xref: /openbmc/qemu/tcg/i386/tcg-target.c.inc (revision 5a5bb0a5a0b879c8f110c6a9bde9146181ef840c)
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/* Used for function call generation. */
26#define TCG_TARGET_STACK_ALIGN 16
27#if defined(_WIN64)
28#define TCG_TARGET_CALL_STACK_OFFSET 32
29#else
30#define TCG_TARGET_CALL_STACK_OFFSET 0
31#endif
32#define TCG_TARGET_CALL_ARG_I32      TCG_CALL_ARG_NORMAL
33#define TCG_TARGET_CALL_ARG_I64      TCG_CALL_ARG_NORMAL
34#if defined(_WIN64)
35# define TCG_TARGET_CALL_ARG_I128    TCG_CALL_ARG_BY_REF
36# define TCG_TARGET_CALL_RET_I128    TCG_CALL_RET_BY_VEC
37#elif TCG_TARGET_REG_BITS == 64
38# define TCG_TARGET_CALL_ARG_I128    TCG_CALL_ARG_NORMAL
39# define TCG_TARGET_CALL_RET_I128    TCG_CALL_RET_NORMAL
40#else
41# define TCG_TARGET_CALL_ARG_I128    TCG_CALL_ARG_NORMAL
42# define TCG_TARGET_CALL_RET_I128    TCG_CALL_RET_BY_REF
43#endif
44
45#ifdef CONFIG_DEBUG_TCG
46static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
47#if TCG_TARGET_REG_BITS == 64
48    "%rax", "%rcx", "%rdx", "%rbx", "%rsp", "%rbp", "%rsi", "%rdi",
49#else
50    "%eax", "%ecx", "%edx", "%ebx", "%esp", "%ebp", "%esi", "%edi",
51#endif
52    "%r8",  "%r9",  "%r10", "%r11", "%r12", "%r13", "%r14", "%r15",
53    "%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5", "%xmm6", "%xmm7",
54#if TCG_TARGET_REG_BITS == 64
55    "%xmm8", "%xmm9", "%xmm10", "%xmm11",
56    "%xmm12", "%xmm13", "%xmm14", "%xmm15",
57#endif
58};
59#endif
60
61static const int tcg_target_reg_alloc_order[] = {
62#if TCG_TARGET_REG_BITS == 64
63    TCG_REG_RBP,
64    TCG_REG_RBX,
65    TCG_REG_R12,
66    TCG_REG_R13,
67    TCG_REG_R14,
68    TCG_REG_R15,
69    TCG_REG_R10,
70    TCG_REG_R11,
71    TCG_REG_R9,
72    TCG_REG_R8,
73    TCG_REG_RCX,
74    TCG_REG_RDX,
75    TCG_REG_RSI,
76    TCG_REG_RDI,
77    TCG_REG_RAX,
78#else
79    TCG_REG_EBX,
80    TCG_REG_ESI,
81    TCG_REG_EDI,
82    TCG_REG_EBP,
83    TCG_REG_ECX,
84    TCG_REG_EDX,
85    TCG_REG_EAX,
86#endif
87    TCG_REG_XMM0,
88    TCG_REG_XMM1,
89    TCG_REG_XMM2,
90    TCG_REG_XMM3,
91    TCG_REG_XMM4,
92    TCG_REG_XMM5,
93#ifndef _WIN64
94    /* The Win64 ABI has xmm6-xmm15 as caller-saves, and we do not save
95       any of them.  Therefore only allow xmm0-xmm5 to be allocated.  */
96    TCG_REG_XMM6,
97    TCG_REG_XMM7,
98#if TCG_TARGET_REG_BITS == 64
99    TCG_REG_XMM8,
100    TCG_REG_XMM9,
101    TCG_REG_XMM10,
102    TCG_REG_XMM11,
103    TCG_REG_XMM12,
104    TCG_REG_XMM13,
105    TCG_REG_XMM14,
106    TCG_REG_XMM15,
107#endif
108#endif
109};
110
111#define TCG_TMP_VEC  TCG_REG_XMM5
112
113static const int tcg_target_call_iarg_regs[] = {
114#if TCG_TARGET_REG_BITS == 64
115#if defined(_WIN64)
116    TCG_REG_RCX,
117    TCG_REG_RDX,
118#else
119    TCG_REG_RDI,
120    TCG_REG_RSI,
121    TCG_REG_RDX,
122    TCG_REG_RCX,
123#endif
124    TCG_REG_R8,
125    TCG_REG_R9,
126#else
127    /* 32 bit mode uses stack based calling convention (GCC default). */
128#endif
129};
130
131static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
132{
133    switch (kind) {
134    case TCG_CALL_RET_NORMAL:
135        tcg_debug_assert(slot >= 0 && slot <= 1);
136        return slot ? TCG_REG_EDX : TCG_REG_EAX;
137#ifdef _WIN64
138    case TCG_CALL_RET_BY_VEC:
139        tcg_debug_assert(slot == 0);
140        return TCG_REG_XMM0;
141#endif
142    default:
143        g_assert_not_reached();
144    }
145}
146
147/* Constants we accept.  */
148#define TCG_CT_CONST_S32 0x100
149#define TCG_CT_CONST_U32 0x200
150#define TCG_CT_CONST_I32 0x400
151#define TCG_CT_CONST_WSZ 0x800
152#define TCG_CT_CONST_TST 0x1000
153#define TCG_CT_CONST_ZERO 0x2000
154
155/* Registers used with L constraint, which are the first argument
156   registers on x86_64, and two random call clobbered registers on
157   i386. */
158#if TCG_TARGET_REG_BITS == 64
159# define TCG_REG_L0 tcg_target_call_iarg_regs[0]
160# define TCG_REG_L1 tcg_target_call_iarg_regs[1]
161#else
162# define TCG_REG_L0 TCG_REG_EAX
163# define TCG_REG_L1 TCG_REG_EDX
164#endif
165
166#if TCG_TARGET_REG_BITS == 64
167# define ALL_GENERAL_REGS      0x0000ffffu
168# define ALL_VECTOR_REGS       0xffff0000u
169# define ALL_BYTEL_REGS        ALL_GENERAL_REGS
170#else
171# define ALL_GENERAL_REGS      0x000000ffu
172# define ALL_VECTOR_REGS       0x00ff0000u
173# define ALL_BYTEL_REGS        0x0000000fu
174#endif
175#define SOFTMMU_RESERVE_REGS \
176    (tcg_use_softmmu ? (1 << TCG_REG_L0) | (1 << TCG_REG_L1) : 0)
177
178#define have_bmi2       (cpuinfo & CPUINFO_BMI2)
179#define have_lzcnt      (cpuinfo & CPUINFO_LZCNT)
180
181static const tcg_insn_unit *tb_ret_addr;
182
183static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
184                        intptr_t value, intptr_t addend)
185{
186    value += addend;
187    switch(type) {
188    case R_386_PC32:
189        value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
190        if (value != (int32_t)value) {
191            return false;
192        }
193        /* FALLTHRU */
194    case R_386_32:
195        tcg_patch32(code_ptr, value);
196        break;
197    case R_386_PC8:
198        value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
199        if (value != (int8_t)value) {
200            return false;
201        }
202        tcg_patch8(code_ptr, value);
203        break;
204    default:
205        g_assert_not_reached();
206    }
207    return true;
208}
209
210/* test if a constant matches the constraint */
211static bool tcg_target_const_match(int64_t val, int ct,
212                                   TCGType type, TCGCond cond, int vece)
213{
214    if (ct & TCG_CT_CONST) {
215        return 1;
216    }
217    if (type == TCG_TYPE_I32) {
218        if (ct & (TCG_CT_CONST_S32 | TCG_CT_CONST_U32 |
219                  TCG_CT_CONST_I32 | TCG_CT_CONST_TST)) {
220            return 1;
221        }
222    } else {
223        if ((ct & TCG_CT_CONST_S32) && val == (int32_t)val) {
224            return 1;
225        }
226        if ((ct & TCG_CT_CONST_U32) && val == (uint32_t)val) {
227            return 1;
228        }
229        if ((ct & TCG_CT_CONST_I32) && ~val == (int32_t)~val) {
230            return 1;
231        }
232        /*
233         * This will be used in combination with TCG_CT_CONST_S32,
234         * so "normal" TESTQ is already matched.  Also accept:
235         *    TESTQ -> TESTL   (uint32_t)
236         *    TESTQ -> BT      (is_power_of_2)
237         */
238        if ((ct & TCG_CT_CONST_TST)
239            && is_tst_cond(cond)
240            && (val == (uint32_t)val || is_power_of_2(val))) {
241            return 1;
242        }
243    }
244    if ((ct & TCG_CT_CONST_WSZ) && val == (type == TCG_TYPE_I32 ? 32 : 64)) {
245        return 1;
246    }
247    if ((ct & TCG_CT_CONST_ZERO) && val == 0) {
248        return 1;
249    }
250    return 0;
251}
252
253# define LOWREGMASK(x)	((x) & 7)
254
255#define P_EXT		0x100		/* 0x0f opcode prefix */
256#define P_EXT38         0x200           /* 0x0f 0x38 opcode prefix */
257#define P_DATA16        0x400           /* 0x66 opcode prefix */
258#define P_VEXW          0x1000          /* Set VEX.W = 1 */
259#if TCG_TARGET_REG_BITS == 64
260# define P_REXW         P_VEXW          /* Set REX.W = 1; match VEXW */
261# define P_REXB_R       0x2000          /* REG field as byte register */
262# define P_REXB_RM      0x4000          /* R/M field as byte register */
263# define P_GS           0x8000          /* gs segment override */
264#else
265# define P_REXW		0
266# define P_REXB_R	0
267# define P_REXB_RM	0
268# define P_GS           0
269#endif
270#define P_EXT3A         0x10000         /* 0x0f 0x3a opcode prefix */
271#define P_SIMDF3        0x20000         /* 0xf3 opcode prefix */
272#define P_SIMDF2        0x40000         /* 0xf2 opcode prefix */
273#define P_VEXL          0x80000         /* Set VEX.L = 1 */
274#define P_EVEX          0x100000        /* Requires EVEX encoding */
275
276#define OPC_ARITH_EbIb	(0x80)
277#define OPC_ARITH_EvIz	(0x81)
278#define OPC_ARITH_EvIb	(0x83)
279#define OPC_ARITH_GvEv	(0x03)		/* ... plus (ARITH_FOO << 3) */
280#define OPC_ANDN        (0xf2 | P_EXT38)
281#define OPC_ADD_GvEv	(OPC_ARITH_GvEv | (ARITH_ADD << 3))
282#define OPC_AND_GvEv    (OPC_ARITH_GvEv | (ARITH_AND << 3))
283#define OPC_BLENDPS     (0x0c | P_EXT3A | P_DATA16)
284#define OPC_BSF         (0xbc | P_EXT)
285#define OPC_BSR         (0xbd | P_EXT)
286#define OPC_BSWAP	(0xc8 | P_EXT)
287#define OPC_CALL_Jz	(0xe8)
288#define OPC_CMOVCC      (0x40 | P_EXT)  /* ... plus condition code */
289#define OPC_CMP_GvEv	(OPC_ARITH_GvEv | (ARITH_CMP << 3))
290#define OPC_DEC_r32	(0x48)
291#define OPC_IMUL_GvEv	(0xaf | P_EXT)
292#define OPC_IMUL_GvEvIb	(0x6b)
293#define OPC_IMUL_GvEvIz	(0x69)
294#define OPC_INC_r32	(0x40)
295#define OPC_JCC_long	(0x80 | P_EXT)	/* ... plus condition code */
296#define OPC_JCC_short	(0x70)		/* ... plus condition code */
297#define OPC_JMP_long	(0xe9)
298#define OPC_JMP_short	(0xeb)
299#define OPC_LEA         (0x8d)
300#define OPC_LZCNT       (0xbd | P_EXT | P_SIMDF3)
301#define OPC_MOVB_EvGv	(0x88)		/* stores, more or less */
302#define OPC_MOVL_EvGv	(0x89)		/* stores, more or less */
303#define OPC_MOVL_GvEv	(0x8b)		/* loads, more or less */
304#define OPC_MOVB_EvIz   (0xc6)
305#define OPC_MOVL_EvIz	(0xc7)
306#define OPC_MOVB_Ib     (0xb0)
307#define OPC_MOVL_Iv     (0xb8)
308#define OPC_MOVBE_GyMy  (0xf0 | P_EXT38)
309#define OPC_MOVBE_MyGy  (0xf1 | P_EXT38)
310#define OPC_MOVD_VyEy   (0x6e | P_EXT | P_DATA16)
311#define OPC_MOVD_EyVy   (0x7e | P_EXT | P_DATA16)
312#define OPC_MOVDDUP     (0x12 | P_EXT | P_SIMDF2)
313#define OPC_MOVDQA_VxWx (0x6f | P_EXT | P_DATA16)
314#define OPC_MOVDQA_WxVx (0x7f | P_EXT | P_DATA16)
315#define OPC_MOVDQU_VxWx (0x6f | P_EXT | P_SIMDF3)
316#define OPC_MOVDQU_WxVx (0x7f | P_EXT | P_SIMDF3)
317#define OPC_MOVQ_VqWq   (0x7e | P_EXT | P_SIMDF3)
318#define OPC_MOVQ_WqVq   (0xd6 | P_EXT | P_DATA16)
319#define OPC_MOVSBL	(0xbe | P_EXT)
320#define OPC_MOVSWL	(0xbf | P_EXT)
321#define OPC_MOVSLQ	(0x63 | P_REXW)
322#define OPC_MOVZBL	(0xb6 | P_EXT)
323#define OPC_MOVZWL	(0xb7 | P_EXT)
324#define OPC_PABSB       (0x1c | P_EXT38 | P_DATA16)
325#define OPC_PABSW       (0x1d | P_EXT38 | P_DATA16)
326#define OPC_PABSD       (0x1e | P_EXT38 | P_DATA16)
327#define OPC_VPABSQ      (0x1f | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
328#define OPC_PACKSSDW    (0x6b | P_EXT | P_DATA16)
329#define OPC_PACKSSWB    (0x63 | P_EXT | P_DATA16)
330#define OPC_PACKUSDW    (0x2b | P_EXT38 | P_DATA16)
331#define OPC_PACKUSWB    (0x67 | P_EXT | P_DATA16)
332#define OPC_PADDB       (0xfc | P_EXT | P_DATA16)
333#define OPC_PADDW       (0xfd | P_EXT | P_DATA16)
334#define OPC_PADDD       (0xfe | P_EXT | P_DATA16)
335#define OPC_PADDQ       (0xd4 | P_EXT | P_DATA16)
336#define OPC_PADDSB      (0xec | P_EXT | P_DATA16)
337#define OPC_PADDSW      (0xed | P_EXT | P_DATA16)
338#define OPC_PADDUB      (0xdc | P_EXT | P_DATA16)
339#define OPC_PADDUW      (0xdd | P_EXT | P_DATA16)
340#define OPC_PAND        (0xdb | P_EXT | P_DATA16)
341#define OPC_PANDN       (0xdf | P_EXT | P_DATA16)
342#define OPC_PBLENDW     (0x0e | P_EXT3A | P_DATA16)
343#define OPC_PCMPEQB     (0x74 | P_EXT | P_DATA16)
344#define OPC_PCMPEQW     (0x75 | P_EXT | P_DATA16)
345#define OPC_PCMPEQD     (0x76 | P_EXT | P_DATA16)
346#define OPC_PCMPEQQ     (0x29 | P_EXT38 | P_DATA16)
347#define OPC_PCMPGTB     (0x64 | P_EXT | P_DATA16)
348#define OPC_PCMPGTW     (0x65 | P_EXT | P_DATA16)
349#define OPC_PCMPGTD     (0x66 | P_EXT | P_DATA16)
350#define OPC_PCMPGTQ     (0x37 | P_EXT38 | P_DATA16)
351#define OPC_PEXTRD      (0x16 | P_EXT3A | P_DATA16)
352#define OPC_PINSRD      (0x22 | P_EXT3A | P_DATA16)
353#define OPC_PMAXSB      (0x3c | P_EXT38 | P_DATA16)
354#define OPC_PMAXSW      (0xee | P_EXT | P_DATA16)
355#define OPC_PMAXSD      (0x3d | P_EXT38 | P_DATA16)
356#define OPC_VPMAXSQ     (0x3d | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
357#define OPC_PMAXUB      (0xde | P_EXT | P_DATA16)
358#define OPC_PMAXUW      (0x3e | P_EXT38 | P_DATA16)
359#define OPC_PMAXUD      (0x3f | P_EXT38 | P_DATA16)
360#define OPC_VPMAXUQ     (0x3f | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
361#define OPC_PMINSB      (0x38 | P_EXT38 | P_DATA16)
362#define OPC_PMINSW      (0xea | P_EXT | P_DATA16)
363#define OPC_PMINSD      (0x39 | P_EXT38 | P_DATA16)
364#define OPC_VPMINSQ     (0x39 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
365#define OPC_PMINUB      (0xda | P_EXT | P_DATA16)
366#define OPC_PMINUW      (0x3a | P_EXT38 | P_DATA16)
367#define OPC_PMINUD      (0x3b | P_EXT38 | P_DATA16)
368#define OPC_VPMINUQ     (0x3b | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
369#define OPC_PMOVSXBW    (0x20 | P_EXT38 | P_DATA16)
370#define OPC_PMOVSXWD    (0x23 | P_EXT38 | P_DATA16)
371#define OPC_PMOVSXDQ    (0x25 | P_EXT38 | P_DATA16)
372#define OPC_PMOVZXBW    (0x30 | P_EXT38 | P_DATA16)
373#define OPC_PMOVZXWD    (0x33 | P_EXT38 | P_DATA16)
374#define OPC_PMOVZXDQ    (0x35 | P_EXT38 | P_DATA16)
375#define OPC_PMULLW      (0xd5 | P_EXT | P_DATA16)
376#define OPC_PMULLD      (0x40 | P_EXT38 | P_DATA16)
377#define OPC_VPMULLQ     (0x40 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
378#define OPC_POR         (0xeb | P_EXT | P_DATA16)
379#define OPC_PSHUFB      (0x00 | P_EXT38 | P_DATA16)
380#define OPC_PSHUFD      (0x70 | P_EXT | P_DATA16)
381#define OPC_PSHUFLW     (0x70 | P_EXT | P_SIMDF2)
382#define OPC_PSHUFHW     (0x70 | P_EXT | P_SIMDF3)
383#define OPC_PSHIFTW_Ib  (0x71 | P_EXT | P_DATA16) /* /2 /6 /4 */
384#define OPC_PSHIFTD_Ib  (0x72 | P_EXT | P_DATA16) /* /1 /2 /6 /4 */
385#define OPC_PSHIFTQ_Ib  (0x73 | P_EXT | P_DATA16) /* /2 /6 /4 */
386#define OPC_PSLLW       (0xf1 | P_EXT | P_DATA16)
387#define OPC_PSLLD       (0xf2 | P_EXT | P_DATA16)
388#define OPC_PSLLQ       (0xf3 | P_EXT | P_DATA16)
389#define OPC_PSRAW       (0xe1 | P_EXT | P_DATA16)
390#define OPC_PSRAD       (0xe2 | P_EXT | P_DATA16)
391#define OPC_VPSRAQ      (0xe2 | P_EXT | P_DATA16 | P_VEXW | P_EVEX)
392#define OPC_PSRLW       (0xd1 | P_EXT | P_DATA16)
393#define OPC_PSRLD       (0xd2 | P_EXT | P_DATA16)
394#define OPC_PSRLQ       (0xd3 | P_EXT | P_DATA16)
395#define OPC_PSUBB       (0xf8 | P_EXT | P_DATA16)
396#define OPC_PSUBW       (0xf9 | P_EXT | P_DATA16)
397#define OPC_PSUBD       (0xfa | P_EXT | P_DATA16)
398#define OPC_PSUBQ       (0xfb | P_EXT | P_DATA16)
399#define OPC_PSUBSB      (0xe8 | P_EXT | P_DATA16)
400#define OPC_PSUBSW      (0xe9 | P_EXT | P_DATA16)
401#define OPC_PSUBUB      (0xd8 | P_EXT | P_DATA16)
402#define OPC_PSUBUW      (0xd9 | P_EXT | P_DATA16)
403#define OPC_PUNPCKLBW   (0x60 | P_EXT | P_DATA16)
404#define OPC_PUNPCKLWD   (0x61 | P_EXT | P_DATA16)
405#define OPC_PUNPCKLDQ   (0x62 | P_EXT | P_DATA16)
406#define OPC_PUNPCKLQDQ  (0x6c | P_EXT | P_DATA16)
407#define OPC_PUNPCKHBW   (0x68 | P_EXT | P_DATA16)
408#define OPC_PUNPCKHWD   (0x69 | P_EXT | P_DATA16)
409#define OPC_PUNPCKHDQ   (0x6a | P_EXT | P_DATA16)
410#define OPC_PUNPCKHQDQ  (0x6d | P_EXT | P_DATA16)
411#define OPC_PXOR        (0xef | P_EXT | P_DATA16)
412#define OPC_POP_r32	(0x58)
413#define OPC_POPCNT      (0xb8 | P_EXT | P_SIMDF3)
414#define OPC_PUSH_r32	(0x50)
415#define OPC_PUSH_Iv	(0x68)
416#define OPC_PUSH_Ib	(0x6a)
417#define OPC_RET		(0xc3)
418#define OPC_SETCC	(0x90 | P_EXT | P_REXB_RM) /* ... plus cc */
419#define OPC_SHIFT_1	(0xd1)
420#define OPC_SHIFT_Ib	(0xc1)
421#define OPC_SHIFT_cl	(0xd3)
422#define OPC_SARX        (0xf7 | P_EXT38 | P_SIMDF3)
423#define OPC_SHUFPS      (0xc6 | P_EXT)
424#define OPC_SHLX        (0xf7 | P_EXT38 | P_DATA16)
425#define OPC_SHRX        (0xf7 | P_EXT38 | P_SIMDF2)
426#define OPC_SHRD_Ib     (0xac | P_EXT)
427#define OPC_TESTB	(0x84)
428#define OPC_TESTL	(0x85)
429#define OPC_TZCNT       (0xbc | P_EXT | P_SIMDF3)
430#define OPC_UD2         (0x0b | P_EXT)
431#define OPC_VPBLENDD    (0x02 | P_EXT3A | P_DATA16)
432#define OPC_VPBLENDVB   (0x4c | P_EXT3A | P_DATA16)
433#define OPC_VPBLENDMB   (0x66 | P_EXT38 | P_DATA16 | P_EVEX)
434#define OPC_VPBLENDMW   (0x66 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
435#define OPC_VPBLENDMD   (0x64 | P_EXT38 | P_DATA16 | P_EVEX)
436#define OPC_VPBLENDMQ   (0x64 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
437#define OPC_VPCMPB      (0x3f | P_EXT3A | P_DATA16 | P_EVEX)
438#define OPC_VPCMPUB     (0x3e | P_EXT3A | P_DATA16 | P_EVEX)
439#define OPC_VPCMPW      (0x3f | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
440#define OPC_VPCMPUW     (0x3e | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
441#define OPC_VPCMPD      (0x1f | P_EXT3A | P_DATA16 | P_EVEX)
442#define OPC_VPCMPUD     (0x1e | P_EXT3A | P_DATA16 | P_EVEX)
443#define OPC_VPCMPQ      (0x1f | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
444#define OPC_VPCMPUQ     (0x1e | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
445#define OPC_VPINSRB     (0x20 | P_EXT3A | P_DATA16)
446#define OPC_VPINSRW     (0xc4 | P_EXT | P_DATA16)
447#define OPC_VBROADCASTSS (0x18 | P_EXT38 | P_DATA16)
448#define OPC_VBROADCASTSD (0x19 | P_EXT38 | P_DATA16)
449#define OPC_VPBROADCASTB (0x78 | P_EXT38 | P_DATA16)
450#define OPC_VPBROADCASTW (0x79 | P_EXT38 | P_DATA16)
451#define OPC_VPBROADCASTD (0x58 | P_EXT38 | P_DATA16)
452#define OPC_VPBROADCASTQ (0x59 | P_EXT38 | P_DATA16)
453#define OPC_VPMOVM2B    (0x28 | P_EXT38 | P_SIMDF3 | P_EVEX)
454#define OPC_VPMOVM2W    (0x28 | P_EXT38 | P_SIMDF3 | P_VEXW | P_EVEX)
455#define OPC_VPMOVM2D    (0x38 | P_EXT38 | P_SIMDF3 | P_EVEX)
456#define OPC_VPMOVM2Q    (0x38 | P_EXT38 | P_SIMDF3 | P_VEXW | P_EVEX)
457#define OPC_VPERMQ      (0x00 | P_EXT3A | P_DATA16 | P_VEXW)
458#define OPC_VPERM2I128  (0x46 | P_EXT3A | P_DATA16 | P_VEXL)
459#define OPC_VPROLVD     (0x15 | P_EXT38 | P_DATA16 | P_EVEX)
460#define OPC_VPROLVQ     (0x15 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
461#define OPC_VPRORVD     (0x14 | P_EXT38 | P_DATA16 | P_EVEX)
462#define OPC_VPRORVQ     (0x14 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
463#define OPC_VPSHLDW     (0x70 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
464#define OPC_VPSHLDD     (0x71 | P_EXT3A | P_DATA16 | P_EVEX)
465#define OPC_VPSHLDQ     (0x71 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
466#define OPC_VPSHLDVW    (0x70 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
467#define OPC_VPSHLDVD    (0x71 | P_EXT38 | P_DATA16 | P_EVEX)
468#define OPC_VPSHLDVQ    (0x71 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
469#define OPC_VPSHRDVW    (0x72 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
470#define OPC_VPSHRDVD    (0x73 | P_EXT38 | P_DATA16 | P_EVEX)
471#define OPC_VPSHRDVQ    (0x73 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
472#define OPC_VPSLLVW     (0x12 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
473#define OPC_VPSLLVD     (0x47 | P_EXT38 | P_DATA16)
474#define OPC_VPSLLVQ     (0x47 | P_EXT38 | P_DATA16 | P_VEXW)
475#define OPC_VPSRAVW     (0x11 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
476#define OPC_VPSRAVD     (0x46 | P_EXT38 | P_DATA16)
477#define OPC_VPSRAVQ     (0x46 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
478#define OPC_VPSRLVW     (0x10 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
479#define OPC_VPSRLVD     (0x45 | P_EXT38 | P_DATA16)
480#define OPC_VPSRLVQ     (0x45 | P_EXT38 | P_DATA16 | P_VEXW)
481#define OPC_VPTERNLOGQ  (0x25 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
482#define OPC_VPTESTMB    (0x26 | P_EXT38 | P_DATA16 | P_EVEX)
483#define OPC_VPTESTMW    (0x26 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
484#define OPC_VPTESTMD    (0x27 | P_EXT38 | P_DATA16 | P_EVEX)
485#define OPC_VPTESTMQ    (0x27 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
486#define OPC_VPTESTNMB   (0x26 | P_EXT38 | P_SIMDF3 | P_EVEX)
487#define OPC_VPTESTNMW   (0x26 | P_EXT38 | P_SIMDF3 | P_VEXW | P_EVEX)
488#define OPC_VPTESTNMD   (0x27 | P_EXT38 | P_SIMDF3 | P_EVEX)
489#define OPC_VPTESTNMQ   (0x27 | P_EXT38 | P_SIMDF3 | P_VEXW | P_EVEX)
490#define OPC_VZEROUPPER  (0x77 | P_EXT)
491#define OPC_XCHG_ax_r32	(0x90)
492#define OPC_XCHG_EvGv   (0x87)
493
494#define OPC_GRP3_Eb     (0xf6)
495#define OPC_GRP3_Ev     (0xf7)
496#define OPC_GRP5        (0xff)
497#define OPC_GRP14       (0x73 | P_EXT | P_DATA16)
498#define OPC_GRPBT       (0xba | P_EXT)
499
500#define OPC_GRPBT_BT    4
501#define OPC_GRPBT_BTS   5
502#define OPC_GRPBT_BTR   6
503#define OPC_GRPBT_BTC   7
504
505/* Group 1 opcode extensions for 0x80-0x83.
