format.  */
757        tcg_out8(s, mod | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
758    } else {
759        /* Two byte MODRM+SIB format.  */
760
761        /* Note that the encoding that would place %esp into the index
762           field indicates no index register.  In 64-bit mode, the REX.X
763           bit counts, so %r12 can be used as the index.  */
764        if (index < 0) {
765            index = 4;
766        } else {
767            tcg_debug_assert(index != TCG_REG_ESP);
768        }
769
770        tcg_out8(s, mod | (LOWREGMASK(r) << 3) | 4);
771        tcg_out8(s, (shift << 6) | (LOWREGMASK(index) << 3) | LOWREGMASK(rm));
772    }
773
774    if (len == 1) {
775        tcg_out8(s, offset);
776    } else if (len == 4) {
777        tcg_out32(s, offset);
778    }
779}
780
781static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm,
782                                     int index, int shift, intptr_t offset)
783{
784    tcg_out_opc(s, opc, r, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
785    tcg_out_sib_offset(s, r, rm, index, shift, offset);
786}
787
788static void tcg_out_vex_modrm_sib_offset(TCGContext *s, int opc, int r, int v,
789                                         int rm, int index, int shift,
790                                         intptr_t offset)
791{
792    tcg_out_vex_opc(s, opc, r, v, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
793    tcg_out_sib_offset(s, r, rm, index, shift, offset);
794}
795
796/* A simplification of the above with no index or shift.  */
797static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r,
798                                        int rm, intptr_t offset)
799{
800    tcg_out_modrm_sib_offset(s, opc, r, rm, -1, 0, offset);
801}
802
803static inline void tcg_out_vex_modrm_offset(TCGContext *s, int opc, int r,
804                                            int v, int rm, intptr_t offset)
805{
806    tcg_out_vex_modrm_sib_offset(s, opc, r, v, rm, -1, 0, offset);
807}
808
809/* Output an opcode with an expected reference to the constant pool.  */
810static inline void tcg_out_modrm_pool(TCGContext *s, int opc, int r)
811{
812    tcg_out_opc(s, opc, r, 0, 0);
813    /* Absolute for 32-bit, pc-relative for 64-bit.  */
814    tcg_out8(s, LOWREGMASK(r) << 3 | 5);
815    tcg_out32(s, 0);
816}
817
818/* Output an opcode with an expected reference to the constant pool.  */
819static inline void tcg_out_vex_modrm_pool(TCGContext *s, int opc, int r)
820{
821    tcg_out_vex_opc(s, opc, r, 0, 0, 0);
822    /* Absolute for 32-bit, pc-relative for 64-bit.  */
823    tcg_out8(s, LOWREGMASK(r) << 3 | 5);
824    tcg_out32(s, 0);
825}
826
827/* Generate dest op= src.  Uses the same ARITH_* codes as tgen_arithi.  */
828static inline void tgen_arithr(TCGContext *s, int subop, int dest, int src)
829{
830    /* Propagate an opcode prefix, such as P_REXW.  */
831    int ext = subop & ~0x7;
832    subop &= 0x7;
833
834    tcg_out_modrm(s, OPC_ARITH_GvEv + (subop << 3) + ext, dest, src);
835}
836
837static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
838{
839    int rexw = 0;
840
841    if (arg == ret) {
842        return true;
843    }
844    switch (type) {
845    case TCG_TYPE_I64:
846        rexw = P_REXW;
847        /* fallthru */
848    case TCG_TYPE_I32:
849        if (ret < 16) {
850            if (arg < 16) {
851                tcg_out_modrm(s, OPC_MOVL_GvEv + rexw, ret, arg);
852            } else {
853                tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, arg, 0, ret);
854            }
855        } else {
856            if (arg < 16) {
857                tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, ret, 0, arg);
858            } else {
859                tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
860            }
861        }
862        break;
863
864    case TCG_TYPE_V64:
865        tcg_debug_assert(ret >= 16 && arg >= 16);
866        tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
867        break;
868    case TCG_TYPE_V128:
869        tcg_debug_assert(ret >= 16 && arg >= 16);
870        tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx, ret, 0, arg);
871        break;
872    case TCG_TYPE_V256:
873        tcg_debug_assert(ret >= 16 && arg >= 16);
874        tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx | P_VEXL, ret, 0, arg);
875        break;
876
877    default:
878        g_assert_not_reached();
879    }
880    return true;
881}
882
883static const int avx2_dup_insn[4] = {
884    OPC_VPBROADCASTB, OPC_VPBROADCASTW,
885    OPC_VPBROADCASTD, OPC_VPBROADCASTQ,
886};
887
888static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
889                            TCGReg r, TCGReg a)
890{
891    if (have_avx2) {
892        int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
893        tcg_out_vex_modrm(s, avx2_dup_insn[vece] + vex_l, r, 0, a);
894    } else {
895        switch (vece) {
896        case MO_8:
897            /* ??? With zero in a register, use PSHUFB.  */
898            tcg_out_vex_modrm(s, OPC_PUNPCKLBW, r, a, a);
899            a = r;
900            /* FALLTHRU */
901        case MO_16:
902            tcg_out_vex_modrm(s, OPC_PUNPCKLWD, r, a, a);
903            a = r;
904            /* FALLTHRU */
905        case MO_32:
906            tcg_out_vex_modrm(s, OPC_PSHUFD, r, 0, a);
907            /* imm8 operand: all output lanes selected from input lane 0.  */
908            tcg_out8(s, 0);
909            break;
910        case MO_64:
911            tcg_out_vex_modrm(s, OPC_PUNPCKLQDQ, r, a, a);
912            break;
913        default:
914            g_assert_not_reached();
915        }
916    }
917    return true;
918}
919
920static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
921                             TCGReg r, TCGReg base, intptr_t offset)
922{
923    if (have_avx2) {
924        int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
925        tcg_out_vex_modrm_offset(s, avx2_dup_insn[vece] + vex_l,
926                                 r, 0, base, offset);
927    } else {
928        switch (vece) {
929        case MO_64:
930            tcg_out_vex_modrm_offset(s, OPC_MOVDDUP, r, 0, base, offset);
931            break;
932        case MO_32:
933            tcg_out_vex_modrm_offset(s, OPC_VBROADCASTSS, r, 0, base, offset);
934            break;
935        case MO_16:
936            tcg_out_vex_modrm_offset(s, OPC_VPINSRW, r, r, base, offset);
937            tcg_out8(s, 0); /* imm8 */
938            tcg_out_dup_vec(s, type, vece, r, r);
939            break;
940        case MO_8:
941            tcg_out_vex_modrm_offset(s, OPC_VPINSRB, r, r, base, offset);
942            tcg_out8(s, 0); /* imm8 */
943            tcg_out_dup_vec(s, type, vece, r, r);
944            break;
945        default:
946            g_assert_not_reached();
947        }
948    }
949    return true;
950}
951
952static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
953                             TCGReg ret, int64_t arg)
954{
955    int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
956
957    if (arg == 0) {
958        tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
959        return;
960    }
961    if (arg == -1) {
962        tcg_out_vex_modrm(s, OPC_PCMPEQB + vex_l, ret, ret, ret);
963        return;
964    }
965
966    if (TCG_TARGET_REG_BITS == 32 && vece < MO_64) {
967        if (have_avx2) {
968            tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTD + vex_l, ret);
969        } else {
970            tcg_out_vex_modrm_pool(s, OPC_VBROADCASTSS, ret);
971        }
972        new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
973    } else {
974        if (type == TCG_TYPE_V64) {
975            tcg_out_vex_modrm_pool(s, OPC_MOVQ_VqWq, ret);
976        } else if (have_avx2) {
977            tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTQ + vex_l, ret);
978        } else {
979            tcg_out_vex_modrm_pool(s, OPC_MOVDDUP, ret);
980        }
981        if (TCG_TARGET_REG_BITS == 64) {
982            new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
983        } else {
984            new_pool_l2(s, R_386_32, s->code_ptr - 4, 0, arg, arg >> 32);
985        }
986    }
987}
988
989static void tcg_out_movi_vec(TCGContext *s, TCGType type,
990                             TCGReg ret, tcg_target_long arg)
991{
992    if (arg == 0) {
993        tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
994        return;
995    }
996    if (arg == -1) {
997        tcg_out_vex_modrm(s, OPC_PCMPEQB, ret, ret, ret);
998        return;
999    }
1000
1001    int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
1002    tcg_out_vex_modrm_pool(s, OPC_MOVD_VyEy + rexw, ret);
1003    if (TCG_TARGET_REG_BITS == 64) {
1004        new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
1005    } else {
1006        new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
1007    }
1008}
1009
1010static void tcg_out_movi_int(TCGContext *s, TCGType type,
1011                             TCGReg ret, tcg_target_long arg)
1012{
1013    tcg_target_long diff;
1014
1015    if (arg == 0) {
1016        tgen_arithr(s, ARITH_XOR, ret, ret);
1017        return;
1018    }
1019    if (arg == (uint32_t)arg || type == TCG_TYPE_I32) {
1020        tcg_out_opc(s, OPC_MOVL_Iv + LOWREGMASK(ret), 0, ret, 0);
1021        tcg_out32(s, arg);
1022        return;
1023    }
1024    if (arg == (int32_t)arg) {
1025        tcg_out_modrm(s, OPC_MOVL_EvIz + P_REXW, 0, ret);
1026        tcg_out32(s, arg);
1027        return;
1028    }
1029
1030    /* Try a 7 byte pc-relative lea before the 10 byte movq.  */
1031    diff = tcg_pcrel_diff(s, (const void *)arg) - 7;
1032    if (diff == (int32_t)diff) {
1033        tcg_out_opc(s, OPC_LEA | P_REXW, ret, 0, 0);
1034        tcg_out8(s, (LOWREGMASK(ret) << 3) | 5);
1035        tcg_out32(s, diff);
1036        return;
1037    }
1038
1039    tcg_out_opc(s, OPC_MOVL_Iv + P_REXW + LOWREGMASK(ret), 0, ret, 0);
1040    tcg_out64(s, arg);
1041}
1042
1043static void tcg_out_movi(TCGContext *s, TCGType type,
1044                         TCGReg ret, tcg_target_long arg)
1045{
1046    switch (type) {
1047    case TCG_TYPE_I32:
1048#if TCG_TARGET_REG_BITS == 64
1049    case TCG_TYPE_I64:
1050#endif
1051        if (ret < 16) {
1052            tcg_out_movi_int(s, type, ret, arg);
1053        } else {
1054            tcg_out_movi_vec(s, type, ret, arg);
1055        }
1056        break;
1057    default:
1058        g_assert_not_reached();
1059    }
1060}
1061
1062static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
1063{
1064    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1065    tcg_out_modrm(s, OPC_XCHG_EvGv + rexw, r1, r2);
1066    return true;
1067}
1068
1069static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
1070                             tcg_target_long imm)
1071{
1072    /* This function is only used for passing structs by reference. */
1073    tcg_debug_assert(imm == (int32_t)imm);
1074    tcg_out_modrm_offset(s, OPC_LEA | P_REXW, rd, rs, imm);
1075}
1076
1077static inline void tcg_out_pushi(TCGContext *s, tcg_target_long val)
1078{
1079    if (val == (int8_t)val) {
1080        tcg_out_opc(s, OPC_PUSH_Ib, 0, 0, 0);
1081        tcg_out8(s, val);
1082    } else if (val == (int32_t)val) {
1083        tcg_out_opc(s, OPC_PUSH_Iv, 0, 0, 0);
1084        tcg_out32(s, val);
1085    } else {
1086        g_assert_not_reached();
1087    }
1088}
1089
1090static inline void tcg_out_mb(TCGContext *s, TCGArg a0)
1091{
1092    /* Given the strength of x86 memory ordering, we only need care for
1093       store-load ordering.  Experimentally, "lock orl $0,0(%esp)" is
1094       faster than "mfence", so don't bother with the sse insn.  */
1095    if (a0 & TCG_MO_ST_LD) {
1096        tcg_out8(s, 0xf0);
1097        tcg_out_modrm_offset(s, OPC_ARITH_EvIb, ARITH_OR, TCG_REG_ESP, 0);
1098        tcg_out8(s, 0);
1099    }
1100}
1101
1102static inline void tcg_out_push(TCGContext *s, int reg)
1103{
1104    tcg_out_opc(s, OPC_PUSH_r32 + LOWREGMASK(reg), 0, reg, 0);
1105}
1106
1107static inline void tcg_out_pop(TCGContext *s, int reg)
1108{
1109    tcg_out_opc(s, OPC_POP_r32 + LOWREGMASK(reg), 0, reg, 0);
1110}
1111
1112static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret,
1113                       TCGReg arg1, intptr_t arg2)
1114{
1115    switch (type) {
1116    case TCG_TYPE_I32:
1117        if (ret < 16) {
1118            tcg_out_modrm_offset(s, OPC_MOVL_GvEv, ret, arg1, arg2);
1119        } else {
1120            tcg_out_vex_modrm_offset(s, OPC_MOVD_VyEy, ret, 0, arg1, arg2);
1121        }
1122        break;
1123    case TCG_TYPE_I64:
1124        if (ret < 16) {
1125            tcg_out_modrm_offset(s, OPC_MOVL_GvEv | P_REXW, ret, arg1, arg2);
1126            break;
1127        }
1128        /* FALLTHRU */
1129    case TCG_TYPE_V64:
1130        /* There is no instruction that can validate 8-byte alignment.  */
1131        tcg_debug_assert(ret >= 16);
1132        tcg_out_vex_modrm_offset(s, OPC_MOVQ_VqWq, ret, 0, arg1, arg2);
1133        break;
1134    case TCG_TYPE_V128:
1135        /*
1136         * The gvec infrastructure is asserts that v128 vector loads
1137         * and stores use a 16-byte aligned offset.  Validate that the
1138         * final pointer is aligned by using an insn that will SIGSEGV.
1139         */
1140        tcg_debug_assert(ret >= 16);
1141        tcg_out_vex_modrm_offset(s, OPC_MOVDQA_VxWx, ret, 0, arg1, arg2);
1142        break;
1143    case TCG_TYPE_V256:
1144        /*
1145         * The gvec infrastructure only requires 16-byte alignment,
1146         * so here we must use an unaligned load.
1147         */
1148        tcg_debug_assert(ret >= 16);
1149        tcg_out_vex_modrm_offset(s, OPC_MOVDQU_VxWx | P_VEXL,
1150                                 ret, 0, arg1, arg2);
1151        break;
1152    default:
1153        g_assert_not_reached();
1154    }
1155}
1156
1157static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
1158                       TCGReg arg1, intptr_t arg2)
1159{
1160    switch (type) {
1161    case TCG_TYPE_I32:
1162        if (arg < 16) {
1163            tcg_out_modrm_offset(s, OPC_MOVL_EvGv, arg, arg1, arg2);
1164        } else {
1165            tcg_out_vex_modrm_offset(s, OPC_MOVD_EyVy, arg, 0, arg1, arg2);
1166        }
1167        break;
1168    case TCG_TYPE_I64:
1169        if (arg < 16) {
1170            tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_REXW, arg, arg1, arg2);
1171            break;
1172        }
1173        /* FALLTHRU */
1174    case TCG_TYPE_V64:
1175        /* There is no instruction that can validate 8-byte alignment.  */
1176        tcg_debug_assert(arg >= 16);
1177        tcg_out_vex_modrm_offset(s, OPC_MOVQ_WqVq, arg, 0, arg1, arg2);
1178        break;
1179    case TCG_TYPE_V128:
1180        /*
1181         * The gvec infrastructure is asserts that v128 vector loads
1182         * and stores use a 16-byte aligned offset.  Validate that the
1183         * final pointer is aligned by using an insn that will SIGSEGV.
1184         *
1185         * This specific instance is also used by TCG_CALL_RET_BY_VEC,
1186         * for _WIN64, which must have SSE2 but may not have AVX.
1187         */
1188        tcg_debug_assert(arg >= 16);
1189        if (have_avx1) {
1190            tcg_out_vex_modrm_offset(s, OPC_MOVDQA_WxVx, arg, 0, arg1, arg2);
1191        } else {
1192            tcg_out_modrm_offset(s, OPC_MOVDQA_WxVx, arg, arg1, arg2);
1193        }
1194        break;
1195    case TCG_TYPE_V256:
1196        /*
1197         * The gvec infrastructure only requires 16-byte alignment,
1198         * so here we must use an unaligned store.
