xref: /openbmc/qemu/tcg/i386/tcg-target.c.inc (revision ffe98631)
1/*
2 * Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2008 Fabrice Bellard
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to deal
8 * in the Software without restriction, including without limitation the rights
9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10 * copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22 * THE SOFTWARE.
23 */
24
25#include "../tcg-ldst.c.inc"
26#include "../tcg-pool.c.inc"
27
28#ifdef CONFIG_DEBUG_TCG
29static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
30#if TCG_TARGET_REG_BITS == 64
31    "%rax", "%rcx", "%rdx", "%rbx", "%rsp", "%rbp", "%rsi", "%rdi",
32#else
33    "%eax", "%ecx", "%edx", "%ebx", "%esp", "%ebp", "%esi", "%edi",
34#endif
35    "%r8",  "%r9",  "%r10", "%r11", "%r12", "%r13", "%r14", "%r15",
36    "%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5", "%xmm6", "%xmm7",
37#if TCG_TARGET_REG_BITS == 64
38    "%xmm8", "%xmm9", "%xmm10", "%xmm11",
39    "%xmm12", "%xmm13", "%xmm14", "%xmm15",
40#endif
41};
42#endif
43
44static const int tcg_target_reg_alloc_order[] = {
45#if TCG_TARGET_REG_BITS == 64
46    TCG_REG_RBP,
47    TCG_REG_RBX,
48    TCG_REG_R12,
49    TCG_REG_R13,
50    TCG_REG_R14,
51    TCG_REG_R15,
52    TCG_REG_R10,
53    TCG_REG_R11,
54    TCG_REG_R9,
55    TCG_REG_R8,
56    TCG_REG_RCX,
57    TCG_REG_RDX,
58    TCG_REG_RSI,
59    TCG_REG_RDI,
60    TCG_REG_RAX,
61#else
62    TCG_REG_EBX,
63    TCG_REG_ESI,
64    TCG_REG_EDI,
65    TCG_REG_EBP,
66    TCG_REG_ECX,
67    TCG_REG_EDX,
68    TCG_REG_EAX,
69#endif
70    TCG_REG_XMM0,
71    TCG_REG_XMM1,
72    TCG_REG_XMM2,
73    TCG_REG_XMM3,
74    TCG_REG_XMM4,
75    TCG_REG_XMM5,
76#ifndef _WIN64
77    /* The Win64 ABI has xmm6-xmm15 as caller-saves, and we do not save
78       any of them.  Therefore only allow xmm0-xmm5 to be allocated.  */
79    TCG_REG_XMM6,
80    TCG_REG_XMM7,
81#if TCG_TARGET_REG_BITS == 64
82    TCG_REG_XMM8,
83    TCG_REG_XMM9,
84    TCG_REG_XMM10,
85    TCG_REG_XMM11,
86    TCG_REG_XMM12,
87    TCG_REG_XMM13,
88    TCG_REG_XMM14,
89    TCG_REG_XMM15,
90#endif
91#endif
92};
93
94static const int tcg_target_call_iarg_regs[] = {
95#if TCG_TARGET_REG_BITS == 64
96#if defined(_WIN64)
97    TCG_REG_RCX,
98    TCG_REG_RDX,
99#else
100    TCG_REG_RDI,
101    TCG_REG_RSI,
102    TCG_REG_RDX,
103    TCG_REG_RCX,
104#endif
105    TCG_REG_R8,
106    TCG_REG_R9,
107#else
108    /* 32 bit mode uses stack based calling convention (GCC default). */
109#endif
110};
111
112static const int tcg_target_call_oarg_regs[] = {
113    TCG_REG_EAX,
114#if TCG_TARGET_REG_BITS == 32
115    TCG_REG_EDX
116#endif
117};
118
119/* Constants we accept.  */
120#define TCG_CT_CONST_S32 0x100
121#define TCG_CT_CONST_U32 0x200
122#define TCG_CT_CONST_I32 0x400
123#define TCG_CT_CONST_WSZ 0x800
124
125/* Registers used with L constraint, which are the first argument
126   registers on x86_64, and two random call clobbered registers on
127   i386. */
128#if TCG_TARGET_REG_BITS == 64
129# define TCG_REG_L0 tcg_target_call_iarg_regs[0]
130# define TCG_REG_L1 tcg_target_call_iarg_regs[1]
131#else
132# define TCG_REG_L0 TCG_REG_EAX
133# define TCG_REG_L1 TCG_REG_EDX
134#endif
135
136#define ALL_BYTEH_REGS         0x0000000fu
137#if TCG_TARGET_REG_BITS == 64
138# define ALL_GENERAL_REGS      0x0000ffffu
139# define ALL_VECTOR_REGS       0xffff0000u
140# define ALL_BYTEL_REGS        ALL_GENERAL_REGS
141#else
142# define ALL_GENERAL_REGS      0x000000ffu
143# define ALL_VECTOR_REGS       0x00ff0000u
144# define ALL_BYTEL_REGS        ALL_BYTEH_REGS
145#endif
146#ifdef CONFIG_SOFTMMU
147# define SOFTMMU_RESERVE_REGS  ((1 << TCG_REG_L0) | (1 << TCG_REG_L1))
148#else
149# define SOFTMMU_RESERVE_REGS  0
150#endif
151
152/* The host compiler should supply <cpuid.h> to enable runtime features
153   detection, as we're not going to go so far as our own inline assembly.
154   If not available, default values will be assumed.  */
155#if defined(CONFIG_CPUID_H)
156#include "qemu/cpuid.h"
157#endif
158
159/* For 64-bit, we always know that CMOV is available.  */
160#if TCG_TARGET_REG_BITS == 64
161# define have_cmov 1
162#elif defined(CONFIG_CPUID_H)
163static bool have_cmov;
164#else
165# define have_cmov 0
166#endif
167
168/* We need these symbols in tcg-target.h, and we can't properly conditionalize
169   it there.  Therefore we always define the variable.  */
170bool have_bmi1;
171bool have_popcnt;
172bool have_avx1;
173bool have_avx2;
174bool have_avx512bw;
175bool have_avx512dq;
176bool have_avx512vbmi2;
177bool have_avx512vl;
178bool have_movbe;
179
180#ifdef CONFIG_CPUID_H
181static bool have_bmi2;
182static bool have_lzcnt;
183#else
184# define have_bmi2 0
185# define have_lzcnt 0
186#endif
187
188static const tcg_insn_unit *tb_ret_addr;
189
190static bool patch_reloc(tcg_insn_unit *code_ptr, int type,
191                        intptr_t value, intptr_t addend)
192{
193    value += addend;
194    switch(type) {
195    case R_386_PC32:
196        value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
197        if (value != (int32_t)value) {
198            return false;
199        }
200        /* FALLTHRU */
201    case R_386_32:
202        tcg_patch32(code_ptr, value);
203        break;
204    case R_386_PC8:
205        value -= (uintptr_t)tcg_splitwx_to_rx(code_ptr);
206        if (value != (int8_t)value) {
207            return false;
208        }
209        tcg_patch8(code_ptr, value);
210        break;
211    default:
212        tcg_abort();
213    }
214    return true;
215}
216
217/* test if a constant matches the constraint */
218static bool tcg_target_const_match(int64_t val, TCGType type, int ct)
219{
220    if (ct & TCG_CT_CONST) {
221        return 1;
222    }
223    if (type == TCG_TYPE_I32) {
224        if (ct & (TCG_CT_CONST_S32 | TCG_CT_CONST_U32 | TCG_CT_CONST_I32)) {
225            return 1;
226        }
227    } else {
228        if ((ct & TCG_CT_CONST_S32) && val == (int32_t)val) {
229            return 1;
230        }
231        if ((ct & TCG_CT_CONST_U32) && val == (uint32_t)val) {
232            return 1;
233        }
234        if ((ct & TCG_CT_CONST_I32) && ~val == (int32_t)~val) {
235            return 1;
236        }
237    }
238    if ((ct & TCG_CT_CONST_WSZ) && val == (type == TCG_TYPE_I32 ? 32 : 64)) {
239        return 1;
240    }
241    return 0;
242}
243
244# define LOWREGMASK(x)	((x) & 7)
245
246#define P_EXT		0x100		/* 0x0f opcode prefix */
247#define P_EXT38         0x200           /* 0x0f 0x38 opcode prefix */
248#define P_DATA16        0x400           /* 0x66 opcode prefix */
249#define P_VEXW          0x1000          /* Set VEX.W = 1 */
250#if TCG_TARGET_REG_BITS == 64
251# define P_REXW         P_VEXW          /* Set REX.W = 1; match VEXW */
252# define P_REXB_R       0x2000          /* REG field as byte register */
253# define P_REXB_RM      0x4000          /* R/M field as byte register */
254# define P_GS           0x8000          /* gs segment override */
255#else
256# define P_REXW		0
257# define P_REXB_R	0
258# define P_REXB_RM	0
259# define P_GS           0
260#endif
261#define P_EXT3A         0x10000         /* 0x0f 0x3a opcode prefix */
262#define P_SIMDF3        0x20000         /* 0xf3 opcode prefix */
263#define P_SIMDF2        0x40000         /* 0xf2 opcode prefix */
264#define P_VEXL          0x80000         /* Set VEX.L = 1 */
265#define P_EVEX          0x100000        /* Requires EVEX encoding */
266
267#define OPC_ARITH_EvIz	(0x81)
268#define OPC_ARITH_EvIb	(0x83)
269#define OPC_ARITH_GvEv	(0x03)		/* ... plus (ARITH_FOO << 3) */
270#define OPC_ANDN        (0xf2 | P_EXT38)
271#define OPC_ADD_GvEv	(OPC_ARITH_GvEv | (ARITH_ADD << 3))
272#define OPC_AND_GvEv    (OPC_ARITH_GvEv | (ARITH_AND << 3))
273#define OPC_BLENDPS     (0x0c | P_EXT3A | P_DATA16)
274#define OPC_BSF         (0xbc | P_EXT)
275#define OPC_BSR         (0xbd | P_EXT)
276#define OPC_BSWAP	(0xc8 | P_EXT)
277#define OPC_CALL_Jz	(0xe8)
278#define OPC_CMOVCC      (0x40 | P_EXT)  /* ... plus condition code */
279#define OPC_CMP_GvEv	(OPC_ARITH_GvEv | (ARITH_CMP << 3))
280#define OPC_DEC_r32	(0x48)
281#define OPC_IMUL_GvEv	(0xaf | P_EXT)
282#define OPC_IMUL_GvEvIb	(0x6b)
283#define OPC_IMUL_GvEvIz	(0x69)
284#define OPC_INC_r32	(0x40)
285#define OPC_JCC_long	(0x80 | P_EXT)	/* ... plus condition code */
286#define OPC_JCC_short	(0x70)		/* ... plus condition code */
287#define OPC_JMP_long	(0xe9)
288#define OPC_JMP_short	(0xeb)
289#define OPC_LEA         (0x8d)
290#define OPC_LZCNT       (0xbd | P_EXT | P_SIMDF3)
291#define OPC_MOVB_EvGv	(0x88)		/* stores, more or less */
292#define OPC_MOVL_EvGv	(0x89)		/* stores, more or less */
293#define OPC_MOVL_GvEv	(0x8b)		/* loads, more or less */
294#define OPC_MOVB_EvIz   (0xc6)
295#define OPC_MOVL_EvIz	(0xc7)
296#define OPC_MOVL_Iv     (0xb8)
297#define OPC_MOVBE_GyMy  (0xf0 | P_EXT38)
298#define OPC_MOVBE_MyGy  (0xf1 | P_EXT38)
299#define OPC_MOVD_VyEy   (0x6e | P_EXT | P_DATA16)
300#define OPC_MOVD_EyVy   (0x7e | P_EXT | P_DATA16)
301#define OPC_MOVDDUP     (0x12 | P_EXT | P_SIMDF2)
302#define OPC_MOVDQA_VxWx (0x6f | P_EXT | P_DATA16)
303#define OPC_MOVDQA_WxVx (0x7f | P_EXT | P_DATA16)
304#define OPC_MOVDQU_VxWx (0x6f | P_EXT | P_SIMDF3)
305#define OPC_MOVDQU_WxVx (0x7f | P_EXT | P_SIMDF3)
306#define OPC_MOVQ_VqWq   (0x7e | P_EXT | P_SIMDF3)
307#define OPC_MOVQ_WqVq   (0xd6 | P_EXT | P_DATA16)
308#define OPC_MOVSBL	(0xbe | P_EXT)
309#define OPC_MOVSWL	(0xbf | P_EXT)
310#define OPC_MOVSLQ	(0x63 | P_REXW)
311#define OPC_MOVZBL	(0xb6 | P_EXT)
312#define OPC_MOVZWL	(0xb7 | P_EXT)
313#define OPC_PABSB       (0x1c | P_EXT38 | P_DATA16)
314#define OPC_PABSW       (0x1d | P_EXT38 | P_DATA16)
315#define OPC_PABSD       (0x1e | P_EXT38 | P_DATA16)
316#define OPC_VPABSQ      (0x1f | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
317#define OPC_PACKSSDW    (0x6b | P_EXT | P_DATA16)
318#define OPC_PACKSSWB    (0x63 | P_EXT | P_DATA16)
319#define OPC_PACKUSDW    (0x2b | P_EXT38 | P_DATA16)
320#define OPC_PACKUSWB    (0x67 | P_EXT | P_DATA16)
321#define OPC_PADDB       (0xfc | P_EXT | P_DATA16)
322#define OPC_PADDW       (0xfd | P_EXT | P_DATA16)
323#define OPC_PADDD       (0xfe | P_EXT | P_DATA16)
324#define OPC_PADDQ       (0xd4 | P_EXT | P_DATA16)
325#define OPC_PADDSB      (0xec | P_EXT | P_DATA16)
326#define OPC_PADDSW      (0xed | P_EXT | P_DATA16)
327#define OPC_PADDUB      (0xdc | P_EXT | P_DATA16)
328#define OPC_PADDUW      (0xdd | P_EXT | P_DATA16)
329#define OPC_PAND        (0xdb | P_EXT | P_DATA16)
330#define OPC_PANDN       (0xdf | P_EXT | P_DATA16)
331#define OPC_PBLENDW     (0x0e | P_EXT3A | P_DATA16)
332#define OPC_PCMPEQB     (0x74 | P_EXT | P_DATA16)
333#define OPC_PCMPEQW     (0x75 | P_EXT | P_DATA16)
334#define OPC_PCMPEQD     (0x76 | P_EXT | P_DATA16)
335#define OPC_PCMPEQQ     (0x29 | P_EXT38 | P_DATA16)
336#define OPC_PCMPGTB     (0x64 | P_EXT | P_DATA16)
337#define OPC_PCMPGTW     (0x65 | P_EXT | P_DATA16)
338#define OPC_PCMPGTD     (0x66 | P_EXT | P_DATA16)
339#define OPC_PCMPGTQ     (0x37 | P_EXT38 | P_DATA16)
340#define OPC_PMAXSB      (0x3c | P_EXT38 | P_DATA16)
341#define OPC_PMAXSW      (0xee | P_EXT | P_DATA16)
342#define OPC_PMAXSD      (0x3d | P_EXT38 | P_DATA16)
343#define OPC_VPMAXSQ     (0x3d | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
344#define OPC_PMAXUB      (0xde | P_EXT | P_DATA16)
345#define OPC_PMAXUW      (0x3e | P_EXT38 | P_DATA16)
346#define OPC_PMAXUD      (0x3f | P_EXT38 | P_DATA16)
347#define OPC_VPMAXUQ     (0x3f | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
348#define OPC_PMINSB      (0x38 | P_EXT38 | P_DATA16)
349#define OPC_PMINSW      (0xea | P_EXT | P_DATA16)
350#define OPC_PMINSD      (0x39 | P_EXT38 | P_DATA16)
351#define OPC_VPMINSQ     (0x39 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
352#define OPC_PMINUB      (0xda | P_EXT | P_DATA16)
353#define OPC_PMINUW      (0x3a | P_EXT38 | P_DATA16)
354#define OPC_PMINUD      (0x3b | P_EXT38 | P_DATA16)
355#define OPC_VPMINUQ     (0x3b | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
356#define OPC_PMOVSXBW    (0x20 | P_EXT38 | P_DATA16)
357#define OPC_PMOVSXWD    (0x23 | P_EXT38 | P_DATA16)
358#define OPC_PMOVSXDQ    (0x25 | P_EXT38 | P_DATA16)
359#define OPC_PMOVZXBW    (0x30 | P_EXT38 | P_DATA16)
360#define OPC_PMOVZXWD    (0x33 | P_EXT38 | P_DATA16)
361#define OPC_PMOVZXDQ    (0x35 | P_EXT38 | P_DATA16)
362#define OPC_PMULLW      (0xd5 | P_EXT | P_DATA16)
363#define OPC_PMULLD      (0x40 | P_EXT38 | P_DATA16)
364#define OPC_VPMULLQ     (0x40 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
365#define OPC_POR         (0xeb | P_EXT | P_DATA16)
366#define OPC_PSHUFB      (0x00 | P_EXT38 | P_DATA16)
367#define OPC_PSHUFD      (0x70 | P_EXT | P_DATA16)
368#define OPC_PSHUFLW     (0x70 | P_EXT | P_SIMDF2)
369#define OPC_PSHUFHW     (0x70 | P_EXT | P_SIMDF3)
370#define OPC_PSHIFTW_Ib  (0x71 | P_EXT | P_DATA16) /* /2 /6 /4 */
371#define OPC_PSHIFTD_Ib  (0x72 | P_EXT | P_DATA16) /* /1 /2 /6 /4 */
372#define OPC_PSHIFTQ_Ib  (0x73 | P_EXT | P_DATA16) /* /2 /6 /4 */
373#define OPC_PSLLW       (0xf1 | P_EXT | P_DATA16)
374#define OPC_PSLLD       (0xf2 | P_EXT | P_DATA16)
375#define OPC_PSLLQ       (0xf3 | P_EXT | P_DATA16)
376#define OPC_PSRAW       (0xe1 | P_EXT | P_DATA16)
377#define OPC_PSRAD       (0xe2 | P_EXT | P_DATA16)
378#define OPC_VPSRAQ      (0xe2 | P_EXT | P_DATA16 | P_VEXW | P_EVEX)
379#define OPC_PSRLW       (0xd1 | P_EXT | P_DATA16)
380#define OPC_PSRLD       (0xd2 | P_EXT | P_DATA16)
381#define OPC_PSRLQ       (0xd3 | P_EXT | P_DATA16)
382#define OPC_PSUBB       (0xf8 | P_EXT | P_DATA16)
383#define OPC_PSUBW       (0xf9 | P_EXT | P_DATA16)
384#define OPC_PSUBD       (0xfa | P_EXT | P_DATA16)
385#define OPC_PSUBQ       (0xfb | P_EXT | P_DATA16)
386#define OPC_PSUBSB      (0xe8 | P_EXT | P_DATA16)
387#define OPC_PSUBSW      (0xe9 | P_EXT | P_DATA16)
388#define OPC_PSUBUB      (0xd8 | P_EXT | P_DATA16)
389#define OPC_PSUBUW      (0xd9 | P_EXT | P_DATA16)
390#define OPC_PUNPCKLBW   (0x60 | P_EXT | P_DATA16)
391#define OPC_PUNPCKLWD   (0x61 | P_EXT | P_DATA16)
392#define OPC_PUNPCKLDQ   (0x62 | P_EXT | P_DATA16)
393#define OPC_PUNPCKLQDQ  (0x6c | P_EXT | P_DATA16)
394#define OPC_PUNPCKHBW   (0x68 | P_EXT | P_DATA16)
395#define OPC_PUNPCKHWD   (0x69 | P_EXT | P_DATA16)
396#define OPC_PUNPCKHDQ   (0x6a | P_EXT | P_DATA16)
397#define OPC_PUNPCKHQDQ  (0x6d | P_EXT | P_DATA16)
398#define OPC_PXOR        (0xef | P_EXT | P_DATA16)
399#define OPC_POP_r32	(0x58)
400#define OPC_POPCNT      (0xb8 | P_EXT | P_SIMDF3)
401#define OPC_PUSH_r32	(0x50)
402#define OPC_PUSH_Iv	(0x68)
403#define OPC_PUSH_Ib	(0x6a)
404#define OPC_RET		(0xc3)
405#define OPC_SETCC	(0x90 | P_EXT | P_REXB_RM) /* ... plus cc */
406#define OPC_SHIFT_1	(0xd1)
407#define OPC_SHIFT_Ib	(0xc1)
408#define OPC_SHIFT_cl	(0xd3)
409#define OPC_SARX        (0xf7 | P_EXT38 | P_SIMDF3)
410#define OPC_SHUFPS      (0xc6 | P_EXT)
411#define OPC_SHLX        (0xf7 | P_EXT38 | P_DATA16)
412#define OPC_SHRX        (0xf7 | P_EXT38 | P_SIMDF2)
413#define OPC_SHRD_Ib     (0xac | P_EXT)
414#define OPC_TESTL	(0x85)
415#define OPC_TZCNT       (0xbc | P_EXT | P_SIMDF3)
416#define OPC_UD2         (0x0b | P_EXT)
417#define OPC_VPBLENDD    (0x02 | P_EXT3A | P_DATA16)
418#define OPC_VPBLENDVB   (0x4c | P_EXT3A | P_DATA16)
419#define OPC_VPINSRB     (0x20 | P_EXT3A | P_DATA16)
420#define OPC_VPINSRW     (0xc4 | P_EXT | P_DATA16)
421#define OPC_VBROADCASTSS (0x18 | P_EXT38 | P_DATA16)
422#define OPC_VBROADCASTSD (0x19 | P_EXT38 | P_DATA16)
423#define OPC_VPBROADCASTB (0x78 | P_EXT38 | P_DATA16)
424#define OPC_VPBROADCASTW (0x79 | P_EXT38 | P_DATA16)
425#define OPC_VPBROADCASTD (0x58 | P_EXT38 | P_DATA16)
426#define OPC_VPBROADCASTQ (0x59 | P_EXT38 | P_DATA16)
427#define OPC_VPERMQ      (0x00 | P_EXT3A | P_DATA16 | P_VEXW)
428#define OPC_VPERM2I128  (0x46 | P_EXT3A | P_DATA16 | P_VEXL)
429#define OPC_VPROLVD     (0x15 | P_EXT38 | P_DATA16 | P_EVEX)
430#define OPC_VPROLVQ     (0x15 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
431#define OPC_VPRORVD     (0x14 | P_EXT38 | P_DATA16 | P_EVEX)
432#define OPC_VPRORVQ     (0x14 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
433#define OPC_VPSHLDW     (0x70 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
434#define OPC_VPSHLDD     (0x71 | P_EXT3A | P_DATA16 | P_EVEX)
435#define OPC_VPSHLDQ     (0x71 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
436#define OPC_VPSHLDVW    (0x70 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
437#define OPC_VPSHLDVD    (0x71 | P_EXT38 | P_DATA16 | P_EVEX)
438#define OPC_VPSHLDVQ    (0x71 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
439#define OPC_VPSHRDVW    (0x72 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
440#define OPC_VPSHRDVD    (0x73 | P_EXT38 | P_DATA16 | P_EVEX)
441#define OPC_VPSHRDVQ    (0x73 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
442#define OPC_VPSLLVW     (0x12 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
443#define OPC_VPSLLVD     (0x47 | P_EXT38 | P_DATA16)
444#define OPC_VPSLLVQ     (0x47 | P_EXT38 | P_DATA16 | P_VEXW)
445#define OPC_VPSRAVW     (0x11 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
446#define OPC_VPSRAVD     (0x46 | P_EXT38 | P_DATA16)
447#define OPC_VPSRAVQ     (0x46 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
448#define OPC_VPSRLVW     (0x10 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
449#define OPC_VPSRLVD     (0x45 | P_EXT38 | P_DATA16)
450#define OPC_VPSRLVQ     (0x45 | P_EXT38 | P_DATA16 | P_VEXW)
451#define OPC_VPTERNLOGQ  (0x25 | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
452#define OPC_VZEROUPPER  (0x77 | P_EXT)
453#define OPC_XCHG_ax_r32	(0x90)
454
455#define OPC_GRP3_Eb     (0xf6)
456#define OPC_GRP3_Ev     (0xf7)
457#define OPC_GRP5        (0xff)
458#define OPC_GRP14       (0x73 | P_EXT | P_DATA16)
459
460/* Group 1 opcode extensions for 0x80-0x83.
