IEMAllN8veExecMem.cpp@ 105413

最後變更在這個檔案從105413是 105413,由 vboxsync 提交於 8 月前
VMM/IEM: When pruning exec memory set the hint to the start of the pruned region for faster subsequent allocation, bugref:10653
屬性 svn:eol-style 設為 `native` 屬性 svn:keywords 設為 `Author Date Id Revision`
檔案大小: 67.1 KB

行
1	/* $Id: IEMAllN8veExecMem.cpp 105413 2024-07-19 07:18:53Z vboxsync $ */
2	/** @file
3	* IEM - Native Recompiler
4	*
5	* Logging group IEM_RE_NATIVE assignments:
6	* - Level 1 (Log) : ...
7	* - Flow (LogFlow) : ...
8	* - Level 2 (Log2) : Details calls as they're recompiled.
9	* - Level 3 (Log3) : Disassemble native code after recompiling.
10	* - Level 4 (Log4) : ...
11	* - Level 5 (Log5) : ...
12	* - Level 6 (Log6) : ...
13	* - Level 7 (Log7) : ...
14	* - Level 8 (Log8) : ...
15	* - Level 9 (Log9) : ...
16	* - Level 10 (Log10): ...
17	* - Level 11 (Log11): Variable allocator.
18	* - Level 12 (Log12): Register allocator.
19	*/
20
21	/*
22	* Copyright (C) 2023 Oracle and/or its affiliates.
23	*
24	* This file is part of VirtualBox base platform packages, as
25	* available from https://www.alldomusa.eu.org.
26	*
27	* This program is free software; you can redistribute it and/or
28	* modify it under the terms of the GNU General Public License
29	* as published by the Free Software Foundation, in version 3 of the
30	* License.
31	*
32	* This program is distributed in the hope that it will be useful, but
33	* WITHOUT ANY WARRANTY; without even the implied warranty of
34	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
35	* General Public License for more details.
36	*
37	* You should have received a copy of the GNU General Public License
38	* along with this program; if not, see <https://www.gnu.org/licenses>.
39	*
40	* SPDX-License-Identifier: GPL-3.0-only
41	*/
42
43
44	/*********************************************************************************************************************************
45	* Header Files *
46	*********************************************************************************************************************************/
47	#define LOG_GROUP LOG_GROUP_IEM_RE_NATIVE
48	#define IEM_WITH_OPAQUE_DECODER_STATE
49	#define VMM_INCLUDED_SRC_include_IEMMc_h /* block IEMMc.h inclusion. */
50	#include <VBox/vmm/iem.h>
51	#include <VBox/vmm/cpum.h>
52	#include "IEMInternal.h"
53	#include <VBox/vmm/vmcc.h>
54	#include <VBox/log.h>
55	#include <VBox/err.h>
56	#include <VBox/param.h>
57	#include <iprt/assert.h>
58	#include <iprt/mem.h>
59	#include <iprt/string.h>
60	#if defined(RT_ARCH_AMD64)
61	# include <iprt/x86.h>
62	#elif defined(RT_ARCH_ARM64)
63	# include <iprt/armv8.h>
64	#endif
65
66	#ifdef RT_OS_WINDOWS
67	# include <iprt/formats/pecoff.h> /* this is incomaptible with windows.h, thus: */
68	extern "C" DECLIMPORT(uint8_t) __cdecl RtlAddFunctionTable(void *pvFunctionTable, uint32_t cEntries, uintptr_t uBaseAddress);
69	extern "C" DECLIMPORT(uint8_t) __cdecl RtlDelFunctionTable(void *pvFunctionTable);
70	#else
71	# include <iprt/formats/dwarf.h>
72	# if defined(RT_OS_DARWIN)
73	# include <libkern/OSCacheControl.h>
74	# include <mach/mach.h>
75	# include <mach/mach_vm.h>
76	# define IEMNATIVE_USE_LIBUNWIND
77	extern "C" void __register_frame(const void *pvFde);
78	extern "C" void __deregister_frame(const void *pvFde);
79	# else
80	# ifdef DEBUG_bird /** @todo not thread safe yet */
81	# define IEMNATIVE_USE_GDB_JIT
82	# endif
83	# ifdef IEMNATIVE_USE_GDB_JIT
84	# include <iprt/critsect.h>
85	# include <iprt/once.h>
86	# include <iprt/formats/elf64.h>
87	# endif
88	extern "C" void __register_frame_info(void pvBegin, void pvObj); /* found no header for these two */
89	extern "C" void __deregister_frame_info(void pvBegin); /* (returns pvObj from __register_frame_info call) */
90	# endif
91	#endif
92
93	#include "IEMN8veRecompiler.h"
94
95
96	/*********************************************************************************************************************************
97	* Executable Memory Allocator *
98	*********************************************************************************************************************************/
99	/** The chunk sub-allocation unit size in bytes. */
100	#define IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE 256
101	/** The chunk sub-allocation unit size as a shift factor. */
102	#define IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT 8
103	/** Enables adding a header to the sub-allocator allocations.
104	* This is useful for freeing up executable memory among other things. */
105	#define IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
106	/** Use alternative pruning. */
107	#define IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING
108
109
110	#if defined(IN_RING3) && !defined(RT_OS_WINDOWS)
111	# ifdef IEMNATIVE_USE_GDB_JIT
112	# define IEMNATIVE_USE_GDB_JIT_ET_DYN
113
114	/** GDB JIT: Code entry. */
115	typedef struct GDBJITCODEENTRY
116	{
117	struct GDBJITCODEENTRY *pNext;
118	struct GDBJITCODEENTRY *pPrev;
119	uint8_t *pbSymFile;
120	uint64_t cbSymFile;
121	} GDBJITCODEENTRY;
122
123	/** GDB JIT: Actions. */
124	typedef enum GDBJITACTIONS : uint32_t
125	{
126	kGdbJitaction_NoAction = 0, kGdbJitaction_Register, kGdbJitaction_Unregister
127	} GDBJITACTIONS;
128
129	/** GDB JIT: Descriptor. */
130	typedef struct GDBJITDESCRIPTOR
131	{
132	uint32_t uVersion;
133	GDBJITACTIONS enmAction;
134	GDBJITCODEENTRY *pRelevant;
135	GDBJITCODEENTRY *pHead;
136	/** Our addition: */
137	GDBJITCODEENTRY *pTail;
138	} GDBJITDESCRIPTOR;
139
140	/** GDB JIT: Our simple symbol file data. */
141	typedef struct GDBJITSYMFILE
142	{
143	Elf64_Ehdr EHdr;
144	# ifndef IEMNATIVE_USE_GDB_JIT_ET_DYN
145	Elf64_Shdr aShdrs[5];
146	# else
147	Elf64_Shdr aShdrs[7];
148	Elf64_Phdr aPhdrs[2];
149	# endif
150	/** The dwarf ehframe data for the chunk. */
151	uint8_t abEhFrame[512];
152	char szzStrTab[128];
153	Elf64_Sym aSymbols[3];
154	# ifdef IEMNATIVE_USE_GDB_JIT_ET_DYN
155	Elf64_Sym aDynSyms[2];
156	Elf64_Dyn aDyn[6];
157	# endif
158	} GDBJITSYMFILE;
159
160	extern "C" GDBJITDESCRIPTOR __jit_debug_descriptor;
161	extern "C" DECLEXPORT(void) __jit_debug_register_code(void);
162
163	/** Init once for g_IemNativeGdbJitLock. */
164	static RTONCE g_IemNativeGdbJitOnce = RTONCE_INITIALIZER;
165	/** Init once for the critical section. */
166	static RTCRITSECT g_IemNativeGdbJitLock;
167
168	/** GDB reads the info here. */
169	GDBJITDESCRIPTOR __jit_debug_descriptor = { 1, kGdbJitaction_NoAction, NULL, NULL };
170
171	/** GDB sets a breakpoint on this and checks __jit_debug_descriptor when hit. */
172	DECL_NO_INLINE(RT_NOTHING, DECLEXPORT(void)) __jit_debug_register_code(void)
173	{
174	ASMNopPause();
175	}
176
177	/** @callback_method_impl{FNRTONCE} */
178	static DECLCALLBACK(int32_t) iemNativeGdbJitInitOnce(void *pvUser)
179	{
180	RT_NOREF(pvUser);
181	return RTCritSectInit(&g_IemNativeGdbJitLock);
182	}
183
184
185	# endif /* IEMNATIVE_USE_GDB_JIT */
186
187	/**
188	* Per-chunk unwind info for non-windows hosts.
189	*/
190	typedef struct IEMEXECMEMCHUNKEHFRAME
191	{
192	# ifdef IEMNATIVE_USE_LIBUNWIND
193	/** The offset of the FDA into abEhFrame. */
194	uintptr_t offFda;
195	# else
196	/** 'struct object' storage area. */
197	uint8_t abObject[1024];
198	# endif
199	# ifdef IEMNATIVE_USE_GDB_JIT
200	# if 0
201	/** The GDB JIT 'symbol file' data. */
202	GDBJITSYMFILE GdbJitSymFile;
203	# endif
204	/** The GDB JIT list entry. */
205	GDBJITCODEENTRY GdbJitEntry;
206	# endif
207	/** The dwarf ehframe data for the chunk. */
208	uint8_t abEhFrame[512];
209	} IEMEXECMEMCHUNKEHFRAME;
210	/** Pointer to per-chunk info info for non-windows hosts. */
211	typedef IEMEXECMEMCHUNKEHFRAME *PIEMEXECMEMCHUNKEHFRAME;
212	#endif
213
214
215	/**
216	* An chunk of executable memory.
217	*/
218	typedef struct IEMEXECMEMCHUNK
219	{
220	/** Number of free items in this chunk. */
221	uint32_t cFreeUnits;
222	/** Hint were to start searching for free space in the allocation bitmap. */
223	uint32_t idxFreeHint;
224	/** Pointer to the readable/writeable view of the memory chunk. */
225	void *pvChunkRw;
226	/** Pointer to the readable/executable view of the memory chunk. */
227	void *pvChunkRx;
228	#ifdef IEMNATIVE_WITH_RECOMPILER_PER_CHUNK_TAIL_CODE
229	/** Pointer to the context structure detailing the per chunk common code. */
230	PCIEMNATIVEPERCHUNKCTX pCtx;
231	#endif
232	#ifdef IN_RING3
233	/**
234	* Pointer to the unwind information.
235	*
236	* This is used during C++ throw and longjmp (windows and probably most other
237	* platforms). Some debuggers (windbg) makes use of it as well.
238	*
239	* Windows: This is allocated from hHeap on windows because (at least for
240	* AMD64) the UNWIND_INFO structure address in the
241	* RUNTIME_FUNCTION entry is an RVA and the chunk is the "image".
242	*
243	* Others: Allocated from the regular heap to avoid unnecessary executable data
244	* structures. This points to an IEMEXECMEMCHUNKEHFRAME structure. */
245	void *pvUnwindInfo;
246	#elif defined(IN_RING0)
247	/** Allocation handle. */
248	RTR0MEMOBJ hMemObj;
249	#endif
250	} IEMEXECMEMCHUNK;
251	/** Pointer to a memory chunk. */
252	typedef IEMEXECMEMCHUNK *PIEMEXECMEMCHUNK;
253
254
255	/**
256	* Executable memory allocator for the native recompiler.