1199         */
1200        tcg_debug_assert(arg >= 16);
1201        tcg_out_vex_modrm_offset(s, OPC_MOVDQU_WxVx | P_VEXL,
1202                                 arg, 0, arg1, arg2);
1203        break;
1204    default:
1205        g_assert_not_reached();
1206    }
1207}
1208
1209static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
1210                        TCGReg base, intptr_t ofs)
1211{
1212    int rexw = 0;
1213    if (TCG_TARGET_REG_BITS == 64 && type == TCG_TYPE_I64) {
1214        if (val != (int32_t)val) {
1215            return false;
1216        }
1217        rexw = P_REXW;
1218    } else if (type != TCG_TYPE_I32) {
1219        return false;
1220    }
1221    tcg_out_modrm_offset(s, OPC_MOVL_EvIz | rexw, 0, base, ofs);
1222    tcg_out32(s, val);
1223    return true;
1224}
1225
1226static void tcg_out_shifti(TCGContext *s, int subopc, int reg, int count)
1227{
1228    /* Propagate an opcode prefix, such as P_DATA16.  */
1229    int ext = subopc & ~0x7;
1230    subopc &= 0x7;
1231
1232    if (count == 1) {
1233        tcg_out_modrm(s, OPC_SHIFT_1 + ext, subopc, reg);
1234    } else {
1235        tcg_out_modrm(s, OPC_SHIFT_Ib + ext, subopc, reg);
1236        tcg_out8(s, count);
1237    }
1238}
1239
1240static inline void tcg_out_bswap32(TCGContext *s, int reg)
1241{
1242    tcg_out_opc(s, OPC_BSWAP + LOWREGMASK(reg), 0, reg, 0);
1243}
1244
1245static inline void tcg_out_rolw_8(TCGContext *s, int reg)
1246{
1247    tcg_out_shifti(s, SHIFT_ROL + P_DATA16, reg, 8);
1248}
1249
1250static void tcg_out_ext8u(TCGContext *s, TCGReg dest, TCGReg src)
1251{
1252    /* movzbl */
1253    tcg_debug_assert(src < 4 || TCG_TARGET_REG_BITS == 64);
1254    tcg_out_modrm(s, OPC_MOVZBL + P_REXB_RM, dest, src);
1255}
1256
1257static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg dest, TCGReg src)
1258{
1259    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1260    /* movsbl */
1261    tcg_debug_assert(src < 4 || TCG_TARGET_REG_BITS == 64);
1262    tcg_out_modrm(s, OPC_MOVSBL + P_REXB_RM + rexw, dest, src);
1263}
1264
1265static void tcg_out_ext16u(TCGContext *s, TCGReg dest, TCGReg src)
1266{
1267    /* movzwl */
1268    tcg_out_modrm(s, OPC_MOVZWL, dest, src);
1269}
1270
1271static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg dest, TCGReg src)
1272{
1273    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1274    /* movsw[lq] */
1275    tcg_out_modrm(s, OPC_MOVSWL + rexw, dest, src);
1276}
1277
1278static void tcg_out_ext32u(TCGContext *s, TCGReg dest, TCGReg src)
1279{
1280    /* 32-bit mov zero extends.  */
1281    tcg_out_modrm(s, OPC_MOVL_GvEv, dest, src);
1282}
1283
1284static void tcg_out_ext32s(TCGContext *s, TCGReg dest, TCGReg src)
1285{
1286    tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
1287    tcg_out_modrm(s, OPC_MOVSLQ, dest, src);
1288}
1289
1290static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg dest, TCGReg src)
1291{
1292    tcg_out_ext32s(s, dest, src);
1293}
1294
1295static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg dest, TCGReg src)
1296{
1297    if (dest != src) {
1298        tcg_out_ext32u(s, dest, src);
1299    }
1300}
1301
1302static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg dest, TCGReg src)
1303{
1304    tcg_out_ext32u(s, dest, src);
1305}
1306
1307static inline void tcg_out_bswap64(TCGContext *s, int reg)
1308{
1309    tcg_out_opc(s, OPC_BSWAP + P_REXW + LOWREGMASK(reg), 0, reg, 0);
1310}
1311
1312static void tgen_arithi(TCGContext *s, int c, int r0,
1313                        tcg_target_long val, int cf)
1314{
1315    int rexw = 0;
1316
1317    if (TCG_TARGET_REG_BITS == 64) {
1318        rexw = c & -8;
1319        c &= 7;
1320    }
1321
1322    /* ??? While INC is 2 bytes shorter than ADDL $1, they also induce
1323       partial flags update stalls on Pentium4 and are not recommended
1324       by current Intel optimization manuals.  */
1325    if (!cf && (c == ARITH_ADD || c == ARITH_SUB) && (val == 1 || val == -1)) {
1326        int is_inc = (c == ARITH_ADD) ^ (val < 0);
1327        if (TCG_TARGET_REG_BITS == 64) {
1328            /* The single-byte increment encodings are re-tasked as the
1329               REX prefixes.  Use the MODRM encoding.  */
1330            tcg_out_modrm(s, OPC_GRP5 + rexw,
1331                          (is_inc ? EXT5_INC_Ev : EXT5_DEC_Ev), r0);
1332        } else {
1333            tcg_out8(s, (is_inc ? OPC_INC_r32 : OPC_DEC_r32) + r0);
1334        }
1335        return;
1336    }
1337
1338    if (c == ARITH_AND) {
1339        if (TCG_TARGET_REG_BITS == 64) {
1340            if (val == 0xffffffffu) {
1341                tcg_out_ext32u(s, r0, r0);
1342                return;
1343            }
1344            if (val == (uint32_t)val) {
1345                /* AND with no high bits set can use a 32-bit operation.  */
1346                rexw = 0;
1347            }
1348        }
1349        if (val == 0xffu && (r0 < 4 || TCG_TARGET_REG_BITS == 64)) {
1350            tcg_out_ext8u(s, r0, r0);
1351            return;
1352        }
1353        if (val == 0xffffu) {
1354            tcg_out_ext16u(s, r0, r0);
1355            return;
1356        }
1357    }
1358
1359    if (val == (int8_t)val) {
1360        tcg_out_modrm(s, OPC_ARITH_EvIb + rexw, c, r0);
1361        tcg_out8(s, val);
1362        return;
1363    }
1364    if (rexw == 0 || val == (int32_t)val) {
1365        tcg_out_modrm(s, OPC_ARITH_EvIz + rexw, c, r0);
1366        tcg_out32(s, val);
1367        return;
1368    }
1369
1370    g_assert_not_reached();
1371}
1372
1373static void tcg_out_addi(TCGContext *s, int reg, tcg_target_long val)
1374{
1375    if (val != 0) {
1376        tgen_arithi(s, ARITH_ADD + P_REXW, reg, val, 0);
1377    }
1378}
1379
1380/* Set SMALL to force a short forward branch.  */
1381static void tcg_out_jxx(TCGContext *s, int opc, TCGLabel *l, bool small)
1382{
1383    int32_t val, val1;
1384
1385    if (l->has_value) {
1386        val = tcg_pcrel_diff(s, l->u.value_ptr);
1387        val1 = val - 2;
1388        if ((int8_t)val1 == val1) {
1389            if (opc == -1) {
1390                tcg_out8(s, OPC_JMP_short);
1391            } else {
1392                tcg_out8(s, OPC_JCC_short + opc);
1393            }
1394            tcg_out8(s, val1);
1395        } else {
1396            tcg_debug_assert(!small);
1397            if (opc == -1) {
1398                tcg_out8(s, OPC_JMP_long);
1399                tcg_out32(s, val - 5);
1400            } else {
1401                tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1402                tcg_out32(s, val - 6);
1403            }
1404        }
1405    } else if (small) {
1406        if (opc == -1) {
1407            tcg_out8(s, OPC_JMP_short);
1408        } else {
1409            tcg_out8(s, OPC_JCC_short + opc);
1410        }
1411        tcg_out_reloc(s, s->code_ptr, R_386_PC8, l, -1);
1412        s->code_ptr += 1;
1413    } else {
1414        if (opc == -1) {
1415            tcg_out8(s, OPC_JMP_long);
1416        } else {
1417            tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1418        }
1419        tcg_out_reloc(s, s->code_ptr, R_386_PC32, l, -4);
1420        s->code_ptr += 4;
1421    }
1422}
1423
1424static void tcg_out_cmp(TCGContext *s, TCGArg arg1, TCGArg arg2,
1425                        int const_arg2, int rexw)
1426{
1427    if (const_arg2) {
1428        if (arg2 == 0) {
1429            /* test r, r */
1430            tcg_out_modrm(s, OPC_TESTL + rexw, arg1, arg1);
1431        } else {
1432            tgen_arithi(s, ARITH_CMP + rexw, arg1, arg2, 0);
1433        }
1434    } else {
1435        tgen_arithr(s, ARITH_CMP + rexw, arg1, arg2);
1436    }
1437}
1438
1439static void tcg_out_brcond32(TCGContext *s, TCGCond cond,
1440                             TCGArg arg1, TCGArg arg2, int const_arg2,
1441                             TCGLabel *label, int small)
1442{
1443    tcg_out_cmp(s, arg1, arg2, const_arg2, 0);
1444    tcg_out_jxx(s, tcg_cond_to_jcc[cond], label, small);
1445}
1446
1447#if TCG_TARGET_REG_BITS == 64
1448static void tcg_out_brcond64(TCGContext *s, TCGCond cond,
1449                             TCGArg arg1, TCGArg arg2, int const_arg2,
1450                             TCGLabel *label, int small)
1451{
1452    tcg_out_cmp(s, arg1, arg2, const_arg2, P_REXW);
1453    tcg_out_jxx(s, tcg_cond_to_jcc[cond], label, small);
1454}
1455#else
1456/* XXX: we implement it at the target level to avoid having to
1457   handle cross basic blocks temporaries */
1458static void tcg_out_brcond2(TCGContext *s, const TCGArg *args,
1459                            const int *const_args, int small)
1460{
1461    TCGLabel *label_next = gen_new_label();
1462    TCGLabel *label_this = arg_label(args[5]);
1463
1464    switch(args[4]) {
1465    case TCG_COND_EQ:
1466        tcg_out_brcond32(s, TCG_COND_NE, args[0], args[2], const_args[2],
1467                         label_next, 1);
1468        tcg_out_brcond32(s, TCG_COND_EQ, args[1], args[3], const_args[3],
1469                         label_this, small);
1470        break;
1471    case TCG_COND_NE:
1472        tcg_out_brcond32(s, TCG_COND_NE, args[0], args[2], const_args[2],
1473                         label_this, small);
1474        tcg_out_brcond32(s, TCG_COND_NE, args[1], args[3], const_args[3],
1475                         label_this, small);
1476        break;
1477    case TCG_COND_LT:
1478        tcg_out_brcond32(s, TCG_COND_LT, args[1], args[3], const_args[3],
1479                         label_this, small);
1480        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1481        tcg_out_brcond32(s, TCG_COND_LTU, args[0], args[2], const_args[2],
1482                         label_this, small);
1483        break;
1484    case TCG_COND_LE:
1485        tcg_out_brcond32(s, TCG_COND_LT, args[1], args[3], const_args[3],
1486                         label_this, small);
1487        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1488        tcg_out_brcond32(s, TCG_COND_LEU, args[0], args[2], const_args[2],
1489                         label_this, small);
1490        break;
1491    case TCG_COND_GT:
1492        tcg_out_brcond32(s, TCG_COND_GT, args[1], args[3], const_args[3],
1493                         label_this, small);
1494        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1495        tcg_out_brcond32(s, TCG_COND_GTU, args[0], args[2], const_args[2],
1496                         label_this, small);
1497        break;
1498    case TCG_COND_GE:
1499        tcg_out_brcond32(s, TCG_COND_GT, args[1], args[3], const_args[3],
1500                         label_this, small);
1501        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1502        tcg_out_brcond32(s, TCG_COND_GEU, args[0], args[2], const_args[2],
1503                         label_this, small);
1504        break;
1505    case TCG_COND_LTU:
1506        tcg_out_brcond32(s, TCG_COND_LTU, args[1], args[3], const_args[3],
1507                         label_this, small);
1508        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1509        tcg_out_brcond32(s, TCG_COND_LTU, args[0], args[2], const_args[2],
1510                         label_this, small);
1511        break;
1512    case TCG_COND_LEU:
1513        tcg_out_brcond32(s, TCG_COND_LTU, args[1], args[3], const_args[3],
1514                         label_this, small);
1515        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1516        tcg_out_brcond32(s, TCG_COND_LEU, args[0], args[2], const_args[2],
1517                         label_this, small);
1518        break;
1519    case TCG_COND_GTU:
1520        tcg_out_brcond32(s, TCG_COND_GTU, args[1], args[3], const_args[3],
1521                         label_this, small);
1522        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1523        tcg_out_brcond32(s, TCG_COND_GTU, args[0], args[2], const_args[2],
1524                         label_this, small);
1525        break;
1526    case TCG_COND_GEU:
1527        tcg_out_brcond32(s, TCG_COND_GTU, args[1], args[3], const_args[3],
1528                         label_this, small);
1529        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1530        tcg_out_brcond32(s, TCG_COND_GEU, args[0], args[2], const_args[2],
1531                         label_this, small);
1532        break;
1533    default:
1534        g_assert_not_reached();
1535    }
1536    tcg_out_label(s, label_next);
1537}
1538#endif
1539
1540static void tcg_out_setcond32(TCGContext *s, TCGCond cond, TCGArg dest,
1541                              TCGArg arg1, TCGArg arg2, int const_arg2)
1542{
1543    tcg_out_cmp(s, arg1, arg2, const_arg2, 0);
1544    tcg_out_modrm(s, OPC_SETCC | tcg_cond_to_jcc[cond], 0, dest);
1545    tcg_out_ext8u(s, dest, dest);
1546}
1547
1548#if TCG_TARGET_REG_BITS == 64
1549static void tcg_out_setcond64(TCGContext *s, TCGCond cond, TCGArg dest,
1550                              TCGArg arg1, TCGArg arg2, int const_arg2)
1551{
1552    tcg_out_cmp(s, arg1, arg2, const_arg2, P_REXW);
1553    tcg_out_modrm(s, OPC_SETCC | tcg_cond_to_jcc[cond], 0, dest);
1554    tcg_out_ext8u(s, dest, dest);
1555}
1556#else
1557static void tcg_out_setcond2(TCGContext *s, const TCGArg *args,
1558                             const int *const_args)
1559{
1560    TCGArg new_args[6];
1561    TCGLabel *label_true, *label_over;
1562
1563    memcpy(new_args, args+1, 5*sizeof(TCGArg));
1564
1565    if (args[0] == args[1] || args[0] == args[2]
1566        || (!const_args[3] && args[0] == args[3])
1567        || (!const_args[4] && args[0] == args[4])) {
1568        /* When the destination overlaps with one of the argument
1569           registers, don't do anything tricky.  */
1570        label_true = gen_new_label();
1571        label_over = gen_new_label();
1572
1573        new_args[5] = label_arg(label_true);
1574        tcg_out_brcond2(s, new_args, const_args+1, 1);
1575
1576        tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1577        tcg_out_jxx(s, JCC_JMP, label_over, 1);
1578        tcg_out_label(s, label_true);
1579
1580        tcg_out_movi(s, TCG_TYPE_I32, args[0], 1);
1581        tcg_out_label(s, label_over);
1582    } else {
1583        /* When the destination does not overlap one of the arguments,
1584           clear the destination first, jump if cond false, and emit an
1585           increment in the true case.  This results in smaller code.  */
1586
1587        tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1588
1589        label_over = gen_new_label();
1590        new_args[4] = tcg_invert_cond(new_args[4]);
1591        new_args[5] = label_arg(label_over);
1592        tcg_out_brcond2(s, new_args, const_args+1, 1);
1593
1594        tgen_arithi(s, ARITH_ADD, args[0], 1, 0);
1595        tcg_out_label(s, label_over);
1596    }
1597}
1598#endif
1599
1600static void tcg_out_cmov(TCGContext *s, TCGCond cond, int rexw,
1601                         TCGReg dest, TCGReg v1)
1602{
1603    if (have_cmov) {
1604        tcg_out_modrm(s, OPC_CMOVCC | tcg_cond_to_jcc[cond] | rexw, dest, v1);
1605    } else {
1606        TCGLabel *over = gen_new_label();
1607        tcg_out_jxx(s, tcg_cond_to_jcc[tcg_invert_cond(cond)], over, 1);
1608        tcg_out_mov(s, TCG_TYPE_I32, dest, v1);
1609        tcg_out_label(s, over);
1610    }
1611}
1612
1613static void tcg_out_movcond32(TCGContext *s, TCGCond cond, TCGReg dest,
1614                              TCGReg c1, TCGArg c2, int const_c2,
1615                              TCGReg v1)
1616{
1617    tcg_out_cmp(s, c1, c2, const_c2, 0);
1618    tcg_out_cmov(s, cond, 0, dest, v1);
1619}
1620
1621#if TCG_TARGET_REG_BITS == 64
1622static void tcg_out_movcond64(TCGContext *s, TCGCond cond, TCGReg dest,
1623                              TCGReg c1, TCGArg c2, int const_c2,
1624                              TCGReg v1)
1625{
1626    tcg_out_cmp(s, c1, c2, const_c2, P_REXW);
1627    tcg_out_cmov(s, cond, P_REXW, dest, v1);
1628}
1629#endif
1630
1631static void tcg_out_ctz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1632                        TCGArg arg2, bool const_a2)
1633{
1634    if (have_bmi1) {
1635        tcg_out_modrm(s, OPC_TZCNT + rexw, dest, arg1);
1636        if (const_a2) {
1637            tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1638        } else {
1639            tcg_debug_assert(dest != arg2);
1640            tcg_out_cmov(s, TCG_COND_LTU, rexw, dest, arg2);
1641        }
1642    } else {
1643        tcg_debug_assert(dest != arg2);
1644        tcg_out_modrm(s, OPC_BSF + rexw, dest, arg1);
1645        tcg_out_cmov(s, TCG_COND_EQ, rexw, dest, arg2);
1646    }
1647}
1648
1649static void tcg_out_clz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1650                        TCGArg arg2, bool const_a2)
1651{
1652    if (have_lzcnt) {
1653        tcg_out_modrm(s, OPC_LZCNT + rexw, dest, arg1);
1654        if (const_a2) {
1655            tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1656        } else {
1657            tcg_debug_assert(dest != arg2);
1658            tcg_out_cmov(s, TCG_COND_LTU, rexw, dest, arg2);
1659        }
1660    } else {
1661        tcg_debug_assert(!const_a2);
1662        tcg_debug_assert(dest != arg1);
1663        tcg_debug_assert(dest != arg2);
1664
1665        /* Recall that the output of BSR is the index not the count.  */
1666        tcg_out_modrm(s, OPC_BSR + rexw, dest, arg1);
1667        tgen_arithi(s, ARITH_XOR + rexw, dest, rexw ? 63 : 31, 0);
1668
1669        /* Since we have destroyed the flags from BSR, we have to re-test.  */
1670        tcg_out_cmp(s, arg1, 0, 1, rexw);
1671        tcg_out_cmov(s, TCG_COND_EQ, rexw, dest, arg2);
1672    }
1673}
1674
1675static void tcg_out_branch(TCGContext *s, int call, const tcg_insn_unit *dest)
1676{
1677    intptr_t disp = tcg_pcrel_diff(s, dest) - 5;
1678
1679    if (disp == (int32_t)disp) {
1680        tcg_out_opc(s, call ? OPC_CALL_Jz : OPC_JMP_long, 0, 0, 0);
1681        tcg_out32(s, disp);
1682    } else {
1683        /* rip-relative addressing into the constant pool.
