xref: /openbmc/qemu/tcg/i386/tcg-target.c.inc (revision 07d5d502f2b4a8eedda3c6bdfcab31dc36d1d1d5)
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
1645static void tgen_brcond(TCGContext *s, TCGType type, TCGCond cond,
1646                        TCGReg arg1, TCGReg arg2, TCGLabel *label)
1647{
1648    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1649    tcg_out_brcond(s, rexw, cond, arg1, arg2, false, label, false);
1650}
1651
1652static void tgen_brcondi(TCGContext *s, TCGType type, TCGCond cond,
1653                         TCGReg arg1, tcg_target_long arg2, TCGLabel *label)
1654{
1655    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1656    tcg_out_brcond(s, rexw, cond, arg1, arg2, true, label, false);
1657}
1658
1659static const TCGOutOpBrcond outop_brcond = {
1660    .base.static_constraint = C_O0_I2(r, reT),
1661    .out_rr = tgen_brcond,
1662    .out_ri = tgen_brcondi,
1663};
1664
1665static void tcg_out_brcond2(TCGContext *s, TCGCond cond, TCGReg al,
1666                            TCGReg ah, TCGArg bl, bool blconst,
1667                            TCGArg bh, bool bhconst,
1668                            TCGLabel *label_this, bool small)
1669{
1670    TCGLabel *label_next = gen_new_label();
1671
1672    switch (cond) {
1673    case TCG_COND_EQ:
1674    case TCG_COND_TSTEQ:
1675        tcg_out_brcond(s, 0, tcg_invert_cond(cond),
1676                       al, bl, blconst, label_next, true);
1677        tcg_out_brcond(s, 0, cond, ah, bh, bhconst, label_this, small);
1678        break;
1679
1680    case TCG_COND_NE:
1681    case TCG_COND_TSTNE:
1682        tcg_out_brcond(s, 0, cond, al, bl, blconst, label_this, small);
1683        tcg_out_brcond(s, 0, cond, ah, bh, bhconst, label_this, small);
1684        break;
1685
1686    default:
1687        tcg_out_brcond(s, 0, tcg_high_cond(cond),
1688                       ah, bh, bhconst, label_this, small);
1689        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1690        tcg_out_brcond(s, 0, tcg_unsigned_cond(cond),
1691                       al, bl, blconst, label_this, small);
1692        break;
1693    }
1694    tcg_out_label(s, label_next);
1695}
1696
1697static void tgen_brcond2(TCGContext *s, TCGCond cond, TCGReg al,
1698                         TCGReg ah, TCGArg bl, bool blconst,
1699                         TCGArg bh, bool bhconst, TCGLabel *l)
1700{
1701    tcg_out_brcond2(s, cond, al, ah, bl, blconst, bh, bhconst, l, false);
1702}
1703
1704#if TCG_TARGET_REG_BITS != 32
1705__attribute__((unused))
1706#endif
1707static const TCGOutOpBrcond2 outop_brcond2 = {
1708    .base.static_constraint = C_O0_I4(r, r, ri, ri),
1709    .out = tgen_brcond2,
1710};
1711
1712static void tcg_out_setcond(TCGContext *s, TCGType type, TCGCond cond,
1713                            TCGReg dest, TCGReg arg1, TCGArg arg2,
1714                            bool const_arg2, bool neg)
1715{
1716    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1717    int cmp_rexw = rexw;
1718    bool inv = false;
1719    bool cleared;
1720    int jcc;
1721
1722    switch (cond) {
1723    case TCG_COND_NE:
1724        inv = true;
1725        /* fall through */
1726    case TCG_COND_EQ:
1727        /* If arg2 is 0, convert to LTU/GEU vs 1. */
1728        if (const_arg2 && arg2 == 0) {
1729            arg2 = 1;
1730            goto do_ltu;
1731        }
1732        break;
1733
1734    case TCG_COND_TSTNE:
1735        inv = true;
1736        /* fall through */
1737    case TCG_COND_TSTEQ:
1738        /* If arg2 is -1, convert to LTU/GEU vs 1. */
1739        if (const_arg2 && arg2 == 0xffffffffu) {
1740            arg2 = 1;
1741            cmp_rexw = 0;
1742            goto do_ltu;
1743        }
1744        break;
1745
1746    case TCG_COND_LEU:
1747        inv = true;
1748        /* fall through */
1749    case TCG_COND_GTU:
1750        /* If arg2 is a register, swap for LTU/GEU. */
1751        if (!const_arg2) {
1752            TCGReg t = arg1;
1753            arg1 = arg2;
1754            arg2 = t;
1755            goto do_ltu;
1756        }
1757        break;
1758
1759    case TCG_COND_GEU:
1760        inv = true;
1761        /* fall through */
1762    case TCG_COND_LTU:
1763    do_ltu:
1764        /*
1765         * Relying on the carry bit, use SBB to produce -1 if LTU, 0 if GEU.
1766         * We can then use NEG or INC to produce the desired result.
1767         * This is always smaller than the SETCC expansion.
1768         */
1769        tcg_out_cmp(s, TCG_COND_LTU, arg1, arg2, const_arg2, cmp_rexw);
1770
1771        /* X - X - C = -C = (C ? -1 : 0) */
1772        tgen_arithr(s, ARITH_SBB + (neg ? rexw : 0), dest, dest);
1773        if (inv && neg) {
1774            /* ~(C ? -1 : 0) = (C ? 0 : -1) */
1775            tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, dest);
1776        } else if (inv) {
1777            /* (C ? -1 : 0) + 1 = (C ? 0 : 1) */
1778            tgen_arithi(s, ARITH_ADD, dest, 1, 0);
1779        } else if (!neg) {
1780            /* -(C ? -1 : 0) = (C ? 1 : 0) */
1781            tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_NEG, dest);
1782        }
1783        return;
1784
1785    case TCG_COND_GE:
1786        inv = true;
1787        /* fall through */
1788    case TCG_COND_LT:
1789        /* If arg2 is 0, extract the sign bit. */
1790        if (const_arg2 && arg2 == 0) {
1791            tcg_out_mov(s, type, dest, arg1);
1792            if (inv) {
1793                tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, dest);
1794            }
1795            tcg_out_shifti(s, (neg ? SHIFT_SAR : SHIFT_SHR) + rexw,
1796                           dest, rexw ? 63 : 31);
1797            return;
1798        }
1799        break;
1800
1801    default:
1802        break;
1803    }
1804
1805    /*
1806     * If dest does not overlap the inputs, clearing it first is preferred.
1807     * The XOR breaks any false dependency for the low-byte write to dest,
1808     * and is also one byte smaller than MOVZBL.
1809     */
1810    cleared = false;
1811    if (dest != arg1 && (const_arg2 || dest != arg2)) {
1812        tgen_arithr(s, ARITH_XOR, dest, dest);
1813        cleared = true;
1814    }
1815
1816    jcc = tcg_out_cmp(s, cond, arg1, arg2, const_arg2, cmp_rexw);
1817    tcg_out_modrm(s, OPC_SETCC | jcc, 0, dest);
1818
1819    if (!cleared) {
1820        tcg_out_ext8u(s, dest, dest);
1821    }
1822    if (neg) {
1823        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, dest);
1824    }
1825}
1826
1827static void tgen_setcond(TCGContext *s, TCGType type, TCGCond cond,
1828                         TCGReg dest, TCGReg arg1, TCGReg arg2)
1829{
1830    tcg_out_setcond(s, type, cond, dest, arg1, arg2, false, false);
1831}
1832
1833static void tgen_setcondi(TCGContext *s, TCGType type, TCGCond cond,
1834                          TCGReg dest, TCGReg arg1, tcg_target_long arg2)
1835{
1836    tcg_out_setcond(s, type, cond, dest, arg1, arg2, true, false);
1837}
1838
1839static const TCGOutOpSetcond outop_setcond = {
1840    .base.static_constraint = C_O1_I2(q, r, reT),
1841    .out_rrr = tgen_setcond,
1842    .out_rri = tgen_setcondi,
1843};
1844
1845static void tgen_negsetcond(TCGContext *s, TCGType type, TCGCond cond,
1846                            TCGReg dest, TCGReg arg1, TCGReg arg2)
1847{
1848    tcg_out_setcond(s, type, cond, dest, arg1, arg2, false, true);
1849}
1850
1851static void tgen_negsetcondi(TCGContext *s, TCGType type, TCGCond cond,
1852                             TCGReg dest, TCGReg arg1, tcg_target_long arg2)
1853{
1854    tcg_out_setcond(s, type, cond, dest, arg1, arg2, true, true);
1855}
1856
1857static const TCGOutOpSetcond outop_negsetcond = {
1858    .base.static_constraint = C_O1_I2(q, r, reT),
1859    .out_rrr = tgen_negsetcond,
1860    .out_rri = tgen_negsetcondi,
1861};
1862
1863static void tgen_setcond2(TCGContext *s, TCGCond cond, TCGReg ret,
1864                          TCGReg al, TCGReg ah,
1865                          TCGArg bl, bool const_bl,
1866                          TCGArg bh, bool const_bh)
1867{
1868    TCGLabel *label_over = gen_new_label();
1869
1870    if (ret == al || ret == ah
1871        || (!const_bl && ret == bl)
1872        || (!const_bh && ret == bh)) {
1873        /*
1874         * When the destination overlaps with one of the argument
1875         * registers, don't do anything tricky.
1876         */
1877        TCGLabel *label_true = gen_new_label();
1878
1879        tcg_out_brcond2(s, cond, al, ah, bl, const_bl,
1880                        bh, const_bh, label_true, true);
1881
1882        tcg_out_movi(s, TCG_TYPE_I32, ret, 0);
1883        tcg_out_jxx(s, JCC_JMP, label_over, 1);
1884        tcg_out_label(s, label_true);
1885
1886        tcg_out_movi(s, TCG_TYPE_I32, ret, 1);
1887    } else {
1888        /*
1889         * When the destination does not overlap one of the arguments,
1890         * clear the destination first, jump if cond false, and emit an
1891         * increment in the true case.  This results in smaller code.
1892         */
1893        tcg_out_movi(s, TCG_TYPE_I32, ret, 0);
1894
1895        tcg_out_brcond2(s, tcg_invert_cond(cond), al, ah, bl, const_bl,
1896                        bh, const_bh, label_over, true);
1897
1898        tgen_arithi(s, ARITH_ADD, ret, 1, 0);
1899    }
1900    tcg_out_label(s, label_over);
1901}
1902
1903#if TCG_TARGET_REG_BITS != 32
1904__attribute__((unused))
1905#endif
1906static const TCGOutOpSetcond2 outop_setcond2 = {
1907    .base.static_constraint = C_O1_I4(r, r, r, ri, ri),
1908    .out = tgen_setcond2,
1909};
1910
1911static void tcg_out_cmov(TCGContext *s, int jcc, int rexw,
1912                         TCGReg dest, TCGReg v1)
1913{
1914    tcg_out_modrm(s, OPC_CMOVCC | jcc | rexw, dest, v1);
1915}
1916
1917static void tgen_movcond(TCGContext *s, TCGType type, TCGCond cond,
1918                         TCGReg dest, TCGReg c1, TCGArg c2, bool const_c2,
1919                         TCGArg vt, bool const_vt,
1920                         TCGArg vf, bool consf_vf)
1921{
1922    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
1923    int jcc = tcg_out_cmp(s, cond, c1, c2, const_c2, rexw);
1924    tcg_out_cmov(s, jcc, rexw, dest, vt);
1925}
1926
1927static const TCGOutOpMovcond outop_movcond = {
1928    .base.static_constraint = C_O1_I4(r, r, reT, r, 0),
1929    .out = tgen_movcond,
1930};
1931
1932static void tcg_out_branch(TCGContext *s, int call, const tcg_insn_unit *dest)
1933{
1934    intptr_t disp = tcg_pcrel_diff(s, dest) - 5;
1935
1936    if (disp == (int32_t)disp) {
1937        tcg_out_opc(s, call ? OPC_CALL_Jz : OPC_JMP_long, 0, 0, 0);
1938        tcg_out32(s, disp);
1939    } else {
1940        /* rip-relative addressing into the constant pool.
1941           This is 6 + 8 = 14 bytes, as compared to using an
1942           immediate load 10 + 6 = 16 bytes, plus we may
1943           be able to re-use the pool constant for more calls.  */
1944        tcg_out_opc(s, OPC_GRP5, 0, 0, 0);
1945        tcg_out8(s, (call ? EXT5_CALLN_Ev : EXT5_JMPN_Ev) << 3 | 5);
1946        new_pool_label(s, (uintptr_t)dest, R_386_PC32, s->code_ptr, -4);
1947        tcg_out32(s, 0);
1948    }
1949}
1950
1951static void tcg_out_call(TCGContext *s, const tcg_insn_unit *dest,
1952                         const TCGHelperInfo *info)
1953{
1954    tcg_out_branch(s, 1, dest);
1955
1956#ifndef _WIN32
1957    if (TCG_TARGET_REG_BITS == 32 && info->out_kind == TCG_CALL_RET_BY_REF) {
1958        /*
1959         * The sysv i386 abi for struct return places a reference as the
1960         * first argument of the stack, and pops that argument with the
1961         * return statement.  Since we want to retain the aligned stack
1962         * pointer for the callee, we do not want to actually push that
1963         * argument before the call but rely on the normal store to the
1964         * stack slot.  But we do need to compensate for the pop in order
1965         * to reset our correct stack pointer value.
1966         * Pushing a garbage value back onto the stack is quickest.
1967         */
1968        tcg_out_push(s, TCG_REG_EAX);
1969    }
1970#endif
1971}
1972
1973static void tcg_out_jmp(TCGContext *s, const tcg_insn_unit *dest)
1974{
1975    tcg_out_branch(s, 0, dest);
1976}
1977
1978static void tcg_out_nopn(TCGContext *s, int n)
1979{
1980    int i;
1981    /* Emit 1 or 2 operand size prefixes for the standard one byte nop,
1982     * "xchg %eax,%eax", forming "xchg %ax,%ax". All cores accept the
1983     * duplicate prefix, and all of the interesting recent cores can
1984     * decode and discard the duplicates in a single cycle.
1985     */
1986    tcg_debug_assert(n >= 1);
1987    for (i = 1; i < n; ++i) {
1988        tcg_out8(s, 0x66);
1989    }
1990    tcg_out8(s, 0x90);
1991}
1992
1993typedef struct {
1994    TCGReg base;
1995    int index;
1996    int ofs;
1997    int seg;
1998    TCGAtomAlign aa;
1999} HostAddress;
2000
2001bool tcg_target_has_memory_bswap(MemOp memop)
2002{
2003    TCGAtomAlign aa;
2004
2005    if (!have_movbe) {
2006        return false;
2007    }
2008    if ((memop & MO_SIZE) < MO_128) {
2009        return true;
2010    }
2011
2012    /*
2013     * Reject 16-byte memop with 16-byte atomicity, i.e. VMOVDQA,
2014     * but do allow a pair of 64-bit operations, i.e. MOVBEQ.
2015     */
2016    aa = atom_and_align_for_opc(tcg_ctx, memop, MO_ATOM_IFALIGN, true);
2017    return aa.atom < MO_128;
2018}
2019
2020/*
2021 * Because i686 has no register parameters and because x86_64 has xchg
2022 * to handle addr/data register overlap, we have placed all input arguments
2023 * before we need might need a scratch reg.
2024 *
2025 * Even then, a scratch is only needed for l->raddr.  Rather than expose
2026 * a general-purpose scratch when we don't actually know it's available,
2027 * use the ra_gen hook to load into RAX if needed.