257	*/
258	typedef struct IEMEXECMEMALLOCATOR
259	{
260	/** Magic value (IEMEXECMEMALLOCATOR_MAGIC). */
261	uint32_t uMagic;
262
263	/** The chunk size. */
264	uint32_t cbChunk;
265	/** The maximum number of chunks. */
266	uint32_t cMaxChunks;
267	/** The current number of chunks. */
268	uint32_t cChunks;
269	/** Hint where to start looking for available memory. */
270	uint32_t idxChunkHint;
271	/** Statistics: Current number of allocations. */
272	uint32_t cAllocations;
273
274	/** The total amount of memory available. */
275	uint64_t cbTotal;
276	/** Total amount of free memory. */
277	uint64_t cbFree;
278	/** Total amount of memory allocated. */
279	uint64_t cbAllocated;
280
281	/** Pointer to the allocation bitmaps for all the chunks (follows aChunks).
282	*
283	* Since the chunk size is a power of two and the minimum chunk size is a lot
284	* higher than the IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE, each chunk will always
285	* require a whole number of uint64_t elements in the allocation bitmap. So,
286	* for sake of simplicity, they are allocated as one continous chunk for
287	* simplicity/laziness. */
288	uint64_t *pbmAlloc;
289	/** Number of units (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE) per chunk. */
290	uint32_t cUnitsPerChunk;
291	/** Number of bitmap elements per chunk (for quickly locating the bitmap
292	* portion corresponding to an chunk). */
293	uint32_t cBitmapElementsPerChunk;
294
295	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING
296	/** The next chunk to prune in. */
297	uint32_t idxChunkPrune;
298	/** Where in chunk offset to start pruning at. */
299	uint32_t offChunkPrune;
300	/** Profiling the pruning code. */
301	STAMPROFILE StatPruneProf;
302	/** Number of bytes recovered by the pruning. */
303	STAMPROFILE StatPruneRecovered;
304	#endif
305
306	#ifdef VBOX_WITH_STATISTICS
307	STAMPROFILE StatAlloc;
308	#endif
309
310
311	#if defined(IN_RING3) && !defined(RT_OS_WINDOWS)
312	/** Pointer to the array of unwind info running parallel to aChunks (same
313	* allocation as this structure, located after the bitmaps).
314	* (For Windows, the structures must reside in 32-bit RVA distance to the
315	* actual chunk, so they are allocated off the chunk.) */
316	PIEMEXECMEMCHUNKEHFRAME paEhFrames;
317	#endif
318
319	/** The allocation chunks. */
320	RT_FLEXIBLE_ARRAY_EXTENSION
321	IEMEXECMEMCHUNK aChunks[RT_FLEXIBLE_ARRAY];
322	} IEMEXECMEMALLOCATOR;
323	/** Pointer to an executable memory allocator. */
324	typedef IEMEXECMEMALLOCATOR *PIEMEXECMEMALLOCATOR;
325
326	/** Magic value for IEMEXECMEMALLOCATOR::uMagic (Scott Frederick Turow). */
327	#define IEMEXECMEMALLOCATOR_MAGIC UINT32_C(0x19490412)
328
329
330	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
331	/**
332	* Allocation header.
333	*/
334	typedef struct IEMEXECMEMALLOCHDR
335	{
336	/** Magic value / eyecatcher (IEMEXECMEMALLOCHDR_MAGIC). */
337	uint32_t uMagic;
338	/** The allocation chunk (for speeding up freeing). */
339	uint32_t idxChunk;
340	/** Pointer to the translation block the allocation belongs to.
341	* This is the whole point of the header. */
342	PIEMTB pTb;
343	} IEMEXECMEMALLOCHDR;
344	/** Pointer to an allocation header. */
345	typedef IEMEXECMEMALLOCHDR *PIEMEXECMEMALLOCHDR;
346	/** Magic value for IEMEXECMEMALLOCHDR ('ExeM'). */
347	# define IEMEXECMEMALLOCHDR_MAGIC UINT32_C(0x4d657845)
348	#endif
349
350
351	static int iemExecMemAllocatorGrow(PVMCPUCC pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator);
352
353
354	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING
355	/**
356	* Frees up executable memory when we're out space.
357	*
358	* This is an alternative to iemTbAllocatorFreeupNativeSpace() that frees up
359	* space in a more linear fashion from the allocator's point of view. It may
360	* also defragment if implemented & enabled
361	*/
362	static void iemExecMemAllocatorPrune(PVMCPU pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator)
363	{
364	# ifndef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
365	# error "IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING requires IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER"
366	# endif
367	STAM_REL_PROFILE_START(&pExecMemAllocator->StatPruneProf, a);
368
369	/*
370	* Before we can start, we must process delayed frees.
371	*/
372	iemTbAllocatorProcessDelayedFrees(pVCpu, pVCpu->iem.s.pTbAllocatorR3);
373
374	AssertCompile(RT_IS_POWER_OF_TWO(IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE));
375
376	uint32_t const cbChunk = pExecMemAllocator->cbChunk;
377	AssertReturnVoid(RT_IS_POWER_OF_TWO(cbChunk));
378	AssertReturnVoid(cbChunk >= _1M && cbChunk <= _256M); /* see iemExecMemAllocatorInit */
379
380	uint32_t const cChunks = pExecMemAllocator->cChunks;
381	AssertReturnVoid(cChunks == pExecMemAllocator->cMaxChunks);
382	AssertReturnVoid(cChunks >= 1);
383
384	Assert(!pVCpu->iem.s.pCurTbR3);
385
386	/*
387	* Decide how much to prune. The chunk is is a multiple of two, so we'll be
388	* scanning a multiple of two here as well.
389	*/
390	uint32_t cbToPrune = cbChunk;
391
392	/* Never more than 25%. */
393	if (cChunks < 4)
394	cbToPrune /= cChunks == 1 ? 4 : 2;
395
396	/* Upper limit. In a debug build a 4MB limit averages out at ~0.6ms per call. */
397	if (cbToPrune > _4M)
398	cbToPrune = _4M;
399
400	/*
401	* Adjust the pruning chunk and offset accordingly.
402	*/
403	uint32_t idxChunk = pExecMemAllocator->idxChunkPrune;
404	uint32_t offChunk = pExecMemAllocator->offChunkPrune;
405	offChunk &= ~(uint32_t)(IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1U);
406	if (offChunk >= cbChunk)
407	{
408	offChunk = 0;
409	idxChunk += 1;
410	}
411	if (idxChunk >= cChunks)
412	{
413	offChunk = 0;
414	idxChunk = 0;
415	}
416
417	uint32_t const offPruneEnd = RT_MIN(offChunk + cbToPrune, cbChunk);
418
419	/*
420	* Do the pruning. The current approach is the sever kind.
421	*/
422	uint64_t cbPruned = 0;
423	uint8_t * const pbChunk = (uint8_t *)pExecMemAllocator->aChunks[idxChunk].pvChunkRx;
424	while (offChunk < offPruneEnd)
425	{
426	PIEMEXECMEMALLOCHDR pHdr = (PIEMEXECMEMALLOCHDR)&pbChunk[offChunk];
427
428	/* Is this the start of an allocation block for TB? (We typically have
429	one allocation at the start of each chunk for the unwind info where
430	pTb is NULL.) */
431	if ( pHdr->uMagic == IEMEXECMEMALLOCHDR_MAGIC
432	&& pHdr->pTb != NULL
433	&& pHdr->idxChunk == idxChunk)
434	{
435	PIEMTB const pTb = pHdr->pTb;
436	AssertPtr(pTb);
437
438	uint32_t const cbBlock = RT_ALIGN_32(pTb->Native.cInstructions * sizeof(IEMNATIVEINSTR) + sizeof(*pHdr),
439	IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE);
440	AssertBreakStmt(offChunk + cbBlock <= cbChunk, offChunk += IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE); /* paranoia */
441
442	iemTbAllocatorFree(pVCpu, pTb);
443
444	cbPruned += cbBlock;
445	offChunk += cbBlock;
446	}
447	else
448	offChunk += IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE;
449	}
450	STAM_REL_PROFILE_ADD_PERIOD(&pExecMemAllocator->StatPruneRecovered, cbPruned);
451
452	/*
453	* Save the current pruning point.
454	*/
455	pExecMemAllocator->offChunkPrune = offChunk;
456	pExecMemAllocator->idxChunkPrune = idxChunk;
457
458	/* Set the hint to the start of the pruned region. */
459	pExecMemAllocator->idxChunkHint = idxChunk;
460	pExecMemAllocator->aChunks[idxChunk].idxFreeHint = offChunk / IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE;
461
462	STAM_REL_PROFILE_STOP(&pExecMemAllocator->StatPruneProf, a);
463	}
464	#endif /* IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING */
465
466
467	/**
468	* Try allocate a block of @a cReqUnits in the chunk @a idxChunk.
469	*/
470	static void *
471	iemExecMemAllocatorAllocInChunkInt(PIEMEXECMEMALLOCATOR pExecMemAllocator, uint64_t *pbmAlloc, uint32_t idxFirst,
472	uint32_t cToScan, uint32_t cReqUnits, uint32_t idxChunk, PIEMTB pTb,
473	void *ppvExec, PCIEMNATIVEPERCHUNKCTX ppChunkCtx)
474	{
475	/*
476	* Shift the bitmap to the idxFirst bit so we can use ASMBitFirstClear.
477	*/
478	Assert(!(cToScan & 63));
479	Assert(!(idxFirst & 63));
480	Assert(cToScan + idxFirst <= pExecMemAllocator->cUnitsPerChunk);
481	pbmAlloc += idxFirst / 64;
482
483	/*
484	* Scan the bitmap for cReqUnits of consequtive clear bits
485	*/
486	/** @todo This can probably be done more efficiently for non-x86 systems. */
487	int iBit = ASMBitFirstClear(pbmAlloc, cToScan);
488	while (iBit >= 0 && (uint32_t)iBit <= cToScan - cReqUnits)
489	{
490	uint32_t idxAddBit = 1;
491	while (idxAddBit < cReqUnits && !ASMBitTest(pbmAlloc, (uint32_t)iBit + idxAddBit))
492	idxAddBit++;
493	if (idxAddBit >= cReqUnits)
494	{
495	ASMBitSetRange(pbmAlloc, (uint32_t)iBit, (uint32_t)iBit + cReqUnits);
496
497	PIEMEXECMEMCHUNK const pChunk = &pExecMemAllocator->aChunks[idxChunk];
498	pChunk->cFreeUnits -= cReqUnits;
499	pChunk->idxFreeHint = (uint32_t)iBit + cReqUnits;
500
501	pExecMemAllocator->cAllocations += 1;
502	uint32_t const cbReq = cReqUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
503	pExecMemAllocator->cbAllocated += cbReq;
504	pExecMemAllocator->cbFree -= cbReq;
505	pExecMemAllocator->idxChunkHint = idxChunk;
506
507	void * const pvMemRw = (uint8_t *)pChunk->pvChunkRw
508	+ ((idxFirst + (uint32_t)iBit) << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT);
509
510	if (ppChunkCtx)
511	*ppChunkCtx = pChunk->pCtx;
512
513	/*
514	* Initialize the header and return.