1684           This is 6 + 8 = 14 bytes, as compared to using an
1685           immediate load 10 + 6 = 16 bytes, plus we may
1686           be able to re-use the pool constant for more calls.  */
1687        tcg_out_opc(s, OPC_GRP5, 0, 0, 0);
1688        tcg_out8(s, (call ? EXT5_CALLN_Ev : EXT5_JMPN_Ev) << 3 | 5);
1689        new_pool_label(s, (uintptr_t)dest, R_386_PC32, s->code_ptr, -4);
1690        tcg_out32(s, 0);
1691    }
1692}
1693
1694static void tcg_out_call(TCGContext *s, const tcg_insn_unit *dest,
1695                         const TCGHelperInfo *info)
1696{
1697    tcg_out_branch(s, 1, dest);
1698
1699#ifndef _WIN32
1700    if (TCG_TARGET_REG_BITS == 32 && info->out_kind == TCG_CALL_RET_BY_REF) {
1701        /*
1702         * The sysv i386 abi for struct return places a reference as the
1703         * first argument of the stack, and pops that argument with the
1704         * return statement.  Since we want to retain the aligned stack
1705         * pointer for the callee, we do not want to actually push that
1706         * argument before the call but rely on the normal store to the
1707         * stack slot.  But we do need to compensate for the pop in order
1708         * to reset our correct stack pointer value.
1709         * Pushing a garbage value back onto the stack is quickest.
1710         */
1711        tcg_out_push(s, TCG_REG_EAX);
1712    }
1713#endif
1714}
1715
1716static void tcg_out_jmp(TCGContext *s, const tcg_insn_unit *dest)
1717{
1718    tcg_out_branch(s, 0, dest);
1719}
1720
1721static void tcg_out_nopn(TCGContext *s, int n)
1722{
1723    int i;
1724    /* Emit 1 or 2 operand size prefixes for the standard one byte nop,
1725     * "xchg %eax,%eax", forming "xchg %ax,%ax". All cores accept the
1726     * duplicate prefix, and all of the interesting recent cores can
1727     * decode and discard the duplicates in a single cycle.
1728     */
1729    tcg_debug_assert(n >= 1);
1730    for (i = 1; i < n; ++i) {
1731        tcg_out8(s, 0x66);
1732    }
1733    tcg_out8(s, 0x90);
1734}
1735
1736/* Test register R vs immediate bits I, setting Z flag for EQ/NE. */
1737static void __attribute__((unused))
1738tcg_out_testi(TCGContext *s, TCGReg r, uint32_t i)
1739{
1740    /*
1741     * This is used for testing alignment, so we can usually use testb.
1742     * For i686, we have to use testl for %esi/%edi.
1743     */
1744    if (i <= 0xff && (TCG_TARGET_REG_BITS == 64 || r < 4)) {
1745        tcg_out_modrm(s, OPC_GRP3_Eb | P_REXB_RM, EXT3_TESTi, r);
1746        tcg_out8(s, i);
1747    } else {
1748        tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_TESTi, r);
1749        tcg_out32(s, i);
1750    }
1751}
1752
1753typedef struct {
1754    TCGReg base;
1755    int index;
1756    int ofs;
1757    int seg;
1758    TCGAtomAlign aa;
1759} HostAddress;
1760
1761bool tcg_target_has_memory_bswap(MemOp memop)
1762{
1763    TCGAtomAlign aa;
1764
1765    if (!have_movbe) {
1766        return false;
1767    }
1768    if ((memop & MO_SIZE) < MO_128) {
1769        return true;
1770    }
1771
1772    /*
1773     * Reject 16-byte memop with 16-byte atomicity, i.e. VMOVDQA,
1774     * but do allow a pair of 64-bit operations, i.e. MOVBEQ.
1775     */
1776    aa = atom_and_align_for_opc(tcg_ctx, memop, MO_ATOM_IFALIGN, true);
1777    return aa.atom < MO_128;
1778}
1779
1780/*
1781 * Because i686 has no register parameters and because x86_64 has xchg
1782 * to handle addr/data register overlap, we have placed all input arguments
1783 * before we need might need a scratch reg.
1784 *
1785 * Even then, a scratch is only needed for l->raddr.  Rather than expose
1786 * a general-purpose scratch when we don't actually know it's available,
1787 * use the ra_gen hook to load into RAX if needed.
1788 */
1789#if TCG_TARGET_REG_BITS == 64
1790static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
1791{
1792    if (arg < 0) {
1793        arg = TCG_REG_RAX;
1794    }
1795    tcg_out_movi(s, TCG_TYPE_PTR, arg, (uintptr_t)l->raddr);
1796    return arg;
1797}
1798static const TCGLdstHelperParam ldst_helper_param = {
1799    .ra_gen = ldst_ra_gen
1800};
1801#else
1802static const TCGLdstHelperParam ldst_helper_param = { };
1803#endif
1804
1805static void tcg_out_vec_to_pair(TCGContext *s, TCGType type,
1806                                TCGReg l, TCGReg h, TCGReg v)
1807{
1808    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1809
1810    /* vpmov{d,q} %v, %l */
1811    tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, v, 0, l);
1812    /* vpextr{d,q} $1, %v, %h */
1813    tcg_out_vex_modrm(s, OPC_PEXTRD + rexw, v, 0, h);
1814    tcg_out8(s, 1);
1815}
1816
1817static void tcg_out_pair_to_vec(TCGContext *s, TCGType type,
1818                                TCGReg v, TCGReg l, TCGReg h)
1819{
1820    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1821
1822    /* vmov{d,q} %l, %v */
1823    tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, v, 0, l);
1824    /* vpinsr{d,q} $1, %h, %v, %v */
1825    tcg_out_vex_modrm(s, OPC_PINSRD + rexw, v, v, h);
1826    tcg_out8(s, 1);
1827}
1828
1829/*
1830 * Generate code for the slow path for a load at the end of block
1831 */
1832static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
1833{
1834    MemOp opc = get_memop(l->oi);
1835    tcg_insn_unit **label_ptr = &l->label_ptr[0];
1836
1837    /* resolve label address */
1838    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
1839    if (label_ptr[1]) {
1840        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
1841    }
1842
1843    tcg_out_ld_helper_args(s, l, &ldst_helper_param);
1844    tcg_out_branch(s, 1, qemu_ld_helpers[opc & MO_SIZE]);
1845    tcg_out_ld_helper_ret(s, l, false, &ldst_helper_param);
1846
1847    tcg_out_jmp(s, l->raddr);
1848    return true;
1849}
1850
1851/*
1852 * Generate code for the slow path for a store at the end of block
1853 */
1854static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
1855{
1856    MemOp opc = get_memop(l->oi);
1857    tcg_insn_unit **label_ptr = &l->label_ptr[0];
1858
1859    /* resolve label address */
1860    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
1861    if (label_ptr[1]) {
1862        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
1863    }
1864
1865    tcg_out_st_helper_args(s, l, &ldst_helper_param);
1866    tcg_out_branch(s, 1, qemu_st_helpers[opc & MO_SIZE]);
1867
1868    tcg_out_jmp(s, l->raddr);
1869    return true;
1870}
1871
1872#ifndef CONFIG_SOFTMMU
1873static HostAddress x86_guest_base = {
1874    .index = -1
1875};
1876
1877#if defined(__x86_64__) && defined(__linux__)
1878# include <asm/prctl.h>
1879# include <sys/prctl.h>
1880int arch_prctl(int code, unsigned long addr);
1881static inline int setup_guest_base_seg(void)
1882{
1883    if (arch_prctl(ARCH_SET_GS, guest_base) == 0) {
1884        return P_GS;
1885    }
1886    return 0;
1887}
1888#elif defined(__x86_64__) && \
1889      (defined (__FreeBSD__) || defined (__FreeBSD_kernel__))
1890# include <machine/sysarch.h>
1891static inline int setup_guest_base_seg(void)
1892{
1893    if (sysarch(AMD64_SET_GSBASE, &guest_base) == 0) {
1894        return P_GS;
1895    }
1896    return 0;
1897}
1898#else
1899static inline int setup_guest_base_seg(void)
1900{
1901    return 0;
1902}
1903#endif /* setup_guest_base_seg */
1904#endif /* !SOFTMMU */
1905
1906#define MIN_TLB_MASK_TABLE_OFS  INT_MIN
1907
1908/*
1909 * For softmmu, perform the TLB load and compare.
1910 * For useronly, perform any required alignment tests.
1911 * In both cases, return a TCGLabelQemuLdst structure if the slow path
1912 * is required and fill in @h with the host address for the fast path.
1913 */
1914static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
1915                                           TCGReg addrlo, TCGReg addrhi,
1916                                           MemOpIdx oi, bool is_ld)
1917{
1918    TCGLabelQemuLdst *ldst = NULL;
1919    MemOp opc = get_memop(oi);
1920    MemOp s_bits = opc & MO_SIZE;
1921    unsigned a_mask;
1922
1923#ifdef CONFIG_SOFTMMU
1924    h->index = TCG_REG_L0;
1925    h->ofs = 0;
1926    h->seg = 0;
1927#else
1928    *h = x86_guest_base;
1929#endif
1930    h->base = addrlo;
1931    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, s_bits == MO_128);
1932    a_mask = (1 << h->aa.align) - 1;
1933
1934#ifdef CONFIG_SOFTMMU
1935    int cmp_ofs = is_ld ? offsetof(CPUTLBEntry, addr_read)
1936                        : offsetof(CPUTLBEntry, addr_write);
1937    TCGType ttype = TCG_TYPE_I32;
1938    TCGType tlbtype = TCG_TYPE_I32;
1939    int trexw = 0, hrexw = 0, tlbrexw = 0;
1940    unsigned mem_index = get_mmuidx(oi);
1941    unsigned s_mask = (1 << s_bits) - 1;
1942    int fast_ofs = tlb_mask_table_ofs(s, mem_index);
1943    int tlb_mask;
1944
1945    ldst = new_ldst_label(s);
1946    ldst->is_ld = is_ld;
1947    ldst->oi = oi;
1948    ldst->addrlo_reg = addrlo;
1949    ldst->addrhi_reg = addrhi;
1950
1951    if (TCG_TARGET_REG_BITS == 64) {
1952        ttype = s->addr_type;
1953        trexw = (ttype == TCG_TYPE_I32 ? 0 : P_REXW);
1954        if (TCG_TYPE_PTR == TCG_TYPE_I64) {
1955            hrexw = P_REXW;
1956            if (s->page_bits + s->tlb_dyn_max_bits > 32) {
1957                tlbtype = TCG_TYPE_I64;
1958                tlbrexw = P_REXW;
1959            }
1960        }
1961    }
1962
1963    tcg_out_mov(s, tlbtype, TCG_REG_L0, addrlo);
1964    tcg_out_shifti(s, SHIFT_SHR + tlbrexw, TCG_REG_L0,
1965                   s->page_bits - CPU_TLB_ENTRY_BITS);
1966
1967    tcg_out_modrm_offset(s, OPC_AND_GvEv + trexw, TCG_REG_L0, TCG_AREG0,
1968                         fast_ofs + offsetof(CPUTLBDescFast, mask));
1969
1970    tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, TCG_REG_L0, TCG_AREG0,
1971                         fast_ofs + offsetof(CPUTLBDescFast, table));
1972
1973    /*
1974     * If the required alignment is at least as large as the access, simply
1975     * copy the address and mask.  For lesser alignments, check that we don't
1976     * cross pages for the complete access.
1977     */
1978    if (a_mask >= s_mask) {
1979        tcg_out_mov(s, ttype, TCG_REG_L1, addrlo);
1980    } else {
1981        tcg_out_modrm_offset(s, OPC_LEA + trexw, TCG_REG_L1,
1982                             addrlo, s_mask - a_mask);
1983    }
1984    tlb_mask = s->page_mask | a_mask;
1985    tgen_arithi(s, ARITH_AND + trexw, TCG_REG_L1, tlb_mask, 0);
1986
1987    /* cmp 0(TCG_REG_L0), TCG_REG_L1 */
1988    tcg_out_modrm_offset(s, OPC_CMP_GvEv + trexw,
1989                         TCG_REG_L1, TCG_REG_L0, cmp_ofs);
1990
1991    /* jne slow_path */
1992    tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
1993    ldst->label_ptr[0] = s->code_ptr;
1994    s->code_ptr += 4;
1995
1996    if (TCG_TARGET_REG_BITS == 32 && s->addr_type == TCG_TYPE_I64) {
1997        /* cmp 4(TCG_REG_L0), addrhi */
1998        tcg_out_modrm_offset(s, OPC_CMP_GvEv, addrhi, TCG_REG_L0, cmp_ofs + 4);
1999
2000        /* jne slow_path */
2001        tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
2002        ldst->label_ptr[1] = s->code_ptr;
2003        s->code_ptr += 4;
2004    }
2005
2006    /* TLB Hit.  */
2007    tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_L0, TCG_REG_L0,
2008               offsetof(CPUTLBEntry, addend));
2009#else
2010    if (a_mask) {
2011        ldst = new_ldst_label(s);
2012
2013        ldst->is_ld = is_ld;
2014        ldst->oi = oi;
2015        ldst->addrlo_reg = addrlo;
2016        ldst->addrhi_reg = addrhi;
2017
2018        tcg_out_testi(s, addrlo, a_mask);
2019        /* jne slow_path */
2020        tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
2021        ldst->label_ptr[0] = s->code_ptr;
2022        s->code_ptr += 4;
2023    }
2024#endif
2025
2026    return ldst;
2027}
2028
2029static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2030                                   HostAddress h, TCGType type, MemOp memop)
2031{
2032    bool use_movbe = false;
2033    int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
2034    int movop = OPC_MOVL_GvEv;
2035
2036    /* Do big-endian loads with movbe.  */
2037    if (memop & MO_BSWAP) {
2038        tcg_debug_assert(have_movbe);
2039        use_movbe = true;
2040        movop = OPC_MOVBE_GyMy;
2041    }
2042
2043    switch (memop & MO_SSIZE) {
2044    case MO_UB:
2045        tcg_out_modrm_sib_offset(s, OPC_MOVZBL + h.seg, datalo,
2046                                 h.base, h.index, 0, h.ofs);
2047        break;
2048    case MO_SB:
2049        tcg_out_modrm_sib_offset(s, OPC_MOVSBL + rexw + h.seg, datalo,
2050                                 h.base, h.index, 0, h.ofs);
2051        break;
2052    case MO_UW:
2053        if (use_movbe) {
2054            /* There is no extending movbe; only low 16-bits are modified.  */
2055            if (datalo != h.base && datalo != h.index) {
2056                /* XOR breaks dependency chains.  */
2057                tgen_arithr(s, ARITH_XOR, datalo, datalo);
2058                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2059                                         datalo, h.base, h.index, 0, h.ofs);
2060            } else {
2061                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2062                                         datalo, h.base, h.index, 0, h.ofs);
2063                tcg_out_ext16u(s, datalo, datalo);
2064            }
2065        } else {
2066            tcg_out_modrm_sib_offset(s, OPC_MOVZWL + h.seg, datalo,
2067                                     h.base, h.index, 0, h.ofs);
2068        }
2069        break;
2070    case MO_SW:
2071        if (use_movbe) {
2072            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2073                                     datalo, h.base, h.index, 0, h.ofs);
2074            tcg_out_ext16s(s, type, datalo, datalo);
2075        } else {
2076            tcg_out_modrm_sib_offset(s, OPC_MOVSWL + rexw + h.seg,
2077                                     datalo, h.base, h.index, 0, h.ofs);
2078        }
2079        break;
2080    case MO_UL:
2081        tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2082                                 h.base, h.index, 0, h.ofs);
2083        break;
2084#if TCG_TARGET_REG_BITS == 64
2085    case MO_SL:
2086        if (use_movbe) {
2087            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + h.seg, datalo,
2088                                     h.base, h.index, 0, h.ofs);
2089            tcg_out_ext32s(s, datalo, datalo);
2090        } else {
2091            tcg_out_modrm_sib_offset(s, OPC_MOVSLQ + h.seg, datalo,
2092                                     h.base, h.index, 0, h.ofs);
2093        }
2094        break;
2095#endif
2096    case MO_UQ:
2097        if (TCG_TARGET_REG_BITS == 64) {
2098            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2099                                     h.base, h.index, 0, h.ofs);
2100            break;
2101        }
2102        if (use_movbe) {
2103            TCGReg t = datalo;
2104            datalo = datahi;
2105            datahi = t;
2106        }
2107        if (h.base == datalo || h.index == datalo) {
2108            tcg_out_modrm_sib_offset(s, OPC_LEA, datahi,
2109                                     h.base, h.index, 0, h.ofs);
2110            tcg_out_modrm_offset(s, movop + h.seg, datalo, datahi, 0);
2111            tcg_out_modrm_offset(s, movop + h.seg, datahi, datahi, 4);
2112        } else {
2113            tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2114                                     h.base, h.index, 0, h.ofs);
2115            tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2116                                     h.base, h.index, 0, h.ofs + 4);
2117        }
2118        break;
2119
2120    case MO_128:
2121        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2122
2123        /*
2124         * Without 16-byte atomicity, use integer regs.