506   These are also used as modifiers for OPC_ARITH.  */
507#define ARITH_ADD 0
508#define ARITH_OR  1
509#define ARITH_ADC 2
510#define ARITH_SBB 3
511#define ARITH_AND 4
512#define ARITH_SUB 5
513#define ARITH_XOR 6
514#define ARITH_CMP 7
515
516/* Group 2 opcode extensions for 0xc0, 0xc1, 0xd0-0xd3.  */
517#define SHIFT_ROL 0
518#define SHIFT_ROR 1
519#define SHIFT_SHL 4
520#define SHIFT_SHR 5
521#define SHIFT_SAR 7
522
523/* Group 3 opcode extensions for 0xf6, 0xf7.  To be used with OPC_GRP3.  */
524#define EXT3_TESTi 0
525#define EXT3_NOT   2
526#define EXT3_NEG   3
527#define EXT3_MUL   4
528#define EXT3_IMUL  5
529#define EXT3_DIV   6
530#define EXT3_IDIV  7
531
532/* Group 5 opcode extensions for 0xff.  To be used with OPC_GRP5.  */
533#define EXT5_INC_Ev	0
534#define EXT5_DEC_Ev	1
535#define EXT5_CALLN_Ev	2
536#define EXT5_JMPN_Ev	4
537
538/* Condition codes to be added to OPC_JCC_{long,short}.  */
539#define JCC_JMP (-1)
540#define JCC_JO  0x0
541#define JCC_JNO 0x1
542#define JCC_JB  0x2
543#define JCC_JAE 0x3
544#define JCC_JE  0x4
545#define JCC_JNE 0x5
546#define JCC_JBE 0x6
547#define JCC_JA  0x7
548#define JCC_JS  0x8
549#define JCC_JNS 0x9
550#define JCC_JP  0xa
551#define JCC_JNP 0xb
552#define JCC_JL  0xc
553#define JCC_JGE 0xd
554#define JCC_JLE 0xe
555#define JCC_JG  0xf
556
557static const uint8_t tcg_cond_to_jcc[] = {
558    [TCG_COND_EQ] = JCC_JE,
559    [TCG_COND_NE] = JCC_JNE,
560    [TCG_COND_LT] = JCC_JL,
561    [TCG_COND_GE] = JCC_JGE,
562    [TCG_COND_LE] = JCC_JLE,
563    [TCG_COND_GT] = JCC_JG,
564    [TCG_COND_LTU] = JCC_JB,
565    [TCG_COND_GEU] = JCC_JAE,
566    [TCG_COND_LEU] = JCC_JBE,
567    [TCG_COND_GTU] = JCC_JA,
568    [TCG_COND_TSTEQ] = JCC_JE,
569    [TCG_COND_TSTNE] = JCC_JNE,
570};
571
572#if TCG_TARGET_REG_BITS == 64
573static void tcg_out_opc(TCGContext *s, int opc, int r, int rm, int x)
574{
575    int rex;
576
577    if (opc & P_GS) {
578        tcg_out8(s, 0x65);
579    }
580    if (opc & P_DATA16) {
581        /* We should never be asking for both 16 and 64-bit operation.  */
582        tcg_debug_assert((opc & P_REXW) == 0);
583        tcg_out8(s, 0x66);
584    }
585    if (opc & P_SIMDF3) {
586        tcg_out8(s, 0xf3);
587    } else if (opc & P_SIMDF2) {
588        tcg_out8(s, 0xf2);
589    }
590
591    rex = 0;
592    rex |= (opc & P_REXW) ? 0x8 : 0x0;  /* REX.W */
593    rex |= (r & 8) >> 1;                /* REX.R */
594    rex |= (x & 8) >> 2;                /* REX.X */
595    rex |= (rm & 8) >> 3;               /* REX.B */
596
597    /* P_REXB_{R,RM} indicates that the given register is the low byte.
598       For %[abcd]l we need no REX prefix, but for %{si,di,bp,sp}l we do,
599       as otherwise the encoding indicates %[abcd]h.  Note that the values
600       that are ORed in merely indicate that the REX byte must be present;
601       those bits get discarded in output.  */
602    rex |= opc & (r >= 4 ? P_REXB_R : 0);
603    rex |= opc & (rm >= 4 ? P_REXB_RM : 0);
604
605    if (rex) {
606        tcg_out8(s, (uint8_t)(rex | 0x40));
607    }
608
609    if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
610        tcg_out8(s, 0x0f);
611        if (opc & P_EXT38) {
612            tcg_out8(s, 0x38);
613        } else if (opc & P_EXT3A) {
614            tcg_out8(s, 0x3a);
615        }
616    }
617
618    tcg_out8(s, opc);
619}
620#else
621static void tcg_out_opc(TCGContext *s, int opc)
622{
623    if (opc & P_DATA16) {
624        tcg_out8(s, 0x66);
625    }
626    if (opc & P_SIMDF3) {
627        tcg_out8(s, 0xf3);
628    } else if (opc & P_SIMDF2) {
629        tcg_out8(s, 0xf2);
630    }
631    if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
632        tcg_out8(s, 0x0f);
633        if (opc & P_EXT38) {
634            tcg_out8(s, 0x38);
635        } else if (opc & P_EXT3A) {
636            tcg_out8(s, 0x3a);
637        }
638    }
639    tcg_out8(s, opc);
640}
641/* Discard the register arguments to tcg_out_opc early, so as not to penalize
642   the 32-bit compilation paths.  This method works with all versions of gcc,
643   whereas relying on optimization may not be able to exclude them.  */
644#define tcg_out_opc(s, opc, r, rm, x)  (tcg_out_opc)(s, opc)
645#endif
646
647static void tcg_out_modrm(TCGContext *s, int opc, int r, int rm)
648{
649    tcg_out_opc(s, opc, r, rm, 0);
650    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
651}
652
653static void tcg_out_vex_opc(TCGContext *s, int opc, int r, int v,
654                            int rm, int index)
655{
656    int tmp;
657
658    if (opc & P_GS) {
659        tcg_out8(s, 0x65);
660    }
661    /* Use the two byte form if possible, which cannot encode
662       VEX.W, VEX.B, VEX.X, or an m-mmmm field other than P_EXT.  */
663    if ((opc & (P_EXT | P_EXT38 | P_EXT3A | P_VEXW)) == P_EXT
664        && ((rm | index) & 8) == 0) {
665        /* Two byte VEX prefix.  */
666        tcg_out8(s, 0xc5);
667
668        tmp = (r & 8 ? 0 : 0x80);              /* VEX.R */
669    } else {
670        /* Three byte VEX prefix.  */
671        tcg_out8(s, 0xc4);
672
673        /* VEX.m-mmmm */
674        if (opc & P_EXT3A) {
675            tmp = 3;
676        } else if (opc & P_EXT38) {
677            tmp = 2;
678        } else if (opc & P_EXT) {
679            tmp = 1;
680        } else {
681            g_assert_not_reached();
682        }
683        tmp |= (r & 8 ? 0 : 0x80);             /* VEX.R */
684        tmp |= (index & 8 ? 0 : 0x40);         /* VEX.X */
685        tmp |= (rm & 8 ? 0 : 0x20);            /* VEX.B */
686        tcg_out8(s, tmp);
687
688        tmp = (opc & P_VEXW ? 0x80 : 0);       /* VEX.W */
689    }
690
691    tmp |= (opc & P_VEXL ? 0x04 : 0);      /* VEX.L */
692    /* VEX.pp */
693    if (opc & P_DATA16) {
694        tmp |= 1;                          /* 0x66 */
695    } else if (opc & P_SIMDF3) {
696        tmp |= 2;                          /* 0xf3 */
697    } else if (opc & P_SIMDF2) {
698        tmp |= 3;                          /* 0xf2 */
699    }
700    tmp |= (~v & 15) << 3;                 /* VEX.vvvv */
701    tcg_out8(s, tmp);
702    tcg_out8(s, opc);
703}
704
705static void tcg_out_evex_opc(TCGContext *s, int opc, int r, int v,
706                             int rm, int index, int aaa, bool z)
707{
708    /* The entire 4-byte evex prefix; with R' and V' set. */
709    uint32_t p = 0x08041062;
710    int mm, pp;
711
712    tcg_debug_assert(have_avx512vl);
713
714    /* EVEX.mm */
715    if (opc & P_EXT3A) {
716        mm = 3;
717    } else if (opc & P_EXT38) {
718        mm = 2;
719    } else if (opc & P_EXT) {
720        mm = 1;
721    } else {
722        g_assert_not_reached();
723    }
724
725    /* EVEX.pp */
726    if (opc & P_DATA16) {
727        pp = 1;                          /* 0x66 */
728    } else if (opc & P_SIMDF3) {
729        pp = 2;                          /* 0xf3 */
730    } else if (opc & P_SIMDF2) {
731        pp = 3;                          /* 0xf2 */
732    } else {
733        pp = 0;
734    }
735
736    p = deposit32(p, 8, 2, mm);
737    p = deposit32(p, 13, 1, (rm & 8) == 0);             /* EVEX.RXB.B */
738    p = deposit32(p, 14, 1, (index & 8) == 0);          /* EVEX.RXB.X */
739    p = deposit32(p, 15, 1, (r & 8) == 0);              /* EVEX.RXB.R */
740    p = deposit32(p, 16, 2, pp);
741    p = deposit32(p, 19, 4, ~v);
742    p = deposit32(p, 23, 1, (opc & P_VEXW) != 0);
743    p = deposit32(p, 24, 3, aaa);
744    p = deposit32(p, 29, 2, (opc & P_VEXL) != 0);
745    p = deposit32(p, 31, 1, z);
746
747    tcg_out32(s, p);
748    tcg_out8(s, opc);
749}
750
751static void tcg_out_vex_modrm(TCGContext *s, int opc, int r, int v, int rm)
752{
753    if (opc & P_EVEX) {
754        tcg_out_evex_opc(s, opc, r, v, rm, 0, 0, false);
755    } else {
756        tcg_out_vex_opc(s, opc, r, v, rm, 0);
757    }
758    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
759}
760
761static void tcg_out_vex_modrm_type(TCGContext *s, int opc,
762                                   int r, int v, int rm, TCGType type)
763{
764    if (type == TCG_TYPE_V256) {
765        opc |= P_VEXL;
766    }
767    tcg_out_vex_modrm(s, opc, r, v, rm);
768}
769
770static void tcg_out_evex_modrm_type(TCGContext *s, int opc, int r, int v,
771                                    int rm, int aaa, bool z, TCGType type)
772{
773    if (type == TCG_TYPE_V256) {
774        opc |= P_VEXL;
775    }
776    tcg_out_evex_opc(s, opc, r, v, rm, 0, aaa, z);
777    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
778}
779
780/* Output an opcode with a full "rm + (index<<shift) + offset" address mode.
781   We handle either RM and INDEX missing with a negative value.  In 64-bit
782   mode for absolute addresses, ~RM is the size of the immediate operand
783   that will follow the instruction.  */
784
785static void tcg_out_sib_offset(TCGContext *s, int r, int rm, int index,
786                               int shift, intptr_t offset)
787{
788    int mod, len;
789
790    if (index < 0 && rm < 0) {
791        if (TCG_TARGET_REG_BITS == 64) {
792            /* Try for a rip-relative addressing mode.  This has replaced
793               the 32-bit-mode absolute addressing encoding.  */
794            intptr_t pc = (intptr_t)s->code_ptr + 5 + ~rm;
795            intptr_t disp = offset - pc;
796            if (disp == (int32_t)disp) {
797                tcg_out8(s, (LOWREGMASK(r) << 3) | 5);
798                tcg_out32(s, disp);
799                return;
800            }
801
802            /* Try for an absolute address encoding.  This requires the
803               use of the MODRM+SIB encoding and is therefore larger than
804               rip-relative addressing.  */
805            if (offset == (int32_t)offset) {
806                tcg_out8(s, (LOWREGMASK(r) << 3) | 4);
807                tcg_out8(s, (4 << 3) | 5);
808                tcg_out32(s, offset);
809                return;
810            }
811
812            /* ??? The memory isn't directly addressable.  */
813            g_assert_not_reached();
814        } else {
815            /* Absolute address.  */
816            tcg_out8(s, (r << 3) | 5);
817            tcg_out32(s, offset);
818            return;
819        }
820    }
821
822    /* Find the length of the immediate addend.  Note that the encoding
823       that would be used for (%ebp) indicates absolute addressing.  */
824    if (rm < 0) {
825        mod = 0, len = 4, rm = 5;
826    } else if (offset == 0 && LOWREGMASK(rm) != TCG_REG_EBP) {
827        mod = 0, len = 0;
828    } else if (offset == (int8_t)offset) {
829        mod = 0x40, len = 1;
830    } else {
831        mod = 0x80, len = 4;
832    }
833
834    /* Use a single byte MODRM format if possible.  Note that the encoding
835       that would be used for %esp is the escape to the two byte form.  */
836    if (index < 0 && LOWREGMASK(rm) != TCG_REG_ESP) {
837        /* Single byte MODRM format.  */
838        tcg_out8(s, mod | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
839    } else {
840        /* Two byte MODRM+SIB format.  */
841
842        /* Note that the encoding that would place %esp into the index
843           field indicates no index register.  In 64-bit mode, the REX.X
844           bit counts, so %r12 can be used as the index.  */
845        if (index < 0) {
846            index = 4;
847        } else {
848            tcg_debug_assert(index != TCG_REG_ESP);
849        }
850
851        tcg_out8(s, mod | (LOWREGMASK(r) << 3) | 4);
852        tcg_out8(s, (shift << 6) | (LOWREGMASK(index) << 3) | LOWREGMASK(rm));
853    }
854
855    if (len == 1) {
856        tcg_out8(s, offset);
857    } else if (len == 4) {
858        tcg_out32(s, offset);
859    }
860}
861
862static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm,
863                                     int index, int shift, intptr_t offset)
864{
865    tcg_out_opc(s, opc, r, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
866    tcg_out_sib_offset(s, r, rm, index, shift, offset);
867}
868
869static void tcg_out_vex_modrm_sib_offset(TCGContext *s, int opc, int r, int v,
870                                         int rm, int index, int shift,
871                                         intptr_t offset)
872{
873    tcg_out_vex_opc(s, opc, r, v, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
874    tcg_out_sib_offset(s, r, rm, index, shift, offset);
875}
876
877/* A simplification of the above with no index or shift.  */
878static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r,
879                                        int rm, intptr_t offset)
880{
881    tcg_out_modrm_sib_offset(s, opc, r, rm, -1, 0, offset);
882}
883
884static inline void tcg_out_vex_modrm_offset(TCGContext *s, int opc, int r,
885                                            int v, int rm, intptr_t offset)
886{
887    tcg_out_vex_modrm_sib_offset(s, opc, r, v, rm, -1, 0, offset);
888}
889
890/* Output an opcode with an expected reference to the constant pool.  */
891static inline void tcg_out_modrm_pool(TCGContext *s, int opc, int r)
892{
893    tcg_out_opc(s, opc, r, 0, 0);
894    /* Absolute for 32-bit, pc-relative for 64-bit.  */
895    tcg_out8(s, LOWREGMASK(r) << 3 | 5);
896    tcg_out32(s, 0);
897}
898
899/* Output an opcode with an expected reference to the constant pool.  */
900static inline void tcg_out_vex_modrm_pool(TCGContext *s, int opc, int r)
901{
902    tcg_out_vex_opc(s, opc, r, 0, 0, 0);
903    /* Absolute for 32-bit, pc-relative for 64-bit.  */
904    tcg_out8(s, LOWREGMASK(r) << 3 | 5);
905    tcg_out32(s, 0);
906}
907
908/* Generate dest op= src.  Uses the same ARITH_* codes as tgen_arithi.  */
909static inline void tgen_arithr(TCGContext *s, int subop, int dest, int src)
910{
911    /* Propagate an opcode prefix, such as P_REXW.  */
912    int ext = subop & ~0x7;
913    subop &= 0x7;
914
915    tcg_out_modrm(s, OPC_ARITH_GvEv + (subop << 3) + ext, dest, src);
916}
917
918static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
919{
920    int rexw = 0;
921
922    if (arg == ret) {
923        return true;
924    }
925    switch (type) {
926    case TCG_TYPE_I64:
927        rexw = P_REXW;
928        /* fallthru */
929    case TCG_TYPE_I32:
930        if (ret < 16) {
931            if (arg < 16) {
932                tcg_out_modrm(s, OPC_MOVL_GvEv + rexw, ret, arg);
933            } else {
934                tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, arg, 0, ret);
935            }
936        } else {
937            if (arg < 16) {
938                tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, ret, 0, arg);
939            } else {
940                tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
941            }
942        }
943        break;
944
945    case TCG_TYPE_V64:
946        tcg_debug_assert(ret >= 16 && arg >= 16);
947        tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
948        break;
949    case TCG_TYPE_V128:
950        tcg_debug_assert(ret >= 16 && arg >= 16);
951        tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx, ret, 0, arg);
952        break;
953    case TCG_TYPE_V256:
954        tcg_debug_assert(ret >= 16 && arg >= 16);
955        tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx | P_VEXL, ret, 0, arg);
956        break;
957
958    default:
959        g_assert_not_reached();
960    }
961    return true;
962}
963
964static const int avx2_dup_insn[4] = {
965    OPC_VPBROADCASTB, OPC_VPBROADCASTW,
966    OPC_VPBROADCASTD, OPC_VPBROADCASTQ,
967};
968
969static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
970                            TCGReg r, TCGReg a)
971{
972    if (have_avx2) {
973        tcg_out_vex_modrm_type(s, avx2_dup_insn[vece], r, 0, a, type);
974    } else {
975        switch (vece) {
976        case MO_8:
977            /* ??? With zero in a register, use PSHUFB.  */
978            tcg_out_vex_modrm(s, OPC_PUNPCKLBW, r, a, a);
979            a = r;
980            /* FALLTHRU */
981        case MO_16:
982            tcg_out_vex_modrm(s, OPC_PUNPCKLWD, r, a, a);
983            a = r;
984            /* FALLTHRU */
985        case MO_32:
986            tcg_out_vex_modrm(s, OPC_PSHUFD, r, 0, a);
987            /* imm8 operand: all output lanes selected from input lane 0.  */
988            tcg_out8(s, 0);
989            break;
990        case MO_64:
991            tcg_out_vex_modrm(s, OPC_PUNPCKLQDQ, r, a, a);
992            break;
993        default:
994            g_assert_not_reached();
995        }
996    }
997    return true;
998}
999
1000static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
1001                             TCGReg r, TCGReg base, intptr_t offset)
1002{
1003    if (have_avx2) {
1004        int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
1005        tcg_out_vex_modrm_offset(s, avx2_dup_insn[vece] + vex_l,
1006                                 r, 0, base, offset);
1007    } else {
1008        switch (vece) {
1009        case MO_64:
1010            tcg_out_vex_modrm_offset(s, OPC_MOVDDUP, r, 0, base, offset);
1011            break;
1012        case MO_32:
1013            tcg_out_vex_modrm_offset(s, OPC_VBROADCASTSS, r, 0, base, offset);
1014            break;
1015        case MO_16:
1016            tcg_out_vex_modrm_offset(s, OPC_VPINSRW, r, r, base, offset);
1017            tcg_out8(s, 0); /* imm8 */
1018            tcg_out_dup_vec(s, type, vece, r, r);
1019            break;
1020        case MO_8:
1021            tcg_out_vex_modrm_offset(s, OPC_VPINSRB, r, r, base, offset);
1022            tcg_out8(s, 0); /* imm8 */
1023            tcg_out_dup_vec(s, type, vece, r, r);
1024            break;
1025        default:
1026            g_assert_not_reached();
1027        }
1028    }
1029    return true;
1030}
1031
1032static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
1033                             TCGReg ret, int64_t arg)
1034{
1035    int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
1036
1037    if (arg == 0) {
1038        tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
1039        return;
1040    }
1041    if (arg == -1) {
1042        tcg_out_vex_modrm(s, OPC_PCMPEQB + vex_l, ret, ret, ret);
1043        return;
1044    }
1045
1046    if (TCG_TARGET_REG_BITS == 32 && vece < MO_64) {
1047        if (have_avx2) {
1048            tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTD + vex_l, ret);
1049        } else {
1050            tcg_out_vex_modrm_pool(s, OPC_VBROADCASTSS, ret);
1051        }
1052        new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
1053    } else {
1054        if (type == TCG_TYPE_V64) {
1055            tcg_out_vex_modrm_pool(s, OPC_MOVQ_VqWq, ret);
1056        } else if (have_avx2) {
1057            tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTQ + vex_l, ret);
1058        } else {
1059            tcg_out_vex_modrm_pool(s, OPC_MOVDDUP, ret);
1060        }
1061        if (TCG_TARGET_REG_BITS == 64) {
1062            new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
1063        } else {
1064            new_pool_l2(s, R_386_32, s->code_ptr - 4, 0, arg, arg >> 32);
1065        }
1066    }
1067}
1068
1069static void tcg_out_movi_vec(TCGContext *s, TCGType type,
1070                             TCGReg ret, tcg_target_long arg)
1071{
1072    if (arg == 0) {
1073        tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
1074        return;
1075    }
1076    if (arg == -1) {
1077        tcg_out_vex_modrm(s, OPC_PCMPEQB, ret, ret, ret);
1078        return;
1079    }
1080
1081    int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
1082    tcg_out_vex_modrm_pool(s, OPC_MOVD_VyEy + rexw, ret);
1083    if (TCG_TARGET_REG_BITS == 64) {
1084        new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
1085    } else {
1086        new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
1087    }
1088}
1089
1090static void tcg_out_movi_int(TCGContext *s, TCGType type,
1091                             TCGReg ret, tcg_target_long arg)
1092{
1093    tcg_target_long diff;
1094
1095    if (arg == 0) {
1096        tgen_arithr(s, ARITH_XOR, ret, ret);
1097        return;
1098    }
1099    if (arg == (uint32_t)arg || type == TCG_TYPE_I32) {
1100        tcg_out_opc(s, OPC_MOVL_Iv + LOWREGMASK(ret), 0, ret, 0);
1101        tcg_out32(s, arg);
1102        return;
1103    }
1104    if (arg == (int32_t)arg) {
1105        tcg_out_modrm(s, OPC_MOVL_EvIz + P_REXW, 0, ret);
1106        tcg_out32(s, arg);
1107        return;
1108    }
1109
1110    /* Try a 7 byte pc-relative lea before the 10 byte movq.  */
1111    diff = tcg_pcrel_diff(s, (const void *)arg) - 7;
1112    if (diff == (int32_t)diff) {
1113        tcg_out_opc(s, OPC_LEA | P_REXW, ret, 0, 0);