515	*/
516	# ifdef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
517	PIEMEXECMEMALLOCHDR const pHdr = (PIEMEXECMEMALLOCHDR)pvMemRw;
518	pHdr->uMagic = IEMEXECMEMALLOCHDR_MAGIC;
519	pHdr->idxChunk = idxChunk;
520	pHdr->pTb = pTb;
521
522	if (ppvExec)
523	ppvExec = (uint8_t )pChunk->pvChunkRx
524	+ ((idxFirst + (uint32_t)iBit) << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT)
525	+ sizeof(*pHdr);
526
527	return pHdr + 1;
528	#else
529	if (ppvExec)
530	ppvExec = (uint8_t )pChunk->pvChunkRx
531	+ ((idxFirst + (uint32_t)iBit) << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT);
532
533	RT_NOREF(pTb);
534	return pvMem;
535	#endif
536	}
537
538	iBit = ASMBitNextClear(pbmAlloc, cToScan, iBit + idxAddBit - 1);
539	}
540	return NULL;
541	}
542
543
544	static PIEMNATIVEINSTR
545	iemExecMemAllocatorAllocInChunk(PIEMEXECMEMALLOCATOR pExecMemAllocator, uint32_t idxChunk, uint32_t cbReq, PIEMTB pTb,
546	PIEMNATIVEINSTR ppaExec, PCIEMNATIVEPERCHUNKCTX ppChunkCtx)
547	{
548	/*
549	* Figure out how much to allocate.
550	*/
551	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
552	uint32_t const cReqUnits = (cbReq + sizeof(IEMEXECMEMALLOCHDR) + IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1)
553	#else
554	uint32_t const cReqUnits = (cbReq + IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1)
555	#endif
556	>> IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
557	if (cReqUnits <= pExecMemAllocator->aChunks[idxChunk].cFreeUnits)
558	{
559	uint64_t * const pbmAlloc = &pExecMemAllocator->pbmAlloc[pExecMemAllocator->cBitmapElementsPerChunk * idxChunk];
560	uint32_t const idxHint = pExecMemAllocator->aChunks[idxChunk].idxFreeHint & ~(uint32_t)63;
561	if (idxHint + cReqUnits <= pExecMemAllocator->cUnitsPerChunk)
562	{
563	void *pvRet = iemExecMemAllocatorAllocInChunkInt(pExecMemAllocator, pbmAlloc, idxHint,
564	pExecMemAllocator->cUnitsPerChunk - idxHint,
565	cReqUnits, idxChunk, pTb, (void **)ppaExec, ppChunkCtx);
566	if (pvRet)
567	return (PIEMNATIVEINSTR)pvRet;
568	}
569	return (PIEMNATIVEINSTR)iemExecMemAllocatorAllocInChunkInt(pExecMemAllocator, pbmAlloc, 0,
570	RT_MIN(pExecMemAllocator->cUnitsPerChunk,
571	RT_ALIGN_32(idxHint + cReqUnits, 64)),
572	cReqUnits, idxChunk, pTb, (void **)ppaExec, ppChunkCtx);
573	}
574	return NULL;
575	}
576
577
578	/**
579	* Allocates @a cbReq bytes of executable memory.
580	*
581	* @returns Pointer to the readable/writeable memory, NULL if out of memory or other problem
582	* encountered.
583	* @param pVCpu The cross context virtual CPU structure of the
584	* calling thread.
585	* @param cbReq How many bytes are required.
586	* @param pTb The translation block that will be using the allocation.
587	* @param ppaExec Where to return the pointer to executable view of
588	* the allocated memory, optional.
589	* @param ppChunkCtx Where to return the per chunk attached context
590	* if available, optional.
591	*/
592	DECLHIDDEN(PIEMNATIVEINSTR) iemExecMemAllocatorAlloc(PVMCPU pVCpu, uint32_t cbReq, PIEMTB pTb,
593	PIEMNATIVEINSTR ppaExec, PCIEMNATIVEPERCHUNKCTX ppChunkCtx) RT_NOEXCEPT
594	{
595	PIEMEXECMEMALLOCATOR pExecMemAllocator = pVCpu->iem.s.pExecMemAllocatorR3;
596	AssertReturn(pExecMemAllocator && pExecMemAllocator->uMagic == IEMEXECMEMALLOCATOR_MAGIC, NULL);
597	AssertMsgReturn(cbReq > 32 && cbReq < _512K, ("%#x\n", cbReq), NULL);
598	STAM_PROFILE_START(&pExecMemAllocator->StatAlloc, a);
599
600	for (unsigned iIteration = 0;; iIteration++)
601	{
602	if (cbReq <= pExecMemAllocator->cbFree)
603	{
604	uint32_t const cChunks = pExecMemAllocator->cChunks;
605	uint32_t const idxChunkHint = pExecMemAllocator->idxChunkHint < cChunks ? pExecMemAllocator->idxChunkHint : 0;
606	for (uint32_t idxChunk = idxChunkHint; idxChunk < cChunks; idxChunk++)
607	{
608	PIEMNATIVEINSTR const pRet = iemExecMemAllocatorAllocInChunk(pExecMemAllocator, idxChunk, cbReq, pTb,
609	ppaExec, ppChunkCtx);
610	if (pRet)
611	{
612	STAM_PROFILE_STOP(&pExecMemAllocator->StatAlloc, a);
613	return pRet;
614	}
615	}
616	for (uint32_t idxChunk = 0; idxChunk < idxChunkHint; idxChunk++)
617	{
618	PIEMNATIVEINSTR const pRet = iemExecMemAllocatorAllocInChunk(pExecMemAllocator, idxChunk, cbReq, pTb,
619	ppaExec, ppChunkCtx);
620	if (pRet)
621	{
622	STAM_PROFILE_STOP(&pExecMemAllocator->StatAlloc, a);
623	return pRet;
624	}
625	}
626	}
627
628	/*
629	* Can we grow it with another chunk?
630	*/
631	if (pExecMemAllocator->cChunks < pExecMemAllocator->cMaxChunks)
632	{
633	int rc = iemExecMemAllocatorGrow(pVCpu, pExecMemAllocator);
634	AssertLogRelRCReturn(rc, NULL);
635
636	uint32_t const idxChunk = pExecMemAllocator->cChunks - 1;
637	PIEMNATIVEINSTR const pRet = iemExecMemAllocatorAllocInChunk(pExecMemAllocator, idxChunk, cbReq, pTb,
638	ppaExec, ppChunkCtx);
639	if (pRet)
640	{
641	STAM_PROFILE_STOP(&pExecMemAllocator->StatAlloc, a);
642	return pRet;
643	}
644	AssertFailed();
645	}
646
647	/*
648	* Try prune native TBs once.
649	*/
650	if (iIteration == 0)
651	{
652	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING
653	iemExecMemAllocatorPrune(pVCpu, pExecMemAllocator);
654	#else
655	/* No header included in the instruction count here. */
656	uint32_t const cNeededInstrs = RT_ALIGN_32(cbReq, IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE) / sizeof(IEMNATIVEINSTR);
657	iemTbAllocatorFreeupNativeSpace(pVCpu, cNeededInstrs);
658	#endif
659	}
660	else
661	{
662	STAM_REL_COUNTER_INC(&pVCpu->iem.s.StatNativeExecMemInstrBufAllocFailed);
663	STAM_PROFILE_STOP(&pExecMemAllocator->StatAlloc, a);
664	return NULL;
665	}
666	}
667	}
668
669
670	/** This is a hook to ensure the instruction cache is properly flushed before the code in the memory
671	* given by @a pv and @a cb is executed */
672	DECLHIDDEN(void) iemExecMemAllocatorReadyForUse(PVMCPUCC pVCpu, void *pv, size_t cb) RT_NOEXCEPT
673	{
674	#ifdef RT_OS_DARWIN
675	/*
676	* Flush the instruction cache:
677	* https://developer.apple.com/documentation/apple-silicon/porting-just-in-time-compilers-to-apple-silicon
678	*/
679	/* sys_dcache_flush(pv, cb); - not necessary */
680	sys_icache_invalidate(pv, cb);
681	RT_NOREF(pVCpu);
682
683	#elif defined(RT_OS_LINUX) && defined(RT_ARCH_ARM64)
684	RT_NOREF(pVCpu);
685
686	/* There is __builtin___clear_cache() but it flushes both the instruction and data cache, so do it manually. */
687	static uint32_t s_u32CtrEl0 = 0;
688	if (!s_u32CtrEl0)
689	asm volatile ("mrs %0, ctr_el0":"=r" (s_u32CtrEl0));
690	uintptr_t cbICacheLine = (uintptr_t)4 << (s_u32CtrEl0 & 0xf);
691
692	uintptr_t pb = (uintptr_t)pv & ~(cbICacheLine - 1);
693	for (; pb < (uintptr_t)pv + cb; pb += cbICacheLine)
694	asm volatile ("ic ivau, %0" : : "r" (pb) : "memory");
695
696	asm volatile ("dsb ish\n\t isb\n\t" : : : "memory");
697
698	#else
699	RT_NOREF(pVCpu, pv, cb);
700	#endif
701	}
702
703
704	/**
705	* Frees executable memory.