2125         * That is where we want the data, and it allows bswaps.
2126         */
2127        if (h.aa.atom < MO_128) {
2128            if (use_movbe) {
2129                TCGReg t = datalo;
2130                datalo = datahi;
2131                datahi = t;
2132            }
2133            if (h.base == datalo || h.index == datalo) {
2134                tcg_out_modrm_sib_offset(s, OPC_LEA + P_REXW, datahi,
2135                                         h.base, h.index, 0, h.ofs);
2136                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2137                                     datalo, datahi, 0);
2138                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2139                                     datahi, datahi, 8);
2140            } else {
2141                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2142                                         h.base, h.index, 0, h.ofs);
2143                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2144                                         h.base, h.index, 0, h.ofs + 8);
2145            }
2146            break;
2147        }
2148
2149        /*
2150         * With 16-byte atomicity, a vector load is required.
2151         * If we already have 16-byte alignment, then VMOVDQA always works.
2152         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2153         * Else use we require a runtime test for alignment for VMOVDQA;
2154         * use VMOVDQU on the unaligned nonatomic path for simplicity.
2155         */
2156        if (h.aa.align >= MO_128) {
2157            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2158                                         TCG_TMP_VEC, 0,
2159                                         h.base, h.index, 0, h.ofs);
2160        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2161            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2162                                         TCG_TMP_VEC, 0,
2163                                         h.base, h.index, 0, h.ofs);
2164        } else {
2165            TCGLabel *l1 = gen_new_label();
2166            TCGLabel *l2 = gen_new_label();
2167
2168            tcg_out_testi(s, h.base, 15);
2169            tcg_out_jxx(s, JCC_JNE, l1, true);
2170
2171            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2172                                         TCG_TMP_VEC, 0,
2173                                         h.base, h.index, 0, h.ofs);
2174            tcg_out_jxx(s, JCC_JMP, l2, true);
2175
2176            tcg_out_label(s, l1);
2177            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2178                                         TCG_TMP_VEC, 0,
2179                                         h.base, h.index, 0, h.ofs);
2180            tcg_out_label(s, l2);
2181        }
2182        tcg_out_vec_to_pair(s, TCG_TYPE_I64, datalo, datahi, TCG_TMP_VEC);
2183        break;
2184
2185    default:
2186        g_assert_not_reached();
2187    }
2188}
2189
2190static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
2191                            TCGReg addrlo, TCGReg addrhi,
2192                            MemOpIdx oi, TCGType data_type)
2193{
2194    TCGLabelQemuLdst *ldst;
2195    HostAddress h;
2196
2197    ldst = prepare_host_addr(s, &h, addrlo, addrhi, oi, true);
2198    tcg_out_qemu_ld_direct(s, datalo, datahi, h, data_type, get_memop(oi));
2199
2200    if (ldst) {
2201        ldst->type = data_type;
2202        ldst->datalo_reg = datalo;
2203        ldst->datahi_reg = datahi;
2204        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2205    }
2206}
2207
2208static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2209                                   HostAddress h, MemOp memop)
2210{
2211    bool use_movbe = false;
2212    int movop = OPC_MOVL_EvGv;
2213
2214    /*
2215     * Do big-endian stores with movbe or softmmu.
2216     * User-only without movbe will have its swapping done generically.
2217     */
2218    if (memop & MO_BSWAP) {
2219        tcg_debug_assert(have_movbe);
2220        use_movbe = true;
2221        movop = OPC_MOVBE_MyGy;
2222    }
2223
2224    switch (memop & MO_SIZE) {
2225    case MO_8:
2226        /* This is handled with constraints on INDEX_op_qemu_st8_i32. */
2227        tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || datalo < 4);
2228        tcg_out_modrm_sib_offset(s, OPC_MOVB_EvGv + P_REXB_R + h.seg,
2229                                 datalo, h.base, h.index, 0, h.ofs);
2230        break;
2231    case MO_16:
2232        tcg_out_modrm_sib_offset(s, movop + P_DATA16 + h.seg, datalo,
2233                                 h.base, h.index, 0, h.ofs);
2234        break;
2235    case MO_32:
2236        tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2237                                 h.base, h.index, 0, h.ofs);
2238        break;
2239    case MO_64:
2240        if (TCG_TARGET_REG_BITS == 64) {
2241            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2242                                     h.base, h.index, 0, h.ofs);
2243        } else {
2244            if (use_movbe) {
2245                TCGReg t = datalo;
2246                datalo = datahi;
2247                datahi = t;
2248            }
2249            tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2250                                     h.base, h.index, 0, h.ofs);
2251            tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2252                                     h.base, h.index, 0, h.ofs + 4);
2253        }
2254        break;
2255
2256    case MO_128:
2257        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2258
2259        /*
2260         * Without 16-byte atomicity, use integer regs.
2261         * That is where we have the data, and it allows bswaps.
2262         */
2263        if (h.aa.atom < MO_128) {
2264            if (use_movbe) {
2265                TCGReg t = datalo;
2266                datalo = datahi;
2267                datahi = t;
2268            }
2269            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2270                                     h.base, h.index, 0, h.ofs);
2271            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2272                                     h.base, h.index, 0, h.ofs + 8);
2273            break;
2274        }
2275
2276        /*
2277         * With 16-byte atomicity, a vector store is required.
2278         * If we already have 16-byte alignment, then VMOVDQA always works.
2279         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2280         * Else use we require a runtime test for alignment for VMOVDQA;
2281         * use VMOVDQU on the unaligned nonatomic path for simplicity.
2282         */
2283        tcg_out_pair_to_vec(s, TCG_TYPE_I64, TCG_TMP_VEC, datalo, datahi);
2284        if (h.aa.align >= MO_128) {
2285            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2286                                         TCG_TMP_VEC, 0,
2287                                         h.base, h.index, 0, h.ofs);
2288        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2289            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2290                                         TCG_TMP_VEC, 0,
2291                                         h.base, h.index, 0, h.ofs);
2292        } else {
2293            TCGLabel *l1 = gen_new_label();
2294            TCGLabel *l2 = gen_new_label();
2295
2296            tcg_out_testi(s, h.base, 15);
2297            tcg_out_jxx(s, JCC_JNE, l1, true);
2298
2299            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2300                                         TCG_TMP_VEC, 0,
2301                                         h.base, h.index, 0, h.ofs);
2302            tcg_out_jxx(s, JCC_JMP, l2, true);
2303
2304            tcg_out_label(s, l1);
2305            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2306                                         TCG_TMP_VEC, 0,
2307                                         h.base, h.index, 0, h.ofs);
2308            tcg_out_label(s, l2);
2309        }
2310        break;
2311
2312    default:
2313        g_assert_not_reached();
2314    }
2315}
2316
2317static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
2318                            TCGReg addrlo, TCGReg addrhi,
2319                            MemOpIdx oi, TCGType data_type)
2320{
2321    TCGLabelQemuLdst *ldst;
2322    HostAddress h;
2323
2324    ldst = prepare_host_addr(s, &h, addrlo, addrhi, oi, false);
2325    tcg_out_qemu_st_direct(s, datalo, datahi, h, get_memop(oi));
2326
2327    if (ldst) {
2328        ldst->type = data_type;
2329        ldst->datalo_reg = datalo;
2330        ldst->datahi_reg = datahi;
2331        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2332    }
2333}
2334
2335static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
2336{
2337    /* Reuse the zeroing that exists for goto_ptr.  */
2338    if (a0 == 0) {
2339        tcg_out_jmp(s, tcg_code_gen_epilogue);
2340    } else {
2341        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_EAX, a0);
2342        tcg_out_jmp(s, tb_ret_addr);
2343    }
2344}
2345
2346static void tcg_out_goto_tb(TCGContext *s, int which)
2347{
2348    /*
2349     * Jump displacement must be aligned for atomic patching;
2350     * see if we need to add extra nops before jump
2351     */
2352    int gap = QEMU_ALIGN_PTR_UP(s->code_ptr + 1, 4) - s->code_ptr;
2353    if (gap != 1) {
2354        tcg_out_nopn(s, gap - 1);
2355    }
2356    tcg_out8(s, OPC_JMP_long); /* jmp im */
2357    set_jmp_insn_offset(s, which);
2358    tcg_out32(s, 0);
2359    set_jmp_reset_offset(s, which);
2360}
2361
2362void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
2363                              uintptr_t jmp_rx, uintptr_t jmp_rw)
2364{
2365    /* patch the branch destination */
2366    uintptr_t addr = tb->jmp_target_addr[n];
2367    qatomic_set((int32_t *)jmp_rw, addr - (jmp_rx + 4));
2368    /* no need to flush icache explicitly */
2369}
2370
2371static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
2372                              const TCGArg args[TCG_MAX_OP_ARGS],
2373                              const int const_args[TCG_MAX_OP_ARGS])
2374{
2375    TCGArg a0, a1, a2;
2376    int c, const_a2, vexop, rexw = 0;
2377
2378#if TCG_TARGET_REG_BITS == 64
2379# define OP_32_64(x) \
2380        case glue(glue(INDEX_op_, x), _i64): \
2381            rexw = P_REXW; /* FALLTHRU */    \
2382        case glue(glue(INDEX_op_, x), _i32)
2383#else
2384# define OP_32_64(x) \
2385        case glue(glue(INDEX_op_, x), _i32)
2386#endif
2387
2388    /* Hoist the loads of the most common arguments.  */
2389    a0 = args[0];
2390    a1 = args[1];
2391    a2 = args[2];
2392    const_a2 = const_args[2];
2393
2394    switch (opc) {
2395    case INDEX_op_goto_ptr:
2396        /* jmp to the given host address (could be epilogue) */
2397        tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, a0);
2398        break;
2399    case INDEX_op_br:
2400        tcg_out_jxx(s, JCC_JMP, arg_label(a0), 0);
2401        break;
2402    OP_32_64(ld8u):
2403        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
2404        tcg_out_modrm_offset(s, OPC_MOVZBL, a0, a1, a2);
2405        break;
2406    OP_32_64(ld8s):
2407        tcg_out_modrm_offset(s, OPC_MOVSBL + rexw, a0, a1, a2);
2408        break;
2409    OP_32_64(ld16u):
2410        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
2411        tcg_out_modrm_offset(s, OPC_MOVZWL, a0, a1, a2);
2412        break;
2413    OP_32_64(ld16s):
2414        tcg_out_modrm_offset(s, OPC_MOVSWL + rexw, a0, a1, a2);
2415        break;
2416#if TCG_TARGET_REG_BITS == 64
2417    case INDEX_op_ld32u_i64:
2418#endif
2419    case INDEX_op_ld_i32:
2420        tcg_out_ld(s, TCG_TYPE_I32, a0, a1, a2);
2421        break;
2422
2423    OP_32_64(st8):
2424        if (const_args[0]) {
2425            tcg_out_modrm_offset(s, OPC_MOVB_EvIz, 0, a1, a2);
2426            tcg_out8(s, a0);
2427        } else {
2428            tcg_out_modrm_offset(s, OPC_MOVB_EvGv | P_REXB_R, a0, a1, a2);
2429        }
2430        break;
2431    OP_32_64(st16):
2432        if (const_args[0]) {
2433            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_DATA16, 0, a1, a2);
2434            tcg_out16(s, a0);
2435        } else {
2436            tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_DATA16, a0, a1, a2);
2437        }
2438        break;
2439#if TCG_TARGET_REG_BITS == 64
2440    case INDEX_op_st32_i64:
2441#endif
2442    case INDEX_op_st_i32:
2443        if (const_args[0]) {
2444            tcg_out_modrm_offset(s, OPC_MOVL_EvIz, 0, a1, a2);
2445            tcg_out32(s, a0);
2446        } else {
2447            tcg_out_st(s, TCG_TYPE_I32, a0, a1, a2);
2448        }
2449        break;
2450
2451    OP_32_64(add):
2452        /* For 3-operand addition, use LEA.  */
2453        if (a0 != a1) {
2454            TCGArg c3 = 0;
2455            if (const_a2) {
2456                c3 = a2, a2 = -1;
2457            } else if (a0 == a2) {
2458                /* Watch out for dest = src + dest, since we've removed
2459                   the matching constraint on the add.  */
2460                tgen_arithr(s, ARITH_ADD + rexw, a0, a1);
2461                break;
2462            }
2463
2464            tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a2, 0, c3);
2465            break;
2466        }
2467        c = ARITH_ADD;
2468        goto gen_arith;
2469    OP_32_64(sub):
2470        c = ARITH_SUB;
2471        goto gen_arith;
2472    OP_32_64(and):
2473        c = ARITH_AND;
2474        goto gen_arith;
2475    OP_32_64(or):
2476        c = ARITH_OR;
2477        goto gen_arith;
2478    OP_32_64(xor):
2479        c = ARITH_XOR;
2480        goto gen_arith;
2481    gen_arith:
2482        if (const_a2) {
2483            tgen_arithi(s, c + rexw, a0, a2, 0);
2484        } else {
2485            tgen_arithr(s, c + rexw, a0, a2);
2486        }
2487        break;
2488
2489    OP_32_64(andc):
2490        if (const_a2) {
2491            tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, a0, a1);
2492            tgen_arithi(s, ARITH_AND + rexw, a0, ~a2, 0);
2493        } else {
2494            tcg_out_vex_modrm(s, OPC_ANDN + rexw, a0, a2, a1);
2495        }
2496        break;
2497
2498    OP_32_64(mul):
2499        if (const_a2) {
2500            int32_t val;