1114        tcg_out8(s, (LOWREGMASK(ret) << 3) | 5);
1115        tcg_out32(s, diff);
1116        return;
1117    }
1118
1119    tcg_out_opc(s, OPC_MOVL_Iv + P_REXW + LOWREGMASK(ret), 0, ret, 0);
1120    tcg_out64(s, arg);
1121}
1122
1123static void tcg_out_movi(TCGContext *s, TCGType type,
1124                         TCGReg ret, tcg_target_long arg)
1125{
1126    switch (type) {
1127    case TCG_TYPE_I32:
1128#if TCG_TARGET_REG_BITS == 64
1129    case TCG_TYPE_I64:
1130#endif
1131        if (ret < 16) {
1132            tcg_out_movi_int(s, type, ret, arg);
1133        } else {
1134            tcg_out_movi_vec(s, type, ret, arg);
1135        }
1136        break;
1137    default:
1138        g_assert_not_reached();
1139    }
1140}
1141
1142static bool tcg_out_xchg(TCGContext *s, TCGType type, TCGReg r1, TCGReg r2)
1143{
1144    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1145    tcg_out_modrm(s, OPC_XCHG_EvGv + rexw, r1, r2);
1146    return true;
1147}
1148
1149static void tcg_out_addi_ptr(TCGContext *s, TCGReg rd, TCGReg rs,
1150                             tcg_target_long imm)
1151{
1152    /* This function is only used for passing structs by reference. */
1153    tcg_debug_assert(imm == (int32_t)imm);
1154    tcg_out_modrm_offset(s, OPC_LEA | P_REXW, rd, rs, imm);
1155}
1156
1157static inline void tcg_out_pushi(TCGContext *s, tcg_target_long val)
1158{
1159    if (val == (int8_t)val) {
1160        tcg_out_opc(s, OPC_PUSH_Ib, 0, 0, 0);
1161        tcg_out8(s, val);
1162    } else if (val == (int32_t)val) {
1163        tcg_out_opc(s, OPC_PUSH_Iv, 0, 0, 0);
1164        tcg_out32(s, val);
1165    } else {
1166        g_assert_not_reached();
1167    }
1168}
1169
1170static inline void tcg_out_mb(TCGContext *s, TCGArg a0)
1171{
1172    /* Given the strength of x86 memory ordering, we only need care for
1173       store-load ordering.  Experimentally, "lock orl $0,0(%esp)" is
1174       faster than "mfence", so don't bother with the sse insn.  */
1175    if (a0 & TCG_MO_ST_LD) {
1176        tcg_out8(s, 0xf0);
1177        tcg_out_modrm_offset(s, OPC_ARITH_EvIb, ARITH_OR, TCG_REG_ESP, 0);
1178        tcg_out8(s, 0);
1179    }
1180}
1181
1182static inline void tcg_out_push(TCGContext *s, int reg)
1183{
1184    tcg_out_opc(s, OPC_PUSH_r32 + LOWREGMASK(reg), 0, reg, 0);
1185}
1186
1187static inline void tcg_out_pop(TCGContext *s, int reg)
1188{
1189    tcg_out_opc(s, OPC_POP_r32 + LOWREGMASK(reg), 0, reg, 0);
1190}
1191
1192static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret,
1193                       TCGReg arg1, intptr_t arg2)
1194{
1195    switch (type) {
1196    case TCG_TYPE_I32:
1197        if (ret < 16) {
1198            tcg_out_modrm_offset(s, OPC_MOVL_GvEv, ret, arg1, arg2);
1199        } else {
1200            tcg_out_vex_modrm_offset(s, OPC_MOVD_VyEy, ret, 0, arg1, arg2);
1201        }
1202        break;
1203    case TCG_TYPE_I64:
1204        if (ret < 16) {
1205            tcg_out_modrm_offset(s, OPC_MOVL_GvEv | P_REXW, ret, arg1, arg2);
1206            break;
1207        }
1208        /* FALLTHRU */
1209    case TCG_TYPE_V64:
1210        /* There is no instruction that can validate 8-byte alignment.  */
1211        tcg_debug_assert(ret >= 16);
1212        tcg_out_vex_modrm_offset(s, OPC_MOVQ_VqWq, ret, 0, arg1, arg2);
1213        break;
1214    case TCG_TYPE_V128:
1215        /*
1216         * The gvec infrastructure is asserts that v128 vector loads
1217         * and stores use a 16-byte aligned offset.  Validate that the
1218         * final pointer is aligned by using an insn that will SIGSEGV.
1219         */
1220        tcg_debug_assert(ret >= 16);
1221        tcg_out_vex_modrm_offset(s, OPC_MOVDQA_VxWx, ret, 0, arg1, arg2);
1222        break;
1223    case TCG_TYPE_V256:
1224        /*
1225         * The gvec infrastructure only requires 16-byte alignment,
1226         * so here we must use an unaligned load.
1227         */
1228        tcg_debug_assert(ret >= 16);
1229        tcg_out_vex_modrm_offset(s, OPC_MOVDQU_VxWx | P_VEXL,
1230                                 ret, 0, arg1, arg2);
1231        break;
1232    default:
1233        g_assert_not_reached();
1234    }
1235}
1236
1237static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
1238                       TCGReg arg1, intptr_t arg2)
1239{
1240    switch (type) {
1241    case TCG_TYPE_I32:
1242        if (arg < 16) {
1243            tcg_out_modrm_offset(s, OPC_MOVL_EvGv, arg, arg1, arg2);
1244        } else {
1245            tcg_out_vex_modrm_offset(s, OPC_MOVD_EyVy, arg, 0, arg1, arg2);
1246        }
1247        break;
1248    case TCG_TYPE_I64:
1249        if (arg < 16) {
1250            tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_REXW, arg, arg1, arg2);
1251            break;
1252        }
1253        /* FALLTHRU */
1254    case TCG_TYPE_V64:
1255        /* There is no instruction that can validate 8-byte alignment.  */
1256        tcg_debug_assert(arg >= 16);
1257        tcg_out_vex_modrm_offset(s, OPC_MOVQ_WqVq, arg, 0, arg1, arg2);
1258        break;
1259    case TCG_TYPE_V128:
1260        /*
1261         * The gvec infrastructure is asserts that v128 vector loads
1262         * and stores use a 16-byte aligned offset.  Validate that the
1263         * final pointer is aligned by using an insn that will SIGSEGV.
1264         *
1265         * This specific instance is also used by TCG_CALL_RET_BY_VEC,
1266         * for _WIN64, which must have SSE2 but may not have AVX.
1267         */
1268        tcg_debug_assert(arg >= 16);
1269        if (have_avx1) {
1270            tcg_out_vex_modrm_offset(s, OPC_MOVDQA_WxVx, arg, 0, arg1, arg2);
1271        } else {
1272            tcg_out_modrm_offset(s, OPC_MOVDQA_WxVx, arg, arg1, arg2);
1273        }
1274        break;
1275    case TCG_TYPE_V256:
1276        /*
1277         * The gvec infrastructure only requires 16-byte alignment,
1278         * so here we must use an unaligned store.
1279         */
1280        tcg_debug_assert(arg >= 16);
1281        tcg_out_vex_modrm_offset(s, OPC_MOVDQU_WxVx | P_VEXL,
1282                                 arg, 0, arg1, arg2);
1283        break;
1284    default:
1285        g_assert_not_reached();
1286    }
1287}
1288
1289static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
1290                        TCGReg base, intptr_t ofs)
1291{
1292    int rexw = 0;
1293    if (TCG_TARGET_REG_BITS == 64 && type == TCG_TYPE_I64) {
1294        if (val != (int32_t)val) {
1295            return false;
1296        }
1297        rexw = P_REXW;
1298    } else if (type != TCG_TYPE_I32) {
1299        return false;
1300    }
1301    tcg_out_modrm_offset(s, OPC_MOVL_EvIz | rexw, 0, base, ofs);
1302    tcg_out32(s, val);
1303    return true;
1304}
1305
1306static void tcg_out_shifti(TCGContext *s, int subopc, int reg, int count)
1307{
1308    /* Propagate an opcode prefix, such as P_DATA16.  */
1309    int ext = subopc & ~0x7;
1310    subopc &= 0x7;
1311
1312    if (count == 1) {
1313        tcg_out_modrm(s, OPC_SHIFT_1 + ext, subopc, reg);
1314    } else {
1315        tcg_out_modrm(s, OPC_SHIFT_Ib + ext, subopc, reg);
1316        tcg_out8(s, count);
1317    }
1318}
1319
1320static inline void tcg_out_bswap32(TCGContext *s, int reg)
1321{
1322    tcg_out_opc(s, OPC_BSWAP + LOWREGMASK(reg), 0, reg, 0);
1323}
1324
1325static inline void tcg_out_rolw_8(TCGContext *s, int reg)
1326{
1327    tcg_out_shifti(s, SHIFT_ROL + P_DATA16, reg, 8);
1328}
1329
1330static void tcg_out_ext8u(TCGContext *s, TCGReg dest, TCGReg src)
1331{
1332    if (TCG_TARGET_REG_BITS == 32 && src >= 4) {
1333        tcg_out_mov(s, TCG_TYPE_I32, dest, src);
1334        if (dest >= 4) {
1335            tcg_out_modrm(s, OPC_ARITH_EvIz, ARITH_AND, dest);
1336            tcg_out32(s, 0xff);
1337            return;
1338        }
1339        src = dest;
1340    }
1341    tcg_out_modrm(s, OPC_MOVZBL + P_REXB_RM, dest, src);
1342}
1343
1344static void tcg_out_ext8s(TCGContext *s, TCGType type, TCGReg dest, TCGReg src)
1345{
1346    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1347
1348    if (TCG_TARGET_REG_BITS == 32 && src >= 4) {
1349        tcg_out_mov(s, TCG_TYPE_I32, dest, src);
1350        if (dest >= 4) {
1351            tcg_out_shifti(s, SHIFT_SHL, dest, 24);
1352            tcg_out_shifti(s, SHIFT_SAR, dest, 24);
1353            return;
1354        }
1355        src = dest;
1356    }
1357    tcg_out_modrm(s, OPC_MOVSBL + P_REXB_RM + rexw, dest, src);
1358}
1359
1360static void tcg_out_ext16u(TCGContext *s, TCGReg dest, TCGReg src)
1361{
1362    /* movzwl */
1363    tcg_out_modrm(s, OPC_MOVZWL, dest, src);
1364}
1365
1366static void tcg_out_ext16s(TCGContext *s, TCGType type, TCGReg dest, TCGReg src)
1367{
1368    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1369    /* movsw[lq] */
1370    tcg_out_modrm(s, OPC_MOVSWL + rexw, dest, src);
1371}
1372
1373static void tcg_out_ext32u(TCGContext *s, TCGReg dest, TCGReg src)
1374{
1375    /* 32-bit mov zero extends.  */
1376    tcg_out_modrm(s, OPC_MOVL_GvEv, dest, src);
1377}
1378
1379static void tcg_out_ext32s(TCGContext *s, TCGReg dest, TCGReg src)
1380{
1381    tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
1382    tcg_out_modrm(s, OPC_MOVSLQ, dest, src);
1383}
1384
1385static void tcg_out_exts_i32_i64(TCGContext *s, TCGReg dest, TCGReg src)
1386{
1387    tcg_out_ext32s(s, dest, src);
1388}
1389
1390static void tcg_out_extu_i32_i64(TCGContext *s, TCGReg dest, TCGReg src)
1391{
1392    if (dest != src) {
1393        tcg_out_ext32u(s, dest, src);
1394    }
1395}
1396
1397static void tcg_out_extrl_i64_i32(TCGContext *s, TCGReg dest, TCGReg src)
1398{
1399    tcg_out_ext32u(s, dest, src);
1400}
1401
1402static inline void tcg_out_bswap64(TCGContext *s, int reg)
1403{
1404    tcg_out_opc(s, OPC_BSWAP + P_REXW + LOWREGMASK(reg), 0, reg, 0);
1405}
1406
1407static void tgen_arithi(TCGContext *s, int c, int r0,
1408                        tcg_target_long val, int cf)
1409{
1410    int rexw = 0;
1411
1412    if (TCG_TARGET_REG_BITS == 64) {
1413        rexw = c & -8;
1414        c &= 7;
1415    }
1416
1417    switch (c) {
1418    case ARITH_ADD:
1419    case ARITH_SUB:
1420        if (!cf) {
1421            /*
1422             * ??? While INC is 2 bytes shorter than ADDL $1, they also induce
1423             * partial flags update stalls on Pentium4 and are not recommended
1424             * by current Intel optimization manuals.
1425             */
1426            if (val == 1 || val == -1) {
1427                int is_inc = (c == ARITH_ADD) ^ (val < 0);
1428                if (TCG_TARGET_REG_BITS == 64) {
1429                    /*
1430                     * The single-byte increment encodings are re-tasked
1431                     * as the REX prefixes.  Use the MODRM encoding.
1432                     */
1433                    tcg_out_modrm(s, OPC_GRP5 + rexw,
1434                                  (is_inc ? EXT5_INC_Ev : EXT5_DEC_Ev), r0);
1435                } else {
1436                    tcg_out8(s, (is_inc ? OPC_INC_r32 : OPC_DEC_r32) + r0);
1437                }
1438                return;
1439            }
1440            if (val == 128) {
1441                /*
1442                 * Facilitate using an 8-bit immediate.  Carry is inverted
1443                 * by this transformation, so do it only if cf == 0.
1444                 */
1445                c ^= ARITH_ADD ^ ARITH_SUB;
1446                val = -128;
1447            }
1448        }
1449        break;
1450
1451    case ARITH_AND:
1452        if (TCG_TARGET_REG_BITS == 64) {
1453            if (val == 0xffffffffu) {
1454                tcg_out_ext32u(s, r0, r0);
1455                return;
1456            }
1457            if (val == (uint32_t)val) {
1458                /* AND with no high bits set can use a 32-bit operation.  */
1459                rexw = 0;
1460            }
1461        }
1462        if (val == 0xffu && (r0 < 4 || TCG_TARGET_REG_BITS == 64)) {
1463            tcg_out_ext8u(s, r0, r0);
1464            return;
1465        }
1466        if (val == 0xffffu) {
1467            tcg_out_ext16u(s, r0, r0);
1468            return;
1469        }
1470        break;
1471
1472    case ARITH_OR:
1473    case ARITH_XOR:
1474        if (val >= 0x80 && val <= 0xff
1475            && (r0 < 4 || TCG_TARGET_REG_BITS == 64)) {
1476            tcg_out_modrm(s, OPC_ARITH_EbIb + P_REXB_RM, c, r0);
1477            tcg_out8(s, val);
1478            return;
1479        }
1480        break;
1481    }
1482
1483    if (val == (int8_t)val) {
1484        tcg_out_modrm(s, OPC_ARITH_EvIb + rexw, c, r0);
1485        tcg_out8(s, val);
1486        return;
1487    }
1488    if (rexw == 0 || val == (int32_t)val) {
1489        tcg_out_modrm(s, OPC_ARITH_EvIz + rexw, c, r0);
1490        tcg_out32(s, val);
1491        return;
1492    }
1493
1494    g_assert_not_reached();
1495}
1496
1497static void tcg_out_addi(TCGContext *s, int reg, tcg_target_long val)
1498{
1499    if (val != 0) {
1500        tgen_arithi(s, ARITH_ADD + P_REXW, reg, val, 0);
1501    }
1502}
1503
1504/* Set SMALL to force a short forward branch.  */
1505static void tcg_out_jxx(TCGContext *s, int opc, TCGLabel *l, bool small)
1506{
1507    int32_t val, val1;
1508
1509    if (l->has_value) {
1510        val = tcg_pcrel_diff(s, l->u.value_ptr);
1511        val1 = val - 2;
1512        if ((int8_t)val1 == val1) {
1513            if (opc == -1) {
1514                tcg_out8(s, OPC_JMP_short);
1515            } else {
1516                tcg_out8(s, OPC_JCC_short + opc);
1517            }
1518            tcg_out8(s, val1);
1519        } else {
1520            tcg_debug_assert(!small);
1521            if (opc == -1) {
1522                tcg_out8(s, OPC_JMP_long);
1523                tcg_out32(s, val - 5);
1524            } else {
1525                tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1526                tcg_out32(s, val - 6);
1527            }
1528        }
1529    } else if (small) {
1530        if (opc == -1) {
1531            tcg_out8(s, OPC_JMP_short);
1532        } else {
1533            tcg_out8(s, OPC_JCC_short + opc);
1534        }
1535        tcg_out_reloc(s, s->code_ptr, R_386_PC8, l, -1);
1536        s->code_ptr += 1;
1537    } else {
1538        if (opc == -1) {
1539            tcg_out8(s, OPC_JMP_long);
1540        } else {
1541            tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1542        }
1543        tcg_out_reloc(s, s->code_ptr, R_386_PC32, l, -4);
1544        s->code_ptr += 4;
1545    }
1546}
1547
1548static int tcg_out_cmp(TCGContext *s, TCGCond cond, TCGArg arg1,
1549                       TCGArg arg2, int const_arg2, int rexw)
1550{
1551    int jz, js;
1552
1553    if (!is_tst_cond(cond)) {
1554        if (!const_arg2) {
1555            tgen_arithr(s, ARITH_CMP + rexw, arg1, arg2);
1556        } else if (arg2 == 0) {
1557            tcg_out_modrm(s, OPC_TESTL + rexw, arg1, arg1);
1558        } else {
1559            tcg_debug_assert(!rexw || arg2 == (int32_t)arg2);
1560            tgen_arithi(s, ARITH_CMP + rexw, arg1, arg2, 0);
1561        }
1562        return tcg_cond_to_jcc[cond];
1563    }
1564
1565    jz = tcg_cond_to_jcc[cond];
1566    js = (cond == TCG_COND_TSTNE ? JCC_JS : JCC_JNS);
1567
1568    if (!const_arg2) {
1569        tcg_out_modrm(s, OPC_TESTL + rexw, arg1, arg2);
1570        return jz;
1571    }
1572
1573    if (arg2 <= 0xff && (TCG_TARGET_REG_BITS == 64 || arg1 < 4)) {
1574        if (arg2 == 0x80) {
1575            tcg_out_modrm(s, OPC_TESTB | P_REXB_R, arg1, arg1);
1576            return js;
1577        }
1578        if (arg2 == 0xff) {
1579            tcg_out_modrm(s, OPC_TESTB | P_REXB_R, arg1, arg1);
1580            return jz;
1581        }
1582        tcg_out_modrm(s, OPC_GRP3_Eb | P_REXB_RM, EXT3_TESTi, arg1);
1583        tcg_out8(s, arg2);
1584        return jz;
1585    }
1586
1587    if ((arg2 & ~0xff00) == 0 && arg1 < 4) {
1588        if (arg2 == 0x8000) {
1589            tcg_out_modrm(s, OPC_TESTB, arg1 + 4, arg1 + 4);
1590            return js;
1591        }
1592        if (arg2 == 0xff00) {
1593            tcg_out_modrm(s, OPC_TESTB, arg1 + 4, arg1 + 4);
1594            return jz;
1595        }
1596        tcg_out_modrm(s, OPC_GRP3_Eb, EXT3_TESTi, arg1 + 4);
1597        tcg_out8(s, arg2 >> 8);
1598        return jz;
1599    }
1600
1601    if (arg2 == 0xffff) {
1602        tcg_out_modrm(s, OPC_TESTL | P_DATA16, arg1, arg1);
1603        return jz;
1604    }
1605    if (arg2 == 0xffffffffu) {
1606        tcg_out_modrm(s, OPC_TESTL, arg1, arg1);
1607        return jz;
1608    }
1609
1610    if (is_power_of_2(rexw ? arg2 : (uint32_t)arg2)) {
1611        int jc = (cond == TCG_COND_TSTNE ? JCC_JB : JCC_JAE);
1612        int sh = ctz64(arg2);
1613
1614        rexw = (sh & 32 ? P_REXW : 0);
1615        if ((sh & 31) == 31) {
1616            tcg_out_modrm(s, OPC_TESTL | rexw, arg1, arg1);
1617            return js;
1618        } else {
1619            tcg_out_modrm(s, OPC_GRPBT | rexw, OPC_GRPBT_BT, arg1);
1620            tcg_out8(s, sh);
1621            return jc;
1622        }
1623    }
1624
1625    if (rexw) {
1626        if (arg2 == (uint32_t)arg2) {
1627            rexw = 0;
1628        } else {
1629            tcg_debug_assert(arg2 == (int32_t)arg2);
1630        }
1631    }
1632    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_TESTi, arg1);
1633    tcg_out32(s, arg2);
1634    return jz;
1635}
1636
1637static void tcg_out_brcond(TCGContext *s, int rexw, TCGCond cond,
1638                           TCGArg arg1, TCGArg arg2, int const_arg2,
1639                           TCGLabel *label, bool small)
1640{
1641    int jcc = tcg_out_cmp(s, cond, arg1, arg2, const_arg2, rexw);
1642    tcg_out_jxx(s, jcc, label, small);
1643}
1644
1645#if TCG_TARGET_REG_BITS == 32
1646static void tcg_out_brcond2(TCGContext *s, const TCGArg *args,
1647                            const int *const_args, bool small)
1648{
1649    TCGLabel *label_next = gen_new_label();
1650    TCGLabel *label_this = arg_label(args[5]);
1651    TCGCond cond = args[4];
1652
1653    switch (cond) {
1654    case TCG_COND_EQ:
1655    case TCG_COND_TSTEQ:
1656        tcg_out_brcond(s, 0, tcg_invert_cond(cond),
1657                       args[0], args[2], const_args[2], label_next, 1);
1658        tcg_out_brcond(s, 0, cond, args[1], args[3], const_args[3],
1659                       label_this, small);
1660        break;
1661
1662    case TCG_COND_NE:
1663    case TCG_COND_TSTNE:
1664        tcg_out_brcond(s, 0, cond, args[0], args[2], const_args[2],
1665                       label_this, small);
1666        tcg_out_brcond(s, 0, cond, args[1], args[3], const_args[3],
1667                       label_this, small);
1668        break;
1669
1670    default:
1671        tcg_out_brcond(s, 0, tcg_high_cond(cond), args[1],
1672                       args[3], const_args[3], label_this, small);
1673        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1674        tcg_out_brcond(s, 0, tcg_unsigned_cond(cond), args[0],
1675                       args[2], const_args[2], label_this, small);
1676        break;
1677    }
1678    tcg_out_label(s, label_next);
1679}
1680#endif
1681
1682static void tcg_out_setcond(TCGContext *s, int rexw, TCGCond cond,
1683                            TCGArg dest, TCGArg arg1, TCGArg arg2,
1684                            int const_arg2, bool neg)
1685{
1686    int cmp_rexw = rexw;
1687    bool inv = false;
1688    bool cleared;
1689    int jcc;
1690
1691    switch (cond) {
1692    case TCG_COND_NE:
1693        inv = true;
1694        /* fall through */
1695    case TCG_COND_EQ:
1696        /* If arg2 is 0, convert to LTU/GEU vs 1. */
1697        if (const_arg2 && arg2 == 0) {
1698            arg2 = 1;
1699            goto do_ltu;
1700        }
1701        break;
1702
1703    case TCG_COND_TSTNE:
1704        inv = true;
1705        /* fall through */
1706    case TCG_COND_TSTEQ:
1707        /* If arg2 is -1, convert to LTU/GEU vs 1. */
1708        if (const_arg2 && arg2 == 0xffffffffu) {
1709            arg2 = 1;
1710            cmp_rexw = 0;
1711            goto do_ltu;
1712        }
1713        break;
1714
1715    case TCG_COND_LEU:
1716        inv = true;
1717        /* fall through */
1718    case TCG_COND_GTU:
1719        /* If arg2 is a register, swap for LTU/GEU. */
1720        if (!const_arg2) {
1721            TCGReg t = arg1;
1722            arg1 = arg2;
1723            arg2 = t;
1724            goto do_ltu;
1725        }
1726        break;
1727
1728    case TCG_COND_GEU:
1729        inv = true;
1730        /* fall through */
1731    case TCG_COND_LTU:
1732    do_ltu:
1733        /*
1734         * Relying on the carry bit, use SBB to produce -1 if LTU, 0 if GEU.