706	*/
707	DECLHIDDEN(void) iemExecMemAllocatorFree(PVMCPU pVCpu, void *pv, size_t cb) RT_NOEXCEPT
708	{
709	PIEMEXECMEMALLOCATOR pExecMemAllocator = pVCpu->iem.s.pExecMemAllocatorR3;
710	Assert(pExecMemAllocator && pExecMemAllocator->uMagic == IEMEXECMEMALLOCATOR_MAGIC);
711	AssertPtr(pv);
712	#ifndef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
713	Assert(!((uintptr_t)pv & (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1)));
714
715	/* Align the size as we did when allocating the block. */
716	cb = RT_ALIGN_Z(cb, IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE);
717
718	#else
719	PIEMEXECMEMALLOCHDR pHdr = (PIEMEXECMEMALLOCHDR)pv - 1;
720	Assert(!((uintptr_t)pHdr & (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1)));
721	AssertReturnVoid(pHdr->uMagic == IEMEXECMEMALLOCHDR_MAGIC);
722	uint32_t const idxChunk = pHdr->idxChunk;
723	AssertReturnVoid(idxChunk < pExecMemAllocator->cChunks);
724	pv = pHdr;
725
726	/* Adjust and align the size to cover the whole allocation area. */
727	cb = RT_ALIGN_Z(cb + sizeof(*pHdr), IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE);
728	#endif
729
730	/* Free it / assert sanity. */
731	bool fFound = false;
732	uint32_t const cbChunk = pExecMemAllocator->cbChunk;
733	#ifndef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
734	uint32_t const cChunks = pExecMemAllocator->cChunks;
735	for (uint32_t idxChunk = 0; idxChunk < cChunks; idxChunk++)
736	#endif
737	{
738	uintptr_t const offChunk = (uintptr_t)pv - (uintptr_t)pExecMemAllocator->aChunks[idxChunk].pvChunkRx;
739	fFound = offChunk < cbChunk;
740	if (fFound)
741	{
742	uint32_t const idxFirst = (uint32_t)offChunk >> IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
743	uint32_t const cReqUnits = (uint32_t)cb >> IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
744
745	/* Check that it's valid and free it. */
746	uint64_t * const pbmAlloc = &pExecMemAllocator->pbmAlloc[pExecMemAllocator->cBitmapElementsPerChunk * idxChunk];
747	AssertReturnVoid(ASMBitTest(pbmAlloc, idxFirst));
748	for (uint32_t i = 1; i < cReqUnits; i++)
749	AssertReturnVoid(ASMBitTest(pbmAlloc, idxFirst + i));
750	ASMBitClearRange(pbmAlloc, idxFirst, idxFirst + cReqUnits);
751
752	/* Invalidate the header using the writeable memory view. */
753	pHdr = (PIEMEXECMEMALLOCHDR)((uintptr_t)pExecMemAllocator->aChunks[idxChunk].pvChunkRw + offChunk);
754	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
755	pHdr->uMagic = 0;
756	pHdr->idxChunk = 0;
757	pHdr->pTb = NULL;
758	#endif
759	pExecMemAllocator->aChunks[idxChunk].cFreeUnits += cReqUnits;
760	pExecMemAllocator->aChunks[idxChunk].idxFreeHint = idxFirst;
761
762	/* Update the stats. */
763	pExecMemAllocator->cbAllocated -= cb;
764	pExecMemAllocator->cbFree += cb;
765	pExecMemAllocator->cAllocations -= 1;
766	return;
767	}
768	}
769	AssertFailed();
770	}
771
772
773	/**
774	* Interface used by iemNativeRecompileAttachExecMemChunkCtx and unwind info
775	* generators.
776	*/
777	DECLHIDDEN(PIEMNATIVEINSTR)
778	iemExecMemAllocatorAllocFromChunk(PVMCPU pVCpu, uint32_t idxChunk, uint32_t cbReq, PIEMNATIVEINSTR *ppaExec)
779	{
780	PIEMEXECMEMALLOCATOR pExecMemAllocator = pVCpu->iem.s.pExecMemAllocatorR3;
781	AssertReturn(idxChunk < pExecMemAllocator->cChunks, NULL);
782	Assert(cbReq < _1M);
783	return iemExecMemAllocatorAllocInChunk(pExecMemAllocator, idxChunk, cbReq, NULL /pTb/, ppaExec, NULL /ppChunkCtx/);
784	}
785
786
787	#ifdef IEMNATIVE_WITH_RECOMPILER_PER_CHUNK_TAIL_CODE
788	/**
789	* For getting the per-chunk context detailing common code for a TB.
790	*
791	* This is for use by the disassembler.
792	*/
793	DECLHIDDEN(PCIEMNATIVEPERCHUNKCTX) iemExecMemGetTbChunkCtx(PVMCPU pVCpu, PCIEMTB pTb)
794	{
795	PIEMEXECMEMALLOCATOR pExecMemAllocator = pVCpu->iem.s.pExecMemAllocatorR3;
796	if ((pTb->fFlags & IEMTB_F_TYPE_MASK) == IEMTB_F_TYPE_NATIVE)
797	{
798	uintptr_t const uAddress = (uintptr_t)pTb->Native.paInstructions;
799	uint32_t const cbChunk = pExecMemAllocator->cbChunk;
800	uint32_t idxChunk = pExecMemAllocator->cChunks;
801	while (idxChunk-- > 0)
802	if (uAddress - (uintptr_t)pExecMemAllocator->aChunks[idxChunk].pvChunkRx < cbChunk)
803	return pExecMemAllocator->aChunks[idxChunk].pCtx;
804	}
805	return NULL;
806	}
807	#endif
808
809
810	#ifdef IN_RING3
811	# ifdef RT_OS_WINDOWS
812
813	/**
814	* Initializes the unwind info structures for windows hosts.
815	*/
816	static int
817	iemExecMemAllocatorInitAndRegisterUnwindInfoForChunk(PVMCPUCC pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator,
818	void *pvChunk, uint32_t idxChunk)
819	{
820	RT_NOREF(pVCpu);
821
822	/*
823	* The AMD64 unwind opcodes.
824	*
825	* This is a program that starts with RSP after a RET instruction that
826	* ends up in recompiled code, and the operations we describe here will
827	* restore all non-volatile registers and bring RSP back to where our
828	* RET address is. This means it's reverse order from what happens in
829	* the prologue.
830	*
831	* Note! Using a frame register approach here both because we have one
832	* and but mainly because the UWOP_ALLOC_LARGE argument values
833	* would be a pain to write initializers for. On the positive
834	* side, we're impervious to changes in the the stack variable
835	* area can can deal with dynamic stack allocations if necessary.
836	*/
837	static const IMAGE_UNWIND_CODE s_aOpcodes[] =
838	{
839	{ { 16, IMAGE_AMD64_UWOP_SET_FPREG, 0 } }, /* RSP = RBP - FrameOffset * 10 (0x60) */
840	{ { 16, IMAGE_AMD64_UWOP_ALLOC_SMALL, 0 } }, /* RSP += 8; */
841	{ { 14, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_x15 } }, /* R15 = [RSP]; RSP += 8; */
842	{ { 12, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_x14 } }, /* R14 = [RSP]; RSP += 8; */
843	{ { 10, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_x13 } }, /* R13 = [RSP]; RSP += 8; */
844	{ { 8, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_x12 } }, /* R12 = [RSP]; RSP += 8; */
845	{ { 7, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_xDI } }, /* RDI = [RSP]; RSP += 8; */
846	{ { 6, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_xSI } }, /* RSI = [RSP]; RSP += 8; */
847	{ { 5, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_xBX } }, /* RBX = [RSP]; RSP += 8; */
848	{ { 4, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_xBP } }, /* RBP = [RSP]; RSP += 8; */
849	};
850	union
851	{
852	IMAGE_UNWIND_INFO Info;
853	uint8_t abPadding[RT_UOFFSETOF(IMAGE_UNWIND_INFO, aOpcodes) + 16];
854	} s_UnwindInfo =
855	{
856	{
857	/* .Version = */ 1,
858	/* .Flags = */ 0,
859	/* .SizeOfProlog = / 16, / whatever */
860	/* .CountOfCodes = */ RT_ELEMENTS(s_aOpcodes),
861	/* .FrameRegister = */ X86_GREG_xBP,
862	/* .FrameOffset = / (-IEMNATIVE_FP_OFF_LAST_PUSH + 8) / 16 / we're off by one slot. sigh. */,
863	}
864	};
865	AssertCompile(-IEMNATIVE_FP_OFF_LAST_PUSH < 240 && -IEMNATIVE_FP_OFF_LAST_PUSH > 0);
866	AssertCompile((-IEMNATIVE_FP_OFF_LAST_PUSH & 0xf) == 8);
867
868	/*
869	* Calc how much space we need and allocate it off the exec heap.
870	*/
871	unsigned const cFunctionEntries = 1;
872	unsigned const cbUnwindInfo = sizeof(s_aOpcodes) + RT_UOFFSETOF(IMAGE_UNWIND_INFO, aOpcodes);
873	unsigned const cbNeeded = sizeof(IMAGE_RUNTIME_FUNCTION_ENTRY) * cFunctionEntries + cbUnwindInfo;
874	PIMAGE_RUNTIME_FUNCTION_ENTRY const paFunctions
875	= (PIMAGE_RUNTIME_FUNCTION_ENTRY)iemExecMemAllocatorAllocInChunk(pExecMemAllocator, idxChunk, cbNeeded, NULL, NULL, NULL);
876	AssertReturn(paFunctions, VERR_INTERNAL_ERROR_5);
877	pExecMemAllocator->aChunks[idxChunk].pvUnwindInfo = paFunctions;
878
879	/*
880	* Initialize the structures.
881	*/
882	PIMAGE_UNWIND_INFO const pInfo = (PIMAGE_UNWIND_INFO)&paFunctions[cFunctionEntries];
883
884	paFunctions[0].BeginAddress = 0;
885	paFunctions[0].EndAddress = pExecMemAllocator->cbChunk;
886	paFunctions[0].UnwindInfoAddress = (uint32_t)((uintptr_t)pInfo - (uintptr_t)pvChunk);
887
888	memcpy(pInfo, &s_UnwindInfo, RT_UOFFSETOF(IMAGE_UNWIND_INFO, aOpcodes));
889	memcpy(&pInfo->aOpcodes[0], s_aOpcodes, sizeof(s_aOpcodes));
890
891	/*
892	* Register it.
893	*/
894	uint8_t fRet = RtlAddFunctionTable(paFunctions, cFunctionEntries, (uintptr_t)pvChunk);
895	AssertReturn(fRet, VERR_INTERNAL_ERROR_3); /* Nothing to clean up on failure, since its within the chunk itself. */
896
897	return VINF_SUCCESS;
898	}
899
900
901	# else /* !RT_OS_WINDOWS */
902
903	/**
904	* Emits a LEB128 encoded value between -0x2000 and 0x2000 (both exclusive).
905	*/
906	DECLINLINE(RTPTRUNION) iemDwarfPutLeb128(RTPTRUNION Ptr, int32_t iValue)
907	{
908	if (iValue >= 64)
909	{
910	Assert(iValue < 0x2000);
911	*Ptr.pb++ = ((uint8_t)iValue & 0x7f) \| 0x80;
912	*Ptr.pb++ = (uint8_t)(iValue >> 7) & 0x3f;
913	}
914	else if (iValue >= 0)
915	*Ptr.pb++ = (uint8_t)iValue;
916	else if (iValue > -64)
917	*Ptr.pb++ = ((uint8_t)iValue & 0x3f) \| 0x40;
918	else
919	{
920	Assert(iValue > -0x2000);
921	*Ptr.pb++ = ((uint8_t)iValue & 0x7f) \| 0x80;
922	*Ptr.pb++ = ((uint8_t)(iValue >> 7) & 0x3f) \| 0x40;
923	}
924	return Ptr;
925	}
926
927
928	/**
929	* Emits an ULEB128 encoded value (up to 64-bit wide).
930	*/
931	DECLINLINE(RTPTRUNION) iemDwarfPutUleb128(RTPTRUNION Ptr, uint64_t uValue)
932	{
933	while (uValue >= 0x80)
934	{
935	*Ptr.pb++ = ((uint8_t)uValue & 0x7f) \| 0x80;
936	uValue >>= 7;
937	}
938	*Ptr.pb++ = (uint8_t)uValue;
939	return Ptr;
940	}
941
942
943	/**
944	* Emits a CFA rule as register @a uReg + offset @a off.