461   These are also used as modifiers for OPC_ARITH.  */
462#define ARITH_ADD 0
463#define ARITH_OR  1
464#define ARITH_ADC 2
465#define ARITH_SBB 3
466#define ARITH_AND 4
467#define ARITH_SUB 5
468#define ARITH_XOR 6
469#define ARITH_CMP 7
470
471/* Group 2 opcode extensions for 0xc0, 0xc1, 0xd0-0xd3.  */
472#define SHIFT_ROL 0
473#define SHIFT_ROR 1
474#define SHIFT_SHL 4
475#define SHIFT_SHR 5
476#define SHIFT_SAR 7
477
478/* Group 3 opcode extensions for 0xf6, 0xf7.  To be used with OPC_GRP3.  */
479#define EXT3_TESTi 0
480#define EXT3_NOT   2
481#define EXT3_NEG   3
482#define EXT3_MUL   4
483#define EXT3_IMUL  5
484#define EXT3_DIV   6
485#define EXT3_IDIV  7
486
487/* Group 5 opcode extensions for 0xff.  To be used with OPC_GRP5.  */
488#define EXT5_INC_Ev	0
489#define EXT5_DEC_Ev	1
490#define EXT5_CALLN_Ev	2
491#define EXT5_JMPN_Ev	4
492
493/* Condition codes to be added to OPC_JCC_{long,short}.  */
494#define JCC_JMP (-1)
495#define JCC_JO  0x0
496#define JCC_JNO 0x1
497#define JCC_JB  0x2
498#define JCC_JAE 0x3
499#define JCC_JE  0x4
500#define JCC_JNE 0x5
501#define JCC_JBE 0x6
502#define JCC_JA  0x7
503#define JCC_JS  0x8
504#define JCC_JNS 0x9
505#define JCC_JP  0xa
506#define JCC_JNP 0xb
507#define JCC_JL  0xc
508#define JCC_JGE 0xd
509#define JCC_JLE 0xe
510#define JCC_JG  0xf
511
512static const uint8_t tcg_cond_to_jcc[] = {
513    [TCG_COND_EQ] = JCC_JE,
514    [TCG_COND_NE] = JCC_JNE,
515    [TCG_COND_LT] = JCC_JL,
516    [TCG_COND_GE] = JCC_JGE,
517    [TCG_COND_LE] = JCC_JLE,
518    [TCG_COND_GT] = JCC_JG,
519    [TCG_COND_LTU] = JCC_JB,
520    [TCG_COND_GEU] = JCC_JAE,
521    [TCG_COND_LEU] = JCC_JBE,
522    [TCG_COND_GTU] = JCC_JA,
523};
524
525#if TCG_TARGET_REG_BITS == 64
526static void tcg_out_opc(TCGContext *s, int opc, int r, int rm, int x)
527{
528    int rex;
529
530    if (opc & P_GS) {
531        tcg_out8(s, 0x65);
532    }
533    if (opc & P_DATA16) {
534        /* We should never be asking for both 16 and 64-bit operation.  */
535        tcg_debug_assert((opc & P_REXW) == 0);
536        tcg_out8(s, 0x66);
537    }
538    if (opc & P_SIMDF3) {
539        tcg_out8(s, 0xf3);
540    } else if (opc & P_SIMDF2) {
541        tcg_out8(s, 0xf2);
542    }
543
544    rex = 0;
545    rex |= (opc & P_REXW) ? 0x8 : 0x0;  /* REX.W */
546    rex |= (r & 8) >> 1;                /* REX.R */
547    rex |= (x & 8) >> 2;                /* REX.X */
548    rex |= (rm & 8) >> 3;               /* REX.B */
549
550    /* P_REXB_{R,RM} indicates that the given register is the low byte.
551       For %[abcd]l we need no REX prefix, but for %{si,di,bp,sp}l we do,
552       as otherwise the encoding indicates %[abcd]h.  Note that the values
553       that are ORed in merely indicate that the REX byte must be present;
554       those bits get discarded in output.  */
555    rex |= opc & (r >= 4 ? P_REXB_R : 0);
556    rex |= opc & (rm >= 4 ? P_REXB_RM : 0);
557
558    if (rex) {
559        tcg_out8(s, (uint8_t)(rex | 0x40));
560    }
561
562    if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
563        tcg_out8(s, 0x0f);
564        if (opc & P_EXT38) {
565            tcg_out8(s, 0x38);
566        } else if (opc & P_EXT3A) {
567            tcg_out8(s, 0x3a);
568        }
569    }
570
571    tcg_out8(s, opc);
572}
573#else
574static void tcg_out_opc(TCGContext *s, int opc)
575{
576    if (opc & P_DATA16) {
577        tcg_out8(s, 0x66);
578    }
579    if (opc & P_SIMDF3) {
580        tcg_out8(s, 0xf3);
581    } else if (opc & P_SIMDF2) {
582        tcg_out8(s, 0xf2);
583    }
584    if (opc & (P_EXT | P_EXT38 | P_EXT3A)) {
585        tcg_out8(s, 0x0f);
586        if (opc & P_EXT38) {
587            tcg_out8(s, 0x38);
588        } else if (opc & P_EXT3A) {
589            tcg_out8(s, 0x3a);
590        }
591    }
592    tcg_out8(s, opc);
593}
594/* Discard the register arguments to tcg_out_opc early, so as not to penalize
595   the 32-bit compilation paths.  This method works with all versions of gcc,
596   whereas relying on optimization may not be able to exclude them.  */
597#define tcg_out_opc(s, opc, r, rm, x)  (tcg_out_opc)(s, opc)
598#endif
599
600static void tcg_out_modrm(TCGContext *s, int opc, int r, int rm)
601{
602    tcg_out_opc(s, opc, r, rm, 0);
603    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
604}
605
606static void tcg_out_vex_opc(TCGContext *s, int opc, int r, int v,
607                            int rm, int index)
608{
609    int tmp;
610
611    /* Use the two byte form if possible, which cannot encode
612       VEX.W, VEX.B, VEX.X, or an m-mmmm field other than P_EXT.  */
613    if ((opc & (P_EXT | P_EXT38 | P_EXT3A | P_VEXW)) == P_EXT
614        && ((rm | index) & 8) == 0) {
615        /* Two byte VEX prefix.  */
616        tcg_out8(s, 0xc5);
617
618        tmp = (r & 8 ? 0 : 0x80);              /* VEX.R */
619    } else {
620        /* Three byte VEX prefix.  */
621        tcg_out8(s, 0xc4);
622
623        /* VEX.m-mmmm */
624        if (opc & P_EXT3A) {
625            tmp = 3;
626        } else if (opc & P_EXT38) {
627            tmp = 2;
628        } else if (opc & P_EXT) {
629            tmp = 1;
630        } else {
631            g_assert_not_reached();
632        }
633        tmp |= (r & 8 ? 0 : 0x80);             /* VEX.R */
634        tmp |= (index & 8 ? 0 : 0x40);         /* VEX.X */
635        tmp |= (rm & 8 ? 0 : 0x20);            /* VEX.B */
636        tcg_out8(s, tmp);
637
638        tmp = (opc & P_VEXW ? 0x80 : 0);       /* VEX.W */
639    }
640
641    tmp |= (opc & P_VEXL ? 0x04 : 0);      /* VEX.L */
642    /* VEX.pp */
643    if (opc & P_DATA16) {
644        tmp |= 1;                          /* 0x66 */
645    } else if (opc & P_SIMDF3) {
646        tmp |= 2;                          /* 0xf3 */
647    } else if (opc & P_SIMDF2) {
648        tmp |= 3;                          /* 0xf2 */
649    }
650    tmp |= (~v & 15) << 3;                 /* VEX.vvvv */
651    tcg_out8(s, tmp);
652    tcg_out8(s, opc);
653}
654
655static void tcg_out_evex_opc(TCGContext *s, int opc, int r, int v,
656                             int rm, int index)
657{
658    /* The entire 4-byte evex prefix; with R' and V' set. */
659    uint32_t p = 0x08041062;
660    int mm, pp;
661
662    tcg_debug_assert(have_avx512vl);
663
664    /* EVEX.mm */
665    if (opc & P_EXT3A) {
666        mm = 3;
667    } else if (opc & P_EXT38) {
668        mm = 2;
669    } else if (opc & P_EXT) {
670        mm = 1;
671    } else {
672        g_assert_not_reached();
673    }
674
675    /* EVEX.pp */
676    if (opc & P_DATA16) {
677        pp = 1;                          /* 0x66 */
678    } else if (opc & P_SIMDF3) {
679        pp = 2;                          /* 0xf3 */
680    } else if (opc & P_SIMDF2) {
681        pp = 3;                          /* 0xf2 */
682    } else {
683        pp = 0;
684    }
685
686    p = deposit32(p, 8, 2, mm);
687    p = deposit32(p, 13, 1, (rm & 8) == 0);             /* EVEX.RXB.B */
688    p = deposit32(p, 14, 1, (index & 8) == 0);          /* EVEX.RXB.X */
689    p = deposit32(p, 15, 1, (r & 8) == 0);              /* EVEX.RXB.R */
690    p = deposit32(p, 16, 2, pp);
691    p = deposit32(p, 19, 4, ~v);
692    p = deposit32(p, 23, 1, (opc & P_VEXW) != 0);
693    p = deposit32(p, 29, 2, (opc & P_VEXL) != 0);
694
695    tcg_out32(s, p);
696    tcg_out8(s, opc);
697}
698
699static void tcg_out_vex_modrm(TCGContext *s, int opc, int r, int v, int rm)
700{
701    if (opc & P_EVEX) {
702        tcg_out_evex_opc(s, opc, r, v, rm, 0);
703    } else {
704        tcg_out_vex_opc(s, opc, r, v, rm, 0);
705    }
706    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
707}
708
709/* Output an opcode with a full "rm + (index<<shift) + offset" address mode.
710   We handle either RM and INDEX missing with a negative value.  In 64-bit
711   mode for absolute addresses, ~RM is the size of the immediate operand
712   that will follow the instruction.  */
713
714static void tcg_out_sib_offset(TCGContext *s, int r, int rm, int index,
715                               int shift, intptr_t offset)
716{
717    int mod, len;
718
719    if (index < 0 && rm < 0) {
720        if (TCG_TARGET_REG_BITS == 64) {
721            /* Try for a rip-relative addressing mode.  This has replaced
722               the 32-bit-mode absolute addressing encoding.  */
723            intptr_t pc = (intptr_t)s->code_ptr + 5 + ~rm;
724            intptr_t disp = offset - pc;
725            if (disp == (int32_t)disp) {
726                tcg_out8(s, (LOWREGMASK(r) << 3) | 5);
727                tcg_out32(s, disp);
728                return;
729            }
730
731            /* Try for an absolute address encoding.  This requires the
732               use of the MODRM+SIB encoding and is therefore larger than
733               rip-relative addressing.  */
734            if (offset == (int32_t)offset) {
735                tcg_out8(s, (LOWREGMASK(r) << 3) | 4);
736                tcg_out8(s, (4 << 3) | 5);
737                tcg_out32(s, offset);
738                return;
739            }
740
741            /* ??? The memory isn't directly addressable.  */
742            g_assert_not_reached();
743        } else {
744            /* Absolute address.  */
745            tcg_out8(s, (r << 3) | 5);
746            tcg_out32(s, offset);
747            return;
748        }
749    }
750
751    /* Find the length of the immediate addend.  Note that the encoding
752       that would be used for (%ebp) indicates absolute addressing.  */
753    if (rm < 0) {
754        mod = 0, len = 4, rm = 5;
755    } else if (offset == 0 && LOWREGMASK(rm) != TCG_REG_EBP) {
756        mod = 0, len = 0;
757    } else if (offset == (int8_t)offset) {
758        mod = 0x40, len = 1;
759    } else {
760        mod = 0x80, len = 4;
761    }
762
763    /* Use a single byte MODRM format if possible.  Note that the encoding
764       that would be used for %esp is the escape to the two byte form.  */
765    if (index < 0 && LOWREGMASK(rm) != TCG_REG_ESP) {
766        /* Single byte MODRM format.  */
767        tcg_out8(s, mod | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
768    } else {
769        /* Two byte MODRM+SIB format.  */
770
771        /* Note that the encoding that would place %esp into the index
772           field indicates no index register.  In 64-bit mode, the REX.X
773           bit counts, so %r12 can be used as the index.  */
774        if (index < 0) {
775            index = 4;
776        } else {
777            tcg_debug_assert(index != TCG_REG_ESP);
778        }
779
780        tcg_out8(s, mod | (LOWREGMASK(r) << 3) | 4);
781        tcg_out8(s, (shift << 6) | (LOWREGMASK(index) << 3) | LOWREGMASK(rm));
782    }
783
784    if (len == 1) {
785        tcg_out8(s, offset);
786    } else if (len == 4) {
787        tcg_out32(s, offset);
788    }
789}
790
791static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm,
792                                     int index, int shift, intptr_t offset)
793{
794    tcg_out_opc(s, opc, r, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
795    tcg_out_sib_offset(s, r, rm, index, shift, offset);
796}
797
798static void tcg_out_vex_modrm_sib_offset(TCGContext *s, int opc, int r, int v,
799                                         int rm, int index, int shift,
800                                         intptr_t offset)
801{
802    tcg_out_vex_opc(s, opc, r, v, rm < 0 ? 0 : rm, index < 0 ? 0 : index);
803    tcg_out_sib_offset(s, r, rm, index, shift, offset);
804}
805
806/* A simplification of the above with no index or shift.  */
807static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r,
808                                        int rm, intptr_t offset)
809{
810    tcg_out_modrm_sib_offset(s, opc, r, rm, -1, 0, offset);
811}
812
813static inline void tcg_out_vex_modrm_offset(TCGContext *s, int opc, int r,
814                                            int v, int rm, intptr_t offset)
815{
816    tcg_out_vex_modrm_sib_offset(s, opc, r, v, rm, -1, 0, offset);
817}
818
819/* Output an opcode with an expected reference to the constant pool.  */
820static inline void tcg_out_modrm_pool(TCGContext *s, int opc, int r)
821{
822    tcg_out_opc(s, opc, r, 0, 0);
823    /* Absolute for 32-bit, pc-relative for 64-bit.  */
824    tcg_out8(s, LOWREGMASK(r) << 3 | 5);
825    tcg_out32(s, 0);
826}
827
828/* Output an opcode with an expected reference to the constant pool.  */
829static inline void tcg_out_vex_modrm_pool(TCGContext *s, int opc, int r)
830{
831    tcg_out_vex_opc(s, opc, r, 0, 0, 0);
832    /* Absolute for 32-bit, pc-relative for 64-bit.  */
833    tcg_out8(s, LOWREGMASK(r) << 3 | 5);
834    tcg_out32(s, 0);
835}
836
837/* Generate dest op= src.  Uses the same ARITH_* codes as tgen_arithi.  */
838static inline void tgen_arithr(TCGContext *s, int subop, int dest, int src)
839{
840    /* Propagate an opcode prefix, such as P_REXW.  */
841    int ext = subop & ~0x7;
842    subop &= 0x7;
843
844    tcg_out_modrm(s, OPC_ARITH_GvEv + (subop << 3) + ext, dest, src);
845}
846
847static bool tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg)
848{
849    int rexw = 0;
850
851    if (arg == ret) {
852        return true;
853    }
854    switch (type) {
855    case TCG_TYPE_I64:
856        rexw = P_REXW;
857        /* fallthru */
858    case TCG_TYPE_I32:
859        if (ret < 16) {
860            if (arg < 16) {
861                tcg_out_modrm(s, OPC_MOVL_GvEv + rexw, ret, arg);
862            } else {
863                tcg_out_vex_modrm(s, OPC_MOVD_EyVy + rexw, arg, 0, ret);
864            }
865        } else {
866            if (arg < 16) {
867                tcg_out_vex_modrm(s, OPC_MOVD_VyEy + rexw, ret, 0, arg);
868            } else {
869                tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
870            }
871        }
872        break;
873
874    case TCG_TYPE_V64:
875        tcg_debug_assert(ret >= 16 && arg >= 16);
876        tcg_out_vex_modrm(s, OPC_MOVQ_VqWq, ret, 0, arg);
877        break;
878    case TCG_TYPE_V128:
879        tcg_debug_assert(ret >= 16 && arg >= 16);
880        tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx, ret, 0, arg);
881        break;
882    case TCG_TYPE_V256:
883        tcg_debug_assert(ret >= 16 && arg >= 16);
884        tcg_out_vex_modrm(s, OPC_MOVDQA_VxWx | P_VEXL, ret, 0, arg);
885        break;
886
887    default:
888        g_assert_not_reached();
889    }
890    return true;
891}
892
893static const int avx2_dup_insn[4] = {
894    OPC_VPBROADCASTB, OPC_VPBROADCASTW,
895    OPC_VPBROADCASTD, OPC_VPBROADCASTQ,
896};
897
898static bool tcg_out_dup_vec(TCGContext *s, TCGType type, unsigned vece,
899                            TCGReg r, TCGReg a)
900{
901    if (have_avx2) {
902        int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
903        tcg_out_vex_modrm(s, avx2_dup_insn[vece] + vex_l, r, 0, a);
904    } else {
905        switch (vece) {
906        case MO_8:
907            /* ??? With zero in a register, use PSHUFB.  */
908            tcg_out_vex_modrm(s, OPC_PUNPCKLBW, r, a, a);
909            a = r;
910            /* FALLTHRU */
911        case MO_16:
912            tcg_out_vex_modrm(s, OPC_PUNPCKLWD, r, a, a);
913            a = r;
914            /* FALLTHRU */
915        case MO_32:
916            tcg_out_vex_modrm(s, OPC_PSHUFD, r, 0, a);
917            /* imm8 operand: all output lanes selected from input lane 0.  */
918            tcg_out8(s, 0);
919            break;
920        case MO_64:
921            tcg_out_vex_modrm(s, OPC_PUNPCKLQDQ, r, a, a);
922            break;
923        default:
924            g_assert_not_reached();
925        }
926    }
927    return true;
928}
929
930static bool tcg_out_dupm_vec(TCGContext *s, TCGType type, unsigned vece,
931                             TCGReg r, TCGReg base, intptr_t offset)
932{
933    if (have_avx2) {
934        int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
935        tcg_out_vex_modrm_offset(s, avx2_dup_insn[vece] + vex_l,
936                                 r, 0, base, offset);
937    } else {
938        switch (vece) {
939        case MO_64:
940            tcg_out_vex_modrm_offset(s, OPC_MOVDDUP, r, 0, base, offset);
941            break;
942        case MO_32:
943            tcg_out_vex_modrm_offset(s, OPC_VBROADCASTSS, r, 0, base, offset);
944            break;
945        case MO_16:
946            tcg_out_vex_modrm_offset(s, OPC_VPINSRW, r, r, base, offset);
947            tcg_out8(s, 0); /* imm8 */
948            tcg_out_dup_vec(s, type, vece, r, r);
949            break;
950        case MO_8:
951            tcg_out_vex_modrm_offset(s, OPC_VPINSRB, r, r, base, offset);
952            tcg_out8(s, 0); /* imm8 */
953            tcg_out_dup_vec(s, type, vece, r, r);
954            break;
955        default:
956            g_assert_not_reached();
957        }
958    }
959    return true;
960}
961
962static void tcg_out_dupi_vec(TCGContext *s, TCGType type, unsigned vece,
963                             TCGReg ret, int64_t arg)
964{
965    int vex_l = (type == TCG_TYPE_V256 ? P_VEXL : 0);
966
967    if (arg == 0) {
968        tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
969        return;
970    }
971    if (arg == -1) {
972        tcg_out_vex_modrm(s, OPC_PCMPEQB + vex_l, ret, ret, ret);
973        return;
974    }
975
976    if (TCG_TARGET_REG_BITS == 32 && vece < MO_64) {
977        if (have_avx2) {
978            tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTD + vex_l, ret);
979        } else {
980            tcg_out_vex_modrm_pool(s, OPC_VBROADCASTSS, ret);
981        }
982        new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
983    } else {
984        if (type == TCG_TYPE_V64) {
985            tcg_out_vex_modrm_pool(s, OPC_MOVQ_VqWq, ret);
986        } else if (have_avx2) {
987            tcg_out_vex_modrm_pool(s, OPC_VPBROADCASTQ + vex_l, ret);
988        } else {
989            tcg_out_vex_modrm_pool(s, OPC_MOVDDUP, ret);
990        }
991        if (TCG_TARGET_REG_BITS == 64) {
992            new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
993        } else {
994            new_pool_l2(s, R_386_32, s->code_ptr - 4, 0, arg, arg >> 32);
995        }
996    }
997}
998
999static void tcg_out_movi_vec(TCGContext *s, TCGType type,
1000                             TCGReg ret, tcg_target_long arg)
1001{
1002    if (arg == 0) {
1003        tcg_out_vex_modrm(s, OPC_PXOR, ret, ret, ret);
1004        return;
1005    }
1006    if (arg == -1) {
1007        tcg_out_vex_modrm(s, OPC_PCMPEQB, ret, ret, ret);
1008        return;
1009    }
1010
1011    int rexw = (type == TCG_TYPE_I32 ? 0 : P_REXW);
1012    tcg_out_vex_modrm_pool(s, OPC_MOVD_VyEy + rexw, ret);
1013    if (TCG_TARGET_REG_BITS == 64) {
1014        new_pool_label(s, arg, R_386_PC32, s->code_ptr - 4, -4);
1015    } else {
1016        new_pool_label(s, arg, R_386_32, s->code_ptr - 4, 0);
1017    }
1018}
1019
1020static void tcg_out_movi_int(TCGContext *s, TCGType type,
1021                             TCGReg ret, tcg_target_long arg)
1022{
1023    tcg_target_long diff;
1024
1025    if (arg == 0) {
1026        tgen_arithr(s, ARITH_XOR, ret, ret);
1027        return;
1028    }
1029    if (arg == (uint32_t)arg || type == TCG_TYPE_I32) {
1030        tcg_out_opc(s, OPC_MOVL_Iv + LOWREGMASK(ret), 0, ret, 0);
1031        tcg_out32(s, arg);
1032        return;
1033    }
1034    if (arg == (int32_t)arg) {
1035        tcg_out_modrm(s, OPC_MOVL_EvIz + P_REXW, 0, ret);
1036        tcg_out32(s, arg);
1037        return;
1038    }
1039
1040    /* Try a 7 byte pc-relative lea before the 10 byte movq.  */
1041    diff = tcg_pcrel_diff(s, (const void *)arg) - 7;
1042    if (diff == (int32_t)diff) {
1043        tcg_out_opc(s, OPC_LEA | P_REXW, ret, 0, 0);
1044        tcg_out8(s, (LOWREGMASK(ret) << 3) | 5);
1045        tcg_out32(s, diff);
1046        return;
1047    }
1048
1049    tcg_out_opc(s, OPC_MOVL_Iv + P_REXW + LOWREGMASK(ret), 0, ret, 0);
1050    tcg_out64(s, arg);
1051}
1052
1053static void tcg_out_movi(TCGContext *s, TCGType type,
1054                         TCGReg ret, tcg_target_long arg)
1055{
1056    switch (type) {
1057    case TCG_TYPE_I32:
1058#if TCG_TARGET_REG_BITS == 64
1059    case TCG_TYPE_I64:
1060#endif
1061        if (ret < 16) {
1062            tcg_out_movi_int(s, type, ret, arg);
1063        } else {
1064            tcg_out_movi_vec(s, type, ret, arg);
1065        }
1066        break;
1067    default:
1068        g_assert_not_reached();
1069    }
1070}
1071
1072static inline void tcg_out_pushi(TCGContext *s, tcg_target_long val)
1073{
1074    if (val == (int8_t)val) {
1075        tcg_out_opc(s, OPC_PUSH_Ib, 0, 0, 0);
1076        tcg_out8(s, val);
1077    } else if (val == (int32_t)val) {
1078        tcg_out_opc(s, OPC_PUSH_Iv, 0, 0, 0);
1079        tcg_out32(s, val);
1080    } else {
1081        tcg_abort();
1082    }
1083}
1084
1085static inline void tcg_out_mb(TCGContext *s, TCGArg a0)
1086{
1087    /* Given the strength of x86 memory ordering, we only need care for
1088       store-load ordering.  Experimentally, "lock orl $0,0(%esp)" is
1089       faster than "mfence", so don't bother with the sse insn.  */
1090    if (a0 & TCG_MO_ST_LD) {
1091        tcg_out8(s, 0xf0);
1092        tcg_out_modrm_offset(s, OPC_ARITH_EvIb, ARITH_OR, TCG_REG_ESP, 0);
1093        tcg_out8(s, 0);
1094    }
1095}
1096
1097static inline void tcg_out_push(TCGContext *s, int reg)
1098{
1099    tcg_out_opc(s, OPC_PUSH_r32 + LOWREGMASK(reg), 0, reg, 0);
1100}
1101
1102static inline void tcg_out_pop(TCGContext *s, int reg)
1103{
1104    tcg_out_opc(s, OPC_POP_r32 + LOWREGMASK(reg), 0, reg, 0);
1105}
1106
1107static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret,
1108                       TCGReg arg1, intptr_t arg2)
1109{
1110    switch (type) {
1111    case TCG_TYPE_I32:
1112        if (ret < 16) {
1113            tcg_out_modrm_offset(s, OPC_MOVL_GvEv, ret, arg1, arg2);
1114        } else {
1115            tcg_out_vex_modrm_offset(s, OPC_MOVD_VyEy, ret, 0, arg1, arg2);
1116        }
1117        break;
1118    case TCG_TYPE_I64:
1119        if (ret < 16) {
1120            tcg_out_modrm_offset(s, OPC_MOVL_GvEv | P_REXW, ret, arg1, arg2);
1121            break;
1122        }
1123        /* FALLTHRU */
1124    case TCG_TYPE_V64:
1125        /* There is no instruction that can validate 8-byte alignment.  */
1126        tcg_debug_assert(ret >= 16);
1127        tcg_out_vex_modrm_offset(s, OPC_MOVQ_VqWq, ret, 0, arg1, arg2);
1128        break;
1129    case TCG_TYPE_V128:
1130        /*
1131         * The gvec infrastructure is asserts that v128 vector loads
1132         * and stores use a 16-byte aligned offset.  Validate that the
1133         * final pointer is aligned by using an insn that will SIGSEGV.
1134         */
1135        tcg_debug_assert(ret >= 16);
1136        tcg_out_vex_modrm_offset(s, OPC_MOVDQA_VxWx, ret, 0, arg1, arg2);
1137        break;
1138    case TCG_TYPE_V256:
1139        /*
1140         * The gvec infrastructure only requires 16-byte alignment,
1141         * so here we must use an unaligned load.
1142         */
1143        tcg_debug_assert(ret >= 16);
1144        tcg_out_vex_modrm_offset(s, OPC_MOVDQU_VxWx | P_VEXL,
1145                                 ret, 0, arg1, arg2);
1146        break;
1147    default:
1148        g_assert_not_reached();
1149    }
1150}
1151
1152static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
1153                       TCGReg arg1, intptr_t arg2)
1154{
1155    switch (type) {
1156    case TCG_TYPE_I32:
1157        if (arg < 16) {
1158            tcg_out_modrm_offset(s, OPC_MOVL_EvGv, arg, arg1, arg2);
1159        } else {
1160            tcg_out_vex_modrm_offset(s, OPC_MOVD_EyVy, arg, 0, arg1, arg2);
1161        }
1162        break;
1163    case TCG_TYPE_I64:
1164        if (arg < 16) {
1165            tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_REXW, arg, arg1, arg2);
1166            break;
1167        }
1168        /* FALLTHRU */
1169    case TCG_TYPE_V64:
1170        /* There is no instruction that can validate 8-byte alignment.  */
1171        tcg_debug_assert(arg >= 16);
1172        tcg_out_vex_modrm_offset(s, OPC_MOVQ_WqVq, arg, 0, arg1, arg2);
1173        break;
1174    case TCG_TYPE_V128:
1175        /*
1176         * The gvec infrastructure is asserts that v128 vector loads
1177         * and stores use a 16-byte aligned offset.  Validate that the
1178         * final pointer is aligned by using an insn that will SIGSEGV.
1179         */
1180        tcg_debug_assert(arg >= 16);
1181        tcg_out_vex_modrm_offset(s, OPC_MOVDQA_WxVx, arg, 0, arg1, arg2);
1182        break;
1183    case TCG_TYPE_V256:
1184        /*
1185         * The gvec infrastructure only requires 16-byte alignment,
1186         * so here we must use an unaligned store.
1187         */
1188        tcg_debug_assert(arg >= 16);
1189        tcg_out_vex_modrm_offset(s, OPC_MOVDQU_WxVx | P_VEXL,
1190                                 arg, 0, arg1, arg2);
1191        break;
1192    default:
1193        g_assert_not_reached();
1194    }
1195}
1196
1197static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val,
1198                        TCGReg base, intptr_t ofs)
1199{
1200    int rexw = 0;
1201    if (TCG_TARGET_REG_BITS == 64 && type == TCG_TYPE_I64) {
1202        if (val != (int32_t)val) {
1203            return false;
1204        }
1205        rexw = P_REXW;
1206    } else if (type != TCG_TYPE_I32) {
1207        return false;
1208    }
1209    tcg_out_modrm_offset(s, OPC_MOVL_EvIz | rexw, 0, base, ofs);
1210    tcg_out32(s, val);
1211    return true;
1212}
1213
1214static void tcg_out_shifti(TCGContext *s, int subopc, int reg, int count)
1215{
1216    /* Propagate an opcode prefix, such as P_DATA16.  */
1217    int ext = subopc & ~0x7;
1218    subopc &= 0x7;
1219
1220    if (count == 1) {
1221        tcg_out_modrm(s, OPC_SHIFT_1 + ext, subopc, reg);
1222    } else {
1223        tcg_out_modrm(s, OPC_SHIFT_Ib + ext, subopc, reg);
1224        tcg_out8(s, count);
1225    }
1226}
1227
1228static inline void tcg_out_bswap32(TCGContext *s, int reg)
1229{
1230    tcg_out_opc(s, OPC_BSWAP + LOWREGMASK(reg), 0, reg, 0);
1231}
1232
1233static inline void tcg_out_rolw_8(TCGContext *s, int reg)
1234{
1235    tcg_out_shifti(s, SHIFT_ROL + P_DATA16, reg, 8);
1236}
1237
1238static inline void tcg_out_ext8u(TCGContext *s, int dest, int src)
1239{
1240    /* movzbl */
1241    tcg_debug_assert(src < 4 || TCG_TARGET_REG_BITS == 64);
1242    tcg_out_modrm(s, OPC_MOVZBL + P_REXB_RM, dest, src);
1243}
1244
1245static void tcg_out_ext8s(TCGContext *s, int dest, int src, int rexw)
1246{
1247    /* movsbl */
1248    tcg_debug_assert(src < 4 || TCG_TARGET_REG_BITS == 64);
1249    tcg_out_modrm(s, OPC_MOVSBL + P_REXB_RM + rexw, dest, src);
1250}
1251
1252static inline void tcg_out_ext16u(TCGContext *s, int dest, int src)
1253{
1254    /* movzwl */
1255    tcg_out_modrm(s, OPC_MOVZWL, dest, src);
1256}
1257
1258static inline void tcg_out_ext16s(TCGContext *s, int dest, int src, int rexw)
1259{
1260    /* movsw[lq] */
1261    tcg_out_modrm(s, OPC_MOVSWL + rexw, dest, src);
1262}
1263
1264static inline void tcg_out_ext32u(TCGContext *s, int dest, int src)
1265{
1266    /* 32-bit mov zero extends.  */
1267    tcg_out_modrm(s, OPC_MOVL_GvEv, dest, src);
1268}
1269
1270static inline void tcg_out_ext32s(TCGContext *s, int dest, int src)
1271{
1272    tcg_out_modrm(s, OPC_MOVSLQ, dest, src);
1273}
1274
1275static inline void tcg_out_bswap64(TCGContext *s, int reg)
1276{
1277    tcg_out_opc(s, OPC_BSWAP + P_REXW + LOWREGMASK(reg), 0, reg, 0);
1278}
1279
1280static void tgen_arithi(TCGContext *s, int c, int r0,
1281                        tcg_target_long val, int cf)
1282{
1283    int rexw = 0;
1284
1285    if (TCG_TARGET_REG_BITS == 64) {
1286        rexw = c & -8;
1287        c &= 7;
1288    }
1289
1290    /* ??? While INC is 2 bytes shorter than ADDL $1, they also induce
1291       partial flags update stalls on Pentium4 and are not recommended
1292       by current Intel optimization manuals.  */
1293    if (!cf && (c == ARITH_ADD || c == ARITH_SUB) && (val == 1 || val == -1)) {
1294        int is_inc = (c == ARITH_ADD) ^ (val < 0);
1295        if (TCG_TARGET_REG_BITS == 64) {
1296            /* The single-byte increment encodings are re-tasked as the
1297               REX prefixes.  Use the MODRM encoding.  */
1298            tcg_out_modrm(s, OPC_GRP5 + rexw,
1299                          (is_inc ? EXT5_INC_Ev : EXT5_DEC_Ev), r0);
1300        } else {
1301            tcg_out8(s, (is_inc ? OPC_INC_r32 : OPC_DEC_r32) + r0);
1302        }
1303        return;
1304    }
1305
1306    if (c == ARITH_AND) {
1307        if (TCG_TARGET_REG_BITS == 64) {
1308            if (val == 0xffffffffu) {
1309                tcg_out_ext32u(s, r0, r0);
1310                return;
1311            }
1312            if (val == (uint32_t)val) {
1313                /* AND with no high bits set can use a 32-bit operation.  */
1314                rexw = 0;
1315            }
1316        }
1317        if (val == 0xffu && (r0 < 4 || TCG_TARGET_REG_BITS == 64)) {
1318            tcg_out_ext8u(s, r0, r0);
1319            return;
1320        }
1321        if (val == 0xffffu) {
1322            tcg_out_ext16u(s, r0, r0);
1323            return;
1324        }
1325    }
1326
1327    if (val == (int8_t)val) {
1328        tcg_out_modrm(s, OPC_ARITH_EvIb + rexw, c, r0);
1329        tcg_out8(s, val);
1330        return;
1331    }
1332    if (rexw == 0 || val == (int32_t)val) {
1333        tcg_out_modrm(s, OPC_ARITH_EvIz + rexw, c, r0);
1334        tcg_out32(s, val);
1335        return;
1336    }
1337
1338    tcg_abort();
1339}
1340
1341static void tcg_out_addi(TCGContext *s, int reg, tcg_target_long val)
1342{
1343    if (val != 0) {
1344        tgen_arithi(s, ARITH_ADD + P_REXW, reg, val, 0);
1345    }
1346}
1347
1348/* Use SMALL != 0 to force a short forward branch.  */
1349static void tcg_out_jxx(TCGContext *s, int opc, TCGLabel *l, int small)
1350{
1351    int32_t val, val1;
1352
1353    if (l->has_value) {
1354        val = tcg_pcrel_diff(s, l->u.value_ptr);
1355        val1 = val - 2;
1356        if ((int8_t)val1 == val1) {
1357            if (opc == -1) {
1358                tcg_out8(s, OPC_JMP_short);
1359            } else {
1360                tcg_out8(s, OPC_JCC_short + opc);
1361            }
1362            tcg_out8(s, val1);
1363        } else {
1364            if (small) {
1365                tcg_abort();
1366            }
1367            if (opc == -1) {
1368                tcg_out8(s, OPC_JMP_long);
1369                tcg_out32(s, val - 5);
1370            } else {
1371                tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1372                tcg_out32(s, val - 6);
1373            }
1374        }
1375    } else if (small) {
1376        if (opc == -1) {
1377            tcg_out8(s, OPC_JMP_short);
1378        } else {
1379            tcg_out8(s, OPC_JCC_short + opc);
1380        }
1381        tcg_out_reloc(s, s->code_ptr, R_386_PC8, l, -1);
1382        s->code_ptr += 1;
1383    } else {
1384        if (opc == -1) {
1385            tcg_out8(s, OPC_JMP_long);
1386        } else {
1387            tcg_out_opc(s, OPC_JCC_long + opc, 0, 0, 0);
1388        }
1389        tcg_out_reloc(s, s->code_ptr, R_386_PC32, l, -4);
1390        s->code_ptr += 4;
1391    }
1392}
1393
1394static void tcg_out_cmp(TCGContext *s, TCGArg arg1, TCGArg arg2,
1395                        int const_arg2, int rexw)
1396{
1397    if (const_arg2) {
1398        if (arg2 == 0) {
1399            /* test r, r */
1400            tcg_out_modrm(s, OPC_TESTL + rexw, arg1, arg1);
1401        } else {
1402            tgen_arithi(s, ARITH_CMP + rexw, arg1, arg2, 0);
1403        }
1404    } else {
1405        tgen_arithr(s, ARITH_CMP + rexw, arg1, arg2);
1406    }
1407}
1408
1409static void tcg_out_brcond32(TCGContext *s, TCGCond cond,
1410                             TCGArg arg1, TCGArg arg2, int const_arg2,
1411                             TCGLabel *label, int small)
1412{
1413    tcg_out_cmp(s, arg1, arg2, const_arg2, 0);
1414    tcg_out_jxx(s, tcg_cond_to_jcc[cond], label, small);
1415}
1416
1417#if TCG_TARGET_REG_BITS == 64
1418static void tcg_out_brcond64(TCGContext *s, TCGCond cond,
1419                             TCGArg arg1, TCGArg arg2, int const_arg2,
1420                             TCGLabel *label, int small)
1421{
1422    tcg_out_cmp(s, arg1, arg2, const_arg2, P_REXW);
1423    tcg_out_jxx(s, tcg_cond_to_jcc[cond], label, small);
1424}
1425#else
1426/* XXX: we implement it at the target level to avoid having to
1427   handle cross basic blocks temporaries */
1428static void tcg_out_brcond2(TCGContext *s, const TCGArg *args,
1429                            const int *const_args, int small)
1430{
1431    TCGLabel *label_next = gen_new_label();
1432    TCGLabel *label_this = arg_label(args[5]);
1433
1434    switch(args[4]) {
1435    case TCG_COND_EQ:
1436        tcg_out_brcond32(s, TCG_COND_NE, args[0], args[2], const_args[2],
1437                         label_next, 1);
1438        tcg_out_brcond32(s, TCG_COND_EQ, args[1], args[3], const_args[3],
1439                         label_this, small);
1440        break;
1441    case TCG_COND_NE:
1442        tcg_out_brcond32(s, TCG_COND_NE, args[0], args[2], const_args[2],
1443                         label_this, small);
1444        tcg_out_brcond32(s, TCG_COND_NE, args[1], args[3], const_args[3],
1445                         label_this, small);
1446        break;
1447    case TCG_COND_LT:
1448        tcg_out_brcond32(s, TCG_COND_LT, args[1], args[3], const_args[3],
1449                         label_this, small);
1450        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1451        tcg_out_brcond32(s, TCG_COND_LTU, args[0], args[2], const_args[2],
1452                         label_this, small);
1453        break;
1454    case TCG_COND_LE:
1455        tcg_out_brcond32(s, TCG_COND_LT, args[1], args[3], const_args[3],
1456                         label_this, small);
1457        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1458        tcg_out_brcond32(s, TCG_COND_LEU, args[0], args[2], const_args[2],
1459                         label_this, small);
1460        break;
1461    case TCG_COND_GT:
1462        tcg_out_brcond32(s, TCG_COND_GT, args[1], args[3], const_args[3],
1463                         label_this, small);
1464        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1465        tcg_out_brcond32(s, TCG_COND_GTU, args[0], args[2], const_args[2],
1466                         label_this, small);
1467        break;
1468    case TCG_COND_GE:
1469        tcg_out_brcond32(s, TCG_COND_GT, args[1], args[3], const_args[3],
1470                         label_this, small);
1471        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1472        tcg_out_brcond32(s, TCG_COND_GEU, args[0], args[2], const_args[2],
1473                         label_this, small);
1474        break;
1475    case TCG_COND_LTU:
1476        tcg_out_brcond32(s, TCG_COND_LTU, args[1], args[3], const_args[3],
1477                         label_this, small);
1478        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1479        tcg_out_brcond32(s, TCG_COND_LTU, args[0], args[2], const_args[2],
1480                         label_this, small);
1481        break;
1482    case TCG_COND_LEU:
1483        tcg_out_brcond32(s, TCG_COND_LTU, args[1], args[3], const_args[3],
1484                         label_this, small);
1485        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1486        tcg_out_brcond32(s, TCG_COND_LEU, args[0], args[2], const_args[2],
1487                         label_this, small);
1488        break;
1489    case TCG_COND_GTU:
1490        tcg_out_brcond32(s, TCG_COND_GTU, args[1], args[3], const_args[3],
1491                         label_this, small);
1492        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1493        tcg_out_brcond32(s, TCG_COND_GTU, args[0], args[2], const_args[2],
1494                         label_this, small);
1495        break;
1496    case TCG_COND_GEU:
1497        tcg_out_brcond32(s, TCG_COND_GTU, args[1], args[3], const_args[3],
1498                         label_this, small);
1499        tcg_out_jxx(s, JCC_JNE, label_next, 1);
1500        tcg_out_brcond32(s, TCG_COND_GEU, args[0], args[2], const_args[2],
1501                         label_this, small);
1502        break;
1503    default:
1504        tcg_abort();
1505    }
1506    tcg_out_label(s, label_next);
1507}
1508#endif
1509
1510static void tcg_out_setcond32(TCGContext *s, TCGCond cond, TCGArg dest,
1511                              TCGArg arg1, TCGArg arg2, int const_arg2)
1512{
1513    tcg_out_cmp(s, arg1, arg2, const_arg2, 0);
1514    tcg_out_modrm(s, OPC_SETCC | tcg_cond_to_jcc[cond], 0, dest);
1515    tcg_out_ext8u(s, dest, dest);
1516}
1517
1518#if TCG_TARGET_REG_BITS == 64
1519static void tcg_out_setcond64(TCGContext *s, TCGCond cond, TCGArg dest,
1520                              TCGArg arg1, TCGArg arg2, int const_arg2)
1521{
1522    tcg_out_cmp(s, arg1, arg2, const_arg2, P_REXW);
1523    tcg_out_modrm(s, OPC_SETCC | tcg_cond_to_jcc[cond], 0, dest);
1524    tcg_out_ext8u(s, dest, dest);
1525}
1526#else
1527static void tcg_out_setcond2(TCGContext *s, const TCGArg *args,
1528                             const int *const_args)
1529{
1530    TCGArg new_args[6];
1531    TCGLabel *label_true, *label_over;
1532
1533    memcpy(new_args, args+1, 5*sizeof(TCGArg));
1534
1535    if (args[0] == args[1] || args[0] == args[2]
1536        || (!const_args[3] && args[0] == args[3])
1537        || (!const_args[4] && args[0] == args[4])) {
1538        /* When the destination overlaps with one of the argument
1539           registers, don't do anything tricky.  */
1540        label_true = gen_new_label();
1541        label_over = gen_new_label();
1542
1543        new_args[5] = label_arg(label_true);
1544        tcg_out_brcond2(s, new_args, const_args+1, 1);
1545
1546        tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1547        tcg_out_jxx(s, JCC_JMP, label_over, 1);
1548        tcg_out_label(s, label_true);
1549
1550        tcg_out_movi(s, TCG_TYPE_I32, args[0], 1);
1551        tcg_out_label(s, label_over);
1552    } else {
1553        /* When the destination does not overlap one of the arguments,
1554           clear the destination first, jump if cond false, and emit an
1555           increment in the true case.  This results in smaller code.  */
1556
1557        tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
1558
1559        label_over = gen_new_label();
1560        new_args[4] = tcg_invert_cond(new_args[4]);
1561        new_args[5] = label_arg(label_over);
1562        tcg_out_brcond2(s, new_args, const_args+1, 1);
1563
1564        tgen_arithi(s, ARITH_ADD, args[0], 1, 0);
1565        tcg_out_label(s, label_over);
1566    }
1567}
1568#endif
1569
1570static void tcg_out_cmov(TCGContext *s, TCGCond cond, int rexw,
1571                         TCGReg dest, TCGReg v1)
1572{
1573    if (have_cmov) {
1574        tcg_out_modrm(s, OPC_CMOVCC | tcg_cond_to_jcc[cond] | rexw, dest, v1);
1575    } else {
1576        TCGLabel *over = gen_new_label();
1577        tcg_out_jxx(s, tcg_cond_to_jcc[tcg_invert_cond(cond)], over, 1);
1578        tcg_out_mov(s, TCG_TYPE_I32, dest, v1);
1579        tcg_out_label(s, over);
1580    }
1581}
1582
1583static void tcg_out_movcond32(TCGContext *s, TCGCond cond, TCGReg dest,
1584                              TCGReg c1, TCGArg c2, int const_c2,
1585                              TCGReg v1)
1586{
1587    tcg_out_cmp(s, c1, c2, const_c2, 0);
1588    tcg_out_cmov(s, cond, 0, dest, v1);
1589}
1590
1591#if TCG_TARGET_REG_BITS == 64
1592static void tcg_out_movcond64(TCGContext *s, TCGCond cond, TCGReg dest,
1593                              TCGReg c1, TCGArg c2, int const_c2,
1594                              TCGReg v1)
1595{
1596    tcg_out_cmp(s, c1, c2, const_c2, P_REXW);
1597    tcg_out_cmov(s, cond, P_REXW, dest, v1);
1598}
1599#endif
1600
1601static void tcg_out_ctz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1602                        TCGArg arg2, bool const_a2)
1603{
1604    if (have_bmi1) {
1605        tcg_out_modrm(s, OPC_TZCNT + rexw, dest, arg1);
1606        if (const_a2) {
1607            tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1608        } else {
1609            tcg_debug_assert(dest != arg2);
1610            tcg_out_cmov(s, TCG_COND_LTU, rexw, dest, arg2);
1611        }
1612    } else {
1613        tcg_debug_assert(dest != arg2);
1614        tcg_out_modrm(s, OPC_BSF + rexw, dest, arg1);
1615        tcg_out_cmov(s, TCG_COND_EQ, rexw, dest, arg2);
1616    }
1617}
1618
1619static void tcg_out_clz(TCGContext *s, int rexw, TCGReg dest, TCGReg arg1,
1620                        TCGArg arg2, bool const_a2)
1621{
1622    if (have_lzcnt) {
1623        tcg_out_modrm(s, OPC_LZCNT + rexw, dest, arg1);
1624        if (const_a2) {
1625            tcg_debug_assert(arg2 == (rexw ? 64 : 32));
1626        } else {
1627            tcg_debug_assert(dest != arg2);
1628            tcg_out_cmov(s, TCG_COND_LTU, rexw, dest, arg2);
1629        }
1630    } else {
1631        tcg_debug_assert(!const_a2);
1632        tcg_debug_assert(dest != arg1);
1633        tcg_debug_assert(dest != arg2);
1634
1635        /* Recall that the output of BSR is the index not the count.  */
1636        tcg_out_modrm(s, OPC_BSR + rexw, dest, arg1);
1637        tgen_arithi(s, ARITH_XOR + rexw, dest, rexw ? 63 : 31, 0);
1638
1639        /* Since we have destroyed the flags from BSR, we have to re-test.  */
1640        tcg_out_cmp(s, arg1, 0, 1, rexw);
1641        tcg_out_cmov(s, TCG_COND_EQ, rexw, dest, arg2);
1642    }
1643}
1644
1645static void tcg_out_branch(TCGContext *s, int call, const tcg_insn_unit *dest)
1646{
1647    intptr_t disp = tcg_pcrel_diff(s, dest) - 5;
1648
1649    if (disp == (int32_t)disp) {
1650        tcg_out_opc(s, call ? OPC_CALL_Jz : OPC_JMP_long, 0, 0, 0);
1651        tcg_out32(s, disp);
1652    } else {
1653        /* rip-relative addressing into the constant pool.
1654           This is 6 + 8 = 14 bytes, as compared to using an
1655           immediate load 10 + 6 = 16 bytes, plus we may
1656           be able to re-use the pool constant for more calls.  */
1657        tcg_out_opc(s, OPC_GRP5, 0, 0, 0);
1658        tcg_out8(s, (call ? EXT5_CALLN_Ev : EXT5_JMPN_Ev) << 3 | 5);
1659        new_pool_label(s, (uintptr_t)dest, R_386_PC32, s->code_ptr, -4);
1660        tcg_out32(s, 0);
1661    }
1662}
1663
1664static void tcg_out_call(TCGContext *s, const tcg_insn_unit *dest,
1665                         const TCGHelperInfo *info)
1666{
1667    tcg_out_branch(s, 1, dest);
1668}
1669
1670static void tcg_out_jmp(TCGContext *s, const tcg_insn_unit *dest)
1671{
1672    tcg_out_branch(s, 0, dest);
1673}
1674
1675static void tcg_out_nopn(TCGContext *s, int n)
1676{
1677    int i;
1678    /* Emit 1 or 2 operand size prefixes for the standard one byte nop,
1679     * "xchg %eax,%eax", forming "xchg %ax,%ax". All cores accept the
1680     * duplicate prefix, and all of the interesting recent cores can
1681     * decode and discard the duplicates in a single cycle.
1682     */
1683    tcg_debug_assert(n >= 1);
1684    for (i = 1; i < n; ++i) {
1685        tcg_out8(s, 0x66);
1686    }
1687    tcg_out8(s, 0x90);
1688}
1689
1690#if defined(CONFIG_SOFTMMU)
1691/* helper signature: helper_ret_ld_mmu(CPUState *env, target_ulong addr,
1692 *                                     int mmu_idx, uintptr_t ra)
1693 */
1694static void * const qemu_ld_helpers[(MO_SIZE | MO_BSWAP) + 1] = {
1695    [MO_UB]   = helper_ret_ldub_mmu,
1696    [MO_LEUW] = helper_le_lduw_mmu,
1697    [MO_LEUL] = helper_le_ldul_mmu,
1698    [MO_LEUQ] = helper_le_ldq_mmu,
1699    [MO_BEUW] = helper_be_lduw_mmu,
1700    [MO_BEUL] = helper_be_ldul_mmu,
1701    [MO_BEUQ] = helper_be_ldq_mmu,
1702};
1703
1704/* helper signature: helper_ret_st_mmu(CPUState *env, target_ulong addr,
1705 *                                     uintxx_t val, int mmu_idx, uintptr_t ra)
1706 */
1707static void * const qemu_st_helpers[(MO_SIZE | MO_BSWAP) + 1] = {
1708    [MO_UB]   = helper_ret_stb_mmu,
1709    [MO_LEUW] = helper_le_stw_mmu,
1710    [MO_LEUL] = helper_le_stl_mmu,
1711    [MO_LEUQ] = helper_le_stq_mmu,
1712    [MO_BEUW] = helper_be_stw_mmu,
1713    [MO_BEUL] = helper_be_stl_mmu,
1714    [MO_BEUQ] = helper_be_stq_mmu,
1715};
1716
1717/* Perform the TLB load and compare.
1718
1719   Inputs:
1720   ADDRLO and ADDRHI contain the low and high part of the address.
1721
1722   MEM_INDEX and S_BITS are the memory context and log2 size of the load.
1723
1724   WHICH is the offset into the CPUTLBEntry structure of the slot to read.
1725   This should be offsetof addr_read or addr_write.
1726
1727   Outputs:
1728   LABEL_PTRS is filled with 1 (32-bit addresses) or 2 (64-bit addresses)
1729   positions of the displacements of forward jumps to the TLB miss case.
1730
1731   Second argument register is loaded with the low part of the address.
1732   In the TLB hit case, it has been adjusted as indicated by the TLB
1733   and so is a host address.  In the TLB miss case, it continues to
1734   hold a guest address.
1735
1736   First argument register is clobbered.  */
1737
1738static inline void tcg_out_tlb_load(TCGContext *s, TCGReg addrlo, TCGReg addrhi,
1739                                    int mem_index, MemOp opc,
1740                                    tcg_insn_unit **label_ptr, int which)
1741{
1742    const TCGReg r0 = TCG_REG_L0;
1743    const TCGReg r1 = TCG_REG_L1;
1744    TCGType ttype = TCG_TYPE_I32;
1745    TCGType tlbtype = TCG_TYPE_I32;
1746    int trexw = 0, hrexw = 0, tlbrexw = 0;
1747    unsigned a_bits = get_alignment_bits(opc);
1748    unsigned s_bits = opc & MO_SIZE;
1749    unsigned a_mask = (1 << a_bits) - 1;
1750    unsigned s_mask = (1 << s_bits) - 1;
1751    target_ulong tlb_mask;
1752
1753    if (TCG_TARGET_REG_BITS == 64) {
1754        if (TARGET_LONG_BITS == 64) {
1755            ttype = TCG_TYPE_I64;
1756            trexw = P_REXW;
1757        }
1758        if (TCG_TYPE_PTR == TCG_TYPE_I64) {
1759            hrexw = P_REXW;
1760            if (TARGET_PAGE_BITS + CPU_TLB_DYN_MAX_BITS > 32) {
1761                tlbtype = TCG_TYPE_I64;
1762                tlbrexw = P_REXW;
1763            }
1764        }
1765    }
1766
1767    tcg_out_mov(s, tlbtype, r0, addrlo);
1768    tcg_out_shifti(s, SHIFT_SHR + tlbrexw, r0,
1769                   TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS);
1770
1771    tcg_out_modrm_offset(s, OPC_AND_GvEv + trexw, r0, TCG_AREG0,
1772                         TLB_MASK_TABLE_OFS(mem_index) +
1773                         offsetof(CPUTLBDescFast, mask));
1774
1775    tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, r0, TCG_AREG0,
1776                         TLB_MASK_TABLE_OFS(mem_index) +
1777                         offsetof(CPUTLBDescFast, table));
1778
1779    /* If the required alignment is at least as large as the access, simply
1780       copy the address and mask.  For lesser alignments, check that we don't
1781       cross pages for the complete access.  */
1782    if (a_bits >= s_bits) {
1783        tcg_out_mov(s, ttype, r1, addrlo);
1784    } else {
1785        tcg_out_modrm_offset(s, OPC_LEA + trexw, r1, addrlo, s_mask - a_mask);
1786    }
1787    tlb_mask = (target_ulong)TARGET_PAGE_MASK | a_mask;
1788    tgen_arithi(s, ARITH_AND + trexw, r1, tlb_mask, 0);
1789
1790    /* cmp 0(r0), r1 */
1791    tcg_out_modrm_offset(s, OPC_CMP_GvEv + trexw, r1, r0, which);
1792
1793    /* Prepare for both the fast path add of the tlb addend, and the slow
1794       path function argument setup.  */
1795    tcg_out_mov(s, ttype, r1, addrlo);
1796
1797    /* jne slow_path */
1798    tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
1799    label_ptr[0] = s->code_ptr;
1800    s->code_ptr += 4;
1801
1802    if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
1803        /* cmp 4(r0), addrhi */
1804        tcg_out_modrm_offset(s, OPC_CMP_GvEv, addrhi, r0, which + 4);
1805
1806        /* jne slow_path */
1807        tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
1808        label_ptr[1] = s->code_ptr;
1809        s->code_ptr += 4;
1810    }
1811
1812    /* TLB Hit.  */
1813
1814    /* add addend(r0), r1 */
1815    tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, r1, r0,
1816                         offsetof(CPUTLBEntry, addend));
1817}
1818
1819/*
1820 * Record the context of a call to the out of line helper code for the slow path
1821 * for a load or store, so that we can later generate the correct helper code
1822 */
1823static void add_qemu_ldst_label(TCGContext *s, bool is_ld, bool is_64,
1824                                MemOpIdx oi,
1825                                TCGReg datalo, TCGReg datahi,
1826                                TCGReg addrlo, TCGReg addrhi,
1827                                tcg_insn_unit *raddr,
1828                                tcg_insn_unit **label_ptr)
1829{
1830    TCGLabelQemuLdst *label = new_ldst_label(s);
1831
1832    label->is_ld = is_ld;
1833    label->oi = oi;
1834    label->type = is_64 ? TCG_TYPE_I64 : TCG_TYPE_I32;
1835    label->datalo_reg = datalo;
1836    label->datahi_reg = datahi;
1837    label->addrlo_reg = addrlo;
1838    label->addrhi_reg = addrhi;
1839    label->raddr = tcg_splitwx_to_rx(raddr);
1840    label->label_ptr[0] = label_ptr[0];
1841    if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
1842        label->label_ptr[1] = label_ptr[1];
1843    }
1844}
1845
1846/*
1847 * Generate code for the slow path for a load at the end of block
1848 */
1849static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
1850{
1851    MemOpIdx oi = l->oi;
1852    MemOp opc = get_memop(oi);
1853    TCGReg data_reg;
1854    tcg_insn_unit **label_ptr = &l->label_ptr[0];
1855    int rexw = (l->type == TCG_TYPE_I64 ? P_REXW : 0);
1856
1857    /* resolve label address */
1858    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
1859    if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
1860        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
1861    }
1862
1863    if (TCG_TARGET_REG_BITS == 32) {
1864        int ofs = 0;
1865
1866        tcg_out_st(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP, ofs);
1867        ofs += 4;
1868
1869        tcg_out_st(s, TCG_TYPE_I32, l->addrlo_reg, TCG_REG_ESP, ofs);
1870        ofs += 4;
1871
1872        if (TARGET_LONG_BITS == 64) {
1873            tcg_out_st(s, TCG_TYPE_I32, l->addrhi_reg, TCG_REG_ESP, ofs);
1874            ofs += 4;
1875        }
1876
1877        tcg_out_sti(s, TCG_TYPE_I32, oi, TCG_REG_ESP, ofs);
1878        ofs += 4;
1879
1880        tcg_out_sti(s, TCG_TYPE_PTR, (uintptr_t)l->raddr, TCG_REG_ESP, ofs);
1881    } else {
1882        tcg_out_mov(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[0], TCG_AREG0);
1883        /* The second argument is already loaded with addrlo.  */
1884        tcg_out_movi(s, TCG_TYPE_I32, tcg_target_call_iarg_regs[2], oi);
1885        tcg_out_movi(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[3],
1886                     (uintptr_t)l->raddr);
1887    }
1888
1889    tcg_out_branch(s, 1, qemu_ld_helpers[opc & (MO_BSWAP | MO_SIZE)]);
1890
1891    data_reg = l->datalo_reg;
1892    switch (opc & MO_SSIZE) {
1893    case MO_SB:
1894        tcg_out_ext8s(s, data_reg, TCG_REG_EAX, rexw);
1895        break;
1896    case MO_SW:
1897        tcg_out_ext16s(s, data_reg, TCG_REG_EAX, rexw);
1898        break;
1899#if TCG_TARGET_REG_BITS == 64
1900    case MO_SL:
1901        tcg_out_ext32s(s, data_reg, TCG_REG_EAX);
1902        break;
1903#endif
1904    case MO_UB:
1905    case MO_UW:
1906        /* Note that the helpers have zero-extended to tcg_target_long.  */
1907    case MO_UL:
1908        tcg_out_mov(s, TCG_TYPE_I32, data_reg, TCG_REG_EAX);
1909        break;
1910    case MO_UQ:
1911        if (TCG_TARGET_REG_BITS == 64) {
1912            tcg_out_mov(s, TCG_TYPE_I64, data_reg, TCG_REG_RAX);
1913        } else if (data_reg == TCG_REG_EDX) {
1914            /* xchg %edx, %eax */
1915            tcg_out_opc(s, OPC_XCHG_ax_r32 + TCG_REG_EDX, 0, 0, 0);
1916            tcg_out_mov(s, TCG_TYPE_I32, l->datahi_reg, TCG_REG_EAX);
1917        } else {
1918            tcg_out_mov(s, TCG_TYPE_I32, data_reg, TCG_REG_EAX);
1919            tcg_out_mov(s, TCG_TYPE_I32, l->datahi_reg, TCG_REG_EDX);
1920        }
1921        break;
1922    default:
1923        tcg_abort();
1924    }
1925
1926    /* Jump to the code corresponding to next IR of qemu_st */
1927    tcg_out_jmp(s, l->raddr);
1928    return true;
1929}
1930
1931/*
1932 * Generate code for the slow path for a store at the end of block
1933 */
1934static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
1935{
1936    MemOpIdx oi = l->oi;
1937    MemOp opc = get_memop(oi);
1938    MemOp s_bits = opc & MO_SIZE;
1939    tcg_insn_unit **label_ptr = &l->label_ptr[0];
1940    TCGReg retaddr;
1941
1942    /* resolve label address */
1943    tcg_patch32(label_ptr[0], s->code_ptr - label_ptr[0] - 4);
1944    if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
1945        tcg_patch32(label_ptr[1], s->code_ptr - label_ptr[1] - 4);
1946    }
1947
1948    if (TCG_TARGET_REG_BITS == 32) {
1949        int ofs = 0;
1950
1951        tcg_out_st(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP, ofs);
1952        ofs += 4;
1953
1954        tcg_out_st(s, TCG_TYPE_I32, l->addrlo_reg, TCG_REG_ESP, ofs);
1955        ofs += 4;
1956
1957        if (TARGET_LONG_BITS == 64) {
1958            tcg_out_st(s, TCG_TYPE_I32, l->addrhi_reg, TCG_REG_ESP, ofs);
1959            ofs += 4;
1960        }
1961
1962        tcg_out_st(s, TCG_TYPE_I32, l->datalo_reg, TCG_REG_ESP, ofs);
1963        ofs += 4;
1964
1965        if (s_bits == MO_64) {
1966            tcg_out_st(s, TCG_TYPE_I32, l->datahi_reg, TCG_REG_ESP, ofs);
1967            ofs += 4;
1968        }
1969
1970        tcg_out_sti(s, TCG_TYPE_I32, oi, TCG_REG_ESP, ofs);
1971        ofs += 4;
1972
1973        retaddr = TCG_REG_EAX;
1974        tcg_out_movi(s, TCG_TYPE_PTR, retaddr, (uintptr_t)l->raddr);
1975        tcg_out_st(s, TCG_TYPE_PTR, retaddr, TCG_REG_ESP, ofs);
1976    } else {
1977        tcg_out_mov(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[0], TCG_AREG0);
1978        /* The second argument is already loaded with addrlo.  */
1979        tcg_out_mov(s, (s_bits == MO_64 ? TCG_TYPE_I64 : TCG_TYPE_I32),
1980                    tcg_target_call_iarg_regs[2], l->datalo_reg);
1981        tcg_out_movi(s, TCG_TYPE_I32, tcg_target_call_iarg_regs[3], oi);
1982
1983        if (ARRAY_SIZE(tcg_target_call_iarg_regs) > 4) {
1984            retaddr = tcg_target_call_iarg_regs[4];
1985            tcg_out_movi(s, TCG_TYPE_PTR, retaddr, (uintptr_t)l->raddr);
1986        } else {
1987            retaddr = TCG_REG_RAX;
1988            tcg_out_movi(s, TCG_TYPE_PTR, retaddr, (uintptr_t)l->raddr);
1989            tcg_out_st(s, TCG_TYPE_PTR, retaddr, TCG_REG_ESP,
1990                       TCG_TARGET_CALL_STACK_OFFSET);
1991        }
1992    }
1993
1994    /* "Tail call" to the helper, with the return address back inline.  */
1995    tcg_out_push(s, retaddr);
1996    tcg_out_jmp(s, qemu_st_helpers[opc & (MO_BSWAP | MO_SIZE)]);
1997    return true;
1998}
1999#else
2000
2001static void tcg_out_test_alignment(TCGContext *s, bool is_ld, TCGReg addrlo,
2002                                   TCGReg addrhi, unsigned a_bits)
2003{
2004    unsigned a_mask = (1 << a_bits) - 1;
2005    TCGLabelQemuLdst *label;
2006
2007    /*
2008     * We are expecting a_bits to max out at 7, so we can usually use testb.
2009     * For i686, we have to use testl for %esi/%edi.
2010     */
2011    if (a_mask <= 0xff && (TCG_TARGET_REG_BITS == 64 || addrlo < 4)) {
2012        tcg_out_modrm(s, OPC_GRP3_Eb | P_REXB_RM, EXT3_TESTi, addrlo);
2013        tcg_out8(s, a_mask);
2014    } else {
2015        tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_TESTi, addrlo);
2016        tcg_out32(s, a_mask);
2017    }
2018
2019    /* jne slow_path */
2020    tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
2021
2022    label = new_ldst_label(s);
2023    label->is_ld = is_ld;
2024    label->addrlo_reg = addrlo;
2025    label->addrhi_reg = addrhi;
2026    label->raddr = tcg_splitwx_to_rx(s->code_ptr + 4);
2027    label->label_ptr[0] = s->code_ptr;
2028
2029    s->code_ptr += 4;
2030}
2031
2032static bool tcg_out_fail_alignment(TCGContext *s, TCGLabelQemuLdst *l)
2033{
2034    /* resolve label address */
2035    tcg_patch32(l->label_ptr[0], s->code_ptr - l->label_ptr[0] - 4);
2036
2037    if (TCG_TARGET_REG_BITS == 32) {
2038        int ofs = 0;
2039
2040        tcg_out_st(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP, ofs);
2041        ofs += 4;
2042
2043        tcg_out_st(s, TCG_TYPE_I32, l->addrlo_reg, TCG_REG_ESP, ofs);
2044        ofs += 4;
2045        if (TARGET_LONG_BITS == 64) {
2046            tcg_out_st(s, TCG_TYPE_I32, l->addrhi_reg, TCG_REG_ESP, ofs);
2047            ofs += 4;
2048        }
2049
2050        tcg_out_pushi(s, (uintptr_t)l->raddr);
2051    } else {
2052        tcg_out_mov(s, TCG_TYPE_TL, tcg_target_call_iarg_regs[1],
2053                    l->addrlo_reg);
2054        tcg_out_mov(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[0], TCG_AREG0);
2055
2056        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_RAX, (uintptr_t)l->raddr);
2057        tcg_out_push(s, TCG_REG_RAX);
2058    }
2059
2060    /* "Tail call" to the helper, with the return address back inline. */
2061    tcg_out_jmp(s, (const void *)(l->is_ld ? helper_unaligned_ld
2062                                  : helper_unaligned_st));
2063    return true;
2064}
2065
2066static bool tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
2067{
2068    return tcg_out_fail_alignment(s, l);
2069}
2070
2071static bool tcg_out_qemu_st_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
2072{
2073    return tcg_out_fail_alignment(s, l);
2074}
2075
2076#if TCG_TARGET_REG_BITS == 32
2077# define x86_guest_base_seg     0
2078# define x86_guest_base_index   -1
2079# define x86_guest_base_offset  guest_base
2080#else
2081static int x86_guest_base_seg;
2082static int x86_guest_base_index = -1;
2083static int32_t x86_guest_base_offset;
2084# if defined(__x86_64__) && defined(__linux__)
2085#  include <asm/prctl.h>
2086#  include <sys/prctl.h>
2087int arch_prctl(int code, unsigned long addr);
2088static inline int setup_guest_base_seg(void)
2089{
2090    if (arch_prctl(ARCH_SET_GS, guest_base) == 0) {
2091        return P_GS;
2092    }
2093    return 0;
2094}
2095# elif defined (__FreeBSD__) || defined (__FreeBSD_kernel__)
2096#  include <machine/sysarch.h>
2097static inline int setup_guest_base_seg(void)
2098{
2099    if (sysarch(AMD64_SET_GSBASE, &guest_base) == 0) {
2100        return P_GS;
2101    }
2102    return 0;
2103}
2104# else
2105static inline int setup_guest_base_seg(void)
2106{
2107    return 0;
2108}
2109# endif
2110#endif
2111#endif /* SOFTMMU */
2112
2113static void tcg_out_qemu_ld_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2114                                   TCGReg base, int index, intptr_t ofs,
2115                                   int seg, bool is64, MemOp memop)
2116{
2117    bool use_movbe = false;
2118    int rexw = is64 * P_REXW;
2119    int movop = OPC_MOVL_GvEv;
2120
2121    /* Do big-endian loads with movbe.  */
2122    if (memop & MO_BSWAP) {
2123        tcg_debug_assert(have_movbe);
2124        use_movbe = true;
2125        movop = OPC_MOVBE_GyMy;
2126    }
2127
2128    switch (memop & MO_SSIZE) {
2129    case MO_UB:
2130        tcg_out_modrm_sib_offset(s, OPC_MOVZBL + seg, datalo,
2131                                 base, index, 0, ofs);
2132        break;
2133    case MO_SB:
2134        tcg_out_modrm_sib_offset(s, OPC_MOVSBL + rexw + seg, datalo,
2135                                 base, index, 0, ofs);
2136        break;
2137    case MO_UW:
2138        if (use_movbe) {
2139            /* There is no extending movbe; only low 16-bits are modified.  */
2140            if (datalo != base && datalo != index) {
2141                /* XOR breaks dependency chains.  */
2142                tgen_arithr(s, ARITH_XOR, datalo, datalo);
2143                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + seg,
2144                                         datalo, base, index, 0, ofs);
2145            } else {
2146                tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + seg,
2147                                         datalo, base, index, 0, ofs);
2148                tcg_out_ext16u(s, datalo, datalo);
2149            }
2150        } else {
2151            tcg_out_modrm_sib_offset(s, OPC_MOVZWL + seg, datalo,
2152                                     base, index, 0, ofs);
2153        }
2154        break;
2155    case MO_SW:
2156        if (use_movbe) {
2157            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + P_DATA16 + seg,
2158                                     datalo, base, index, 0, ofs);
2159            tcg_out_ext16s(s, datalo, datalo, rexw);
2160        } else {
2161            tcg_out_modrm_sib_offset(s, OPC_MOVSWL + rexw + seg,
2162                                     datalo, base, index, 0, ofs);
2163        }
2164        break;
2165    case MO_UL:
2166        tcg_out_modrm_sib_offset(s, movop + seg, datalo, base, index, 0, ofs);
2167        break;
2168#if TCG_TARGET_REG_BITS == 64
2169    case MO_SL:
2170        if (use_movbe) {
2171            tcg_out_modrm_sib_offset(s, OPC_MOVBE_GyMy + seg, datalo,
2172                                     base, index, 0, ofs);
2173            tcg_out_ext32s(s, datalo, datalo);
2174        } else {
2175            tcg_out_modrm_sib_offset(s, OPC_MOVSLQ + seg, datalo,
2176                                     base, index, 0, ofs);
2177        }
2178        break;
2179#endif
2180    case MO_UQ:
2181        if (TCG_TARGET_REG_BITS == 64) {
2182            tcg_out_modrm_sib_offset(s, movop + P_REXW + seg, datalo,
2183                                     base, index, 0, ofs);
2184        } else {
2185            if (use_movbe) {
2186                TCGReg t = datalo;
2187                datalo = datahi;
2188                datahi = t;
2189            }
2190            if (base != datalo) {
2191                tcg_out_modrm_sib_offset(s, movop + seg, datalo,
2192                                         base, index, 0, ofs);
2193                tcg_out_modrm_sib_offset(s, movop + seg, datahi,
2194                                         base, index, 0, ofs + 4);
2195            } else {
2196                tcg_out_modrm_sib_offset(s, movop + seg, datahi,
2197                                         base, index, 0, ofs + 4);
2198                tcg_out_modrm_sib_offset(s, movop + seg, datalo,
2199                                         base, index, 0, ofs);
2200            }
2201        }
2202        break;
2203    default:
2204        g_assert_not_reached();
2205    }
2206}
2207
2208/* XXX: qemu_ld and qemu_st could be modified to clobber only EDX and
2209   EAX. It will be useful once fixed registers globals are less
2210   common. */
2211static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args, bool is64)
2212{
2213    TCGReg datalo, datahi, addrlo;
2214    TCGReg addrhi __attribute__((unused));
2215    MemOpIdx oi;
2216    MemOp opc;
2217#if defined(CONFIG_SOFTMMU)
2218    int mem_index;
2219    tcg_insn_unit *label_ptr[2];
2220#else
2221    unsigned a_bits;
2222#endif
2223
2224    datalo = *args++;
2225    datahi = (TCG_TARGET_REG_BITS == 32 && is64 ? *args++ : 0);
2226    addrlo = *args++;
2227    addrhi = (TARGET_LONG_BITS > TCG_TARGET_REG_BITS ? *args++ : 0);
2228    oi = *args++;
2229    opc = get_memop(oi);
2230
2231#if defined(CONFIG_SOFTMMU)
2232    mem_index = get_mmuidx(oi);
2233
2234    tcg_out_tlb_load(s, addrlo, addrhi, mem_index, opc,
2235                     label_ptr, offsetof(CPUTLBEntry, addr_read));
2236
2237    /* TLB Hit.  */
2238    tcg_out_qemu_ld_direct(s, datalo, datahi, TCG_REG_L1, -1, 0, 0, is64, opc);
2239
2240    /* Record the current context of a load into ldst label */
2241    add_qemu_ldst_label(s, true, is64, oi, datalo, datahi, addrlo, addrhi,
2242                        s->code_ptr, label_ptr);
2243#else
2244    a_bits = get_alignment_bits(opc);
2245    if (a_bits) {
2246        tcg_out_test_alignment(s, true, addrlo, addrhi, a_bits);
2247    }
2248
2249    tcg_out_qemu_ld_direct(s, datalo, datahi, addrlo, x86_guest_base_index,
2250                           x86_guest_base_offset, x86_guest_base_seg,
2251                           is64, opc);
2252#endif
2253}
2254
2255static void tcg_out_qemu_st_direct(TCGContext *s, TCGReg datalo, TCGReg datahi,
2256                                   TCGReg base, int index, intptr_t ofs,
2257                                   int seg, MemOp memop)
2258{
2259    bool use_movbe = false;
2260    int movop = OPC_MOVL_EvGv;
2261
2262    /*
2263     * Do big-endian stores with movbe or softmmu.
2264     * User-only without movbe will have its swapping done generically.
2265     */
2266    if (memop & MO_BSWAP) {
2267        tcg_debug_assert(have_movbe);
2268        use_movbe = true;
2269        movop = OPC_MOVBE_MyGy;
2270    }
2271
2272    switch (memop & MO_SIZE) {
2273    case MO_8:
2274        /* This is handled with constraints on INDEX_op_qemu_st8_i32. */
2275        tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || datalo < 4);
2276        tcg_out_modrm_sib_offset(s, OPC_MOVB_EvGv + P_REXB_R + seg,
2277                                 datalo, base, index, 0, ofs);
2278        break;
2279    case MO_16:
2280        tcg_out_modrm_sib_offset(s, movop + P_DATA16 + seg, datalo,
2281                                 base, index, 0, ofs);
2282        break;
2283    case MO_32:
2284        tcg_out_modrm_sib_offset(s, movop + seg, datalo, base, index, 0, ofs);
2285        break;
2286    case MO_64:
2287        if (TCG_TARGET_REG_BITS == 64) {
2288            tcg_out_modrm_sib_offset(s, movop + P_REXW + seg, datalo,
2289                                     base, index, 0, ofs);
2290        } else {
2291            if (use_movbe) {
2292                TCGReg t = datalo;
2293                datalo = datahi;
2294                datahi = t;
2295            }
2296            tcg_out_modrm_sib_offset(s, movop + seg, datalo,
2297                                     base, index, 0, ofs);
2298            tcg_out_modrm_sib_offset(s, movop + seg, datahi,
2299                                     base, index, 0, ofs + 4);
2300        }
2301        break;
2302    default:
2303        g_assert_not_reached();
2304    }
2305}
2306
2307static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is64)
2308{
2309    TCGReg datalo, datahi, addrlo;
2310    TCGReg addrhi __attribute__((unused));
2311    MemOpIdx oi;
2312    MemOp opc;
2313#if defined(CONFIG_SOFTMMU)
2314    int mem_index;
2315    tcg_insn_unit *label_ptr[2];
2316#else
2317    unsigned a_bits;
2318#endif
2319
2320    datalo = *args++;
2321    datahi = (TCG_TARGET_REG_BITS == 32 && is64 ? *args++ : 0);
2322    addrlo = *args++;
2323    addrhi = (TARGET_LONG_BITS > TCG_TARGET_REG_BITS ? *args++ : 0);
2324    oi = *args++;
2325    opc = get_memop(oi);
2326
2327#if defined(CONFIG_SOFTMMU)
2328    mem_index = get_mmuidx(oi);
2329
2330    tcg_out_tlb_load(s, addrlo, addrhi, mem_index, opc,
2331                     label_ptr, offsetof(CPUTLBEntry, addr_write));
2332
2333    /* TLB Hit.  */
2334    tcg_out_qemu_st_direct(s, datalo, datahi, TCG_REG_L1, -1, 0, 0, opc);
2335
2336    /* Record the current context of a store into ldst label */
2337    add_qemu_ldst_label(s, false, is64, oi, datalo, datahi, addrlo, addrhi,
2338                        s->code_ptr, label_ptr);
2339#else
2340    a_bits = get_alignment_bits(opc);
2341    if (a_bits) {
2342        tcg_out_test_alignment(s, false, addrlo, addrhi, a_bits);
2343    }
2344
2345    tcg_out_qemu_st_direct(s, datalo, datahi, addrlo, x86_guest_base_index,
2346                           x86_guest_base_offset, x86_guest_base_seg, opc);
2347#endif
2348}
2349
2350static void tcg_out_exit_tb(TCGContext *s, uintptr_t a0)
2351{
2352    /* Reuse the zeroing that exists for goto_ptr.  */
2353    if (a0 == 0) {
2354        tcg_out_jmp(s, tcg_code_gen_epilogue);
2355    } else {
2356        tcg_out_movi(s, TCG_TYPE_PTR, TCG_REG_EAX, a0);
2357        tcg_out_jmp(s, tb_ret_addr);
2358    }
2359}
2360
2361static void tcg_out_goto_tb(TCGContext *s, int which)
2362{
2363    /*
2364     * Jump displacement must be aligned for atomic patching;
2365     * see if we need to add extra nops before jump
2366     */
2367    int gap = QEMU_ALIGN_PTR_UP(s->code_ptr + 1, 4) - s->code_ptr;
2368    if (gap != 1) {
2369        tcg_out_nopn(s, gap - 1);
2370    }
2371    tcg_out8(s, OPC_JMP_long); /* jmp im */
2372    set_jmp_insn_offset(s, which);
2373    tcg_out32(s, 0);
2374    set_jmp_reset_offset(s, which);
2375}
2376
2377void tb_target_set_jmp_target(const TranslationBlock *tb, int n,
2378                              uintptr_t jmp_rx, uintptr_t jmp_rw)
2379{
2380    /* patch the branch destination */
2381    uintptr_t addr = tb->jmp_target_addr[n];
2382    qatomic_set((int32_t *)jmp_rw, addr - (jmp_rx + 4));
2383    /* no need to flush icache explicitly */
2384}
2385
2386static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
2387                              const TCGArg args[TCG_MAX_OP_ARGS],
2388                              const int const_args[TCG_MAX_OP_ARGS])
2389{
2390    TCGArg a0, a1, a2;
2391    int c, const_a2, vexop, rexw = 0;
2392
2393#if TCG_TARGET_REG_BITS == 64
2394# define OP_32_64(x) \
2395        case glue(glue(INDEX_op_, x), _i64): \
2396            rexw = P_REXW; /* FALLTHRU */    \
2397        case glue(glue(INDEX_op_, x), _i32)
2398#else
2399# define OP_32_64(x) \
2400        case glue(glue(INDEX_op_, x), _i32)
2401#endif
2402
2403    /* Hoist the loads of the most common arguments.  */
2404    a0 = args[0];
2405    a1 = args[1];
2406    a2 = args[2];
2407    const_a2 = const_args[2];
2408
2409    switch (opc) {
2410    case INDEX_op_goto_ptr:
2411        /* jmp to the given host address (could be epilogue) */
2412        tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, a0);
2413        break;
2414    case INDEX_op_br:
2415        tcg_out_jxx(s, JCC_JMP, arg_label(a0), 0);
2416        break;
2417    OP_32_64(ld8u):
2418        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
2419        tcg_out_modrm_offset(s, OPC_MOVZBL, a0, a1, a2);
2420        break;
2421    OP_32_64(ld8s):
2422        tcg_out_modrm_offset(s, OPC_MOVSBL + rexw, a0, a1, a2);
2423        break;
2424    OP_32_64(ld16u):
2425        /* Note that we can ignore REXW for the zero-extend to 64-bit.  */
2426        tcg_out_modrm_offset(s, OPC_MOVZWL, a0, a1, a2);
2427        break;
2428    OP_32_64(ld16s):
2429        tcg_out_modrm_offset(s, OPC_MOVSWL + rexw, a0, a1, a2);
2430        break;
2431#if TCG_TARGET_REG_BITS == 64
2432    case INDEX_op_ld32u_i64:
2433#endif
2434    case INDEX_op_ld_i32:
2435        tcg_out_ld(s, TCG_TYPE_I32, a0, a1, a2);
2436        break;
2437
2438    OP_32_64(st8):
2439        if (const_args[0]) {
2440            tcg_out_modrm_offset(s, OPC_MOVB_EvIz, 0, a1, a2);
2441            tcg_out8(s, a0);
2442        } else {
2443            tcg_out_modrm_offset(s, OPC_MOVB_EvGv | P_REXB_R, a0, a1, a2);
2444        }
2445        break;
2446    OP_32_64(st16):
2447        if (const_args[0]) {
2448            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_DATA16, 0, a1, a2);
2449            tcg_out16(s, a0);
2450        } else {
2451            tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_DATA16, a0, a1, a2);
2452        }
2453        break;
2454#if TCG_TARGET_REG_BITS == 64
2455    case INDEX_op_st32_i64:
2456#endif
2457    case INDEX_op_st_i32:
2458        if (const_args[0]) {
2459            tcg_out_modrm_offset(s, OPC_MOVL_EvIz, 0, a1, a2);
2460            tcg_out32(s, a0);
2461        } else {
2462            tcg_out_st(s, TCG_TYPE_I32, a0, a1, a2);
2463        }
2464        break;
2465
2466    OP_32_64(add):
2467        /* For 3-operand addition, use LEA.  */
2468        if (a0 != a1) {
2469            TCGArg c3 = 0;
2470            if (const_a2) {
2471                c3 = a2, a2 = -1;
2472            } else if (a0 == a2) {
2473                /* Watch out for dest = src + dest, since we've removed
2474                   the matching constraint on the add.  */
2475                tgen_arithr(s, ARITH_ADD + rexw, a0, a1);
2476                break;
2477            }
2478
2479            tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a2, 0, c3);
2480            break;
2481        }
2482        c = ARITH_ADD;
2483        goto gen_arith;
2484    OP_32_64(sub):
2485        c = ARITH_SUB;
2486        goto gen_arith;
2487    OP_32_64(and):
2488        c = ARITH_AND;
2489        goto gen_arith;
2490    OP_32_64(or):
2491        c = ARITH_OR;
2492        goto gen_arith;
2493    OP_32_64(xor):
2494        c = ARITH_XOR;
2495        goto gen_arith;
2496    gen_arith:
2497        if (const_a2) {
2498            tgen_arithi(s, c + rexw, a0, a2, 0);
2499        } else {
2500            tgen_arithr(s, c + rexw, a0, a2);
2501        }
2502        break;
2503
2504    OP_32_64(andc):
2505        if (const_a2) {
2506            tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, a0, a1);
2507            tgen_arithi(s, ARITH_AND + rexw, a0, ~a2, 0);
2508        } else {
2509            tcg_out_vex_modrm(s, OPC_ANDN + rexw, a0, a2, a1);
2510        }
2511        break;
2512
2513    OP_32_64(mul):
2514        if (const_a2) {
2515            int32_t val;
2516            val = a2;
2517            if (val == (int8_t)val) {
2518                tcg_out_modrm(s, OPC_IMUL_GvEvIb + rexw, a0, a0);
2519                tcg_out8(s, val);
2520            } else {
2521                tcg_out_modrm(s, OPC_IMUL_GvEvIz + rexw, a0, a0);
2522                tcg_out32(s, val);
2523            }
2524        } else {
2525            tcg_out_modrm(s, OPC_IMUL_GvEv + rexw, a0, a2);
2526        }
2527        break;
2528
2529    OP_32_64(div2):
2530        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IDIV, args[4]);
2531        break;
2532    OP_32_64(divu2):
2533        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_DIV, args[4]);
2534        break;
2535
2536    OP_32_64(shl):
2537        /* For small constant 3-operand shift, use LEA.  */
2538        if (const_a2 && a0 != a1 && (a2 - 1) < 3) {
2539            if (a2 - 1 == 0) {
2540                /* shl $1,a1,a0 -> lea (a1,a1),a0 */
2541                tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, a1, a1, 0, 0);
2542            } else {
2543                /* shl $n,a1,a0 -> lea 0(,a1,n),a0 */
2544                tcg_out_modrm_sib_offset(s, OPC_LEA + rexw, a0, -1, a1, a2, 0);
2545            }
2546            break;
2547        }
2548        c = SHIFT_SHL;
2549        vexop = OPC_SHLX;
2550        goto gen_shift_maybe_vex;
2551    OP_32_64(shr):
2552        c = SHIFT_SHR;
2553        vexop = OPC_SHRX;
2554        goto gen_shift_maybe_vex;
2555    OP_32_64(sar):
2556        c = SHIFT_SAR;
2557        vexop = OPC_SARX;
2558        goto gen_shift_maybe_vex;
2559    OP_32_64(rotl):
2560        c = SHIFT_ROL;
2561        goto gen_shift;
2562    OP_32_64(rotr):
2563        c = SHIFT_ROR;
2564        goto gen_shift;
2565    gen_shift_maybe_vex:
2566        if (have_bmi2) {
2567            if (!const_a2) {
2568                tcg_out_vex_modrm(s, vexop + rexw, a0, a2, a1);
2569                break;
2570            }
2571            tcg_out_mov(s, rexw ? TCG_TYPE_I64 : TCG_TYPE_I32, a0, a1);
2572        }
2573        /* FALLTHRU */
2574    gen_shift:
2575        if (const_a2) {
2576            tcg_out_shifti(s, c + rexw, a0, a2);
2577        } else {
2578            tcg_out_modrm(s, OPC_SHIFT_cl + rexw, c, a0);
2579        }
2580        break;
2581
2582    OP_32_64(ctz):
2583        tcg_out_ctz(s, rexw, args[0], args[1], args[2], const_args[2]);
2584        break;
2585    OP_32_64(clz):
2586        tcg_out_clz(s, rexw, args[0], args[1], args[2], const_args[2]);
2587        break;
2588    OP_32_64(ctpop):
2589        tcg_out_modrm(s, OPC_POPCNT + rexw, a0, a1);
2590        break;
2591
2592    case INDEX_op_brcond_i32:
2593        tcg_out_brcond32(s, a2, a0, a1, const_args[1], arg_label(args[3]), 0);
2594        break;
2595    case INDEX_op_setcond_i32:
2596        tcg_out_setcond32(s, args[3], a0, a1, a2, const_a2);
2597        break;
2598    case INDEX_op_movcond_i32:
2599        tcg_out_movcond32(s, args[5], a0, a1, a2, const_a2, args[3]);
2600        break;
2601
2602    OP_32_64(bswap16):
2603        if (a2 & TCG_BSWAP_OS) {
2604            /* Output must be sign-extended. */
2605            if (rexw) {
2606                tcg_out_bswap64(s, a0);
2607                tcg_out_shifti(s, SHIFT_SAR + rexw, a0, 48);
2608            } else {
2609                tcg_out_bswap32(s, a0);
2610                tcg_out_shifti(s, SHIFT_SAR, a0, 16);
2611            }
2612        } else if ((a2 & (TCG_BSWAP_IZ | TCG_BSWAP_OZ)) == TCG_BSWAP_OZ) {
2613            /* Output must be zero-extended, but input isn't. */
2614            tcg_out_bswap32(s, a0);
2615            tcg_out_shifti(s, SHIFT_SHR, a0, 16);
2616        } else {
2617            tcg_out_rolw_8(s, a0);
2618        }
2619        break;
2620    OP_32_64(bswap32):
2621        tcg_out_bswap32(s, a0);
2622        if (rexw && (a2 & TCG_BSWAP_OS)) {
2623            tcg_out_ext32s(s, a0, a0);
2624        }
2625        break;
2626
2627    OP_32_64(neg):
2628        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NEG, a0);
2629        break;
2630    OP_32_64(not):
2631        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_NOT, a0);
2632        break;
2633
2634    OP_32_64(ext8s):
2635        tcg_out_ext8s(s, a0, a1, rexw);
2636        break;
2637    OP_32_64(ext16s):
2638        tcg_out_ext16s(s, a0, a1, rexw);
2639        break;
2640    OP_32_64(ext8u):
2641        tcg_out_ext8u(s, a0, a1);
2642        break;
2643    OP_32_64(ext16u):
2644        tcg_out_ext16u(s, a0, a1);
2645        break;
2646
2647    case INDEX_op_qemu_ld_i32:
2648        tcg_out_qemu_ld(s, args, 0);
2649        break;
2650    case INDEX_op_qemu_ld_i64:
2651        tcg_out_qemu_ld(s, args, 1);
2652        break;
2653    case INDEX_op_qemu_st_i32:
2654    case INDEX_op_qemu_st8_i32:
2655        tcg_out_qemu_st(s, args, 0);
2656        break;
2657    case INDEX_op_qemu_st_i64:
2658        tcg_out_qemu_st(s, args, 1);
2659        break;
2660
2661    OP_32_64(mulu2):
2662        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_MUL, args[3]);
2663        break;
2664    OP_32_64(muls2):
2665        tcg_out_modrm(s, OPC_GRP3_Ev + rexw, EXT3_IMUL, args[3]);
2666        break;
2667    OP_32_64(add2):
2668        if (const_args[4]) {
2669            tgen_arithi(s, ARITH_ADD + rexw, a0, args[4], 1);
2670        } else {
2671            tgen_arithr(s, ARITH_ADD + rexw, a0, args[4]);
2672        }
2673        if (const_args[5]) {
2674            tgen_arithi(s, ARITH_ADC + rexw, a1, args[5], 1);
2675        } else {
2676            tgen_arithr(s, ARITH_ADC + rexw, a1, args[5]);
2677        }
2678        break;
2679    OP_32_64(sub2):
2680        if (const_args[4]) {
2681            tgen_arithi(s, ARITH_SUB + rexw, a0, args[4], 1);
2682        } else {
2683            tgen_arithr(s, ARITH_SUB + rexw, a0, args[4]);
2684        }
2685        if (const_args[5]) {
2686            tgen_arithi(s, ARITH_SBB + rexw, a1, args[5], 1);
2687        } else {
2688            tgen_arithr(s, ARITH_SBB + rexw, a1, args[5]);
2689        }
2690        break;
2691
2692#if TCG_TARGET_REG_BITS == 32
2693    case INDEX_op_brcond2_i32:
2694        tcg_out_brcond2(s, args, const_args, 0);
2695        break;
2696    case INDEX_op_setcond2_i32:
2697        tcg_out_setcond2(s, args, const_args);
2698        break;
2699#else /* TCG_TARGET_REG_BITS == 64 */
2700    case INDEX_op_ld32s_i64:
2701        tcg_out_modrm_offset(s, OPC_MOVSLQ, a0, a1, a2);
2702        break;
2703    case INDEX_op_ld_i64:
2704        tcg_out_ld(s, TCG_TYPE_I64, a0, a1, a2);
2705        break;
2706    case INDEX_op_st_i64:
2707        if (const_args[0]) {
2708            tcg_out_modrm_offset(s, OPC_MOVL_EvIz | P_REXW, 0, a1, a2);
2709            tcg_out32(s, a0);
2710        } else {
2711            tcg_out_st(s, TCG_TYPE_I64, a0, a1, a2);
2712        }
2713        break;
2714
2715    case INDEX_op_brcond_i64:
2716        tcg_out_brcond64(s, a2, a0, a1, const_args[1], arg_label(args[3]), 0);
2717        break;
2718    case INDEX_op_setcond_i64:
2719        tcg_out_setcond64(s, args[3], a0, a1, a2, const_a2);
2720        break;
2721    case INDEX_op_movcond_i64:
2722        tcg_out_movcond64(s, args[5], a0, a1, a2, const_a2, args[3]);
2723        break;
2724
2725    case INDEX_op_bswap64_i64:
2726        tcg_out_bswap64(s, a0);
2727        break;
2728    case INDEX_op_extu_i32_i64:
2729    case INDEX_op_ext32u_i64:
2730    case INDEX_op_extrl_i64_i32:
2731        tcg_out_ext32u(s, a0, a1);
2732        break;
2733    case INDEX_op_ext_i32_i64:
2734    case INDEX_op_ext32s_i64:
2735        tcg_out_ext32s(s, a0, a1);
2736        break;
2737    case INDEX_op_extrh_i64_i32:
2738        tcg_out_shifti(s, SHIFT_SHR + P_REXW, a0, 32);
2739        break;
2740#endif
2741
2742    OP_32_64(deposit):
2743        if (args[3] == 0 && args[4] == 8) {
2744            /* load bits 0..7 */
2745            tcg_out_modrm(s, OPC_MOVB_EvGv | P_REXB_R | P_REXB_RM, a2, a0);
2746        } else if (args[3] == 8 && args[4] == 8) {
2747            /* load bits 8..15 */
2748            tcg_out_modrm(s, OPC_MOVB_EvGv, a2, a0 + 4);
2749        } else if (args[3] == 0 && args[4] == 16) {
2750            /* load bits 0..15 */
2751            tcg_out_modrm(s, OPC_MOVL_EvGv | P_DATA16, a2, a0);
2752        } else {
2753            tcg_abort();
2754        }
2755        break;
2756
2757    case INDEX_op_extract_i64:
2758        if (a2 + args[3] == 32) {
2759            /* This is a 32-bit zero-extending right shift.  */
2760            tcg_out_mov(s, TCG_TYPE_I32, a0, a1);
2761            tcg_out_shifti(s, SHIFT_SHR, a0, a2);
2762            break;
2763        }
2764        /* FALLTHRU */
2765    case INDEX_op_extract_i32:
2766        /* On the off-chance that we can use the high-byte registers.
2767           Otherwise we emit the same ext16 + shift pattern that we
2768           would have gotten from the normal tcg-op.c expansion.  */
2769        tcg_debug_assert(a2 == 8 && args[3] == 8);
2770        if (a1 < 4 && a0 < 8) {
2771            tcg_out_modrm(s, OPC_MOVZBL, a0, a1 + 4);
2772        } else {
2773            tcg_out_ext16u(s, a0, a1);
2774            tcg_out_shifti(s, SHIFT_SHR, a0, 8);
2775        }
2776        break;
2777
2778    case INDEX_op_sextract_i32:
2779        /* We don't implement sextract_i64, as we cannot sign-extend to
2780           64-bits without using the REX prefix that explicitly excludes
2781           access to the high-byte registers.  */
2782        tcg_debug_assert(a2 == 8 && args[3] == 8);
2783        if (a1 < 4 && a0 < 8) {
2784            tcg_out_modrm(s, OPC_MOVSBL, a0, a1 + 4);
2785        } else {
2786            tcg_out_ext16s(s, a0, a1, 0);
2787            tcg_out_shifti(s, SHIFT_SAR, a0, 8);
2788        }
2789        break;
2790
2791    OP_32_64(extract2):
2792        /* Note that SHRD outputs to the r/m operand.  */
2793        tcg_out_modrm(s, OPC_SHRD_Ib + rexw, a2, a0);
2794        tcg_out8(s, args[3]);
2795        break;
2796
2797    case INDEX_op_mb:
2798        tcg_out_mb(s, a0);
2799        break;
2800    case INDEX_op_mov_i32:  /* Always emitted via tcg_out_mov.  */
2801    case INDEX_op_mov_i64:
2802    case INDEX_op_call:     /* Always emitted via tcg_out_call.  */
2803    case INDEX_op_exit_tb:  /* Always emitted via tcg_out_exit_tb.  */
2804    case INDEX_op_goto_tb:  /* Always emitted via tcg_out_goto_tb.  */
2805    default:
2806        tcg_abort();
2807    }
2808
2809#undef OP_32_64
2810}
2811
2812static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
2813                           unsigned vecl, unsigned vece,
2814                           const TCGArg args[TCG_MAX_OP_ARGS],
2815                           const int const_args[TCG_MAX_OP_ARGS])
2816{
2817    static int const add_insn[4] = {
2818        OPC_PADDB, OPC_PADDW, OPC_PADDD, OPC_PADDQ
2819    };
2820    static int const ssadd_insn[4] = {
2821        OPC_PADDSB, OPC_PADDSW, OPC_UD2, OPC_UD2
2822    };
2823    static int const usadd_insn[4] = {
2824        OPC_PADDUB, OPC_PADDUW, OPC_UD2, OPC_UD2
2825    };
2826    static int const sub_insn[4] = {
2827        OPC_PSUBB, OPC_PSUBW, OPC_PSUBD, OPC_PSUBQ
2828    };
2829    static int const sssub_insn[4] = {
2830        OPC_PSUBSB, OPC_PSUBSW, OPC_UD2, OPC_UD2
2831    };
2832    static int const ussub_insn[4] = {
2833        OPC_PSUBUB, OPC_PSUBUW, OPC_UD2, OPC_UD2
2834    };
2835    static int const mul_insn[4] = {
2836        OPC_UD2, OPC_PMULLW, OPC_PMULLD, OPC_VPMULLQ
2837    };
2838    static int const shift_imm_insn[4] = {
2839        OPC_UD2, OPC_PSHIFTW_Ib, OPC_PSHIFTD_Ib, OPC_PSHIFTQ_Ib
2840    };
2841    static int const cmpeq_insn[4] = {
2842        OPC_PCMPEQB, OPC_PCMPEQW, OPC_PCMPEQD, OPC_PCMPEQQ
2843    };
2844    static int const cmpgt_insn[4] = {
2845        OPC_PCMPGTB, OPC_PCMPGTW, OPC_PCMPGTD, OPC_PCMPGTQ
2846    };
2847    static int const punpckl_insn[4] = {
2848        OPC_PUNPCKLBW, OPC_PUNPCKLWD, OPC_PUNPCKLDQ, OPC_PUNPCKLQDQ
2849    };
2850    static int const punpckh_insn[4] = {
2851        OPC_PUNPCKHBW, OPC_PUNPCKHWD, OPC_PUNPCKHDQ, OPC_PUNPCKHQDQ
2852    };
2853    static int const packss_insn[4] = {
2854        OPC_PACKSSWB, OPC_PACKSSDW, OPC_UD2, OPC_UD2
2855    };
2856    static int const packus_insn[4] = {
2857        OPC_PACKUSWB, OPC_PACKUSDW, OPC_UD2, OPC_UD2
2858    };
2859    static int const smin_insn[4] = {
2860        OPC_PMINSB, OPC_PMINSW, OPC_PMINSD, OPC_VPMINSQ
2861    };
2862    static int const smax_insn[4] = {
2863        OPC_PMAXSB, OPC_PMAXSW, OPC_PMAXSD, OPC_VPMAXSQ
2864    };
2865    static int const umin_insn[4] = {
2866        OPC_PMINUB, OPC_PMINUW, OPC_PMINUD, OPC_VPMINUQ
2867    };
2868    static int const umax_insn[4] = {
2869        OPC_PMAXUB, OPC_PMAXUW, OPC_PMAXUD, OPC_VPMAXUQ
2870    };
2871    static int const rotlv_insn[4] = {
2872        OPC_UD2, OPC_UD2, OPC_VPROLVD, OPC_VPROLVQ
2873    };
2874    static int const rotrv_insn[4] = {
2875        OPC_UD2, OPC_UD2, OPC_VPRORVD, OPC_VPRORVQ
2876    };
2877    static int const shlv_insn[4] = {
2878        OPC_UD2, OPC_VPSLLVW, OPC_VPSLLVD, OPC_VPSLLVQ
2879    };
2880    static int const shrv_insn[4] = {
2881        OPC_UD2, OPC_VPSRLVW, OPC_VPSRLVD, OPC_VPSRLVQ
2882    };
2883    static int const sarv_insn[4] = {
2884        OPC_UD2, OPC_VPSRAVW, OPC_VPSRAVD, OPC_VPSRAVQ
2885    };
2886    static int const shls_insn[4] = {
2887        OPC_UD2, OPC_PSLLW, OPC_PSLLD, OPC_PSLLQ
2888    };
2889    static int const shrs_insn[4] = {
2890        OPC_UD2, OPC_PSRLW, OPC_PSRLD, OPC_PSRLQ
2891    };
2892    static int const sars_insn[4] = {
2893        OPC_UD2, OPC_PSRAW, OPC_PSRAD, OPC_VPSRAQ
2894    };
2895    static int const vpshldi_insn[4] = {
2896        OPC_UD2, OPC_VPSHLDW, OPC_VPSHLDD, OPC_VPSHLDQ
2897    };
2898    static int const vpshldv_insn[4] = {
2899        OPC_UD2, OPC_VPSHLDVW, OPC_VPSHLDVD, OPC_VPSHLDVQ
2900    };
2901    static int const vpshrdv_insn[4] = {
2902        OPC_UD2, OPC_VPSHRDVW, OPC_VPSHRDVD, OPC_VPSHRDVQ
2903    };
2904    static int const abs_insn[4] = {
2905        OPC_PABSB, OPC_PABSW, OPC_PABSD, OPC_VPABSQ
2906    };
2907
2908    TCGType type = vecl + TCG_TYPE_V64;
2909    int insn, sub;
2910    TCGArg a0, a1, a2, a3;
2911
2912    a0 = args[0];
2913    a1 = args[1];
2914    a2 = args[2];
2915
2916    switch (opc) {
2917    case INDEX_op_add_vec:
2918        insn = add_insn[vece];
2919        goto gen_simd;
2920    case INDEX_op_ssadd_vec:
2921        insn = ssadd_insn[vece];
2922        goto gen_simd;
2923    case INDEX_op_usadd_vec:
2924        insn = usadd_insn[vece];
2925        goto gen_simd;
2926    case INDEX_op_sub_vec:
2927        insn = sub_insn[vece];
2928        goto gen_simd;
2929    case INDEX_op_sssub_vec:
2930        insn = sssub_insn[vece];
2931        goto gen_simd;
2932    case INDEX_op_ussub_vec:
2933        insn = ussub_insn[vece];
2934        goto gen_simd;
2935    case INDEX_op_mul_vec:
2936        insn = mul_insn[vece];
2937        goto gen_simd;
2938    case INDEX_op_and_vec:
2939        insn = OPC_PAND;
2940        goto gen_simd;
2941    case INDEX_op_or_vec:
2942        insn = OPC_POR;
2943        goto gen_simd;
2944    case INDEX_op_xor_vec:
2945        insn = OPC_PXOR;
2946        goto gen_simd;
2947    case INDEX_op_smin_vec:
2948        insn = smin_insn[vece];
2949        goto gen_simd;
2950    case INDEX_op_umin_vec:
2951        insn = umin_insn[vece];
2952        goto gen_simd;
2953    case INDEX_op_smax_vec:
2954        insn = smax_insn[vece];
2955        goto gen_simd;
2956    case INDEX_op_umax_vec:
2957        insn = umax_insn[vece];
2958        goto gen_simd;
2959    case INDEX_op_shlv_vec:
2960        insn = shlv_insn[vece];
2961        goto gen_simd;
2962    case INDEX_op_shrv_vec:
2963        insn = shrv_insn[vece];
2964        goto gen_simd;
2965    case INDEX_op_sarv_vec:
2966        insn = sarv_insn[vece];
2967        goto gen_simd;
2968    case INDEX_op_rotlv_vec:
2969        insn = rotlv_insn[vece];
2970        goto gen_simd;
2971    case INDEX_op_rotrv_vec:
2972        insn = rotrv_insn[vece];
2973        goto gen_simd;
2974    case INDEX_op_shls_vec:
2975        insn = shls_insn[vece];
2976        goto gen_simd;
2977    case INDEX_op_shrs_vec:
2978        insn = shrs_insn[vece];
2979        goto gen_simd;
2980    case INDEX_op_sars_vec:
2981        insn = sars_insn[vece];
2982        goto gen_simd;
2983    case INDEX_op_x86_punpckl_vec:
2984        insn = punpckl_insn[vece];
2985        goto gen_simd;
2986    case INDEX_op_x86_punpckh_vec:
2987        insn = punpckh_insn[vece];
2988        goto gen_simd;
2989    case INDEX_op_x86_packss_vec:
2990        insn = packss_insn[vece];
2991        goto gen_simd;
2992    case INDEX_op_x86_packus_vec:
2993        insn = packus_insn[vece];
2994        goto gen_simd;
2995    case INDEX_op_x86_vpshldv_vec:
2996        insn = vpshldv_insn[vece];
2997        a1 = a2;
2998        a2 = args[3];
2999        goto gen_simd;
3000    case INDEX_op_x86_vpshrdv_vec:
3001        insn = vpshrdv_insn[vece];
3002        a1 = a2;
3003        a2 = args[3];
3004        goto gen_simd;
3005#if TCG_TARGET_REG_BITS == 32
3006    case INDEX_op_dup2_vec:
3007        /* First merge the two 32-bit inputs to a single 64-bit element. */
3008        tcg_out_vex_modrm(s, OPC_PUNPCKLDQ, a0, a1, a2);
3009        /* Then replicate the 64-bit elements across the rest of the vector. */
3010        if (type != TCG_TYPE_V64) {
3011            tcg_out_dup_vec(s, type, MO_64, a0, a0);
3012        }
3013        break;
3014#endif
3015    case INDEX_op_abs_vec:
3016        insn = abs_insn[vece];
3017        a2 = a1;
3018        a1 = 0;
3019        goto gen_simd;
3020    gen_simd:
3021        tcg_debug_assert(insn != OPC_UD2);
3022        if (type == TCG_TYPE_V256) {
3023            insn |= P_VEXL;
3024        }
3025        tcg_out_vex_modrm(s, insn, a0, a1, a2);
3026        break;
3027
3028    case INDEX_op_cmp_vec:
3029        sub = args[3];
3030        if (sub == TCG_COND_EQ) {
3031            insn = cmpeq_insn[vece];
3032        } else if (sub == TCG_COND_GT) {
3033            insn = cmpgt_insn[vece];
3034        } else {
3035            g_assert_not_reached();
3036        }
3037        goto gen_simd;
3038
3039    case INDEX_op_andc_vec:
3040        insn = OPC_PANDN;
3041        if (type == TCG_TYPE_V256) {
3042            insn |= P_VEXL;
3043        }
3044        tcg_out_vex_modrm(s, insn, a0, a2, a1);
3045        break;
3046
3047    case INDEX_op_shli_vec:
3048        insn = shift_imm_insn[vece];
3049        sub = 6;
3050        goto gen_shift;
3051    case INDEX_op_shri_vec:
3052        insn = shift_imm_insn[vece];
3053        sub = 2;
3054        goto gen_shift;
3055    case INDEX_op_sari_vec:
3056        if (vece == MO_64) {
3057            insn = OPC_PSHIFTD_Ib | P_VEXW | P_EVEX;
3058        } else {
3059            insn = shift_imm_insn[vece];
3060        }
3061        sub = 4;
3062        goto gen_shift;
3063    case INDEX_op_rotli_vec:
3064        insn = OPC_PSHIFTD_Ib | P_EVEX;  /* VPROL[DQ] */
3065        if (vece == MO_64) {
3066            insn |= P_VEXW;
3067        }
3068        sub = 1;
3069        goto gen_shift;
3070    gen_shift:
3071        tcg_debug_assert(vece != MO_8);
3072        if (type == TCG_TYPE_V256) {
3073            insn |= P_VEXL;
3074        }
3075        tcg_out_vex_modrm(s, insn, sub, a0, a1);
3076        tcg_out8(s, a2);
3077        break;
3078
3079    case INDEX_op_ld_vec:
3080        tcg_out_ld(s, type, a0, a1, a2);
3081        break;
3082    case INDEX_op_st_vec:
3083        tcg_out_st(s, type, a0, a1, a2);
3084        break;
3085    case INDEX_op_dupm_vec:
3086        tcg_out_dupm_vec(s, type, vece, a0, a1, a2);
3087        break;
3088
3089    case INDEX_op_x86_shufps_vec:
3090        insn = OPC_SHUFPS;
3091        sub = args[3];
3092        goto gen_simd_imm8;
3093    case INDEX_op_x86_blend_vec:
3094        if (vece == MO_16) {
3095            insn = OPC_PBLENDW;
3096        } else if (vece == MO_32) {
3097            insn = (have_avx2 ? OPC_VPBLENDD : OPC_BLENDPS);
3098        } else {
3099            g_assert_not_reached();
3100        }
3101        sub = args[3];
3102        goto gen_simd_imm8;
3103    case INDEX_op_x86_vperm2i128_vec:
3104        insn = OPC_VPERM2I128;
3105        sub = args[3];
3106        goto gen_simd_imm8;
3107    case INDEX_op_x86_vpshldi_vec:
3108        insn = vpshldi_insn[vece];
3109        sub = args[3];
3110        goto gen_simd_imm8;
3111
3112    case INDEX_op_not_vec:
3113        insn = OPC_VPTERNLOGQ;
3114        a2 = a1;
3115        sub = 0x33; /* !B */
3116        goto gen_simd_imm8;
3117    case INDEX_op_nor_vec:
3118        insn = OPC_VPTERNLOGQ;
3119        sub = 0x11; /* norCB */
3120        goto gen_simd_imm8;
3121    case INDEX_op_nand_vec:
3122        insn = OPC_VPTERNLOGQ;
3123        sub = 0x77; /* nandCB */
3124        goto gen_simd_imm8;
3125    case INDEX_op_eqv_vec:
3126        insn = OPC_VPTERNLOGQ;
3127        sub = 0x99; /* xnorCB */
3128        goto gen_simd_imm8;
3129    case INDEX_op_orc_vec:
3130        insn = OPC_VPTERNLOGQ;
3131        sub = 0xdd; /* orB!C */
3132        goto gen_simd_imm8;
3133
3134    case INDEX_op_bitsel_vec:
3135        insn = OPC_VPTERNLOGQ;
3136        a3 = args[3];
3137        if (a0 == a1) {
3138            a1 = a2;
3139            a2 = a3;
3140            sub = 0xca; /* A?B:C */
3141        } else if (a0 == a2) {
3142            a2 = a3;
3143            sub = 0xe2; /* B?A:C */
3144        } else {
3145            tcg_out_mov(s, type, a0, a3);
3146            sub = 0xb8; /* B?C:A */
3147        }
3148        goto gen_simd_imm8;
3149
3150    gen_simd_imm8:
3151        tcg_debug_assert(insn != OPC_UD2);
3152        if (type == TCG_TYPE_V256) {
3153            insn |= P_VEXL;
3154        }
3155        tcg_out_vex_modrm(s, insn, a0, a1, a2);
3156        tcg_out8(s, sub);
3157        break;
3158
3159    case INDEX_op_x86_vpblendvb_vec:
3160        insn = OPC_VPBLENDVB;
3161        if (type == TCG_TYPE_V256) {
3162            insn |= P_VEXL;
3163        }
3164        tcg_out_vex_modrm(s, insn, a0, a1, a2);
3165        tcg_out8(s, args[3] << 4);
3166        break;
3167
3168    case INDEX_op_x86_psrldq_vec:
3169        tcg_out_vex_modrm(s, OPC_GRP14, 3, a0, a1);
3170        tcg_out8(s, a2);
3171        break;
3172
3173    case INDEX_op_mov_vec:  /* Always emitted via tcg_out_mov.  */
3174    case INDEX_op_dup_vec:  /* Always emitted via tcg_out_dup_vec.  */
3175    default:
3176        g_assert_not_reached();
3177    }
3178}
3179
3180static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op)
3181{
3182    switch (op) {
3183    case INDEX_op_goto_ptr:
3184        return C_O0_I1(r);
3185
3186    case INDEX_op_ld8u_i32:
3187    case INDEX_op_ld8u_i64:
3188    case INDEX_op_ld8s_i32:
3189    case INDEX_op_ld8s_i64:
3190    case INDEX_op_ld16u_i32:
3191    case INDEX_op_ld16u_i64:
3192    case INDEX_op_ld16s_i32:
3193    case INDEX_op_ld16s_i64:
3194    case INDEX_op_ld_i32:
3195    case INDEX_op_ld32u_i64:
3196    case INDEX_op_ld32s_i64:
3197    case INDEX_op_ld_i64:
3198        return C_O1_I1(r, r);
3199
3200    case INDEX_op_st8_i32:
3201    case INDEX_op_st8_i64:
3202        return C_O0_I2(qi, r);
3203
3204    case INDEX_op_st16_i32:
3205    case INDEX_op_st16_i64:
3206    case INDEX_op_st_i32:
3207    case INDEX_op_st32_i64:
3208        return C_O0_I2(ri, r);
3209
3210    case INDEX_op_st_i64:
3211        return C_O0_I2(re, r);
3212
3213    case INDEX_op_add_i32:
3214    case INDEX_op_add_i64:
3215        return C_O1_I2(r, r, re);
3216
3217    case INDEX_op_sub_i32:
3218    case INDEX_op_sub_i64:
3219    case INDEX_op_mul_i32:
3220    case INDEX_op_mul_i64:
3221    case INDEX_op_or_i32:
3222    case INDEX_op_or_i64:
3223    case INDEX_op_xor_i32:
3224    case INDEX_op_xor_i64:
3225        return C_O1_I2(r, 0, re);
3226
3227    case INDEX_op_and_i32:
3228    case INDEX_op_and_i64:
3229        return C_O1_I2(r, 0, reZ);
3230
3231    case INDEX_op_andc_i32:
3232    case INDEX_op_andc_i64:
3233        return C_O1_I2(r, r, rI);
3234
3235    case INDEX_op_shl_i32:
3236    case INDEX_op_shl_i64:
3237    case INDEX_op_shr_i32:
3238    case INDEX_op_shr_i64:
3239    case INDEX_op_sar_i32:
3240    case INDEX_op_sar_i64:
3241        return have_bmi2 ? C_O1_I2(r, r, ri) : C_O1_I2(r, 0, ci);
3242
3243    case INDEX_op_rotl_i32:
3244    case INDEX_op_rotl_i64:
3245    case INDEX_op_rotr_i32:
3246    case INDEX_op_rotr_i64:
3247        return C_O1_I2(r, 0, ci);
3248
3249    case INDEX_op_brcond_i32:
3250    case INDEX_op_brcond_i64:
3251        return C_O0_I2(r, re);
3252
3253    case INDEX_op_bswap16_i32:
3254    case INDEX_op_bswap16_i64:
3255    case INDEX_op_bswap32_i32:
3256    case INDEX_op_bswap32_i64:
3257    case INDEX_op_bswap64_i64:
3258    case INDEX_op_neg_i32:
3259    case INDEX_op_neg_i64:
3260    case INDEX_op_not_i32:
3261    case INDEX_op_not_i64:
3262    case INDEX_op_extrh_i64_i32:
3263        return C_O1_I1(r, 0);
3264
3265    case INDEX_op_ext8s_i32:
3266    case INDEX_op_ext8s_i64:
3267    case INDEX_op_ext8u_i32:
3268    case INDEX_op_ext8u_i64:
3269        return C_O1_I1(r, q);
3270
3271    case INDEX_op_ext16s_i32:
3272    case INDEX_op_ext16s_i64:
3273    case INDEX_op_ext16u_i32:
3274    case INDEX_op_ext16u_i64:
3275    case INDEX_op_ext32s_i64:
3276    case INDEX_op_ext32u_i64:
3277    case INDEX_op_ext_i32_i64:
3278    case INDEX_op_extu_i32_i64:
3279    case INDEX_op_extrl_i64_i32:
3280    case INDEX_op_extract_i32:
3281    case INDEX_op_extract_i64:
3282    case INDEX_op_sextract_i32:
3283    case INDEX_op_ctpop_i32:
3284    case INDEX_op_ctpop_i64:
3285        return C_O1_I1(r, r);
3286
3287    case INDEX_op_extract2_i32:
3288    case INDEX_op_extract2_i64:
3289        return C_O1_I2(r, 0, r);
3290
3291    case INDEX_op_deposit_i32:
3292    case INDEX_op_deposit_i64:
3293        return C_O1_I2(Q, 0, Q);
3294
3295    case INDEX_op_setcond_i32:
3296    case INDEX_op_setcond_i64:
3297        return C_O1_I2(q, r, re);
3298
3299    case INDEX_op_movcond_i32:
3300    case INDEX_op_movcond_i64:
3301        return C_O1_I4(r, r, re, r, 0);
3302
3303    case INDEX_op_div2_i32:
3304    case INDEX_op_div2_i64:
3305    case INDEX_op_divu2_i32:
3306    case INDEX_op_divu2_i64:
3307        return C_O2_I3(a, d, 0, 1, r);
3308
3309    case INDEX_op_mulu2_i32:
3310    case INDEX_op_mulu2_i64:
3311    case INDEX_op_muls2_i32:
3312    case INDEX_op_muls2_i64:
3313        return C_O2_I2(a, d, a, r);
3314
3315    case INDEX_op_add2_i32:
3316    case INDEX_op_add2_i64:
3317    case INDEX_op_sub2_i32:
3318    case INDEX_op_sub2_i64:
3319        return C_O2_I4(r, r, 0, 1, re, re);
3320
3321    case INDEX_op_ctz_i32:
3322    case INDEX_op_ctz_i64:
3323        return have_bmi1 ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
3324
3325    case INDEX_op_clz_i32:
3326    case INDEX_op_clz_i64:
3327        return have_lzcnt ? C_N1_I2(r, r, rW) : C_N1_I2(r, r, r);
3328
3329    case INDEX_op_qemu_ld_i32:
3330        return (TARGET_LONG_BITS <= TCG_TARGET_REG_BITS
3331                ? C_O1_I1(r, L) : C_O1_I2(r, L, L));
3332
3333    case INDEX_op_qemu_st_i32:
3334        return (TARGET_LONG_BITS <= TCG_TARGET_REG_BITS
3335                ? C_O0_I2(L, L) : C_O0_I3(L, L, L));
3336    case INDEX_op_qemu_st8_i32:
3337        return (TARGET_LONG_BITS <= TCG_TARGET_REG_BITS
3338                ? C_O0_I2(s, L) : C_O0_I3(s, L, L));
3339
3340    case INDEX_op_qemu_ld_i64:
3341        return (TCG_TARGET_REG_BITS == 64 ? C_O1_I1(r, L)
3342                : TARGET_LONG_BITS <= TCG_TARGET_REG_BITS ? C_O2_I1(r, r, L)
3343                : C_O2_I2(r, r, L, L));
3344
3345    case INDEX_op_qemu_st_i64:
3346        return (TCG_TARGET_REG_BITS == 64 ? C_O0_I2(L, L)
3347                : TARGET_LONG_BITS <= TCG_TARGET_REG_BITS ? C_O0_I3(L, L, L)
3348                : C_O0_I4(L, L, L, L));
3349
3350    case INDEX_op_brcond2_i32:
3351        return C_O0_I4(r, r, ri, ri);
3352
3353    case INDEX_op_setcond2_i32:
3354        return C_O1_I4(r, r, r, ri, ri);
3355
3356    case INDEX_op_ld_vec:
3357    case INDEX_op_dupm_vec:
3358        return C_O1_I1(x, r);
3359
3360    case INDEX_op_st_vec:
3361        return C_O0_I2(x, r);
3362
3363    case INDEX_op_add_vec:
3364    case INDEX_op_sub_vec:
3365    case INDEX_op_mul_vec:
3366    case INDEX_op_and_vec:
3367    case INDEX_op_or_vec:
3368    case INDEX_op_xor_vec:
3369    case INDEX_op_andc_vec:
3370    case INDEX_op_orc_vec:
3371    case INDEX_op_nand_vec:
3372    case INDEX_op_nor_vec:
3373    case INDEX_op_eqv_vec:
3374    case INDEX_op_ssadd_vec:
3375    case INDEX_op_usadd_vec:
3376    case INDEX_op_sssub_vec:
3377    case INDEX_op_ussub_vec:
3378    case INDEX_op_smin_vec:
3379    case INDEX_op_umin_vec:
3380    case INDEX_op_smax_vec:
3381    case INDEX_op_umax_vec:
3382    case INDEX_op_shlv_vec:
3383    case INDEX_op_shrv_vec:
3384    case INDEX_op_sarv_vec:
3385    case INDEX_op_rotlv_vec:
3386    case INDEX_op_rotrv_vec:
3387    case INDEX_op_shls_vec:
3388    case INDEX_op_shrs_vec:
3389    case INDEX_op_sars_vec:
3390    case INDEX_op_cmp_vec:
3391    case INDEX_op_x86_shufps_vec:
3392    case INDEX_op_x86_blend_vec:
3393    case INDEX_op_x86_packss_vec:
3394    case INDEX_op_x86_packus_vec:
3395    case INDEX_op_x86_vperm2i128_vec:
3396    case INDEX_op_x86_punpckl_vec:
3397    case INDEX_op_x86_punpckh_vec:
3398    case INDEX_op_x86_vpshldi_vec:
3399#if TCG_TARGET_REG_BITS == 32
3400    case INDEX_op_dup2_vec:
3401#endif
3402        return C_O1_I2(x, x, x);
3403
3404    case INDEX_op_abs_vec:
3405    case INDEX_op_dup_vec:
3406    case INDEX_op_not_vec:
3407    case INDEX_op_shli_vec:
3408    case INDEX_op_shri_vec:
3409    case INDEX_op_sari_vec:
3410    case INDEX_op_rotli_vec:
3411    case INDEX_op_x86_psrldq_vec:
3412        return C_O1_I1(x, x);
3413
3414    case INDEX_op_x86_vpshldv_vec:
3415    case INDEX_op_x86_vpshrdv_vec:
3416        return C_O1_I3(x, 0, x, x);
3417
3418    case INDEX_op_bitsel_vec:
3419    case INDEX_op_x86_vpblendvb_vec:
3420        return C_O1_I3(x, x, x, x);
3421
3422    default:
3423        g_assert_not_reached();
3424    }
3425}
3426
3427int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
3428{
3429    switch (opc) {
3430    case INDEX_op_add_vec:
3431    case INDEX_op_sub_vec:
3432    case INDEX_op_and_vec:
3433    case INDEX_op_or_vec:
3434    case INDEX_op_xor_vec:
3435    case INDEX_op_andc_vec:
3436    case INDEX_op_orc_vec:
3437    case INDEX_op_nand_vec:
3438    case INDEX_op_nor_vec:
3439    case INDEX_op_eqv_vec:
3440    case INDEX_op_not_vec:
3441    case INDEX_op_bitsel_vec:
3442        return 1;
3443    case INDEX_op_cmp_vec:
3444    case INDEX_op_cmpsel_vec:
3445        return -1;
3446
3447    case INDEX_op_rotli_vec:
3448        return have_avx512vl && vece >= MO_32 ? 1 : -1;
3449
3450    case INDEX_op_shli_vec:
3451    case INDEX_op_shri_vec:
3452        /* We must expand the operation for MO_8.  */
3453        return vece == MO_8 ? -1 : 1;
3454
3455    case INDEX_op_sari_vec:
3456        switch (vece) {
3457        case MO_8:
3458            return -1;
3459        case MO_16:
3460        case MO_32:
3461            return 1;
3462        case MO_64:
3463            if (have_avx512vl) {
3464                return 1;
3465            }
3466            /*
3467             * We can emulate this for MO_64, but it does not pay off
3468             * unless we're producing at least 4 values.
3469             */
3470            return type >= TCG_TYPE_V256 ? -1 : 0;
3471        }
3472        return 0;
3473
3474    case INDEX_op_shls_vec:
3475    case INDEX_op_shrs_vec:
3476        return vece >= MO_16;
3477    case INDEX_op_sars_vec:
3478        switch (vece) {
3479        case MO_16:
3480        case MO_32:
3481            return 1;
3482        case MO_64:
3483            return have_avx512vl;
3484        }
3485        return 0;
3486    case INDEX_op_rotls_vec:
3487        return vece >= MO_16 ? -1 : 0;
3488
3489    case INDEX_op_shlv_vec:
3490    case INDEX_op_shrv_vec:
3491        switch (vece) {
3492        case MO_16:
3493            return have_avx512bw;
3494        case MO_32:
3495        case MO_64:
3496            return have_avx2;
3497        }
3498        return 0;
3499    case INDEX_op_sarv_vec:
3500        switch (vece) {
3501        case MO_16:
3502            return have_avx512bw;
3503        case MO_32:
3504            return have_avx2;
3505        case MO_64:
3506            return have_avx512vl;
3507        }
3508        return 0;
3509    case INDEX_op_rotlv_vec:
3510    case INDEX_op_rotrv_vec:
3511        switch (vece) {
3512        case MO_16:
3513            return have_avx512vbmi2 ? -1 : 0;
3514        case MO_32:
3515        case MO_64:
3516            return have_avx512vl ? 1 : have_avx2 ? -1 : 0;
3517        }
3518        return 0;
3519
3520    case INDEX_op_mul_vec:
3521        switch (vece) {
3522        case MO_8:
3523            return -1;
3524        case MO_64:
3525            return have_avx512dq;
3526        }
3527        return 1;
3528
3529    case INDEX_op_ssadd_vec:
3530    case INDEX_op_usadd_vec:
3531    case INDEX_op_sssub_vec:
3532    case INDEX_op_ussub_vec:
3533        return vece <= MO_16;
3534    case INDEX_op_smin_vec:
3535    case INDEX_op_smax_vec:
3536    case INDEX_op_umin_vec:
3537    case INDEX_op_umax_vec:
3538    case INDEX_op_abs_vec:
3539        return vece <= MO_32 || have_avx512vl;
3540
3541    default:
3542        return 0;
3543    }
3544}
3545
3546static void expand_vec_shi(TCGType type, unsigned vece, TCGOpcode opc,
3547                           TCGv_vec v0, TCGv_vec v1, TCGArg imm)
3548{
3549    TCGv_vec t1, t2;
3550
3551    tcg_debug_assert(vece == MO_8);
3552
3553    t1 = tcg_temp_new_vec(type);
3554    t2 = tcg_temp_new_vec(type);
3555
3556    /*
3557     * Unpack to W, shift, and repack.  Tricky bits:
3558     * (1) Use punpck*bw x,x to produce DDCCBBAA,
3559     *     i.e. duplicate in other half of the 16-bit lane.
3560     * (2) For right-shift, add 8 so that the high half of the lane
3561     *     becomes zero.  For left-shift, and left-rotate, we must
3562     *     shift up and down again.
3563     * (3) Step 2 leaves high half zero such that PACKUSWB
3564     *     (pack with unsigned saturation) does not modify
3565     *     the quantity.
3566     */
3567    vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3568              tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3569    vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3570              tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3571
3572    if (opc != INDEX_op_rotli_vec) {
3573        imm += 8;
3574    }
3575    if (opc == INDEX_op_shri_vec) {
3576        tcg_gen_shri_vec(MO_16, t1, t1, imm);
3577        tcg_gen_shri_vec(MO_16, t2, t2, imm);
3578    } else {
3579        tcg_gen_shli_vec(MO_16, t1, t1, imm);
3580        tcg_gen_shli_vec(MO_16, t2, t2, imm);
3581        tcg_gen_shri_vec(MO_16, t1, t1, 8);
3582        tcg_gen_shri_vec(MO_16, t2, t2, 8);
3583    }
3584
3585    vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
3586              tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
3587    tcg_temp_free_vec(t1);
3588    tcg_temp_free_vec(t2);
3589}
3590
3591static void expand_vec_sari(TCGType type, unsigned vece,
3592                            TCGv_vec v0, TCGv_vec v1, TCGArg imm)
3593{
3594    TCGv_vec t1, t2;
3595
3596    switch (vece) {
3597    case MO_8:
3598        /* Unpack to W, shift, and repack, as in expand_vec_shi.  */
3599        t1 = tcg_temp_new_vec(type);
3600        t2 = tcg_temp_new_vec(type);
3601        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3602                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3603        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3604                  tcgv_vec_arg(t2), tcgv_vec_arg(v1), tcgv_vec_arg(v1));
3605        tcg_gen_sari_vec(MO_16, t1, t1, imm + 8);
3606        tcg_gen_sari_vec(MO_16, t2, t2, imm + 8);
3607        vec_gen_3(INDEX_op_x86_packss_vec, type, MO_8,
3608                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t2));
3609        tcg_temp_free_vec(t1);
3610        tcg_temp_free_vec(t2);
3611        break;
3612
3613    case MO_64:
3614        if (imm <= 32) {
3615            /*
3616             * We can emulate a small sign extend by performing an arithmetic
3617             * 32-bit shift and overwriting the high half of a 64-bit logical
3618             * shift.  Note that the ISA says shift of 32 is valid, but TCG
3619             * does not, so we have to bound the smaller shift -- we get the
3620             * same result in the high half either way.
3621             */
3622            t1 = tcg_temp_new_vec(type);
3623            tcg_gen_sari_vec(MO_32, t1, v1, MIN(imm, 31));
3624            tcg_gen_shri_vec(MO_64, v0, v1, imm);
3625            vec_gen_4(INDEX_op_x86_blend_vec, type, MO_32,
3626                      tcgv_vec_arg(v0), tcgv_vec_arg(v0),
3627                      tcgv_vec_arg(t1), 0xaa);
3628            tcg_temp_free_vec(t1);
3629        } else {
3630            /* Otherwise we will need to use a compare vs 0 to produce
3631             * the sign-extend, shift and merge.
3632             */
3633            t1 = tcg_const_zeros_vec(type);
3634            tcg_gen_cmp_vec(TCG_COND_GT, MO_64, t1, t1, v1);
3635            tcg_gen_shri_vec(MO_64, v0, v1, imm);
3636            tcg_gen_shli_vec(MO_64, t1, t1, 64 - imm);
3637            tcg_gen_or_vec(MO_64, v0, v0, t1);
3638            tcg_temp_free_vec(t1);
3639        }
3640        break;
3641
3642    default:
3643        g_assert_not_reached();
3644    }
3645}
3646
3647static void expand_vec_rotli(TCGType type, unsigned vece,
3648                             TCGv_vec v0, TCGv_vec v1, TCGArg imm)
3649{
3650    TCGv_vec t;
3651
3652    if (vece == MO_8) {
3653        expand_vec_shi(type, vece, INDEX_op_rotli_vec, v0, v1, imm);
3654        return;
3655    }
3656
3657    if (have_avx512vbmi2) {
3658        vec_gen_4(INDEX_op_x86_vpshldi_vec, type, vece,
3659                  tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v1), imm);
3660        return;
3661    }
3662
3663    t = tcg_temp_new_vec(type);
3664    tcg_gen_shli_vec(vece, t, v1, imm);
3665    tcg_gen_shri_vec(vece, v0, v1, (8 << vece) - imm);
3666    tcg_gen_or_vec(vece, v0, v0, t);
3667    tcg_temp_free_vec(t);
3668}
3669
3670static void expand_vec_rotv(TCGType type, unsigned vece, TCGv_vec v0,
3671                            TCGv_vec v1, TCGv_vec sh, bool right)
3672{
3673    TCGv_vec t;
3674
3675    if (have_avx512vbmi2) {
3676        vec_gen_4(right ? INDEX_op_x86_vpshrdv_vec : INDEX_op_x86_vpshldv_vec,
3677                  type, vece, tcgv_vec_arg(v0), tcgv_vec_arg(v1),
3678                  tcgv_vec_arg(v1), tcgv_vec_arg(sh));
3679        return;
3680    }
3681
3682    t = tcg_temp_new_vec(type);
3683    tcg_gen_dupi_vec(vece, t, 8 << vece);
3684    tcg_gen_sub_vec(vece, t, t, sh);
3685    if (right) {
3686        tcg_gen_shlv_vec(vece, t, v1, t);
3687        tcg_gen_shrv_vec(vece, v0, v1, sh);
3688    } else {
3689        tcg_gen_shrv_vec(vece, t, v1, t);
3690        tcg_gen_shlv_vec(vece, v0, v1, sh);
3691    }
3692    tcg_gen_or_vec(vece, v0, v0, t);
3693    tcg_temp_free_vec(t);
3694}
3695
3696static void expand_vec_rotls(TCGType type, unsigned vece,
3697                             TCGv_vec v0, TCGv_vec v1, TCGv_i32 lsh)
3698{
3699    TCGv_vec t = tcg_temp_new_vec(type);
3700
3701    tcg_debug_assert(vece != MO_8);
3702
3703    if (vece >= MO_32 ? have_avx512vl : have_avx512vbmi2) {
3704        tcg_gen_dup_i32_vec(vece, t, lsh);
3705        if (vece >= MO_32) {
3706            tcg_gen_rotlv_vec(vece, v0, v1, t);
3707        } else {
3708            expand_vec_rotv(type, vece, v0, v1, t, false);
3709        }
3710    } else {
3711        TCGv_i32 rsh = tcg_temp_new_i32();
3712
3713        tcg_gen_neg_i32(rsh, lsh);
3714        tcg_gen_andi_i32(rsh, rsh, (8 << vece) - 1);
3715        tcg_gen_shls_vec(vece, t, v1, lsh);
3716        tcg_gen_shrs_vec(vece, v0, v1, rsh);
3717        tcg_gen_or_vec(vece, v0, v0, t);
3718
3719        tcg_temp_free_i32(rsh);
3720    }
3721
3722    tcg_temp_free_vec(t);
3723}
3724
3725static void expand_vec_mul(TCGType type, unsigned vece,
3726                           TCGv_vec v0, TCGv_vec v1, TCGv_vec v2)
3727{
3728    TCGv_vec t1, t2, t3, t4, zero;
3729
3730    tcg_debug_assert(vece == MO_8);
3731
3732    /*
3733     * Unpack v1 bytes to words, 0 | x.
3734     * Unpack v2 bytes to words, y | 0.
3735     * This leaves the 8-bit result, x * y, with 8 bits of right padding.
3736     * Shift logical right by 8 bits to clear the high 8 bytes before
3737     * using an unsigned saturated pack.
3738     *
3739     * The difference between the V64, V128 and V256 cases is merely how
3740     * we distribute the expansion between temporaries.
3741     */
3742    switch (type) {
3743    case TCG_TYPE_V64:
3744        t1 = tcg_temp_new_vec(TCG_TYPE_V128);
3745        t2 = tcg_temp_new_vec(TCG_TYPE_V128);
3746        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
3747        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
3748                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
3749        vec_gen_3(INDEX_op_x86_punpckl_vec, TCG_TYPE_V128, MO_8,
3750                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
3751        tcg_gen_mul_vec(MO_16, t1, t1, t2);
3752        tcg_gen_shri_vec(MO_16, t1, t1, 8);
3753        vec_gen_3(INDEX_op_x86_packus_vec, TCG_TYPE_V128, MO_8,
3754                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t1));
3755        tcg_temp_free_vec(t1);
3756        tcg_temp_free_vec(t2);
3757        break;
3758
3759    case TCG_TYPE_V128:
3760    case TCG_TYPE_V256:
3761        t1 = tcg_temp_new_vec(type);
3762        t2 = tcg_temp_new_vec(type);
3763        t3 = tcg_temp_new_vec(type);
3764        t4 = tcg_temp_new_vec(type);
3765        zero = tcg_constant_vec(TCG_TYPE_V128, MO_8, 0);
3766        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3767                  tcgv_vec_arg(t1), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
3768        vec_gen_3(INDEX_op_x86_punpckl_vec, type, MO_8,
3769                  tcgv_vec_arg(t2), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
3770        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3771                  tcgv_vec_arg(t3), tcgv_vec_arg(v1), tcgv_vec_arg(zero));
3772        vec_gen_3(INDEX_op_x86_punpckh_vec, type, MO_8,
3773                  tcgv_vec_arg(t4), tcgv_vec_arg(zero), tcgv_vec_arg(v2));
3774        tcg_gen_mul_vec(MO_16, t1, t1, t2);
3775        tcg_gen_mul_vec(MO_16, t3, t3, t4);
3776        tcg_gen_shri_vec(MO_16, t1, t1, 8);
3777        tcg_gen_shri_vec(MO_16, t3, t3, 8);
3778        vec_gen_3(INDEX_op_x86_packus_vec, type, MO_8,
3779                  tcgv_vec_arg(v0), tcgv_vec_arg(t1), tcgv_vec_arg(t3));
3780        tcg_temp_free_vec(t1);
3781        tcg_temp_free_vec(t2);
3782        tcg_temp_free_vec(t3);
3783        tcg_temp_free_vec(t4);
3784        break;
3785
3786    default:
3787        g_assert_not_reached();
3788    }
3789}
3790
3791static bool expand_vec_cmp_noinv(TCGType type, unsigned vece, TCGv_vec v0,
3792                                 TCGv_vec v1, TCGv_vec v2, TCGCond cond)
3793{
3794    enum {
3795        NEED_INV  = 1,
3796        NEED_SWAP = 2,
3797        NEED_BIAS = 4,
3798        NEED_UMIN = 8,
3799        NEED_UMAX = 16,
3800    };
3801    TCGv_vec t1, t2, t3;
3802    uint8_t fixup;
3803
3804    switch (cond) {
3805    case TCG_COND_EQ:
3806    case TCG_COND_GT:
3807        fixup = 0;
3808        break;
3809    case TCG_COND_NE:
3810    case TCG_COND_LE:
3811        fixup = NEED_INV;
3812        break;
3813    case TCG_COND_LT:
3814        fixup = NEED_SWAP;
3815        break;
3816    case TCG_COND_GE:
3817        fixup = NEED_SWAP | NEED_INV;
3818        break;
3819    case TCG_COND_LEU:
3820        if (tcg_can_emit_vec_op(INDEX_op_umin_vec, type, vece)) {
3821            fixup = NEED_UMIN;
3822        } else {
3823            fixup = NEED_BIAS | NEED_INV;
3824        }
3825        break;
3826    case TCG_COND_GTU:
3827        if (tcg_can_emit_vec_op(INDEX_op_umin_vec, type, vece)) {
3828            fixup = NEED_UMIN | NEED_INV;
3829        } else {
3830            fixup = NEED_BIAS;
3831        }
3832        break;
3833    case TCG_COND_GEU:
3834        if (tcg_can_emit_vec_op(INDEX_op_umax_vec, type, vece)) {
3835            fixup = NEED_UMAX;
3836        } else {
3837            fixup = NEED_BIAS | NEED_SWAP | NEED_INV;
3838        }
3839        break;
3840    case TCG_COND_LTU:
3841        if (tcg_can_emit_vec_op(INDEX_op_umax_vec, type, vece)) {
3842            fixup = NEED_UMAX | NEED_INV;
3843        } else {
3844            fixup = NEED_BIAS | NEED_SWAP;
3845        }
3846        break;
3847    default:
3848        g_assert_not_reached();
3849    }
3850
3851    if (fixup & NEED_INV) {
3852        cond = tcg_invert_cond(cond);
3853    }
3854    if (fixup & NEED_SWAP) {
3855        t1 = v1, v1 = v2, v2 = t1;
3856        cond = tcg_swap_cond(cond);
3857    }
3858
3859    t1 = t2 = NULL;
3860    if (fixup & (NEED_UMIN | NEED_UMAX)) {
3861        t1 = tcg_temp_new_vec(type);
3862        if (fixup & NEED_UMIN) {
3863            tcg_gen_umin_vec(vece, t1, v1, v2);
3864        } else {
3865            tcg_gen_umax_vec(vece, t1, v1, v2);
3866        }
3867        v2 = t1;
3868        cond = TCG_COND_EQ;
3869    } else if (fixup & NEED_BIAS) {
3870        t1 = tcg_temp_new_vec(type);
3871        t2 = tcg_temp_new_vec(type);
3872        t3 = tcg_constant_vec(type, vece, 1ull << ((8 << vece) - 1));
3873        tcg_gen_sub_vec(vece, t1, v1, t3);
3874        tcg_gen_sub_vec(vece, t2, v2, t3);
3875        v1 = t1;
3876        v2 = t2;
3877        cond = tcg_signed_cond(cond);
3878    }
3879
3880    tcg_debug_assert(cond == TCG_COND_EQ || cond == TCG_COND_GT);
3881    /* Expand directly; do not recurse.  */
3882    vec_gen_4(INDEX_op_cmp_vec, type, vece,
3883              tcgv_vec_arg(v0), tcgv_vec_arg(v1), tcgv_vec_arg(v2), cond);
3884
3885    if (t1) {
3886        tcg_temp_free_vec(t1);
3887        if (t2) {
3888            tcg_temp_free_vec(t2);
3889        }
3890    }
3891    return fixup & NEED_INV;
3892}
3893
3894static void expand_vec_cmp(TCGType type, unsigned vece, TCGv_vec v0,
3895                           TCGv_vec v1, TCGv_vec v2, TCGCond cond)
3896{
3897    if (expand_vec_cmp_noinv(type, vece, v0, v1, v2, cond)) {
3898        tcg_gen_not_vec(vece, v0, v0);
3899    }
3900}
3901
3902static void expand_vec_cmpsel(TCGType type, unsigned vece, TCGv_vec v0,
3903                              TCGv_vec c1, TCGv_vec c2,
3904                              TCGv_vec v3, TCGv_vec v4, TCGCond cond)
3905{
3906    TCGv_vec t = tcg_temp_new_vec(type);
3907
3908    if (expand_vec_cmp_noinv(type, vece, t, c1, c2, cond)) {
3909        /* Invert the sense of the compare by swapping arguments.  */
3910        TCGv_vec x;
3911        x = v3, v3 = v4, v4 = x;
3912    }
3913    vec_gen_4(INDEX_op_x86_vpblendvb_vec, type, vece,
3914              tcgv_vec_arg(v0), tcgv_vec_arg(v4),
3915              tcgv_vec_arg(v3), tcgv_vec_arg(t));
3916    tcg_temp_free_vec(t);
3917}
3918
3919void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
3920                       TCGArg a0, ...)
3921{
3922    va_list va;
3923    TCGArg a2;
3924    TCGv_vec v0, v1, v2, v3, v4;
3925
3926    va_start(va, a0);
3927    v0 = temp_tcgv_vec(arg_temp(a0));
3928    v1 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
3929    a2 = va_arg(va, TCGArg);
3930
3931    switch (opc) {
3932    case INDEX_op_shli_vec:
3933    case INDEX_op_shri_vec:
3934        expand_vec_shi(type, vece, opc, v0, v1, a2);
3935        break;
3936
3937    case INDEX_op_sari_vec:
3938        expand_vec_sari(type, vece, v0, v1, a2);
3939        break;
3940
3941    case INDEX_op_rotli_vec:
3942        expand_vec_rotli(type, vece, v0, v1, a2);
3943        break;
3944
3945    case INDEX_op_rotls_vec:
3946        expand_vec_rotls(type, vece, v0, v1, temp_tcgv_i32(arg_temp(a2)));
3947        break;
3948
3949    case INDEX_op_rotlv_vec:
3950        v2 = temp_tcgv_vec(arg_temp(a2));
3951        expand_vec_rotv(type, vece, v0, v1, v2, false);
3952        break;
3953    case INDEX_op_rotrv_vec:
3954        v2 = temp_tcgv_vec(arg_temp(a2));
3955        expand_vec_rotv(type, vece, v0, v1, v2, true);
3956        break;
3957
3958    case INDEX_op_mul_vec:
3959        v2 = temp_tcgv_vec(arg_temp(a2));
3960        expand_vec_mul(type, vece, v0, v1, v2);
3961        break;
3962
3963    case INDEX_op_cmp_vec:
3964        v2 = temp_tcgv_vec(arg_temp(a2));
3965        expand_vec_cmp(type, vece, v0, v1, v2, va_arg(va, TCGArg));
3966        break;
3967
3968    case INDEX_op_cmpsel_vec:
3969        v2 = temp_tcgv_vec(arg_temp(a2));
3970        v3 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
3971        v4 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
3972        expand_vec_cmpsel(type, vece, v0, v1, v2, v3, v4, va_arg(va, TCGArg));
3973        break;
3974
3975    default:
3976        break;
3977    }
3978
3979    va_end(va);
3980}
3981
3982static const int tcg_target_callee_save_regs[] = {
3983#if TCG_TARGET_REG_BITS == 64
3984    TCG_REG_RBP,
3985    TCG_REG_RBX,
3986#if defined(_WIN64)
3987    TCG_REG_RDI,
3988    TCG_REG_RSI,
3989#endif
3990    TCG_REG_R12,
3991    TCG_REG_R13,
3992    TCG_REG_R14, /* Currently used for the global env. */
3993    TCG_REG_R15,
3994#else
3995    TCG_REG_EBP, /* Currently used for the global env. */
3996    TCG_REG_EBX,
3997    TCG_REG_ESI,
3998    TCG_REG_EDI,
3999#endif
4000};
4001
4002/* Compute frame size via macros, to share between tcg_target_qemu_prologue
4003   and tcg_register_jit.  */
4004
4005#define PUSH_SIZE \
4006    ((1 + ARRAY_SIZE(tcg_target_callee_save_regs)) \
4007     * (TCG_TARGET_REG_BITS / 8))
4008
4009#define FRAME_SIZE \
4010    ((PUSH_SIZE \
4011      + TCG_STATIC_CALL_ARGS_SIZE \
4012      + CPU_TEMP_BUF_NLONGS * sizeof(long) \
4013      + TCG_TARGET_STACK_ALIGN - 1) \
4014     & ~(TCG_TARGET_STACK_ALIGN - 1))
4015
4016/* Generate global QEMU prologue and epilogue code */
4017static void tcg_target_qemu_prologue(TCGContext *s)
4018{
4019    int i, stack_addend;
4020
4021    /* TB prologue */
4022
4023    /* Reserve some stack space, also for TCG temps.  */
4024    stack_addend = FRAME_SIZE - PUSH_SIZE;
4025    tcg_set_frame(s, TCG_REG_CALL_STACK, TCG_STATIC_CALL_ARGS_SIZE,
4026                  CPU_TEMP_BUF_NLONGS * sizeof(long));
4027
4028    /* Save all callee saved registers.  */
4029    for (i = 0; i < ARRAY_SIZE(tcg_target_callee_save_regs); i++) {
4030        tcg_out_push(s, tcg_target_callee_save_regs[i]);
4031    }
4032
4033#if TCG_TARGET_REG_BITS == 32
4034    tcg_out_ld(s, TCG_TYPE_PTR, TCG_AREG0, TCG_REG_ESP,
4035               (ARRAY_SIZE(tcg_target_callee_save_regs) + 1) * 4);
4036    tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4037    /* jmp *tb.  */
4038    tcg_out_modrm_offset(s, OPC_GRP5, EXT5_JMPN_Ev, TCG_REG_ESP,
4039                         (ARRAY_SIZE(tcg_target_callee_save_regs) + 2) * 4
4040                         + stack_addend);
4041#else
4042# if !defined(CONFIG_SOFTMMU) && TCG_TARGET_REG_BITS == 64
4043    if (guest_base) {
4044        int seg = setup_guest_base_seg();
4045        if (seg != 0) {
4046            x86_guest_base_seg = seg;
4047        } else if (guest_base == (int32_t)guest_base) {
4048            x86_guest_base_offset = guest_base;
4049        } else {
4050            /* Choose R12 because, as a base, it requires a SIB byte. */
4051            x86_guest_base_index = TCG_REG_R12;
4052            tcg_out_movi(s, TCG_TYPE_PTR, x86_guest_base_index, guest_base);
4053            tcg_regset_set_reg(s->reserved_regs, x86_guest_base_index);
4054        }
4055    }
4056# endif
4057    tcg_out_mov(s, TCG_TYPE_PTR, TCG_AREG0, tcg_target_call_iarg_regs[0]);
4058    tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
4059    /* jmp *tb.  */
4060    tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, tcg_target_call_iarg_regs[1]);
4061#endif
4062
4063    /*
4064     * Return path for goto_ptr. Set return value to 0, a-la exit_tb,
4065     * and fall through to the rest of the epilogue.
4066     */
4067    tcg_code_gen_epilogue = tcg_splitwx_to_rx(s->code_ptr);
4068    tcg_out_movi(s, TCG_TYPE_REG, TCG_REG_EAX, 0);
4069
4070    /* TB epilogue */
4071    tb_ret_addr = tcg_splitwx_to_rx(s->code_ptr);
4072
4073    tcg_out_addi(s, TCG_REG_CALL_STACK, stack_addend);
4074
4075    if (have_avx2) {
4076        tcg_out_vex_opc(s, OPC_VZEROUPPER, 0, 0, 0, 0);
4077    }
4078    for (i = ARRAY_SIZE(tcg_target_callee_save_regs) - 1; i >= 0; i--) {
4079        tcg_out_pop(s, tcg_target_callee_save_regs[i]);
4080    }
4081    tcg_out_opc(s, OPC_RET, 0, 0, 0);
4082}
4083
4084static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
4085{
4086    memset(p, 0x90, count);
4087}
4088
4089static void tcg_target_init(TCGContext *s)
4090{
4091#ifdef CONFIG_CPUID_H
4092    unsigned a, b, c, d, b7 = 0, c7 = 0;
4093    unsigned max = __get_cpuid_max(0, 0);
4094
4095    if (max >= 7) {
4096        /* BMI1 is available on AMD Piledriver and Intel Haswell CPUs.  */
4097        __cpuid_count(7, 0, a, b7, c7, d);
4098        have_bmi1 = (b7 & bit_BMI) != 0;
4099        have_bmi2 = (b7 & bit_BMI2) != 0;
4100    }
4101
4102    if (max >= 1) {
4103        __cpuid(1, a, b, c, d);
4104#ifndef have_cmov
4105        /* For 32-bit, 99% certainty that we're running on hardware that
4106           supports cmov, but we still need to check.  In case cmov is not
4107           available, we'll use a small forward branch.  */
4108        have_cmov = (d & bit_CMOV) != 0;
4109#endif
4110
4111        /* MOVBE is only available on Intel Atom and Haswell CPUs, so we
4112           need to probe for it.  */
4113        have_movbe = (c & bit_MOVBE) != 0;
4114        have_popcnt = (c & bit_POPCNT) != 0;
4115
4116        /* There are a number of things we must check before we can be
4117           sure of not hitting invalid opcode.  */
4118        if (c & bit_OSXSAVE) {
4119            unsigned xcrl, xcrh;
4120            /* The xgetbv instruction is not available to older versions of
4121             * the assembler, so we encode the instruction manually.
4122             */
4123            asm(".byte 0x0f, 0x01, 0xd0" : "=a" (xcrl), "=d" (xcrh) : "c" (0));
4124            if ((xcrl & 6) == 6) {
4125                have_avx1 = (c & bit_AVX) != 0;
4126                have_avx2 = (b7 & bit_AVX2) != 0;
4127
4128                /*
4129                 * There are interesting instructions in AVX512, so long
4130                 * as we have AVX512VL, which indicates support for EVEX
4131                 * on sizes smaller than 512 bits.  We are required to
4132                 * check that OPMASK and all extended ZMM state are enabled
4133                 * even if we're not using them -- the insns will fault.
4134                 */
4135                if ((xcrl & 0xe0) == 0xe0
4136                    && (b7 & bit_AVX512F)
4137                    && (b7 & bit_AVX512VL)) {
4138                    have_avx512vl = true;
4139                    have_avx512bw = (b7 & bit_AVX512BW) != 0;
4140                    have_avx512dq = (b7 & bit_AVX512DQ) != 0;
4141                    have_avx512vbmi2 = (c7 & bit_AVX512VBMI2) != 0;
4142                }
4143            }
4144        }
4145    }
4146
4147    max = __get_cpuid_max(0x8000000, 0);
4148    if (max >= 1) {
4149        __cpuid(0x80000001, a, b, c, d);
4150        /* LZCNT was introduced with AMD Barcelona and Intel Haswell CPUs.  */
4151        have_lzcnt = (c & bit_LZCNT) != 0;
4152    }
4153#endif /* CONFIG_CPUID_H */
4154
4155    tcg_target_available_regs[TCG_TYPE_I32] = ALL_GENERAL_REGS;
4156    if (TCG_TARGET_REG_BITS == 64) {
4157        tcg_target_available_regs[TCG_TYPE_I64] = ALL_GENERAL_REGS;
4158    }
4159    if (have_avx1) {
4160        tcg_target_available_regs[TCG_TYPE_V64] = ALL_VECTOR_REGS;
4161        tcg_target_available_regs[TCG_TYPE_V128] = ALL_VECTOR_REGS;
4162    }
4163    if (have_avx2) {
4164        tcg_target_available_regs[TCG_TYPE_V256] = ALL_VECTOR_REGS;
4165    }
4166
4167    tcg_target_call_clobber_regs = ALL_VECTOR_REGS;
4168    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EAX);
4169    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EDX);
4170    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_ECX);
4171    if (TCG_TARGET_REG_BITS == 64) {
4172#if !defined(_WIN64)
4173        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RDI);
4174        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_RSI);
4175#endif
4176        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R8);
4177        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R9);
4178        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R10);
4179        tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_R11);
4180    }
4181
4182    s->reserved_regs = 0;
4183    tcg_regset_set_reg(s->reserved_regs, TCG_REG_CALL_STACK);
4184}
4185
4186typedef struct {
4187    DebugFrameHeader h;
4188    uint8_t fde_def_cfa[4];
4189    uint8_t fde_reg_ofs[14];
4190} DebugFrame;
4191
4192/* We're expecting a 2 byte uleb128 encoded value.  */
4193QEMU_BUILD_BUG_ON(FRAME_SIZE >= (1 << 14));
4194
4195#if !defined(__ELF__)
4196    /* Host machine without ELF. */
4197#elif TCG_TARGET_REG_BITS == 64
4198#define ELF_HOST_MACHINE EM_X86_64
4199static const DebugFrame debug_frame = {
4200    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4201    .h.cie.id = -1,
4202    .h.cie.version = 1,
4203    .h.cie.code_align = 1,
4204    .h.cie.data_align = 0x78,             /* sleb128 -8 */
4205    .h.cie.return_column = 16,
4206
4207    /* Total FDE size does not include the "len" member.  */
4208    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4209
4210    .fde_def_cfa = {
4211        12, 7,                          /* DW_CFA_def_cfa %rsp, ... */
4212        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4213        (FRAME_SIZE >> 7)
4214    },
4215    .fde_reg_ofs = {
4216        0x90, 1,                        /* DW_CFA_offset, %rip, -8 */
4217        /* The following ordering must match tcg_target_callee_save_regs.  */
4218        0x86, 2,                        /* DW_CFA_offset, %rbp, -16 */
4219        0x83, 3,                        /* DW_CFA_offset, %rbx, -24 */
4220        0x8c, 4,                        /* DW_CFA_offset, %r12, -32 */
4221        0x8d, 5,                        /* DW_CFA_offset, %r13, -40 */
4222        0x8e, 6,                        /* DW_CFA_offset, %r14, -48 */
4223        0x8f, 7,                        /* DW_CFA_offset, %r15, -56 */
4224    }
4225};
4226#else
4227#define ELF_HOST_MACHINE EM_386
4228static const DebugFrame debug_frame = {
4229    .h.cie.len = sizeof(DebugFrameCIE)-4, /* length after .len member */
4230    .h.cie.id = -1,
4231    .h.cie.version = 1,
4232    .h.cie.code_align = 1,
4233    .h.cie.data_align = 0x7c,             /* sleb128 -4 */
4234    .h.cie.return_column = 8,
4235
4236    /* Total FDE size does not include the "len" member.  */
4237    .h.fde.len = sizeof(DebugFrame) - offsetof(DebugFrame, h.fde.cie_offset),
4238
4239    .fde_def_cfa = {
4240        12, 4,                          /* DW_CFA_def_cfa %esp, ... */
4241        (FRAME_SIZE & 0x7f) | 0x80,     /* ... uleb128 FRAME_SIZE */
4242        (FRAME_SIZE >> 7)
4243    },
4244    .fde_reg_ofs = {
4245        0x88, 1,                        /* DW_CFA_offset, %eip, -4 */
4246        /* The following ordering must match tcg_target_callee_save_regs.  */
4247        0x85, 2,                        /* DW_CFA_offset, %ebp, -8 */
4248        0x83, 3,                        /* DW_CFA_offset, %ebx, -12 */
4249        0x86, 4,                        /* DW_CFA_offset, %esi, -16 */
4250        0x87, 5,                        /* DW_CFA_offset, %edi, -20 */
4251    }
4252};
4253#endif
4254
4255#if defined(ELF_HOST_MACHINE)
4256void tcg_register_jit(const void *buf, size_t buf_size)
4257{
4258    tcg_register_jit_int(buf, buf_size, &debug_frame, sizeof(debug_frame));
4259}
4260#endif
4261