1735         * We can then use NEG or INC to produce the desired result.
1736         * This is always smaller than the SETCC expansion.
1737         */
1738        tcg_out_cmp(s, TCG_COND_LTU, arg1, arg2, const_arg2, cmp_rexw);
1739
1740        /* X - X - C = -C = (C ? -1 : 0) */
1741        tgen_arithr(s, ARITH_SBB + (neg ? rexw : 0), dest, dest);
1742        if (inv && neg) {
1743            /* ~(C ? -1 : 0) = (C ? 0 : -1) */
1744            tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, dest);
1745        } else if (inv) {
1746            /* (C ? -1 : 0) + 1 = (C ? 0 : 1) */
1747            tgen_arithi(s, ARITH_ADD, dest, 1, 0);
1748        } else if (!neg) {
1749            /* -(C ? -1 : 0) = (C ? 1 : 0) */
1750            tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_NEG, dest);
1751        }
1752        return;
1753
1754    case TCG_COND_GE:
1755        inv = true;
1756        /* fall through */
1757    case TCG_COND_LT:
1758        /* If arg2 is 0, extract the sign bit. */
1759        if (const_arg2 && arg2 == 0) {
1760            tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, dest, arg1);
1761            if (inv) {
1762                tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, dest);
1763            }
1764            tcg_out_shifti(s, (neg ? SHIFT_SAR : SHIFT_SHR) + rexw,
1765                           dest, rexw ? 63 : 31);
1766            return;
1767        }
1768        break;
1769
1770    default:
1771        break;
1772    }
1773
1774    /*
1775     * If dest does not overlap the inputs, clearing it first is preferred.
1776     * The XOR breaks any false dependency for the low-byte write to dest,
1777     * and is also one byte smaller than MOVZBL.
1778     */
1779    cleared = false;
1780    if (dest != arg1 && (const_arg2 || dest != arg2)) {
1781        tgen_arithr(s, ARITH_XOR, dest, dest);
1782        cleared = true;
1783    }
1784
1785    jcc = tcg_out_cmp(s, cond, arg1, arg2, const_arg2, cmp_rexw);
1786    tcg_out_modrm(s, OPC_SETCC | jcc, 0, dest);
1787
1788    if (!cleared) {
1789        tcg_out_ext8u(s, dest, dest);
1790    }
1791    if (neg) {
1792        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, dest);
1793    }
1794}
1795
1796#if TCG_TARGET_REG_BITS == 32
1797static void tcg_out_setcond2(TCGContext *s, const TCGArg *args,
1798                             const int *const_args)
1799{
1800    TCGArg new_args[6];
1801    TCGLabel *label_true, *label_over;
1802
1803    memcpy(new_args, args+1, 5*sizeof(TCGArg));
1804
1805    if (args[0] == args[1] || args[0] == args[2]
1806        || (!const_args[3] && args[0] == args[3])
1807        || (!const_args[4] && args[0] == args[4])) {
1808        /* When the destination overlaps with one of the argument
1809           registers, don't do anything tricky.  */
1810        label_true = gen_new_label();
1811        label_over = gen_new_label();
1812
1813        new_args[5] = label_arg(label_true);
1814        tcg_out_brcond2(s, new_args, const_args+1, 1);
1815
1816        tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1817        tcg_out_jxx(s, JCC_JMP, label_over, 1);
1818        tcg_out_label(s, label_true);
1819
1820        tcg_out_movi(s, TCG_TYPE_I32, args[0], 1);
1821        tcg_out_label(s, label_over);
1822    } else {
1823        /* When the destination does not overlap one of the arguments,
1824           clear the destination first, jump if cond false, and emit an
1825           increment in the true case.  This results in smaller code.  */
1826
1827        tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1828
1829        label_over = gen_new_label();
1830        new_args[4] = tcg_invert_cond(new_args[4]);
1831        new_args[5] = label_arg(label_over);
1832        tcg_out_brcond2(s, new_args, const_args+1, 1);
1833
1834        tgen_arithi(s, ARITH_ADD, args[0], 1, 0);
1835        tcg_out_label(s, label_over);
1836    }
1837}
1838#endif
1839
1840static void tcg_out_cmov(TCGContext *s, int jcc, int rexw,
1841                         TCGReg dest, TCGReg v1)
1842{
1843    tcg_out_modrm(s, OPC_CMOVCC | jcc | rexw, dest, v1);
1844}
1845
1846static void tcg_out_movcond(TCGContext *s, int rexw, TCGCond cond,
1847                            TCGReg dest, TCGReg c1, TCGArg c2, int const_c2,
1848                            TCGReg v1)
1849{
1850    int jcc = tcg_out_cmp(s, cond, c1, c2, const_c2, rexw);
1851    tcg_out_cmov(s, jcc, rexw, dest, v1);
1852}
1853
1854static void tcg_out_ctz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1855                        TCGArg arg2, bool const_a2)
1856{
1857    if (have_bmi1) {
1858        tcg_out_modrm(s, OPC_TZCNT + rexw, dest, arg1);
1859        if (const_a2) {
1860            tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1861        } else {
1862            tcg_debug_assert(dest != arg2);
1863            tcg_out_cmov(s, JCC_JB, rexw, dest, arg2);
1864        }
1865    } else {
1866        tcg_debug_assert(dest != arg2);
1867        tcg_out_modrm(s, OPC_BSF + rexw, dest, arg1);
1868        tcg_out_cmov(s, JCC_JE, rexw, dest, arg2);
1869    }
1870}
1871
1872static void tcg_out_branch(TCGContext *s, int call, const tcg_insn_unit *dest)
1873{
1874    intptr_t disp = tcg_pcrel_diff(s, dest) - 5;
1875
1876    if (disp == (int32_t)disp) {
1877        tcg_out_opc(s, call ? OPC_CALL_Jz : OPC_JMP_long, 0, 0, 0);
1878        tcg_out32(s, disp);
1879    } else {
1880        /* rip-relative addressing into the constant pool.
1881           This is 6 + 8 = 14 bytes, as compared to using an
1882           immediate load 10 + 6 = 16 bytes, plus we may
1883           be able to re-use the pool constant for more calls.  */
1884        tcg_out_opc(s, OPC_GRP5, 0, 0, 0);
1885        tcg_out8(s, (call ? EXT5_CALLN_Ev : EXT5_JMPN_Ev) << 3 | 5);
1886        new_pool_label(s, (uintptr_t)dest, R_386_PC32, s->code_ptr, -4);
1887        tcg_out32(s, 0);
1888    }
1889}
1890
1891static void tcg_out_call(TCGContext *s, const tcg_insn_unit *dest,
1892                         const TCGHelperInfo *info)
1893{
1894    tcg_out_branch(s, 1, dest);
1895
1896#ifndef _WIN32
1897    if (TCG_TARGET_REG_BITS == 32 && info->out_kind == TCG_CALL_RET_BY_REF) {
1898        /*
1899         * The sysv i386 abi for struct return places a reference as the
1900         * first argument of the stack, and pops that argument with the
1901         * return statement.  Since we want to retain the aligned stack
1902         * pointer for the callee, we do not want to actually push that
1903         * argument before the call but rely on the normal store to the
1904         * stack slot.  But we do need to compensate for the pop in order
1905         * to reset our correct stack pointer value.
1906         * Pushing a garbage value back onto the stack is quickest.
1907         */
1908        tcg_out_push(s, TCG_REG_EAX);
1909    }
1910#endif
1911}
1912
1913static void tcg_out_jmp(TCGContext *s, const tcg_insn_unit *dest)
1914{
1915    tcg_out_branch(s, 0, dest);
1916}
1917
1918static void tcg_out_nopn(TCGContext *s, int n)
1919{
1920    int i;
1921    /* Emit 1 or 2 operand size prefixes for the standard one byte nop,
1922     * "xchg %eax,%eax", forming "xchg %ax,%ax". All cores accept the
1923     * duplicate prefix, and all of the interesting recent cores can
1924     * decode and discard the duplicates in a single cycle.
1925     */
1926    tcg_debug_assert(n >= 1);
1927    for (i = 1; i < n; ++i) {
1928        tcg_out8(s, 0x66);
1929    }
1930    tcg_out8(s, 0x90);
1931}
1932
1933typedef struct {
1934    TCGReg base;
1935    int index;
1936    int ofs;
1937    int seg;
1938    TCGAtomAlign aa;
1939} HostAddress;
1940
1941bool tcg_target_has_memory_bswap(MemOp memop)
1942{
1943    TCGAtomAlign aa;
1944
1945    if (!have_movbe) {
1946        return false;
1947    }
1948    if ((memop & MO_SIZE) < MO_128) {
1949        return true;
1950    }
1951
1952    /*
1953     * Reject 16-byte memop with 16-byte atomicity, i.e. VMOVDQA,
1954     * but do allow a pair of 64-bit operations, i.e. MOVBEQ.
1955     */
1956    aa = atom_and_align_for_opc(tcg_ctx, memop, MO_ATOM_IFALIGN, true);
1957    return aa.atom < MO_128;
1958}
1959
1960/*
1961 * Because i686 has no register parameters and because x86_64 has xchg
1962 * to handle addr/data register overlap, we have placed all input arguments
1963 * before we need might need a scratch reg.
1964 *
1965 * Even then, a scratch is only needed for l->raddr.  Rather than expose
1966 * a general-purpose scratch when we don't actually know it's available,
1967 * use the ra_gen hook to load into RAX if needed.
1968 */
1969#if TCG_TARGET_REG_BITS == 64
1970static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
1971{
1972    if (arg < 0) {
1973        arg = TCG_REG_RAX;
1974    }
1975    tcg_out_movi(s, TCG_TYPE_PTR, arg, (uintptr_t)l->raddr);
1976    return arg;
1977}
1978static const TCGLdstHelperParam ldst_helper_param = {
1979    .ra_gen = ldst_ra_gen
1980};
1981#else
1982static const TCGLdstHelperParam ldst_helper_param = { };
1983#endif
1984
1985static void tcg_out_vec_to_pair(TCGContext *s, TCGType type,
1986                                TCGReg l, TCGReg h, TCGReg v)
1987{
1988    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1989
1990    /* vpmov{d,q} %v, %l */
1991    tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, v, 0, l);
1992    /* vpextr{d,q} $1, %v, %h */
1993    tcg_out_vex_modrm(s, OPC_PEXTRD + rexw, v, 0, h);
1994    tcg_out8(s, 1);
1995}
1996
1997static void tcg_out_pair_to_vec(TCGContext *s, TCGType type,
1998                                TCGReg v, TCGReg l, TCGReg h)
1999{
2000    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2001
2002    /* vmov{d,q} %l, %v */
2003    tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, v, 0, l);
2004    /* vpinsr{d,q} $1, %h, %v, %v */
2005    tcg_out_vex_modrm(s, OPC_PINSRD + rexw, v, v, h);
2006    tcg_out8(s, 1);
2007}
2008
2009/*
2010 * Generate code for the slow path for a load at the end of block
2011 */
2012static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
2013{
2014    MemOp opc = get_memop(l->oi);
2015    tcg_insn_unit **label_ptr = &l->label_ptr[0];
2016
2017    /* resolve label address */
2018    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
2019    if (label_ptr[1]) {
2020        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
2021    }
2022
2023    tcg_out_ld_helper_args(s, l, &ldst_helper_param);
2024    tcg_out_branch(s, 1, qemu_ld_helpers[opc & MO_SIZE]);
2025    tcg_out_ld_helper_ret(s, l, false, &ldst_helper_param);
2026
2027    tcg_out_jmp(s, l->raddr);
2028    return true;
2029}
2030
2031/*
2032 * Generate code for the slow path for a store at the end of block
2033 */
2034static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
2035{
2036    MemOp opc = get_memop(l->oi);
2037    tcg_insn_unit **label_ptr = &l->label_ptr[0];
2038
2039    /* resolve label address */
2040    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
2041    if (label_ptr[1]) {
2042        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
2043    }
2044
2045    tcg_out_st_helper_args(s, l, &ldst_helper_param);
2046    tcg_out_branch(s, 1, qemu_st_helpers[opc & MO_SIZE]);
2047
2048    tcg_out_jmp(s, l->raddr);
2049    return true;
2050}
2051
2052#ifdef CONFIG_USER_ONLY
2053static HostAddress x86_guest_base = {
2054    .index = -1
2055};
2056
2057#if defined(__x86_64__) && defined(__linux__)
2058# include <asm/prctl.h>
2059# include <sys/prctl.h>
2060int arch_prctl(int code, unsigned long addr);
2061static inline int setup_guest_base_seg(void)
2062{
2063    if (arch_prctl(ARCH_SET_GS, guest_base) == 0) {
2064        return P_GS;
2065    }
2066    return 0;
2067}
2068#define setup_guest_base_seg  setup_guest_base_seg
2069#elif defined(__x86_64__) && \
2070      (defined (__FreeBSD__) || defined (__FreeBSD_kernel__))
2071# include <machine/sysarch.h>
2072static inline int setup_guest_base_seg(void)
2073{
2074    if (sysarch(AMD64_SET_GSBASE, &guest_base) == 0) {
2075        return P_GS;
2076    }
2077    return 0;
2078}
2079#define setup_guest_base_seg  setup_guest_base_seg
2080#endif
2081#else
2082# define x86_guest_base (*(HostAddress *)({ qemu_build_not_reached(); NULL; }))
2083#endif /* CONFIG_USER_ONLY */
2084#ifndef setup_guest_base_seg
2085# define setup_guest_base_seg()  0
2086#endif
2087
2088#define MIN_TLB_MASK_TABLE_OFS  INT_MIN
2089
2090/*
2091 * For softmmu, perform the TLB load and compare.
2092 * For useronly, perform any required alignment tests.
2093 * In both cases, return a TCGLabelQemuLdst structure if the slow path
2094 * is required and fill in @h with the host address for the fast path.
2095 */
2096static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
2097                                           TCGReg addr, MemOpIdx oi, bool is_ld)
2098{
2099    TCGLabelQemuLdst *ldst = NULL;
2100    MemOp opc = get_memop(oi);
2101    MemOp s_bits = opc & MO_SIZE;
2102    unsigned a_mask;
2103
2104    if (tcg_use_softmmu) {
2105        h->index = TCG_REG_L0;
2106        h->ofs = 0;
2107        h->seg = 0;
2108    } else {
2109        *h = x86_guest_base;
2110    }
2111    h->base = addr;
2112    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, s_bits == MO_128);
2113    a_mask = (1 << h->aa.align) - 1;
2114
2115    if (tcg_use_softmmu) {
2116        int cmp_ofs = is_ld ? offsetof(CPUTLBEntry, addr_read)
2117                            : offsetof(CPUTLBEntry, addr_write);
2118        TCGType ttype = TCG_TYPE_I32;
2119        TCGType tlbtype = TCG_TYPE_I32;
2120        int trexw = 0, hrexw = 0, tlbrexw = 0;
2121        unsigned mem_index = get_mmuidx(oi);
2122        unsigned s_mask = (1 << s_bits) - 1;
2123        int fast_ofs = tlb_mask_table_ofs(s, mem_index);
2124        int tlb_mask;
2125
2126        ldst = new_ldst_label(s);
2127        ldst->is_ld = is_ld;
2128        ldst->oi = oi;
2129        ldst->addr_reg = addr;
2130
2131        if (TCG_TARGET_REG_BITS == 64) {
2132            ttype = s->addr_type;
2133            trexw = (ttype == TCG_TYPE_I32 ? 0 : P_REXW);
2134            if (TCG_TYPE_PTR == TCG_TYPE_I64) {
2135                hrexw = P_REXW;
2136                if (s->page_bits + s->tlb_dyn_max_bits > 32) {
2137                    tlbtype = TCG_TYPE_I64;
2138                    tlbrexw = P_REXW;
2139                }
2140            }
2141        }
2142
2143        tcg_out_mov(s, tlbtype, TCG_REG_L0, addr);
2144        tcg_out_shifti(s, SHIFT_SHR + tlbrexw, TCG_REG_L0,
2145                       s->page_bits - CPU_TLB_ENTRY_BITS);
2146
2147        tcg_out_modrm_offset(s, OPC_AND_GvEv + trexw, TCG_REG_L0, TCG_AREG0,
2148                             fast_ofs + offsetof(CPUTLBDescFast, mask));
2149
2150        tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, TCG_REG_L0, TCG_AREG0,
2151                             fast_ofs + offsetof(CPUTLBDescFast, table));
2152
2153        /*
2154         * If the required alignment is at least as large as the access,
2155         * simply copy the address and mask.  For lesser alignments,
2156         * check that we don't cross pages for the complete access.
2157         */
2158        if (a_mask >= s_mask) {
2159            tcg_out_mov(s, ttype, TCG_REG_L1, addr);
2160        } else {
2161            tcg_out_modrm_offset(s, OPC_LEA + trexw, TCG_REG_L1,
2162                                 addr, s_mask - a_mask);
2163        }
2164        tlb_mask = s->page_mask | a_mask;
2165        tgen_arithi(s, ARITH_AND + trexw, TCG_REG_L1, tlb_mask, 0);
2166
2167        /* cmp 0(TCG_REG_L0), TCG_REG_L1 */
2168        tcg_out_modrm_offset(s, OPC_CMP_GvEv + trexw,
2169                             TCG_REG_L1, TCG_REG_L0, cmp_ofs);
2170
2171        /* jne slow_path */
2172        tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
2173        ldst->label_ptr[0] = s->code_ptr;
2174        s->code_ptr += 4;
2175
2176        /* TLB Hit.  */
2177        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_L0, TCG_REG_L0,
2178                   offsetof(CPUTLBEntry, addend));
2179    } else if (a_mask) {
2180        int jcc;
2181
2182        ldst = new_ldst_label(s);
2183        ldst->is_ld = is_ld;
2184        ldst->oi = oi;
2185        ldst->addr_reg = addr;
2186
2187        /* jne slow_path */
2188        jcc = tcg_out_cmp(s, TCG_COND_TSTNE, addr, a_mask, true, false);
2189        tcg_out_opc(s, OPC_JCC_long + jcc, 0, 0, 0);
2190        ldst->label_ptr[0] = s->code_ptr;
2191        s->code_ptr += 4;
2192    }
2193
2194    return ldst;
2195}
2196
2197static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2198                                   HostAddress h, TCGType type, MemOp memop)
2199{
2200    bool use_movbe = false;
2201    int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
2202    int movop = OPC_MOVL_GvEv;
2203
2204    /* Do big-endian loads with movbe.  */
2205    if (memop & MO_BSWAP) {
2206        tcg_debug_assert(have_movbe);
2207        use_movbe = true;
2208        movop = OPC_MOVBE_GyMy;
2209    }
2210
2211    switch (memop & MO_SSIZE) {
2212    case MO_UB:
2213        tcg_out_modrm_sib_offset(s, OPC_MOVZBL + h.seg, datalo,
2214                                 h.base, h.index, 0, h.ofs);
2215        break;
2216    case MO_SB:
2217        tcg_out_modrm_sib_offset(s, OPC_MOVSBL + rexw + h.seg, datalo,
2218                                 h.base, h.index, 0, h.ofs);
2219        break;
2220    case MO_UW:
2221        if (use_movbe) {
2222            /* There is no extending movbe; only low 16-bits are modified.  */
2223            if (datalo != h.base && datalo != h.index) {
2224                /* XOR breaks dependency chains.  */
2225                tgen_arithr(s, ARITH_XOR, datalo, datalo);
2226                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2227                                         datalo, h.base, h.index, 0, h.ofs);
2228            } else {
2229                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2230                                         datalo, h.base, h.index, 0, h.ofs);
2231                tcg_out_ext16u(s, datalo, datalo);
2232            }
2233        } else {
2234            tcg_out_modrm_sib_offset(s, OPC_MOVZWL + h.seg, datalo,
2235                                     h.base, h.index, 0, h.ofs);
2236        }
2237        break;
2238    case MO_SW:
2239        if (use_movbe) {
2240            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2241                                     datalo, h.base, h.index, 0, h.ofs);
2242            tcg_out_ext16s(s, type, datalo, datalo);
2243        } else {
2244            tcg_out_modrm_sib_offset(s, OPC_MOVSWL + rexw + h.seg,
2245                                     datalo, h.base, h.index, 0, h.ofs);
2246        }
2247        break;
2248    case MO_UL:
2249        tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2250                                 h.base, h.index, 0, h.ofs);
2251        break;
2252#if TCG_TARGET_REG_BITS == 64
2253    case MO_SL:
2254        if (use_movbe) {
2255            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + h.seg, datalo,
2256                                     h.base, h.index, 0, h.ofs);
2257            tcg_out_ext32s(s, datalo, datalo);
2258        } else {
2259            tcg_out_modrm_sib_offset(s, OPC_MOVSLQ + h.seg, datalo,
2260                                     h.base, h.index, 0, h.ofs);
2261        }
2262        break;
2263#endif
2264    case MO_UQ:
2265        if (TCG_TARGET_REG_BITS == 64) {
2266            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2267                                     h.base, h.index, 0, h.ofs);
2268            break;
2269        }
2270        if (use_movbe) {
2271            TCGReg t = datalo;
2272            datalo = datahi;
2273            datahi = t;
2274        }
2275        if (h.base == datalo || h.index == datalo) {
2276            tcg_out_modrm_sib_offset(s, OPC_LEA, datahi,
2277                                     h.base, h.index, 0, h.ofs);
2278            tcg_out_modrm_offset(s, movop + h.seg, datalo, datahi, 0);
2279            tcg_out_modrm_offset(s, movop + h.seg, datahi, datahi, 4);
2280        } else {
2281            tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2282                                     h.base, h.index, 0, h.ofs);
2283            tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2284                                     h.base, h.index, 0, h.ofs + 4);
2285        }
2286        break;
2287
2288    case MO_128:
2289        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2290
2291        /*
2292         * Without 16-byte atomicity, use integer regs.