945	*/
946	DECLINLINE(RTPTRUNION) iemDwarfPutCfaDefCfa(RTPTRUNION Ptr, uint32_t uReg, uint32_t off)
947	{
948	*Ptr.pb++ = DW_CFA_def_cfa;
949	Ptr = iemDwarfPutUleb128(Ptr, uReg);
950	Ptr = iemDwarfPutUleb128(Ptr, off);
951	return Ptr;
952	}
953
954
955	/**
956	* Emits a register (@a uReg) save location:
957	* CFA + @a off * data_alignment_factor
958	*/
959	DECLINLINE(RTPTRUNION) iemDwarfPutCfaOffset(RTPTRUNION Ptr, uint32_t uReg, uint32_t off)
960	{
961	if (uReg < 0x40)
962	*Ptr.pb++ = DW_CFA_offset \| uReg;
963	else
964	{
965	*Ptr.pb++ = DW_CFA_offset_extended;
966	Ptr = iemDwarfPutUleb128(Ptr, uReg);
967	}
968	Ptr = iemDwarfPutUleb128(Ptr, off);
969	return Ptr;
970	}
971
972
973	# if 0 /* unused */
974	/**
975	* Emits a register (@a uReg) save location, using signed offset:
976	* CFA + @a offSigned * data_alignment_factor
977	*/
978	DECLINLINE(RTPTRUNION) iemDwarfPutCfaSignedOffset(RTPTRUNION Ptr, uint32_t uReg, int32_t offSigned)
979	{
980	*Ptr.pb++ = DW_CFA_offset_extended_sf;
981	Ptr = iemDwarfPutUleb128(Ptr, uReg);
982	Ptr = iemDwarfPutLeb128(Ptr, offSigned);
983	return Ptr;
984	}
985	# endif
986
987
988	/**
989	* Initializes the unwind info section for non-windows hosts.
990	*/
991	static int
992	iemExecMemAllocatorInitAndRegisterUnwindInfoForChunk(PVMCPUCC pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator,
993	void *pvChunk, uint32_t idxChunk)
994	{
995	PIEMEXECMEMCHUNKEHFRAME const pEhFrame = &pExecMemAllocator->paEhFrames[idxChunk];
996	pExecMemAllocator->aChunks[idxChunk].pvUnwindInfo = pEhFrame; /* not necessary, but whatever */
997
998	RTPTRUNION Ptr = { pEhFrame->abEhFrame };
999
1000	/*
1001	* Generate the CIE first.
1002	*/
1003	# ifdef IEMNATIVE_USE_LIBUNWIND /* libunwind (llvm, darwin) only supports v1 and v3. */
1004	uint8_t const iDwarfVer = 3;
1005	# else
1006	uint8_t const iDwarfVer = 4;
1007	# endif
1008	RTPTRUNION const PtrCie = Ptr;
1009	Ptr.pu32++ = 123; / The CIE length will be determined later. */
1010	Ptr.pu32++ = 0 /UINT32_MAX/; / I'm a CIE in .eh_frame speak. */
1011	Ptr.pb++ = iDwarfVer; / DwARF version */
1012	Ptr.pb++ = 0; / Augmentation. */
1013	if (iDwarfVer >= 4)
1014	{
1015	Ptr.pb++ = sizeof(uintptr_t); / Address size. */
1016	Ptr.pb++ = 0; / Segment selector size. */
1017	}
1018	# ifdef RT_ARCH_AMD64
1019	Ptr = iemDwarfPutLeb128(Ptr, 1); /* Code alignment factor (LEB128 = 1). */
1020	# else
1021	Ptr = iemDwarfPutLeb128(Ptr, 4); /* Code alignment factor (LEB128 = 4). */
1022	# endif
1023	Ptr = iemDwarfPutLeb128(Ptr, -8); /* Data alignment factor (LEB128 = -8). */
1024	# ifdef RT_ARCH_AMD64
1025	Ptr = iemDwarfPutUleb128(Ptr, DWREG_AMD64_RA); /* Return address column (ULEB128) */
1026	# elif defined(RT_ARCH_ARM64)
1027	Ptr = iemDwarfPutUleb128(Ptr, DWREG_ARM64_LR); /* Return address column (ULEB128) */
1028	# else
1029	# error "port me"
1030	# endif
1031	/* Initial instructions: */
1032	# ifdef RT_ARCH_AMD64
1033	Ptr = iemDwarfPutCfaDefCfa(Ptr, DWREG_AMD64_RBP, 16); /* CFA = RBP + 0x10 - first stack parameter */
1034	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_RA, 1); /* Ret RIP = [CFA + 1-8] /
1035	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_RBP, 2); /* RBP = [CFA + 2-8] /
1036	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_RBX, 3); /* RBX = [CFA + 3-8] /
1037	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_R12, 4); /* R12 = [CFA + 4-8] /
1038	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_R13, 5); /* R13 = [CFA + 5-8] /
1039	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_R14, 6); /* R14 = [CFA + 6-8] /
1040	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_R15, 7); /* R15 = [CFA + 7-8] /
1041	# elif defined(RT_ARCH_ARM64)
1042	# if 1
1043	Ptr = iemDwarfPutCfaDefCfa(Ptr, DWREG_ARM64_BP, 16); /* CFA = BP + 0x10 - first stack parameter */
1044	# else
1045	Ptr = iemDwarfPutCfaDefCfa(Ptr, DWREG_ARM64_SP, IEMNATIVE_FRAME_VAR_SIZE + IEMNATIVE_FRAME_SAVE_REG_SIZE);
1046	# endif
1047	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_LR, 1); /* Ret PC = [CFA + 1-8] /
1048	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_BP, 2); /* Ret BP = [CFA + 2-8] /
1049	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X28, 3); /* X28 = [CFA + 3-8] /
1050	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X27, 4); /* X27 = [CFA + 4-8] /
1051	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X26, 5); /* X26 = [CFA + 5-8] /
1052	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X25, 6); /* X25 = [CFA + 6-8] /
1053	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X24, 7); /* X24 = [CFA + 7-8] /
1054	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X23, 8); /* X23 = [CFA + 8-8] /
1055	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X22, 9); /* X22 = [CFA + 9-8] /
1056	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X21, 10); /* X21 = [CFA +10-8] /
1057	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X20, 11); /* X20 = [CFA +11-8] /
1058	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X19, 12); /* X19 = [CFA +12-8] /
1059	AssertCompile(IEMNATIVE_FRAME_SAVE_REG_SIZE / 8 == 12);
1060	/** @todo we we need to do something about clearing DWREG_ARM64_RA_SIGN_STATE or something? */
1061	# else
1062	# error "port me"
1063	# endif
1064	while ((Ptr.u - PtrCie.u) & 3)
1065	*Ptr.pb++ = DW_CFA_nop;
1066	/* Finalize the CIE size. */
1067	*PtrCie.pu32 = Ptr.u - PtrCie.u - sizeof(uint32_t);
1068
1069	/*
1070	* Generate an FDE for the whole chunk area.
1071	*/
1072	# ifdef IEMNATIVE_USE_LIBUNWIND
1073	pEhFrame->offFda = Ptr.u - (uintptr_t)&pEhFrame->abEhFrame[0];
1074	# endif
1075	RTPTRUNION const PtrFde = Ptr;
1076	Ptr.pu32++ = 123; / The CIE length will be determined later. */
1077	Ptr.pu32 = Ptr.u - PtrCie.u; / Negated self relative CIE address. */
1078	Ptr.pu32++;
1079	Ptr.pu64++ = (uintptr_t)pvChunk; / Absolute start PC of this FDE. */
1080	Ptr.pu64++ = pExecMemAllocator->cbChunk; / PC range length for this PDE. */
1081	# if 0 /* not requried for recent libunwind.dylib nor recent libgcc/glib. */
1082	*Ptr.pb++ = DW_CFA_nop;
1083	# endif
1084	while ((Ptr.u - PtrFde.u) & 3)
1085	*Ptr.pb++ = DW_CFA_nop;
1086	/* Finalize the FDE size. */
1087	*PtrFde.pu32 = Ptr.u - PtrFde.u - sizeof(uint32_t);
1088
1089	/* Terminator entry. */
1090	*Ptr.pu32++ = 0;
1091	Ptr.pu32++ = 0; / just to be sure... */
1092	Assert(Ptr.u - (uintptr_t)&pEhFrame->abEhFrame[0] <= sizeof(pEhFrame->abEhFrame));
1093
1094	/*
1095	* Register it.
1096	*/
1097	# ifdef IEMNATIVE_USE_LIBUNWIND
1098	__register_frame(&pEhFrame->abEhFrame[pEhFrame->offFda]);
1099	# else
1100	memset(pEhFrame->abObject, 0xf6, sizeof(pEhFrame->abObject)); /* color the memory to better spot usage */
1101	__register_frame_info(pEhFrame->abEhFrame, pEhFrame->abObject);
1102	# endif
1103
1104	# ifdef IEMNATIVE_USE_GDB_JIT
1105	/*
1106	* Now for telling GDB about this (experimental).
1107	*
1108	* This seems to work best with ET_DYN.
1109	*/
1110	GDBJITSYMFILE * const pSymFile = (GDBJITSYMFILE *)iemExecMemAllocatorAllocInChunk(pExecMemAllocator, idxChunk,
1111	sizeof(GDBJITSYMFILE), NULL, NULL, NULL);
1112	AssertReturn(pSymFile, VERR_INTERNAL_ERROR_5);
1113	unsigned const offSymFileInChunk = (uintptr_t)pSymFile - (uintptr_t)pvChunk;
1114
1115	RT_ZERO(*pSymFile);
1116
1117	/*
1118	* The ELF header:
1119	*/
1120	pSymFile->EHdr.e_ident[0] = ELFMAG0;
1121	pSymFile->EHdr.e_ident[1] = ELFMAG1;
1122	pSymFile->EHdr.e_ident[2] = ELFMAG2;
1123	pSymFile->EHdr.e_ident[3] = ELFMAG3;
1124	pSymFile->EHdr.e_ident[EI_VERSION] = EV_CURRENT;
1125	pSymFile->EHdr.e_ident[EI_CLASS] = ELFCLASS64;
1126	pSymFile->EHdr.e_ident[EI_DATA] = ELFDATA2LSB;
1127	pSymFile->EHdr.e_ident[EI_OSABI] = ELFOSABI_NONE;
1128	# ifdef IEMNATIVE_USE_GDB_JIT_ET_DYN
1129	pSymFile->EHdr.e_type = ET_DYN;
1130	# else
1131	pSymFile->EHdr.e_type = ET_REL;
1132	# endif
1133	# ifdef RT_ARCH_AMD64
1134	pSymFile->EHdr.e_machine = EM_AMD64;
1135	# elif defined(RT_ARCH_ARM64)
1136	pSymFile->EHdr.e_machine = EM_AARCH64;
1137	# else
1138	# error "port me"
1139	# endif
1140	pSymFile->EHdr.e_version = 1; /?/
1141	pSymFile->EHdr.e_entry = 0;
1142	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1143	pSymFile->EHdr.e_phoff = RT_UOFFSETOF(GDBJITSYMFILE, aPhdrs);
1144	# else
1145	pSymFile->EHdr.e_phoff = 0;
1146	# endif
1147	pSymFile->EHdr.e_shoff = sizeof(pSymFile->EHdr);
1148	pSymFile->EHdr.e_flags = 0;
1149	pSymFile->EHdr.e_ehsize = sizeof(pSymFile->EHdr);
1150	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1151	pSymFile->EHdr.e_phentsize = sizeof(pSymFile->aPhdrs[0]);
1152	pSymFile->EHdr.e_phnum = RT_ELEMENTS(pSymFile->aPhdrs);
1153	# else
1154	pSymFile->EHdr.e_phentsize = 0;
1155	pSymFile->EHdr.e_phnum = 0;
1156	# endif
1157	pSymFile->EHdr.e_shentsize = sizeof(pSymFile->aShdrs[0]);
1158	pSymFile->EHdr.e_shnum = RT_ELEMENTS(pSymFile->aShdrs);
1159	pSymFile->EHdr.e_shstrndx = 0; /* set later */
1160
1161	uint32_t offStrTab = 0;
1162	#define APPEND_STR(a_szStr) do { \
1163	memcpy(&pSymFile->szzStrTab[offStrTab], a_szStr, sizeof(a_szStr)); \
1164	offStrTab += sizeof(a_szStr); \
1165	Assert(offStrTab < sizeof(pSymFile->szzStrTab)); \
1166	} while (0)
1167	#define APPEND_STR_FMT(a_szStr, ...) do { \
1168	offStrTab += RTStrPrintf(&pSymFile->szzStrTab[offStrTab], sizeof(pSymFile->szzStrTab) - offStrTab, a_szStr, __VA_ARGS__); \
1169	offStrTab++; \
1170	Assert(offStrTab < sizeof(pSymFile->szzStrTab)); \
1171	} while (0)
1172
1173	/*
1174	* Section headers.
