1
2 /*---------------------------------------------------------------*/
3 /*--- begin libvex.h ---*/
4 /*---------------------------------------------------------------*/
5
6 /*
7 This file is part of Valgrind, a dynamic binary instrumentation
8 framework.
9
10 Copyright (C) 2004-2010 OpenWorks LLP
11 info@open-works.net
12
13 This program is free software; you can redistribute it and/or
14 modify it under the terms of the GNU General Public License as
15 published by the Free Software Foundation; either version 2 of the
16 License, or (at your option) any later version.
17
18 This program is distributed in the hope that it will be useful, but
19 WITHOUT ANY WARRANTY; without even the implied warranty of
20 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
21 General Public License for more details.
22
23 You should have received a copy of the GNU General Public License
24 along with this program; if not, write to the Free Software
25 Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
26 02110-1301, USA.
27
28 The GNU General Public License is contained in the file COPYING.
29
30 Neither the names of the U.S. Department of Energy nor the
31 University of California nor the names of its contributors may be
32 used to endorse or promote products derived from this software
33 without prior written permission.
34 */
35
36 #ifndef __LIBVEX_H
37 #define __LIBVEX_H
38
39
40 #include "libvex_basictypes.h"
41 #include "libvex_ir.h"
42
43
44 /*---------------------------------------------------------------*/
45 /*--- This file defines the top-level interface to LibVEX. ---*/
46 /*---------------------------------------------------------------*/
47
48 /*-------------------------------------------------------*/
49 /*--- Architectures, variants, and other arch info ---*/
50 /*-------------------------------------------------------*/
51
52 typedef
53 enum {
54 VexArch_INVALID,
55 VexArchX86,
56 VexArchAMD64,
57 VexArchARM,
58 VexArchPPC32,
59 VexArchPPC64
60 }
61 VexArch;
62
63
64 /* For a given architecture, these specify extra capabilities beyond
65 the minimum supported (baseline) capabilities. They may be OR'd
66 together, although some combinations don't make sense. (eg, SSE2
67 but not SSE1). LibVEX_Translate will check for nonsensical
68 combinations. */
69
70 /* x86: baseline capability is Pentium-1 (FPU, MMX, but no SSE), with
71 cmpxchg8b. */
72 #define VEX_HWCAPS_X86_SSE1 (1<<1) /* SSE1 support (Pentium III) */
73 #define VEX_HWCAPS_X86_SSE2 (1<<2) /* SSE2 support (Pentium 4) */
74 #define VEX_HWCAPS_X86_SSE3 (1<<3) /* SSE3 support (>= Prescott) */
75 #define VEX_HWCAPS_X86_LZCNT (1<<4) /* SSE4a LZCNT insn */
76
77 /* amd64: baseline capability is SSE2, with cmpxchg8b but not
78 cmpxchg16b. */
79 #define VEX_HWCAPS_AMD64_SSE3 (1<<5) /* SSE3 support */
80 #define VEX_HWCAPS_AMD64_CX16 (1<<6) /* cmpxchg16b support */
81 #define VEX_HWCAPS_AMD64_LZCNT (1<<7) /* SSE4a LZCNT insn */
82
83 /* ppc32: baseline capability is integer only */
84 #define VEX_HWCAPS_PPC32_F (1<<8) /* basic (non-optional) FP */
85 #define VEX_HWCAPS_PPC32_V (1<<9) /* Altivec (VMX) */
86 #define VEX_HWCAPS_PPC32_FX (1<<10) /* FP extns (fsqrt, fsqrts) */
87 #define VEX_HWCAPS_PPC32_GX (1<<11) /* Graphics extns
88 (fres,frsqrte,fsel,stfiwx) */
89
90 /* ppc64: baseline capability is integer and basic FP insns */
91 #define VEX_HWCAPS_PPC64_V (1<<12) /* Altivec (VMX) */
92 #define VEX_HWCAPS_PPC64_FX (1<<13) /* FP extns (fsqrt, fsqrts) */
93 #define VEX_HWCAPS_PPC64_GX (1<<14) /* Graphics extns
94 (fres,frsqrte,fsel,stfiwx) */