2293         * That is where we want the data, and it allows bswaps.
2294         */
2295        if (h.aa.atom < MO_128) {
2296            if (use_movbe) {
2297                TCGReg t = datalo;
2298                datalo = datahi;
2299                datahi = t;
2300            }
2301            if (h.base == datalo || h.index == datalo) {
2302                tcg_out_modrm_sib_offset(s, OPC_LEA + P_REXW, datahi,
2303                                         h.base, h.index, 0, h.ofs);
2304                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2305                                     datalo, datahi, 0);
2306                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2307                                     datahi, datahi, 8);
2308            } else {
2309                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2310                                         h.base, h.index, 0, h.ofs);
2311                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2312                                         h.base, h.index, 0, h.ofs + 8);
2313            }
2314            break;
2315        }
2316
2317        /*
2318         * With 16-byte atomicity, a vector load is required.
2319         * If we already have 16-byte alignment, then VMOVDQA always works.
2320         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2321         * Else use we require a runtime test for alignment for VMOVDQA;
2322         * use VMOVDQU on the unaligned nonatomic path for simplicity.
2323         */
2324        if (h.aa.align >= MO_128) {
2325            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2326                                         TCG_TMP_VEC, 0,
2327                                         h.base, h.index, 0, h.ofs);
2328        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2329            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2330                                         TCG_TMP_VEC, 0,
2331                                         h.base, h.index, 0, h.ofs);
2332        } else {
2333            TCGLabel *l1 = gen_new_label();
2334            TCGLabel *l2 = gen_new_label();
2335            int jcc;
2336
2337            jcc = tcg_out_cmp(s, TCG_COND_TSTNE, h.base, 15, true, false);
2338            tcg_out_jxx(s, jcc, l1, true);
2339
2340            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2341                                         TCG_TMP_VEC, 0,
2342                                         h.base, h.index, 0, h.ofs);
2343            tcg_out_jxx(s, JCC_JMP, l2, true);
2344
2345            tcg_out_label(s, l1);
2346            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2347                                         TCG_TMP_VEC, 0,
2348                                         h.base, h.index, 0, h.ofs);
2349            tcg_out_label(s, l2);
2350        }
2351        tcg_out_vec_to_pair(s, TCG_TYPE_I64, datalo, datahi, TCG_TMP_VEC);
2352        break;
2353
2354    default:
2355        g_assert_not_reached();
2356    }
2357}
2358
2359static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
2360                            TCGReg addr, MemOpIdx oi, TCGType data_type)
2361{
2362    TCGLabelQemuLdst *ldst;
2363    HostAddress h;
2364
2365    ldst = prepare_host_addr(s, &h, addr, oi, true);
2366    tcg_out_qemu_ld_direct(s, datalo, datahi, h, data_type, get_memop(oi));
2367
2368    if (ldst) {
2369        ldst->type = data_type;
2370        ldst->datalo_reg = datalo;
2371        ldst->datahi_reg = datahi;
2372        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2373    }
2374}
2375
2376static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2377                                   HostAddress h, MemOp memop)
2378{
2379    bool use_movbe = false;
2380    int movop = OPC_MOVL_EvGv;
2381
2382    /*
2383     * Do big-endian stores with movbe or system-mode.
2384     * User-only without movbe will have its swapping done generically.
2385     */
2386    if (memop & MO_BSWAP) {
2387        tcg_debug_assert(have_movbe);
2388        use_movbe = true;
2389        movop = OPC_MOVBE_MyGy;
2390    }
2391
2392    switch (memop & MO_SIZE) {
2393    case MO_8:
2394        /* This is handled with constraints on INDEX_op_qemu_st8_i32. */
2395        tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || datalo < 4);
2396        tcg_out_modrm_sib_offset(s, OPC_MOVB_EvGv + P_REXB_R + h.seg,
2397                                 datalo, h.base, h.index, 0, h.ofs);
2398        break;
2399    case MO_16:
2400        tcg_out_modrm_sib_offset(s, movop + P_DATA16 + h.seg, datalo,
2401                                 h.base, h.index, 0, h.ofs);
2402        break;
2403    case MO_32:
2404        tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2405                                 h.base, h.index, 0, h.ofs);
2406        break;
2407    case MO_64:
2408        if (TCG_TARGET_REG_BITS == 64) {
2409            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2410                                     h.base, h.index, 0, h.ofs);
2411        } else {
2412            if (use_movbe) {
2413                TCGReg t = datalo;
2414                datalo = datahi;
2415                datahi = t;
2416            }
2417            tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2418                                     h.base, h.index, 0, h.ofs);
2419            tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2420                                     h.base, h.index, 0, h.ofs + 4);
2421        }
2422        break;
2423
2424    case MO_128:
2425        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2426
2427        /*
2428         * Without 16-byte atomicity, use integer regs.
2429         * That is where we have the data, and it allows bswaps.
2430         */
2431        if (h.aa.atom < MO_128) {
2432            if (use_movbe) {
2433                TCGReg t = datalo;
2434                datalo = datahi;
2435                datahi = t;
2436            }
2437            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2438                                     h.base, h.index, 0, h.ofs);
2439            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2440                                     h.base, h.index, 0, h.ofs + 8);
2441            break;
2442        }
2443
2444        /*
2445         * With 16-byte atomicity, a vector store is required.
2446         * If we already have 16-byte alignment, then VMOVDQA always works.
2447         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2448         * Else use we require a runtime test for alignment for VMOVDQA;
2449         * use VMOVDQU on the unaligned nonatomic path for simplicity.
2450         */
2451        tcg_out_pair_to_vec(s, TCG_TYPE_I64, TCG_TMP_VEC, datalo, datahi);
2452        if (h.aa.align >= MO_128) {
2453            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2454                                         TCG_TMP_VEC, 0,
2455                                         h.base, h.index, 0, h.ofs);
2456        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2457            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2458                                         TCG_TMP_VEC, 0,
2459                                         h.base, h.index, 0, h.ofs);
2460        } else {
2461            TCGLabel *l1 = gen_new_label();
2462            TCGLabel *l2 = gen_new_label();
2463            int jcc;
2464
2465            jcc = tcg_out_cmp(s, TCG_COND_TSTNE, h.base, 15, true, false);
2466            tcg_out_jxx(s, jcc, l1, true);
2467
2468            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2469                                         TCG_TMP_VEC, 0,
2470                                         h.base, h.index, 0, h.ofs);
2471            tcg_out_jxx(s, JCC_JMP, l2, true);
2472
2473            tcg_out_label(s, l1);
2474            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2475                                         TCG_TMP_VEC, 0,
2476                                         h.base, h.index, 0, h.ofs);
2477            tcg_out_label(s, l2);
2478        }
2479        break;
2480
2481    default:
2482        g_assert_not_reached();
2483    }
2484}
2485
2486static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
2487                            TCGReg addr, MemOpIdx oi, TCGType data_type)
2488{
2489    TCGLabelQemuLdst *ldst;
2490    HostAddress h;
2491
2492    ldst = prepare_host_addr(s, &h, addr, oi, false);
2493    tcg_out_qemu_st_direct(s, datalo, datahi, h, get_memop(oi));
2494
2495    if (ldst) {
2496        ldst->type = data_type;
2497        ldst->datalo_reg = datalo;
2498        ldst->datahi_reg = datahi;
2499        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2500    }
2501}
2502
2503static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
2504{
2505    /* Reuse the zeroing that exists for goto_ptr.  */
2506    if (a0 == 0) {
2507        tcg_out_jmp(s, tcg_code_gen_epilogue);
2508    } else {
2509        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_EAX, a0);
2510        tcg_out_jmp(s, tb_ret_addr);
2511    }
2512}
2513
2514static void tcg_out_goto_tb(TCGContext *s, int which)
2515{
2516    /*
2517     * Jump displacement must be aligned for atomic patching;
2518     * see if we need to add extra nops before jump
2519     */
2520    int gap = QEMU_ALIGN_PTR_UP(s->code_ptr + 1, 4) - s->code_ptr;
2521    if (gap != 1) {
2522        tcg_out_nopn(s, gap - 1);
2523    }
2524    tcg_out8(s, OPC_JMP_long); /* jmp im */
2525    set_jmp_insn_offset(s, which);
2526    tcg_out32(s, 0);
2527    set_jmp_reset_offset(s, which);
2528}
2529
2530void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
2531                              uintptr_t jmp_rx, uintptr_t jmp_rw)
2532{
2533    /* patch the branch destination */
2534    uintptr_t addr = tb->jmp_target_addr[n];
2535    qatomic_set((int32_t *)jmp_rw, addr - (jmp_rx + 4));
2536    /* no need to flush icache explicitly */
2537}
2538
2539
2540static void tgen_add(TCGContext *s, TCGType type,
2541                     TCGReg a0, TCGReg a1, TCGReg a2)
2542{
2543    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2544
2545    if (a0 == a1) {
2546        tgen_arithr(s, ARITH_ADD + rexw, a0, a2);
2547    } else if (a0 == a2) {
2548        tgen_arithr(s, ARITH_ADD + rexw, a0, a1);
2549    } else {
2550        tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a2, 0, 0);
2551    }
2552}
2553
2554static void tgen_addi(TCGContext *s, TCGType type,
2555                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2556{
2557    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2558
2559    if (a0 == a1) {
2560        tgen_arithi(s, ARITH_ADD + rexw, a0, a2, false);
2561    } else {
2562        tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, -1, 0, a2);
2563    }
2564}
2565
2566static const TCGOutOpBinary outop_add = {
2567    .base.static_constraint = C_O1_I2(r, r, re),
2568    .out_rrr = tgen_add,
2569    .out_rri = tgen_addi,
2570};
2571
2572static void tgen_and(TCGContext *s, TCGType type,
2573                     TCGReg a0, TCGReg a1, TCGReg a2)
2574{
2575    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2576    tgen_arithr(s, ARITH_AND + rexw, a0, a2);
2577}
2578
2579static void tgen_andi(TCGContext *s, TCGType type,
2580                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2581{
2582    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2583    tgen_arithi(s, ARITH_AND + rexw, a0, a2, false);
2584}
2585
2586static const TCGOutOpBinary outop_and = {
2587    .base.static_constraint = C_O1_I2(r, 0, reZ),
2588    .out_rrr = tgen_and,
2589    .out_rri = tgen_andi,
2590};
2591
2592static void tgen_andc(TCGContext *s, TCGType type,
2593                      TCGReg a0, TCGReg a1, TCGReg a2)
2594{
2595    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2596    tcg_out_vex_modrm(s, OPC_ANDN + rexw, a0, a2, a1);
2597}
2598
2599static TCGConstraintSetIndex cset_andc(TCGType type, unsigned flags)
2600{
2601    return have_bmi1 ? C_O1_I2(r, r, r) : C_NotImplemented;
2602}
2603
2604static const TCGOutOpBinary outop_andc = {
2605    .base.static_constraint = C_Dynamic,
2606    .base.dynamic_constraint = cset_andc,
2607    .out_rrr = tgen_andc,
2608};
2609
2610static void tgen_clz(TCGContext *s, TCGType type,
2611                     TCGReg a0, TCGReg a1, TCGReg a2)
2612{
2613    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2614    int jcc;
2615
2616    if (have_lzcnt) {
2617        tcg_out_modrm(s, OPC_LZCNT + rexw, a0, a1);
2618        jcc = JCC_JB;
2619    } else {
2620        /* Recall that the output of BSR is the index not the count.  */
2621        tcg_out_modrm(s, OPC_BSR + rexw, a0, a1);
2622        tgen_arithi(s, ARITH_XOR + rexw, a0, rexw ? 63 : 31, 0);
2623
2624        /* Since we have destroyed the flags from BSR, we have to re-test.  */
2625        jcc = tcg_out_cmp(s, TCG_COND_EQ, a1, 0, 1, rexw);
2626    }
2627    tcg_out_cmov(s, jcc, rexw, a0, a2);
2628}
2629
2630static void tgen_clzi(TCGContext *s, TCGType type,
2631                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2632{
2633    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2634    tcg_out_modrm(s, OPC_LZCNT + rexw, a0, a1);
2635}
2636
2637static TCGConstraintSetIndex cset_clz(TCGType type, unsigned flags)
2638{
2639    return have_lzcnt ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
2640}
2641
2642static const TCGOutOpBinary outop_clz = {
2643    .base.static_constraint = C_Dynamic,
2644    .base.dynamic_constraint = cset_clz,
2645    .out_rrr = tgen_clz,
2646    .out_rri = tgen_clzi,
2647};
2648
2649static const TCGOutOpBinary outop_divs = {
2650    .base.static_constraint = C_NotImplemented,
2651};
2652
2653static void tgen_divs2(TCGContext *s, TCGType type,
2654                       TCGReg a0, TCGReg a1, TCGReg a4)
2655{
2656    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2657    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IDIV, a4);
2658}
2659
2660static const TCGOutOpDivRem outop_divs2 = {
2661    .base.static_constraint = C_O2_I3(a, d, 0, 1, r),
2662    .out_rr01r = tgen_divs2,
2663};
2664
2665static const TCGOutOpBinary outop_divu = {
2666    .base.static_constraint = C_NotImplemented,
2667};
2668
2669static void tgen_divu2(TCGContext *s, TCGType type,
2670                       TCGReg a0, TCGReg a1, TCGReg a4)
2671{
2672    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2673    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_DIV, a4);
2674}
2675
2676static const TCGOutOpDivRem outop_divu2 = {
2677    .base.static_constraint = C_O2_I3(a, d, 0, 1, r),
2678    .out_rr01r = tgen_divu2,
2679};
2680
2681static const TCGOutOpBinary outop_eqv = {
2682    .base.static_constraint = C_NotImplemented,
2683};
2684
2685static void tgen_mul(TCGContext *s, TCGType type,
2686                     TCGReg a0, TCGReg a1, TCGReg a2)
2687{
2688    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2689    tcg_out_modrm(s, OPC_IMUL_GvEv + rexw, a0, a2);
2690}
2691
2692static void tgen_muli(TCGContext *s, TCGType type,
2693                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2694{
2695    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2696
2697    if (a2 == (int8_t)a2) {
2698        tcg_out_modrm(s, OPC_IMUL_GvEvIb + rexw, a0, a0);
2699        tcg_out8(s, a2);
2700    } else {
2701        tcg_out_modrm(s, OPC_IMUL_GvEvIz + rexw, a0, a0);
2702        tcg_out32(s, a2);
2703    }
2704}
2705
2706static const TCGOutOpBinary outop_mul = {
2707    .base.static_constraint = C_O1_I2(r, 0, re),
2708    .out_rrr = tgen_mul,
2709    .out_rri = tgen_muli,
2710};
2711
2712static const TCGOutOpBinary outop_mulsh = {
2713    .base.static_constraint = C_NotImplemented,
2714};
2715
2716static const TCGOutOpBinary outop_muluh = {
2717    .base.static_constraint = C_NotImplemented,
2718};
2719
2720static const TCGOutOpBinary outop_nand = {
2721    .base.static_constraint = C_NotImplemented,
2722};
2723
2724static const TCGOutOpBinary outop_nor = {
2725    .base.static_constraint = C_NotImplemented,
2726};
2727
2728static void tgen_or(TCGContext *s, TCGType type,
2729                    TCGReg a0, TCGReg a1, TCGReg a2)
2730{
2731    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2732    tgen_arithr(s, ARITH_OR + rexw, a0, a2);
2733}
2734
2735static void tgen_ori(TCGContext *s, TCGType type,
2736                     TCGReg a0, TCGReg a1, tcg_target_long a2)
2737{
2738    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2739    tgen_arithi(s, ARITH_OR + rexw, a0, a2, false);
2740}
2741
2742static const TCGOutOpBinary outop_or = {
2743    .base.static_constraint = C_O1_I2(r, 0, re),
2744    .out_rrr = tgen_or,
2745    .out_rri = tgen_ori,
2746};
2747
2748static const TCGOutOpBinary outop_orc = {
2749    .base.static_constraint = C_NotImplemented,
2750};
2751
2752static const TCGOutOpBinary outop_rems = {
2753    .base.static_constraint = C_NotImplemented,
2754};
2755
2756static const TCGOutOpBinary outop_remu = {
2757    .base.static_constraint = C_NotImplemented,
2758};
2759
2760static void tgen_rotl(TCGContext *s, TCGType type,
2761                     TCGReg a0, TCGReg a1, TCGReg a2)
2762{
2763    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2764    tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_ROL, a0);
2765}
2766
2767static void tgen_rotli(TCGContext *s, TCGType type,
2768                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2769{
2770    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2771    tcg_out_shifti(s, SHIFT_ROL + rexw, a0, a2);
2772}
2773
2774static const TCGOutOpBinary outop_rotl = {
2775    .base.static_constraint = C_O1_I2(r, 0, ci),
2776    .out_rrr = tgen_rotl,
2777    .out_rri = tgen_rotli,
2778};
2779
2780static void tgen_rotr(TCGContext *s, TCGType type,
2781                     TCGReg a0, TCGReg a1, TCGReg a2)
2782{
2783    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2784    tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_ROR, a0);
2785}
2786
2787static void tgen_rotri(TCGContext *s, TCGType type,
2788                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2789{
2790    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2791    tcg_out_shifti(s, SHIFT_ROR + rexw, a0, a2);
2792}
2793
2794static const TCGOutOpBinary outop_rotr = {
2795    .base.static_constraint = C_O1_I2(r, 0, ci),
2796    .out_rrr = tgen_rotr,
2797    .out_rri = tgen_rotri,
2798};
2799
2800static TCGConstraintSetIndex cset_shift(TCGType type, unsigned flags)
2801{
2802    return have_bmi2 ? C_O1_I2(r, r, ri) : C_O1_I2(r, 0, ci);
2803}
2804
2805static void tgen_sar(TCGContext *s, TCGType type,
2806                     TCGReg a0, TCGReg a1, TCGReg a2)
2807{
2808    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2809    if (have_bmi2) {
2810        tcg_out_vex_modrm(s, OPC_SARX + rexw, a0, a2, a1);
2811    } else {
2812        tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_SAR, a0);
2813    }
2814}
2815
2816static void tgen_sari(TCGContext *s, TCGType type,
2817                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2818{
2819    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2820
2821    tcg_out_mov(s, type, a0, a1);
2822    tcg_out_shifti(s, SHIFT_SAR + rexw, a0, a2);
2823}
2824
2825static const TCGOutOpBinary outop_sar = {
2826    .base.static_constraint = C_Dynamic,
2827    .base.dynamic_constraint = cset_shift,
2828    .out_rrr = tgen_sar,
2829    .out_rri = tgen_sari,
2830};
2831
2832static void tgen_shl(TCGContext *s, TCGType type,
2833                     TCGReg a0, TCGReg a1, TCGReg a2)
2834{
2835    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2836    if (have_bmi2) {
2837        tcg_out_vex_modrm(s, OPC_SHLX + rexw, a0, a2, a1);
2838    } else {
2839        tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_SHL, a0);
2840    }
2841}
2842
2843static void tgen_shli(TCGContext *s, TCGType type,
2844                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2845{
2846    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2847
2848    /* For small constant 3-operand shift, use LEA.  */
2849    if (a0 != a1 && a2 >= 1 && a2 <= 3) {
2850        if (a2 == 1) {
2851            /* shl $1,a1,a0 -> lea (a1,a1),a0 */
2852            tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a1, 0, 0);
2853        } else {
2854            /* shl $n,a1,a0 -> lea 0(,a1,n),a0 */
2855            tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, -1, a1, a2, 0);
2856        }
2857        return;
2858    }
2859    tcg_out_mov(s, type, a0, a1);
2860    tcg_out_shifti(s, SHIFT_SHL + rexw, a0, a2);
2861}
2862
2863static const TCGOutOpBinary outop_shl = {
2864    .base.static_constraint = C_Dynamic,
2865    .base.dynamic_constraint = cset_shift,
2866    .out_rrr = tgen_shl,
2867    .out_rri = tgen_shli,
2868};
2869
2870static void tgen_shr(TCGContext *s, TCGType type,
2871                     TCGReg a0, TCGReg a1, TCGReg a2)
2872{
2873    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2874    if (have_bmi2) {
2875        tcg_out_vex_modrm(s, OPC_SHRX + rexw, a0, a2, a1);