2028 */
2029#if TCG_TARGET_REG_BITS == 64
2030static TCGReg ldst_ra_gen(TCGContext *s, const TCGLabelQemuLdst *l, int arg)
2031{
2032    if (arg < 0) {
2033        arg = TCG_REG_RAX;
2034    }
2035    tcg_out_movi(s, TCG_TYPE_PTR, arg, (uintptr_t)l->raddr);
2036    return arg;
2037}
2038static const TCGLdstHelperParam ldst_helper_param = {
2039    .ra_gen = ldst_ra_gen
2040};
2041#else
2042static const TCGLdstHelperParam ldst_helper_param = { };
2043#endif
2044
2045static void tcg_out_vec_to_pair(TCGContext *s, TCGType type,
2046                                TCGReg l, TCGReg h, TCGReg v)
2047{
2048    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2049
2050    /* vpmov{d,q} %v, %l */
2051    tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, v, 0, l);
2052    /* vpextr{d,q} $1, %v, %h */
2053    tcg_out_vex_modrm(s, OPC_PEXTRD + rexw, v, 0, h);
2054    tcg_out8(s, 1);
2055}
2056
2057static void tcg_out_pair_to_vec(TCGContext *s, TCGType type,
2058                                TCGReg v, TCGReg l, TCGReg h)
2059{
2060    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2061
2062    /* vmov{d,q} %l, %v */
2063    tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, v, 0, l);
2064    /* vpinsr{d,q} $1, %h, %v, %v */
2065    tcg_out_vex_modrm(s, OPC_PINSRD + rexw, v, v, h);
2066    tcg_out8(s, 1);
2067}
2068
2069/*
2070 * Generate code for the slow path for a load at the end of block
2071 */
2072static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
2073{
2074    MemOp opc = get_memop(l->oi);
2075    tcg_insn_unit **label_ptr = &l->label_ptr[0];
2076
2077    /* resolve label address */
2078    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
2079    if (label_ptr[1]) {
2080        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
2081    }
2082
2083    tcg_out_ld_helper_args(s, l, &ldst_helper_param);
2084    tcg_out_branch(s, 1, qemu_ld_helpers[opc & MO_SIZE]);
2085    tcg_out_ld_helper_ret(s, l, false, &ldst_helper_param);
2086
2087    tcg_out_jmp(s, l->raddr);
2088    return true;
2089}
2090
2091/*
2092 * Generate code for the slow path for a store at the end of block
2093 */
2094static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
2095{
2096    MemOp opc = get_memop(l->oi);
2097    tcg_insn_unit **label_ptr = &l->label_ptr[0];
2098
2099    /* resolve label address */
2100    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
2101    if (label_ptr[1]) {
2102        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
2103    }
2104
2105    tcg_out_st_helper_args(s, l, &ldst_helper_param);
2106    tcg_out_branch(s, 1, qemu_st_helpers[opc & MO_SIZE]);
2107
2108    tcg_out_jmp(s, l->raddr);
2109    return true;
2110}
2111
2112#ifdef CONFIG_USER_ONLY
2113static HostAddress x86_guest_base = {
2114    .index = -1
2115};
2116
2117#if defined(__x86_64__) && defined(__linux__)
2118# include <asm/prctl.h>
2119# include <sys/prctl.h>
2120int arch_prctl(int code, unsigned long addr);
2121static inline int setup_guest_base_seg(void)
2122{
2123    if (arch_prctl(ARCH_SET_GS, guest_base) == 0) {
2124        return P_GS;
2125    }
2126    return 0;
2127}
2128#define setup_guest_base_seg  setup_guest_base_seg
2129#elif defined(__x86_64__) && \
2130      (defined (__FreeBSD__) || defined (__FreeBSD_kernel__))
2131# include <machine/sysarch.h>
2132static inline int setup_guest_base_seg(void)
2133{
2134    if (sysarch(AMD64_SET_GSBASE, &guest_base) == 0) {
2135        return P_GS;
2136    }
2137    return 0;
2138}
2139#define setup_guest_base_seg  setup_guest_base_seg
2140#endif
2141#else
2142# define x86_guest_base (*(HostAddress *)({ qemu_build_not_reached(); NULL; }))
2143#endif /* CONFIG_USER_ONLY */
2144#ifndef setup_guest_base_seg
2145# define setup_guest_base_seg()  0
2146#endif
2147
2148#define MIN_TLB_MASK_TABLE_OFS  INT_MIN
2149
2150/*
2151 * For softmmu, perform the TLB load and compare.
2152 * For useronly, perform any required alignment tests.
2153 * In both cases, return a TCGLabelQemuLdst structure if the slow path
2154 * is required and fill in @h with the host address for the fast path.
2155 */
2156static TCGLabelQemuLdst *prepare_host_addr(TCGContext *s, HostAddress *h,
2157                                           TCGReg addr, MemOpIdx oi, bool is_ld)
2158{
2159    TCGLabelQemuLdst *ldst = NULL;
2160    MemOp opc = get_memop(oi);
2161    MemOp s_bits = opc & MO_SIZE;
2162    unsigned a_mask;
2163
2164    if (tcg_use_softmmu) {
2165        h->index = TCG_REG_L0;
2166        h->ofs = 0;
2167        h->seg = 0;
2168    } else {
2169        *h = x86_guest_base;
2170    }
2171    h->base = addr;
2172    h->aa = atom_and_align_for_opc(s, opc, MO_ATOM_IFALIGN, s_bits == MO_128);
2173    a_mask = (1 << h->aa.align) - 1;
2174
2175    if (tcg_use_softmmu) {
2176        int cmp_ofs = is_ld ? offsetof(CPUTLBEntry, addr_read)
2177                            : offsetof(CPUTLBEntry, addr_write);
2178        TCGType ttype = TCG_TYPE_I32;
2179        TCGType tlbtype = TCG_TYPE_I32;
2180        int trexw = 0, hrexw = 0, tlbrexw = 0;
2181        unsigned mem_index = get_mmuidx(oi);
2182        unsigned s_mask = (1 << s_bits) - 1;
2183        int fast_ofs = tlb_mask_table_ofs(s, mem_index);
2184        int tlb_mask;
2185
2186        ldst = new_ldst_label(s);
2187        ldst->is_ld = is_ld;
2188        ldst->oi = oi;
2189        ldst->addr_reg = addr;
2190
2191        if (TCG_TARGET_REG_BITS == 64) {
2192            ttype = s->addr_type;
2193            trexw = (ttype == TCG_TYPE_I32 ? 0 : P_REXW);
2194            if (TCG_TYPE_PTR == TCG_TYPE_I64) {
2195                hrexw = P_REXW;
2196                if (s->page_bits + s->tlb_dyn_max_bits > 32) {
2197                    tlbtype = TCG_TYPE_I64;
2198                    tlbrexw = P_REXW;
2199                }
2200            }
2201        }
2202
2203        tcg_out_mov(s, tlbtype, TCG_REG_L0, addr);
2204        tcg_out_shifti(s, SHIFT_SHR + tlbrexw, TCG_REG_L0,
2205                       s->page_bits - CPU_TLB_ENTRY_BITS);
2206
2207        tcg_out_modrm_offset(s, OPC_AND_GvEv + trexw, TCG_REG_L0, TCG_AREG0,
2208                             fast_ofs + offsetof(CPUTLBDescFast, mask));
2209
2210        tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, TCG_REG_L0, TCG_AREG0,
2211                             fast_ofs + offsetof(CPUTLBDescFast, table));
2212
2213        /*
2214         * If the required alignment is at least as large as the access,
2215         * simply copy the address and mask.  For lesser alignments,
2216         * check that we don't cross pages for the complete access.
2217         */
2218        if (a_mask >= s_mask) {
2219            tcg_out_mov(s, ttype, TCG_REG_L1, addr);
2220        } else {
2221            tcg_out_modrm_offset(s, OPC_LEA + trexw, TCG_REG_L1,
2222                                 addr, s_mask - a_mask);
2223        }
2224        tlb_mask = s->page_mask | a_mask;
2225        tgen_arithi(s, ARITH_AND + trexw, TCG_REG_L1, tlb_mask, 0);
2226
2227        /* cmp 0(TCG_REG_L0), TCG_REG_L1 */
2228        tcg_out_modrm_offset(s, OPC_CMP_GvEv + trexw,
2229                             TCG_REG_L1, TCG_REG_L0, cmp_ofs);
2230
2231        /* jne slow_path */
2232        tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
2233        ldst->label_ptr[0] = s->code_ptr;
2234        s->code_ptr += 4;
2235
2236        /* TLB Hit.  */
2237        tcg_out_ld(s, TCG_TYPE_PTR, TCG_REG_L0, TCG_REG_L0,
2238                   offsetof(CPUTLBEntry, addend));
2239    } else if (a_mask) {
2240        int jcc;
2241
2242        ldst = new_ldst_label(s);
2243        ldst->is_ld = is_ld;
2244        ldst->oi = oi;
2245        ldst->addr_reg = addr;
2246
2247        /* jne slow_path */
2248        jcc = tcg_out_cmp(s, TCG_COND_TSTNE, addr, a_mask, true, false);
2249        tcg_out_opc(s, OPC_JCC_long + jcc, 0, 0, 0);
2250        ldst->label_ptr[0] = s->code_ptr;
2251        s->code_ptr += 4;
2252    }
2253
2254    return ldst;
2255}
2256
2257static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2258                                   HostAddress h, TCGType type, MemOp memop)
2259{
2260    bool use_movbe = false;
2261    int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
2262    int movop = OPC_MOVL_GvEv;
2263
2264    /* Do big-endian loads with movbe.  */
2265    if (memop & MO_BSWAP) {
2266        tcg_debug_assert(have_movbe);
2267        use_movbe = true;
2268        movop = OPC_MOVBE_GyMy;
2269    }
2270
2271    switch (memop & MO_SSIZE) {
2272    case MO_UB:
2273        tcg_out_modrm_sib_offset(s, OPC_MOVZBL + h.seg, datalo,
2274                                 h.base, h.index, 0, h.ofs);
2275        break;
2276    case MO_SB:
2277        tcg_out_modrm_sib_offset(s, OPC_MOVSBL + rexw + h.seg, datalo,
2278                                 h.base, h.index, 0, h.ofs);
2279        break;
2280    case MO_UW:
2281        if (use_movbe) {
2282            /* There is no extending movbe; only low 16-bits are modified.  */
2283            if (datalo != h.base && datalo != h.index) {
2284                /* XOR breaks dependency chains.  */
2285                tgen_arithr(s, ARITH_XOR, datalo, datalo);
2286                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2287                                         datalo, h.base, h.index, 0, h.ofs);
2288            } else {
2289                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2290                                         datalo, h.base, h.index, 0, h.ofs);
2291                tcg_out_ext16u(s, datalo, datalo);
2292            }
2293        } else {
2294            tcg_out_modrm_sib_offset(s, OPC_MOVZWL + h.seg, datalo,
2295                                     h.base, h.index, 0, h.ofs);
2296        }
2297        break;
2298    case MO_SW:
2299        if (use_movbe) {
2300            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + h.seg,
2301                                     datalo, h.base, h.index, 0, h.ofs);
2302            tcg_out_ext16s(s, type, datalo, datalo);
2303        } else {
2304            tcg_out_modrm_sib_offset(s, OPC_MOVSWL + rexw + h.seg,
2305                                     datalo, h.base, h.index, 0, h.ofs);
2306        }
2307        break;
2308    case MO_UL:
2309        tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2310                                 h.base, h.index, 0, h.ofs);
2311        break;
2312#if TCG_TARGET_REG_BITS == 64
2313    case MO_SL:
2314        if (use_movbe) {
2315            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + h.seg, datalo,
2316                                     h.base, h.index, 0, h.ofs);
2317            tcg_out_ext32s(s, datalo, datalo);
2318        } else {
2319            tcg_out_modrm_sib_offset(s, OPC_MOVSLQ + h.seg, datalo,
2320                                     h.base, h.index, 0, h.ofs);
2321        }
2322        break;
2323#endif
2324    case MO_UQ:
2325        if (TCG_TARGET_REG_BITS == 64) {
2326            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2327                                     h.base, h.index, 0, h.ofs);
2328            break;
2329        }
2330        if (use_movbe) {
2331            TCGReg t = datalo;
2332            datalo = datahi;
2333            datahi = t;
2334        }
2335        if (h.base == datalo || h.index == datalo) {
2336            tcg_out_modrm_sib_offset(s, OPC_LEA, datahi,
2337                                     h.base, h.index, 0, h.ofs);
2338            tcg_out_modrm_offset(s, movop + h.seg, datalo, datahi, 0);
2339            tcg_out_modrm_offset(s, movop + h.seg, datahi, datahi, 4);
2340        } else {
2341            tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2342                                     h.base, h.index, 0, h.ofs);
2343            tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2344                                     h.base, h.index, 0, h.ofs + 4);
2345        }
2346        break;
2347
2348    case MO_128:
2349        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2350
2351        /*
2352         * Without 16-byte atomicity, use integer regs.
2353         * That is where we want the data, and it allows bswaps.
2354         */
2355        if (h.aa.atom < MO_128) {
2356            if (use_movbe) {
2357                TCGReg t = datalo;
2358                datalo = datahi;
2359                datahi = t;
2360            }
2361            if (h.base == datalo || h.index == datalo) {
2362                tcg_out_modrm_sib_offset(s, OPC_LEA + P_REXW, datahi,
2363                                         h.base, h.index, 0, h.ofs);
2364                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2365                                     datalo, datahi, 0);
2366                tcg_out_modrm_offset(s, movop + P_REXW + h.seg,
2367                                     datahi, datahi, 8);
2368            } else {
2369                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2370                                         h.base, h.index, 0, h.ofs);
2371                tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2372                                         h.base, h.index, 0, h.ofs + 8);
2373            }
2374            break;
2375        }
2376
2377        /*
2378         * With 16-byte atomicity, a vector load is required.
2379         * If we already have 16-byte alignment, then VMOVDQA always works.
2380         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2381         * Else use we require a runtime test for alignment for VMOVDQA;
2382         * use VMOVDQU on the unaligned nonatomic path for simplicity.
2383         */
2384        if (h.aa.align >= MO_128) {
2385            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2386                                         TCG_TMP_VEC, 0,
2387                                         h.base, h.index, 0, h.ofs);
2388        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2389            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2390                                         TCG_TMP_VEC, 0,
2391                                         h.base, h.index, 0, h.ofs);
2392        } else {
2393            TCGLabel *l1 = gen_new_label();
2394            TCGLabel *l2 = gen_new_label();
2395            int jcc;
2396
2397            jcc = tcg_out_cmp(s, TCG_COND_TSTNE, h.base, 15, true, false);
2398            tcg_out_jxx(s, jcc, l1, true);
2399
2400            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_VxWx + h.seg,
2401                                         TCG_TMP_VEC, 0,
2402                                         h.base, h.index, 0, h.ofs);
2403            tcg_out_jxx(s, JCC_JMP, l2, true);
2404
2405            tcg_out_label(s, l1);
2406            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_VxWx + h.seg,
2407                                         TCG_TMP_VEC, 0,
2408                                         h.base, h.index, 0, h.ofs);
2409            tcg_out_label(s, l2);
2410        }
2411        tcg_out_vec_to_pair(s, TCG_TYPE_I64, datalo, datahi, TCG_TMP_VEC);
2412        break;
2413
2414    default:
2415        g_assert_not_reached();
2416    }
2417}
2418
2419static void tcg_out_qemu_ld(TCGContext *s, TCGReg datalo, TCGReg datahi,
2420                            TCGReg addr, MemOpIdx oi, TCGType data_type)
2421{
2422    TCGLabelQemuLdst *ldst;
2423    HostAddress h;
2424
2425    ldst = prepare_host_addr(s, &h, addr, oi, true);
2426    tcg_out_qemu_ld_direct(s, datalo, datahi, h, data_type, get_memop(oi));
2427
2428    if (ldst) {
2429        ldst->type = data_type;
2430        ldst->datalo_reg = datalo;
2431        ldst->datahi_reg = datahi;
2432        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2433    }
2434}
2435
2436static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2437                                   HostAddress h, MemOp memop)
2438{
2439    bool use_movbe = false;
2440    int movop = OPC_MOVL_EvGv;
2441
2442    /*
2443     * Do big-endian stores with movbe or system-mode.