2501            val = a2;
2502            if (val == (int8_t)val) {
2503                tcg_out_modrm(s, OPC_IMUL_GvEvIb + rexw, a0, a0);
2504                tcg_out8(s, val);
2505            } else {
2506                tcg_out_modrm(s, OPC_IMUL_GvEvIz + rexw, a0, a0);
2507                tcg_out32(s, val);
2508            }
2509        } else {
2510            tcg_out_modrm(s, OPC_IMUL_GvEv + rexw, a0, a2);
2511        }
2512        break;
2513
2514    OP_32_64(div2):
2515        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IDIV, args[4]);
2516        break;
2517    OP_32_64(divu2):
2518        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_DIV, args[4]);
2519        break;
2520
2521    OP_32_64(shl):
2522        /* For small constant 3-operand shift, use LEA.  */
2523        if (const_a2 && a0 != a1 && (a2 - 1) < 3) {
2524            if (a2 - 1 == 0) {
2525                /* shl $1,a1,a0 -> lea (a1,a1),a0 */
2526                tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a1, 0, 0);
2527            } else {
2528                /* shl $n,a1,a0 -> lea 0(,a1,n),a0 */
2529                tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, -1, a1, a2, 0);
2530            }
2531            break;
2532        }
2533        c = SHIFT_SHL;
2534        vexop = OPC_SHLX;
2535        goto gen_shift_maybe_vex;
2536    OP_32_64(shr):
2537        c = SHIFT_SHR;
2538        vexop = OPC_SHRX;
2539        goto gen_shift_maybe_vex;
2540    OP_32_64(sar):
2541        c = SHIFT_SAR;
2542        vexop = OPC_SARX;
2543        goto gen_shift_maybe_vex;
2544    OP_32_64(rotl):
2545        c = SHIFT_ROL;
2546        goto gen_shift;
2547    OP_32_64(rotr):
2548        c = SHIFT_ROR;
2549        goto gen_shift;
2550    gen_shift_maybe_vex:
2551        if (have_bmi2) {
2552            if (!const_a2) {
2553                tcg_out_vex_modrm(s, vexop + rexw, a0, a2, a1);
2554                break;
2555            }
2556            tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, a0, a1);
2557        }
2558        /* FALLTHRU */
2559    gen_shift:
2560        if (const_a2) {
2561            tcg_out_shifti(s, c + rexw, a0, a2);
2562        } else {
2563            tcg_out_modrm(s, OPC_SHIFT_cl + rexw, c, a0);
2564        }
2565        break;
2566
2567    OP_32_64(ctz):
2568        tcg_out_ctz(s, rexw, args[0], args[1], args[2], const_args[2]);
2569        break;
2570    OP_32_64(clz):
2571        tcg_out_clz(s, rexw, args[0], args[1], args[2], const_args[2]);
2572        break;
2573    OP_32_64(ctpop):
2574        tcg_out_modrm(s, OPC_POPCNT + rexw, a0, a1);
2575        break;
2576
2577    case INDEX_op_brcond_i32:
2578        tcg_out_brcond32(s, a2, a0, a1, const_args[1], arg_label(args[3]), 0);
2579        break;
2580    case INDEX_op_setcond_i32:
2581        tcg_out_setcond32(s, args[3], a0, a1, a2, const_a2);
2582        break;
2583    case INDEX_op_movcond_i32:
2584        tcg_out_movcond32(s, args[5], a0, a1, a2, const_a2, args[3]);
2585        break;
2586
2587    OP_32_64(bswap16):
2588        if (a2 & TCG_BSWAP_OS) {
2589            /* Output must be sign-extended. */
2590            if (rexw) {
2591                tcg_out_bswap64(s, a0);
2592                tcg_out_shifti(s, SHIFT_SAR + rexw, a0, 48);
2593            } else {
2594                tcg_out_bswap32(s, a0);
2595                tcg_out_shifti(s, SHIFT_SAR, a0, 16);
2596            }
2597        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
2598            /* Output must be zero-extended, but input isn't. */
2599            tcg_out_bswap32(s, a0);
2600            tcg_out_shifti(s, SHIFT_SHR, a0, 16);
2601        } else {
2602            tcg_out_rolw_8(s, a0);
2603        }
2604        break;
2605    OP_32_64(bswap32):
2606        tcg_out_bswap32(s, a0);
2607        if (rexw && (a2 & TCG_BSWAP_OS)) {
2608            tcg_out_ext32s(s, a0, a0);
2609        }
2610        break;
2611
2612    OP_32_64(neg):
2613        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, a0);
2614        break;
2615    OP_32_64(not):
2616        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, a0);
2617        break;
2618
2619    case INDEX_op_qemu_ld_a64_i32:
2620        if (TCG_TARGET_REG_BITS == 32) {
2621            tcg_out_qemu_ld(s, a0, -1, a1, a2, args[3], TCG_TYPE_I32);
2622            break;
2623        }
2624        /* fall through */
2625    case INDEX_op_qemu_ld_a32_i32:
2626        tcg_out_qemu_ld(s, a0, -1, a1, -1, a2, TCG_TYPE_I32);
2627        break;
2628    case INDEX_op_qemu_ld_a32_i64:
2629        if (TCG_TARGET_REG_BITS == 64) {
2630            tcg_out_qemu_ld(s, a0, -1, a1, -1, a2, TCG_TYPE_I64);
2631        } else {
2632            tcg_out_qemu_ld(s, a0, a1, a2, -1, args[3], TCG_TYPE_I64);
2633        }
2634        break;
2635    case INDEX_op_qemu_ld_a64_i64:
2636        if (TCG_TARGET_REG_BITS == 64) {
2637            tcg_out_qemu_ld(s, a0, -1, a1, -1, a2, TCG_TYPE_I64);
2638        } else {
2639            tcg_out_qemu_ld(s, a0, a1, a2, args[3], args[4], TCG_TYPE_I64);
2640        }
2641        break;
2642    case INDEX_op_qemu_ld_a32_i128:
2643    case INDEX_op_qemu_ld_a64_i128:
2644        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2645        tcg_out_qemu_ld(s, a0, a1, a2, -1, args[3], TCG_TYPE_I128);
2646        break;
2647
2648    case INDEX_op_qemu_st_a64_i32:
2649    case INDEX_op_qemu_st8_a64_i32:
2650        if (TCG_TARGET_REG_BITS == 32) {
2651            tcg_out_qemu_st(s, a0, -1, a1, a2, args[3], TCG_TYPE_I32);
2652            break;
2653        }
2654        /* fall through */
2655    case INDEX_op_qemu_st_a32_i32:
2656    case INDEX_op_qemu_st8_a32_i32:
2657        tcg_out_qemu_st(s, a0, -1, a1, -1, a2, TCG_TYPE_I32);
2658        break;
2659    case INDEX_op_qemu_st_a32_i64:
2660        if (TCG_TARGET_REG_BITS == 64) {
2661            tcg_out_qemu_st(s, a0, -1, a1, -1, a2, TCG_TYPE_I64);
2662        } else {
2663            tcg_out_qemu_st(s, a0, a1, a2, -1, args[3], TCG_TYPE_I64);
2664        }
2665        break;
2666    case INDEX_op_qemu_st_a64_i64:
2667        if (TCG_TARGET_REG_BITS == 64) {
2668            tcg_out_qemu_st(s, a0, -1, a1, -1, a2, TCG_TYPE_I64);
2669        } else {
2670            tcg_out_qemu_st(s, a0, a1, a2, args[3], args[4], TCG_TYPE_I64);
2671        }
2672        break;
2673    case INDEX_op_qemu_st_a32_i128:
2674    case INDEX_op_qemu_st_a64_i128:
2675        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2676        tcg_out_qemu_st(s, a0, a1, a2, -1, args[3], TCG_TYPE_I128);
2677        break;
2678
2679    OP_32_64(mulu2):
2680        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_MUL, args[3]);
2681        break;
2682    OP_32_64(muls2):
2683        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IMUL, args[3]);
2684        break;
2685    OP_32_64(add2):
2686        if (const_args[4]) {
2687            tgen_arithi(s, ARITH_ADD + rexw, a0, args[4], 1);
2688        } else {
2689            tgen_arithr(s, ARITH_ADD + rexw, a0, args[4]);
2690        }
2691        if (const_args[5]) {
2692            tgen_arithi(s, ARITH_ADC + rexw, a1, args[5], 1);
2693        } else {
2694            tgen_arithr(s, ARITH_ADC + rexw, a1, args[5]);
2695        }
2696        break;
2697    OP_32_64(sub2):
2698        if (const_args[4]) {
2699            tgen_arithi(s, ARITH_SUB + rexw, a0, args[4], 1);
2700        } else {
2701            tgen_arithr(s, ARITH_SUB + rexw, a0, args[4]);
2702        }
2703        if (const_args[5]) {
2704            tgen_arithi(s, ARITH_SBB + rexw, a1, args[5], 1);
2705        } else {
2706            tgen_arithr(s, ARITH_SBB + rexw, a1, args[5]);
2707        }
2708        break;
2709
2710#if TCG_TARGET_REG_BITS == 32
2711    case INDEX_op_brcond2_i32:
2712        tcg_out_brcond2(s, args, const_args, 0);
2713        break;
2714    case INDEX_op_setcond2_i32:
2715        tcg_out_setcond2(s, args, const_args);
2716        break;
2717#else /* TCG_TARGET_REG_BITS == 64 */
2718    case INDEX_op_ld32s_i64:
2719        tcg_out_modrm_offset(s, OPC_MOVSLQ, a0, a1, a2);
2720        break;
2721    case INDEX_op_ld_i64:
2722        tcg_out_ld(s, TCG_TYPE_I64, a0, a1, a2);
2723        break;
2724    case INDEX_op_st_i64:
2725        if (const_args[0]) {
2726            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_REXW, 0, a1, a2);
2727            tcg_out32(s, a0);
2728        } else {
2729            tcg_out_st(s, TCG_TYPE_I64, a0, a1, a2);
2730        }
2731        break;
2732
2733    case INDEX_op_brcond_i64:
2734        tcg_out_brcond64(s, a2, a0, a1, const_args[1], arg_label(args[3]), 0);
2735        break;
2736    case INDEX_op_setcond_i64:
2737        tcg_out_setcond64(s, args[3], a0, a1, a2, const_a2);
2738        break;
2739    case INDEX_op_movcond_i64:
2740        tcg_out_movcond64(s, args[5], a0, a1, a2, const_a2, args[3]);
2741        break;
2742
2743    case INDEX_op_bswap64_i64:
2744        tcg_out_bswap64(s, a0);
2745        break;
2746    case INDEX_op_extrh_i64_i32:
2747        tcg_out_shifti(s, SHIFT_SHR + P_REXW, a0, 32);
2748        break;
2749#endif
2750
2751    OP_32_64(deposit):
2752        if (args[3] == 0 && args[4] == 8) {
2753            /* load bits 0..7 */
2754            tcg_out_modrm(s, OPC_MOVB_EvGv | P_REXB_R | P_REXB_RM, a2, a0);
2755        } else if (args[3] == 8 && args[4] == 8) {
2756            /* load bits 8..15 */
2757            tcg_out_modrm(s, OPC_MOVB_EvGv, a2, a0 + 4);
2758        } else if (args[3] == 0 && args[4] == 16) {
2759            /* load bits 0..15 */
2760            tcg_out_modrm(s, OPC_MOVL_EvGv | P_DATA16, a2, a0);
2761        } else {
2762            g_assert_not_reached();
2763        }
2764        break;
2765
2766    case INDEX_op_extract_i64:
2767        if (a2 + args[3] == 32) {
2768            /* This is a 32-bit zero-extending right shift.  */
2769            tcg_out_mov(s, TCG_TYPE_I32, a0, a1);
2770            tcg_out_shifti(s, SHIFT_SHR, a0, a2);
2771            break;
2772        }
2773        /* FALLTHRU */
2774    case INDEX_op_extract_i32:
2775        /* On the off-chance that we can use the high-byte registers.
2776           Otherwise we emit the same ext16 + shift pattern that we
2777           would have gotten from the normal tcg-op.c expansion.  */
2778        tcg_debug_assert(a2 == 8 && args[3] == 8);
2779        if (a1 < 4 && a0 < 8) {
2780            tcg_out_modrm(s, OPC_MOVZBL, a0, a1 + 4);
2781        } else {
2782            tcg_out_ext16u(s, a0, a1);
2783            tcg_out_shifti(s, SHIFT_SHR, a0, 8);
2784        }
2785        break;
2786
2787    case INDEX_op_sextract_i32:
2788        /* We don't implement sextract_i64, as we cannot sign-extend to
2789           64-bits without using the REX prefix that explicitly excludes
2790           access to the high-byte registers.  */
2791        tcg_debug_assert(a2 == 8 && args[3] == 8);
2792        if (a1 < 4 && a0 < 8) {
2793            tcg_out_modrm(s, OPC_MOVSBL, a0, a1 + 4);
2794        } else {
2795            tcg_out_ext16s(s, TCG_TYPE_I32, a0, a1);
2796            tcg_out_shifti(s, SHIFT_SAR, a0, 8);
2797        }
2798        break;
2799
2800    OP_32_64(extract2):
2801        /* Note that SHRD outputs to the r/m operand.  */
2802        tcg_out_modrm(s, OPC_SHRD_Ib + rexw, a2, a0);
2803        tcg_out8(s, args[3]);
2804        break;
2805
2806    case INDEX_op_mb:
2807        tcg_out_mb(s, a0);
2808        break;
2809    case INDEX_op_mov_i32:  /* Always emitted via tcg_out_mov.  */
2810    case INDEX_op_mov_i64:
2811    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
2812    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
2813    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
2814    case INDEX_op_ext8s_i32:  /* Always emitted via tcg_reg_alloc_op.  */
2815    case INDEX_op_ext8s_i64:
2816    case INDEX_op_ext8u_i32:
2817    case INDEX_op_ext8u_i64:
2818    case INDEX_op_ext16s_i32:
2819    case INDEX_op_ext16s_i64:
2820    case INDEX_op_ext16u_i32:
2821    case INDEX_op_ext16u_i64:
2822    case INDEX_op_ext32s_i64:
2823    case INDEX_op_ext32u_i64:
2824    case INDEX_op_ext_i32_i64:
2825    case INDEX_op_extu_i32_i64:
2826    case INDEX_op_extrl_i64_i32:
2827    default:
2828        g_assert_not_reached();
2829    }
2830
2831#undef OP_32_64
2832}
2833
2834static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2835                           unsigned vecl, unsigned vece,
2836                           const TCGArg args[TCG_MAX_OP_ARGS],
2837                           const int const_args[TCG_MAX_OP_ARGS])
2838{
2839    static int const add_insn[4] = {
2840        OPC_PADDB, OPC_PADDW, OPC_PADDD, OPC_PADDQ
2841    };
2842    static int const ssadd_insn[4] = {
2843        OPC_PADDSB, OPC_PADDSW, OPC_UD2, OPC_UD2
2844    };
2845    static int const usadd_insn[4] = {
2846        OPC_PADDUB, OPC_PADDUW, OPC_UD2, OPC_UD2
2847    };
2848    static int const sub_insn[4] = {
2849        OPC_PSUBB, OPC_PSUBW, OPC_PSUBD, OPC_PSUBQ
2850    };
2851    static int const sssub_insn[4] = {
2852        OPC_PSUBSB, OPC_PSUBSW, OPC_UD2, OPC_UD2
2853    };
2854    static int const ussub_insn[4] = {
2855        OPC_PSUBUB, OPC_PSUBUW, OPC_UD2, OPC_UD2
2856    };
2857    static int const mul_insn[4] = {
2858        OPC_UD2, OPC_PMULLW, OPC_PMULLD, OPC_VPMULLQ
2859    };
2860    static int const shift_imm_insn[4] = {
2861        OPC_UD2, OPC_PSHIFTW_Ib, OPC_PSHIFTD_Ib, OPC_PSHIFTQ_Ib
2862    };
2863    static int const cmpeq_insn[4] = {
2864        OPC_PCMPEQB, OPC_PCMPEQW, OPC_PCMPEQD, OPC_PCMPEQQ
2865    };
2866    static int const cmpgt_insn[4] = {
2867        OPC_PCMPGTB, OPC_PCMPGTW, OPC_PCMPGTD, OPC_PCMPGTQ
2868    };
2869    static int const punpckl_insn[4] = {
2870        OPC_PUNPCKLBW, OPC_PUNPCKLWD, OPC_PUNPCKLDQ, OPC_PUNPCKLQDQ
2871    };
2872    static int const punpckh_insn[4] = {
2873        OPC_PUNPCKHBW, OPC_PUNPCKHWD, OPC_PUNPCKHDQ, OPC_PUNPCKHQDQ
2874    };
2875    static int const packss_insn[4] = {
2876        OPC_PACKSSWB, OPC_PACKSSDW, OPC_UD2, OPC_UD2
2877    };
2878    static int const packus_insn[4] = {
2879        OPC_PACKUSWB, OPC_PACKUSDW, OPC_UD2, OPC_UD2
2880    };
2881    static int const smin_insn[4] = {
2882        OPC_PMINSB, OPC_PMINSW, OPC_PMINSD, OPC_VPMINSQ
2883    };
2884    static int const smax_insn[4] = {
2885        OPC_PMAXSB, OPC_PMAXSW, OPC_PMAXSD, OPC_VPMAXSQ
2886    };
2887    static int const umin_insn[4] = {