2876    } else {
2877        tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_SHR, a0);
2878    }
2879}
2880
2881static void tgen_shri(TCGContext *s, TCGType type,
2882                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2883{
2884    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2885
2886    tcg_out_mov(s, type, a0, a1);
2887    tcg_out_shifti(s, SHIFT_SHR + rexw, a0, a2);
2888}
2889
2890static const TCGOutOpBinary outop_shr = {
2891    .base.static_constraint = C_Dynamic,
2892    .base.dynamic_constraint = cset_shift,
2893    .out_rrr = tgen_shr,
2894    .out_rri = tgen_shri,
2895};
2896
2897static void tgen_sub(TCGContext *s, TCGType type,
2898                      TCGReg a0, TCGReg a1, TCGReg a2)
2899{
2900    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2901    tgen_arithr(s, ARITH_SUB + rexw, a0, a2);
2902}
2903
2904static const TCGOutOpSubtract outop_sub = {
2905    .base.static_constraint = C_O1_I2(r, 0, r),
2906    .out_rrr = tgen_sub,
2907};
2908
2909static void tgen_xor(TCGContext *s, TCGType type,
2910                     TCGReg a0, TCGReg a1, TCGReg a2)
2911{
2912    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2913    tgen_arithr(s, ARITH_XOR + rexw, a0, a2);
2914}
2915
2916static void tgen_xori(TCGContext *s, TCGType type,
2917                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2918{
2919    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2920    tgen_arithi(s, ARITH_XOR + rexw, a0, a2, false);
2921}
2922
2923static const TCGOutOpBinary outop_xor = {
2924    .base.static_constraint = C_O1_I2(r, 0, re),
2925    .out_rrr = tgen_xor,
2926    .out_rri = tgen_xori,
2927};
2928
2929static void tgen_neg(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
2930{
2931    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2932    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, a0);
2933}
2934
2935static const TCGOutOpUnary outop_neg = {
2936    .base.static_constraint = C_O1_I1(r, 0),
2937    .out_rr = tgen_neg,
2938};
2939
2940static void tgen_not(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
2941{
2942    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2943    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, a0);
2944}
2945
2946static const TCGOutOpUnary outop_not = {
2947    .base.static_constraint = C_O1_I1(r, 0),
2948    .out_rr = tgen_not,
2949};
2950
2951
2952static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
2953                       const TCGArg args[TCG_MAX_OP_ARGS],
2954                       const int const_args[TCG_MAX_OP_ARGS])
2955{
2956    TCGArg a0, a1, a2;
2957    int const_a2, rexw;
2958
2959#if TCG_TARGET_REG_BITS == 64
2960# define OP_32_64(x) \
2961        case glue(glue(INDEX_op_, x), _i64): \
2962        case glue(glue(INDEX_op_, x), _i32)
2963#else
2964# define OP_32_64(x) \
2965        case glue(glue(INDEX_op_, x), _i32)
2966#endif
2967
2968    /* Hoist the loads of the most common arguments.  */
2969    a0 = args[0];
2970    a1 = args[1];
2971    a2 = args[2];
2972    const_a2 = const_args[2];
2973    rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2974
2975    switch (opc) {
2976    case INDEX_op_goto_ptr:
2977        /* jmp to the given host address (could be epilogue) */
2978        tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, a0);
2979        break;
2980    case INDEX_op_br:
2981        tcg_out_jxx(s, JCC_JMP, arg_label(a0), 0);
2982        break;
2983    OP_32_64(ld8u):
2984        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
2985        tcg_out_modrm_offset(s, OPC_MOVZBL, a0, a1, a2);
2986        break;
2987    OP_32_64(ld8s):
2988        tcg_out_modrm_offset(s, OPC_MOVSBL + rexw, a0, a1, a2);
2989        break;
2990    OP_32_64(ld16u):
2991        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
2992        tcg_out_modrm_offset(s, OPC_MOVZWL, a0, a1, a2);
2993        break;
2994    OP_32_64(ld16s):
2995        tcg_out_modrm_offset(s, OPC_MOVSWL + rexw, a0, a1, a2);
2996        break;
2997#if TCG_TARGET_REG_BITS == 64
2998    case INDEX_op_ld32u_i64:
2999#endif
3000    case INDEX_op_ld_i32:
3001        tcg_out_ld(s, TCG_TYPE_I32, a0, a1, a2);
3002        break;
3003
3004    OP_32_64(st8):
3005        if (const_args[0]) {
3006            tcg_out_modrm_offset(s, OPC_MOVB_EvIz, 0, a1, a2);
3007            tcg_out8(s, a0);
3008        } else {
3009            tcg_out_modrm_offset(s, OPC_MOVB_EvGv | P_REXB_R, a0, a1, a2);
3010        }
3011        break;
3012    OP_32_64(st16):
3013        if (const_args[0]) {
3014            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_DATA16, 0, a1, a2);
3015            tcg_out16(s, a0);
3016        } else {
3017            tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_DATA16, a0, a1, a2);
3018        }
3019        break;
3020#if TCG_TARGET_REG_BITS == 64
3021    case INDEX_op_st32_i64:
3022#endif
3023    case INDEX_op_st_i32:
3024        if (const_args[0]) {
3025            tcg_out_modrm_offset(s, OPC_MOVL_EvIz, 0, a1, a2);
3026            tcg_out32(s, a0);
3027        } else {
3028            tcg_out_st(s, TCG_TYPE_I32, a0, a1, a2);
3029        }
3030        break;
3031
3032    OP_32_64(ctz):
3033        tcg_out_ctz(s, rexw, args[0], args[1], args[2], const_args[2]);
3034        break;
3035    OP_32_64(ctpop):
3036        tcg_out_modrm(s, OPC_POPCNT + rexw, a0, a1);
3037        break;
3038
3039    OP_32_64(brcond):
3040        tcg_out_brcond(s, rexw, a2, a0, a1, const_args[1],
3041                       arg_label(args[3]), 0);
3042        break;
3043    OP_32_64(setcond):
3044        tcg_out_setcond(s, rexw, args[3], a0, a1, a2, const_a2, false);
3045        break;
3046    OP_32_64(negsetcond):
3047        tcg_out_setcond(s, rexw, args[3], a0, a1, a2, const_a2, true);
3048        break;
3049    OP_32_64(movcond):
3050        tcg_out_movcond(s, rexw, args[5], a0, a1, a2, const_a2, args[3]);
3051        break;
3052
3053    OP_32_64(bswap16):
3054        if (a2 & TCG_BSWAP_OS) {
3055            /* Output must be sign-extended. */
3056            if (rexw) {
3057                tcg_out_bswap64(s, a0);
3058                tcg_out_shifti(s, SHIFT_SAR + rexw, a0, 48);
3059            } else {
3060                tcg_out_bswap32(s, a0);
3061                tcg_out_shifti(s, SHIFT_SAR, a0, 16);
3062            }
3063        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
3064            /* Output must be zero-extended, but input isn't. */
3065            tcg_out_bswap32(s, a0);
3066            tcg_out_shifti(s, SHIFT_SHR, a0, 16);
3067        } else {
3068            tcg_out_rolw_8(s, a0);
3069        }
3070        break;
3071    OP_32_64(bswap32):
3072        tcg_out_bswap32(s, a0);
3073        if (rexw && (a2 & TCG_BSWAP_OS)) {
3074            tcg_out_ext32s(s, a0, a0);
3075        }
3076        break;
3077
3078    case INDEX_op_qemu_ld_i32:
3079        tcg_out_qemu_ld(s, a0, -1, a1, a2, TCG_TYPE_I32);
3080        break;
3081    case INDEX_op_qemu_ld_i64:
3082        if (TCG_TARGET_REG_BITS == 64) {
3083            tcg_out_qemu_ld(s, a0, -1, a1, a2, TCG_TYPE_I64);
3084        } else {
3085            tcg_out_qemu_ld(s, a0, a1, a2, args[3], TCG_TYPE_I64);
3086        }
3087        break;
3088    case INDEX_op_qemu_ld_i128:
3089        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3090        tcg_out_qemu_ld(s, a0, a1, a2, args[3], TCG_TYPE_I128);
3091        break;
3092
3093    case INDEX_op_qemu_st_i32:
3094    case INDEX_op_qemu_st8_i32:
3095        tcg_out_qemu_st(s, a0, -1, a1, a2, TCG_TYPE_I32);
3096        break;
3097    case INDEX_op_qemu_st_i64:
3098        if (TCG_TARGET_REG_BITS == 64) {
3099            tcg_out_qemu_st(s, a0, -1, a1, a2, TCG_TYPE_I64);
3100        } else {
3101            tcg_out_qemu_st(s, a0, a1, a2, args[3], TCG_TYPE_I64);
3102        }
3103        break;
3104    case INDEX_op_qemu_st_i128:
3105        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3106        tcg_out_qemu_st(s, a0, a1, a2, args[3], TCG_TYPE_I128);
3107        break;
3108
3109    OP_32_64(mulu2):
3110        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_MUL, args[3]);
3111        break;
3112    OP_32_64(muls2):
3113        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IMUL, args[3]);
3114        break;
3115    OP_32_64(add2):
3116        if (const_args[4]) {
3117            tgen_arithi(s, ARITH_ADD + rexw, a0, args[4], 1);
3118        } else {
3119            tgen_arithr(s, ARITH_ADD + rexw, a0, args[4]);
3120        }
3121        if (const_args[5]) {
3122            tgen_arithi(s, ARITH_ADC + rexw, a1, args[5], 1);
3123        } else {
3124            tgen_arithr(s, ARITH_ADC + rexw, a1, args[5]);
3125        }
3126        break;
3127    OP_32_64(sub2):
3128        if (const_args[4]) {
3129            tgen_arithi(s, ARITH_SUB + rexw, a0, args[4], 1);
3130        } else {
3131            tgen_arithr(s, ARITH_SUB + rexw, a0, args[4]);
3132        }
3133        if (const_args[5]) {
3134            tgen_arithi(s, ARITH_SBB + rexw, a1, args[5], 1);
3135        } else {
3136            tgen_arithr(s, ARITH_SBB + rexw, a1, args[5]);
3137        }
3138        break;
3139
3140#if TCG_TARGET_REG_BITS == 32
3141    case INDEX_op_brcond2_i32:
3142        tcg_out_brcond2(s, args, const_args, 0);
3143        break;
3144    case INDEX_op_setcond2_i32:
3145        tcg_out_setcond2(s, args, const_args);
3146        break;
3147#else /* TCG_TARGET_REG_BITS == 64 */
3148    case INDEX_op_ld32s_i64:
3149        tcg_out_modrm_offset(s, OPC_MOVSLQ, a0, a1, a2);
3150        break;
3151    case INDEX_op_ld_i64:
3152        tcg_out_ld(s, TCG_TYPE_I64, a0, a1, a2);
3153        break;
3154    case INDEX_op_st_i64:
3155        if (const_args[0]) {
3156            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_REXW, 0, a1, a2);
3157            tcg_out32(s, a0);
3158        } else {
3159            tcg_out_st(s, TCG_TYPE_I64, a0, a1, a2);
3160        }
3161        break;
3162
3163    case INDEX_op_bswap64_i64:
3164        tcg_out_bswap64(s, a0);
3165        break;
3166    case INDEX_op_extrh_i64_i32:
3167        tcg_out_shifti(s, SHIFT_SHR + P_REXW, a0, 32);
3168        break;
3169#endif
3170
3171    OP_32_64(deposit):
3172        if (args[3] == 0 && args[4] == 8) {
3173            /* load bits 0..7 */
3174            if (const_a2) {
3175                tcg_out_opc(s, OPC_MOVB_Ib | P_REXB_RM | LOWREGMASK(a0),
3176                            0, a0, 0);
3177                tcg_out8(s, a2);
3178            } else {
3179                tcg_out_modrm(s, OPC_MOVB_EvGv | P_REXB_R | P_REXB_RM, a2, a0);
3180            }
3181        } else if (TCG_TARGET_REG_BITS == 32 && args[3] == 8 && args[4] == 8) {
3182            /* load bits 8..15 */
3183            if (const_a2) {
3184                tcg_out8(s, OPC_MOVB_Ib + a0 + 4);
3185                tcg_out8(s, a2);
3186            } else {
3187                tcg_out_modrm(s, OPC_MOVB_EvGv, a2, a0 + 4);
3188            }
3189        } else if (args[3] == 0 && args[4] == 16) {
3190            /* load bits 0..15 */
3191            if (const_a2) {
3192                tcg_out_opc(s, OPC_MOVL_Iv | P_DATA16 | LOWREGMASK(a0),
3193                            0, a0, 0);
3194                tcg_out16(s, a2);
3195            } else {
3196                tcg_out_modrm(s, OPC_MOVL_EvGv | P_DATA16, a2, a0);
3197            }
3198        } else {
3199            g_assert_not_reached();
3200        }
3201        break;
3202
3203    case INDEX_op_extract_i64:
3204        if (a2 + args[3] == 32) {
3205            if (a2 == 0) {
3206                tcg_out_ext32u(s, a0, a1);
3207                break;
3208            }
3209            /* This is a 32-bit zero-extending right shift.  */
3210            tcg_out_mov(s, TCG_TYPE_I32, a0, a1);
3211            tcg_out_shifti(s, SHIFT_SHR, a0, a2);
3212            break;
3213        }
3214        /* FALLTHRU */
3215    case INDEX_op_extract_i32:
3216        if (a2 == 0 && args[3] == 8) {
3217            tcg_out_ext8u(s, a0, a1);
3218        } else if (a2 == 0 && args[3] == 16) {
3219            tcg_out_ext16u(s, a0, a1);
3220        } else if (a2 == 8 && args[3] == 8) {
3221            /*
3222             * On the off-chance that we can use the high-byte registers.
3223             * Otherwise we emit the same ext16 + shift pattern that we
3224             * would have gotten from the normal tcg-op.c expansion.
3225             */
3226            if (a1 < 4 && a0 < 8) {
3227                tcg_out_modrm(s, OPC_MOVZBL, a0, a1 + 4);
3228            } else {
3229                tcg_out_ext16u(s, a0, a1);
3230                tcg_out_shifti(s, SHIFT_SHR, a0, 8);
3231            }
3232        } else {
3233            g_assert_not_reached();
3234        }
3235        break;
3236
3237    case INDEX_op_sextract_i64:
3238        if (a2 == 0 && args[3] == 8) {
3239            tcg_out_ext8s(s, TCG_TYPE_I64, a0, a1);
3240        } else if (a2 == 0 && args[3] == 16) {
3241            tcg_out_ext16s(s, TCG_TYPE_I64, a0, a1);
3242        } else if (a2 == 0 && args[3] == 32) {
3243            tcg_out_ext32s(s, a0, a1);
3244        } else {
3245            g_assert_not_reached();
3246        }
3247        break;
3248
3249    case INDEX_op_sextract_i32:
3250        if (a2 == 0 && args[3] == 8) {
3251            tcg_out_ext8s(s, TCG_TYPE_I32, a0, a1);
3252        } else if (a2 == 0 && args[3] == 16) {
3253            tcg_out_ext16s(s, TCG_TYPE_I32, a0, a1);
3254        } else if (a2 == 8 && args[3] == 8) {
3255            if (a1 < 4 && a0 < 8) {
3256                tcg_out_modrm(s, OPC_MOVSBL, a0, a1 + 4);
3257            } else {
3258                tcg_out_ext16s(s, TCG_TYPE_I32, a0, a1);
3259                tcg_out_shifti(s, SHIFT_SAR, a0, 8);
3260            }
3261        } else {
3262            g_assert_not_reached();
3263        }
3264        break;
3265
3266    OP_32_64(extract2):
3267        /* Note that SHRD outputs to the r/m operand.  */
3268        tcg_out_modrm(s, OPC_SHRD_Ib + rexw, a2, a0);
3269        tcg_out8(s, args[3]);
3270        break;
3271
3272    case INDEX_op_mb:
3273        tcg_out_mb(s, a0);
3274        break;
3275    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
3276    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
3277    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
3278    case INDEX_op_ext_i32_i64:  /* Always emitted via tcg_reg_alloc_op.  */
3279    case INDEX_op_extu_i32_i64:
3280    case INDEX_op_extrl_i64_i32:
3281    default:
3282        g_assert_not_reached();
3283    }
3284
3285#undef OP_32_64
3286}
3287
3288static int const umin_insn[4] = {
3289    OPC_PMINUB, OPC_PMINUW, OPC_PMINUD, OPC_VPMINUQ
3290};
3291
3292static int const umax_insn[4] = {
3293    OPC_PMAXUB, OPC_PMAXUW, OPC_PMAXUD, OPC_VPMAXUQ
3294};
3295
3296static bool tcg_out_cmp_vec_noinv(TCGContext *s, TCGType type, unsigned vece,
3297                                  TCGReg v0, TCGReg v1, TCGReg v2, TCGCond cond)
3298{
3299    static int const cmpeq_insn[4] = {
3300        OPC_PCMPEQB, OPC_PCMPEQW, OPC_PCMPEQD, OPC_PCMPEQQ
3301    };
3302    static int const cmpgt_insn[4] = {
3303        OPC_PCMPGTB, OPC_PCMPGTW, OPC_PCMPGTD, OPC_PCMPGTQ
3304    };
3305
3306    enum {
3307        NEED_INV  = 1,
3308        NEED_SWAP = 2,
3309        NEED_UMIN = 4,
3310        NEED_UMAX = 8,
3311        INVALID   = 16,
3312    };
3313    static const uint8_t cond_fixup[16] = {
3314        [0 ... 15] = INVALID,
3315        [TCG_COND_EQ] = 0,
3316        [TCG_COND_GT] = 0,
3317        [TCG_COND_NE] = NEED_INV,
3318        [TCG_COND_LE] = NEED_INV,
3319        [TCG_COND_LT] = NEED_SWAP,
3320        [TCG_COND_GE] = NEED_SWAP | NEED_INV,
3321        [TCG_COND_LEU] = NEED_UMIN,
3322        [TCG_COND_GTU] = NEED_UMIN | NEED_INV,
3323        [TCG_COND_GEU] = NEED_UMAX,
3324        [TCG_COND_LTU] = NEED_UMAX | NEED_INV,
3325    };
3326    int fixup = cond_fixup[cond];
3327
3328    assert(!(fixup & INVALID));
3329
3330    if (fixup & NEED_INV) {
3331        cond = tcg_invert_cond(cond);
3332    }
3333
3334    if (fixup & NEED_SWAP) {
3335        TCGReg swap = v1;
3336        v1 = v2;
3337        v2 = swap;
3338        cond = tcg_swap_cond(cond);
3339    }
3340
3341    if (fixup & (NEED_UMIN | NEED_UMAX)) {
3342        int op = (fixup & NEED_UMIN ? umin_insn[vece] : umax_insn[vece]);
3343
3344        /* avx2 does not have 64-bit min/max; adjusted during expand. */
3345        assert(vece <= MO_32);
3346
3347        tcg_out_vex_modrm_type(s, op, TCG_TMP_VEC, v1, v2, type);
3348        v2 = TCG_TMP_VEC;
3349        cond = TCG_COND_EQ;
3350    }
3351
3352    switch (cond) {
3353    case TCG_COND_EQ:
3354        tcg_out_vex_modrm_type(s, cmpeq_insn[vece], v0, v1, v2, type);
3355        break;
3356    case TCG_COND_GT:
3357        tcg_out_vex_modrm_type(s, cmpgt_insn[vece], v0, v1, v2, type);
3358        break;
3359    default:
3360        g_assert_not_reached();
3361    }
3362    return fixup & NEED_INV;
3363}
3364
3365static void tcg_out_cmp_vec_k1(TCGContext *s, TCGType type, unsigned vece,
3366                               TCGReg v1, TCGReg v2, TCGCond cond)
3367{
3368    static const int cmpm_insn[2][4] = {
3369        { OPC_VPCMPB, OPC_VPCMPW, OPC_VPCMPD, OPC_VPCMPQ },
3370        { OPC_VPCMPUB, OPC_VPCMPUW, OPC_VPCMPUD, OPC_VPCMPUQ }
3371    };
3372    static const int testm_insn[4] = {
3373        OPC_VPTESTMB, OPC_VPTESTMW, OPC_VPTESTMD, OPC_VPTESTMQ
3374    };
3375    static const int testnm_insn[4] = {
3376        OPC_VPTESTNMB, OPC_VPTESTNMW, OPC_VPTESTNMD, OPC_VPTESTNMQ
3377    };
3378
3379    static const int cond_ext[16] = {
3380        [TCG_COND_EQ] = 0,
3381        [TCG_COND_NE] = 4,
3382        [TCG_COND_LT] = 1,
3383        [TCG_COND_LTU] = 1,
3384        [TCG_COND_LE] = 2,
3385        [TCG_COND_LEU] = 2,
3386        [TCG_COND_NEVER] = 3,
3387        [TCG_COND_GE] = 5,
3388        [TCG_COND_GEU] = 5,
3389        [TCG_COND_GT] = 6,
3390        [TCG_COND_GTU] = 6,
3391        [TCG_COND_ALWAYS] = 7,
3392    };
3393
3394    switch (cond) {
3395    case TCG_COND_TSTNE:
3396        tcg_out_vex_modrm_type(s, testm_insn[vece], /* k1 */ 1, v1, v2, type);
3397        break;
3398    case TCG_COND_TSTEQ:
3399        tcg_out_vex_modrm_type(s, testnm_insn[vece], /* k1 */ 1, v1, v2, type);
3400        break;
3401    default:
3402        tcg_out_vex_modrm_type(s, cmpm_insn[is_unsigned_cond(cond)][vece],
3403                               /* k1 */ 1, v1, v2, type);
3404        tcg_out8(s, cond_ext[cond]);
3405        break;
3406    }
3407}
3408
3409static void tcg_out_k1_to_vec(TCGContext *s, TCGType type,
3410                              unsigned vece, TCGReg dest)
3411{
3412    static const int movm_insn[] = {
3413        OPC_VPMOVM2B, OPC_VPMOVM2W, OPC_VPMOVM2D, OPC_VPMOVM2Q
3414    };
3415    tcg_out_vex_modrm_type(s, movm_insn[vece], dest, 0, /* k1 */ 1, type);
3416}
3417
3418static void tcg_out_cmp_vec(TCGContext *s, TCGType type, unsigned vece,
3419                            TCGReg v0, TCGReg v1, TCGReg v2, TCGCond cond)
3420{
3421    /*
3422     * With avx512, we have a complete set of comparisons into mask.
3423     * Unless there's a single insn expansion for the comparision,
3424     * expand via a mask in k1.
3425     */
3426    if ((vece <= MO_16 ? have_avx512bw : have_avx512dq)
3427        && cond != TCG_COND_EQ
3428        && cond != TCG_COND_LT
3429        && cond != TCG_COND_GT) {
3430        tcg_out_cmp_vec_k1(s, type, vece, v1, v2, cond);
3431        tcg_out_k1_to_vec(s, type, vece, v0);
3432        return;
3433    }
3434
3435    if (tcg_out_cmp_vec_noinv(s, type, vece, v0, v1, v2, cond)) {
3436        tcg_out_dupi_vec(s, type, vece, TCG_TMP_VEC, -1);
3437        tcg_out_vex_modrm_type(s, OPC_PXOR, v0, v0, TCG_TMP_VEC, type);
3438    }
3439}
3440
3441static void tcg_out_cmpsel_vec_k1(TCGContext *s, TCGType type, unsigned vece,
3442                                  TCGReg v0, TCGReg c1, TCGReg c2,
3443                                  TCGReg v3, TCGReg v4, TCGCond cond)
3444{
3445    static const int vpblendm_insn[] = {
3446        OPC_VPBLENDMB, OPC_VPBLENDMW, OPC_VPBLENDMD, OPC_VPBLENDMQ
3447    };
3448    bool z = false;
3449
3450    /* Swap to place constant in V4 to take advantage of zero-masking. */
3451    if (!v3) {
3452        z = true;
3453        v3 = v4;
3454        cond = tcg_invert_cond(cond);
3455    }
3456
3457    tcg_out_cmp_vec_k1(s, type, vece, c1, c2, cond);
3458    tcg_out_evex_modrm_type(s, vpblendm_insn[vece], v0, v4, v3,
3459                            /* k1 */1, z, type);
3460}
3461
3462static void tcg_out_cmpsel_vec(TCGContext *s, TCGType type, unsigned vece,
3463                               TCGReg v0, TCGReg c1, TCGReg c2,
3464                               TCGReg v3, TCGReg v4, TCGCond cond)
3465{
3466    bool inv;
3467
3468    if (vece <= MO_16 ? have_avx512bw : have_avx512vl) {
3469        tcg_out_cmpsel_vec_k1(s, type, vece, v0, c1, c2, v3, v4, cond);
3470        return;
3471    }
3472
3473    inv = tcg_out_cmp_vec_noinv(s, type, vece, TCG_TMP_VEC, c1, c2, cond);
3474
3475    /*
3476     * Since XMM0 is 16, the only way we get 0 into V3
3477     * is via the constant zero constraint.