1175	*/
1176	/* Section header #0: NULL */
1177	unsigned i = 0;
1178	APPEND_STR("");
1179	RT_ZERO(pSymFile->aShdrs[i]);
1180	i++;
1181
1182	/* Section header: .eh_frame */
1183	pSymFile->aShdrs[i].sh_name = offStrTab;
1184	APPEND_STR(".eh_frame");
1185	pSymFile->aShdrs[i].sh_type = SHT_PROGBITS;
1186	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC \| SHF_EXECINSTR;
1187	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN) \|\| defined(IEMNATIVE_USE_GDB_JIT_ELF_RVAS)
1188	pSymFile->aShdrs[i].sh_offset
1189	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, abEhFrame);
1190	# else
1191	pSymFile->aShdrs[i].sh_addr = (uintptr_t)&pSymFile->abEhFrame[0];
1192	pSymFile->aShdrs[i].sh_offset = 0;
1193	# endif
1194
1195	pSymFile->aShdrs[i].sh_size = sizeof(pEhFrame->abEhFrame);
1196	pSymFile->aShdrs[i].sh_link = 0;
1197	pSymFile->aShdrs[i].sh_info = 0;
1198	pSymFile->aShdrs[i].sh_addralign = 1;
1199	pSymFile->aShdrs[i].sh_entsize = 0;
1200	memcpy(pSymFile->abEhFrame, pEhFrame->abEhFrame, sizeof(pEhFrame->abEhFrame));
1201	i++;
1202
1203	/* Section header: .shstrtab */
1204	unsigned const iShStrTab = i;
1205	pSymFile->EHdr.e_shstrndx = iShStrTab;
1206	pSymFile->aShdrs[i].sh_name = offStrTab;
1207	APPEND_STR(".shstrtab");
1208	pSymFile->aShdrs[i].sh_type = SHT_STRTAB;
1209	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC;
1210	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN) \|\| defined(IEMNATIVE_USE_GDB_JIT_ELF_RVAS)
1211	pSymFile->aShdrs[i].sh_offset
1212	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, szzStrTab);
1213	# else
1214	pSymFile->aShdrs[i].sh_addr = (uintptr_t)&pSymFile->szzStrTab[0];
1215	pSymFile->aShdrs[i].sh_offset = 0;
1216	# endif
1217	pSymFile->aShdrs[i].sh_size = sizeof(pSymFile->szzStrTab);
1218	pSymFile->aShdrs[i].sh_link = 0;
1219	pSymFile->aShdrs[i].sh_info = 0;
1220	pSymFile->aShdrs[i].sh_addralign = 1;
1221	pSymFile->aShdrs[i].sh_entsize = 0;
1222	i++;
1223
1224	/* Section header: .symbols */
1225	pSymFile->aShdrs[i].sh_name = offStrTab;
1226	APPEND_STR(".symtab");
1227	pSymFile->aShdrs[i].sh_type = SHT_SYMTAB;
1228	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC;
1229	pSymFile->aShdrs[i].sh_offset
1230	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, aSymbols);
1231	pSymFile->aShdrs[i].sh_size = sizeof(pSymFile->aSymbols);
1232	pSymFile->aShdrs[i].sh_link = iShStrTab;
1233	pSymFile->aShdrs[i].sh_info = RT_ELEMENTS(pSymFile->aSymbols);
1234	pSymFile->aShdrs[i].sh_addralign = sizeof(pSymFile->aSymbols[0].st_value);
1235	pSymFile->aShdrs[i].sh_entsize = sizeof(pSymFile->aSymbols[0]);
1236	i++;
1237
1238	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1239	/* Section header: .symbols */
1240	pSymFile->aShdrs[i].sh_name = offStrTab;
1241	APPEND_STR(".dynsym");
1242	pSymFile->aShdrs[i].sh_type = SHT_DYNSYM;
1243	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC;
1244	pSymFile->aShdrs[i].sh_offset
1245	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, aDynSyms);
1246	pSymFile->aShdrs[i].sh_size = sizeof(pSymFile->aDynSyms);
1247	pSymFile->aShdrs[i].sh_link = iShStrTab;
1248	pSymFile->aShdrs[i].sh_info = RT_ELEMENTS(pSymFile->aDynSyms);
1249	pSymFile->aShdrs[i].sh_addralign = sizeof(pSymFile->aDynSyms[0].st_value);
1250	pSymFile->aShdrs[i].sh_entsize = sizeof(pSymFile->aDynSyms[0]);
1251	i++;
1252	# endif
1253
1254	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1255	/* Section header: .dynamic */
1256	pSymFile->aShdrs[i].sh_name = offStrTab;
1257	APPEND_STR(".dynamic");
1258	pSymFile->aShdrs[i].sh_type = SHT_DYNAMIC;
1259	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC;
1260	pSymFile->aShdrs[i].sh_offset
1261	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, aDyn);
1262	pSymFile->aShdrs[i].sh_size = sizeof(pSymFile->aDyn);
1263	pSymFile->aShdrs[i].sh_link = iShStrTab;
1264	pSymFile->aShdrs[i].sh_info = 0;
1265	pSymFile->aShdrs[i].sh_addralign = 1;
1266	pSymFile->aShdrs[i].sh_entsize = sizeof(pSymFile->aDyn[0]);
1267	i++;
1268	# endif
1269
1270	/* Section header: .text */
1271	unsigned const iShText = i;
1272	pSymFile->aShdrs[i].sh_name = offStrTab;
1273	APPEND_STR(".text");
1274	pSymFile->aShdrs[i].sh_type = SHT_PROGBITS;
1275	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC \| SHF_EXECINSTR;
1276	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN) \|\| defined(IEMNATIVE_USE_GDB_JIT_ELF_RVAS)
1277	pSymFile->aShdrs[i].sh_offset
1278	= pSymFile->aShdrs[i].sh_addr = sizeof(GDBJITSYMFILE);
1279	# else
1280	pSymFile->aShdrs[i].sh_addr = (uintptr_t)(pSymFile + 1);
1281	pSymFile->aShdrs[i].sh_offset = 0;
1282	# endif
1283	pSymFile->aShdrs[i].sh_size = pExecMemAllocator->cbChunk - offSymFileInChunk - sizeof(GDBJITSYMFILE);
1284	pSymFile->aShdrs[i].sh_link = 0;
1285	pSymFile->aShdrs[i].sh_info = 0;
1286	pSymFile->aShdrs[i].sh_addralign = 1;
1287	pSymFile->aShdrs[i].sh_entsize = 0;
1288	i++;
1289
1290	Assert(i == RT_ELEMENTS(pSymFile->aShdrs));
1291
1292	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1293	/*
1294	* The program headers:
1295	*/
1296	/* Everything in a single LOAD segment: */
1297	i = 0;
1298	pSymFile->aPhdrs[i].p_type = PT_LOAD;
1299	pSymFile->aPhdrs[i].p_flags = PF_X \| PF_R;
1300	pSymFile->aPhdrs[i].p_offset
1301	= pSymFile->aPhdrs[i].p_vaddr
1302	= pSymFile->aPhdrs[i].p_paddr = 0;
1303	pSymFile->aPhdrs[i].p_filesz /* Size of segment in file. */
1304	= pSymFile->aPhdrs[i].p_memsz = pExecMemAllocator->cbChunk - offSymFileInChunk;
1305	pSymFile->aPhdrs[i].p_align = HOST_PAGE_SIZE;
1306	i++;
1307	/* The .dynamic segment. */
1308	pSymFile->aPhdrs[i].p_type = PT_DYNAMIC;
1309	pSymFile->aPhdrs[i].p_flags = PF_R;
1310	pSymFile->aPhdrs[i].p_offset
1311	= pSymFile->aPhdrs[i].p_vaddr
1312	= pSymFile->aPhdrs[i].p_paddr = RT_UOFFSETOF(GDBJITSYMFILE, aDyn);
1313	pSymFile->aPhdrs[i].p_filesz /* Size of segment in file. */
1314	= pSymFile->aPhdrs[i].p_memsz = sizeof(pSymFile->aDyn);
1315	pSymFile->aPhdrs[i].p_align = sizeof(pSymFile->aDyn[0].d_tag);
1316	i++;
1317
1318	Assert(i == RT_ELEMENTS(pSymFile->aPhdrs));
1319
1320	/*
1321	* The dynamic section:
1322	*/
1323	i = 0;
1324	pSymFile->aDyn[i].d_tag = DT_SONAME;
1325	pSymFile->aDyn[i].d_un.d_val = offStrTab;
1326	APPEND_STR_FMT("iem-exec-chunk-%u-%u", pVCpu->idCpu, idxChunk);
1327	i++;
1328	pSymFile->aDyn[i].d_tag = DT_STRTAB;
1329	pSymFile->aDyn[i].d_un.d_ptr = RT_UOFFSETOF(GDBJITSYMFILE, szzStrTab);
1330	i++;
1331	pSymFile->aDyn[i].d_tag = DT_STRSZ;
1332	pSymFile->aDyn[i].d_un.d_val = sizeof(pSymFile->szzStrTab);
1333	i++;
1334	pSymFile->aDyn[i].d_tag = DT_SYMTAB;
1335	pSymFile->aDyn[i].d_un.d_ptr = RT_UOFFSETOF(GDBJITSYMFILE, aDynSyms);
1336	i++;
1337	pSymFile->aDyn[i].d_tag = DT_SYMENT;
1338	pSymFile->aDyn[i].d_un.d_val = sizeof(pSymFile->aDynSyms[0]);
1339	i++;
1340	pSymFile->aDyn[i].d_tag = DT_NULL;
1341	i++;
1342	Assert(i == RT_ELEMENTS(pSymFile->aDyn));
1343	# endif /* IEMNATIVE_USE_GDB_JIT_ET_DYN */
1344
1345	/*
1346	* Symbol tables:
1347	*/
1348	/** @todo gdb doesn't seem to really like this ... */
1349	i = 0;
1350	pSymFile->aSymbols[i].st_name = 0;
1351	pSymFile->aSymbols[i].st_shndx = SHN_UNDEF;
1352	pSymFile->aSymbols[i].st_value = 0;
1353	pSymFile->aSymbols[i].st_size = 0;
1354	pSymFile->aSymbols[i].st_info = ELF64_ST_INFO(STB_LOCAL, STT_NOTYPE);
1355	pSymFile->aSymbols[i].st_other = 0 /* STV_DEFAULT */;
1356	# ifdef IEMNATIVE_USE_GDB_JIT_ET_DYN
1357	pSymFile->aDynSyms[0] = pSymFile->aSymbols[i];
1358	# endif
1359	i++;
1360
1361	pSymFile->aSymbols[i].st_name = 0;
1362	pSymFile->aSymbols[i].st_shndx = SHN_ABS;
1363	pSymFile->aSymbols[i].st_value = 0;
1364	pSymFile->aSymbols[i].st_size = 0;
1365	pSymFile->aSymbols[i].st_info = ELF64_ST_INFO(STB_LOCAL, STT_FILE);
1366	pSymFile->aSymbols[i].st_other = 0 /* STV_DEFAULT */;
1367	i++;
1368
1369	pSymFile->aSymbols[i].st_name = offStrTab;
1370	APPEND_STR_FMT("iem_exec_chunk_%u_%u", pVCpu->idCpu, idxChunk);
1371	# if 0
1372	pSymFile->aSymbols[i].st_shndx = iShText;
1373	pSymFile->aSymbols[i].st_value = 0;
1374	# else
1375	pSymFile->aSymbols[i].st_shndx = SHN_ABS;
1376	pSymFile->aSymbols[i].st_value = (uintptr_t)(pSymFile + 1);
1377	# endif
1378	pSymFile->aSymbols[i].st_size = pSymFile->aShdrs[iShText].sh_size;
1379	pSymFile->aSymbols[i].st_info = ELF64_ST_INFO(STB_GLOBAL, STT_FUNC);
1380	pSymFile->aSymbols[i].st_other = 0 /* STV_DEFAULT */;
1381	# ifdef IEMNATIVE_USE_GDB_JIT_ET_DYN
1382	pSymFile->aDynSyms[1] = pSymFile->aSymbols[i];
1383	pSymFile->aDynSyms[1].st_value = (uintptr_t)(pSymFile + 1);
1384	# endif
1385	i++;
1386
1387	Assert(i == RT_ELEMENTS(pSymFile->aSymbols));
1388	Assert(offStrTab < sizeof(pSymFile->szzStrTab));
1389
1390	/*
1391	* The GDB JIT entry and informing GDB.