2888        OPC_PMINUB, OPC_PMINUW, OPC_PMINUD, OPC_VPMINUQ
2889    };
2890    static int const umax_insn[4] = {
2891        OPC_PMAXUB, OPC_PMAXUW, OPC_PMAXUD, OPC_VPMAXUQ
2892    };
2893    static int const rotlv_insn[4] = {
2894        OPC_UD2, OPC_UD2, OPC_VPROLVD, OPC_VPROLVQ
2895    };
2896    static int const rotrv_insn[4] = {
2897        OPC_UD2, OPC_UD2, OPC_VPRORVD, OPC_VPRORVQ
2898    };
2899    static int const shlv_insn[4] = {
2900        OPC_UD2, OPC_VPSLLVW, OPC_VPSLLVD, OPC_VPSLLVQ
2901    };
2902    static int const shrv_insn[4] = {
2903        OPC_UD2, OPC_VPSRLVW, OPC_VPSRLVD, OPC_VPSRLVQ
2904    };
2905    static int const sarv_insn[4] = {
2906        OPC_UD2, OPC_VPSRAVW, OPC_VPSRAVD, OPC_VPSRAVQ
2907    };
2908    static int const shls_insn[4] = {
2909        OPC_UD2, OPC_PSLLW, OPC_PSLLD, OPC_PSLLQ
2910    };
2911    static int const shrs_insn[4] = {
2912        OPC_UD2, OPC_PSRLW, OPC_PSRLD, OPC_PSRLQ
2913    };
2914    static int const sars_insn[4] = {
2915        OPC_UD2, OPC_PSRAW, OPC_PSRAD, OPC_VPSRAQ
2916    };
2917    static int const vpshldi_insn[4] = {
2918        OPC_UD2, OPC_VPSHLDW, OPC_VPSHLDD, OPC_VPSHLDQ
2919    };
2920    static int const vpshldv_insn[4] = {
2921        OPC_UD2, OPC_VPSHLDVW, OPC_VPSHLDVD, OPC_VPSHLDVQ
2922    };
2923    static int const vpshrdv_insn[4] = {
2924        OPC_UD2, OPC_VPSHRDVW, OPC_VPSHRDVD, OPC_VPSHRDVQ
2925    };
2926    static int const abs_insn[4] = {
2927        OPC_PABSB, OPC_PABSW, OPC_PABSD, OPC_VPABSQ
2928    };
2929
2930    TCGType type = vecl + TCG_TYPE_V64;
2931    int insn, sub;
2932    TCGArg a0, a1, a2, a3;
2933
2934    a0 = args[0];
2935    a1 = args[1];
2936    a2 = args[2];
2937
2938    switch (opc) {
2939    case INDEX_op_add_vec:
2940        insn = add_insn[vece];
2941        goto gen_simd;
2942    case INDEX_op_ssadd_vec:
2943        insn = ssadd_insn[vece];
2944        goto gen_simd;
2945    case INDEX_op_usadd_vec:
2946        insn = usadd_insn[vece];
2947        goto gen_simd;
2948    case INDEX_op_sub_vec:
2949        insn = sub_insn[vece];
2950        goto gen_simd;
2951    case INDEX_op_sssub_vec:
2952        insn = sssub_insn[vece];
2953        goto gen_simd;
2954    case INDEX_op_ussub_vec:
2955        insn = ussub_insn[vece];
2956        goto gen_simd;
2957    case INDEX_op_mul_vec:
2958        insn = mul_insn[vece];
2959        goto gen_simd;
2960    case INDEX_op_and_vec:
2961        insn = OPC_PAND;
2962        goto gen_simd;
2963    case INDEX_op_or_vec:
2964        insn = OPC_POR;
2965        goto gen_simd;
2966    case INDEX_op_xor_vec:
2967        insn = OPC_PXOR;
2968        goto gen_simd;
2969    case INDEX_op_smin_vec:
2970        insn = smin_insn[vece];
2971        goto gen_simd;
2972    case INDEX_op_umin_vec:
2973        insn = umin_insn[vece];
2974        goto gen_simd;
2975    case INDEX_op_smax_vec:
2976        insn = smax_insn[vece];
2977        goto gen_simd;
2978    case INDEX_op_umax_vec:
2979        insn = umax_insn[vece];
2980        goto gen_simd;
2981    case INDEX_op_shlv_vec:
2982        insn = shlv_insn[vece];
2983        goto gen_simd;
2984    case INDEX_op_shrv_vec:
2985        insn = shrv_insn[vece];
2986        goto gen_simd;
2987    case INDEX_op_sarv_vec:
2988        insn = sarv_insn[vece];
2989        goto gen_simd;
2990    case INDEX_op_rotlv_vec:
2991        insn = rotlv_insn[vece];
2992        goto gen_simd;
2993    case INDEX_op_rotrv_vec:
2994        insn = rotrv_insn[vece];
2995        goto gen_simd;
2996    case INDEX_op_shls_vec:
2997        insn = shls_insn[vece];
2998        goto gen_simd;
2999    case INDEX_op_shrs_vec:
3000        insn = shrs_insn[vece];
3001        goto gen_simd;
3002    case INDEX_op_sars_vec:
3003        insn = sars_insn[vece];
3004        goto gen_simd;
3005    case INDEX_op_x86_punpckl_vec:
3006        insn = punpckl_insn[vece];
3007        goto gen_simd;
3008    case INDEX_op_x86_punpckh_vec:
3009        insn = punpckh_insn[vece];
3010        goto gen_simd;
3011    case INDEX_op_x86_packss_vec:
3012        insn = packss_insn[vece];
3013        goto gen_simd;
3014    case INDEX_op_x86_packus_vec:
3015        insn = packus_insn[vece];
3016        goto gen_simd;
3017    case INDEX_op_x86_vpshldv_vec:
3018        insn = vpshldv_insn[vece];
3019        a1 = a2;
3020        a2 = args[3];
3021        goto gen_simd;
3022    case INDEX_op_x86_vpshrdv_vec:
3023        insn = vpshrdv_insn[vece];
3024        a1 = a2;
3025        a2 = args[3];
3026        goto gen_simd;
3027#if TCG_TARGET_REG_BITS == 32
3028    case INDEX_op_dup2_vec:
3029        /* First merge the two 32-bit inputs to a single 64-bit element. */
3030        tcg_out_vex_modrm(s, OPC_PUNPCKLDQ, a0, a1, a2);
3031        /* Then replicate the 64-bit elements across the rest of the vector. */
3032        if (type != TCG_TYPE_V64) {
3033            tcg_out_dup_vec(s, type, MO_64, a0, a0);
3034        }
3035        break;
3036#endif
3037    case INDEX_op_abs_vec:
3038        insn = abs_insn[vece];
3039        a2 = a1;
3040        a1 = 0;
3041        goto gen_simd;
3042    gen_simd:
3043        tcg_debug_assert(insn != OPC_UD2);
3044        if (type == TCG_TYPE_V256) {
3045            insn |= P_VEXL;
3046        }
3047        tcg_out_vex_modrm(s, insn, a0, a1, a2);
3048        break;
3049
3050    case INDEX_op_cmp_vec:
3051        sub = args[3];
3052        if (sub == TCG_COND_EQ) {
3053            insn = cmpeq_insn[vece];
3054        } else if (sub == TCG_COND_GT) {
3055            insn = cmpgt_insn[vece];
3056        } else {
3057            g_assert_not_reached();
3058        }
3059        goto gen_simd;
3060
3061    case INDEX_op_andc_vec:
3062        insn = OPC_PANDN;
3063        if (type == TCG_TYPE_V256) {
3064            insn |= P_VEXL;
3065        }
3066        tcg_out_vex_modrm(s, insn, a0, a2, a1);
3067        break;
3068
3069    case INDEX_op_shli_vec:
3070        insn = shift_imm_insn[vece];
3071        sub = 6;
3072        goto gen_shift;
3073    case INDEX_op_shri_vec:
3074        insn = shift_imm_insn[vece];
3075        sub = 2;
3076        goto gen_shift;
3077    case INDEX_op_sari_vec:
3078        if (vece == MO_64) {
3079            insn = OPC_PSHIFTD_Ib | P_VEXW | P_EVEX;
3080        } else {
3081            insn = shift_imm_insn[vece];
3082        }
3083        sub = 4;
3084        goto gen_shift;
3085    case INDEX_op_rotli_vec:
3086        insn = OPC_PSHIFTD_Ib | P_EVEX;  /* VPROL[DQ] */
3087        if (vece == MO_64) {
3088            insn |= P_VEXW;
3089        }
3090        sub = 1;
3091        goto gen_shift;
3092    gen_shift:
3093        tcg_debug_assert(vece != MO_8);
3094        if (type == TCG_TYPE_V256) {
3095            insn |= P_VEXL;
3096        }
3097        tcg_out_vex_modrm(s, insn, sub, a0, a1);
3098        tcg_out8(s, a2);
3099        break;
3100
3101    case INDEX_op_ld_vec:
3102        tcg_out_ld(s, type, a0, a1, a2);
3103        break;
3104    case INDEX_op_st_vec:
3105        tcg_out_st(s, type, a0, a1, a2);
3106        break;
3107    case INDEX_op_dupm_vec:
3108        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
3109        break;
3110
3111    case INDEX_op_x86_shufps_vec:
3112        insn = OPC_SHUFPS;
3113        sub = args[3];
3114        goto gen_simd_imm8;
3115    case INDEX_op_x86_blend_vec:
3116        if (vece == MO_16) {
3117            insn = OPC_PBLENDW;
3118        } else if (vece == MO_32) {
3119            insn = (have_avx2 ? OPC_VPBLENDD : OPC_BLENDPS);
3120        } else {
3121            g_assert_not_reached();
3122        }
3123        sub = args[3];
3124        goto gen_simd_imm8;
3125    case INDEX_op_x86_vperm2i128_vec:
3126        insn = OPC_VPERM2I128;
3127        sub = args[3];
3128        goto gen_simd_imm8;
3129    case INDEX_op_x86_vpshldi_vec:
3130        insn = vpshldi_insn[vece];
3131        sub = args[3];
3132        goto gen_simd_imm8;
3133
3134    case INDEX_op_not_vec:
3135        insn = OPC_VPTERNLOGQ;
3136        a2 = a1;
3137        sub = 0x33; /* !B */
3138        goto gen_simd_imm8;
3139    case INDEX_op_nor_vec:
3140        insn = OPC_VPTERNLOGQ;
3141        sub = 0x11; /* norCB */
3142        goto gen_simd_imm8;
3143    case INDEX_op_nand_vec:
3144        insn = OPC_VPTERNLOGQ;
3145        sub = 0x77; /* nandCB */
3146        goto gen_simd_imm8;
3147    case INDEX_op_eqv_vec:
3148        insn = OPC_VPTERNLOGQ;
3149        sub = 0x99; /* xnorCB */
3150        goto gen_simd_imm8;
3151    case INDEX_op_orc_vec:
3152        insn = OPC_VPTERNLOGQ;
3153        sub = 0xdd; /* orB!C */
3154        goto gen_simd_imm8;
3155
3156    case INDEX_op_bitsel_vec:
3157        insn = OPC_VPTERNLOGQ;
3158        a3 = args[3];
3159        if (a0 == a1) {
3160            a1 = a2;
3161            a2 = a3;
3162            sub = 0xca; /* A?B:C */
3163        } else if (a0 == a2) {
3164            a2 = a3;
3165            sub = 0xe2; /* B?A:C */
3166        } else {
3167            tcg_out_mov(s, type, a0, a3);
3168            sub = 0xb8; /* B?C:A */
3169        }
3170        goto gen_simd_imm8;
3171
3172    gen_simd_imm8:
3173        tcg_debug_assert(insn != OPC_UD2);
3174        if (type == TCG_TYPE_V256) {
3175            insn |= P_VEXL;
3176        }
3177        tcg_out_vex_modrm(s, insn, a0, a1, a2);
3178        tcg_out8(s, sub);
3179        break;
3180
3181    case INDEX_op_x86_vpblendvb_vec:
3182        insn = OPC_VPBLENDVB;
3183        if (type == TCG_TYPE_V256) {
3184            insn |= P_VEXL;
3185        }
3186        tcg_out_vex_modrm(s, insn, a0, a1, a2);
3187        tcg_out8(s, args[3] << 4);
3188        break;
3189
3190    case INDEX_op_x86_psrldq_vec:
3191        tcg_out_vex_modrm(s, OPC_GRP14, 3, a0, a1);
3192        tcg_out8(s, a2);
3193        break;
3194
3195    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
3196    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
3197    default:
3198        g_assert_not_reached();
3199    }
3200}
3201
3202static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
3203{
3204    switch (op) {
3205    case INDEX_op_goto_ptr:
3206        return C_O0_I1(r);
3207
3208    case INDEX_op_ld8u_i32:
3209    case INDEX_op_ld8u_i64:
3210    case INDEX_op_ld8s_i32:
3211    case INDEX_op_ld8s_i64:
3212    case INDEX_op_ld16u_i32:
3213    case INDEX_op_ld16u_i64:
3214    case INDEX_op_ld16s_i32:
3215    case INDEX_op_ld16s_i64:
3216    case INDEX_op_ld_i32:
3217    case INDEX_op_ld32u_i64:
3218    case INDEX_op_ld32s_i64:
3219    case INDEX_op_ld_i64:
3220        return C_O1_I1(r, r);
3221
3222    case INDEX_op_st8_i32:
3223    case INDEX_op_st8_i64:
3224        return C_O0_I2(qi, r);
3225
3226    case INDEX_op_st16_i32:
3227    case INDEX_op_st16_i64:
3228    case INDEX_op_st_i32:
3229    case INDEX_op_st32_i64:
3230        return C_O0_I2(ri, r);
3231
3232    case INDEX_op_st_i64:
3233        return C_O0_I2(re, r);
3234
3235    case INDEX_op_add_i32:
3236    case INDEX_op_add_i64:
3237        return C_O1_I2(r, r, re);
3238
3239    case INDEX_op_sub_i32:
3240    case INDEX_op_sub_i64:
3241    case INDEX_op_mul_i32:
3242    case INDEX_op_mul_i64:
3243    case INDEX_op_or_i32:
3244    case INDEX_op_or_i64:
3245    case INDEX_op_xor_i32:
3246    case INDEX_op_xor_i64:
3247        return C_O1_I2(r, 0, re);
3248
3249    case INDEX_op_and_i32:
3250    case INDEX_op_and_i64:
3251        return C_O1_I2(r, 0, reZ);
3252
3253    case INDEX_op_andc_i32:
3254    case INDEX_op_andc_i64:
3255        return C_O1_I2(r, r, rI);
3256
3257    case INDEX_op_shl_i32:
3258    case INDEX_op_shl_i64:
3259    case INDEX_op_shr_i32:
3260    case INDEX_op_shr_i64:
3261    case INDEX_op_sar_i32:
3262    case INDEX_op_sar_i64:
3263        return have_bmi2 ? C_O1_I2(r, r, ri) : C_O1_I2(r, 0, ci);
3264
3265    case INDEX_op_rotl_i32:
3266    case INDEX_op_rotl_i64:
3267    case INDEX_op_rotr_i32:
3268    case INDEX_op_rotr_i64:
3269        return C_O1_I2(r, 0, ci);
3270
3271    case INDEX_op_brcond_i32:
3272    case INDEX_op_brcond_i64:
3273        return C_O0_I2(r, re);
3274
3275    case INDEX_op_bswap16_i32:
3276    case INDEX_op_bswap16_i64:
3277    case INDEX_op_bswap32_i32:
3278    case INDEX_op_bswap32_i64:
3279    case INDEX_op_bswap64_i64:
3280    case INDEX_op_neg_i32:
3281    case INDEX_op_neg_i64:
3282    case INDEX_op_not_i32:
3283    case INDEX_op_not_i64:
3284    case INDEX_op_extrh_i64_i32:
3285        return C_O1_I1(r, 0);
3286
3287    case INDEX_op_ext8s_i32:
3288    case INDEX_op_ext8s_i64:
3289    case INDEX_op_ext8u_i32:
3290    case INDEX_op_ext8u_i64:
3291        return C_O1_I1(r, q);
3292
3293    case INDEX_op_ext16s_i32:
3294    case INDEX_op_ext16s_i64:
3295    case INDEX_op_ext16u_i32:
3296    case INDEX_op_ext16u_i64:
3297    case INDEX_op_ext32s_i64:
3298    case INDEX_op_ext32u_i64:
3299    case INDEX_op_ext_i32_i64:
3300    case INDEX_op_extu_i32_i64:
3301    case INDEX_op_extrl_i64_i32:
3302    case INDEX_op_extract_i32:
3303    case INDEX_op_extract_i64:
3304    case INDEX_op_sextract_i32:
3305    case INDEX_op_ctpop_i32:
3306    case INDEX_op_ctpop_i64:
3307        return C_O1_I1(r, r);
3308
3309    case INDEX_op_extract2_i32:
3310    case INDEX_op_extract2_i64:
3311        return C_O1_I2(r, 0, r);
3312
3313    case INDEX_op_deposit_i32:
3314    case INDEX_op_deposit_i64:
3315        return C_O1_I2(Q, 0, Q);
3316
3317    case INDEX_op_setcond_i32:
3318    case INDEX_op_setcond_i64:
3319        return C_O1_I2(q, r, re);
3320
3321    case INDEX_op_movcond_i32:
3322    case INDEX_op_movcond_i64:
3323        return C_O1_I4(r, r, re, r, 0);
3324
3325    case INDEX_op_div2_i32:
3326    case INDEX_op_div2_i64:
3327    case INDEX_op_divu2_i32:
3328    case INDEX_op_divu2_i64:
3329        return C_O2_I3(a, d, 0, 1, r);
3330
3331    case INDEX_op_mulu2_i32:
3332    case INDEX_op_mulu2_i64:
3333    case INDEX_op_muls2_i32:
3334    case INDEX_op_muls2_i64:
3335        return C_O2_I2(a, d, a, r);
3336
3337    case INDEX_op_add2_i32:
3338    case INDEX_op_add2_i64:
3339    case INDEX_op_sub2_i32:
3340    case INDEX_op_sub2_i64:
3341        return C_N1_O1_I4(r, r, 0, 1, re, re);
3342
3343    case INDEX_op_ctz_i32:
3344    case INDEX_op_ctz_i64:
3345        return have_bmi1 ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
3346
3347    case INDEX_op_clz_i32:
3348    case INDEX_op_clz_i64:
3349        return have_lzcnt ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
3350
3351    case INDEX_op_qemu_ld_a32_i32:
3352        return C_O1_I1(r, L);
3353    case INDEX_op_qemu_ld_a64_i32:
3354        return TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L) : C_O1_I2(r, L, L);
3355
3356    case INDEX_op_qemu_st_a32_i32:
3357        return C_O0_I2(L, L);
3358    case INDEX_op_qemu_st_a64_i32:
3359        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I3(L, L, L);
3360    case INDEX_op_qemu_st8_a32_i32:
3361        return C_O0_I2(s, L);
3362    case INDEX_op_qemu_st8_a64_i32:
3363        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(s, L) : C_O0_I3(s, L, L);
3364
3365    case INDEX_op_qemu_ld_a32_i64:
3366        return TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L) : C_O2_I1(r, r, L);
3367    case INDEX_op_qemu_ld_a64_i64:
3368        return TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L) : C_O2_I2(r, r, L, L);
3369
3370    case INDEX_op_qemu_st_a32_i64:
3371        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I3(L, L, L);
3372    case INDEX_op_qemu_st_a64_i64:
3373        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I4(L, L, L, L);
3374
3375    case INDEX_op_qemu_ld_a32_i128:
3376    case INDEX_op_qemu_ld_a64_i128:
3377        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3378        return C_O2_I1(r, r, L);
3379    case INDEX_op_qemu_st_a32_i128:
3380    case INDEX_op_qemu_st_a64_i128:
3381        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3382        return C_O0_I3(L, L, L);
3383
3384    case INDEX_op_brcond2_i32:
3385        return C_O0_I4(r, r, ri, ri);
3386
3387    case INDEX_op_setcond2_i32:
3388        return C_O1_I4(r, r, r, ri, ri);
3389
3390    case INDEX_op_ld_vec:
3391    case INDEX_op_dupm_vec:
3392        return C_O1_I1(x, r);
3393
3394    case INDEX_op_st_vec:
3395        return C_O0_I2(x, r);
3396
3397    case INDEX_op_add_vec:
3398    case INDEX_op_sub_vec:
3399    case INDEX_op_mul_vec:
3400    case INDEX_op_and_vec:
3401    case INDEX_op_or_vec:
3402    case INDEX_op_xor_vec:
3403    case INDEX_op_andc_vec:
3404    case INDEX_op_orc_vec:
3405    case INDEX_op_nand_vec:
3406    case INDEX_op_nor_vec:
3407    case INDEX_op_eqv_vec:
3408    case INDEX_op_ssadd_vec:
3409    case INDEX_op_usadd_vec:
3410    case INDEX_op_sssub_vec:
3411    case INDEX_op_ussub_vec:
3412    case INDEX_op_smin_vec:
3413    case INDEX_op_umin_vec:
3414    case INDEX_op_smax_vec:
3415    case INDEX_op_umax_vec:
3416    case INDEX_op_shlv_vec:
3417    case INDEX_op_shrv_vec:
3418    case INDEX_op_sarv_vec:
3419    case INDEX_op_rotlv_vec:
3420    case INDEX_op_rotrv_vec:
3421    case INDEX_op_shls_vec:
3422    case INDEX_op_shrs_vec:
3423    case INDEX_op_sars_vec:
3424    case INDEX_op_cmp_vec:
3425    case INDEX_op_x86_shufps_vec:
3426    case INDEX_op_x86_blend_vec:
3427    case INDEX_op_x86_packss_vec:
3428    case INDEX_op_x86_packus_vec:
3429    case INDEX_op_x86_vperm2i128_vec:
3430    case INDEX_op_x86_punpckl_vec:
3431    case INDEX_op_x86_punpckh_vec:
3432    case INDEX_op_x86_vpshldi_vec:
3433#if TCG_TARGET_REG_BITS == 32
3434    case INDEX_op_dup2_vec:
3435#endif
3436        return C_O1_I2(x, x, x);
3437
3438    case INDEX_op_abs_vec:
3439    case INDEX_op_dup_vec:
3440    case INDEX_op_not_vec:
3441    case INDEX_op_shli_vec:
3442    case INDEX_op_shri_vec:
3443    case INDEX_op_sari_vec:
3444    case INDEX_op_rotli_vec:
3445    case INDEX_op_x86_psrldq_vec:
3446        return C_O1_I1(x, x);
3447
3448    case INDEX_op_x86_vpshldv_vec:
3449    case INDEX_op_x86_vpshrdv_vec:
3450        return C_O1_I3(x, 0, x, x);
3451
3452    case INDEX_op_bitsel_vec:
3453    case INDEX_op_x86_vpblendvb_vec:
3454        return C_O1_I3(x, x, x, x);
3455
3456    default:
3457        g_assert_not_reached();
3458    }
3459}
3460
3461int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
3462{
3463    switch (opc) {
3464    case INDEX_op_add_vec:
3465    case INDEX_op_sub_vec:
3466    case INDEX_op_and_vec:
3467    case INDEX_op_or_vec:
3468    case INDEX_op_xor_vec:
3469    case INDEX_op_andc_vec:
3470    case INDEX_op_orc_vec:
3471    case INDEX_op_nand_vec:
3472    case INDEX_op_nor_vec:
3473    case INDEX_op_eqv_vec:
3474    case INDEX_op_not_vec:
3475    case INDEX_op_bitsel_vec:
3476        return 1;
3477    case INDEX_op_cmp_vec:
3478    case INDEX_op_cmpsel_vec:
3479        return -1;
3480
3481    case INDEX_op_rotli_vec:
3482        return have_avx512vl && vece >= MO_32 ? 1 : -1;
3483
3484    case INDEX_op_shli_vec:
3485    case INDEX_op_shri_vec:
3486        /* We must expand the operation for MO_8.  */
3487        return vece == MO_8 ? -1 : 1;
3488
3489    case INDEX_op_sari_vec:
3490        switch (vece) {
3491        case MO_8:
3492            return -1;
3493        case MO_16:
3494        case MO_32:
3495            return 1;
3496        case MO_64:
3497            if (have_avx512vl) {
3498                return 1;
3499            }
3500            /*
3501             * We can emulate this for MO_64, but it does not pay off
3502             * unless we're producing at least 4 values.
3503             */
3504            return type >= TCG_TYPE_V256 ? -1 : 0;
3505        }
3506        return 0;
3507
3508    case INDEX_op_shls_vec:
3509    case INDEX_op_shrs_vec:
3510        return vece >= MO_16;
3511    case INDEX_op_sars_vec:
3512        switch (vece) {
3513        case MO_16:
3514        case MO_32:
3515            return 1;
3516        case MO_64:
3517            return have_avx512vl;
3518        }
3519        return 0;
3520    case INDEX_op_rotls_vec:
3521        return vece >= MO_16 ? -1 : 0;
3522
3523    case INDEX_op_shlv_vec:
3524    case INDEX_op_shrv_vec:
3525        switch (vece) {
3526        case MO_16:
3527            return have_avx512bw;
3528        case MO_32:
3529        case MO_64:
3530            return have_avx2;
3531        }
3532        return 0;
3533    case INDEX_op_sarv_vec:
3534        switch (vece) {
3535        case MO_16:
3536            return have_avx512bw;
3537        case MO_32:
3538            return have_avx2;
3539        case MO_64:
3540            return have_avx512vl;
3541        }
3542        return 0;
3543    case INDEX_op_rotlv_vec:
3544    case INDEX_op_rotrv_vec:
3545        switch (vece) {
3546        case MO_16:
3547            return have_avx512vbmi2 ? -1 : 0;
3548        case MO_32:
3549        case MO_64:
3550            return have_avx512vl ? 1 : have_avx2 ? -1 : 0;
3551        }
3552        return 0;
3553
3554    case INDEX_op_mul_vec:
3555        switch (vece) {
3556        case MO_8:
3557            return -1;
3558        case MO_64:
3559            return have_avx512dq;
3560        }
3561        return 1;
3562
3563    case INDEX_op_ssadd_vec:
3564    case INDEX_op_usadd_vec:
3565    case INDEX_op_sssub_vec:
3566    case INDEX_op_ussub_vec:
3567        return vece <= MO_16;
3568    case INDEX_op_smin_vec:
3569    case INDEX_op_smax_vec:
3570    case INDEX_op_umin_vec:
3571    case INDEX_op_umax_vec:
3572    case INDEX_op_abs_vec:
3573        return vece <= MO_32 || have_avx512vl;
3574
3575    default:
3576        return 0;
3577    }
3578}
3579
3580static void expand_vec_shi(TCGType type, unsigned vece, TCGOpcode opc,
3581                           TCGv_vec v0, TCGv_vec v1, TCGArg imm)
3582{
3583    TCGv_vec t1, t2;
3584
3585    tcg_debug_assert(vece == MO_8);
3586
3587    t1 = tcg_temp_new_vec(type);
3588    t2 = tcg_temp_new_vec(type);
3589
3590    /*
3591     * Unpack to W, shift, and repack.  Tricky bits:
3592     * (1) Use punpck*bw x,x to produce DDCCBBAA,
3593     *     i.e. duplicate in other half of the 16-bit lane.
3594     * (2) For right-shift, add 8 so that the high half of the lane
3595     *     becomes zero.  For left-shift, and left-rotate, we must
3596     *     shift up and down again.
3597     * (3) Step 2 leaves high half zero such that PACKUSWB
3598     *     (pack with unsigned saturation) does not modify
3599     *     the quantity.
3600     */
3601    vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3602              tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3603    vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3604              tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3605
3606    if (opc != INDEX_op_rotli_vec) {
3607        imm += 8;
3608    }
3609    if (opc == INDEX_op_shri_vec) {
3610        tcg_gen_shri_vec(MO_16, t1, t1, imm);
3611        tcg_gen_shri_vec(MO_16, t2, t2, imm);
3612    } else {
3613        tcg_gen_shli_vec(MO_16, t1, t1, imm);
3614        tcg_gen_shli_vec(MO_16, t2, t2, imm);
3615        tcg_gen_shri_vec(MO_16, t1, t1, 8);
3616        tcg_gen_shri_vec(MO_16, t2, t2, 8);
3617    }
3618
3619    vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
3620              tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
3621    tcg_temp_free_vec(t1);
3622    tcg_temp_free_vec(t2);
3623}
3624
3625static void expand_vec_sari(TCGType type, unsigned vece,
3626                            TCGv_vec v0, TCGv_vec v1, TCGArg imm)
3627{
3628    TCGv_vec t1, t2;
3629
3630    switch (vece) {
3631    case MO_8:
3632        /* Unpack to W, shift, and repack, as in expand_vec_shi.  */
3633        t1 = tcg_temp_new_vec(type);
3634        t2 = tcg_temp_new_vec(type);
3635        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3636                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3637        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3638                  tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3639        tcg_gen_sari_vec(MO_16, t1, t1, imm + 8);
3640        tcg_gen_sari_vec(MO_16, t2, t2, imm + 8);
3641        vec_gen_3(INDEX_op_x86_packss_vec, type, MO_8,
3642                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
3643        tcg_temp_free_vec(t1);
3644        tcg_temp_free_vec(t2);
3645        break;
3646
3647    case MO_64:
3648        t1 = tcg_temp_new_vec(type);
3649        if (imm <= 32) {
3650            /*
3651             * We can emulate a small sign extend by performing an arithmetic
3652             * 32-bit shift and overwriting the high half of a 64-bit logical
3653             * shift.  Note that the ISA says shift of 32 is valid, but TCG
3654             * does not, so we have to bound the smaller shift -- we get the
3655             * same result in the high half either way.
3656             */
3657            tcg_gen_sari_vec(MO_32, t1, v1, MIN(imm, 31));
3658            tcg_gen_shri_vec(MO_64, v0, v1, imm);
3659            vec_gen_4(INDEX_op_x86_blend_vec, type, MO_32,
3660                      tcgv_vec_arg(v0), tcgv_vec_arg(v0),
3661                      tcgv_vec_arg(t1), 0xaa);
3662        } else {
3663            /* Otherwise we will need to use a compare vs 0 to produce
3664             * the sign-extend, shift and merge.
3665             */
3666            tcg_gen_cmp_vec(TCG_COND_GT, MO_64, t1,
3667                            tcg_constant_vec(type, MO_64, 0), v1);
3668            tcg_gen_shri_vec(MO_64, v0, v1, imm);
3669            tcg_gen_shli_vec(MO_64, t1, t1, 64 - imm);
3670            tcg_gen_or_vec(MO_64, v0, v0, t1);
3671        }
3672        tcg_temp_free_vec(t1);
3673        break;
3674
3675    default:
3676        g_assert_not_reached();
3677    }
3678}
3679
3680static void expand_vec_rotli(TCGType type, unsigned vece,
3681                             TCGv_vec v0, TCGv_vec v1, TCGArg imm)
3682{
3683    TCGv_vec t;
3684
3685    if (vece == MO_8) {
3686        expand_vec_shi(type, vece, INDEX_op_rotli_vec, v0, v1, imm);
3687        return;
3688    }
3689
3690    if (have_avx512vbmi2) {
3691        vec_gen_4(INDEX_op_x86_vpshldi_vec, type, vece,
3692                  tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v1), imm);
3693        return;
3694    }
3695
3696    t = tcg_temp_new_vec(type);
3697    tcg_gen_shli_vec(vece, t, v1, imm);
3698    tcg_gen_shri_vec(vece, v0, v1, (8 << vece) - imm);
3699    tcg_gen_or_vec(vece, v0, v0, t);
3700    tcg_temp_free_vec(t);
3701}
3702
3703static void expand_vec_rotv(TCGType type, unsigned vece, TCGv_vec v0,
3704                            TCGv_vec v1, TCGv_vec sh, bool right)
3705{
3706    TCGv_vec t;
3707
3708    if (have_avx512vbmi2) {
3709        vec_gen_4(right ? INDEX_op_x86_vpshrdv_vec : INDEX_op_x86_vpshldv_vec,
3710                  type, vece, tcgv_vec_arg(v0), tcgv_vec_arg(v1),
3711                  tcgv_vec_arg(v1), tcgv_vec_arg(sh));
3712        return;
3713    }
3714
3715    t = tcg_temp_new_vec(type);
3716    tcg_gen_dupi_vec(vece, t, 8 << vece);
3717    tcg_gen_sub_vec(vece, t, t, sh);
3718    if (right) {
3719        tcg_gen_shlv_vec(vece, t, v1, t);
3720        tcg_gen_shrv_vec(vece, v0, v1, sh);
3721    } else {
3722        tcg_gen_shrv_vec(vece, t, v1, t);
3723        tcg_gen_shlv_vec(vece, v0, v1, sh);
3724    }
3725    tcg_gen_or_vec(vece, v0, v0, t);
3726    tcg_temp_free_vec(t);
3727}
3728
3729static void expand_vec_rotls(TCGType type, unsigned vece,
3730                             TCGv_vec v0, TCGv_vec v1, TCGv_i32 lsh)
3731{
3732    TCGv_vec t = tcg_temp_new_vec(type);
3733
3734    tcg_debug_assert(vece != MO_8);
3735
3736    if (vece >= MO_32 ? have_avx512vl : have_avx512vbmi2) {
3737        tcg_gen_dup_i32_vec(vece, t, lsh);
3738        if (vece >= MO_32) {
3739            tcg_gen_rotlv_vec(vece, v0, v1, t);
3740        } else {
3741            expand_vec_rotv(type, vece, v0, v1, t, false);
3742        }
3743    } else {
3744        TCGv_i32 rsh = tcg_temp_new_i32();
3745
3746        tcg_gen_neg_i32(rsh, lsh);
3747        tcg_gen_andi_i32(rsh, rsh, (8 << vece) - 1);
3748        tcg_gen_shls_vec(vece, t, v1, lsh);
3749        tcg_gen_shrs_vec(vece, v0, v1, rsh);
3750        tcg_gen_or_vec(vece, v0, v0, t);
3751
3752        tcg_temp_free_i32(rsh);
3753    }
3754
3755    tcg_temp_free_vec(t);
3756}
3757
3758static void expand_vec_mul(TCGType type, unsigned vece,
3759                           TCGv_vec v0, TCGv_vec v1, TCGv_vec v2)
3760{
3761    TCGv_vec t1, t2, t3, t4, zero;
3762
3763    tcg_debug_assert(vece == MO_8);
3764
3765    /*
3766     * Unpack v1 bytes to words, 0 | x.
3767     * Unpack v2 bytes to words, y | 0.
3768     * This leaves the 8-bit result, x * y, with 8 bits of right padding.
3769     * Shift logical right by 8 bits to clear the high 8 bytes before
3770     * using an unsigned saturated pack.
3771     *
3772     * The difference between the V64, V128 and V256 cases is merely how
3773     * we distribute the expansion between temporaries.