3478     */
3479    if (!v3) {
3480        if (inv) {
3481            tcg_out_vex_modrm_type(s, OPC_PAND, v0, TCG_TMP_VEC, v4, type);
3482        } else {
3483            tcg_out_vex_modrm_type(s, OPC_PANDN, v0, TCG_TMP_VEC, v4, type);
3484        }
3485    } else {
3486        if (inv) {
3487            TCGReg swap = v3;
3488            v3 = v4;
3489            v4 = swap;
3490        }
3491        tcg_out_vex_modrm_type(s, OPC_VPBLENDVB, v0, v4, v3, type);
3492        tcg_out8(s, (TCG_TMP_VEC - TCG_REG_XMM0) << 4);
3493    }
3494}
3495
3496static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
3497                           unsigned vecl, unsigned vece,
3498                           const TCGArg args[TCG_MAX_OP_ARGS],
3499                           const int const_args[TCG_MAX_OP_ARGS])
3500{
3501    static int const add_insn[4] = {
3502        OPC_PADDB, OPC_PADDW, OPC_PADDD, OPC_PADDQ
3503    };
3504    static int const ssadd_insn[4] = {
3505        OPC_PADDSB, OPC_PADDSW, OPC_UD2, OPC_UD2
3506    };
3507    static int const usadd_insn[4] = {
3508        OPC_PADDUB, OPC_PADDUW, OPC_UD2, OPC_UD2
3509    };
3510    static int const sub_insn[4] = {
3511        OPC_PSUBB, OPC_PSUBW, OPC_PSUBD, OPC_PSUBQ
3512    };
3513    static int const sssub_insn[4] = {
3514        OPC_PSUBSB, OPC_PSUBSW, OPC_UD2, OPC_UD2
3515    };
3516    static int const ussub_insn[4] = {
3517        OPC_PSUBUB, OPC_PSUBUW, OPC_UD2, OPC_UD2
3518    };
3519    static int const mul_insn[4] = {
3520        OPC_UD2, OPC_PMULLW, OPC_PMULLD, OPC_VPMULLQ
3521    };
3522    static int const shift_imm_insn[4] = {
3523        OPC_UD2, OPC_PSHIFTW_Ib, OPC_PSHIFTD_Ib, OPC_PSHIFTQ_Ib
3524    };
3525    static int const punpckl_insn[4] = {
3526        OPC_PUNPCKLBW, OPC_PUNPCKLWD, OPC_PUNPCKLDQ, OPC_PUNPCKLQDQ
3527    };
3528    static int const punpckh_insn[4] = {
3529        OPC_PUNPCKHBW, OPC_PUNPCKHWD, OPC_PUNPCKHDQ, OPC_PUNPCKHQDQ
3530    };
3531    static int const packss_insn[4] = {
3532        OPC_PACKSSWB, OPC_PACKSSDW, OPC_UD2, OPC_UD2
3533    };
3534    static int const packus_insn[4] = {
3535        OPC_PACKUSWB, OPC_PACKUSDW, OPC_UD2, OPC_UD2
3536    };
3537    static int const smin_insn[4] = {
3538        OPC_PMINSB, OPC_PMINSW, OPC_PMINSD, OPC_VPMINSQ
3539    };
3540    static int const smax_insn[4] = {
3541        OPC_PMAXSB, OPC_PMAXSW, OPC_PMAXSD, OPC_VPMAXSQ
3542    };
3543    static int const rotlv_insn[4] = {
3544        OPC_UD2, OPC_UD2, OPC_VPROLVD, OPC_VPROLVQ
3545    };
3546    static int const rotrv_insn[4] = {
3547        OPC_UD2, OPC_UD2, OPC_VPRORVD, OPC_VPRORVQ
3548    };
3549    static int const shlv_insn[4] = {
3550        OPC_UD2, OPC_VPSLLVW, OPC_VPSLLVD, OPC_VPSLLVQ
3551    };
3552    static int const shrv_insn[4] = {
3553        OPC_UD2, OPC_VPSRLVW, OPC_VPSRLVD, OPC_VPSRLVQ
3554    };
3555    static int const sarv_insn[4] = {
3556        OPC_UD2, OPC_VPSRAVW, OPC_VPSRAVD, OPC_VPSRAVQ
3557    };
3558    static int const shls_insn[4] = {
3559        OPC_UD2, OPC_PSLLW, OPC_PSLLD, OPC_PSLLQ
3560    };
3561    static int const shrs_insn[4] = {
3562        OPC_UD2, OPC_PSRLW, OPC_PSRLD, OPC_PSRLQ
3563    };
3564    static int const sars_insn[4] = {
3565        OPC_UD2, OPC_PSRAW, OPC_PSRAD, OPC_VPSRAQ
3566    };
3567    static int const vpshldi_insn[4] = {
3568        OPC_UD2, OPC_VPSHLDW, OPC_VPSHLDD, OPC_VPSHLDQ
3569    };
3570    static int const vpshldv_insn[4] = {
3571        OPC_UD2, OPC_VPSHLDVW, OPC_VPSHLDVD, OPC_VPSHLDVQ
3572    };
3573    static int const vpshrdv_insn[4] = {
3574        OPC_UD2, OPC_VPSHRDVW, OPC_VPSHRDVD, OPC_VPSHRDVQ
3575    };
3576    static int const abs_insn[4] = {
3577        OPC_PABSB, OPC_PABSW, OPC_PABSD, OPC_VPABSQ
3578    };
3579
3580    TCGType type = vecl + TCG_TYPE_V64;
3581    int insn, sub;
3582    TCGArg a0, a1, a2, a3;
3583
3584    a0 = args[0];
3585    a1 = args[1];
3586    a2 = args[2];
3587
3588    switch (opc) {
3589    case INDEX_op_add_vec:
3590        insn = add_insn[vece];
3591        goto gen_simd;
3592    case INDEX_op_ssadd_vec:
3593        insn = ssadd_insn[vece];
3594        goto gen_simd;
3595    case INDEX_op_usadd_vec:
3596        insn = usadd_insn[vece];
3597        goto gen_simd;
3598    case INDEX_op_sub_vec:
3599        insn = sub_insn[vece];
3600        goto gen_simd;
3601    case INDEX_op_sssub_vec:
3602        insn = sssub_insn[vece];
3603        goto gen_simd;
3604    case INDEX_op_ussub_vec:
3605        insn = ussub_insn[vece];
3606        goto gen_simd;
3607    case INDEX_op_mul_vec:
3608        insn = mul_insn[vece];
3609        goto gen_simd;
3610    case INDEX_op_and_vec:
3611        insn = OPC_PAND;
3612        goto gen_simd;
3613    case INDEX_op_or_vec:
3614        insn = OPC_POR;
3615        goto gen_simd;
3616    case INDEX_op_xor_vec:
3617        insn = OPC_PXOR;
3618        goto gen_simd;
3619    case INDEX_op_smin_vec:
3620        insn = smin_insn[vece];
3621        goto gen_simd;
3622    case INDEX_op_umin_vec:
3623        insn = umin_insn[vece];
3624        goto gen_simd;
3625    case INDEX_op_smax_vec:
3626        insn = smax_insn[vece];
3627        goto gen_simd;
3628    case INDEX_op_umax_vec:
3629        insn = umax_insn[vece];
3630        goto gen_simd;
3631    case INDEX_op_shlv_vec:
3632        insn = shlv_insn[vece];
3633        goto gen_simd;
3634    case INDEX_op_shrv_vec:
3635        insn = shrv_insn[vece];
3636        goto gen_simd;
3637    case INDEX_op_sarv_vec:
3638        insn = sarv_insn[vece];
3639        goto gen_simd;
3640    case INDEX_op_rotlv_vec:
3641        insn = rotlv_insn[vece];
3642        goto gen_simd;
3643    case INDEX_op_rotrv_vec:
3644        insn = rotrv_insn[vece];
3645        goto gen_simd;
3646    case INDEX_op_shls_vec:
3647        insn = shls_insn[vece];
3648        goto gen_simd;
3649    case INDEX_op_shrs_vec:
3650        insn = shrs_insn[vece];
3651        goto gen_simd;
3652    case INDEX_op_sars_vec:
3653        insn = sars_insn[vece];
3654        goto gen_simd;
3655    case INDEX_op_x86_punpckl_vec:
3656        insn = punpckl_insn[vece];
3657        goto gen_simd;
3658    case INDEX_op_x86_punpckh_vec:
3659        insn = punpckh_insn[vece];
3660        goto gen_simd;
3661    case INDEX_op_x86_packss_vec:
3662        insn = packss_insn[vece];
3663        goto gen_simd;
3664    case INDEX_op_x86_packus_vec:
3665        insn = packus_insn[vece];
3666        goto gen_simd;
3667    case INDEX_op_x86_vpshldv_vec:
3668        insn = vpshldv_insn[vece];
3669        a1 = a2;
3670        a2 = args[3];
3671        goto gen_simd;
3672    case INDEX_op_x86_vpshrdv_vec:
3673        insn = vpshrdv_insn[vece];
3674        a1 = a2;
3675        a2 = args[3];
3676        goto gen_simd;
3677#if TCG_TARGET_REG_BITS == 32
3678    case INDEX_op_dup2_vec:
3679        /* First merge the two 32-bit inputs to a single 64-bit element. */
3680        tcg_out_vex_modrm(s, OPC_PUNPCKLDQ, a0, a1, a2);
3681        /* Then replicate the 64-bit elements across the rest of the vector. */
3682        if (type != TCG_TYPE_V64) {
3683            tcg_out_dup_vec(s, type, MO_64, a0, a0);
3684        }
3685        break;
3686#endif
3687    case INDEX_op_abs_vec:
3688        insn = abs_insn[vece];
3689        a2 = a1;
3690        a1 = 0;
3691        goto gen_simd;
3692    gen_simd:
3693        tcg_debug_assert(insn != OPC_UD2);
3694        tcg_out_vex_modrm_type(s, insn, a0, a1, a2, type);
3695        break;
3696
3697    case INDEX_op_cmp_vec:
3698        tcg_out_cmp_vec(s, type, vece, a0, a1, a2, args[3]);
3699        break;
3700
3701    case INDEX_op_cmpsel_vec:
3702        tcg_out_cmpsel_vec(s, type, vece, a0, a1, a2,
3703                           args[3], args[4], args[5]);
3704        break;
3705
3706    case INDEX_op_andc_vec:
3707        insn = OPC_PANDN;
3708        tcg_out_vex_modrm_type(s, insn, a0, a2, a1, type);
3709        break;
3710
3711    case INDEX_op_shli_vec:
3712        insn = shift_imm_insn[vece];
3713        sub = 6;
3714        goto gen_shift;
3715    case INDEX_op_shri_vec:
3716        insn = shift_imm_insn[vece];
3717        sub = 2;
3718        goto gen_shift;
3719    case INDEX_op_sari_vec:
3720        if (vece == MO_64) {
3721            insn = OPC_PSHIFTD_Ib | P_VEXW | P_EVEX;
3722        } else {
3723            insn = shift_imm_insn[vece];
3724        }
3725        sub = 4;
3726        goto gen_shift;
3727    case INDEX_op_rotli_vec:
3728        insn = OPC_PSHIFTD_Ib | P_EVEX;  /* VPROL[DQ] */
3729        if (vece == MO_64) {
3730            insn |= P_VEXW;
3731        }
3732        sub = 1;
3733        goto gen_shift;
3734    gen_shift:
3735        tcg_debug_assert(vece != MO_8);
3736        tcg_out_vex_modrm_type(s, insn, sub, a0, a1, type);
3737        tcg_out8(s, a2);
3738        break;
3739
3740    case INDEX_op_ld_vec:
3741        tcg_out_ld(s, type, a0, a1, a2);
3742        break;
3743    case INDEX_op_st_vec:
3744        tcg_out_st(s, type, a0, a1, a2);
3745        break;
3746    case INDEX_op_dupm_vec:
3747        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
3748        break;
3749
3750    case INDEX_op_x86_shufps_vec:
3751        insn = OPC_SHUFPS;
3752        sub = args[3];
3753        goto gen_simd_imm8;
3754    case INDEX_op_x86_blend_vec:
3755        if (vece == MO_16) {
3756            insn = OPC_PBLENDW;
3757        } else if (vece == MO_32) {
3758            insn = (have_avx2 ? OPC_VPBLENDD : OPC_BLENDPS);
3759        } else {
3760            g_assert_not_reached();
3761        }
3762        sub = args[3];
3763        goto gen_simd_imm8;
3764    case INDEX_op_x86_vperm2i128_vec:
3765        insn = OPC_VPERM2I128;
3766        sub = args[3];
3767        goto gen_simd_imm8;
3768    case INDEX_op_x86_vpshldi_vec:
3769        insn = vpshldi_insn[vece];
3770        sub = args[3];
3771        goto gen_simd_imm8;
3772
3773    case INDEX_op_not_vec:
3774        insn = OPC_VPTERNLOGQ;
3775        a2 = a1;
3776        sub = 0x33; /* !B */
3777        goto gen_simd_imm8;
3778    case INDEX_op_nor_vec:
3779        insn = OPC_VPTERNLOGQ;
3780        sub = 0x11; /* norCB */
3781        goto gen_simd_imm8;
3782    case INDEX_op_nand_vec:
3783        insn = OPC_VPTERNLOGQ;
3784        sub = 0x77; /* nandCB */
3785        goto gen_simd_imm8;
3786    case INDEX_op_eqv_vec:
3787        insn = OPC_VPTERNLOGQ;
3788        sub = 0x99; /* xnorCB */
3789        goto gen_simd_imm8;
3790    case INDEX_op_orc_vec:
3791        insn = OPC_VPTERNLOGQ;
3792        sub = 0xdd; /* orB!C */
3793        goto gen_simd_imm8;
3794
3795    case INDEX_op_bitsel_vec:
3796        insn = OPC_VPTERNLOGQ;
3797        a3 = args[3];
3798        if (a0 == a1) {
3799            a1 = a2;
3800            a2 = a3;
3801            sub = 0xca; /* A?B:C */
3802        } else if (a0 == a2) {
3803            a2 = a3;
3804            sub = 0xe2; /* B?A:C */
3805        } else {
3806            tcg_out_mov(s, type, a0, a3);
3807            sub = 0xb8; /* B?C:A */
3808        }
3809        goto gen_simd_imm8;
3810
3811    gen_simd_imm8:
3812        tcg_debug_assert(insn != OPC_UD2);
3813        tcg_out_vex_modrm_type(s, insn, a0, a1, a2, type);
3814        tcg_out8(s, sub);
3815        break;
3816
3817    case INDEX_op_x86_psrldq_vec:
3818        tcg_out_vex_modrm(s, OPC_GRP14, 3, a0, a1);
3819        tcg_out8(s, a2);
3820        break;
3821
3822    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
3823    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
3824    default:
3825        g_assert_not_reached();
3826    }
3827}
3828
3829static TCGConstraintSetIndex
3830tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
3831{
3832    switch (op) {
3833    case INDEX_op_goto_ptr:
3834        return C_O0_I1(r);
3835
3836    case INDEX_op_ld8u_i32:
3837    case INDEX_op_ld8u_i64:
3838    case INDEX_op_ld8s_i32:
3839    case INDEX_op_ld8s_i64:
3840    case INDEX_op_ld16u_i32:
3841    case INDEX_op_ld16u_i64:
3842    case INDEX_op_ld16s_i32:
3843    case INDEX_op_ld16s_i64:
3844    case INDEX_op_ld_i32:
3845    case INDEX_op_ld32u_i64:
3846    case INDEX_op_ld32s_i64:
3847    case INDEX_op_ld_i64:
3848        return C_O1_I1(r, r);
3849
3850    case INDEX_op_st8_i32:
3851    case INDEX_op_st8_i64:
3852        return C_O0_I2(qi, r);
3853
3854    case INDEX_op_st16_i32:
3855    case INDEX_op_st16_i64:
3856    case INDEX_op_st_i32:
3857    case INDEX_op_st32_i64:
3858        return C_O0_I2(ri, r);
3859
3860    case INDEX_op_st_i64:
3861        return C_O0_I2(re, r);
3862
3863    case INDEX_op_brcond_i32:
3864    case INDEX_op_brcond_i64:
3865        return C_O0_I2(r, reT);
3866
3867    case INDEX_op_bswap16_i32:
3868    case INDEX_op_bswap16_i64:
3869    case INDEX_op_bswap32_i32:
3870    case INDEX_op_bswap32_i64:
3871    case INDEX_op_bswap64_i64:
3872    case INDEX_op_extrh_i64_i32:
3873        return C_O1_I1(r, 0);
3874
3875    case INDEX_op_ext_i32_i64:
3876    case INDEX_op_extu_i32_i64:
3877    case INDEX_op_extrl_i64_i32:
3878    case INDEX_op_extract_i32:
3879    case INDEX_op_extract_i64:
3880    case INDEX_op_sextract_i32:
3881    case INDEX_op_sextract_i64:
3882    case INDEX_op_ctpop_i32:
3883    case INDEX_op_ctpop_i64:
3884        return C_O1_I1(r, r);
3885
3886    case INDEX_op_extract2_i32:
3887    case INDEX_op_extract2_i64:
3888        return C_O1_I2(r, 0, r);
3889
3890    case INDEX_op_deposit_i32:
3891    case INDEX_op_deposit_i64:
3892        return C_O1_I2(q, 0, qi);
3893
3894    case INDEX_op_setcond_i32:
3895    case INDEX_op_setcond_i64:
3896    case INDEX_op_negsetcond_i32:
3897    case INDEX_op_negsetcond_i64:
3898        return C_O1_I2(q, r, reT);
3899
3900    case INDEX_op_movcond_i32:
3901    case INDEX_op_movcond_i64:
3902        return C_O1_I4(r, r, reT, r, 0);
3903
3904    case INDEX_op_mulu2_i32:
3905    case INDEX_op_mulu2_i64:
3906    case INDEX_op_muls2_i32:
3907    case INDEX_op_muls2_i64:
3908        return C_O2_I2(a, d, a, r);
3909
3910    case INDEX_op_add2_i32:
3911    case INDEX_op_add2_i64:
3912    case INDEX_op_sub2_i32:
3913    case INDEX_op_sub2_i64:
3914        return C_N1_O1_I4(r, r, 0, 1, re, re);
3915
3916    case INDEX_op_ctz_i32:
3917    case INDEX_op_ctz_i64:
3918        return have_bmi1 ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
3919
3920    case INDEX_op_qemu_ld_i32:
3921        return C_O1_I1(r, L);
3922
3923    case INDEX_op_qemu_st_i32:
3924        return C_O0_I2(L, L);
3925    case INDEX_op_qemu_st8_i32:
3926        return C_O0_I2(s, L);
3927
3928    case INDEX_op_qemu_ld_i64:
3929        return TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L) : C_O2_I1(r, r, L);
3930
3931    case INDEX_op_qemu_st_i64:
3932        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I3(L, L, L);
3933
3934    case INDEX_op_qemu_ld_i128:
3935        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3936        return C_O2_I1(r, r, L);
3937    case INDEX_op_qemu_st_i128:
3938        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3939        return C_O0_I3(L, L, L);
3940
3941    case INDEX_op_brcond2_i32:
3942        return C_O0_I4(r, r, ri, ri);
3943
3944    case INDEX_op_setcond2_i32:
3945        return C_O1_I4(r, r, r, ri, ri);
3946
3947    case INDEX_op_ld_vec:
3948    case INDEX_op_dupm_vec:
3949        return C_O1_I1(x, r);
3950
3951    case INDEX_op_st_vec:
3952        return C_O0_I2(x, r);
3953
3954    case INDEX_op_add_vec:
3955    case INDEX_op_sub_vec:
3956    case INDEX_op_mul_vec:
3957    case INDEX_op_and_vec:
3958    case INDEX_op_or_vec:
3959    case INDEX_op_xor_vec:
3960    case INDEX_op_andc_vec:
3961    case INDEX_op_orc_vec:
3962    case INDEX_op_nand_vec:
3963    case INDEX_op_nor_vec:
3964    case INDEX_op_eqv_vec:
3965    case INDEX_op_ssadd_vec:
3966    case INDEX_op_usadd_vec:
3967    case INDEX_op_sssub_vec:
3968    case INDEX_op_ussub_vec:
3969    case INDEX_op_smin_vec:
3970    case INDEX_op_umin_vec:
3971    case INDEX_op_smax_vec:
3972    case INDEX_op_umax_vec:
3973    case INDEX_op_shlv_vec:
3974    case INDEX_op_shrv_vec:
3975    case INDEX_op_sarv_vec:
3976    case INDEX_op_rotlv_vec:
3977    case INDEX_op_rotrv_vec:
3978    case INDEX_op_shls_vec:
3979    case INDEX_op_shrs_vec:
3980    case INDEX_op_sars_vec:
3981    case INDEX_op_cmp_vec:
3982    case INDEX_op_x86_shufps_vec:
3983    case INDEX_op_x86_blend_vec:
3984    case INDEX_op_x86_packss_vec:
3985    case INDEX_op_x86_packus_vec:
3986    case INDEX_op_x86_vperm2i128_vec:
3987    case INDEX_op_x86_punpckl_vec:
3988    case INDEX_op_x86_punpckh_vec:
3989    case INDEX_op_x86_vpshldi_vec:
3990#if TCG_TARGET_REG_BITS == 32
3991    case INDEX_op_dup2_vec:
3992#endif
3993        return C_O1_I2(x, x, x);
3994
3995    case INDEX_op_abs_vec:
3996    case INDEX_op_dup_vec:
3997    case INDEX_op_not_vec:
3998    case INDEX_op_shli_vec:
3999    case INDEX_op_shri_vec:
4000    case INDEX_op_sari_vec:
4001    case INDEX_op_rotli_vec:
4002    case INDEX_op_x86_psrldq_vec:
4003        return C_O1_I1(x, x);
4004
4005    case INDEX_op_x86_vpshldv_vec:
4006    case INDEX_op_x86_vpshrdv_vec:
4007        return C_O1_I3(x, 0, x, x);
4008
4009    case INDEX_op_bitsel_vec:
4010        return C_O1_I3(x, x, x, x);
4011    case INDEX_op_cmpsel_vec:
4012        return C_O1_I4(x, x, x, xO, x);
4013
4014    default:
4015        return C_NotImplemented;
4016    }
4017}
4018
4019int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
4020{
4021    switch (opc) {
4022    case INDEX_op_add_vec:
4023    case INDEX_op_sub_vec:
4024    case INDEX_op_and_vec:
4025    case INDEX_op_or_vec:
4026    case INDEX_op_xor_vec:
4027    case INDEX_op_andc_vec:
4028    case INDEX_op_orc_vec:
4029    case INDEX_op_nand_vec:
4030    case INDEX_op_nor_vec:
4031    case INDEX_op_eqv_vec:
4032    case INDEX_op_not_vec:
4033    case INDEX_op_bitsel_vec:
4034        return 1;
4035    case INDEX_op_cmp_vec:
4036    case INDEX_op_cmpsel_vec:
4037        return -1;
4038
4039    case INDEX_op_rotli_vec:
4040        return have_avx512vl && vece >= MO_32 ? 1 : -1;
4041
4042    case INDEX_op_shli_vec:
4043    case INDEX_op_shri_vec:
4044        /* We must expand the operation for MO_8.  */
4045        return vece == MO_8 ? -1 : 1;
4046
4047    case INDEX_op_sari_vec:
4048        switch (vece) {
4049        case MO_8:
4050            return -1;
4051        case MO_16:
4052        case MO_32:
4053            return 1;
4054        case MO_64:
4055            if (have_avx512vl) {
4056                return 1;
4057            }
4058            /*
4059             * We can emulate this for MO_64, but it does not pay off
4060             * unless we're producing at least 4 values.