95
96 /* arm: baseline capability is ARMv4 */
97 /* Bits 5:0 - architecture level (e.g. 5 for v5, 6 for v6 etc) */
98 #define VEX_HWCAPS_ARM_VFP (1<<6) /* VFP extension */
99 #define VEX_HWCAPS_ARM_VFP2 (1<<7) /* VFPv2 */
100 #define VEX_HWCAPS_ARM_VFP3 (1<<8) /* VFPv3 */
101 /* Bits 15:10 reserved for (possible) future VFP revisions */
102 #define VEX_HWCAPS_ARM_NEON (1<<16) /* Advanced SIMD also known as NEON */
103
104 /* Get an ARM architecure level from HWCAPS */
105 #define VEX_ARM_ARCHLEVEL(x) ((x) & 0x3f)
106
107 /* These return statically allocated strings. */
108
109 extern const HChar* LibVEX_ppVexArch ( VexArch );
110 extern const HChar* LibVEX_ppVexHwCaps ( VexArch, UInt );
111
112
113 /* This struct is a bit of a hack, but is needed to carry misc
114 important bits of info about an arch. Fields which are meaningless
115 or ignored for the platform in question should be set to zero. */
116
117 typedef
118 struct {
119 /* This is the only mandatory field. */
120 UInt hwcaps;
121 /* PPC32/PPC64 only: size of cache line */
122 Int ppc_cache_line_szB;
123 /* PPC32/PPC64 only: sizes zeroed by the dcbz/dcbzl instructions
124 * (bug#135264) */
125 UInt ppc_dcbz_szB;
126 UInt ppc_dcbzl_szB; /* 0 means unsupported (SIGILL) */
127 }
128 VexArchInfo;
129
130 /* Write default settings info *vai. */
131 extern
132 void LibVEX_default_VexArchInfo ( /*OUT*/VexArchInfo* vai );
133
134
135 /* This struct carries guest and host ABI variant information that may
136 be needed. Fields which are meaningless or ignored for the
137 platform in question should be set to zero.
138
139 Settings which are believed to be correct are:
140
141 guest_stack_redzone_size
142 guest is ppc32-linux ==> 0
143 guest is ppc64-linux ==> 288
144 guest is ppc32-aix5 ==> 220
145 guest is ppc64-aix5 ==> unknown
146 guest is amd64-linux ==> 128
147 guest is other ==> inapplicable
148
149 guest_amd64_assume_fs_is_zero
150 guest is amd64-linux ==> True
151 guest is amd64-darwin ==> False
152 guest is other ==> inapplicable
153
154 guest_amd64_assume_gs_is_0x60
155 guest is amd64-darwin ==> True
156 guest is amd64-linux ==> False
157 guest is other ==> inapplicable
158
159 guest_ppc_zap_RZ_at_blr
160 guest is ppc64-linux ==> True
161 guest is ppc32-linux ==> False
162 guest is ppc64-aix5 ==> unknown
163 guest is ppc32-aix5 ==> False
164 guest is other ==> inapplicable
165
166 guest_ppc_zap_RZ_at_bl
167 guest is ppc64-linux ==> const True
168 guest is ppc32-linux ==> const False
169 guest is ppc64-aix5 ==> unknown
170 guest is ppc32-aix5 ==> True except for calls to
171 millicode, $SAVEFn, $RESTFn
172 guest is other ==> inapplicable
173
174 guest_ppc_sc_continues_at_LR:
175 guest is ppc32-aix5 or ppc64-aix5 ==> True
176 guest is ppc32-linux or ppc64-linux ==> False
177 guest is other ==> inapplicable
178
179 host_ppc_calls_use_fndescrs:
180 host is ppc32-linux ==> False
181 host is ppc64-linux ==> True
182 host is ppc32-aix5 or ppc64-aix5 ==> True
183 host is other ==> inapplicable
184
185 host_ppc32_regalign_int64_args:
186 host is ppc32-linux ==> True
187 host is ppc32-aix5 ==> False
188 host is other ==> inapplicable
189 */
190
191 typedef
192 struct {
193 /* PPC and AMD64 GUESTS only: how many bytes below the
194 stack pointer are validly addressible? */
195 Int guest_stack_redzone_size;
196
197 /* AMD64 GUESTS only: should we translate %fs-prefixed
198 instructions using the assumption that %fs always contains
199 zero? */
200 Bool guest_amd64_assume_fs_is_zero;
201
202 /* AMD64 GUESTS only: should we translate %gs-prefixed
203 instructions using the assumption that %gs always contains
204 0x60? */
205 Bool guest_amd64_assume_gs_is_0x60;
206
207 /* PPC GUESTS only: should we zap the stack red zone at a 'blr'
208 (function return) ? */
209 Bool guest_ppc_zap_RZ_at_blr;
210
211 /* PPC GUESTS only: should we zap the stack red zone at a 'bl'
212 (function call) ? Is supplied with the guest address of the
213 target of the call since that may be significant. If NULL,
214 is assumed equivalent to a fn which always returns False. */
215 Bool (*guest_ppc_zap_RZ_at_bl)(Addr64);
216
217 /* PPC32/PPC64 GUESTS only: where does the kernel resume after
218 'sc'? False => Linux style, at the next insn. True => AIX
219 style, at the address stated in the link register. */
220 Bool guest_ppc_sc_continues_at_LR;
221
222 /* PPC32/PPC64 HOSTS only: does '&f' give us a pointer to a
223 function descriptor on the host, or to the function code
224 itself? True => descriptor, False => code. */
225 Bool host_ppc_calls_use_fndescrs;
226
227 /* PPC32 HOSTS only: when generating code to pass a 64-bit value
228 (actual parameter) in a pair of regs, should we skip an arg
229 reg if it is even-numbered? True => yes, False => no. */
230 Bool host_ppc32_regalign_int64_args;
231 }
232 VexAbiInfo;
233
234 /* Write default settings info *vbi. */
235 extern
236 void LibVEX_default_VexAbiInfo ( /*OUT*/VexAbiInfo* vbi );
237
238
239 /*-------------------------------------------------------*/
240 /*--- Control of Vex's optimiser (iropt). ---*/
241 /*-------------------------------------------------------*/
242
243 /* Control of Vex's optimiser. */
244
245 typedef
246 struct {
247 /* Controls verbosity of iropt. 0 = no output. */
248 Int iropt_verbosity;
249 /* Control aggressiveness of iropt. 0 = no opt, 1 = simple
250 opts, 2 (default) = max optimisation. */
251 Int iropt_level;
252 /* Ensure all integer registers are up to date at potential
253 memory exception points? True(default)=yes, False=no, only
254 the guest's stack pointer. */
255 Bool iropt_precise_memory_exns;
256 /* How aggressive should iropt be in unrolling loops? Higher
257 numbers make it more enthusiastic about loop unrolling.
258 Default=120. A setting of zero disables unrolling. */
259 Int iropt_unroll_thresh;
260 /* What's the maximum basic block length the front end(s) allow?
261 BBs longer than this are split up. Default=50 (guest
262 insns). */
263 Int guest_max_insns;
264 /* How aggressive should front ends be in following
265 unconditional branches to known destinations? Default=10,
266 meaning that if a block contains less than 10 guest insns so
267 far, the front end(s) will attempt to chase into its
268 successor. A setting of zero disables chasing. */
269 Int guest_chase_thresh;
270 /* EXPERIMENTAL: chase across conditional branches? Not all
271 front ends honour this. Default: NO. */
272 Bool guest_chase_cond;
273 }
274 VexControl;
275
276
277 /* Write the default settings into *vcon. */
278
279 extern
280 void LibVEX_default_VexControl ( /*OUT*/ VexControl* vcon );
281
282
283 /*-------------------------------------------------------*/
284 /*--- Storage management control ---*/
285 /*-------------------------------------------------------*/
286
287 /* Allocate in Vex's temporary allocation area. Be careful with this.
288 You can only call it inside an instrumentation or optimisation
289 callback that you have previously specified in a call to
290 LibVEX_Translate. The storage allocated will only stay alive until
291 translation of the current basic block is complete.
292 */
293 extern HChar* private_LibVEX_alloc_first;
294 extern HChar* private_LibVEX_alloc_curr;
295 extern HChar* private_LibVEX_alloc_last;
296 extern void private_LibVEX_alloc_OOM(void) __attribute__((noreturn));
297
LibVEX_Alloc(Int nbytes)298 static inline void* LibVEX_Alloc ( Int nbytes )
299 {
300 #if 0
301 /* Nasty debugging hack, do not use. */
302 return malloc(nbytes);
303 #else
304 HChar* curr;
305 HChar* next;
306 Int ALIGN;
307 ALIGN = sizeof(void*)-1;
308 nbytes = (nbytes + ALIGN) & ~ALIGN;
309 curr = private_LibVEX_alloc_curr;
310 next = curr + nbytes;
311 if (next >= private_LibVEX_alloc_last)
312 private_LibVEX_alloc_OOM();
313 private_LibVEX_alloc_curr = next;
314 return curr;
315 #endif
316 }
317
318 /* Show Vex allocation statistics. */
319 extern void LibVEX_ShowAllocStats ( void );
320
321
322 /*-------------------------------------------------------*/
323 /*--- Describing guest state layout ---*/
324 /*-------------------------------------------------------*/
325
326 /* Describe the guest state enough that the instrumentation
327 functions can work. */
328
329 /* The max number of guest state chunks which we can describe as
330 always defined (for the benefit of Memcheck). */
331 #define VEXGLO_N_ALWAYSDEFD 24
332
333 typedef
334 struct {
335 /* Total size of the guest state, in bytes. Must be
336 8-aligned. */
337 Int total_sizeB;
338 /* Whereabouts is the stack pointer? */
339 Int offset_SP;
340 Int sizeof_SP; /* 4 or 8 */
341 /* Whereabouts is the frame pointer? */
342 Int offset_FP;
343 Int sizeof_FP; /* 4 or 8 */
344 /* Whereabouts is the instruction pointer? */
345 Int offset_IP;
346 Int sizeof_IP; /* 4 or 8 */
347 /* Describe parts of the guest state regarded as 'always
348 defined'. */
349 Int n_alwaysDefd;
350 struct {
351 Int offset;
352 Int size;
353 } alwaysDefd[VEXGLO_N_ALWAYSDEFD];
354 }
355 VexGuestLayout;
356
357 /* A note about guest state layout.
358
359 LibVEX defines the layout for the guest state, in the file
360 pub/libvex_guest_<arch>.h. The struct will have an 16-aligned
361 size. Each translated bb is assumed to be entered with a specified
362 register pointing at such a struct. Beyond that is two copies of
363 the shadow state area with the same size as the struct. Beyond
364 that is a spill area that LibVEX may spill into. It must have size
365 LibVEX_N_SPILL_BYTES, and this must be a 16-aligned number.
366
367 On entry, the baseblock pointer register must be 16-aligned.
368
369 There must be no holes in between the primary guest state, its two
370 copies, and the spill area. In short, all 4 areas must have a
371 16-aligned size and be 16-aligned, and placed back-to-back.
372 */
373
374 #define LibVEX_N_SPILL_BYTES 4096
375
376
377 /*-------------------------------------------------------*/
378 /*--- Initialisation of the library ---*/
379 /*-------------------------------------------------------*/
380
381 /* Initialise the library. You must call this first. */
382
383 extern void LibVEX_Init (
384 /* failure exit function */
385 __attribute__ ((noreturn))
386 void (*failure_exit) ( void ),
387 /* logging output function */
388 void (*log_bytes) ( HChar*, Int nbytes ),
389 /* debug paranoia level */
390 Int debuglevel,
391 /* Are we supporting valgrind checking? */
392 Bool valgrind_support,
393 /* Control ... */
394 /*READONLY*/VexControl* vcon
395 );
396
397
398 /*-------------------------------------------------------*/
399 /*--- Make a translation ---*/
400 /*-------------------------------------------------------*/
401
402 /* Describes the outcome of a translation attempt. */
403 typedef
404 enum {
405 VexTransOK,
406 VexTransAccessFail,
407 VexTransOutputFull
408 }
409 VexTranslateResult;
410
411
412 /* Describes precisely the pieces of guest code that a translation
413 covers. Now that Vex can chase across BB boundaries, the old
414 scheme of describing a chunk of guest code merely by its start
415 address and length is inadequate.
416
417 Hopefully this struct is only 32 bytes long. Space is important as
418 clients will have to store one of these for each translation made.
419 */
420 typedef
421 struct {
422 Addr64 base[3];
423 UShort len[3];
424 UShort n_used;
425 }
426 VexGuestExtents;
427
428
429 /* A structure to carry arguments for LibVEX_Translate. There are so
430 many of them, it seems better to have a structure. */
431 typedef
432 struct {
433 /* IN: The instruction sets we are translating from and to. And
434 guest/host misc info. */
435 VexArch arch_guest;
436 VexArchInfo archinfo_guest;
437 VexArch arch_host;
438 VexArchInfo archinfo_host;
439 VexAbiInfo abiinfo_both;
440
441 /* IN: an opaque value which is passed as the first arg to all
442 callback functions supplied in this struct. Vex has no idea
443 what's at the other end of this pointer. */
444 void* callback_opaque;
445
446 /* IN: the block to translate, and its guest address. */
447 /* where are the actual bytes in the host's address space? */
448 UChar* guest_bytes;
449 /* where do the bytes really come from in the guest's aspace?
450 This is the post-redirection guest address. Not that Vex
451 understands anything about redirection; that is all done on
452 the Valgrind side. */
453 Addr64 guest_bytes_addr;
454
455 /* Is it OK to chase into this guest address? May not be
456 NULL. */
457 Bool (*chase_into_ok) ( /*callback_opaque*/void*, Addr64 );
458
459 /* OUT: which bits of guest code actually got translated */
460 VexGuestExtents* guest_extents;
461
462 /* IN: a place to put the resulting code, and its size */
463 UChar* host_bytes;
464 Int host_bytes_size;
465 /* OUT: how much of the output area is used. */
466 Int* host_bytes_used;
467
468 /* IN: optionally, two instrumentation functions. May be
469 NULL. */
470 IRSB* (*instrument1) ( /*callback_opaque*/void*,
471 IRSB*,
472 VexGuestLayout*,
473 VexGuestExtents*,
474 IRType gWordTy, IRType hWordTy );
475 IRSB* (*instrument2) ( /*callback_opaque*/void*,
476 IRSB*,
477 VexGuestLayout*,
478 VexGuestExtents*,
479 IRType gWordTy, IRType hWordTy );
480
481 IRSB* (*finaltidy) ( IRSB* );
482
483 /* IN: should this translation be self-checking? default: False */
484 Bool do_self_check;
485
486 /* IN: optionally, a callback which allows the caller to add its
487 own IR preamble following the self-check and any other
488 VEX-generated preamble, if any. May be NULL. If non-NULL,
489 the IRSB under construction is handed to this function, which
490 presumably adds IR statements to it. The callback may
491 optionally complete the block and direct bb_to_IR not to
492 disassemble any instructions into it; this is indicated by
493 the callback returning True.
494 */
495 Bool (*preamble_function)(/*callback_opaque*/void*, IRSB*);
496
497 /* IN: debug: trace vex activity at various points */
498 Int traceflags;
499
500 /* IN: address of the dispatcher entry point. Describes the
501 place where generated code should jump to at the end of each
502 bb.
503
504 At the end of each translation, the next guest address is
505 placed in the host's standard return register (x86: %eax,
506 amd64: %rax, ppc32: %r3, ppc64: %r3). Optionally, the guest
507 state pointer register (on host x86: %ebp; amd64: %rbp;
508 ppc32/64: r31) may be set to a VEX_TRC_ value to indicate any
509 special action required before the next block is run.
510
511 Control is then passed back to the dispatcher (beyond Vex's
512 control; caller supplies this) in the following way:
513
514 - On host archs which lack a link register (x86, amd64), by a
515 jump to the host address specified in 'dispatcher', which
516 must be non-NULL.
517
518 - On host archs which have a link register (ppc32, ppc64), by
519 a branch to the link register (which is guaranteed to be
520 unchanged from whatever it was at entry to the
521 translation). 'dispatch' must be NULL.
522
523 The aim is to get back and forth between translations and the
524 dispatcher without creating memory traffic to store return
525 addresses.
526 */
527 void* dispatch;
528 }
529 VexTranslateArgs;
530
531
532 extern
533 VexTranslateResult LibVEX_Translate ( VexTranslateArgs* );
534
535 /* A subtlety re interaction between self-checking translations and
536 bb-chasing. The supplied chase_into_ok function should say NO
537 (False) when presented with any address for which you might want to
538 make a self-checking translation.
539
540 If it doesn't do that, you may end up with Vex chasing from BB #1
541 to BB #2 (fine); but if you wanted checking for #2 and not #1, that
542 would not be the result. Therefore chase_into_ok should disallow
543 following into #2. That will force the caller to eventually
544 request a new translation starting at #2, at which point Vex will
545 correctly observe the make-a-self-check flag. */
546
547
548 /*-------------------------------------------------------*/
549 /*--- Show accumulated statistics ---*/
550 /*-------------------------------------------------------*/
551
552 extern void LibVEX_ShowStats ( void );
553
554
555 /*-------------------------------------------------------*/
556 /*--- Notes ---*/
557 /*-------------------------------------------------------*/
558
559 /* Code generation conventions that need to be recorded somewhere.
560 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
561
562 x86
563 ~~~
564 Generated code should be entered using a JMP instruction. On
565 entry, %ebp should point to the guest state, and %esp should be a
566 valid stack pointer. The generated code may change %eax, %ebx,
567 %ecx, %edx, %esi, %edi, all the FP registers and control state, and
568 all the XMM registers.
569
570 On entry, the FPU control word should be set to 0x027F, and the SSE
571 control word (%mxcsr) should be set to 0x1F80. On exit, they
572 should still have those values (after masking off the lowest 6 bits
573 of %mxcsr). If they don't, there is a bug in VEX-generated code.
574
575 Generated code returns to the scheduler using a JMP instruction, to
576 the address specified in the .dispatch field of VexTranslateArgs.
577 %eax (or %eax:%edx, if simulating a 64-bit target) will contain the
578 guest address of the next block to execute. %ebp may be changed
579 to a VEX_TRC_ value, otherwise it should be as it was at entry.
580
581 CRITICAL ISSUES in x86 code generation. The only known critical
582 issue is that the host FPU and SSE state is not properly saved
583 across calls to helper functions. If any helper references any
584 such state, it is likely (1) to misbehave itself, since the FP
585 stack tags will not be as expected, and (2) after returning to
586 generated code, the generated code is likely to go wrong. This
587 really should be fixed.
588
589 amd64
590 ~~~~~
591 Analogous to x86.
592
593 ppc32
594 ~~~~~
595 On entry, guest state pointer is r31. .dispatch must be NULL.
596 Control is returned with a branch to the link register. Generated
597 code will not change lr. At return, r3 holds the next guest addr
598 (or r3:r4 ?). r31 may be may be changed to a VEX_TRC_ value,
599 otherwise it should be as it was at entry.
600
601 ppc64
602 ~~~~~
603 Same as ppc32.
604
605 ALL GUEST ARCHITECTURES
606 ~~~~~~~~~~~~~~~~~~~~~~~
607 The guest state must contain two pseudo-registers, guest_TISTART
608 and guest_TILEN. These are used to pass the address of areas of
609 guest code, translations of which are to be invalidated, back to
610 the despatcher. Both pseudo-regs must have size equal to the guest
611 word size.
612
613 The architecture must a third pseudo-register, guest_NRADDR, also
614 guest-word-sized. This is used to record the unredirected guest
615 address at the start of a translation whose start has been
616 redirected. By reading this pseudo-register shortly afterwards,
617 the translation can find out what the corresponding no-redirection
618 address was. Note, this is only set for wrap-style redirects, not
619 for replace-style ones.
620 */
621 #endif /* ndef __LIBVEX_H */
622
623 /*---------------------------------------------------------------*/
624 /*--- libvex.h ---*/
625 /*---------------------------------------------------------------*/
626