3774     */
3775    switch (type) {
3776    case TCG_TYPE_V64:
3777        t1 = tcg_temp_new_vec(TCG_TYPE_V128);
3778        t2 = tcg_temp_new_vec(TCG_TYPE_V128);
3779        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
3780        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
3781                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
3782        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
3783                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
3784        tcg_gen_mul_vec(MO_16, t1, t1, t2);
3785        tcg_gen_shri_vec(MO_16, t1, t1, 8);
3786        vec_gen_3(INDEX_op_x86_packus_vec, TCG_TYPE_V128, MO_8,
3787                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t1));
3788        tcg_temp_free_vec(t1);
3789        tcg_temp_free_vec(t2);
3790        break;
3791
3792    case TCG_TYPE_V128:
3793    case TCG_TYPE_V256:
3794        t1 = tcg_temp_new_vec(type);
3795        t2 = tcg_temp_new_vec(type);
3796        t3 = tcg_temp_new_vec(type);
3797        t4 = tcg_temp_new_vec(type);
3798        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
3799        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3800                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
3801        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3802                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
3803        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3804                  tcgv_vec_arg(t3), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
3805        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3806                  tcgv_vec_arg(t4), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
3807        tcg_gen_mul_vec(MO_16, t1, t1, t2);
3808        tcg_gen_mul_vec(MO_16, t3, t3, t4);
3809        tcg_gen_shri_vec(MO_16, t1, t1, 8);
3810        tcg_gen_shri_vec(MO_16, t3, t3, 8);
3811        vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
3812                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t3));
3813        tcg_temp_free_vec(t1);
3814        tcg_temp_free_vec(t2);
3815        tcg_temp_free_vec(t3);
3816        tcg_temp_free_vec(t4);
3817        break;
3818
3819    default:
3820        g_assert_not_reached();
3821    }
3822}
3823
3824static bool expand_vec_cmp_noinv(TCGType type, unsigned vece, TCGv_vec v0,
3825                                 TCGv_vec v1, TCGv_vec v2, TCGCond cond)
3826{
3827    enum {
3828        NEED_INV  = 1,
3829        NEED_SWAP = 2,
3830        NEED_BIAS = 4,
3831        NEED_UMIN = 8,
3832        NEED_UMAX = 16,
3833    };
3834    TCGv_vec t1, t2, t3;
3835    uint8_t fixup;
3836
3837    switch (cond) {
3838    case TCG_COND_EQ:
3839    case TCG_COND_GT:
3840        fixup = 0;
3841        break;
3842    case TCG_COND_NE:
3843    case TCG_COND_LE:
3844        fixup = NEED_INV;
3845        break;
3846    case TCG_COND_LT:
3847        fixup = NEED_SWAP;
3848        break;
3849    case TCG_COND_GE:
3850        fixup = NEED_SWAP | NEED_INV;
3851        break;
3852    case TCG_COND_LEU:
3853        if (tcg_can_emit_vec_op(INDEX_op_umin_vec, type, vece)) {
3854            fixup = NEED_UMIN;
3855        } else {
3856            fixup = NEED_BIAS | NEED_INV;
3857        }
3858        break;
3859    case TCG_COND_GTU:
3860        if (tcg_can_emit_vec_op(INDEX_op_umin_vec, type, vece)) {
3861            fixup = NEED_UMIN | NEED_INV;
3862        } else {
3863            fixup = NEED_BIAS;
3864        }
3865        break;
3866    case TCG_COND_GEU:
3867        if (tcg_can_emit_vec_op(INDEX_op_umax_vec, type, vece)) {
3868            fixup = NEED_UMAX;
3869        } else {
3870            fixup = NEED_BIAS | NEED_SWAP | NEED_INV;
3871        }
3872        break;
3873    case TCG_COND_LTU:
3874        if (tcg_can_emit_vec_op(INDEX_op_umax_vec, type, vece)) {
3875            fixup = NEED_UMAX | NEED_INV;
3876        } else {
3877            fixup = NEED_BIAS | NEED_SWAP;
3878        }
3879        break;
3880    default:
3881        g_assert_not_reached();
3882    }
3883
3884    if (fixup & NEED_INV) {
3885        cond = tcg_invert_cond(cond);
3886    }
3887    if (fixup & NEED_SWAP) {
3888        t1 = v1, v1 = v2, v2 = t1;
3889        cond = tcg_swap_cond(cond);
3890    }
3891
3892    t1 = t2 = NULL;
3893    if (fixup & (NEED_UMIN | NEED_UMAX)) {
3894        t1 = tcg_temp_new_vec(type);
3895        if (fixup & NEED_UMIN) {
3896            tcg_gen_umin_vec(vece, t1, v1, v2);
3897        } else {
3898            tcg_gen_umax_vec(vece, t1, v1, v2);
3899        }
3900        v2 = t1;
3901        cond = TCG_COND_EQ;
3902    } else if (fixup & NEED_BIAS) {
3903        t1 = tcg_temp_new_vec(type);
3904        t2 = tcg_temp_new_vec(type);
3905        t3 = tcg_constant_vec(type, vece, 1ull << ((8 << vece) - 1));
3906        tcg_gen_sub_vec(vece, t1, v1, t3);
3907        tcg_gen_sub_vec(vece, t2, v2, t3);
3908        v1 = t1;
3909        v2 = t2;
3910        cond = tcg_signed_cond(cond);
3911    }
3912
3913    tcg_debug_assert(cond == TCG_COND_EQ || cond == TCG_COND_GT);
3914    /* Expand directly; do not recurse.  */
3915    vec_gen_4(INDEX_op_cmp_vec, type, vece,
3916              tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v2), cond);
3917
3918    if (t1) {
3919        tcg_temp_free_vec(t1);
3920        if (t2) {
3921            tcg_temp_free_vec(t2);
3922        }
3923    }
3924    return fixup & NEED_INV;
3925}
3926
3927static void expand_vec_cmp(TCGType type, unsigned vece, TCGv_vec v0,
3928                           TCGv_vec v1, TCGv_vec v2, TCGCond cond)
3929{
3930    if (expand_vec_cmp_noinv(type, vece, v0, v1, v2, cond)) {
3931        tcg_gen_not_vec(vece, v0, v0);
3932    }
3933}
3934
3935static void expand_vec_cmpsel(TCGType type, unsigned vece, TCGv_vec v0,
3936                              TCGv_vec c1, TCGv_vec c2,
3937                              TCGv_vec v3, TCGv_vec v4, TCGCond cond)
3938{
3939    TCGv_vec t = tcg_temp_new_vec(type);
3940
3941    if (expand_vec_cmp_noinv(type, vece, t, c1, c2, cond)) {
3942        /* Invert the sense of the compare by swapping arguments.  */
3943        TCGv_vec x;
3944        x = v3, v3 = v4, v4 = x;
3945    }
3946    vec_gen_4(INDEX_op_x86_vpblendvb_vec, type, vece,
3947              tcgv_vec_arg(v0), tcgv_vec_arg(v4),
3948              tcgv_vec_arg(v3), tcgv_vec_arg(t));
3949    tcg_temp_free_vec(t);
3950}
3951
3952void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
3953                       TCGArg a0, ...)
3954{
3955    va_list va;
3956    TCGArg a2;
3957    TCGv_vec v0, v1, v2, v3, v4;
3958
3959    va_start(va, a0);
3960    v0 = temp_tcgv_vec(arg_temp(a0));
3961    v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
3962    a2 = va_arg(va, TCGArg);
3963
3964    switch (opc) {
3965    case INDEX_op_shli_vec:
3966    case INDEX_op_shri_vec:
3967        expand_vec_shi(type, vece, opc, v0, v1, a2);
3968        break;
3969
3970    case INDEX_op_sari_vec:
3971        expand_vec_sari(type, vece, v0, v1, a2);
3972        break;
3973
3974    case INDEX_op_rotli_vec:
3975        expand_vec_rotli(type, vece, v0, v1, a2);
3976        break;
3977
3978    case INDEX_op_rotls_vec:
3979        expand_vec_rotls(type, vece, v0, v1, temp_tcgv_i32(arg_temp(a2)));
3980        break;
3981
3982    case INDEX_op_rotlv_vec:
3983        v2 = temp_tcgv_vec(arg_temp(a2));
3984        expand_vec_rotv(type, vece, v0, v1, v2, false);
3985        break;
3986    case INDEX_op_rotrv_vec:
3987        v2 = temp_tcgv_vec(arg_temp(a2));
3988        expand_vec_rotv(type, vece, v0, v1, v2, true);
3989        break;
3990
3991    case INDEX_op_mul_vec:
3992        v2 = temp_tcgv_vec(arg_temp(a2));
3993        expand_vec_mul(type, vece, v0, v1, v2);
3994        break;
3995
3996    case INDEX_op_cmp_vec:
3997        v2 = temp_tcgv_vec(arg_temp(a2));
3998        expand_vec_cmp(type, vece, v0, v1, v2, va_arg(va, TCGArg));
3999        break;
4000
4001    case INDEX_op_cmpsel_vec:
4002        v2 = temp_tcgv_vec(arg_temp(a2));
4003        v3 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
4004        v4 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
4005        expand_vec_cmpsel(type, vece, v0, v1, v2, v3, v4, va_arg(va, TCGArg));
4006        break;
4007
4008    default:
4009        break;
4010    }
4011
4012    va_end(va);
4013}
4014
4015static const int tcg_target_callee_save_regs[] = {
4016#if TCG_TARGET_REG_BITS == 64
4017    TCG_REG_RBP,
4018    TCG_REG_RBX,
4019#if defined(_WIN64)
4020    TCG_REG_RDI,
4021    TCG_REG_RSI,
4022#endif
4023    TCG_REG_R12,
4024    TCG_REG_R13,
4025    TCG_REG_R14, /* Currently used for the global env. */
4026    TCG_REG_R15,
4027#else
4028    TCG_REG_EBP, /* Currently used for the global env. */
4029    TCG_REG_EBX,
4030    TCG_REG_ESI,
4031    TCG_REG_EDI,
4032#endif
4033};
4034
4035/* Compute frame size via macros, to share between tcg_target_qemu_prologue
4036   and tcg_register_jit.  */
4037
4038#define PUSH_SIZE \
4039    ((1 + ARRAY_SIZE(tcg_target_callee_save_regs)) \
4040     * (TCG_TARGET_REG_BITS / 8))
4041
4042#define FRAME_SIZE \
4043    ((PUSH_SIZE \
4044      + TCG_STATIC_CALL_ARGS_SIZE \
4045      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
4046      + TCG_TARGET_STACK_ALIGN - 1) \
4047     & ~(TCG_TARGET_STACK_ALIGN - 1))
4048
4049/* Generate global QEMU prologue and epilogue code */
4050static void tcg_target_qemu_prologue(TCGContext *s)
4051{
4052    int i, stack_addend;
4053
4054    /* TB prologue */
4055
4056    /* Reserve some stack space, also for TCG temps.  */
4057    stack_addend = FRAME_SIZE - PUSH_SIZE;
4058    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
4059                  CPU_TEMP_BUF_NLONGS * sizeof(long));
4060
4061    /* Save all callee saved registers.  */
4062    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
4063        tcg_out_push(s, tcg_target_callee_save_regs[i]);
4064    }
4065
4066#if TCG_TARGET_REG_BITS == 32
4067    tcg_out_ld(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP,
4068               (ARRAY_SIZE(tcg_target_callee_save_regs) + 1) * 4);
4069    tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4070    /* jmp *tb.  */
4071    tcg_out_modrm_offset(s, OPC_GRP5, EXT5_JMPN_Ev, TCG_REG_ESP,
4072                         (ARRAY_SIZE(tcg_target_callee_save_regs) + 2) * 4
4073                         + stack_addend);
4074#else
4075# if !defined(CONFIG_SOFTMMU)
4076    if (guest_base) {
4077        int seg = setup_guest_base_seg();
4078        if (seg != 0) {
4079            x86_guest_base.seg = seg;
4080        } else if (guest_base == (int32_t)guest_base) {
4081            x86_guest_base.ofs = guest_base;
4082        } else {
4083            /* Choose R12 because, as a base, it requires a SIB byte. */
4084            x86_guest_base.index = TCG_REG_R12;
4085            tcg_out_movi(s, TCG_TYPE_PTR, x86_guest_base.index, guest_base);
4086            tcg_regset_set_reg(s->reserved_regs, x86_guest_base.index);
4087        }
4088    }
4089# endif
4090    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
4091    tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4092    /* jmp *tb.  */
4093    tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, tcg_target_call_iarg_regs[1]);
4094#endif
4095
4096    /*
4097     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
4098     * and fall through to the rest of the epilogue.
4099     */
4100    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
4101    tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_EAX, 0);
4102
4103    /* TB epilogue */
4104    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
4105
4106    tcg_out_addi(s, TCG_REG_CALL_STACK, stack_addend);
4107
4108    if (have_avx2) {
4109        tcg_out_vex_opc(s, OPC_VZEROUPPER, 0, 0, 0, 0);
4110    }
4111    for (i = ARRAY_SIZE(tcg_target_callee_save_regs) - 1; i >= 0; i--) {
4112        tcg_out_pop(s, tcg_target_callee_save_regs[i]);
4113    }
4114    tcg_out_opc(s, OPC_RET, 0, 0, 0);
4115}
4116
4117static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
4118{
4119    memset(p, 0x90, count);
4120}
4121
4122static void tcg_target_init(TCGContext *s)
4123{
4124    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
4125    if (TCG_TARGET_REG_BITS == 64) {
4126        tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
4127    }
4128    if (have_avx1) {
4129        tcg_target_available_regs[TCG_TYPE_V64] = ALL_VECTOR_REGS;
4130        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
4131    }
4132    if (have_avx2) {
4133        tcg_target_available_regs[TCG_TYPE_V256] = ALL_VECTOR_REGS;
4134    }
4135
4136    tcg_target_call_clobber_regs = ALL_VECTOR_REGS;
4137    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EAX);
4138    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EDX);
4139    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_ECX);
4140    if (TCG_TARGET_REG_BITS == 64) {
4141#if !defined(_WIN64)
4142        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RDI);
4143        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RSI);
4144#endif
4145        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R8);
4146        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R9);
4147        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R10);
4148        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R11);
4149    }
4150
4151    s->reserved_regs = 0;
4152    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
4153    tcg_regset_set_reg(s->reserved_regs, TCG_TMP_VEC);
4154#ifdef _WIN64
4155    /* These are call saved, and we don't save them, so don't use them. */
4156    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM6);
4157    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM7);
4158    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM8);
4159    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM9);
4160    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM10);
4161    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM11);
4162    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM12);
4163    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM13);
4164    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM14);
4165    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM15);
4166#endif
4167}
4168
4169typedef struct {
4170    DebugFrameHeader h;
4171    uint8_t fde_def_cfa[4];
4172    uint8_t fde_reg_ofs[14];
4173} DebugFrame;
4174
4175/* We're expecting a 2 byte uleb128 encoded value.  */
4176QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
4177
4178#if !defined(__ELF__)
4179    /* Host machine without ELF. */
4180#elif TCG_TARGET_REG_BITS == 64
4181#define ELF_HOST_MACHINE EM_X86_64
4182static const DebugFrame debug_frame = {
4183    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4184    .h.cie.id = -1,
4185    .h.cie.version = 1,
4186    .h.cie.code_align = 1,
4187    .h.cie.data_align = 0x78,             /* sleb128 -8 */
4188    .h.cie.return_column = 16,
4189
4190    /* Total FDE size does not include the "len" member.  */
4191    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4192
4193    .fde_def_cfa = {
4194        12, 7,                          /* DW_CFA_def_cfa %rsp, ... */
4195        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4196        (FRAME_SIZE >> 7)
4197    },
4198    .fde_reg_ofs = {
4199        0x90, 1,                        /* DW_CFA_offset, %rip, -8 */
4200        /* The following ordering must match tcg_target_callee_save_regs.  */
4201        0x86, 2,                        /* DW_CFA_offset, %rbp, -16 */
4202        0x83, 3,                        /* DW_CFA_offset, %rbx, -24 */
4203        0x8c, 4,                        /* DW_CFA_offset, %r12, -32 */
4204        0x8d, 5,                        /* DW_CFA_offset, %r13, -40 */
4205        0x8e, 6,                        /* DW_CFA_offset, %r14, -48 */
4206        0x8f, 7,                        /* DW_CFA_offset, %r15, -56 */
4207    }
4208};
4209#else
4210#define ELF_HOST_MACHINE EM_386
4211static const DebugFrame debug_frame = {
4212    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4213    .h.cie.id = -1,
4214    .h.cie.version = 1,
4215    .h.cie.code_align = 1,
4216    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
4217    .h.cie.return_column = 8,
4218
4219    /* Total FDE size does not include the "len" member.  */
4220    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4221
4222    .fde_def_cfa = {
4223        12, 4,                          /* DW_CFA_def_cfa %esp, ... */
4224        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4225        (FRAME_SIZE >> 7)
4226    },
4227    .fde_reg_ofs = {
4228        0x88, 1,                        /* DW_CFA_offset, %eip, -4 */
4229        /* The following ordering must match tcg_target_callee_save_regs.  */
4230        0x85, 2,                        /* DW_CFA_offset, %ebp, -8 */
4231        0x83, 3,                        /* DW_CFA_offset, %ebx, -12 */
4232        0x86, 4,                        /* DW_CFA_offset, %esi, -16 */
4233        0x87, 5,                        /* DW_CFA_offset, %edi, -20 */
4234    }
4235};
4236#endif
4237
4238#if defined(ELF_HOST_MACHINE)
4239void tcg_register_jit(const void *buf, size_t buf_size)
4240{
4241    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
4242}
4243#endif
4244