2444     * User-only without movbe will have its swapping done generically.
2445     */
2446    if (memop & MO_BSWAP) {
2447        tcg_debug_assert(have_movbe);
2448        use_movbe = true;
2449        movop = OPC_MOVBE_MyGy;
2450    }
2451
2452    switch (memop & MO_SIZE) {
2453    case MO_8:
2454        /* This is handled with constraints on INDEX_op_qemu_st8_i32. */
2455        tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || datalo < 4);
2456        tcg_out_modrm_sib_offset(s, OPC_MOVB_EvGv + P_REXB_R + h.seg,
2457                                 datalo, h.base, h.index, 0, h.ofs);
2458        break;
2459    case MO_16:
2460        tcg_out_modrm_sib_offset(s, movop + P_DATA16 + h.seg, datalo,
2461                                 h.base, h.index, 0, h.ofs);
2462        break;
2463    case MO_32:
2464        tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2465                                 h.base, h.index, 0, h.ofs);
2466        break;
2467    case MO_64:
2468        if (TCG_TARGET_REG_BITS == 64) {
2469            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2470                                     h.base, h.index, 0, h.ofs);
2471        } else {
2472            if (use_movbe) {
2473                TCGReg t = datalo;
2474                datalo = datahi;
2475                datahi = t;
2476            }
2477            tcg_out_modrm_sib_offset(s, movop + h.seg, datalo,
2478                                     h.base, h.index, 0, h.ofs);
2479            tcg_out_modrm_sib_offset(s, movop + h.seg, datahi,
2480                                     h.base, h.index, 0, h.ofs + 4);
2481        }
2482        break;
2483
2484    case MO_128:
2485        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
2486
2487        /*
2488         * Without 16-byte atomicity, use integer regs.
2489         * That is where we have the data, and it allows bswaps.
2490         */
2491        if (h.aa.atom < MO_128) {
2492            if (use_movbe) {
2493                TCGReg t = datalo;
2494                datalo = datahi;
2495                datahi = t;
2496            }
2497            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datalo,
2498                                     h.base, h.index, 0, h.ofs);
2499            tcg_out_modrm_sib_offset(s, movop + P_REXW + h.seg, datahi,
2500                                     h.base, h.index, 0, h.ofs + 8);
2501            break;
2502        }
2503
2504        /*
2505         * With 16-byte atomicity, a vector store is required.
2506         * If we already have 16-byte alignment, then VMOVDQA always works.
2507         * Else if VMOVDQU has atomicity with dynamic alignment, use that.
2508         * Else use we require a runtime test for alignment for VMOVDQA;
2509         * use VMOVDQU on the unaligned nonatomic path for simplicity.
2510         */
2511        tcg_out_pair_to_vec(s, TCG_TYPE_I64, TCG_TMP_VEC, datalo, datahi);
2512        if (h.aa.align >= MO_128) {
2513            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2514                                         TCG_TMP_VEC, 0,
2515                                         h.base, h.index, 0, h.ofs);
2516        } else if (cpuinfo & CPUINFO_ATOMIC_VMOVDQU) {
2517            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2518                                         TCG_TMP_VEC, 0,
2519                                         h.base, h.index, 0, h.ofs);
2520        } else {
2521            TCGLabel *l1 = gen_new_label();
2522            TCGLabel *l2 = gen_new_label();
2523            int jcc;
2524
2525            jcc = tcg_out_cmp(s, TCG_COND_TSTNE, h.base, 15, true, false);
2526            tcg_out_jxx(s, jcc, l1, true);
2527
2528            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQA_WxVx + h.seg,
2529                                         TCG_TMP_VEC, 0,
2530                                         h.base, h.index, 0, h.ofs);
2531            tcg_out_jxx(s, JCC_JMP, l2, true);
2532
2533            tcg_out_label(s, l1);
2534            tcg_out_vex_modrm_sib_offset(s, OPC_MOVDQU_WxVx + h.seg,
2535                                         TCG_TMP_VEC, 0,
2536                                         h.base, h.index, 0, h.ofs);
2537            tcg_out_label(s, l2);
2538        }
2539        break;
2540
2541    default:
2542        g_assert_not_reached();
2543    }
2544}
2545
2546static void tcg_out_qemu_st(TCGContext *s, TCGReg datalo, TCGReg datahi,
2547                            TCGReg addr, MemOpIdx oi, TCGType data_type)
2548{
2549    TCGLabelQemuLdst *ldst;
2550    HostAddress h;
2551
2552    ldst = prepare_host_addr(s, &h, addr, oi, false);
2553    tcg_out_qemu_st_direct(s, datalo, datahi, h, get_memop(oi));
2554
2555    if (ldst) {
2556        ldst->type = data_type;
2557        ldst->datalo_reg = datalo;
2558        ldst->datahi_reg = datahi;
2559        ldst->raddr = tcg_splitwx_to_rx(s->code_ptr);
2560    }
2561}
2562
2563static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
2564{
2565    /* Reuse the zeroing that exists for goto_ptr.  */
2566    if (a0 == 0) {
2567        tcg_out_jmp(s, tcg_code_gen_epilogue);
2568    } else {
2569        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_EAX, a0);
2570        tcg_out_jmp(s, tb_ret_addr);
2571    }
2572}
2573
2574static void tcg_out_goto_tb(TCGContext *s, int which)
2575{
2576    /*
2577     * Jump displacement must be aligned for atomic patching;
2578     * see if we need to add extra nops before jump
2579     */
2580    int gap = QEMU_ALIGN_PTR_UP(s->code_ptr + 1, 4) - s->code_ptr;
2581    if (gap != 1) {
2582        tcg_out_nopn(s, gap - 1);
2583    }
2584    tcg_out8(s, OPC_JMP_long); /* jmp im */
2585    set_jmp_insn_offset(s, which);
2586    tcg_out32(s, 0);
2587    set_jmp_reset_offset(s, which);
2588}
2589
2590void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
2591                              uintptr_t jmp_rx, uintptr_t jmp_rw)
2592{
2593    /* patch the branch destination */
2594    uintptr_t addr = tb->jmp_target_addr[n];
2595    qatomic_set((int32_t *)jmp_rw, addr - (jmp_rx + 4));
2596    /* no need to flush icache explicitly */
2597}
2598
2599
2600static void tgen_add(TCGContext *s, TCGType type,
2601                     TCGReg a0, TCGReg a1, TCGReg a2)
2602{
2603    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2604
2605    if (a0 == a1) {
2606        tgen_arithr(s, ARITH_ADD + rexw, a0, a2);
2607    } else if (a0 == a2) {
2608        tgen_arithr(s, ARITH_ADD + rexw, a0, a1);
2609    } else {
2610        tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a2, 0, 0);
2611    }
2612}
2613
2614static void tgen_addi(TCGContext *s, TCGType type,
2615                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2616{
2617    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2618
2619    if (a0 == a1) {
2620        tgen_arithi(s, ARITH_ADD + rexw, a0, a2, false);
2621    } else {
2622        tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, -1, 0, a2);
2623    }
2624}
2625
2626static const TCGOutOpBinary outop_add = {
2627    .base.static_constraint = C_O1_I2(r, r, re),
2628    .out_rrr = tgen_add,
2629    .out_rri = tgen_addi,
2630};
2631
2632static void tgen_and(TCGContext *s, TCGType type,
2633                     TCGReg a0, TCGReg a1, TCGReg a2)
2634{
2635    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2636    tgen_arithr(s, ARITH_AND + rexw, a0, a2);
2637}
2638
2639static void tgen_andi(TCGContext *s, TCGType type,
2640                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2641{
2642    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2643    tgen_arithi(s, ARITH_AND + rexw, a0, a2, false);
2644}
2645
2646static const TCGOutOpBinary outop_and = {
2647    .base.static_constraint = C_O1_I2(r, 0, reZ),
2648    .out_rrr = tgen_and,
2649    .out_rri = tgen_andi,
2650};
2651
2652static void tgen_andc(TCGContext *s, TCGType type,
2653                      TCGReg a0, TCGReg a1, TCGReg a2)
2654{
2655    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2656    tcg_out_vex_modrm(s, OPC_ANDN + rexw, a0, a2, a1);
2657}
2658
2659static TCGConstraintSetIndex cset_andc(TCGType type, unsigned flags)
2660{
2661    return have_bmi1 ? C_O1_I2(r, r, r) : C_NotImplemented;
2662}
2663
2664static const TCGOutOpBinary outop_andc = {
2665    .base.static_constraint = C_Dynamic,
2666    .base.dynamic_constraint = cset_andc,
2667    .out_rrr = tgen_andc,
2668};
2669
2670static void tgen_clz(TCGContext *s, TCGType type,
2671                     TCGReg a0, TCGReg a1, TCGReg a2)
2672{
2673    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2674    int jcc;
2675
2676    if (have_lzcnt) {
2677        tcg_out_modrm(s, OPC_LZCNT + rexw, a0, a1);
2678        jcc = JCC_JB;
2679    } else {
2680        /* Recall that the output of BSR is the index not the count.  */
2681        tcg_out_modrm(s, OPC_BSR + rexw, a0, a1);
2682        tgen_arithi(s, ARITH_XOR + rexw, a0, rexw ? 63 : 31, 0);
2683
2684        /* Since we have destroyed the flags from BSR, we have to re-test.  */
2685        jcc = tcg_out_cmp(s, TCG_COND_EQ, a1, 0, 1, rexw);
2686    }
2687    tcg_out_cmov(s, jcc, rexw, a0, a2);
2688}
2689
2690static void tgen_clzi(TCGContext *s, TCGType type,
2691                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2692{
2693    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2694    tcg_out_modrm(s, OPC_LZCNT + rexw, a0, a1);
2695}
2696
2697static TCGConstraintSetIndex cset_clz(TCGType type, unsigned flags)
2698{
2699    return have_lzcnt ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
2700}
2701
2702static const TCGOutOpBinary outop_clz = {
2703    .base.static_constraint = C_Dynamic,
2704    .base.dynamic_constraint = cset_clz,
2705    .out_rrr = tgen_clz,
2706    .out_rri = tgen_clzi,
2707};
2708
2709static void tgen_ctpop(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
2710{
2711    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2712    tcg_out_modrm(s, OPC_POPCNT + rexw, a0, a1);
2713}
2714
2715static TCGConstraintSetIndex cset_ctpop(TCGType type, unsigned flags)
2716{
2717    return have_popcnt ? C_O1_I1(r, r) : C_NotImplemented;
2718}
2719
2720static const TCGOutOpUnary outop_ctpop = {
2721    .base.static_constraint = C_Dynamic,
2722    .base.dynamic_constraint = cset_ctpop,
2723    .out_rr = tgen_ctpop,
2724};
2725
2726static void tgen_ctz(TCGContext *s, TCGType type,
2727                     TCGReg a0, TCGReg a1, TCGReg a2)
2728{
2729    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2730    int jcc;
2731
2732    if (have_bmi1) {
2733        tcg_out_modrm(s, OPC_TZCNT + rexw, a0, a1);
2734        jcc = JCC_JB;
2735    } else {
2736        tcg_out_modrm(s, OPC_BSF + rexw, a0, a1);
2737        jcc = JCC_JE;
2738    }
2739    tcg_out_cmov(s, jcc, rexw, a0, a2);
2740}
2741
2742static void tgen_ctzi(TCGContext *s, TCGType type,
2743                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2744{
2745    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2746    tcg_out_modrm(s, OPC_TZCNT + rexw, a0, a1);
2747}
2748
2749static TCGConstraintSetIndex cset_ctz(TCGType type, unsigned flags)
2750{
2751    return have_bmi1 ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
2752}
2753
2754static const TCGOutOpBinary outop_ctz = {
2755    .base.static_constraint = C_Dynamic,
2756    .base.dynamic_constraint = cset_ctz,
2757    .out_rrr = tgen_ctz,
2758    .out_rri = tgen_ctzi,
2759};
2760
2761static const TCGOutOpBinary outop_divs = {
2762    .base.static_constraint = C_NotImplemented,
2763};
2764
2765static void tgen_divs2(TCGContext *s, TCGType type,
2766                       TCGReg a0, TCGReg a1, TCGReg a4)
2767{
2768    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2769    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IDIV, a4);
2770}
2771
2772static const TCGOutOpDivRem outop_divs2 = {
2773    .base.static_constraint = C_O2_I3(a, d, 0, 1, r),
2774    .out_rr01r = tgen_divs2,
2775};
2776
2777static const TCGOutOpBinary outop_divu = {
2778    .base.static_constraint = C_NotImplemented,
2779};
2780
2781static void tgen_divu2(TCGContext *s, TCGType type,
2782                       TCGReg a0, TCGReg a1, TCGReg a4)
2783{
2784    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2785    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_DIV, a4);
2786}
2787
2788static const TCGOutOpDivRem outop_divu2 = {