4061             */
4062            return type >= TCG_TYPE_V256 ? -1 : 0;
4063        }
4064        return 0;
4065
4066    case INDEX_op_shls_vec:
4067    case INDEX_op_shrs_vec:
4068        return vece >= MO_16;
4069    case INDEX_op_sars_vec:
4070        switch (vece) {
4071        case MO_16:
4072        case MO_32:
4073            return 1;
4074        case MO_64:
4075            return have_avx512vl;
4076        }
4077        return 0;
4078    case INDEX_op_rotls_vec:
4079        return vece >= MO_16 ? -1 : 0;
4080
4081    case INDEX_op_shlv_vec:
4082    case INDEX_op_shrv_vec:
4083        switch (vece) {
4084        case MO_16:
4085            return have_avx512bw;
4086        case MO_32:
4087        case MO_64:
4088            return have_avx2;
4089        }
4090        return 0;
4091    case INDEX_op_sarv_vec:
4092        switch (vece) {
4093        case MO_16:
4094            return have_avx512bw;
4095        case MO_32:
4096            return have_avx2;
4097        case MO_64:
4098            return have_avx512vl;
4099        }
4100        return 0;
4101    case INDEX_op_rotlv_vec:
4102    case INDEX_op_rotrv_vec:
4103        switch (vece) {
4104        case MO_16:
4105            return have_avx512vbmi2 ? -1 : 0;
4106        case MO_32:
4107        case MO_64:
4108            return have_avx512vl ? 1 : have_avx2 ? -1 : 0;
4109        }
4110        return 0;
4111
4112    case INDEX_op_mul_vec:
4113        switch (vece) {
4114        case MO_8:
4115            return -1;
4116        case MO_64:
4117            return have_avx512dq;
4118        }
4119        return 1;
4120
4121    case INDEX_op_ssadd_vec:
4122    case INDEX_op_usadd_vec:
4123    case INDEX_op_sssub_vec:
4124    case INDEX_op_ussub_vec:
4125        return vece <= MO_16;
4126    case INDEX_op_smin_vec:
4127    case INDEX_op_smax_vec:
4128    case INDEX_op_umin_vec:
4129    case INDEX_op_umax_vec:
4130    case INDEX_op_abs_vec:
4131        return vece <= MO_32 || have_avx512vl;
4132
4133    default:
4134        return 0;
4135    }
4136}
4137
4138static void expand_vec_shi(TCGType type, unsigned vece, bool right,
4139                           TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4140{
4141    uint8_t mask;
4142
4143    tcg_debug_assert(vece == MO_8);
4144    if (right) {
4145        mask = 0xff >> imm;
4146        tcg_gen_shri_vec(MO_16, v0, v1, imm);
4147    } else {
4148        mask = 0xff << imm;
4149        tcg_gen_shli_vec(MO_16, v0, v1, imm);
4150    }
4151    tcg_gen_and_vec(MO_8, v0, v0, tcg_constant_vec(type, MO_8, mask));
4152}
4153
4154static void expand_vec_sari(TCGType type, unsigned vece,
4155                            TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4156{
4157    TCGv_vec t1, t2;
4158
4159    switch (vece) {
4160    case MO_8:
4161        /* Unpack to 16-bit, shift, and repack.  */
4162        t1 = tcg_temp_new_vec(type);
4163        t2 = tcg_temp_new_vec(type);
4164        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4165                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
4166        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4167                  tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
4168        tcg_gen_sari_vec(MO_16, t1, t1, imm + 8);
4169        tcg_gen_sari_vec(MO_16, t2, t2, imm + 8);
4170        vec_gen_3(INDEX_op_x86_packss_vec, type, MO_8,
4171                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
4172        tcg_temp_free_vec(t1);
4173        tcg_temp_free_vec(t2);
4174        break;
4175
4176    case MO_64:
4177        t1 = tcg_temp_new_vec(type);
4178        if (imm <= 32) {
4179            /*
4180             * We can emulate a small sign extend by performing an arithmetic
4181             * 32-bit shift and overwriting the high half of a 64-bit logical
4182             * shift.  Note that the ISA says shift of 32 is valid, but TCG
4183             * does not, so we have to bound the smaller shift -- we get the
4184             * same result in the high half either way.
4185             */
4186            tcg_gen_sari_vec(MO_32, t1, v1, MIN(imm, 31));
4187            tcg_gen_shri_vec(MO_64, v0, v1, imm);
4188            vec_gen_4(INDEX_op_x86_blend_vec, type, MO_32,
4189                      tcgv_vec_arg(v0), tcgv_vec_arg(v0),
4190                      tcgv_vec_arg(t1), 0xaa);
4191        } else {
4192            /* Otherwise we will need to use a compare vs 0 to produce
4193             * the sign-extend, shift and merge.
4194             */
4195            tcg_gen_cmp_vec(TCG_COND_GT, MO_64, t1,
4196                            tcg_constant_vec(type, MO_64, 0), v1);
4197            tcg_gen_shri_vec(MO_64, v0, v1, imm);
4198            tcg_gen_shli_vec(MO_64, t1, t1, 64 - imm);
4199            tcg_gen_or_vec(MO_64, v0, v0, t1);
4200        }
4201        tcg_temp_free_vec(t1);
4202        break;
4203
4204    default:
4205        g_assert_not_reached();
4206    }
4207}
4208
4209static void expand_vec_rotli(TCGType type, unsigned vece,
4210                             TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4211{
4212    TCGv_vec t;
4213
4214    if (vece != MO_8 && have_avx512vbmi2) {
4215        vec_gen_4(INDEX_op_x86_vpshldi_vec, type, vece,
4216                  tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v1), imm);
4217        return;
4218    }
4219
4220    t = tcg_temp_new_vec(type);
4221    tcg_gen_shli_vec(vece, t, v1, imm);
4222    tcg_gen_shri_vec(vece, v0, v1, (8 << vece) - imm);
4223    tcg_gen_or_vec(vece, v0, v0, t);
4224    tcg_temp_free_vec(t);
4225}
4226
4227static void expand_vec_rotv(TCGType type, unsigned vece, TCGv_vec v0,
4228                            TCGv_vec v1, TCGv_vec sh, bool right)
4229{
4230    TCGv_vec t;
4231
4232    if (have_avx512vbmi2) {
4233        vec_gen_4(right ? INDEX_op_x86_vpshrdv_vec : INDEX_op_x86_vpshldv_vec,
4234                  type, vece, tcgv_vec_arg(v0), tcgv_vec_arg(v1),
4235                  tcgv_vec_arg(v1), tcgv_vec_arg(sh));
4236        return;
4237    }
4238
4239    t = tcg_temp_new_vec(type);
4240    tcg_gen_dupi_vec(vece, t, 8 << vece);
4241    tcg_gen_sub_vec(vece, t, t, sh);
4242    if (right) {
4243        tcg_gen_shlv_vec(vece, t, v1, t);
4244        tcg_gen_shrv_vec(vece, v0, v1, sh);
4245    } else {
4246        tcg_gen_shrv_vec(vece, t, v1, t);
4247        tcg_gen_shlv_vec(vece, v0, v1, sh);
4248    }
4249    tcg_gen_or_vec(vece, v0, v0, t);
4250    tcg_temp_free_vec(t);
4251}
4252
4253static void expand_vec_rotls(TCGType type, unsigned vece,
4254                             TCGv_vec v0, TCGv_vec v1, TCGv_i32 lsh)
4255{
4256    TCGv_vec t = tcg_temp_new_vec(type);
4257
4258    tcg_debug_assert(vece != MO_8);
4259
4260    if (vece >= MO_32 ? have_avx512vl : have_avx512vbmi2) {
4261        tcg_gen_dup_i32_vec(vece, t, lsh);
4262        if (vece >= MO_32) {
4263            tcg_gen_rotlv_vec(vece, v0, v1, t);
4264        } else {
4265            expand_vec_rotv(type, vece, v0, v1, t, false);
4266        }
4267    } else {
4268        TCGv_i32 rsh = tcg_temp_new_i32();
4269
4270        tcg_gen_neg_i32(rsh, lsh);
4271        tcg_gen_andi_i32(rsh, rsh, (8 << vece) - 1);
4272        tcg_gen_shls_vec(vece, t, v1, lsh);
4273        tcg_gen_shrs_vec(vece, v0, v1, rsh);
4274        tcg_gen_or_vec(vece, v0, v0, t);
4275
4276        tcg_temp_free_i32(rsh);
4277    }
4278
4279    tcg_temp_free_vec(t);
4280}
4281
4282static void expand_vec_mul(TCGType type, unsigned vece,
4283                           TCGv_vec v0, TCGv_vec v1, TCGv_vec v2)
4284{
4285    TCGv_vec t1, t2, t3, t4, zero;
4286
4287    tcg_debug_assert(vece == MO_8);
4288
4289    /*
4290     * Unpack v1 bytes to words, 0 | x.
4291     * Unpack v2 bytes to words, y | 0.
4292     * This leaves the 8-bit result, x * y, with 8 bits of right padding.
4293     * Shift logical right by 8 bits to clear the high 8 bytes before
4294     * using an unsigned saturated pack.
4295     *
4296     * The difference between the V64, V128 and V256 cases is merely how
4297     * we distribute the expansion between temporaries.
4298     */
4299    switch (type) {
4300    case TCG_TYPE_V64:
4301        t1 = tcg_temp_new_vec(TCG_TYPE_V128);
4302        t2 = tcg_temp_new_vec(TCG_TYPE_V128);
4303        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
4304        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
4305                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4306        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
4307                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4308        tcg_gen_mul_vec(MO_16, t1, t1, t2);
4309        tcg_gen_shri_vec(MO_16, t1, t1, 8);
4310        vec_gen_3(INDEX_op_x86_packus_vec, TCG_TYPE_V128, MO_8,
4311                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t1));
4312        tcg_temp_free_vec(t1);
4313        tcg_temp_free_vec(t2);
4314        break;
4315
4316    case TCG_TYPE_V128:
4317    case TCG_TYPE_V256:
4318        t1 = tcg_temp_new_vec(type);
4319        t2 = tcg_temp_new_vec(type);
4320        t3 = tcg_temp_new_vec(type);
4321        t4 = tcg_temp_new_vec(type);
4322        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
4323        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4324                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4325        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4326                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4327        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4328                  tcgv_vec_arg(t3), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4329        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4330                  tcgv_vec_arg(t4), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4331        tcg_gen_mul_vec(MO_16, t1, t1, t2);
4332        tcg_gen_mul_vec(MO_16, t3, t3, t4);
4333        tcg_gen_shri_vec(MO_16, t1, t1, 8);
4334        tcg_gen_shri_vec(MO_16, t3, t3, 8);
4335        vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
4336                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t3));
4337        tcg_temp_free_vec(t1);
4338        tcg_temp_free_vec(t2);
4339        tcg_temp_free_vec(t3);
4340        tcg_temp_free_vec(t4);
4341        break;
4342
4343    default:
4344        g_assert_not_reached();
4345    }
4346}
4347
4348static TCGCond expand_vec_cond(TCGType type, unsigned vece,
4349                               TCGArg *a1, TCGArg *a2, TCGCond cond)
4350{
4351    /*
4352     * Without AVX512, there are no 64-bit unsigned comparisons.
4353     * We must bias the inputs so that they become signed.
4354     * All other swapping and inversion are handled during code generation.
4355     */
4356    if (vece == MO_64 && !have_avx512dq && is_unsigned_cond(cond)) {
4357        TCGv_vec v1 = temp_tcgv_vec(arg_temp(*a1));
4358        TCGv_vec v2 = temp_tcgv_vec(arg_temp(*a2));
4359        TCGv_vec t1 = tcg_temp_new_vec(type);
4360        TCGv_vec t2 = tcg_temp_new_vec(type);
4361        TCGv_vec t3 = tcg_constant_vec(type, vece, 1ull << ((8 << vece) - 1));
4362
4363        tcg_gen_sub_vec(vece, t1, v1, t3);
4364        tcg_gen_sub_vec(vece, t2, v2, t3);
4365        *a1 = tcgv_vec_arg(t1);
4366        *a2 = tcgv_vec_arg(t2);
4367        cond = tcg_signed_cond(cond);
4368    }
4369    return cond;
4370}
4371
4372static void expand_vec_cmp(TCGType type, unsigned vece, TCGArg a0,
4373                           TCGArg a1, TCGArg a2, TCGCond cond)
4374{
4375    cond = expand_vec_cond(type, vece, &a1, &a2, cond);
4376    /* Expand directly; do not recurse.  */
4377    vec_gen_4(INDEX_op_cmp_vec, type, vece, a0, a1, a2, cond);
4378}
4379
4380static void expand_vec_cmpsel(TCGType type, unsigned vece, TCGArg a0,
4381                              TCGArg a1, TCGArg a2,
4382                              TCGArg a3, TCGArg a4, TCGCond cond)
4383{
4384    cond = expand_vec_cond(type, vece, &a1, &a2, cond);
4385    /* Expand directly; do not recurse.  */
4386    vec_gen_6(INDEX_op_cmpsel_vec, type, vece, a0, a1, a2, a3, a4, cond);
4387}
4388
4389void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
4390                       TCGArg a0, ...)
4391{
4392    va_list va;
4393    TCGArg a1, a2, a3, a4, a5;
4394    TCGv_vec v0, v1, v2;
4395
4396    va_start(va, a0);
4397    a1 = va_arg(va, TCGArg);
4398    a2 = va_arg(va, TCGArg);
4399    v0 = temp_tcgv_vec(arg_temp(a0));
4400    v1 = temp_tcgv_vec(arg_temp(a1));
4401
4402    switch (opc) {
4403    case INDEX_op_shli_vec:
4404        expand_vec_shi(type, vece, false, v0, v1, a2);
4405        break;
4406    case INDEX_op_shri_vec:
4407        expand_vec_shi(type, vece, true, v0, v1, a2);
4408        break;
4409    case INDEX_op_sari_vec:
4410        expand_vec_sari(type, vece, v0, v1, a2);
4411        break;
4412
4413    case INDEX_op_rotli_vec:
4414        expand_vec_rotli(type, vece, v0, v1, a2);
4415        break;
4416
4417    case INDEX_op_rotls_vec:
4418        expand_vec_rotls(type, vece, v0, v1, temp_tcgv_i32(arg_temp(a2)));
4419        break;
4420
4421    case INDEX_op_rotlv_vec:
4422        v2 = temp_tcgv_vec(arg_temp(a2));
4423        expand_vec_rotv(type, vece, v0, v1, v2, false);
4424        break;
4425    case INDEX_op_rotrv_vec:
4426        v2 = temp_tcgv_vec(arg_temp(a2));
4427        expand_vec_rotv(type, vece, v0, v1, v2, true);
4428        break;
4429
4430    case INDEX_op_mul_vec:
4431        v2 = temp_tcgv_vec(arg_temp(a2));
4432        expand_vec_mul(type, vece, v0, v1, v2);
4433        break;
4434
4435    case INDEX_op_cmp_vec:
4436        a3 = va_arg(va, TCGArg);
4437        expand_vec_cmp(type, vece, a0, a1, a2, a3);
4438        break;
4439
4440    case INDEX_op_cmpsel_vec:
4441        a3 = va_arg(va, TCGArg);
4442        a4 = va_arg(va, TCGArg);
4443        a5 = va_arg(va, TCGArg);
4444        expand_vec_cmpsel(type, vece, a0, a1, a2, a3, a4, a5);
4445        break;
4446
4447    default:
4448        break;
4449    }
4450
4451    va_end(va);
4452}
4453
4454static const int tcg_target_callee_save_regs[] = {
4455#if TCG_TARGET_REG_BITS == 64
4456    TCG_REG_RBP,
4457    TCG_REG_RBX,
4458#if defined(_WIN64)
4459    TCG_REG_RDI,
4460    TCG_REG_RSI,
4461#endif
4462    TCG_REG_R12,
4463    TCG_REG_R13,
4464    TCG_REG_R14, /* Currently used for the global env. */
4465    TCG_REG_R15,
4466#else
4467    TCG_REG_EBP, /* Currently used for the global env. */
4468    TCG_REG_EBX,
4469    TCG_REG_ESI,
4470    TCG_REG_EDI,
4471#endif
4472};
4473
4474/* Compute frame size via macros, to share between tcg_target_qemu_prologue
4475   and tcg_register_jit.  */
4476
4477#define PUSH_SIZE \
4478    ((1 + ARRAY_SIZE(tcg_target_callee_save_regs)) \
4479     * (TCG_TARGET_REG_BITS / 8))
4480
4481#define FRAME_SIZE \
4482    ((PUSH_SIZE \
4483      + TCG_STATIC_CALL_ARGS_SIZE \
4484      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
4485      + TCG_TARGET_STACK_ALIGN - 1) \
4486     & ~(TCG_TARGET_STACK_ALIGN - 1))
4487
4488/* Generate global QEMU prologue and epilogue code */
4489static void tcg_target_qemu_prologue(TCGContext *s)
4490{
4491    int i, stack_addend;
4492
4493    /* TB prologue */
4494
4495    /* Reserve some stack space, also for TCG temps.  */
4496    stack_addend = FRAME_SIZE - PUSH_SIZE;
4497    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
4498                  CPU_TEMP_BUF_NLONGS * sizeof(long));
4499
4500    /* Save all callee saved registers.  */
4501    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
4502        tcg_out_push(s, tcg_target_callee_save_regs[i]);
4503    }
4504
4505    if (!tcg_use_softmmu && guest_base) {
4506        int seg = setup_guest_base_seg();
4507        if (seg != 0) {
4508            x86_guest_base.seg = seg;
4509        } else if (guest_base == (int32_t)guest_base) {
4510            x86_guest_base.ofs = guest_base;
4511        } else {
4512            assert(TCG_TARGET_REG_BITS == 64);
4513            /* Choose R12 because, as a base, it requires a SIB byte. */
4514            x86_guest_base.index = TCG_REG_R12;
4515            tcg_out_movi(s, TCG_TYPE_PTR, x86_guest_base.index, guest_base);
4516            tcg_regset_set_reg(s->reserved_regs, x86_guest_base.index);
4517        }
4518    }
4519
4520    if (TCG_TARGET_REG_BITS == 32) {
4521        tcg_out_ld(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP,
4522                   (ARRAY_SIZE(tcg_target_callee_save_regs) + 1) * 4);
4523        tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4524        /* jmp *tb.  */
4525        tcg_out_modrm_offset(s, OPC_GRP5, EXT5_JMPN_Ev, TCG_REG_ESP,
4526                             (ARRAY_SIZE(tcg_target_callee_save_regs) + 2) * 4
4527                             + stack_addend);
4528    } else {
4529        tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
4530        tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4531        /* jmp *tb.  */
4532        tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, tcg_target_call_iarg_regs[1]);
4533    }
4534
4535    /*
4536     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
4537     * and fall through to the rest of the epilogue.
4538     */
4539    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
4540    tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_EAX, 0);
4541
4542    /* TB epilogue */
4543    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
4544
4545    tcg_out_addi(s, TCG_REG_CALL_STACK, stack_addend);
4546
4547    if (have_avx2) {
4548        tcg_out_vex_opc(s, OPC_VZEROUPPER, 0, 0, 0, 0);
4549    }
4550    for (i = ARRAY_SIZE(tcg_target_callee_save_regs) - 1; i >= 0; i--) {
4551        tcg_out_pop(s, tcg_target_callee_save_regs[i]);
4552    }
4553    tcg_out_opc(s, OPC_RET, 0, 0, 0);
4554}
4555
4556static void tcg_out_tb_start(TCGContext *s)
4557{
4558    /* nothing to do */
4559}
4560
4561static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
4562{
4563    memset(p, 0x90, count);
4564}
4565
4566static void tcg_target_init(TCGContext *s)
4567{
4568    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
4569    if (TCG_TARGET_REG_BITS == 64) {
4570        tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
4571    }
4572    if (have_avx1) {
4573        tcg_target_available_regs[TCG_TYPE_V64] = ALL_VECTOR_REGS;
4574        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
4575    }
4576    if (have_avx2) {
4577        tcg_target_available_regs[TCG_TYPE_V256] = ALL_VECTOR_REGS;
4578    }
4579
4580    tcg_target_call_clobber_regs = ALL_VECTOR_REGS;
4581    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EAX);
4582    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EDX);
4583    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_ECX);
4584    if (TCG_TARGET_REG_BITS == 64) {
4585#if !defined(_WIN64)
4586        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RDI);
4587        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RSI);
4588#endif
4589        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R8);
4590        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R9);
4591        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R10);
4592        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R11);
4593    }
4594
4595    s->reserved_regs = 0;
4596    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
4597    tcg_regset_set_reg(s->reserved_regs, TCG_TMP_VEC);
4598#ifdef _WIN64
4599    /* These are call saved, and we don't save them, so don't use them. */
4600    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM6);
4601    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM7);
4602    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM8);
4603    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM9);
4604    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM10);
4605    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM11);
4606    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM12);
4607    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM13);
4608    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM14);
4609    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM15);
4610#endif
4611}
4612
4613typedef struct {
4614    DebugFrameHeader h;
4615    uint8_t fde_def_cfa[4];
4616    uint8_t fde_reg_ofs[14];
4617} DebugFrame;
4618
4619/* We're expecting a 2 byte uleb128 encoded value.  */
4620QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
4621
4622#if !defined(__ELF__)
4623    /* Host machine without ELF. */
4624#elif TCG_TARGET_REG_BITS == 64
4625#define ELF_HOST_MACHINE EM_X86_64
4626static const DebugFrame debug_frame = {
4627    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4628    .h.cie.id = -1,
4629    .h.cie.version = 1,
4630    .h.cie.code_align = 1,
4631    .h.cie.data_align = 0x78,             /* sleb128 -8 */
4632    .h.cie.return_column = 16,
4633
4634    /* Total FDE size does not include the "len" member.  */
4635    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4636
4637    .fde_def_cfa = {
4638        12, 7,                          /* DW_CFA_def_cfa %rsp, ... */
4639        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4640        (FRAME_SIZE >> 7)
4641    },
4642    .fde_reg_ofs = {
4643        0x90, 1,                        /* DW_CFA_offset, %rip, -8 */
4644        /* The following ordering must match tcg_target_callee_save_regs.  */
4645        0x86, 2,                        /* DW_CFA_offset, %rbp, -16 */
4646        0x83, 3,                        /* DW_CFA_offset, %rbx, -24 */
4647        0x8c, 4,                        /* DW_CFA_offset, %r12, -32 */
4648        0x8d, 5,                        /* DW_CFA_offset, %r13, -40 */
4649        0x8e, 6,                        /* DW_CFA_offset, %r14, -48 */
4650        0x8f, 7,                        /* DW_CFA_offset, %r15, -56 */
4651    }
4652};
4653#else
4654#define ELF_HOST_MACHINE EM_386
4655static const DebugFrame debug_frame = {
4656    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4657    .h.cie.id = -1,
4658    .h.cie.version = 1,
4659    .h.cie.code_align = 1,
4660    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
4661    .h.cie.return_column = 8,
4662
4663    /* Total FDE size does not include the "len" member.  */
4664    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4665
4666    .fde_def_cfa = {
4667        12, 4,                          /* DW_CFA_def_cfa %esp, ... */
4668        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4669        (FRAME_SIZE >> 7)
4670    },
4671    .fde_reg_ofs = {
4672        0x88, 1,                        /* DW_CFA_offset, %eip, -4 */
4673        /* The following ordering must match tcg_target_callee_save_regs.  */
4674        0x85, 2,                        /* DW_CFA_offset, %ebp, -8 */
4675        0x83, 3,                        /* DW_CFA_offset, %ebx, -12 */
4676        0x86, 4,                        /* DW_CFA_offset, %esi, -16 */
4677        0x87, 5,                        /* DW_CFA_offset, %edi, -20 */
4678    }
4679};
4680#endif
4681
4682#if defined(ELF_HOST_MACHINE)
4683void tcg_register_jit(const void *buf, size_t buf_size)
4684{
4685    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
4686}
4687#endif
4688