1392	*/
1393	pEhFrame->GdbJitEntry.pbSymFile = (uint8_t *)pSymFile;
1394	# if 1
1395	pEhFrame->GdbJitEntry.cbSymFile = pExecMemAllocator->cbChunk - ((uintptr_t)pSymFile - (uintptr_t)pvChunk);
1396	# else
1397	pEhFrame->GdbJitEntry.cbSymFile = sizeof(GDBJITSYMFILE);
1398	# endif
1399
1400	RTOnce(&g_IemNativeGdbJitOnce, iemNativeGdbJitInitOnce, NULL);
1401	RTCritSectEnter(&g_IemNativeGdbJitLock);
1402	pEhFrame->GdbJitEntry.pNext = NULL;
1403	pEhFrame->GdbJitEntry.pPrev = __jit_debug_descriptor.pTail;
1404	if (__jit_debug_descriptor.pTail)
1405	__jit_debug_descriptor.pTail->pNext = &pEhFrame->GdbJitEntry;
1406	else
1407	__jit_debug_descriptor.pHead = &pEhFrame->GdbJitEntry;
1408	__jit_debug_descriptor.pTail = &pEhFrame->GdbJitEntry;
1409	__jit_debug_descriptor.pRelevant = &pEhFrame->GdbJitEntry;
1410
1411	/* Notify GDB: */
1412	__jit_debug_descriptor.enmAction = kGdbJitaction_Register;
1413	__jit_debug_register_code();
1414	__jit_debug_descriptor.enmAction = kGdbJitaction_NoAction;
1415	RTCritSectLeave(&g_IemNativeGdbJitLock);
1416
1417	# else /* !IEMNATIVE_USE_GDB_JIT */
1418	RT_NOREF(pVCpu);
1419	# endif /* !IEMNATIVE_USE_GDB_JIT */
1420
1421	return VINF_SUCCESS;
1422	}
1423
1424	# endif /* !RT_OS_WINDOWS */
1425	#endif /* IN_RING3 */
1426
1427
1428	/**
1429	* Adds another chunk to the executable memory allocator.
1430	*
1431	* This is used by the init code for the initial allocation and later by the
1432	* regular allocator function when it's out of memory.
1433	*/
1434	static int iemExecMemAllocatorGrow(PVMCPUCC pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator)
1435	{
1436	/* Check that we've room for growth. */
1437	uint32_t const idxChunk = pExecMemAllocator->cChunks;
1438	AssertLogRelReturn(idxChunk < pExecMemAllocator->cMaxChunks, VERR_OUT_OF_RESOURCES);
1439
1440	/* Allocate a chunk. */
1441	#ifdef RT_OS_DARWIN
1442	void *pvChunk = RTMemPageAllocEx(pExecMemAllocator->cbChunk, 0);
1443	#else
1444	void *pvChunk = RTMemPageAllocEx(pExecMemAllocator->cbChunk, RTMEMPAGEALLOC_F_EXECUTABLE);
1445	#endif
1446	AssertLogRelReturn(pvChunk, VERR_NO_EXEC_MEMORY);
1447
1448	#ifdef RT_OS_DARWIN
1449	/*
1450	* Because it is impossible to have a RWX memory allocation on macOS try to remap the memory
1451	* chunk readable/executable somewhere else so we can save us the hassle of switching between
1452	* protections when exeuctable memory is allocated.
1453	*/
1454	int rc = VERR_NO_EXEC_MEMORY;
1455	mach_port_t hPortTask = mach_task_self();
1456	mach_vm_address_t AddrChunk = (mach_vm_address_t)pvChunk;
1457	mach_vm_address_t AddrRemapped = 0;
1458	vm_prot_t ProtCur = 0;
1459	vm_prot_t ProtMax = 0;
1460	kern_return_t krc = mach_vm_remap(hPortTask, &AddrRemapped, pExecMemAllocator->cbChunk, 0,
1461	VM_FLAGS_ANYWHERE \| VM_FLAGS_RETURN_DATA_ADDR,
1462	hPortTask, AddrChunk, FALSE, &ProtCur, &ProtMax,
1463	VM_INHERIT_NONE);
1464	if (krc == KERN_SUCCESS)
1465	{
1466	krc = mach_vm_protect(mach_task_self(), AddrRemapped, pExecMemAllocator->cbChunk, FALSE, VM_PROT_READ \| VM_PROT_EXECUTE);
1467	if (krc == KERN_SUCCESS)
1468	rc = VINF_SUCCESS;
1469	else
1470	{
1471	AssertLogRelMsgFailed(("mach_vm_protect -> %d (%#x)\n", krc, krc));
1472	krc = mach_vm_deallocate(hPortTask, AddrRemapped, pExecMemAllocator->cbChunk);
1473	Assert(krc == KERN_SUCCESS);
1474	}
1475	}
1476	else
1477	AssertLogRelMsgFailed(("mach_vm_remap -> %d (%#x)\n", krc, krc));
1478	if (RT_FAILURE(rc))
1479	{
1480	RTMemPageFree(pvChunk, pExecMemAllocator->cbChunk);
1481	return rc;
1482	}
1483
1484	void pvChunkRx = (void )AddrRemapped;
1485	#else
1486	# if defined(IN_RING3) \|\| defined(IEMNATIVE_WITH_RECOMPILER_PER_CHUNK_TAIL_CODE)
1487	int rc = VINF_SUCCESS;
1488	# endif
1489	void *pvChunkRx = pvChunk;
1490	#endif
1491
1492	/*
1493	* Add the chunk.
1494	*
1495	* This must be done before the unwind init so windows can allocate
1496	* memory from the chunk when using the alternative sub-allocator.
1497	*/
1498	pExecMemAllocator->aChunks[idxChunk].pvChunkRw = pvChunk;
1499	pExecMemAllocator->aChunks[idxChunk].pvChunkRx = pvChunkRx;
1500	#ifdef IN_RING3
1501	pExecMemAllocator->aChunks[idxChunk].pvUnwindInfo = NULL;
1502	#endif
1503	pExecMemAllocator->aChunks[idxChunk].cFreeUnits = pExecMemAllocator->cUnitsPerChunk;
1504	pExecMemAllocator->aChunks[idxChunk].idxFreeHint = 0;
1505	memset(&pExecMemAllocator->pbmAlloc[pExecMemAllocator->cBitmapElementsPerChunk * idxChunk],
1506	0, sizeof(pExecMemAllocator->pbmAlloc[0]) * pExecMemAllocator->cBitmapElementsPerChunk);
1507
1508	pExecMemAllocator->cChunks = idxChunk + 1;
1509	pExecMemAllocator->idxChunkHint = idxChunk;
1510
1511	pExecMemAllocator->cbTotal += pExecMemAllocator->cbChunk;
1512	pExecMemAllocator->cbFree += pExecMemAllocator->cbChunk;
1513
1514	/* If there is a chunk context init callback call it. */
1515	#ifdef IEMNATIVE_WITH_RECOMPILER_PER_CHUNK_TAIL_CODE
1516	pExecMemAllocator->aChunks[idxChunk].pCtx = iemNativeRecompileAttachExecMemChunkCtx(pVCpu, idxChunk);
1517	if (pExecMemAllocator->aChunks[idxChunk].pCtx)
1518	#endif
1519	{
1520	#ifdef IN_RING3
1521	/*
1522	* Initialize the unwind information (this cannot really fail atm).
1523	* (This sets pvUnwindInfo.)