2789    .base.static_constraint = C_O2_I3(a, d, 0, 1, r),
2790    .out_rr01r = tgen_divu2,
2791};
2792
2793static const TCGOutOpBinary outop_eqv = {
2794    .base.static_constraint = C_NotImplemented,
2795};
2796
2797static void tgen_mul(TCGContext *s, TCGType type,
2798                     TCGReg a0, TCGReg a1, TCGReg a2)
2799{
2800    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2801    tcg_out_modrm(s, OPC_IMUL_GvEv + rexw, a0, a2);
2802}
2803
2804static void tgen_muli(TCGContext *s, TCGType type,
2805                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2806{
2807    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2808
2809    if (a2 == (int8_t)a2) {
2810        tcg_out_modrm(s, OPC_IMUL_GvEvIb + rexw, a0, a0);
2811        tcg_out8(s, a2);
2812    } else {
2813        tcg_out_modrm(s, OPC_IMUL_GvEvIz + rexw, a0, a0);
2814        tcg_out32(s, a2);
2815    }
2816}
2817
2818static const TCGOutOpBinary outop_mul = {
2819    .base.static_constraint = C_O1_I2(r, 0, re),
2820    .out_rrr = tgen_mul,
2821    .out_rri = tgen_muli,
2822};
2823
2824static void tgen_muls2(TCGContext *s, TCGType type,
2825                       TCGReg a0, TCGReg a1, TCGReg a2, TCGReg a3)
2826{
2827    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2828    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IMUL, a3);
2829}
2830
2831static const TCGOutOpMul2 outop_muls2 = {
2832    .base.static_constraint = C_O2_I2(a, d, a, r),
2833    .out_rrrr = tgen_muls2,
2834};
2835
2836static const TCGOutOpBinary outop_mulsh = {
2837    .base.static_constraint = C_NotImplemented,
2838};
2839
2840static const TCGOutOpBinary outop_muluh = {
2841    .base.static_constraint = C_NotImplemented,
2842};
2843
2844static void tgen_mulu2(TCGContext *s, TCGType type,
2845                       TCGReg a0, TCGReg a1, TCGReg a2, TCGReg a3)
2846{
2847    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2848    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_MUL, a3);
2849}
2850
2851static const TCGOutOpMul2 outop_mulu2 = {
2852    .base.static_constraint = C_O2_I2(a, d, a, r),
2853    .out_rrrr = tgen_mulu2,
2854};
2855
2856static const TCGOutOpBinary outop_nand = {
2857    .base.static_constraint = C_NotImplemented,
2858};
2859
2860static const TCGOutOpBinary outop_nor = {
2861    .base.static_constraint = C_NotImplemented,
2862};
2863
2864static void tgen_or(TCGContext *s, TCGType type,
2865                    TCGReg a0, TCGReg a1, TCGReg a2)
2866{
2867    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2868    tgen_arithr(s, ARITH_OR + rexw, a0, a2);
2869}
2870
2871static void tgen_ori(TCGContext *s, TCGType type,
2872                     TCGReg a0, TCGReg a1, tcg_target_long a2)
2873{
2874    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2875    tgen_arithi(s, ARITH_OR + rexw, a0, a2, false);
2876}
2877
2878static const TCGOutOpBinary outop_or = {
2879    .base.static_constraint = C_O1_I2(r, 0, re),
2880    .out_rrr = tgen_or,
2881    .out_rri = tgen_ori,
2882};
2883
2884static const TCGOutOpBinary outop_orc = {
2885    .base.static_constraint = C_NotImplemented,
2886};
2887
2888static const TCGOutOpBinary outop_rems = {
2889    .base.static_constraint = C_NotImplemented,
2890};
2891
2892static const TCGOutOpBinary outop_remu = {
2893    .base.static_constraint = C_NotImplemented,
2894};
2895
2896static void tgen_rotl(TCGContext *s, TCGType type,
2897                     TCGReg a0, TCGReg a1, TCGReg a2)
2898{
2899    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2900    tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_ROL, a0);
2901}
2902
2903static void tgen_rotli(TCGContext *s, TCGType type,
2904                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2905{
2906    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2907    tcg_out_shifti(s, SHIFT_ROL + rexw, a0, a2);
2908}
2909
2910static const TCGOutOpBinary outop_rotl = {
2911    .base.static_constraint = C_O1_I2(r, 0, ci),
2912    .out_rrr = tgen_rotl,
2913    .out_rri = tgen_rotli,
2914};
2915
2916static void tgen_rotr(TCGContext *s, TCGType type,
2917                     TCGReg a0, TCGReg a1, TCGReg a2)
2918{
2919    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2920    tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_ROR, a0);
2921}
2922
2923static void tgen_rotri(TCGContext *s, TCGType type,
2924                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2925{
2926    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2927    tcg_out_shifti(s, SHIFT_ROR + rexw, a0, a2);
2928}
2929
2930static const TCGOutOpBinary outop_rotr = {
2931    .base.static_constraint = C_O1_I2(r, 0, ci),
2932    .out_rrr = tgen_rotr,
2933    .out_rri = tgen_rotri,
2934};
2935
2936static TCGConstraintSetIndex cset_shift(TCGType type, unsigned flags)
2937{
2938    return have_bmi2 ? C_O1_I2(r, r, ri) : C_O1_I2(r, 0, ci);
2939}
2940
2941static void tgen_sar(TCGContext *s, TCGType type,
2942                     TCGReg a0, TCGReg a1, TCGReg a2)
2943{
2944    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2945    if (have_bmi2) {
2946        tcg_out_vex_modrm(s, OPC_SARX + rexw, a0, a2, a1);
2947    } else {
2948        tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_SAR, a0);
2949    }
2950}
2951
2952static void tgen_sari(TCGContext *s, TCGType type,
2953                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2954{
2955    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2956
2957    tcg_out_mov(s, type, a0, a1);
2958    tcg_out_shifti(s, SHIFT_SAR + rexw, a0, a2);
2959}
2960
2961static const TCGOutOpBinary outop_sar = {
2962    .base.static_constraint = C_Dynamic,
2963    .base.dynamic_constraint = cset_shift,
2964    .out_rrr = tgen_sar,
2965    .out_rri = tgen_sari,
2966};
2967
2968static void tgen_shl(TCGContext *s, TCGType type,
2969                     TCGReg a0, TCGReg a1, TCGReg a2)
2970{
2971    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2972    if (have_bmi2) {
2973        tcg_out_vex_modrm(s, OPC_SHLX + rexw, a0, a2, a1);
2974    } else {
2975        tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_SHL, a0);
2976    }
2977}
2978
2979static void tgen_shli(TCGContext *s, TCGType type,
2980                      TCGReg a0, TCGReg a1, tcg_target_long a2)
2981{
2982    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
2983
2984    /* For small constant 3-operand shift, use LEA.  */
2985    if (a0 != a1 && a2 >= 1 && a2 <= 3) {
2986        if (a2 == 1) {
2987            /* shl $1,a1,a0 -> lea (a1,a1),a0 */
2988            tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a1, 0, 0);
2989        } else {
2990            /* shl $n,a1,a0 -> lea 0(,a1,n),a0 */
2991            tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, -1, a1, a2, 0);
2992        }
2993        return;
2994    }
2995    tcg_out_mov(s, type, a0, a1);
2996    tcg_out_shifti(s, SHIFT_SHL + rexw, a0, a2);
2997}
2998
2999static const TCGOutOpBinary outop_shl = {
3000    .base.static_constraint = C_Dynamic,
3001    .base.dynamic_constraint = cset_shift,
3002    .out_rrr = tgen_shl,
3003    .out_rri = tgen_shli,
3004};
3005
3006static void tgen_shr(TCGContext *s, TCGType type,
3007                     TCGReg a0, TCGReg a1, TCGReg a2)
3008{
3009    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3010    if (have_bmi2) {
3011        tcg_out_vex_modrm(s, OPC_SHRX + rexw, a0, a2, a1);
3012    } else {
3013        tcg_out_modrm(s, OPC_SHIFT_cl + rexw, SHIFT_SHR, a0);
3014    }
3015}
3016
3017static void tgen_shri(TCGContext *s, TCGType type,
3018                      TCGReg a0, TCGReg a1, tcg_target_long a2)
3019{
3020    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3021
3022    tcg_out_mov(s, type, a0, a1);
3023    tcg_out_shifti(s, SHIFT_SHR + rexw, a0, a2);
3024}
3025
3026static const TCGOutOpBinary outop_shr = {
3027    .base.static_constraint = C_Dynamic,
3028    .base.dynamic_constraint = cset_shift,
3029    .out_rrr = tgen_shr,
3030    .out_rri = tgen_shri,
3031};
3032
3033static void tgen_sub(TCGContext *s, TCGType type,
3034                      TCGReg a0, TCGReg a1, TCGReg a2)
3035{
3036    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3037    tgen_arithr(s, ARITH_SUB + rexw, a0, a2);
3038}
3039
3040static const TCGOutOpSubtract outop_sub = {
3041    .base.static_constraint = C_O1_I2(r, 0, r),
3042    .out_rrr = tgen_sub,
3043};
3044
3045static void tgen_xor(TCGContext *s, TCGType type,
3046                     TCGReg a0, TCGReg a1, TCGReg a2)
3047{
3048    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3049    tgen_arithr(s, ARITH_XOR + rexw, a0, a2);
3050}
3051
3052static void tgen_xori(TCGContext *s, TCGType type,
3053                      TCGReg a0, TCGReg a1, tcg_target_long a2)
3054{
3055    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3056    tgen_arithi(s, ARITH_XOR + rexw, a0, a2, false);
3057}
3058
3059static const TCGOutOpBinary outop_xor = {
3060    .base.static_constraint = C_O1_I2(r, 0, re),
3061    .out_rrr = tgen_xor,
3062    .out_rri = tgen_xori,
3063};
3064
3065static void tgen_bswap16(TCGContext *s, TCGType type,
3066                         TCGReg a0, TCGReg a1, unsigned flags)
3067{
3068    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3069
3070    if (flags & TCG_BSWAP_OS) {
3071        /* Output must be sign-extended. */
3072        if (rexw) {
3073            tcg_out_bswap64(s, a0);
3074            tcg_out_shifti(s, SHIFT_SAR + rexw, a0, 48);
3075        } else {
3076            tcg_out_bswap32(s, a0);
3077            tcg_out_shifti(s, SHIFT_SAR, a0, 16);
3078        }
3079    } else if ((flags & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
3080        /* Output must be zero-extended, but input isn't. */
3081        tcg_out_bswap32(s, a0);
3082        tcg_out_shifti(s, SHIFT_SHR, a0, 16);
3083    } else {
3084        tcg_out_rolw_8(s, a0);
3085    }
3086}
3087
3088static const TCGOutOpBswap outop_bswap16 = {
3089    .base.static_constraint = C_O1_I1(r, 0),
3090    .out_rr = tgen_bswap16,
3091};
3092
3093static void tgen_bswap32(TCGContext *s, TCGType type,
3094                         TCGReg a0, TCGReg a1, unsigned flags)
3095{
3096    tcg_out_bswap32(s, a0);
3097    if (flags & TCG_BSWAP_OS) {
3098        tcg_out_ext32s(s, a0, a0);
3099    }
3100}
3101
3102static const TCGOutOpBswap outop_bswap32 = {
3103    .base.static_constraint = C_O1_I1(r, 0),
3104    .out_rr = tgen_bswap32,
3105};
3106
3107#if TCG_TARGET_REG_BITS == 64
3108static void tgen_bswap64(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
3109{
3110    tcg_out_bswap64(s, a0);
3111}
3112
3113static const TCGOutOpUnary outop_bswap64 = {
3114    .base.static_constraint = C_O1_I1(r, 0),
3115    .out_rr = tgen_bswap64,
3116};
3117#endif
3118
3119static void tgen_neg(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
3120{
3121    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3122    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, a0);
3123}
3124
3125static const TCGOutOpUnary outop_neg = {
3126    .base.static_constraint = C_O1_I1(r, 0),
3127    .out_rr = tgen_neg,
3128};
3129
3130static void tgen_not(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1)
3131{
3132    int rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3133    tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, a0);
3134}
3135
3136static const TCGOutOpUnary outop_not = {
3137    .base.static_constraint = C_O1_I1(r, 0),
3138    .out_rr = tgen_not,
3139};
3140
3141static void tgen_extract(TCGContext *s, TCGType type, TCGReg a0, TCGReg a1,
3142                         unsigned ofs, unsigned len)
3143{
3144    if (ofs == 0) {
3145        switch (len) {
3146        case 8:
3147            tcg_out_ext8u(s, a0, a1);
3148            return;
3149        case 16:
3150            tcg_out_ext16u(s, a0, a1);
3151            return;
3152        case 32:
3153            tcg_out_ext32u(s, a0, a1);
3154            return;
3155        }
3156    } else if (TCG_TARGET_REG_BITS == 64 && ofs + len == 32) {
3157        /* This is a 32-bit zero-extending right shift.  */
3158        tcg_out_mov(s, TCG_TYPE_I32, a0, a1);
3159        tcg_out_shifti(s, SHIFT_SHR, a0, ofs);
3160        return;
3161    } else if (ofs == 8 && len == 8) {
3162        /*
3163         * On the off-chance that we can use the high-byte registers.
3164         * Otherwise we emit the same ext16 + shift pattern that we
3165         * would have gotten from the normal tcg-op.c expansion.
3166         */
3167        if (a1 < 4 && (TCG_TARGET_REG_BITS == 32 || a0 < 8)) {
3168            tcg_out_modrm(s, OPC_MOVZBL, a0, a1 + 4);
3169        } else {
3170            tcg_out_ext16u(s, a0, a1);
3171            tcg_out_shifti(s, SHIFT_SHR, a0, 8);
3172        }
3173        return;
3174    }
3175    g_assert_not_reached();
3176}
3177
3178static const TCGOutOpExtract outop_extract = {
3179    .base.static_constraint = C_O1_I1(r, r),
3180    .out_rr = tgen_extract,
3181};
3182
3183static void tcg_out_op(TCGContext *s, TCGOpcode opc, TCGType type,
3184                       const TCGArg args[TCG_MAX_OP_ARGS],
3185                       const int const_args[TCG_MAX_OP_ARGS])
3186{
3187    TCGArg a0, a1, a2;
3188    int const_a2, rexw;
3189
3190#if TCG_TARGET_REG_BITS == 64
3191# define OP_32_64(x) \
3192        case glue(glue(INDEX_op_, x), _i64): \
3193        case glue(glue(INDEX_op_, x), _i32)
3194#else
3195# define OP_32_64(x) \
3196        case glue(glue(INDEX_op_, x), _i32)
3197#endif
3198
3199    /* Hoist the loads of the most common arguments.  */
3200    a0 = args[0];
3201    a1 = args[1];
3202    a2 = args[2];
3203    const_a2 = const_args[2];
3204    rexw = type == TCG_TYPE_I32 ? 0 : P_REXW;
3205
3206    switch (opc) {
3207    case INDEX_op_goto_ptr:
3208        /* jmp to the given host address (could be epilogue) */
3209        tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, a0);
3210        break;
3211    case INDEX_op_br:
3212        tcg_out_jxx(s, JCC_JMP, arg_label(a0), 0);
3213        break;
3214    OP_32_64(ld8u):
3215        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
3216        tcg_out_modrm_offset(s, OPC_MOVZBL, a0, a1, a2);
3217        break;
3218    OP_32_64(ld8s):
3219        tcg_out_modrm_offset(s, OPC_MOVSBL + rexw, a0, a1, a2);
3220        break;
3221    OP_32_64(ld16u):
3222        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
3223        tcg_out_modrm_offset(s, OPC_MOVZWL, a0, a1, a2);
3224        break;
3225    OP_32_64(ld16s):
3226        tcg_out_modrm_offset(s, OPC_MOVSWL + rexw, a0, a1, a2);
3227        break;
3228#if TCG_TARGET_REG_BITS == 64
3229    case INDEX_op_ld32u_i64:
3230#endif
3231    case INDEX_op_ld_i32:
3232        tcg_out_ld(s, TCG_TYPE_I32, a0, a1, a2);
3233        break;
3234
3235    OP_32_64(st8):
3236        if (const_args[0]) {
3237            tcg_out_modrm_offset(s, OPC_MOVB_EvIz, 0, a1, a2);
3238            tcg_out8(s, a0);
3239        } else {
3240            tcg_out_modrm_offset(s, OPC_MOVB_EvGv | P_REXB_R, a0, a1, a2);
3241        }
3242        break;
3243    OP_32_64(st16):
3244        if (const_args[0]) {
3245            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_DATA16, 0, a1, a2);
3246            tcg_out16(s, a0);
3247        } else {
3248            tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_DATA16, a0, a1, a2);
3249        }
3250        break;
3251#if TCG_TARGET_REG_BITS == 64
3252    case INDEX_op_st32_i64:
3253#endif
3254    case INDEX_op_st_i32:
3255        if (const_args[0]) {
3256            tcg_out_modrm_offset(s, OPC_MOVL_EvIz, 0, a1, a2);
3257            tcg_out32(s, a0);
3258        } else {
3259            tcg_out_st(s, TCG_TYPE_I32, a0, a1, a2);
3260        }
3261        break;
3262
3263    case INDEX_op_qemu_ld_i32:
3264        tcg_out_qemu_ld(s, a0, -1, a1, a2, TCG_TYPE_I32);
3265        break;
3266    case INDEX_op_qemu_ld_i64:
3267        if (TCG_TARGET_REG_BITS == 64) {
3268            tcg_out_qemu_ld(s, a0, -1, a1, a2, TCG_TYPE_I64);
3269        } else {
3270            tcg_out_qemu_ld(s, a0, a1, a2, args[3], TCG_TYPE_I64);
3271        }
3272        break;
3273    case INDEX_op_qemu_ld_i128:
3274        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3275        tcg_out_qemu_ld(s, a0, a1, a2, args[3], TCG_TYPE_I128);
3276        break;
3277
3278    case INDEX_op_qemu_st_i32:
3279    case INDEX_op_qemu_st8_i32:
3280        tcg_out_qemu_st(s, a0, -1, a1, a2, TCG_TYPE_I32);
3281        break;
3282    case INDEX_op_qemu_st_i64:
3283        if (TCG_TARGET_REG_BITS == 64) {
3284            tcg_out_qemu_st(s, a0, -1, a1, a2, TCG_TYPE_I64);
3285        } else {
3286            tcg_out_qemu_st(s, a0, a1, a2, args[3], TCG_TYPE_I64);
3287        }
3288        break;
3289    case INDEX_op_qemu_st_i128:
3290        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
3291        tcg_out_qemu_st(s, a0, a1, a2, args[3], TCG_TYPE_I128);
3292        break;
3293
3294    OP_32_64(add2):
3295        if (const_args[4]) {
3296            tgen_arithi(s, ARITH_ADD + rexw, a0, args[4], 1);
3297        } else {
3298            tgen_arithr(s, ARITH_ADD + rexw, a0, args[4]);
3299        }
3300        if (const_args[5]) {
3301            tgen_arithi(s, ARITH_ADC + rexw, a1, args[5], 1);
3302        } else {
3303            tgen_arithr(s, ARITH_ADC + rexw, a1, args[5]);
3304        }
3305        break;
3306    OP_32_64(sub2):
3307        if (const_args[4]) {
3308            tgen_arithi(s, ARITH_SUB + rexw, a0, args[4], 1);
3309        } else {
3310            tgen_arithr(s, ARITH_SUB + rexw, a0, args[4]);
3311        }
3312        if (const_args[5]) {
3313            tgen_arithi(s, ARITH_SBB + rexw, a1, args[5], 1);
3314        } else {
3315            tgen_arithr(s, ARITH_SBB + rexw, a1, args[5]);
3316        }
3317        break;
3318
3319#if TCG_TARGET_REG_BITS == 64
3320    case INDEX_op_ld32s_i64:
3321        tcg_out_modrm_offset(s, OPC_MOVSLQ, a0, a1, a2);
3322        break;
3323    case INDEX_op_ld_i64:
3324        tcg_out_ld(s, TCG_TYPE_I64, a0, a1, a2);
3325        break;
3326    case INDEX_op_st_i64:
3327        if (const_args[0]) {
3328            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_REXW, 0, a1, a2);
3329            tcg_out32(s, a0);
3330        } else {
3331            tcg_out_st(s, TCG_TYPE_I64, a0, a1, a2);
3332        }
3333        break;
3334
3335    case INDEX_op_extrh_i64_i32:
3336        tcg_out_shifti(s, SHIFT_SHR + P_REXW, a0, 32);
3337        break;
3338#endif
3339
3340    OP_32_64(deposit):
3341        if (args[3] == 0 && args[4] == 8) {
3342            /* load bits 0..7 */
3343            if (const_a2) {
3344                tcg_out_opc(s, OPC_MOVB_Ib | P_REXB_RM | LOWREGMASK(a0),
3345                            0, a0, 0);
3346                tcg_out8(s, a2);
3347            } else {
3348                tcg_out_modrm(s, OPC_MOVB_EvGv | P_REXB_R | P_REXB_RM, a2, a0);
3349            }
3350        } else if (TCG_TARGET_REG_BITS == 32 && args[3] == 8 && args[4] == 8) {
3351            /* load bits 8..15 */
3352            if (const_a2) {
3353                tcg_out8(s, OPC_MOVB_Ib + a0 + 4);
3354                tcg_out8(s, a2);
3355            } else {
3356                tcg_out_modrm(s, OPC_MOVB_EvGv, a2, a0 + 4);
3357            }
3358        } else if (args[3] == 0 && args[4] == 16) {
3359            /* load bits 0..15 */
3360            if (const_a2) {
3361                tcg_out_opc(s, OPC_MOVL_Iv | P_DATA16 | LOWREGMASK(a0),
3362                            0, a0, 0);
3363                tcg_out16(s, a2);
3364            } else {
3365                tcg_out_modrm(s, OPC_MOVL_EvGv | P_DATA16, a2, a0);
3366            }
3367        } else {
3368            g_assert_not_reached();
3369        }
3370        break;
3371
3372    case INDEX_op_sextract_i64:
3373        if (a2 == 0 && args[3] == 8) {
3374            tcg_out_ext8s(s, TCG_TYPE_I64, a0, a1);
3375        } else if (a2 == 0 && args[3] == 16) {
3376            tcg_out_ext16s(s, TCG_TYPE_I64, a0, a1);
3377        } else if (a2 == 0 && args[3] == 32) {
3378            tcg_out_ext32s(s, a0, a1);
3379        } else {
3380            g_assert_not_reached();
3381        }
3382        break;
3383
3384    case INDEX_op_sextract_i32:
3385        if (a2 == 0 && args[3] == 8) {
3386            tcg_out_ext8s(s, TCG_TYPE_I32, a0, a1);
3387        } else if (a2 == 0 && args[3] == 16) {
3388            tcg_out_ext16s(s, TCG_TYPE_I32, a0, a1);
3389        } else if (a2 == 8 && args[3] == 8) {
3390            if (a1 < 4 && a0 < 8) {
3391                tcg_out_modrm(s, OPC_MOVSBL, a0, a1 + 4);
3392            } else {
3393                tcg_out_ext16s(s, TCG_TYPE_I32, a0, a1);
3394                tcg_out_shifti(s, SHIFT_SAR, a0, 8);
3395            }
3396        } else {
3397            g_assert_not_reached();
3398        }
3399        break;
3400
3401    OP_32_64(extract2):
3402        /* Note that SHRD outputs to the r/m operand.  */
3403        tcg_out_modrm(s, OPC_SHRD_Ib + rexw, a2, a0);
3404        tcg_out8(s, args[3]);
3405        break;
3406
3407    case INDEX_op_mb:
3408        tcg_out_mb(s, a0);
3409        break;
3410    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
3411    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
3412    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
3413    case INDEX_op_ext_i32_i64:  /* Always emitted via tcg_reg_alloc_op.  */
3414    case INDEX_op_extu_i32_i64:
3415    case INDEX_op_extrl_i64_i32:
3416    default:
3417        g_assert_not_reached();
3418    }
3419
3420#undef OP_32_64
3421}
3422
3423static int const umin_insn[4] = {
3424    OPC_PMINUB, OPC_PMINUW, OPC_PMINUD, OPC_VPMINUQ
3425};
3426
3427static int const umax_insn[4] = {
3428    OPC_PMAXUB, OPC_PMAXUW, OPC_PMAXUD, OPC_VPMAXUQ
3429};
3430
3431static bool tcg_out_cmp_vec_noinv(TCGContext *s, TCGType type, unsigned vece,
3432                                  TCGReg v0, TCGReg v1, TCGReg v2, TCGCond cond)
3433{
3434    static int const cmpeq_insn[4] = {
3435        OPC_PCMPEQB, OPC_PCMPEQW, OPC_PCMPEQD, OPC_PCMPEQQ
3436    };
3437    static int const cmpgt_insn[4] = {
3438        OPC_PCMPGTB, OPC_PCMPGTW, OPC_PCMPGTD, OPC_PCMPGTQ
3439    };
3440
3441    enum {
3442        NEED_INV  = 1,
3443        NEED_SWAP = 2,
3444        NEED_UMIN = 4,
3445        NEED_UMAX = 8,
3446        INVALID   = 16,
3447    };
3448    static const uint8_t cond_fixup[16] = {
3449        [0 ... 15] = INVALID,
3450        [TCG_COND_EQ] = 0,
3451        [TCG_COND_GT] = 0,
3452        [TCG_COND_NE] = NEED_INV,
3453        [TCG_COND_LE] = NEED_INV,
3454        [TCG_COND_LT] = NEED_SWAP,
3455        [TCG_COND_GE] = NEED_SWAP | NEED_INV,
3456        [TCG_COND_LEU] = NEED_UMIN,
3457        [TCG_COND_GTU] = NEED_UMIN | NEED_INV,
3458        [TCG_COND_GEU] = NEED_UMAX,
3459        [TCG_COND_LTU] = NEED_UMAX | NEED_INV,
3460    };
3461    int fixup = cond_fixup[cond];
3462
3463    assert(!(fixup & INVALID));
3464
3465    if (fixup & NEED_INV) {
3466        cond = tcg_invert_cond(cond);
3467    }
3468
3469    if (fixup & NEED_SWAP) {
3470        TCGReg swap = v1;
3471        v1 = v2;
3472        v2 = swap;
3473        cond = tcg_swap_cond(cond);
3474    }
3475
3476    if (fixup & (NEED_UMIN | NEED_UMAX)) {
3477        int op = (fixup & NEED_UMIN ? umin_insn[vece] : umax_insn[vece]);
3478
3479        /* avx2 does not have 64-bit min/max; adjusted during expand. */
3480        assert(vece <= MO_32);
3481
3482        tcg_out_vex_modrm_type(s, op, TCG_TMP_VEC, v1, v2, type);
3483        v2 = TCG_TMP_VEC;
3484        cond = TCG_COND_EQ;
3485    }
3486
3487    switch (cond) {
3488    case TCG_COND_EQ:
3489        tcg_out_vex_modrm_type(s, cmpeq_insn[vece], v0, v1, v2, type);
3490        break;
3491    case TCG_COND_GT:
3492        tcg_out_vex_modrm_type(s, cmpgt_insn[vece], v0, v1, v2, type);
3493        break;
3494    default:
3495        g_assert_not_reached();
3496    }
3497    return fixup & NEED_INV;
3498}
3499
3500static void tcg_out_cmp_vec_k1(TCGContext *s, TCGType type, unsigned vece,
3501                               TCGReg v1, TCGReg v2, TCGCond cond)
3502{
3503    static const int cmpm_insn[2][4] = {
3504        { OPC_VPCMPB, OPC_VPCMPW, OPC_VPCMPD, OPC_VPCMPQ },
3505        { OPC_VPCMPUB, OPC_VPCMPUW, OPC_VPCMPUD, OPC_VPCMPUQ }
3506    };
3507    static const int testm_insn[4] = {
3508        OPC_VPTESTMB, OPC_VPTESTMW, OPC_VPTESTMD, OPC_VPTESTMQ
3509    };
3510    static const int testnm_insn[4] = {
3511        OPC_VPTESTNMB, OPC_VPTESTNMW, OPC_VPTESTNMD, OPC_VPTESTNMQ
3512    };
3513
3514    static const int cond_ext[16] = {
3515        [TCG_COND_EQ] = 0,
3516        [TCG_COND_NE] = 4,
3517        [TCG_COND_LT] = 1,
3518        [TCG_COND_LTU] = 1,
3519        [TCG_COND_LE] = 2,
3520        [TCG_COND_LEU] = 2,
3521        [TCG_COND_NEVER] = 3,
3522        [TCG_COND_GE] = 5,
3523        [TCG_COND_GEU] = 5,
3524        [TCG_COND_GT] = 6,
3525        [TCG_COND_GTU] = 6,
3526        [TCG_COND_ALWAYS] = 7,
3527    };
3528
3529    switch (cond) {
3530    case TCG_COND_TSTNE:
3531        tcg_out_vex_modrm_type(s, testm_insn[vece], /* k1 */ 1, v1, v2, type);
3532        break;
3533    case TCG_COND_TSTEQ:
3534        tcg_out_vex_modrm_type(s, testnm_insn[vece], /* k1 */ 1, v1, v2, type);
3535        break;
3536    default:
3537        tcg_out_vex_modrm_type(s, cmpm_insn[is_unsigned_cond(cond)][vece],
3538                               /* k1 */ 1, v1, v2, type);
3539        tcg_out8(s, cond_ext[cond]);
3540        break;
3541    }
3542}
3543
3544static void tcg_out_k1_to_vec(TCGContext *s, TCGType type,
3545                              unsigned vece, TCGReg dest)
3546{
3547    static const int movm_insn[] = {
3548        OPC_VPMOVM2B, OPC_VPMOVM2W, OPC_VPMOVM2D, OPC_VPMOVM2Q
3549    };
3550    tcg_out_vex_modrm_type(s, movm_insn[vece], dest, 0, /* k1 */ 1, type);
3551}
3552
3553static void tcg_out_cmp_vec(TCGContext *s, TCGType type, unsigned vece,
3554                            TCGReg v0, TCGReg v1, TCGReg v2, TCGCond cond)
3555{
3556    /*
3557     * With avx512, we have a complete set of comparisons into mask.
3558     * Unless there's a single insn expansion for the comparision,
3559     * expand via a mask in k1.
3560     */
3561    if ((vece <= MO_16 ? have_avx512bw : have_avx512dq)
3562        && cond != TCG_COND_EQ
3563        && cond != TCG_COND_LT
3564        && cond != TCG_COND_GT) {
3565        tcg_out_cmp_vec_k1(s, type, vece, v1, v2, cond);
3566        tcg_out_k1_to_vec(s, type, vece, v0);
3567        return;
3568    }
3569
3570    if (tcg_out_cmp_vec_noinv(s, type, vece, v0, v1, v2, cond)) {
3571        tcg_out_dupi_vec(s, type, vece, TCG_TMP_VEC, -1);
3572        tcg_out_vex_modrm_type(s, OPC_PXOR, v0, v0, TCG_TMP_VEC, type);
3573    }
3574}
3575
3576static void tcg_out_cmpsel_vec_k1(TCGContext *s, TCGType type, unsigned vece,
3577                                  TCGReg v0, TCGReg c1, TCGReg c2,
3578                                  TCGReg v3, TCGReg v4, TCGCond cond)
3579{
3580    static const int vpblendm_insn[] = {
3581        OPC_VPBLENDMB, OPC_VPBLENDMW, OPC_VPBLENDMD, OPC_VPBLENDMQ
3582    };
3583    bool z = false;
3584
3585    /* Swap to place constant in V4 to take advantage of zero-masking. */
3586    if (!v3) {
3587        z = true;
3588        v3 = v4;
3589        cond = tcg_invert_cond(cond);
3590    }
3591
3592    tcg_out_cmp_vec_k1(s, type, vece, c1, c2, cond);
3593    tcg_out_evex_modrm_type(s, vpblendm_insn[vece], v0, v4, v3,
3594                            /* k1 */1, z, type);
3595}
3596
3597static void tcg_out_cmpsel_vec(TCGContext *s, TCGType type, unsigned vece,
3598                               TCGReg v0, TCGReg c1, TCGReg c2,
3599                               TCGReg v3, TCGReg v4, TCGCond cond)
3600{
3601    bool inv;
3602
3603    if (vece <= MO_16 ? have_avx512bw : have_avx512vl) {
3604        tcg_out_cmpsel_vec_k1(s, type, vece, v0, c1, c2, v3, v4, cond);
3605        return;
3606    }
3607
3608    inv = tcg_out_cmp_vec_noinv(s, type, vece, TCG_TMP_VEC, c1, c2, cond);
3609
3610    /*
3611     * Since XMM0 is 16, the only way we get 0 into V3
3612     * is via the constant zero constraint.
3613     */
3614    if (!v3) {
3615        if (inv) {
3616            tcg_out_vex_modrm_type(s, OPC_PAND, v0, TCG_TMP_VEC, v4, type);
3617        } else {
3618            tcg_out_vex_modrm_type(s, OPC_PANDN, v0, TCG_TMP_VEC, v4, type);
3619        }
3620    } else {
3621        if (inv) {
3622            TCGReg swap = v3;
3623            v3 = v4;
3624            v4 = swap;
3625        }
3626        tcg_out_vex_modrm_type(s, OPC_VPBLENDVB, v0, v4, v3, type);
3627        tcg_out8(s, (TCG_TMP_VEC - TCG_REG_XMM0) << 4);
3628    }
3629}
3630
3631static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
3632                           unsigned vecl, unsigned vece,
3633                           const TCGArg args[TCG_MAX_OP_ARGS],
3634                           const int const_args[TCG_MAX_OP_ARGS])
3635{
3636    static int const add_insn[4] = {
3637        OPC_PADDB, OPC_PADDW, OPC_PADDD, OPC_PADDQ
3638    };
3639    static int const ssadd_insn[4] = {
3640        OPC_PADDSB, OPC_PADDSW, OPC_UD2, OPC_UD2
3641    };
3642    static int const usadd_insn[4] = {
3643        OPC_PADDUB, OPC_PADDUW, OPC_UD2, OPC_UD2
3644    };
3645    static int const sub_insn[4] = {
3646        OPC_PSUBB, OPC_PSUBW, OPC_PSUBD, OPC_PSUBQ
3647    };
3648    static int const sssub_insn[4] = {
3649        OPC_PSUBSB, OPC_PSUBSW, OPC_UD2, OPC_UD2
3650    };
3651    static int const ussub_insn[4] = {
3652        OPC_PSUBUB, OPC_PSUBUW, OPC_UD2, OPC_UD2
3653    };
3654    static int const mul_insn[4] = {
3655        OPC_UD2, OPC_PMULLW, OPC_PMULLD, OPC_VPMULLQ
3656    };
3657    static int const shift_imm_insn[4] = {
3658        OPC_UD2, OPC_PSHIFTW_Ib, OPC_PSHIFTD_Ib, OPC_PSHIFTQ_Ib
3659    };
3660    static int const punpckl_insn[4] = {
3661        OPC_PUNPCKLBW, OPC_PUNPCKLWD, OPC_PUNPCKLDQ, OPC_PUNPCKLQDQ
3662    };
3663    static int const punpckh_insn[4] = {
3664        OPC_PUNPCKHBW, OPC_PUNPCKHWD, OPC_PUNPCKHDQ, OPC_PUNPCKHQDQ
3665    };
3666    static int const packss_insn[4] = {
3667        OPC_PACKSSWB, OPC_PACKSSDW, OPC_UD2, OPC_UD2
3668    };
3669    static int const packus_insn[4] = {
3670        OPC_PACKUSWB, OPC_PACKUSDW, OPC_UD2, OPC_UD2
3671    };
3672    static int const smin_insn[4] = {
3673        OPC_PMINSB, OPC_PMINSW, OPC_PMINSD, OPC_VPMINSQ
3674    };
3675    static int const smax_insn[4] = {
3676        OPC_PMAXSB, OPC_PMAXSW, OPC_PMAXSD, OPC_VPMAXSQ
3677    };
3678    static int const rotlv_insn[4] = {
3679        OPC_UD2, OPC_UD2, OPC_VPROLVD, OPC_VPROLVQ
3680    };
3681    static int const rotrv_insn[4] = {
3682        OPC_UD2, OPC_UD2, OPC_VPRORVD, OPC_VPRORVQ
3683    };
3684    static int const shlv_insn[4] = {
3685        OPC_UD2, OPC_VPSLLVW, OPC_VPSLLVD, OPC_VPSLLVQ
3686    };
3687    static int const shrv_insn[4] = {
3688        OPC_UD2, OPC_VPSRLVW, OPC_VPSRLVD, OPC_VPSRLVQ
3689    };
3690    static int const sarv_insn[4] = {
3691        OPC_UD2, OPC_VPSRAVW, OPC_VPSRAVD, OPC_VPSRAVQ
3692    };
3693    static int const shls_insn[4] = {
3694        OPC_UD2, OPC_PSLLW, OPC_PSLLD, OPC_PSLLQ
3695    };
3696    static int const shrs_insn[4] = {
3697        OPC_UD2, OPC_PSRLW, OPC_PSRLD, OPC_PSRLQ
3698    };
3699    static int const sars_insn[4] = {
3700        OPC_UD2, OPC_PSRAW, OPC_PSRAD, OPC_VPSRAQ
3701    };
3702    static int const vpshldi_insn[4] = {
3703        OPC_UD2, OPC_VPSHLDW, OPC_VPSHLDD, OPC_VPSHLDQ
3704    };
3705    static int const vpshldv_insn[4] = {
3706        OPC_UD2, OPC_VPSHLDVW, OPC_VPSHLDVD, OPC_VPSHLDVQ
3707    };
3708    static int const vpshrdv_insn[4] = {
3709        OPC_UD2, OPC_VPSHRDVW, OPC_VPSHRDVD, OPC_VPSHRDVQ
3710    };
3711    static int const abs_insn[4] = {
3712        OPC_PABSB, OPC_PABSW, OPC_PABSD, OPC_VPABSQ
3713    };
3714
3715    TCGType type = vecl + TCG_TYPE_V64;
3716    int insn, sub;
3717    TCGArg a0, a1, a2, a3;
3718
3719    a0 = args[0];
3720    a1 = args[1];
3721    a2 = args[2];
3722
3723    switch (opc) {
3724    case INDEX_op_add_vec:
3725        insn = add_insn[vece];
3726        goto gen_simd;
3727    case INDEX_op_ssadd_vec:
3728        insn = ssadd_insn[vece];
3729        goto gen_simd;
3730    case INDEX_op_usadd_vec:
3731        insn = usadd_insn[vece];
3732        goto gen_simd;
3733    case INDEX_op_sub_vec:
3734        insn = sub_insn[vece];
3735        goto gen_simd;
3736    case INDEX_op_sssub_vec:
3737        insn = sssub_insn[vece];
3738        goto gen_simd;
3739    case INDEX_op_ussub_vec:
3740        insn = ussub_insn[vece];
3741        goto gen_simd;
3742    case INDEX_op_mul_vec:
3743        insn = mul_insn[vece];
3744        goto gen_simd;
3745    case INDEX_op_and_vec:
3746        insn = OPC_PAND;
3747        goto gen_simd;
3748    case INDEX_op_or_vec:
3749        insn = OPC_POR;
3750        goto gen_simd;
3751    case INDEX_op_xor_vec:
3752        insn = OPC_PXOR;
3753        goto gen_simd;
3754    case INDEX_op_smin_vec:
3755        insn = smin_insn[vece];
3756        goto gen_simd;
3757    case INDEX_op_umin_vec:
3758        insn = umin_insn[vece];
3759        goto gen_simd;
3760    case INDEX_op_smax_vec:
3761        insn = smax_insn[vece];
3762        goto gen_simd;
3763    case INDEX_op_umax_vec:
3764        insn = umax_insn[vece];
3765        goto gen_simd;
3766    case INDEX_op_shlv_vec:
3767        insn = shlv_insn[vece];
3768        goto gen_simd;
3769    case INDEX_op_shrv_vec:
3770        insn = shrv_insn[vece];
3771        goto gen_simd;
3772    case INDEX_op_sarv_vec:
3773        insn = sarv_insn[vece];
3774        goto gen_simd;
3775    case INDEX_op_rotlv_vec:
3776        insn = rotlv_insn[vece];
3777        goto gen_simd;
3778    case INDEX_op_rotrv_vec:
3779        insn = rotrv_insn[vece];
3780        goto gen_simd;
3781    case INDEX_op_shls_vec:
3782        insn = shls_insn[vece];
3783        goto gen_simd;
3784    case INDEX_op_shrs_vec:
3785        insn = shrs_insn[vece];
3786        goto gen_simd;
3787    case INDEX_op_sars_vec:
3788        insn = sars_insn[vece];
3789        goto gen_simd;
3790    case INDEX_op_x86_punpckl_vec:
3791        insn = punpckl_insn[vece];
3792        goto gen_simd;
3793    case INDEX_op_x86_punpckh_vec:
3794        insn = punpckh_insn[vece];
3795        goto gen_simd;
3796    case INDEX_op_x86_packss_vec:
3797        insn = packss_insn[vece];
3798        goto gen_simd;
3799    case INDEX_op_x86_packus_vec:
3800        insn = packus_insn[vece];
3801        goto gen_simd;
3802    case INDEX_op_x86_vpshldv_vec:
3803        insn = vpshldv_insn[vece];
3804        a1 = a2;
3805        a2 = args[3];
3806        goto gen_simd;
3807    case INDEX_op_x86_vpshrdv_vec:
3808        insn = vpshrdv_insn[vece];
3809        a1 = a2;
3810        a2 = args[3];
3811        goto gen_simd;
3812#if TCG_TARGET_REG_BITS == 32
3813    case INDEX_op_dup2_vec:
3814        /* First merge the two 32-bit inputs to a single 64-bit element. */
3815        tcg_out_vex_modrm(s, OPC_PUNPCKLDQ, a0, a1, a2);
3816        /* Then replicate the 64-bit elements across the rest of the vector. */
3817        if (type != TCG_TYPE_V64) {
3818            tcg_out_dup_vec(s, type, MO_64, a0, a0);
3819        }
3820        break;
3821#endif
3822    case INDEX_op_abs_vec:
3823        insn = abs_insn[vece];
3824        a2 = a1;
3825        a1 = 0;
3826        goto gen_simd;
3827    gen_simd:
3828        tcg_debug_assert(insn != OPC_UD2);
3829        tcg_out_vex_modrm_type(s, insn, a0, a1, a2, type);
3830        break;
3831
3832    case INDEX_op_cmp_vec:
3833        tcg_out_cmp_vec(s, type, vece, a0, a1, a2, args[3]);
3834        break;
3835
3836    case INDEX_op_cmpsel_vec:
3837        tcg_out_cmpsel_vec(s, type, vece, a0, a1, a2,
3838                           args[3], args[4], args[5]);
3839        break;
3840
3841    case INDEX_op_andc_vec:
3842        insn = OPC_PANDN;
3843        tcg_out_vex_modrm_type(s, insn, a0, a2, a1, type);
3844        break;
3845
3846    case INDEX_op_shli_vec:
3847        insn = shift_imm_insn[vece];
3848        sub = 6;
3849        goto gen_shift;
3850    case INDEX_op_shri_vec:
3851        insn = shift_imm_insn[vece];
3852        sub = 2;
3853        goto gen_shift;
3854    case INDEX_op_sari_vec:
3855        if (vece == MO_64) {
3856            insn = OPC_PSHIFTD_Ib | P_VEXW | P_EVEX;
3857        } else {
3858            insn = shift_imm_insn[vece];
3859        }
3860        sub = 4;
3861        goto gen_shift;
3862    case INDEX_op_rotli_vec:
3863        insn = OPC_PSHIFTD_Ib | P_EVEX;  /* VPROL[DQ] */
3864        if (vece == MO_64) {
3865            insn |= P_VEXW;
3866        }
3867        sub = 1;
3868        goto gen_shift;
3869    gen_shift:
3870        tcg_debug_assert(vece != MO_8);
3871        tcg_out_vex_modrm_type(s, insn, sub, a0, a1, type);
3872        tcg_out8(s, a2);
3873        break;
3874
3875    case INDEX_op_ld_vec:
3876        tcg_out_ld(s, type, a0, a1, a2);
3877        break;
3878    case INDEX_op_st_vec:
3879        tcg_out_st(s, type, a0, a1, a2);
3880        break;
3881    case INDEX_op_dupm_vec:
3882        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
3883        break;
3884
3885    case INDEX_op_x86_shufps_vec:
3886        insn = OPC_SHUFPS;
3887        sub = args[3];
3888        goto gen_simd_imm8;
3889    case INDEX_op_x86_blend_vec:
3890        if (vece == MO_16) {
3891            insn = OPC_PBLENDW;
3892        } else if (vece == MO_32) {
3893            insn = (have_avx2 ? OPC_VPBLENDD : OPC_BLENDPS);
3894        } else {
3895            g_assert_not_reached();
3896        }
3897        sub = args[3];
3898        goto gen_simd_imm8;
3899    case INDEX_op_x86_vperm2i128_vec:
3900        insn = OPC_VPERM2I128;
3901        sub = args[3];
3902        goto gen_simd_imm8;
3903    case INDEX_op_x86_vpshldi_vec:
3904        insn = vpshldi_insn[vece];
3905        sub = args[3];
3906        goto gen_simd_imm8;
3907
3908    case INDEX_op_not_vec:
3909        insn = OPC_VPTERNLOGQ;
3910        a2 = a1;
3911        sub = 0x33; /* !B */
3912        goto gen_simd_imm8;
3913    case INDEX_op_nor_vec:
3914        insn = OPC_VPTERNLOGQ;
3915        sub = 0x11; /* norCB */
3916        goto gen_simd_imm8;
3917    case INDEX_op_nand_vec:
3918        insn = OPC_VPTERNLOGQ;
3919        sub = 0x77; /* nandCB */
3920        goto gen_simd_imm8;
3921    case INDEX_op_eqv_vec:
3922        insn = OPC_VPTERNLOGQ;
3923        sub = 0x99; /* xnorCB */
3924        goto gen_simd_imm8;
3925    case INDEX_op_orc_vec:
3926        insn = OPC_VPTERNLOGQ;
3927        sub = 0xdd; /* orB!C */
3928        goto gen_simd_imm8;
3929
3930    case INDEX_op_bitsel_vec:
3931        insn = OPC_VPTERNLOGQ;
3932        a3 = args[3];
3933        if (a0 == a1) {
3934            a1 = a2;
3935            a2 = a3;
3936            sub = 0xca; /* A?B:C */
3937        } else if (a0 == a2) {
3938            a2 = a3;
3939            sub = 0xe2; /* B?A:C */
3940        } else {
3941            tcg_out_mov(s, type, a0, a3);
3942            sub = 0xb8; /* B?C:A */
3943        }
3944        goto gen_simd_imm8;
3945
3946    gen_simd_imm8:
3947        tcg_debug_assert(insn != OPC_UD2);
3948        tcg_out_vex_modrm_type(s, insn, a0, a1, a2, type);
3949        tcg_out8(s, sub);
3950        break;
3951
3952    case INDEX_op_x86_psrldq_vec:
3953        tcg_out_vex_modrm(s, OPC_GRP14, 3, a0, a1);
3954        tcg_out8(s, a2);
3955        break;
3956
3957    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
3958    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
3959    default:
3960        g_assert_not_reached();
3961    }
3962}
3963
3964static TCGConstraintSetIndex
3965tcg_target_op_def(TCGOpcode op, TCGType type, unsigned flags)
3966{
3967    switch (op) {
3968    case INDEX_op_goto_ptr:
3969        return C_O0_I1(r);
3970
3971    case INDEX_op_ld8u_i32:
3972    case INDEX_op_ld8u_i64:
3973    case INDEX_op_ld8s_i32:
3974    case INDEX_op_ld8s_i64:
3975    case INDEX_op_ld16u_i32:
3976    case INDEX_op_ld16u_i64:
3977    case INDEX_op_ld16s_i32:
3978    case INDEX_op_ld16s_i64:
3979    case INDEX_op_ld_i32:
3980    case INDEX_op_ld32u_i64:
3981    case INDEX_op_ld32s_i64:
3982    case INDEX_op_ld_i64:
3983        return C_O1_I1(r, r);
3984
3985    case INDEX_op_st8_i32:
3986    case INDEX_op_st8_i64:
3987        return C_O0_I2(qi, r);
3988
3989    case INDEX_op_st16_i32:
3990    case INDEX_op_st16_i64:
3991    case INDEX_op_st_i32:
3992    case INDEX_op_st32_i64:
3993        return C_O0_I2(ri, r);
3994
3995    case INDEX_op_st_i64:
3996        return C_O0_I2(re, r);
3997
3998    case INDEX_op_extrh_i64_i32:
3999        return C_O1_I1(r, 0);
4000
4001    case INDEX_op_ext_i32_i64:
4002    case INDEX_op_extu_i32_i64:
4003    case INDEX_op_extrl_i64_i32:
4004    case INDEX_op_sextract_i32:
4005    case INDEX_op_sextract_i64:
4006        return C_O1_I1(r, r);
4007
4008    case INDEX_op_extract2_i32:
4009    case INDEX_op_extract2_i64:
4010        return C_O1_I2(r, 0, r);
4011
4012    case INDEX_op_deposit_i32:
4013    case INDEX_op_deposit_i64:
4014        return C_O1_I2(q, 0, qi);
4015
4016    case INDEX_op_add2_i32:
4017    case INDEX_op_add2_i64:
4018    case INDEX_op_sub2_i32:
4019    case INDEX_op_sub2_i64:
4020        return C_N1_O1_I4(r, r, 0, 1, re, re);
4021
4022    case INDEX_op_qemu_ld_i32:
4023        return C_O1_I1(r, L);
4024
4025    case INDEX_op_qemu_st_i32:
4026        return C_O0_I2(L, L);
4027    case INDEX_op_qemu_st8_i32:
4028        return C_O0_I2(s, L);
4029
4030    case INDEX_op_qemu_ld_i64:
4031        return TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L) : C_O2_I1(r, r, L);
4032
4033    case INDEX_op_qemu_st_i64:
4034        return TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L) : C_O0_I3(L, L, L);
4035
4036    case INDEX_op_qemu_ld_i128:
4037        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
4038        return C_O2_I1(r, r, L);
4039    case INDEX_op_qemu_st_i128:
4040        tcg_debug_assert(TCG_TARGET_REG_BITS == 64);
4041        return C_O0_I3(L, L, L);
4042
4043    case INDEX_op_ld_vec:
4044    case INDEX_op_dupm_vec:
4045        return C_O1_I1(x, r);
4046
4047    case INDEX_op_st_vec:
4048        return C_O0_I2(x, r);
4049
4050    case INDEX_op_add_vec:
4051    case INDEX_op_sub_vec:
4052    case INDEX_op_mul_vec:
4053    case INDEX_op_and_vec:
4054    case INDEX_op_or_vec:
4055    case INDEX_op_xor_vec:
4056    case INDEX_op_andc_vec:
4057    case INDEX_op_orc_vec:
4058    case INDEX_op_nand_vec:
4059    case INDEX_op_nor_vec:
4060    case INDEX_op_eqv_vec:
4061    case INDEX_op_ssadd_vec:
4062    case INDEX_op_usadd_vec:
4063    case INDEX_op_sssub_vec:
4064    case INDEX_op_ussub_vec:
4065    case INDEX_op_smin_vec:
4066    case INDEX_op_umin_vec:
4067    case INDEX_op_smax_vec:
4068    case INDEX_op_umax_vec:
4069    case INDEX_op_shlv_vec:
4070    case INDEX_op_shrv_vec:
4071    case INDEX_op_sarv_vec:
4072    case INDEX_op_rotlv_vec:
4073    case INDEX_op_rotrv_vec:
4074    case INDEX_op_shls_vec:
4075    case INDEX_op_shrs_vec:
4076    case INDEX_op_sars_vec:
4077    case INDEX_op_cmp_vec:
4078    case INDEX_op_x86_shufps_vec:
4079    case INDEX_op_x86_blend_vec:
4080    case INDEX_op_x86_packss_vec:
4081    case INDEX_op_x86_packus_vec:
4082    case INDEX_op_x86_vperm2i128_vec:
4083    case INDEX_op_x86_punpckl_vec:
4084    case INDEX_op_x86_punpckh_vec:
4085    case INDEX_op_x86_vpshldi_vec:
4086#if TCG_TARGET_REG_BITS == 32
4087    case INDEX_op_dup2_vec:
4088#endif
4089        return C_O1_I2(x, x, x);
4090
4091    case INDEX_op_abs_vec:
4092    case INDEX_op_dup_vec:
4093    case INDEX_op_not_vec:
4094    case INDEX_op_shli_vec:
4095    case INDEX_op_shri_vec:
4096    case INDEX_op_sari_vec:
4097    case INDEX_op_rotli_vec:
4098    case INDEX_op_x86_psrldq_vec:
4099        return C_O1_I1(x, x);
4100
4101    case INDEX_op_x86_vpshldv_vec:
4102    case INDEX_op_x86_vpshrdv_vec:
4103        return C_O1_I3(x, 0, x, x);
4104
4105    case INDEX_op_bitsel_vec:
4106        return C_O1_I3(x, x, x, x);
4107    case INDEX_op_cmpsel_vec:
4108        return C_O1_I4(x, x, x, xO, x);
4109
4110    default:
4111        return C_NotImplemented;
4112    }
4113}
4114
4115int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
4116{
4117    switch (opc) {
4118    case INDEX_op_add_vec:
4119    case INDEX_op_sub_vec:
4120    case INDEX_op_and_vec:
4121    case INDEX_op_or_vec:
4122    case INDEX_op_xor_vec:
4123    case INDEX_op_andc_vec:
4124    case INDEX_op_orc_vec:
4125    case INDEX_op_nand_vec:
4126    case INDEX_op_nor_vec:
4127    case INDEX_op_eqv_vec:
4128    case INDEX_op_not_vec:
4129    case INDEX_op_bitsel_vec:
4130        return 1;
4131    case INDEX_op_cmp_vec:
4132    case INDEX_op_cmpsel_vec:
4133        return -1;
4134
4135    case INDEX_op_rotli_vec:
4136        return have_avx512vl && vece >= MO_32 ? 1 : -1;
4137
4138    case INDEX_op_shli_vec:
4139    case INDEX_op_shri_vec:
4140        /* We must expand the operation for MO_8.  */
4141        return vece == MO_8 ? -1 : 1;
4142
4143    case INDEX_op_sari_vec:
4144        switch (vece) {
4145        case MO_8:
4146            return -1;
4147        case MO_16:
4148        case MO_32:
4149            return 1;
4150        case MO_64:
4151            if (have_avx512vl) {
4152                return 1;
4153            }
4154            /*
4155             * We can emulate this for MO_64, but it does not pay off
4156             * unless we're producing at least 4 values.
4157             */
4158            return type >= TCG_TYPE_V256 ? -1 : 0;
4159        }
4160        return 0;
4161
4162    case INDEX_op_shls_vec:
4163    case INDEX_op_shrs_vec:
4164        return vece >= MO_16;
4165    case INDEX_op_sars_vec:
4166        switch (vece) {
4167        case MO_16:
4168        case MO_32:
4169            return 1;
4170        case MO_64:
4171            return have_avx512vl;
4172        }
4173        return 0;
4174    case INDEX_op_rotls_vec:
4175        return vece >= MO_16 ? -1 : 0;
4176
4177    case INDEX_op_shlv_vec:
4178    case INDEX_op_shrv_vec:
4179        switch (vece) {
4180        case MO_16:
4181            return have_avx512bw;
4182        case MO_32:
4183        case MO_64:
4184            return have_avx2;
4185        }
4186        return 0;
4187    case INDEX_op_sarv_vec:
4188        switch (vece) {
4189        case MO_16:
4190            return have_avx512bw;
4191        case MO_32:
4192            return have_avx2;
4193        case MO_64:
4194            return have_avx512vl;
4195        }
4196        return 0;
4197    case INDEX_op_rotlv_vec:
4198    case INDEX_op_rotrv_vec:
4199        switch (vece) {
4200        case MO_16:
4201            return have_avx512vbmi2 ? -1 : 0;
4202        case MO_32:
4203        case MO_64:
4204            return have_avx512vl ? 1 : have_avx2 ? -1 : 0;
4205        }
4206        return 0;
4207
4208    case INDEX_op_mul_vec:
4209        switch (vece) {
4210        case MO_8:
4211            return -1;
4212        case MO_64:
4213            return have_avx512dq;
4214        }
4215        return 1;
4216
4217    case INDEX_op_ssadd_vec:
4218    case INDEX_op_usadd_vec:
4219    case INDEX_op_sssub_vec:
4220    case INDEX_op_ussub_vec:
4221        return vece <= MO_16;
4222    case INDEX_op_smin_vec:
4223    case INDEX_op_smax_vec:
4224    case INDEX_op_umin_vec:
4225    case INDEX_op_umax_vec:
4226    case INDEX_op_abs_vec:
4227        return vece <= MO_32 || have_avx512vl;
4228
4229    default:
4230        return 0;
4231    }
4232}
4233
4234static void expand_vec_shi(TCGType type, unsigned vece, bool right,
4235                           TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4236{
4237    uint8_t mask;
4238
4239    tcg_debug_assert(vece == MO_8);
4240    if (right) {
4241        mask = 0xff >> imm;
4242        tcg_gen_shri_vec(MO_16, v0, v1, imm);
4243    } else {
4244        mask = 0xff << imm;
4245        tcg_gen_shli_vec(MO_16, v0, v1, imm);
4246    }
4247    tcg_gen_and_vec(MO_8, v0, v0, tcg_constant_vec(type, MO_8, mask));
4248}
4249
4250static void expand_vec_sari(TCGType type, unsigned vece,
4251                            TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4252{
4253    TCGv_vec t1, t2;
4254
4255    switch (vece) {
4256    case MO_8:
4257        /* Unpack to 16-bit, shift, and repack.  */
4258        t1 = tcg_temp_new_vec(type);
4259        t2 = tcg_temp_new_vec(type);
4260        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4261                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
4262        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4263                  tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
4264        tcg_gen_sari_vec(MO_16, t1, t1, imm + 8);
4265        tcg_gen_sari_vec(MO_16, t2, t2, imm + 8);
4266        vec_gen_3(INDEX_op_x86_packss_vec, type, MO_8,
4267                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
4268        tcg_temp_free_vec(t1);
4269        tcg_temp_free_vec(t2);
4270        break;
4271
4272    case MO_64:
4273        t1 = tcg_temp_new_vec(type);
4274        if (imm <= 32) {
4275            /*
4276             * We can emulate a small sign extend by performing an arithmetic
4277             * 32-bit shift and overwriting the high half of a 64-bit logical
4278             * shift.  Note that the ISA says shift of 32 is valid, but TCG
4279             * does not, so we have to bound the smaller shift -- we get the
4280             * same result in the high half either way.
4281             */
4282            tcg_gen_sari_vec(MO_32, t1, v1, MIN(imm, 31));
4283            tcg_gen_shri_vec(MO_64, v0, v1, imm);
4284            vec_gen_4(INDEX_op_x86_blend_vec, type, MO_32,
4285                      tcgv_vec_arg(v0), tcgv_vec_arg(v0),
4286                      tcgv_vec_arg(t1), 0xaa);
4287        } else {
4288            /* Otherwise we will need to use a compare vs 0 to produce
4289             * the sign-extend, shift and merge.
4290             */
4291            tcg_gen_cmp_vec(TCG_COND_GT, MO_64, t1,
4292                            tcg_constant_vec(type, MO_64, 0), v1);
4293            tcg_gen_shri_vec(MO_64, v0, v1, imm);
4294            tcg_gen_shli_vec(MO_64, t1, t1, 64 - imm);
4295            tcg_gen_or_vec(MO_64, v0, v0, t1);
4296        }
4297        tcg_temp_free_vec(t1);
4298        break;
4299
4300    default:
4301        g_assert_not_reached();
4302    }
4303}
4304
4305static void expand_vec_rotli(TCGType type, unsigned vece,
4306                             TCGv_vec v0, TCGv_vec v1, TCGArg imm)
4307{
4308    TCGv_vec t;
4309
4310    if (vece != MO_8 && have_avx512vbmi2) {
4311        vec_gen_4(INDEX_op_x86_vpshldi_vec, type, vece,
4312                  tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v1), imm);
4313        return;
4314    }
4315
4316    t = tcg_temp_new_vec(type);
4317    tcg_gen_shli_vec(vece, t, v1, imm);
4318    tcg_gen_shri_vec(vece, v0, v1, (8 << vece) - imm);
4319    tcg_gen_or_vec(vece, v0, v0, t);
4320    tcg_temp_free_vec(t);
4321}
4322
4323static void expand_vec_rotv(TCGType type, unsigned vece, TCGv_vec v0,
4324                            TCGv_vec v1, TCGv_vec sh, bool right)
4325{
4326    TCGv_vec t;
4327
4328    if (have_avx512vbmi2) {
4329        vec_gen_4(right ? INDEX_op_x86_vpshrdv_vec : INDEX_op_x86_vpshldv_vec,
4330                  type, vece, tcgv_vec_arg(v0), tcgv_vec_arg(v1),
4331                  tcgv_vec_arg(v1), tcgv_vec_arg(sh));
4332        return;
4333    }
4334
4335    t = tcg_temp_new_vec(type);
4336    tcg_gen_dupi_vec(vece, t, 8 << vece);
4337    tcg_gen_sub_vec(vece, t, t, sh);
4338    if (right) {
4339        tcg_gen_shlv_vec(vece, t, v1, t);
4340        tcg_gen_shrv_vec(vece, v0, v1, sh);
4341    } else {
4342        tcg_gen_shrv_vec(vece, t, v1, t);
4343        tcg_gen_shlv_vec(vece, v0, v1, sh);
4344    }
4345    tcg_gen_or_vec(vece, v0, v0, t);
4346    tcg_temp_free_vec(t);
4347}
4348
4349static void expand_vec_rotls(TCGType type, unsigned vece,
4350                             TCGv_vec v0, TCGv_vec v1, TCGv_i32 lsh)
4351{
4352    TCGv_vec t = tcg_temp_new_vec(type);
4353
4354    tcg_debug_assert(vece != MO_8);
4355
4356    if (vece >= MO_32 ? have_avx512vl : have_avx512vbmi2) {
4357        tcg_gen_dup_i32_vec(vece, t, lsh);
4358        if (vece >= MO_32) {
4359            tcg_gen_rotlv_vec(vece, v0, v1, t);
4360        } else {
4361            expand_vec_rotv(type, vece, v0, v1, t, false);
4362        }
4363    } else {
4364        TCGv_i32 rsh = tcg_temp_new_i32();
4365
4366        tcg_gen_neg_i32(rsh, lsh);
4367        tcg_gen_andi_i32(rsh, rsh, (8 << vece) - 1);
4368        tcg_gen_shls_vec(vece, t, v1, lsh);
4369        tcg_gen_shrs_vec(vece, v0, v1, rsh);
4370        tcg_gen_or_vec(vece, v0, v0, t);
4371
4372        tcg_temp_free_i32(rsh);
4373    }
4374
4375    tcg_temp_free_vec(t);
4376}
4377
4378static void expand_vec_mul(TCGType type, unsigned vece,
4379                           TCGv_vec v0, TCGv_vec v1, TCGv_vec v2)
4380{
4381    TCGv_vec t1, t2, t3, t4, zero;
4382
4383    tcg_debug_assert(vece == MO_8);
4384
4385    /*
4386     * Unpack v1 bytes to words, 0 | x.
4387     * Unpack v2 bytes to words, y | 0.
4388     * This leaves the 8-bit result, x * y, with 8 bits of right padding.
4389     * Shift logical right by 8 bits to clear the high 8 bytes before
4390     * using an unsigned saturated pack.
4391     *
4392     * The difference between the V64, V128 and V256 cases is merely how
4393     * we distribute the expansion between temporaries.
4394     */
4395    switch (type) {
4396    case TCG_TYPE_V64:
4397        t1 = tcg_temp_new_vec(TCG_TYPE_V128);
4398        t2 = tcg_temp_new_vec(TCG_TYPE_V128);
4399        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
4400        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
4401                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4402        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
4403                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4404        tcg_gen_mul_vec(MO_16, t1, t1, t2);
4405        tcg_gen_shri_vec(MO_16, t1, t1, 8);
4406        vec_gen_3(INDEX_op_x86_packus_vec, TCG_TYPE_V128, MO_8,
4407                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t1));
4408        tcg_temp_free_vec(t1);
4409        tcg_temp_free_vec(t2);
4410        break;
4411
4412    case TCG_TYPE_V128:
4413    case TCG_TYPE_V256:
4414        t1 = tcg_temp_new_vec(type);
4415        t2 = tcg_temp_new_vec(type);
4416        t3 = tcg_temp_new_vec(type);
4417        t4 = tcg_temp_new_vec(type);
4418        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
4419        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4420                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4421        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
4422                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4423        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4424                  tcgv_vec_arg(t3), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
4425        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
4426                  tcgv_vec_arg(t4), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
4427        tcg_gen_mul_vec(MO_16, t1, t1, t2);
4428        tcg_gen_mul_vec(MO_16, t3, t3, t4);
4429        tcg_gen_shri_vec(MO_16, t1, t1, 8);
4430        tcg_gen_shri_vec(MO_16, t3, t3, 8);
4431        vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
4432                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t3));
4433        tcg_temp_free_vec(t1);
4434        tcg_temp_free_vec(t2);
4435        tcg_temp_free_vec(t3);
4436        tcg_temp_free_vec(t4);
4437        break;
4438
4439    default:
4440        g_assert_not_reached();
4441    }
4442}
4443
4444static TCGCond expand_vec_cond(TCGType type, unsigned vece,
4445                               TCGArg *a1, TCGArg *a2, TCGCond cond)
4446{
4447    /*
4448     * Without AVX512, there are no 64-bit unsigned comparisons.
4449     * We must bias the inputs so that they become signed.
4450     * All other swapping and inversion are handled during code generation.
4451     */
4452    if (vece == MO_64 && !have_avx512dq && is_unsigned_cond(cond)) {
4453        TCGv_vec v1 = temp_tcgv_vec(arg_temp(*a1));
4454        TCGv_vec v2 = temp_tcgv_vec(arg_temp(*a2));
4455        TCGv_vec t1 = tcg_temp_new_vec(type);
4456        TCGv_vec t2 = tcg_temp_new_vec(type);
4457        TCGv_vec t3 = tcg_constant_vec(type, vece, 1ull << ((8 << vece) - 1));
4458
4459        tcg_gen_sub_vec(vece, t1, v1, t3);
4460        tcg_gen_sub_vec(vece, t2, v2, t3);
4461        *a1 = tcgv_vec_arg(t1);
4462        *a2 = tcgv_vec_arg(t2);
4463        cond = tcg_signed_cond(cond);
4464    }
4465    return cond;
4466}
4467
4468static void expand_vec_cmp(TCGType type, unsigned vece, TCGArg a0,
4469                           TCGArg a1, TCGArg a2, TCGCond cond)
4470{
4471    cond = expand_vec_cond(type, vece, &a1, &a2, cond);
4472    /* Expand directly; do not recurse.  */
4473    vec_gen_4(INDEX_op_cmp_vec, type, vece, a0, a1, a2, cond);
4474}
4475
4476static void expand_vec_cmpsel(TCGType type, unsigned vece, TCGArg a0,
4477                              TCGArg a1, TCGArg a2,
4478                              TCGArg a3, TCGArg a4, TCGCond cond)
4479{
4480    cond = expand_vec_cond(type, vece, &a1, &a2, cond);
4481    /* Expand directly; do not recurse.  */
4482    vec_gen_6(INDEX_op_cmpsel_vec, type, vece, a0, a1, a2, a3, a4, cond);
4483}
4484
4485void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
4486                       TCGArg a0, ...)
4487{
4488    va_list va;
4489    TCGArg a1, a2, a3, a4, a5;
4490    TCGv_vec v0, v1, v2;
4491
4492    va_start(va, a0);
4493    a1 = va_arg(va, TCGArg);
4494    a2 = va_arg(va, TCGArg);
4495    v0 = temp_tcgv_vec(arg_temp(a0));
4496    v1 = temp_tcgv_vec(arg_temp(a1));
4497
4498    switch (opc) {
4499    case INDEX_op_shli_vec:
4500        expand_vec_shi(type, vece, false, v0, v1, a2);
4501        break;
4502    case INDEX_op_shri_vec:
4503        expand_vec_shi(type, vece, true, v0, v1, a2);
4504        break;
4505    case INDEX_op_sari_vec:
4506        expand_vec_sari(type, vece, v0, v1, a2);
4507        break;
4508
4509    case INDEX_op_rotli_vec:
4510        expand_vec_rotli(type, vece, v0, v1, a2);
4511        break;
4512
4513    case INDEX_op_rotls_vec:
4514        expand_vec_rotls(type, vece, v0, v1, temp_tcgv_i32(arg_temp(a2)));
4515        break;
4516
4517    case INDEX_op_rotlv_vec:
4518        v2 = temp_tcgv_vec(arg_temp(a2));
4519        expand_vec_rotv(type, vece, v0, v1, v2, false);
4520        break;
4521    case INDEX_op_rotrv_vec:
4522        v2 = temp_tcgv_vec(arg_temp(a2));
4523        expand_vec_rotv(type, vece, v0, v1, v2, true);
4524        break;
4525
4526    case INDEX_op_mul_vec:
4527        v2 = temp_tcgv_vec(arg_temp(a2));
4528        expand_vec_mul(type, vece, v0, v1, v2);
4529        break;
4530
4531    case INDEX_op_cmp_vec:
4532        a3 = va_arg(va, TCGArg);
4533        expand_vec_cmp(type, vece, a0, a1, a2, a3);
4534        break;
4535
4536    case INDEX_op_cmpsel_vec:
4537        a3 = va_arg(va, TCGArg);
4538        a4 = va_arg(va, TCGArg);
4539        a5 = va_arg(va, TCGArg);
4540        expand_vec_cmpsel(type, vece, a0, a1, a2, a3, a4, a5);
4541        break;
4542
4543    default:
4544        break;
4545    }
4546
4547    va_end(va);
4548}
4549
4550static const int tcg_target_callee_save_regs[] = {
4551#if TCG_TARGET_REG_BITS == 64
4552    TCG_REG_RBP,
4553    TCG_REG_RBX,
4554#if defined(_WIN64)
4555    TCG_REG_RDI,
4556    TCG_REG_RSI,
4557#endif
4558    TCG_REG_R12,
4559    TCG_REG_R13,
4560    TCG_REG_R14, /* Currently used for the global env. */
4561    TCG_REG_R15,
4562#else
4563    TCG_REG_EBP, /* Currently used for the global env. */
4564    TCG_REG_EBX,
4565    TCG_REG_ESI,
4566    TCG_REG_EDI,
4567#endif
4568};
4569
4570/* Compute frame size via macros, to share between tcg_target_qemu_prologue
4571   and tcg_register_jit.  */
4572
4573#define PUSH_SIZE \
4574    ((1 + ARRAY_SIZE(tcg_target_callee_save_regs)) \
4575     * (TCG_TARGET_REG_BITS / 8))
4576
4577#define FRAME_SIZE \
4578    ((PUSH_SIZE \
4579      + TCG_STATIC_CALL_ARGS_SIZE \
4580      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
4581      + TCG_TARGET_STACK_ALIGN - 1) \
4582     & ~(TCG_TARGET_STACK_ALIGN - 1))
4583
4584/* Generate global QEMU prologue and epilogue code */
4585static void tcg_target_qemu_prologue(TCGContext *s)
4586{
4587    int i, stack_addend;
4588
4589    /* TB prologue */
4590
4591    /* Reserve some stack space, also for TCG temps.  */
4592    stack_addend = FRAME_SIZE - PUSH_SIZE;
4593    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
4594                  CPU_TEMP_BUF_NLONGS * sizeof(long));
4595
4596    /* Save all callee saved registers.  */
4597    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
4598        tcg_out_push(s, tcg_target_callee_save_regs[i]);
4599    }
4600
4601    if (!tcg_use_softmmu && guest_base) {
4602        int seg = setup_guest_base_seg();
4603        if (seg != 0) {
4604            x86_guest_base.seg = seg;
4605        } else if (guest_base == (int32_t)guest_base) {
4606            x86_guest_base.ofs = guest_base;
4607        } else {
4608            assert(TCG_TARGET_REG_BITS == 64);
4609            /* Choose R12 because, as a base, it requires a SIB byte. */
4610            x86_guest_base.index = TCG_REG_R12;
4611            tcg_out_movi(s, TCG_TYPE_PTR, x86_guest_base.index, guest_base);
4612            tcg_regset_set_reg(s->reserved_regs, x86_guest_base.index);
4613        }
4614    }
4615
4616    if (TCG_TARGET_REG_BITS == 32) {
4617        tcg_out_ld(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP,
4618                   (ARRAY_SIZE(tcg_target_callee_save_regs) + 1) * 4);
4619        tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4620        /* jmp *tb.  */
4621        tcg_out_modrm_offset(s, OPC_GRP5, EXT5_JMPN_Ev, TCG_REG_ESP,
4622                             (ARRAY_SIZE(tcg_target_callee_save_regs) + 2) * 4
4623                             + stack_addend);
4624    } else {
4625        tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
4626        tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4627        /* jmp *tb.  */
4628        tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, tcg_target_call_iarg_regs[1]);
4629    }
4630
4631    /*
4632     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
4633     * and fall through to the rest of the epilogue.
4634     */
4635    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
4636    tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_EAX, 0);
4637
4638    /* TB epilogue */
4639    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
4640
4641    tcg_out_addi(s, TCG_REG_CALL_STACK, stack_addend);
4642
4643    if (have_avx2) {
4644        tcg_out_vex_opc(s, OPC_VZEROUPPER, 0, 0, 0, 0);
4645    }
4646    for (i = ARRAY_SIZE(tcg_target_callee_save_regs) - 1; i >= 0; i--) {
4647        tcg_out_pop(s, tcg_target_callee_save_regs[i]);
4648    }
4649    tcg_out_opc(s, OPC_RET, 0, 0, 0);
4650}
4651
4652static void tcg_out_tb_start(TCGContext *s)
4653{
4654    /* nothing to do */
4655}
4656
4657static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
4658{
4659    memset(p, 0x90, count);
4660}
4661
4662static void tcg_target_init(TCGContext *s)
4663{
4664    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
4665    if (TCG_TARGET_REG_BITS == 64) {
4666        tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
4667    }
4668    if (have_avx1) {
4669        tcg_target_available_regs[TCG_TYPE_V64] = ALL_VECTOR_REGS;
4670        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
4671    }
4672    if (have_avx2) {
4673        tcg_target_available_regs[TCG_TYPE_V256] = ALL_VECTOR_REGS;
4674    }
4675
4676    tcg_target_call_clobber_regs = ALL_VECTOR_REGS;
4677    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EAX);
4678    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EDX);
4679    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_ECX);
4680    if (TCG_TARGET_REG_BITS == 64) {
4681#if !defined(_WIN64)
4682        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RDI);
4683        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RSI);
4684#endif
4685        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R8);
4686        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R9);
4687        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R10);
4688        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R11);
4689    }
4690
4691    s->reserved_regs = 0;
4692    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
4693    tcg_regset_set_reg(s->reserved_regs, TCG_TMP_VEC);
4694#ifdef _WIN64
4695    /* These are call saved, and we don't save them, so don't use them. */
4696    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM6);
4697    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM7);
4698    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM8);
4699    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM9);
4700    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM10);
4701    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM11);
4702    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM12);
4703    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM13);
4704    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM14);
4705    tcg_regset_set_reg(s->reserved_regs, TCG_REG_XMM15);
4706#endif
4707}
4708
4709typedef struct {
4710    DebugFrameHeader h;
4711    uint8_t fde_def_cfa[4];
4712    uint8_t fde_reg_ofs[14];
4713} DebugFrame;
4714
4715/* We're expecting a 2 byte uleb128 encoded value.  */
4716QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
4717
4718#if !defined(__ELF__)
4719    /* Host machine without ELF. */
4720#elif TCG_TARGET_REG_BITS == 64
4721#define ELF_HOST_MACHINE EM_X86_64
4722static const DebugFrame debug_frame = {
4723    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4724    .h.cie.id = -1,
4725    .h.cie.version = 1,
4726    .h.cie.code_align = 1,
4727    .h.cie.data_align = 0x78,             /* sleb128 -8 */
4728    .h.cie.return_column = 16,
4729
4730    /* Total FDE size does not include the "len" member.  */
4731    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4732
4733    .fde_def_cfa = {
4734        12, 7,                          /* DW_CFA_def_cfa %rsp, ... */
4735        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4736        (FRAME_SIZE >> 7)
4737    },
4738    .fde_reg_ofs = {
4739        0x90, 1,                        /* DW_CFA_offset, %rip, -8 */
4740        /* The following ordering must match tcg_target_callee_save_regs.  */
4741        0x86, 2,                        /* DW_CFA_offset, %rbp, -16 */
4742        0x83, 3,                        /* DW_CFA_offset, %rbx, -24 */
4743        0x8c, 4,                        /* DW_CFA_offset, %r12, -32 */
4744        0x8d, 5,                        /* DW_CFA_offset, %r13, -40 */
4745        0x8e, 6,                        /* DW_CFA_offset, %r14, -48 */
4746        0x8f, 7,                        /* DW_CFA_offset, %r15, -56 */
4747    }
4748};
4749#else
4750#define ELF_HOST_MACHINE EM_386
4751static const DebugFrame debug_frame = {
4752    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4753    .h.cie.id = -1,
4754    .h.cie.version = 1,
4755    .h.cie.code_align = 1,
4756    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
4757    .h.cie.return_column = 8,
4758
4759    /* Total FDE size does not include the "len" member.  */
4760    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4761
4762    .fde_def_cfa = {
4763        12, 4,                          /* DW_CFA_def_cfa %esp, ... */
4764        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4765        (FRAME_SIZE >> 7)
4766    },
4767    .fde_reg_ofs = {
4768        0x88, 1,                        /* DW_CFA_offset, %eip, -4 */
4769        /* The following ordering must match tcg_target_callee_save_regs.  */
4770        0x85, 2,                        /* DW_CFA_offset, %ebp, -8 */
4771        0x83, 3,                        /* DW_CFA_offset, %ebx, -12 */
4772        0x86, 4,                        /* DW_CFA_offset, %esi, -16 */
4773        0x87, 5,                        /* DW_CFA_offset, %edi, -20 */
4774    }
4775};
4776#endif
4777
4778#if defined(ELF_HOST_MACHINE)
4779void tcg_register_jit(const void *buf, size_t buf_size)
4780{
4781    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
4782}
4783#endif
4784