1524	*/
1525	rc = iemExecMemAllocatorInitAndRegisterUnwindInfoForChunk(pVCpu, pExecMemAllocator, pvChunkRx, idxChunk);
1526	#endif
1527	}
1528	#if defined(IN_RING3) \|\| defined(IEMNATIVE_WITH_RECOMPILER_PER_CHUNK_TAIL_CODE)
1529	if (RT_SUCCESS(rc))
1530	{ /* likely */ }
1531	else
1532	{
1533	/* Just in case the impossible happens, undo the above up: */
1534	pExecMemAllocator->cbTotal -= pExecMemAllocator->cbChunk;
1535	pExecMemAllocator->cbFree -= pExecMemAllocator->aChunks[idxChunk].cFreeUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
1536	pExecMemAllocator->cChunks = idxChunk;
1537	memset(&pExecMemAllocator->pbmAlloc[pExecMemAllocator->cBitmapElementsPerChunk * idxChunk],
1538	0xff, sizeof(pExecMemAllocator->pbmAlloc[0]) * pExecMemAllocator->cBitmapElementsPerChunk);
1539	pExecMemAllocator->aChunks[idxChunk].pvChunkRw = NULL;
1540	pExecMemAllocator->aChunks[idxChunk].cFreeUnits = 0;
1541
1542	# ifdef RT_OS_DARWIN
1543	krc = mach_vm_deallocate(mach_task_self(), (mach_vm_address_t)pExecMemAllocator->aChunks[idxChunk].pvChunkRx,
1544	pExecMemAllocator->cbChunk);
1545	Assert(krc == KERN_SUCCESS);
1546	# endif
1547
1548	RTMemPageFree(pvChunk, pExecMemAllocator->cbChunk);
1549	return rc;
1550	}
1551	#endif
1552
1553	return VINF_SUCCESS;
1554	}
1555
1556
1557	/**
1558	* Initializes the executable memory allocator for native recompilation on the
1559	* calling EMT.
1560	*
1561	* @returns VBox status code.
1562	* @param pVCpu The cross context virtual CPU structure of the calling
1563	* thread.
1564	* @param cbMax The max size of the allocator.
1565	* @param cbInitial The initial allocator size.
1566	* @param cbChunk The chunk size, 0 or UINT32_MAX for default (@a cbMax
1567	* dependent).
1568	*/
1569	int iemExecMemAllocatorInit(PVMCPU pVCpu, uint64_t cbMax, uint64_t cbInitial, uint32_t cbChunk) RT_NOEXCEPT
1570	{
1571	/*
1572	* Validate input.
1573	*/
1574	AssertLogRelMsgReturn(cbMax >= _1M && cbMax <= _4G+_4G, ("cbMax=%RU64 (%RX64)\n", cbMax, cbMax), VERR_OUT_OF_RANGE);
1575	AssertReturn(cbInitial <= cbMax, VERR_OUT_OF_RANGE);
1576	AssertLogRelMsgReturn( cbChunk != UINT32_MAX
1577	\|\| cbChunk == 0
1578	\|\| ( RT_IS_POWER_OF_TWO(cbChunk)
1579	&& cbChunk >= _1M
1580	&& cbChunk <= _256M
1581	&& cbChunk <= cbMax),
1582	("cbChunk=%RU32 (%RX32) cbMax=%RU64\n", cbChunk, cbChunk, cbMax),
1583	VERR_OUT_OF_RANGE);
1584
1585	/*
1586	* Adjust/figure out the chunk size.
1587	*/
1588	if (cbChunk == 0 \|\| cbChunk == UINT32_MAX)
1589	{
1590	if (cbMax >= _256M)
1591	cbChunk = _64M;
1592	else
1593	{
1594	if (cbMax < _16M)
1595	cbChunk = cbMax >= _4M ? _4M : (uint32_t)cbMax;
1596	else
1597	cbChunk = (uint32_t)cbMax / 4;
1598	if (!RT_IS_POWER_OF_TWO(cbChunk))
1599	cbChunk = RT_BIT_32(ASMBitLastSetU32(cbChunk));
1600	}
1601	}
1602	#ifdef IEMNATIVE_WITH_RECOMPILER_PER_CHUNK_TAIL_CODE
1603	# if defined(RT_OS_AMD64)
1604	Assert(cbChunk <= _2G);
1605	# elif defined(RT_OS_ARM64)
1606	if (cbChunk > _128M)
1607	cbChunk = _128M; /* Max relative branch distance is +/-2^(25+2) = +/-0x8000000 (134 217 728). */
1608	# endif
1609	#endif
1610
1611	if (cbChunk > cbMax)
1612	cbMax = cbChunk;
1613	else
1614	cbMax = (cbMax - 1 + cbChunk) / cbChunk * cbChunk;
1615	uint32_t const cMaxChunks = (uint32_t)(cbMax / cbChunk);
1616	AssertLogRelReturn((uint64_t)cMaxChunks * cbChunk == cbMax, VERR_INTERNAL_ERROR_3);
1617
1618	/*
1619	* Allocate and initialize the allocatore instance.
1620	*/
1621	size_t const offBitmaps = RT_ALIGN_Z(RT_UOFFSETOF_DYN(IEMEXECMEMALLOCATOR, aChunks[cMaxChunks]), RT_CACHELINE_SIZE);
1622	size_t const cbBitmaps = (size_t)(cbChunk >> (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT + 3)) * cMaxChunks;
1623	size_t cbNeeded = offBitmaps + cbBitmaps;
1624	AssertCompile(IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT <= 10);
1625	Assert(cbChunk > RT_BIT_32(IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT + 3));
1626	#if defined(IN_RING3) && !defined(RT_OS_WINDOWS)
1627	size_t const offEhFrames = RT_ALIGN_Z(cbNeeded, RT_CACHELINE_SIZE);
1628	cbNeeded += sizeof(IEMEXECMEMCHUNKEHFRAME) * cMaxChunks;
1629	#endif
1630	PIEMEXECMEMALLOCATOR pExecMemAllocator = (PIEMEXECMEMALLOCATOR)RTMemAllocZ(cbNeeded);
1631	AssertLogRelMsgReturn(pExecMemAllocator, ("cbNeeded=%zx cMaxChunks=%#x cbChunk=%#x\n", cbNeeded, cMaxChunks, cbChunk),
1632	VERR_NO_MEMORY);
1633	pExecMemAllocator->uMagic = IEMEXECMEMALLOCATOR_MAGIC;
1634	pExecMemAllocator->cbChunk = cbChunk;
1635	pExecMemAllocator->cMaxChunks = cMaxChunks;
1636	pExecMemAllocator->cChunks = 0;
1637	pExecMemAllocator->idxChunkHint = 0;
1638	pExecMemAllocator->cAllocations = 0;
1639	pExecMemAllocator->cbTotal = 0;
1640	pExecMemAllocator->cbFree = 0;
1641	pExecMemAllocator->cbAllocated = 0;
1642	pExecMemAllocator->pbmAlloc = (uint64_t *)((uintptr_t)pExecMemAllocator + offBitmaps);
1643	pExecMemAllocator->cUnitsPerChunk = cbChunk >> IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
1644	pExecMemAllocator->cBitmapElementsPerChunk = cbChunk >> (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT + 6);
1645	memset(pExecMemAllocator->pbmAlloc, 0xff, cbBitmaps); /* Mark everything as allocated. Clear when chunks are added. */
1646	#if defined(IN_RING3) && !defined(RT_OS_WINDOWS)
1647	pExecMemAllocator->paEhFrames = (PIEMEXECMEMCHUNKEHFRAME)((uintptr_t)pExecMemAllocator + offEhFrames);
1648	#endif
1649	for (uint32_t i = 0; i < cMaxChunks; i++)
1650	{
1651	pExecMemAllocator->aChunks[i].cFreeUnits = 0;
1652	pExecMemAllocator->aChunks[i].idxFreeHint = 0;
1653	pExecMemAllocator->aChunks[i].pvChunkRw = NULL;
1654	#ifdef IN_RING0
1655	pExecMemAllocator->aChunks[i].hMemObj = NIL_RTR0MEMOBJ;
1656	#else
1657	pExecMemAllocator->aChunks[i].pvUnwindInfo = NULL;
1658	#endif
1659	}
1660	pVCpu->iem.s.pExecMemAllocatorR3 = pExecMemAllocator;
1661
1662	/*
1663	* Do the initial allocations.
1664	*/
1665	while (cbInitial < (uint64_t)pExecMemAllocator->cChunks * pExecMemAllocator->cbChunk)
1666	{
1667	int rc = iemExecMemAllocatorGrow(pVCpu, pExecMemAllocator);
1668	AssertLogRelRCReturn(rc, rc);
1669	}
1670
1671	pExecMemAllocator->idxChunkHint = 0;
1672
1673	/*
1674	* Register statistics.
1675	*/
1676	PUVM const pUVM = pVCpu->pUVCpu->pUVM;
1677	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cAllocations, STAMTYPE_U32, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
1678	"Current number of allocations", "/IEM/CPU%u/re/ExecMem/cAllocations", pVCpu->idCpu);
1679	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cChunks, STAMTYPE_U32, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,
1680	"Currently allocated chunks", "/IEM/CPU%u/re/ExecMem/cChunks", pVCpu->idCpu);
1681	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cMaxChunks, STAMTYPE_U32, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,
1682	"Maximum number of chunks", "/IEM/CPU%u/re/ExecMem/cMaxChunks", pVCpu->idCpu);
1683	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbChunk, STAMTYPE_U32, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
1684	"Allocation chunk size", "/IEM/CPU%u/re/ExecMem/cbChunk", pVCpu->idCpu);
1685	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbAllocated, STAMTYPE_U64, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
1686	"Number of bytes current allocated", "/IEM/CPU%u/re/ExecMem/cbAllocated", pVCpu->idCpu);
1687	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbFree, STAMTYPE_U64, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
1688	"Number of bytes current free", "/IEM/CPU%u/re/ExecMem/cbFree", pVCpu->idCpu);
1689	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbTotal, STAMTYPE_U64, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
1690	"Total number of byte", "/IEM/CPU%u/re/ExecMem/cbTotal", pVCpu->idCpu);
1691	#ifdef VBOX_WITH_STATISTICS
1692	STAMR3RegisterFU(pUVM, &pExecMemAllocator->StatAlloc, STAMTYPE_PROFILE, STAMVISIBILITY_ALWAYS, STAMUNIT_TICKS_PER_CALL,
1693	"Profiling the allocator", "/IEM/CPU%u/re/ExecMem/ProfAlloc", pVCpu->idCpu);
1694	#endif
1695	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING
1696	STAMR3RegisterFU(pUVM, &pExecMemAllocator->StatPruneProf, STAMTYPE_PROFILE, STAMVISIBILITY_ALWAYS, STAMUNIT_TICKS_PER_CALL,
1697	"Pruning executable memory (alt)", "/IEM/CPU%u/re/ExecMem/Pruning", pVCpu->idCpu);
1698	STAMR3RegisterFU(pUVM, &pExecMemAllocator->StatPruneRecovered, STAMTYPE_PROFILE, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES_PER_CALL,
1699	"Bytes recovered while pruning", "/IEM/CPU%u/re/ExecMem/PruningRecovered", pVCpu->idCpu);
1700	#endif
1701
1702	return VINF_SUCCESS;
1703	}
1704

注意: 瀏覽 TracBrowser 來幫助您使用儲存庫瀏覽器

source: vbox/trunk/src/VBox/VMM/VMMAll/IEMAllN8veExecMem.cpp@ 105413

以其他格式下載: