IEMAllN8veExecMem.cpp@ 106378

Last change on this file since 106378 was 106378, checked in by vboxsync, 5 weeks ago
VMM/IEM: Fixed assertion in exec mem allocator. bugref:10720
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 88.0 KB

Line
1	/* $Id: IEMAllN8veExecMem.cpp 106378 2024-10-16 13:43:43Z vboxsync $ */
2	/** @file
3	* IEM - Native Recompiler, Executable Memory Allocator.
4	*/
5
6	/*
7	* Copyright (C) 2023-2024 Oracle and/or its affiliates.
8	*
9	* This file is part of VirtualBox base platform packages, as
10	* available from https://www.virtualbox.org.
11	*
12	* This program is free software; you can redistribute it and/or
13	* modify it under the terms of the GNU General Public License
14	* as published by the Free Software Foundation, in version 3 of the
15	* License.
16	*
17	* This program is distributed in the hope that it will be useful, but
18	* WITHOUT ANY WARRANTY; without even the implied warranty of
19	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20	* General Public License for more details.
21	*
22	* You should have received a copy of the GNU General Public License
23	* along with this program; if not, see <https://www.gnu.org/licenses>.
24	*
25	* SPDX-License-Identifier: GPL-3.0-only
26	*/
27
28
29	/*********************************************************************************************************************************
30	* Header Files *
31	*********************************************************************************************************************************/
32	#define LOG_GROUP LOG_GROUP_IEM_RE_NATIVE
33	#define IEM_WITH_OPAQUE_DECODER_STATE
34	#define VMM_INCLUDED_SRC_include_IEMMc_h /* block IEMMc.h inclusion. */
35	#include <VBox/vmm/iem.h>
36	#include <VBox/vmm/cpum.h>
37	#include "IEMInternal.h"
38	#include <VBox/vmm/vmcc.h>
39	#include <VBox/log.h>
40	#include <VBox/err.h>
41	#include <VBox/param.h>
42	#include <iprt/assert.h>
43	#include <iprt/mem.h>
44	#include <iprt/string.h>
45	#if defined(RT_ARCH_AMD64)
46	# include <iprt/x86.h>
47	#elif defined(RT_ARCH_ARM64)
48	# include <iprt/armv8.h>
49	#endif
50
51	#ifdef RT_OS_WINDOWS
52	# include <iprt/formats/pecoff.h> /* this is incomaptible with windows.h, thus: */
53	extern "C" DECLIMPORT(uint8_t) __cdecl RtlAddFunctionTable(void *pvFunctionTable, uint32_t cEntries, uintptr_t uBaseAddress);
54	extern "C" DECLIMPORT(uint8_t) __cdecl RtlDelFunctionTable(void *pvFunctionTable);
55	#else
56	# include <iprt/formats/dwarf.h>
57	# if defined(RT_OS_DARWIN)
58	# include <libkern/OSCacheControl.h>
59	# include <mach/mach.h>
60	# include <mach/mach_vm.h>
61	# define IEMNATIVE_USE_LIBUNWIND
62	extern "C" void __register_frame(const void *pvFde);
63	extern "C" void __deregister_frame(const void *pvFde);
64	# else
65	# ifdef DEBUG_bird /** @todo not thread safe yet */
66	# define IEMNATIVE_USE_GDB_JIT
67	# endif
68	# ifdef IEMNATIVE_USE_GDB_JIT
69	# include <iprt/critsect.h>
70	# include <iprt/once.h>
71	# include <iprt/formats/elf64.h>
72	# endif
73	extern "C" void __register_frame_info(void pvBegin, void pvObj); /* found no header for these two */
74	extern "C" void __deregister_frame_info(void pvBegin); /* (returns pvObj from __register_frame_info call) */
75	# endif
76	#endif
77
78	#include "IEMN8veRecompiler.h"
79
80
81	/*********************************************************************************************************************************
82	* Executable Memory Allocator *
83	*********************************************************************************************************************************/
84	/** The chunk sub-allocation unit size in bytes. */
85	#define IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE 256
86	/** The chunk sub-allocation unit size as a shift factor. */
87	#define IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT 8
88	/** Enables adding a header to the sub-allocator allocations.
89	* This is useful for freeing up executable memory among other things. */
90	#define IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
91	/** Use alternative pruning. */
92	#define IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING
93
94
95	#if defined(IN_RING3) && !defined(RT_OS_WINDOWS)
96	# ifdef IEMNATIVE_USE_GDB_JIT
97	# define IEMNATIVE_USE_GDB_JIT_ET_DYN
98
99	/** GDB JIT: Code entry. */
100	typedef struct GDBJITCODEENTRY
101	{
102	struct GDBJITCODEENTRY *pNext;
103	struct GDBJITCODEENTRY *pPrev;
104	uint8_t *pbSymFile;
105	uint64_t cbSymFile;
106	} GDBJITCODEENTRY;
107
108	/** GDB JIT: Actions. */
109	typedef enum GDBJITACTIONS : uint32_t
110	{
111	kGdbJitaction_NoAction = 0, kGdbJitaction_Register, kGdbJitaction_Unregister
112	} GDBJITACTIONS;
113
114	/** GDB JIT: Descriptor. */
115	typedef struct GDBJITDESCRIPTOR
116	{
117	uint32_t uVersion;
118	GDBJITACTIONS enmAction;
119	GDBJITCODEENTRY *pRelevant;
120	GDBJITCODEENTRY *pHead;
121	/** Our addition: */
122	GDBJITCODEENTRY *pTail;
123	} GDBJITDESCRIPTOR;
124
125	/** GDB JIT: Our simple symbol file data. */
126	typedef struct GDBJITSYMFILE
127	{
128	Elf64_Ehdr EHdr;
129	# ifndef IEMNATIVE_USE_GDB_JIT_ET_DYN
130	Elf64_Shdr aShdrs[5];
131	# else
132	Elf64_Shdr aShdrs[7];
133	Elf64_Phdr aPhdrs[2];
134	# endif
135	/** The dwarf ehframe data for the chunk. */
136	uint8_t abEhFrame[512];
137	char szzStrTab[128];
138	Elf64_Sym aSymbols[3];
139	# ifdef IEMNATIVE_USE_GDB_JIT_ET_DYN
140	Elf64_Sym aDynSyms[2];
141	Elf64_Dyn aDyn[6];
142	# endif
143	} GDBJITSYMFILE;
144
145	extern "C" GDBJITDESCRIPTOR __jit_debug_descriptor;
146	extern "C" DECLEXPORT(void) __jit_debug_register_code(void);
147
148	/** Init once for g_IemNativeGdbJitLock. */
149	static RTONCE g_IemNativeGdbJitOnce = RTONCE_INITIALIZER;
150	/** Init once for the critical section. */
151	static RTCRITSECT g_IemNativeGdbJitLock;
152
153	/** GDB reads the info here. */
154	GDBJITDESCRIPTOR __jit_debug_descriptor = { 1, kGdbJitaction_NoAction, NULL, NULL };
155
156	/** GDB sets a breakpoint on this and checks __jit_debug_descriptor when hit. */
157	DECL_NO_INLINE(RT_NOTHING, DECLEXPORT(void)) __jit_debug_register_code(void)
158	{
159	ASMNopPause();
160	}
161
162	/** @callback_method_impl{FNRTONCE} */
163	static DECLCALLBACK(int32_t) iemNativeGdbJitInitOnce(void *pvUser)
164	{
165	RT_NOREF(pvUser);
166	return RTCritSectInit(&g_IemNativeGdbJitLock);
167	}
168
169
170	# endif /* IEMNATIVE_USE_GDB_JIT */
171
172	/**
173	* Per-chunk unwind info for non-windows hosts.
174	*/
175	typedef struct IEMEXECMEMCHUNKEHFRAME
176	{
177	# ifdef IEMNATIVE_USE_LIBUNWIND
178	/** The offset of the FDA into abEhFrame. */
179	uintptr_t offFda;
180	# else
181	/** 'struct object' storage area. */
182	uint8_t abObject[1024];
183	# endif
184	# ifdef IEMNATIVE_USE_GDB_JIT
185	# if 0
186	/** The GDB JIT 'symbol file' data. */
187	GDBJITSYMFILE GdbJitSymFile;
188	# endif
189	/** The GDB JIT list entry. */
190	GDBJITCODEENTRY GdbJitEntry;
191	# endif
192	/** The dwarf ehframe data for the chunk. */
193	uint8_t abEhFrame[512];
194	} IEMEXECMEMCHUNKEHFRAME;
195	/** Pointer to per-chunk info info for non-windows hosts. */
196	typedef IEMEXECMEMCHUNKEHFRAME *PIEMEXECMEMCHUNKEHFRAME;
197	#endif
198
199
200	/**
201	* An chunk of executable memory.
202	*/
203	typedef struct IEMEXECMEMCHUNK
204	{
205	/** Number of free items in this chunk. */
206	uint32_t cFreeUnits;
207	/** Hint were to start searching for free space in the allocation bitmap. */
208	uint32_t idxFreeHint;
209	/** Pointer to the readable/writeable view of the memory chunk. */
210	void *pvChunkRw;
211	/** Pointer to the readable/executable view of the memory chunk. */
212	void *pvChunkRx;
213	/** Pointer to the context structure detailing the per chunk common code. */
214	PCIEMNATIVEPERCHUNKCTX pCtx;
215	#ifdef IN_RING3
216	/**
217	* Pointer to the unwind information.
218	*
219	* This is used during C++ throw and longjmp (windows and probably most other
220	* platforms). Some debuggers (windbg) makes use of it as well.
221	*
222	* Windows: This is allocated from hHeap on windows because (at least for
223	* AMD64) the UNWIND_INFO structure address in the
224	* RUNTIME_FUNCTION entry is an RVA and the chunk is the "image".
225	*
226	* Others: Allocated from the regular heap to avoid unnecessary executable data
227	* structures. This points to an IEMEXECMEMCHUNKEHFRAME structure. */
228	void *pvUnwindInfo;
229	#elif defined(IN_RING0)
230	/** Allocation handle. */
231	RTR0MEMOBJ hMemObj;
232	#endif
233	} IEMEXECMEMCHUNK;
234	/** Pointer to a memory chunk. */
235	typedef IEMEXECMEMCHUNK *PIEMEXECMEMCHUNK;
236
237
238	/**
239	* Executable memory allocator for the native recompiler.
240	*/
241	typedef struct IEMEXECMEMALLOCATOR
242	{
243	/** Magic value (IEMEXECMEMALLOCATOR_MAGIC). */
244	uint32_t uMagic;
245
246	/** The chunk size. */
247	uint32_t cbChunk;
248	/** The maximum number of chunks. */
249	uint32_t cMaxChunks;
250	/** The current number of chunks. */
251	uint32_t cChunks;
252	/** Hint where to start looking for available memory. */
253	uint32_t idxChunkHint;
254	/** Statistics: Current number of allocations. */
255	uint32_t cAllocations;
256
257	/** The total amount of memory available. */
258	uint64_t cbTotal;
259	/** Total amount of free memory. */
260	uint64_t cbFree;
261	/** Total amount of memory allocated. */
262	uint64_t cbAllocated;
263
264	/** Pointer to the allocation bitmaps for all the chunks (follows aChunks).
265	*
266	* Since the chunk size is a power of two and the minimum chunk size is a lot
267	* higher than the IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE, each chunk will always
268	* require a whole number of uint64_t elements in the allocation bitmap. So,
269	* for sake of simplicity, they are allocated as one continous chunk for
270	* simplicity/laziness. */
271	uint64_t *pbmAlloc;
272	/** Number of units (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE) per chunk. */
273	uint32_t cUnitsPerChunk;
274	/** Number of bitmap elements per chunk (for quickly locating the bitmap
275	* portion corresponding to an chunk). */
276	uint32_t cBitmapElementsPerChunk;
277
278	/** Number of times we fruitlessly scanned a chunk for free space. */
279	uint64_t cFruitlessChunkScans;
280
281	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING
282	/** The next chunk to prune in. */
283	uint32_t idxChunkPrune;
284	/** Where in chunk offset to start pruning at. */
285	uint32_t offChunkPrune;
286	/** Profiling the pruning code. */
287	STAMPROFILE StatPruneProf;
288	/** Number of bytes recovered by the pruning. */
289	STAMPROFILE StatPruneRecovered;
290	#endif
291
292	#ifdef VBOX_WITH_STATISTICS
293	STAMPROFILE StatAlloc;
294	/** Total amount of memory not being usable currently due to IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE. */
295	uint64_t cbUnusable;
296	/** Allocation size distribution (in alloc units; 0 is the slop bucket). */
297	STAMCOUNTER aStatSizes[16];
298	#endif
299
300	#if defined(IN_RING3) && !defined(RT_OS_WINDOWS)
301	/** Pointer to the array of unwind info running parallel to aChunks (same
302	* allocation as this structure, located after the bitmaps).
303	* (For Windows, the structures must reside in 32-bit RVA distance to the
304	* actual chunk, so they are allocated off the chunk.) */
305	PIEMEXECMEMCHUNKEHFRAME paEhFrames;
306	#endif
307
308	/** The allocation chunks. */
309	RT_FLEXIBLE_ARRAY_EXTENSION
310	IEMEXECMEMCHUNK aChunks[RT_FLEXIBLE_ARRAY];
311	} IEMEXECMEMALLOCATOR;
312	/** Pointer to an executable memory allocator. */
313	typedef IEMEXECMEMALLOCATOR *PIEMEXECMEMALLOCATOR;
314
315	/** Magic value for IEMEXECMEMALLOCATOR::uMagic (Scott Frederick Turow). */
316	#define IEMEXECMEMALLOCATOR_MAGIC UINT32_C(0x19490412)
317
318
319	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
320	/**
321	* Allocation header.
322	*/
323	typedef struct IEMEXECMEMALLOCHDR
324	{
325	union
326	{
327	struct
328	{
329	/** Magic value / eyecatcher (IEMEXECMEMALLOCHDR_MAGIC). */
330	uint32_t uMagic;
331	/** The allocation chunk (for speeding up freeing). */
332	uint32_t idxChunk;
333	};
334	/** Combined magic and chunk index, for the pruning scanner code. */
335	uint64_t u64MagicAndChunkIdx;
336	};
337	/** Pointer to the translation block the allocation belongs to.
338	* This is the whole point of the header. */
339	PIEMTB pTb;
340	} IEMEXECMEMALLOCHDR;
341	/** Pointer to an allocation header. */
342	typedef IEMEXECMEMALLOCHDR *PIEMEXECMEMALLOCHDR;
343	/** Magic value for IEMEXECMEMALLOCHDR ('ExeM'). */
344	# define IEMEXECMEMALLOCHDR_MAGIC UINT32_C(0x4d657845)
345	#endif
346
347
348	static int iemExecMemAllocatorGrow(PVMCPUCC pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator);
349
350
351	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING
352	/**
353	* Frees up executable memory when we're out space.
354	*
355	* This is an alternative to iemTbAllocatorFreeupNativeSpace() that frees up
356	* space in a more linear fashion from the allocator's point of view. It may
357	* also defragment if implemented & enabled
358	*/
359	static void iemExecMemAllocatorPrune(PVMCPU pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator)
360	{
361	# ifndef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
362	# error "IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING requires IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER"
363	# endif
364	STAM_REL_PROFILE_START(&pExecMemAllocator->StatPruneProf, a);
365
366	/*
367	* Before we can start, we must process delayed frees.
368	*/
369	#if 1
370	PIEMTBALLOCATOR const pTbAllocator = iemTbAllocatorFreeBulkStart(pVCpu);
371	#else
372	iemTbAllocatorProcessDelayedFrees(pVCpu, pVCpu->iem.s.pTbAllocatorR3);
373	#endif
374
375	AssertCompile(RT_IS_POWER_OF_TWO(IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE));
376
377	uint32_t const cbChunk = pExecMemAllocator->cbChunk;
378	AssertReturnVoid(RT_IS_POWER_OF_TWO(cbChunk));
379	AssertReturnVoid(cbChunk >= _1M && cbChunk <= _256M); /* see iemExecMemAllocatorInit */
380
381	uint32_t const cChunks = pExecMemAllocator->cChunks;
382	AssertReturnVoid(cChunks == pExecMemAllocator->cMaxChunks);
383	AssertReturnVoid(cChunks >= 1);
384
385	Assert(!pVCpu->iem.s.pCurTbR3);
386
387	/*
388	* Decide how much to prune. The chunk is is a multiple of two, so we'll be
389	* scanning a multiple of two here as well.
390	*/
391	uint32_t cbToPrune = cbChunk;
392
393	/* Never more than 25%. */
394	if (cChunks < 4)
395	cbToPrune /= cChunks == 1 ? 4 : 2;
396
397	/* Upper limit. In a debug build a 4MB limit averages out at ~0.6ms per call. */
398	if (cbToPrune > _4M)
399	cbToPrune = _4M;
400
401	/*
402	* Adjust the pruning chunk and offset accordingly.
403	*/
404	uint32_t idxChunk = pExecMemAllocator->idxChunkPrune;
405	uint32_t offChunk = pExecMemAllocator->offChunkPrune;
406	offChunk &= ~(uint32_t)(IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1U);
407	if (offChunk >= cbChunk)
408	{
409	offChunk = 0;
410	idxChunk += 1;
411	}
412	if (idxChunk >= cChunks)
413	{
414	offChunk = 0;
415	idxChunk = 0;
416	}
417
418	uint32_t const offPruneEnd = RT_MIN(offChunk + cbToPrune, cbChunk);
419
420	/*
421	* Do the pruning. The current approach is the sever kind.
422	*
423	* This is memory bound, as we must load both the allocation header and the
424	* associated TB and then modify them. So, the CPU isn't all that unitilized
425	* here. Try apply some prefetching to speed it up a tiny bit.
426	*/
427	uint64_t cbPruned = 0;
428	uint64_t const u64MagicAndChunkIdx = RT_MAKE_U64(IEMEXECMEMALLOCHDR_MAGIC, idxChunk);
429	uint8_t * const pbChunk = (uint8_t *)pExecMemAllocator->aChunks[idxChunk].pvChunkRx;
430	while (offChunk < offPruneEnd)
431	{
432	PIEMEXECMEMALLOCHDR pHdr = (PIEMEXECMEMALLOCHDR)&pbChunk[offChunk];
433
434	/* Is this the start of an allocation block for a TB? (We typically
435	have one allocation at the start of each chunk for the unwind info
436	where pTb is NULL.) */
437	PIEMTB pTb;
438	if ( pHdr->u64MagicAndChunkIdx == u64MagicAndChunkIdx
439	&& RT_LIKELY((pTb = pHdr->pTb) != NULL))
440	{
441	AssertPtr(pTb);
442
443	uint32_t const cbBlock = RT_ALIGN_32(pTb->Native.cInstructions * sizeof(IEMNATIVEINSTR) + sizeof(*pHdr),
444	IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE);
445
446	/* Prefetch the next header before freeing the current one and its TB. */
447	/** @todo Iff the block size was part of the header in some way, this could be
448	* a tiny bit faster. */
449	offChunk += cbBlock;
450	#if defined(_MSC_VER) && defined(RT_ARCH_AMD64)
451	_mm_prefetch((char *)&pbChunk[offChunk], _MM_HINT_T0);
452	#elif defined(_MSC_VER) && defined(RT_ARCH_ARM64)
453	__prefetch(&pbChunk[offChunk]);
454	#else
455	__builtin_prefetch(&pbChunk[offChunk], 1 /rw/);
456	#endif
457	/* Some paranoia first, though. */
458	AssertBreakStmt(offChunk <= cbChunk, offChunk -= cbBlock - IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE);
459	cbPruned += cbBlock;
460
461	#if 1
462	iemTbAllocatorFreeBulk(pVCpu, pTbAllocator, pTb);
463	#else
464	iemTbAllocatorFree(pVCpu, pTb);
465	#endif
466	}
467	else
468	offChunk += IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE;
469	}
470	STAM_REL_PROFILE_ADD_PERIOD(&pExecMemAllocator->StatPruneRecovered, cbPruned);
471
472	pVCpu->iem.s.ppTbLookupEntryR3 = &pVCpu->iem.s.pTbLookupEntryDummyR3;
473
474	/*
475	* Save the current pruning point.
476	*/
477	pExecMemAllocator->offChunkPrune = offChunk;
478	pExecMemAllocator->idxChunkPrune = idxChunk;
479
480	/* Set the hint to the start of the pruned region. */
481	pExecMemAllocator->idxChunkHint = idxChunk;
482	pExecMemAllocator->aChunks[idxChunk].idxFreeHint = offChunk / IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE;
483
484	STAM_REL_PROFILE_STOP(&pExecMemAllocator->StatPruneProf, a);
485	}
486	#endif /* IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING */
487
488
489	#if defined(VBOX_STRICT) \|\| 0
490	/**
491	* The old bitmap scanner code, for comparison and assertions.
492	*/
493	static uint32_t iemExecMemAllocatorFindReqFreeUnitsOld(uint64_t *pbmAlloc, uint32_t cToScan, uint32_t cReqUnits)
494	{
495	/** @todo This can probably be done more efficiently for non-x86 systems. */
496	int iBit = ASMBitFirstClear(pbmAlloc, cToScan);
497	while (iBit >= 0 && (uint32_t)iBit <= cToScan - cReqUnits)
498	{
499	uint32_t idxAddBit = 1;
500	while (idxAddBit < cReqUnits && !ASMBitTest(pbmAlloc, (uint32_t)iBit + idxAddBit))
501	idxAddBit++;
502	if (idxAddBit >= cReqUnits)
503	return (uint32_t)iBit;
504	iBit = ASMBitNextClear(pbmAlloc, cToScan, iBit + idxAddBit - 1);
505	}
506	return UINT32_MAX;
507	}
508	#endif
509
510
511	/**
512	* Bitmap scanner code that looks for a bunch of @a cReqUnits zero bits.
513	*
514	* Booting win11 with a r165098 release build the average native TB size is
515	* around 9 units (of 256 bytes). So, it is unlikely we need to scan any
516	* subsequent words once we hit a patch of zeros, thus @a a_fBig.
517	*
518	* @todo This needs more tweaking. While it is faster the the old code,
519	* it doens't seem like it's all that much. :/
520	*/
521	template<const bool a_fBig>
522	static uint32_t iemExecMemAllocatorFindReqFreeUnits(uint64_t *pbmAlloc, uint32_t c64WordsToScan, uint32_t cReqUnits)
523	{
524	/*
525	* Scan the (section of the) allocation bitmap in 64-bit words.
526	*/
527	unsigned cPrevLeadingZeros = 0;
528	for (uint32_t off = 0; off < c64WordsToScan; off++)
529	{
530	uint64_t uWord = pbmAlloc[off];
531	if (uWord == UINT64_MAX)
532	{
533	/*
534	* Getting thru patches of UINT64_MAX is a frequent problem when the allocator
535	* fills up, so it's definitely worth optimizing.
536	*
537	* The complicated code below is a bit faster on arm. Reducing the per TB cost
538	* from 4255ns to 4106ns (best run out of 10). On win/amd64 there isn't an
539	* obvious gain here, at least not with the data currently being profiled.
540	*/
541	#if 1
542	off++;
543	uint32_t cQuads = (c64WordsToScan - off) / 4;
544
545	/* Align. */
546	if (cQuads > 1)
547	switch (((uintptr_t)&pbmAlloc[off] / sizeof(uint64_t)) & 3)
548	{
549	case 0:
550	break;
551	case 1:
552	{
553	uWord = pbmAlloc[off];
554	uint64_t uWord1 = pbmAlloc[off + 1];
555	uint64_t uWord2 = pbmAlloc[off + 2];
556	if ((uWord & uWord1 & uWord2) == UINT64_MAX)
557	{
558	off += 3;
559	cQuads = (c64WordsToScan - off) / 4;
560	}
561	else if (uWord == UINT64_MAX)
562	{
563	if (uWord1 != UINT64_MAX)
564	{
565	uWord = uWord1;
566	off += 1;
567	}
568	else
569	{
570	uWord = uWord2;
571	off += 2;
572	}
573	}
574	break;
575	}
576	case 2:
577	{
578	uWord = pbmAlloc[off];
579	uint64_t uWord1 = pbmAlloc[off + 1];
580	if ((uWord & uWord1) == UINT64_MAX)
581	{
582	off += 2;
583	cQuads = (c64WordsToScan - off) / 4;
584	}
585	else if (uWord == UINT64_MAX)
586	{
587	uWord = uWord1;
588	off += 1;
589	}
590	break;
591	}
592	case 3:
593	uWord = pbmAlloc[off];
594	if (uWord == UINT64_MAX)
595	{
596	off++;
597	cQuads = (c64WordsToScan - off) / 4;
598	}
599	break;
600	}
601	if (uWord == UINT64_MAX)
602	{
603	/* Looping over 32 bytes at a time. */
604	for (;;)
605	{
606	if (cQuads-- > 0)
607	{
608	uWord = pbmAlloc[off + 0];
609	uint64_t uWord1 = pbmAlloc[off + 1];
610	uint64_t uWord2 = pbmAlloc[off + 2];
611	uint64_t uWord3 = pbmAlloc[off + 3];
612	if ((uWord & uWord1 & uWord2 & uWord3) == UINT64_MAX)
613	off += 4;
614	else
615	{
616	if (uWord != UINT64_MAX)
617	{ }
618	else if (uWord1 != UINT64_MAX)
619	{
620	uWord = uWord1;
621	off += 1;
622	}
623	else if (uWord2 != UINT64_MAX)
624	{
625	uWord = uWord2;
626	off += 2;
627	}
628	else
629	{
630	uWord = uWord3;
631	off += 3;
632	}
633	break;
634	}
635	}
636	else
637	{
638	if (off < c64WordsToScan)
639	{
640	uWord = pbmAlloc[off];
641	if (uWord != UINT64_MAX)
642	break;
643	off++;
644	if (off < c64WordsToScan)
645	{
646	uWord = pbmAlloc[off];
647	if (uWord != UINT64_MAX)
648	break;
649	off++;
650	if (off < c64WordsToScan)
651	{
652	uWord = pbmAlloc[off];
653	if (uWord != UINT64_MAX)
654	break;
655	Assert(off + 1 == c64WordsToScan);
656	}
657	}
658	}
659	return UINT32_MAX;
660	}
661	}
662	}
663	#else
664	do
665	{
666	off++;
667	if (off < c64WordsToScan)
668	uWord = pbmAlloc[off];
669	else
670	return UINT32_MAX;
671	} while (uWord == UINT64_MAX);
672	#endif
673	cPrevLeadingZeros = 0;
674	}
675
676	/*
677	* If we get down here, we have a word that isn't UINT64_MAX.
678	*/
679	if (uWord != 0)
680	{
681	/*
682	* Fend of large request we cannot satisfy before the first set bit.
683	*/
684	if (!a_fBig \|\| cReqUnits < 64 + cPrevLeadingZeros)
685	{
686	#ifdef __GNUC__
687	unsigned cZerosInWord = __builtin_popcountl(~uWord);
688	#elif defined(_MSC_VER) && defined(RT_ARCH_AMD64)
689	unsigned cZerosInWord = __popcnt64(~uWord);
690	#elif defined(_MSC_VER) && defined(RT_ARCH_ARM64)
691	unsigned cZerosInWord = _CountOneBits64(~uWord);
692	#else
693	# pragma message("need popcount intrinsic or something...")
694	unsigned cZerosInWord = 0;
695	for (uint64_t uTmp = ~uWords; uTmp; cZerosInWord++)
696	uTmp &= uTmp - 1; /* Clears the least significant bit set. */
697	#endif
698	if (cZerosInWord + cPrevLeadingZeros >= cReqUnits)
699	{
700	/* Check if we've got a patch of zeros at the trailing end
701	when joined with the previous word: */
702	#ifdef __GNUC__
703	unsigned cTrailingZeros = __builtin_ctzl(uWord);
704	#else
705	unsigned cTrailingZeros = ASMBitFirstSetU64(uWord) - 1;
706	#endif
707	if (cPrevLeadingZeros + cTrailingZeros >= cReqUnits)
708	return off * 64 - cPrevLeadingZeros;
709
710	/*
711	* Try leading zeros before we get on with the tedious stuff.
712	*/
713	#ifdef __GNUC__
714	cPrevLeadingZeros = __builtin_clzl(uWord);
715	#else
716	cPrevLeadingZeros = 64 - ASMBitLastSetU64(uWord);
717	#endif
718	if (cPrevLeadingZeros >= cReqUnits)
719	return (off + 1) * 64 - cPrevLeadingZeros;
720
721	/*
722	* Check the popcount again sans leading & trailing before looking
723	* inside the word.
724	*/
725	cZerosInWord -= cPrevLeadingZeros + cTrailingZeros;
726	if (cZerosInWord >= cReqUnits)
727	{
728	/* 1; 64 - 0 - 1 = 63; */
729	unsigned const iBitLast = 64 - cPrevLeadingZeros - cReqUnits; /** @todo boundrary */
730	unsigned iBit = cTrailingZeros;
731	uWord >>= cTrailingZeros;
732	do
733	{
734	Assert(uWord & 1);
735	#ifdef __GNUC__
736	unsigned iZeroBit = __builtin_ctzl(~uWord);
737	#else
738	unsigned iZeroBit = ASMBitFirstSetU64(~uWord) - 1;
739	#endif
740	iBit += iZeroBit;
741	uWord >>= iZeroBit;
742	Assert(iBit <= iBitLast);
743	Assert((uWord & 1) == 0);
744	#ifdef __GNUC__
745	unsigned cZeros = __builtin_ctzl(uWord);
746	#else
747	unsigned cZeros = ASMBitFirstSetU64(uWord) - 1;
748	#endif
749	if (cZeros >= cReqUnits)
750	return off * 64 + iBit;
751
752	cZerosInWord -= cZeros; /* (may underflow as we will count shifted in zeros) */
753	iBit += cZeros;
754	uWord >>= cZeros;
755	} while ((int)cZerosInWord >= (int)cReqUnits && iBit < iBitLast);
756	}
757	continue; /* we've already calculated cPrevLeadingZeros */
758	}
759	}
760
761	/* Update the leading (MSB) zero count. */
762	#ifdef __GNUC__
763	cPrevLeadingZeros = __builtin_clzl(uWord);
764	#else
765	cPrevLeadingZeros = 64 - ASMBitLastSetU64(uWord);
766	#endif
767	}
768	/*
769	* uWord == 0
770	*/
771	else
772	{
773	if RT_CONSTEXPR_IF(!a_fBig)
774	return off * 64 - cPrevLeadingZeros;
775	else /* keep else */
776	{
777	if (cPrevLeadingZeros + 64 >= cReqUnits)
778	return off * 64 - cPrevLeadingZeros;
779	for (uint32_t off2 = off + 1;; off2++)
780	{
781	if (off2 < c64WordsToScan)
782	{
783	uWord = pbmAlloc[off2];
784	if (uWord == UINT64_MAX)
785	{
786	cPrevLeadingZeros = 0;
787	break;
788	}
789	if (uWord == 0)
790	{
791	if (cPrevLeadingZeros + (off2 - off + 1) * 64 >= cReqUnits)
792	return off * 64 - cPrevLeadingZeros;
793	}
794	else
795	{
796	#ifdef __GNUC__
797	unsigned cTrailingZeros = __builtin_ctzl(uWord);
798	#else
799	unsigned cTrailingZeros = ASMBitFirstSetU64(uWord) - 1;
800	#endif
801	if (cPrevLeadingZeros + (off2 - off) * 64 + cTrailingZeros >= cReqUnits)
802	return off * 64 - cPrevLeadingZeros;
803	#ifdef __GNUC__
804	cPrevLeadingZeros = __builtin_clzl(uWord);
805	#else
806	cPrevLeadingZeros = 64 - ASMBitLastSetU64(uWord);
807	#endif
808	break;
809	}
810	}
811	else
812	return UINT32_MAX;
813	}
814	}
815	}
816	}
817	return UINT32_MAX;
818	}
819
820
821	/**
822	* Try allocate a block of @a cReqUnits in the chunk @a idxChunk.
823	*/
824	static void *
825	iemExecMemAllocatorAllocInChunkInt(PIEMEXECMEMALLOCATOR pExecMemAllocator, uint64_t *pbmAlloc, uint32_t idxFirst,
826	uint32_t cToScan, uint32_t cReqUnits, uint32_t idxChunk, PIEMTB pTb,
827	void *ppvExec, PCIEMNATIVEPERCHUNKCTX ppChunkCtx)
828	{
829	/*
830	* Shift the bitmap to the idxFirst bit so we can use ASMBitFirstClear.
831	*/
832	Assert(!(cToScan & 63));
833	Assert(!(idxFirst & 63));
834	Assert(cToScan + idxFirst <= pExecMemAllocator->cUnitsPerChunk);
835	pbmAlloc += idxFirst / 64;
836	cToScan += idxFirst & 63;
837	Assert(!(cToScan & 63));
838
839	#if 1
840	uint32_t const iBit = cReqUnits < 64
841	? iemExecMemAllocatorFindReqFreeUnits<false>(pbmAlloc, cToScan / 64, cReqUnits)
842	: iemExecMemAllocatorFindReqFreeUnits<true>( pbmAlloc, cToScan / 64, cReqUnits);
843	# ifdef VBOX_STRICT
844	uint32_t const iBitOld = iemExecMemAllocatorFindReqFreeUnitsOld(pbmAlloc, cToScan, cReqUnits);
845	AssertMsg( iBit == iBitOld
846	\|\| (iBit / 64) == (iBitOld / 64), /* New algorithm will return trailing hit before middle. */
847	("iBit=%#x (%#018RX64); iBitOld=%#x (%#018RX64); cReqUnits=%#x\n",
848	iBit, iBit != UINT32_MAX ? pbmAlloc[iBit / 64] : 0,
849	iBitOld, iBitOld != UINT32_MAX ? pbmAlloc[iBitOld / 64] : 0, cReqUnits));
850	# endif
851	#else
852	uint32_t const iBit = iemExecMemAllocatorFindReqFreeUnitsOld(pbmAlloc, cToScan, cReqUnits);
853	#endif
854	if (iBit != UINT32_MAX)
855	{
856	ASMBitSetRange(pbmAlloc, (uint32_t)iBit, (uint32_t)iBit + cReqUnits);
857
858	PIEMEXECMEMCHUNK const pChunk = &pExecMemAllocator->aChunks[idxChunk];
859	pChunk->cFreeUnits -= cReqUnits;
860	pChunk->idxFreeHint = (uint32_t)iBit + cReqUnits;
861
862	pExecMemAllocator->cAllocations += 1;
863	uint32_t const cbReq = cReqUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
864	pExecMemAllocator->cbAllocated += cbReq;
865	pExecMemAllocator->cbFree -= cbReq;
866	pExecMemAllocator->idxChunkHint = idxChunk;
867
868	void * const pvMemRw = (uint8_t *)pChunk->pvChunkRw
869	+ ((idxFirst + (uint32_t)iBit) << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT);
870
871	if (ppChunkCtx)
872	*ppChunkCtx = pChunk->pCtx;
873
874	/*
875	* Initialize the header and return.
876	*/
877	# ifdef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
878	PIEMEXECMEMALLOCHDR const pHdr = (PIEMEXECMEMALLOCHDR)pvMemRw;
879	pHdr->uMagic = IEMEXECMEMALLOCHDR_MAGIC;
880	pHdr->idxChunk = idxChunk;
881	pHdr->pTb = pTb;
882
883	if (ppvExec)
884	ppvExec = (uint8_t )pChunk->pvChunkRx
885	+ ((idxFirst + (uint32_t)iBit) << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT)
886	+ sizeof(*pHdr);
887
888	return pHdr + 1;
889	#else
890	if (ppvExec)
891	ppvExec = (uint8_t )pChunk->pvChunkRx
892	+ ((idxFirst + (uint32_t)iBit) << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT);
893
894	RT_NOREF(pTb);
895	return pvMem;
896	#endif
897	}
898
899	return NULL;
900	}
901
902
903	/**
904	* Converts requested number of bytes into a unit count.
905	*/
906	DECL_FORCE_INLINE(uint32_t) iemExecMemAllocBytesToUnits(uint32_t cbReq)
907	{
908	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
909	return (cbReq + sizeof(IEMEXECMEMALLOCHDR) + IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1)
910	#else
911	return (cbReq + IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1)
912	#endif
913	>> IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
914	}
915
916
917	DECL_FORCE_INLINE(PIEMNATIVEINSTR)
918	iemExecMemAllocatorAllocUnitsInChunkInner(PIEMEXECMEMALLOCATOR pExecMemAllocator, uint32_t idxChunk, uint32_t cReqUnits,
919	PIEMTB pTb, PIEMNATIVEINSTR ppaExec, PCIEMNATIVEPERCHUNKCTX ppChunkCtx)
920	{
921	uint64_t * const pbmAlloc = &pExecMemAllocator->pbmAlloc[pExecMemAllocator->cBitmapElementsPerChunk * idxChunk];
922	uint32_t const idxHint = pExecMemAllocator->aChunks[idxChunk].idxFreeHint & ~(uint32_t)63;
923	if (idxHint + cReqUnits <= pExecMemAllocator->cUnitsPerChunk)
924	{
925	void *pvRet = iemExecMemAllocatorAllocInChunkInt(pExecMemAllocator, pbmAlloc, idxHint,
926	pExecMemAllocator->cUnitsPerChunk - idxHint,
927	cReqUnits, idxChunk, pTb, (void **)ppaExec, ppChunkCtx);
928	if (pvRet)
929	return (PIEMNATIVEINSTR)pvRet;
930	}
931	void *pvRet = iemExecMemAllocatorAllocInChunkInt(pExecMemAllocator, pbmAlloc, 0,
932	RT_MIN(pExecMemAllocator->cUnitsPerChunk,
933	RT_ALIGN_32(idxHint + cReqUnits, 64*4)),
934	cReqUnits, idxChunk, pTb, (void **)ppaExec, ppChunkCtx);
935	if (pvRet)
936	return (PIEMNATIVEINSTR)pvRet;
937
938	pExecMemAllocator->cFruitlessChunkScans += 1;
939	return NULL;
940	}
941
942
943	DECLINLINE(PIEMNATIVEINSTR)
944	iemExecMemAllocatorAllocBytesInChunk(PIEMEXECMEMALLOCATOR pExecMemAllocator, uint32_t idxChunk, uint32_t cbReq,
945	PIEMNATIVEINSTR *ppaExec)
946	{
947	uint32_t const cReqUnits = iemExecMemAllocBytesToUnits(cbReq);
948	if (cReqUnits <= pExecMemAllocator->aChunks[idxChunk].cFreeUnits)
949	return iemExecMemAllocatorAllocUnitsInChunkInner(pExecMemAllocator, idxChunk, cReqUnits, NULL /pTb/,
950	ppaExec, NULL /ppChunkCtx/);
951	return NULL;
952	}
953
954
955	/**
956	* Allocates @a cbReq bytes of executable memory.
957	*
958	* @returns Pointer to the readable/writeable memory, NULL if out of memory or other problem
959	* encountered.
960	* @param pVCpu The cross context virtual CPU structure of the
961	* calling thread.
962	* @param cbReq How many bytes are required.
963	* @param pTb The translation block that will be using the allocation.
964	* @param ppaExec Where to return the pointer to executable view of
965	* the allocated memory, optional.
966	* @param ppChunkCtx Where to return the per chunk attached context
967	* if available, optional.
968	*/
969	DECLHIDDEN(PIEMNATIVEINSTR) iemExecMemAllocatorAlloc(PVMCPU pVCpu, uint32_t cbReq, PIEMTB pTb,
970	PIEMNATIVEINSTR ppaExec, PCIEMNATIVEPERCHUNKCTX ppChunkCtx) RT_NOEXCEPT
971	{
972	PIEMEXECMEMALLOCATOR pExecMemAllocator = pVCpu->iem.s.pExecMemAllocatorR3;
973	AssertReturn(pExecMemAllocator && pExecMemAllocator->uMagic == IEMEXECMEMALLOCATOR_MAGIC, NULL);
974	AssertMsgReturn(cbReq > 32 && cbReq < _512K, ("%#x\n", cbReq), NULL);
975	STAM_PROFILE_START(&pExecMemAllocator->StatAlloc, a);
976
977	uint32_t const cReqUnits = iemExecMemAllocBytesToUnits(cbReq);
978	STAM_COUNTER_INC(&pExecMemAllocator->aStatSizes[cReqUnits < RT_ELEMENTS(pExecMemAllocator->aStatSizes) ? cReqUnits : 0]);
979	for (unsigned iIteration = 0;; iIteration++)
980	{
981	if ( cbReq * 2 <= pExecMemAllocator->cbFree
982	\|\| (cReqUnits == 1 \|\| pExecMemAllocator->cbFree >= IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE) )
983	{
984	uint32_t const cChunks = pExecMemAllocator->cChunks;
985	uint32_t const idxChunkHint = pExecMemAllocator->idxChunkHint < cChunks ? pExecMemAllocator->idxChunkHint : 0;
986
987	/*
988	* We do two passes here, the first pass we skip chunks with fewer than cReqUnits * 16,
989	* the 2nd pass we skip chunks. The second pass checks the one skipped in the first pass.
990	*/
991	for (uint32_t cMinFreePass = cReqUnits == 1 ? cReqUnits : cReqUnits * 16, cMaxFreePass = UINT32_MAX;;)
992	{
993	for (uint32_t idxChunk = idxChunkHint; idxChunk < cChunks; idxChunk++)
994	if ( pExecMemAllocator->aChunks[idxChunk].cFreeUnits >= cMinFreePass
995	&& pExecMemAllocator->aChunks[idxChunk].cFreeUnits <= cMaxFreePass)
996	{
997	PIEMNATIVEINSTR const pRet = iemExecMemAllocatorAllocUnitsInChunkInner(pExecMemAllocator, idxChunk,
998	cReqUnits, pTb, ppaExec, ppChunkCtx);
999	if (pRet)
1000	{
1001	STAM_PROFILE_STOP(&pExecMemAllocator->StatAlloc, a);
1002	#ifdef VBOX_WITH_STATISTICS
1003	pExecMemAllocator->cbUnusable += (cReqUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT) - cbReq;
1004	#endif
1005	return pRet;
1006	}
1007	}
1008	for (uint32_t idxChunk = 0; idxChunk < idxChunkHint; idxChunk++)
1009	if ( pExecMemAllocator->aChunks[idxChunk].cFreeUnits >= cMinFreePass
1010	&& pExecMemAllocator->aChunks[idxChunk].cFreeUnits <= cMaxFreePass)
1011	{
1012	PIEMNATIVEINSTR const pRet = iemExecMemAllocatorAllocUnitsInChunkInner(pExecMemAllocator, idxChunk,
1013	cReqUnits, pTb, ppaExec, ppChunkCtx);
1014	if (pRet)
1015	{
1016	STAM_PROFILE_STOP(&pExecMemAllocator->StatAlloc, a);
1017	#ifdef VBOX_WITH_STATISTICS
1018	pExecMemAllocator->cbUnusable += (cReqUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT) - cbReq;
1019	#endif
1020	return pRet;
1021	}
1022	}
1023	if (cMinFreePass <= cReqUnits * 2)
1024	break;
1025	cMaxFreePass = cMinFreePass - 1;
1026	cMinFreePass = cReqUnits * 2;
1027	}
1028	}
1029
1030	/*
1031	* Can we grow it with another chunk?
1032	*/
1033	if (pExecMemAllocator->cChunks < pExecMemAllocator->cMaxChunks)
1034	{
1035	int rc = iemExecMemAllocatorGrow(pVCpu, pExecMemAllocator);
1036	AssertLogRelRCReturn(rc, NULL);
1037
1038	uint32_t const idxChunk = pExecMemAllocator->cChunks - 1;
1039	PIEMNATIVEINSTR const pRet = iemExecMemAllocatorAllocUnitsInChunkInner(pExecMemAllocator, idxChunk, cReqUnits, pTb,
1040	ppaExec, ppChunkCtx);
1041	if (pRet)
1042	{
1043	STAM_PROFILE_STOP(&pExecMemAllocator->StatAlloc, a);
1044	#ifdef VBOX_WITH_STATISTICS
1045	pExecMemAllocator->cbUnusable += (cReqUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT) - cbReq;
1046	#endif
1047	return pRet;
1048	}
1049	AssertFailed();
1050	}
1051
1052	/*
1053	* Try prune native TBs once.
1054	*/
1055	if (iIteration == 0)
1056	{
1057	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING
1058	iemExecMemAllocatorPrune(pVCpu, pExecMemAllocator);
1059	#else
1060	/* No header included in the instruction count here. */
1061	uint32_t const cNeededInstrs = RT_ALIGN_32(cbReq, IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE) / sizeof(IEMNATIVEINSTR);
1062	iemTbAllocatorFreeupNativeSpace(pVCpu, cNeededInstrs);
1063	#endif
1064	}
1065	else
1066	{
1067	STAM_REL_COUNTER_INC(&pVCpu->iem.s.StatNativeExecMemInstrBufAllocFailed);
1068	STAM_PROFILE_STOP(&pExecMemAllocator->StatAlloc, a);
1069	return NULL;
1070	}
1071	}
1072	}
1073
1074
1075	/** This is a hook to ensure the instruction cache is properly flushed before the code in the memory
1076	* given by @a pv and @a cb is executed */
1077	DECLHIDDEN(void) iemExecMemAllocatorReadyForUse(PVMCPUCC pVCpu, void *pv, size_t cb) RT_NOEXCEPT
1078	{
1079	#ifdef RT_OS_DARWIN
1080	/*
1081	* We need to synchronize the stuff we wrote to the data cache with the
1082	* instruction cache, since these aren't coherent on arm (or at least not
1083	* on Apple Mn CPUs).
1084	*
1085	* Note! Since we don't any share JIT'ed code with the other CPUs, we don't
1086	* really care whether the dcache is fully flushed back to memory. It
1087	* only needs to hit the level 2 cache, which the level 1 instruction
1088	* and data caches seems to be sharing. In ARM terms, we need to reach
1089	* a point of unification (PoU), rather than a point of coherhency (PoC).
1090	*
1091	* https://developer.apple.com/documentation/apple-silicon/porting-just-in-time-compilers-to-apple-silicon
1092	*
1093	* https://developer.arm.com/documentation/den0013/d/Caches/Point-of-coherency-and-unification
1094	*
1095	* Experimenting with the approach used by sys_icache_invalidate() and
1096	* tweaking it a little, could let us shave off a bit of effort. The thing
1097	* that slows the apple code down on an M2 (runing Sonoma 13.4), seems to
1098	* the 'DSB ISH' instructions performed every 20 icache line flushes.
1099	* Skipping these saves ~100ns or more per TB when profiling the native
1100	* recompiler on the TBs from a win11 full boot-desktop-shutdow sequence.
1101	* Thus we will leave DCACHE_ICACHE_SYNC_WITH_WITH_IVAU_DSB undefined if we
1102	* can.
1103	*
1104	* There appears not to be much difference between DSB options 'ISH',
1105	* 'ISHST', 'NSH' and 'NSHST'. The latter is theoretically all we need, so
1106	* we'll use that one.
1107	*
1108	* See https://developer.arm.com/documentation/100941/0101/Barriers for
1109	* details on the barrier options.
1110	*
1111	* Note! The CFG value "/IEM/HostICacheInvalidationViaHostAPI" can be used
1112	* to disabling the experimental code should it misbehave.
1113	*/
1114	uint8_t const fHostICacheInvalidation = pVCpu->iem.s.fHostICacheInvalidation;
1115	if (!(fHostICacheInvalidation & IEMNATIVE_ICACHE_F_USE_HOST_API))
1116	{
1117	# define DCACHE_ICACHE_SYNC_DSB_OPTION "nshst"
1118	/# define DCACHE_ICACHE_SYNC_WITH_WITH_IVAU_DSB/
1119
1120	/* Skipping this is fine, but doesn't impact perf much. */
1121	__asm__ __volatile__("dsb " DCACHE_ICACHE_SYNC_DSB_OPTION);
1122
1123	/* Invalidate the icache for the range [pv,pv+cb). */
1124	# ifdef DCACHE_ICACHE_SYNC_WITH_WITH_IVAU_DSB
1125	size_t const cIvauDsbEvery= 20;
1126	unsigned cDsb = cIvauDsbEvery;
1127	# endif
1128	size_t const cbCacheLine = 64;
1129	size_t cbInvalidate = cb + ((uintptr_t)pv & (cbCacheLine - 1)) ;
1130	size_t cCacheLines = RT_ALIGN_Z(cbInvalidate, cbCacheLine) / cbCacheLine;
1131	uintptr_t uPtr = (uintptr_t)pv & ~(uintptr_t)(cbCacheLine - 1);
1132	for (;; uPtr += cbCacheLine)
1133	{
1134	__asm__ /__volatile__/("ic ivau, %0" : : "r" (uPtr));
1135	cCacheLines -= 1;
1136	if (!cCacheLines)
1137	break;
1138	# ifdef DCACHE_ICACHE_SYNC_WITH_WITH_IVAU_DSB
1139	cDsb -= 1;
1140	if (cDsb != 0)
1141	{ /* likely */ }
1142	else
1143	{
1144	__asm__ __volatile__("dsb " DCACHE_ICACHE_SYNC_DSB_OPTION);
1145	cDsb = cIvauDsbEvery;
1146	}
1147	# endif
1148	}
1149
1150	/*
1151	* The DSB here is non-optional it seems.
1152	*
1153	* The following ISB can be omitted on M2 without any obvious sideeffects,
1154	* it produces better number in the above mention profiling scenario.
1155	* This could be related to the kHasICDSB flag in cpu_capabilities.h,
1156	* but it doesn't look like that flag is set here (M2, Sonoma 13.4).
1157	*
1158	* I've made the inclusion of the ISH barrier as configurable and with
1159	* a default of skipping it.
1160	*/
1161	if (!(fHostICacheInvalidation & IEMNATIVE_ICACHE_F_END_WITH_ISH))
1162	__asm__ __volatile__("dsb " DCACHE_ICACHE_SYNC_DSB_OPTION
1163	::: "memory");
1164	else
1165	__asm__ __volatile__("dsb " DCACHE_ICACHE_SYNC_DSB_OPTION "\n\t"
1166	"isb"
1167	::: "memory");
1168	}
1169	else
1170	sys_icache_invalidate(pv, cb);
1171
1172	#elif defined(RT_OS_LINUX) && defined(RT_ARCH_ARM64)
1173	RT_NOREF(pVCpu);
1174
1175	/* There is __builtin___clear_cache() but it flushes both the instruction and data cache, so do it manually. */
1176	static uint32_t s_u32CtrEl0 = 0;
1177	if (!s_u32CtrEl0)
1178	asm volatile ("mrs %0, ctr_el0":"=r" (s_u32CtrEl0));
1179	uintptr_t cbICacheLine = (uintptr_t)4 << (s_u32CtrEl0 & 0xf);
1180
1181	uintptr_t pb = (uintptr_t)pv & ~(cbICacheLine - 1);
1182	for (; pb < (uintptr_t)pv + cb; pb += cbICacheLine)
1183	asm volatile ("ic ivau, %0" : : "r" (pb) : "memory");
1184
1185	asm volatile ("dsb ish\n\t isb\n\t" : : : "memory");
1186
1187	#else
1188	RT_NOREF(pVCpu, pv, cb);
1189	#endif
1190	}
1191
1192
1193	/**
1194	* Frees executable memory.
1195	*/
1196	DECLHIDDEN(void) iemExecMemAllocatorFree(PVMCPU pVCpu, void *pv, size_t cb) RT_NOEXCEPT
1197	{
1198	PIEMEXECMEMALLOCATOR pExecMemAllocator = pVCpu->iem.s.pExecMemAllocatorR3;
1199	Assert(pExecMemAllocator && pExecMemAllocator->uMagic == IEMEXECMEMALLOCATOR_MAGIC);
1200	AssertPtr(pv);
1201	#ifdef VBOX_WITH_STATISTICS
1202	size_t const cbOrig = cb;
1203	#endif
1204	#ifndef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
1205	Assert(!((uintptr_t)pv & (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1)));
1206
1207	/* Align the size as we did when allocating the block. */
1208	cb = RT_ALIGN_Z(cb, IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE);
1209
1210	#else
1211	PIEMEXECMEMALLOCHDR pHdr = (PIEMEXECMEMALLOCHDR)pv - 1;
1212	Assert(!((uintptr_t)pHdr & (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1)));
1213	AssertReturnVoid(pHdr->uMagic == IEMEXECMEMALLOCHDR_MAGIC);
1214	uint32_t const idxChunk = pHdr->idxChunk;
1215	AssertReturnVoid(idxChunk < pExecMemAllocator->cChunks);
1216	pv = pHdr;
1217
1218	/* Adjust and align the size to cover the whole allocation area. */
1219	cb = RT_ALIGN_Z(cb + sizeof(*pHdr), IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE);
1220	#endif
1221
1222	/* Free it / assert sanity. */
1223	bool fFound = false;
1224	uint32_t const cbChunk = pExecMemAllocator->cbChunk;
1225	#ifndef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
1226	uint32_t const cChunks = pExecMemAllocator->cChunks;
1227	for (uint32_t idxChunk = 0; idxChunk < cChunks; idxChunk++)
1228	#endif
1229	{
1230	uintptr_t const offChunk = (uintptr_t)pv - (uintptr_t)pExecMemAllocator->aChunks[idxChunk].pvChunkRx;
1231	fFound = offChunk < cbChunk;
1232	if (fFound)
1233	{
1234	uint32_t const idxFirst = (uint32_t)offChunk >> IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
1235	uint32_t const cReqUnits = (uint32_t)cb >> IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
1236
1237	/* Check that it's valid and free it. */
1238	uint64_t * const pbmAlloc = &pExecMemAllocator->pbmAlloc[pExecMemAllocator->cBitmapElementsPerChunk * idxChunk];
1239	AssertReturnVoid(ASMBitTest(pbmAlloc, idxFirst));
1240	for (uint32_t i = 1; i < cReqUnits; i++)
1241	AssertReturnVoid(ASMBitTest(pbmAlloc, idxFirst + i));
1242	ASMBitClearRange(pbmAlloc, idxFirst, idxFirst + cReqUnits);
1243
1244	/* Invalidate the header using the writeable memory view. */
1245	pHdr = (PIEMEXECMEMALLOCHDR)((uintptr_t)pExecMemAllocator->aChunks[idxChunk].pvChunkRw + offChunk);
1246	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
1247	pHdr->uMagic = 0;
1248	pHdr->idxChunk = 0;
1249	pHdr->pTb = NULL;
1250	#endif
1251	pExecMemAllocator->aChunks[idxChunk].cFreeUnits += cReqUnits;
1252	pExecMemAllocator->aChunks[idxChunk].idxFreeHint = idxFirst;
1253
1254	/* Update the stats. */
1255	pExecMemAllocator->cbAllocated -= cb;
1256	pExecMemAllocator->cbFree += cb;
1257	pExecMemAllocator->cAllocations -= 1;
1258	#ifdef VBOX_WITH_STATISTICS
1259	pExecMemAllocator->cbUnusable -= (cReqUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT) - cbOrig;
1260	#endif
1261	return;
1262	}
1263	}
1264	AssertFailed();
1265	}
1266
1267
1268	/**
1269	* Interface used by iemNativeRecompileAttachExecMemChunkCtx and unwind info
1270	* generators.
1271	*/
1272	DECLHIDDEN(PIEMNATIVEINSTR)
1273	iemExecMemAllocatorAllocFromChunk(PVMCPU pVCpu, uint32_t idxChunk, uint32_t cbReq, PIEMNATIVEINSTR *ppaExec)
1274	{
1275	PIEMEXECMEMALLOCATOR pExecMemAllocator = pVCpu->iem.s.pExecMemAllocatorR3;
1276	AssertReturn(idxChunk < pExecMemAllocator->cChunks, NULL);
1277	Assert(cbReq < _1M);
1278	return iemExecMemAllocatorAllocBytesInChunk(pExecMemAllocator, idxChunk, cbReq, ppaExec);
1279	}
1280
1281
1282	/**
1283	* For getting the per-chunk context detailing common code for a TB.
1284	*
1285	* This is for use by the disassembler.
1286	*/
1287	DECLHIDDEN(PCIEMNATIVEPERCHUNKCTX) iemExecMemGetTbChunkCtx(PVMCPU pVCpu, PCIEMTB pTb)
1288	{
1289	PIEMEXECMEMALLOCATOR pExecMemAllocator = pVCpu->iem.s.pExecMemAllocatorR3;
1290	if ((pTb->fFlags & IEMTB_F_TYPE_MASK) == IEMTB_F_TYPE_NATIVE)
1291	{
1292	uintptr_t const uAddress = (uintptr_t)pTb->Native.paInstructions;
1293	uint32_t const cbChunk = pExecMemAllocator->cbChunk;
1294	uint32_t idxChunk = pExecMemAllocator->cChunks;
1295	while (idxChunk-- > 0)
1296	if (uAddress - (uintptr_t)pExecMemAllocator->aChunks[idxChunk].pvChunkRx < cbChunk)
1297	return pExecMemAllocator->aChunks[idxChunk].pCtx;
1298	}
1299	return NULL;
1300	}
1301
1302
1303	#ifdef IN_RING3
1304	# ifdef RT_OS_WINDOWS
1305
1306	/**
1307	* Initializes the unwind info structures for windows hosts.
1308	*/
1309	static int
1310	iemExecMemAllocatorInitAndRegisterUnwindInfoForChunk(PVMCPUCC pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator,
1311	void *pvChunk, uint32_t idxChunk)
1312	{
1313	RT_NOREF(pVCpu);
1314
1315	/*
1316	* The AMD64 unwind opcodes.
1317	*
1318	* This is a program that starts with RSP after a RET instruction that
1319	* ends up in recompiled code, and the operations we describe here will
1320	* restore all non-volatile registers and bring RSP back to where our
1321	* RET address is. This means it's reverse order from what happens in
1322	* the prologue.
1323	*
1324	* Note! Using a frame register approach here both because we have one
1325	* and but mainly because the UWOP_ALLOC_LARGE argument values
1326	* would be a pain to write initializers for. On the positive
1327	* side, we're impervious to changes in the the stack variable
1328	* area can can deal with dynamic stack allocations if necessary.
1329	*/
1330	static const IMAGE_UNWIND_CODE s_aOpcodes[] =
1331	{
1332	{ { 16, IMAGE_AMD64_UWOP_SET_FPREG, 0 } }, /* RSP = RBP - FrameOffset * 10 (0x60) */
1333	{ { 16, IMAGE_AMD64_UWOP_ALLOC_SMALL, 0 } }, /* RSP += 8; */
1334	{ { 14, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_x15 } }, /* R15 = [RSP]; RSP += 8; */
1335	{ { 12, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_x14 } }, /* R14 = [RSP]; RSP += 8; */
1336	{ { 10, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_x13 } }, /* R13 = [RSP]; RSP += 8; */
1337	{ { 8, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_x12 } }, /* R12 = [RSP]; RSP += 8; */
1338	{ { 7, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_xDI } }, /* RDI = [RSP]; RSP += 8; */
1339	{ { 6, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_xSI } }, /* RSI = [RSP]; RSP += 8; */
1340	{ { 5, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_xBX } }, /* RBX = [RSP]; RSP += 8; */
1341	{ { 4, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_xBP } }, /* RBP = [RSP]; RSP += 8; */
1342	};
1343	union
1344	{
1345	IMAGE_UNWIND_INFO Info;
1346	uint8_t abPadding[RT_UOFFSETOF(IMAGE_UNWIND_INFO, aOpcodes) + 16];
1347	} s_UnwindInfo =
1348	{
1349	{
1350	/* .Version = */ 1,
1351	/* .Flags = */ 0,
1352	/* .SizeOfProlog = / 16, / whatever */
1353	/* .CountOfCodes = */ RT_ELEMENTS(s_aOpcodes),
1354	/* .FrameRegister = */ X86_GREG_xBP,
1355	/* .FrameOffset = / (-IEMNATIVE_FP_OFF_LAST_PUSH + 8) / 16 / we're off by one slot. sigh. */,
1356	}
1357	};
1358	AssertCompile(-IEMNATIVE_FP_OFF_LAST_PUSH < 240 && -IEMNATIVE_FP_OFF_LAST_PUSH > 0);
1359	AssertCompile((-IEMNATIVE_FP_OFF_LAST_PUSH & 0xf) == 8);
1360
1361	/*
1362	* Calc how much space we need and allocate it off the exec heap.
1363	*/
1364	unsigned const cFunctionEntries = 1;
1365	unsigned const cbUnwindInfo = sizeof(s_aOpcodes) + RT_UOFFSETOF(IMAGE_UNWIND_INFO, aOpcodes);
1366	unsigned const cbNeeded = sizeof(IMAGE_RUNTIME_FUNCTION_ENTRY) * cFunctionEntries + cbUnwindInfo;
1367	PIMAGE_RUNTIME_FUNCTION_ENTRY const paFunctions
1368	= (PIMAGE_RUNTIME_FUNCTION_ENTRY)iemExecMemAllocatorAllocBytesInChunk(pExecMemAllocator, idxChunk, cbNeeded, NULL);
1369	AssertReturn(paFunctions, VERR_INTERNAL_ERROR_5);
1370	pExecMemAllocator->aChunks[idxChunk].pvUnwindInfo = paFunctions;
1371
1372	/*
1373	* Initialize the structures.
1374	*/
1375	PIMAGE_UNWIND_INFO const pInfo = (PIMAGE_UNWIND_INFO)&paFunctions[cFunctionEntries];
1376
1377	paFunctions[0].BeginAddress = 0;
1378	paFunctions[0].EndAddress = pExecMemAllocator->cbChunk;
1379	paFunctions[0].UnwindInfoAddress = (uint32_t)((uintptr_t)pInfo - (uintptr_t)pvChunk);
1380
1381	memcpy(pInfo, &s_UnwindInfo, RT_UOFFSETOF(IMAGE_UNWIND_INFO, aOpcodes));
1382	memcpy(&pInfo->aOpcodes[0], s_aOpcodes, sizeof(s_aOpcodes));
1383
1384	/*
1385	* Register it.
1386	*/
1387	uint8_t fRet = RtlAddFunctionTable(paFunctions, cFunctionEntries, (uintptr_t)pvChunk);
1388	AssertReturn(fRet, VERR_INTERNAL_ERROR_3); /* Nothing to clean up on failure, since its within the chunk itself. */
1389
1390	return VINF_SUCCESS;
1391	}
1392
1393
1394	# else /* !RT_OS_WINDOWS */
1395
1396	/**
1397	* Emits a LEB128 encoded value between -0x2000 and 0x2000 (both exclusive).
1398	*/
1399	DECLINLINE(RTPTRUNION) iemDwarfPutLeb128(RTPTRUNION Ptr, int32_t iValue)
1400	{
1401	if (iValue >= 64)
1402	{
1403	Assert(iValue < 0x2000);
1404	*Ptr.pb++ = ((uint8_t)iValue & 0x7f) \| 0x80;
1405	*Ptr.pb++ = (uint8_t)(iValue >> 7) & 0x3f;
1406	}
1407	else if (iValue >= 0)
1408	*Ptr.pb++ = (uint8_t)iValue;
1409	else if (iValue > -64)
1410	*Ptr.pb++ = ((uint8_t)iValue & 0x3f) \| 0x40;
1411	else
1412	{
1413	Assert(iValue > -0x2000);
1414	*Ptr.pb++ = ((uint8_t)iValue & 0x7f) \| 0x80;
1415	*Ptr.pb++ = ((uint8_t)(iValue >> 7) & 0x3f) \| 0x40;
1416	}
1417	return Ptr;
1418	}
1419
1420
1421	/**
1422	* Emits an ULEB128 encoded value (up to 64-bit wide).
1423	*/
1424	DECLINLINE(RTPTRUNION) iemDwarfPutUleb128(RTPTRUNION Ptr, uint64_t uValue)
1425	{
1426	while (uValue >= 0x80)
1427	{
1428	*Ptr.pb++ = ((uint8_t)uValue & 0x7f) \| 0x80;
1429	uValue >>= 7;
1430	}
1431	*Ptr.pb++ = (uint8_t)uValue;
1432	return Ptr;
1433	}
1434
1435
1436	/**
1437	* Emits a CFA rule as register @a uReg + offset @a off.
1438	*/
1439	DECLINLINE(RTPTRUNION) iemDwarfPutCfaDefCfa(RTPTRUNION Ptr, uint32_t uReg, uint32_t off)
1440	{
1441	*Ptr.pb++ = DW_CFA_def_cfa;
1442	Ptr = iemDwarfPutUleb128(Ptr, uReg);
1443	Ptr = iemDwarfPutUleb128(Ptr, off);
1444	return Ptr;
1445	}
1446
1447
1448	/**
1449	* Emits a register (@a uReg) save location:
1450	* CFA + @a off * data_alignment_factor
1451	*/
1452	DECLINLINE(RTPTRUNION) iemDwarfPutCfaOffset(RTPTRUNION Ptr, uint32_t uReg, uint32_t off)
1453	{
1454	if (uReg < 0x40)
1455	*Ptr.pb++ = DW_CFA_offset \| uReg;
1456	else
1457	{
1458	*Ptr.pb++ = DW_CFA_offset_extended;
1459	Ptr = iemDwarfPutUleb128(Ptr, uReg);
1460	}
1461	Ptr = iemDwarfPutUleb128(Ptr, off);
1462	return Ptr;
1463	}
1464
1465
1466	# if 0 /* unused */
1467	/**
1468	* Emits a register (@a uReg) save location, using signed offset:
1469	* CFA + @a offSigned * data_alignment_factor
1470	*/
1471	DECLINLINE(RTPTRUNION) iemDwarfPutCfaSignedOffset(RTPTRUNION Ptr, uint32_t uReg, int32_t offSigned)
1472	{
1473	*Ptr.pb++ = DW_CFA_offset_extended_sf;
1474	Ptr = iemDwarfPutUleb128(Ptr, uReg);
1475	Ptr = iemDwarfPutLeb128(Ptr, offSigned);
1476	return Ptr;
1477	}
1478	# endif
1479
1480
1481	/**
1482	* Initializes the unwind info section for non-windows hosts.
1483	*/
1484	static int
1485	iemExecMemAllocatorInitAndRegisterUnwindInfoForChunk(PVMCPUCC pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator,
1486	void *pvChunk, uint32_t idxChunk)
1487	{
1488	PIEMEXECMEMCHUNKEHFRAME const pEhFrame = &pExecMemAllocator->paEhFrames[idxChunk];
1489	pExecMemAllocator->aChunks[idxChunk].pvUnwindInfo = pEhFrame; /* not necessary, but whatever */
1490
1491	RTPTRUNION Ptr = { pEhFrame->abEhFrame };
1492
1493	/*
1494	* Generate the CIE first.
1495	*/
1496	# ifdef IEMNATIVE_USE_LIBUNWIND /* libunwind (llvm, darwin) only supports v1 and v3. */
1497	uint8_t const iDwarfVer = 3;
1498	# else
1499	uint8_t const iDwarfVer = 4;
1500	# endif
1501	RTPTRUNION const PtrCie = Ptr;
1502	Ptr.pu32++ = 123; / The CIE length will be determined later. */
1503	Ptr.pu32++ = 0 /UINT32_MAX/; / I'm a CIE in .eh_frame speak. */
1504	Ptr.pb++ = iDwarfVer; / DwARF version */
1505	Ptr.pb++ = 0; / Augmentation. */
1506	if (iDwarfVer >= 4)
1507	{
1508	Ptr.pb++ = sizeof(uintptr_t); / Address size. */
1509	Ptr.pb++ = 0; / Segment selector size. */
1510	}
1511	# ifdef RT_ARCH_AMD64
1512	Ptr = iemDwarfPutLeb128(Ptr, 1); /* Code alignment factor (LEB128 = 1). */
1513	# else
1514	Ptr = iemDwarfPutLeb128(Ptr, 4); /* Code alignment factor (LEB128 = 4). */
1515	# endif
1516	Ptr = iemDwarfPutLeb128(Ptr, -8); /* Data alignment factor (LEB128 = -8). */
1517	# ifdef RT_ARCH_AMD64
1518	Ptr = iemDwarfPutUleb128(Ptr, DWREG_AMD64_RA); /* Return address column (ULEB128) */
1519	# elif defined(RT_ARCH_ARM64)
1520	Ptr = iemDwarfPutUleb128(Ptr, DWREG_ARM64_LR); /* Return address column (ULEB128) */
1521	# else
1522	# error "port me"
1523	# endif
1524	/* Initial instructions: */
1525	# ifdef RT_ARCH_AMD64
1526	Ptr = iemDwarfPutCfaDefCfa(Ptr, DWREG_AMD64_RBP, 16); /* CFA = RBP + 0x10 - first stack parameter */
1527	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_RA, 1); /* Ret RIP = [CFA + 1-8] /
1528	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_RBP, 2); /* RBP = [CFA + 2-8] /
1529	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_RBX, 3); /* RBX = [CFA + 3-8] /
1530	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_R12, 4); /* R12 = [CFA + 4-8] /
1531	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_R13, 5); /* R13 = [CFA + 5-8] /
1532	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_R14, 6); /* R14 = [CFA + 6-8] /
1533	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_R15, 7); /* R15 = [CFA + 7-8] /
1534	# elif defined(RT_ARCH_ARM64)
1535	# if 1
1536	Ptr = iemDwarfPutCfaDefCfa(Ptr, DWREG_ARM64_BP, 16); /* CFA = BP + 0x10 - first stack parameter */
1537	# else
1538	Ptr = iemDwarfPutCfaDefCfa(Ptr, DWREG_ARM64_SP, IEMNATIVE_FRAME_VAR_SIZE + IEMNATIVE_FRAME_SAVE_REG_SIZE);
1539	# endif
1540	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_LR, 1); /* Ret PC = [CFA + 1-8] /
1541	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_BP, 2); /* Ret BP = [CFA + 2-8] /
1542	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X28, 3); /* X28 = [CFA + 3-8] /
1543	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X27, 4); /* X27 = [CFA + 4-8] /
1544	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X26, 5); /* X26 = [CFA + 5-8] /
1545	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X25, 6); /* X25 = [CFA + 6-8] /
1546	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X24, 7); /* X24 = [CFA + 7-8] /
1547	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X23, 8); /* X23 = [CFA + 8-8] /
1548	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X22, 9); /* X22 = [CFA + 9-8] /
1549	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X21, 10); /* X21 = [CFA +10-8] /
1550	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X20, 11); /* X20 = [CFA +11-8] /
1551	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X19, 12); /* X19 = [CFA +12-8] /
1552	AssertCompile(IEMNATIVE_FRAME_SAVE_REG_SIZE / 8 == 12);
1553	/** @todo we we need to do something about clearing DWREG_ARM64_RA_SIGN_STATE or something? */
1554	# else
1555	# error "port me"
1556	# endif
1557	while ((Ptr.u - PtrCie.u) & 3)
1558	*Ptr.pb++ = DW_CFA_nop;
1559	/* Finalize the CIE size. */
1560	*PtrCie.pu32 = Ptr.u - PtrCie.u - sizeof(uint32_t);
1561
1562	/*
1563	* Generate an FDE for the whole chunk area.
1564	*/
1565	# ifdef IEMNATIVE_USE_LIBUNWIND
1566	pEhFrame->offFda = Ptr.u - (uintptr_t)&pEhFrame->abEhFrame[0];
1567	# endif
1568	RTPTRUNION const PtrFde = Ptr;
1569	Ptr.pu32++ = 123; / The CIE length will be determined later. */
1570	Ptr.pu32 = Ptr.u - PtrCie.u; / Negated self relative CIE address. */
1571	Ptr.pu32++;
1572	Ptr.pu64++ = (uintptr_t)pvChunk; / Absolute start PC of this FDE. */
1573	Ptr.pu64++ = pExecMemAllocator->cbChunk; / PC range length for this PDE. */
1574	# if 0 /* not requried for recent libunwind.dylib nor recent libgcc/glib. */
1575	*Ptr.pb++ = DW_CFA_nop;
1576	# endif
1577	while ((Ptr.u - PtrFde.u) & 3)
1578	*Ptr.pb++ = DW_CFA_nop;
1579	/* Finalize the FDE size. */
1580	*PtrFde.pu32 = Ptr.u - PtrFde.u - sizeof(uint32_t);
1581
1582	/* Terminator entry. */
1583	*Ptr.pu32++ = 0;
1584	Ptr.pu32++ = 0; / just to be sure... */
1585	Assert(Ptr.u - (uintptr_t)&pEhFrame->abEhFrame[0] <= sizeof(pEhFrame->abEhFrame));
1586
1587	/*
1588	* Register it.
1589	*/
1590	# ifdef IEMNATIVE_USE_LIBUNWIND
1591	__register_frame(&pEhFrame->abEhFrame[pEhFrame->offFda]);
1592	# else
1593	memset(pEhFrame->abObject, 0xf6, sizeof(pEhFrame->abObject)); /* color the memory to better spot usage */
1594	__register_frame_info(pEhFrame->abEhFrame, pEhFrame->abObject);
1595	# endif
1596
1597	# ifdef IEMNATIVE_USE_GDB_JIT
1598	/*
1599	* Now for telling GDB about this (experimental).
1600	*
1601	* This seems to work best with ET_DYN.
1602	*/
1603	GDBJITSYMFILE * const pSymFile = (GDBJITSYMFILE *)iemExecMemAllocatorAllocBytesInChunk(pExecMemAllocator, idxChunk,
1604	sizeof(GDBJITSYMFILE), NULL);
1605	AssertReturn(pSymFile, VERR_INTERNAL_ERROR_5);
1606	unsigned const offSymFileInChunk = (uintptr_t)pSymFile - (uintptr_t)pvChunk;
1607
1608	RT_ZERO(*pSymFile);
1609
1610	/*
1611	* The ELF header:
1612	*/
1613	pSymFile->EHdr.e_ident[0] = ELFMAG0;
1614	pSymFile->EHdr.e_ident[1] = ELFMAG1;
1615	pSymFile->EHdr.e_ident[2] = ELFMAG2;
1616	pSymFile->EHdr.e_ident[3] = ELFMAG3;
1617	pSymFile->EHdr.e_ident[EI_VERSION] = EV_CURRENT;
1618	pSymFile->EHdr.e_ident[EI_CLASS] = ELFCLASS64;
1619	pSymFile->EHdr.e_ident[EI_DATA] = ELFDATA2LSB;
1620	pSymFile->EHdr.e_ident[EI_OSABI] = ELFOSABI_NONE;
1621	# ifdef IEMNATIVE_USE_GDB_JIT_ET_DYN
1622	pSymFile->EHdr.e_type = ET_DYN;
1623	# else
1624	pSymFile->EHdr.e_type = ET_REL;
1625	# endif
1626	# ifdef RT_ARCH_AMD64
1627	pSymFile->EHdr.e_machine = EM_AMD64;
1628	# elif defined(RT_ARCH_ARM64)
1629	pSymFile->EHdr.e_machine = EM_AARCH64;
1630	# else
1631	# error "port me"
1632	# endif
1633	pSymFile->EHdr.e_version = 1; /?/
1634	pSymFile->EHdr.e_entry = 0;
1635	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1636	pSymFile->EHdr.e_phoff = RT_UOFFSETOF(GDBJITSYMFILE, aPhdrs);
1637	# else
1638	pSymFile->EHdr.e_phoff = 0;
1639	# endif
1640	pSymFile->EHdr.e_shoff = sizeof(pSymFile->EHdr);
1641	pSymFile->EHdr.e_flags = 0;
1642	pSymFile->EHdr.e_ehsize = sizeof(pSymFile->EHdr);
1643	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1644	pSymFile->EHdr.e_phentsize = sizeof(pSymFile->aPhdrs[0]);
1645	pSymFile->EHdr.e_phnum = RT_ELEMENTS(pSymFile->aPhdrs);
1646	# else
1647	pSymFile->EHdr.e_phentsize = 0;
1648	pSymFile->EHdr.e_phnum = 0;
1649	# endif
1650	pSymFile->EHdr.e_shentsize = sizeof(pSymFile->aShdrs[0]);
1651	pSymFile->EHdr.e_shnum = RT_ELEMENTS(pSymFile->aShdrs);
1652	pSymFile->EHdr.e_shstrndx = 0; /* set later */
1653
1654	uint32_t offStrTab = 0;
1655	#define APPEND_STR(a_szStr) do { \
1656	memcpy(&pSymFile->szzStrTab[offStrTab], a_szStr, sizeof(a_szStr)); \
1657	offStrTab += sizeof(a_szStr); \
1658	Assert(offStrTab < sizeof(pSymFile->szzStrTab)); \
1659	} while (0)
1660	#define APPEND_STR_FMT(a_szStr, ...) do { \
1661	offStrTab += RTStrPrintf(&pSymFile->szzStrTab[offStrTab], sizeof(pSymFile->szzStrTab) - offStrTab, a_szStr, __VA_ARGS__); \
1662	offStrTab++; \
1663	Assert(offStrTab < sizeof(pSymFile->szzStrTab)); \
1664	} while (0)
1665
1666	/*
1667	* Section headers.
1668	*/
1669	/* Section header #0: NULL */
1670	unsigned i = 0;
1671	APPEND_STR("");
1672	RT_ZERO(pSymFile->aShdrs[i]);
1673	i++;
1674
1675	/* Section header: .eh_frame */
1676	pSymFile->aShdrs[i].sh_name = offStrTab;
1677	APPEND_STR(".eh_frame");
1678	pSymFile->aShdrs[i].sh_type = SHT_PROGBITS;
1679	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC \| SHF_EXECINSTR;
1680	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN) \|\| defined(IEMNATIVE_USE_GDB_JIT_ELF_RVAS)
1681	pSymFile->aShdrs[i].sh_offset
1682	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, abEhFrame);
1683	# else
1684	pSymFile->aShdrs[i].sh_addr = (uintptr_t)&pSymFile->abEhFrame[0];
1685	pSymFile->aShdrs[i].sh_offset = 0;
1686	# endif
1687
1688	pSymFile->aShdrs[i].sh_size = sizeof(pEhFrame->abEhFrame);
1689	pSymFile->aShdrs[i].sh_link = 0;
1690	pSymFile->aShdrs[i].sh_info = 0;
1691	pSymFile->aShdrs[i].sh_addralign = 1;
1692	pSymFile->aShdrs[i].sh_entsize = 0;
1693	memcpy(pSymFile->abEhFrame, pEhFrame->abEhFrame, sizeof(pEhFrame->abEhFrame));
1694	i++;
1695
1696	/* Section header: .shstrtab */
1697	unsigned const iShStrTab = i;
1698	pSymFile->EHdr.e_shstrndx = iShStrTab;
1699	pSymFile->aShdrs[i].sh_name = offStrTab;
1700	APPEND_STR(".shstrtab");
1701	pSymFile->aShdrs[i].sh_type = SHT_STRTAB;
1702	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC;
1703	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN) \|\| defined(IEMNATIVE_USE_GDB_JIT_ELF_RVAS)
1704	pSymFile->aShdrs[i].sh_offset
1705	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, szzStrTab);
1706	# else
1707	pSymFile->aShdrs[i].sh_addr = (uintptr_t)&pSymFile->szzStrTab[0];
1708	pSymFile->aShdrs[i].sh_offset = 0;
1709	# endif
1710	pSymFile->aShdrs[i].sh_size = sizeof(pSymFile->szzStrTab);
1711	pSymFile->aShdrs[i].sh_link = 0;
1712	pSymFile->aShdrs[i].sh_info = 0;
1713	pSymFile->aShdrs[i].sh_addralign = 1;
1714	pSymFile->aShdrs[i].sh_entsize = 0;
1715	i++;
1716
1717	/* Section header: .symbols */
1718	pSymFile->aShdrs[i].sh_name = offStrTab;
1719	APPEND_STR(".symtab");
1720	pSymFile->aShdrs[i].sh_type = SHT_SYMTAB;
1721	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC;
1722	pSymFile->aShdrs[i].sh_offset
1723	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, aSymbols);
1724	pSymFile->aShdrs[i].sh_size = sizeof(pSymFile->aSymbols);
1725	pSymFile->aShdrs[i].sh_link = iShStrTab;
1726	pSymFile->aShdrs[i].sh_info = RT_ELEMENTS(pSymFile->aSymbols);
1727	pSymFile->aShdrs[i].sh_addralign = sizeof(pSymFile->aSymbols[0].st_value);
1728	pSymFile->aShdrs[i].sh_entsize = sizeof(pSymFile->aSymbols[0]);
1729	i++;
1730
1731	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1732	/* Section header: .symbols */
1733	pSymFile->aShdrs[i].sh_name = offStrTab;
1734	APPEND_STR(".dynsym");
1735	pSymFile->aShdrs[i].sh_type = SHT_DYNSYM;
1736	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC;
1737	pSymFile->aShdrs[i].sh_offset
1738	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, aDynSyms);
1739	pSymFile->aShdrs[i].sh_size = sizeof(pSymFile->aDynSyms);
1740	pSymFile->aShdrs[i].sh_link = iShStrTab;
1741	pSymFile->aShdrs[i].sh_info = RT_ELEMENTS(pSymFile->aDynSyms);
1742	pSymFile->aShdrs[i].sh_addralign = sizeof(pSymFile->aDynSyms[0].st_value);
1743	pSymFile->aShdrs[i].sh_entsize = sizeof(pSymFile->aDynSyms[0]);
1744	i++;
1745	# endif
1746
1747	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1748	/* Section header: .dynamic */
1749	pSymFile->aShdrs[i].sh_name = offStrTab;
1750	APPEND_STR(".dynamic");
1751	pSymFile->aShdrs[i].sh_type = SHT_DYNAMIC;
1752	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC;
1753	pSymFile->aShdrs[i].sh_offset
1754	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, aDyn);
1755	pSymFile->aShdrs[i].sh_size = sizeof(pSymFile->aDyn);
1756	pSymFile->aShdrs[i].sh_link = iShStrTab;
1757	pSymFile->aShdrs[i].sh_info = 0;
1758	pSymFile->aShdrs[i].sh_addralign = 1;
1759	pSymFile->aShdrs[i].sh_entsize = sizeof(pSymFile->aDyn[0]);
1760	i++;
1761	# endif
1762
1763	/* Section header: .text */
1764	unsigned const iShText = i;
1765	pSymFile->aShdrs[i].sh_name = offStrTab;
1766	APPEND_STR(".text");
1767	pSymFile->aShdrs[i].sh_type = SHT_PROGBITS;
1768	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC \| SHF_EXECINSTR;
1769	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN) \|\| defined(IEMNATIVE_USE_GDB_JIT_ELF_RVAS)
1770	pSymFile->aShdrs[i].sh_offset
1771	= pSymFile->aShdrs[i].sh_addr = sizeof(GDBJITSYMFILE);
1772	# else
1773	pSymFile->aShdrs[i].sh_addr = (uintptr_t)(pSymFile + 1);
1774	pSymFile->aShdrs[i].sh_offset = 0;
1775	# endif
1776	pSymFile->aShdrs[i].sh_size = pExecMemAllocator->cbChunk - offSymFileInChunk - sizeof(GDBJITSYMFILE);
1777	pSymFile->aShdrs[i].sh_link = 0;
1778	pSymFile->aShdrs[i].sh_info = 0;
1779	pSymFile->aShdrs[i].sh_addralign = 1;
1780	pSymFile->aShdrs[i].sh_entsize = 0;
1781	i++;
1782
1783	Assert(i == RT_ELEMENTS(pSymFile->aShdrs));
1784
1785	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1786	/*
1787	* The program headers:
1788	*/
1789	/* Everything in a single LOAD segment: */
1790	i = 0;
1791	pSymFile->aPhdrs[i].p_type = PT_LOAD;
1792	pSymFile->aPhdrs[i].p_flags = PF_X \| PF_R;
1793	pSymFile->aPhdrs[i].p_offset
1794	= pSymFile->aPhdrs[i].p_vaddr
1795	= pSymFile->aPhdrs[i].p_paddr = 0;
1796	pSymFile->aPhdrs[i].p_filesz /* Size of segment in file. */
1797	= pSymFile->aPhdrs[i].p_memsz = pExecMemAllocator->cbChunk - offSymFileInChunk;
1798	pSymFile->aPhdrs[i].p_align = HOST_PAGE_SIZE;
1799	i++;
1800	/* The .dynamic segment. */
1801	pSymFile->aPhdrs[i].p_type = PT_DYNAMIC;
1802	pSymFile->aPhdrs[i].p_flags = PF_R;
1803	pSymFile->aPhdrs[i].p_offset
1804	= pSymFile->aPhdrs[i].p_vaddr
1805	= pSymFile->aPhdrs[i].p_paddr = RT_UOFFSETOF(GDBJITSYMFILE, aDyn);
1806	pSymFile->aPhdrs[i].p_filesz /* Size of segment in file. */
1807	= pSymFile->aPhdrs[i].p_memsz = sizeof(pSymFile->aDyn);
1808	pSymFile->aPhdrs[i].p_align = sizeof(pSymFile->aDyn[0].d_tag);
1809	i++;
1810
1811	Assert(i == RT_ELEMENTS(pSymFile->aPhdrs));
1812
1813	/*
1814	* The dynamic section:
1815	*/
1816	i = 0;
1817	pSymFile->aDyn[i].d_tag = DT_SONAME;
1818	pSymFile->aDyn[i].d_un.d_val = offStrTab;
1819	APPEND_STR_FMT("iem-exec-chunk-%u-%u", pVCpu->idCpu, idxChunk);
1820	i++;
1821	pSymFile->aDyn[i].d_tag = DT_STRTAB;
1822	pSymFile->aDyn[i].d_un.d_ptr = RT_UOFFSETOF(GDBJITSYMFILE, szzStrTab);
1823	i++;
1824	pSymFile->aDyn[i].d_tag = DT_STRSZ;
1825	pSymFile->aDyn[i].d_un.d_val = sizeof(pSymFile->szzStrTab);
1826	i++;
1827	pSymFile->aDyn[i].d_tag = DT_SYMTAB;
1828	pSymFile->aDyn[i].d_un.d_ptr = RT_UOFFSETOF(GDBJITSYMFILE, aDynSyms);
1829	i++;
1830	pSymFile->aDyn[i].d_tag = DT_SYMENT;
1831	pSymFile->aDyn[i].d_un.d_val = sizeof(pSymFile->aDynSyms[0]);
1832	i++;
1833	pSymFile->aDyn[i].d_tag = DT_NULL;
1834	i++;
1835	Assert(i == RT_ELEMENTS(pSymFile->aDyn));
1836	# endif /* IEMNATIVE_USE_GDB_JIT_ET_DYN */
1837
1838	/*
1839	* Symbol tables:
1840	*/
1841	/** @todo gdb doesn't seem to really like this ... */
1842	i = 0;
1843	pSymFile->aSymbols[i].st_name = 0;
1844	pSymFile->aSymbols[i].st_shndx = SHN_UNDEF;
1845	pSymFile->aSymbols[i].st_value = 0;
1846	pSymFile->aSymbols[i].st_size = 0;
1847	pSymFile->aSymbols[i].st_info = ELF64_ST_INFO(STB_LOCAL, STT_NOTYPE);
1848	pSymFile->aSymbols[i].st_other = 0 /* STV_DEFAULT */;
1849	# ifdef IEMNATIVE_USE_GDB_JIT_ET_DYN
1850	pSymFile->aDynSyms[0] = pSymFile->aSymbols[i];
1851	# endif
1852	i++;
1853
1854	pSymFile->aSymbols[i].st_name = 0;
1855	pSymFile->aSymbols[i].st_shndx = SHN_ABS;
1856	pSymFile->aSymbols[i].st_value = 0;
1857	pSymFile->aSymbols[i].st_size = 0;
1858	pSymFile->aSymbols[i].st_info = ELF64_ST_INFO(STB_LOCAL, STT_FILE);
1859	pSymFile->aSymbols[i].st_other = 0 /* STV_DEFAULT */;
1860	i++;
1861
1862	pSymFile->aSymbols[i].st_name = offStrTab;
1863	APPEND_STR_FMT("iem_exec_chunk_%u_%u", pVCpu->idCpu, idxChunk);
1864	# if 0
1865	pSymFile->aSymbols[i].st_shndx = iShText;
1866	pSymFile->aSymbols[i].st_value = 0;
1867	# else
1868	pSymFile->aSymbols[i].st_shndx = SHN_ABS;
1869	pSymFile->aSymbols[i].st_value = (uintptr_t)(pSymFile + 1);
1870	# endif
1871	pSymFile->aSymbols[i].st_size = pSymFile->aShdrs[iShText].sh_size;
1872	pSymFile->aSymbols[i].st_info = ELF64_ST_INFO(STB_GLOBAL, STT_FUNC);
1873	pSymFile->aSymbols[i].st_other = 0 /* STV_DEFAULT */;
1874	# ifdef IEMNATIVE_USE_GDB_JIT_ET_DYN
1875	pSymFile->aDynSyms[1] = pSymFile->aSymbols[i];
1876	pSymFile->aDynSyms[1].st_value = (uintptr_t)(pSymFile + 1);
1877	# endif
1878	i++;
1879
1880	Assert(i == RT_ELEMENTS(pSymFile->aSymbols));
1881	Assert(offStrTab < sizeof(pSymFile->szzStrTab));
1882
1883	/*
1884	* The GDB JIT entry and informing GDB.
1885	*/
1886	pEhFrame->GdbJitEntry.pbSymFile = (uint8_t *)pSymFile;
1887	# if 1
1888	pEhFrame->GdbJitEntry.cbSymFile = pExecMemAllocator->cbChunk - ((uintptr_t)pSymFile - (uintptr_t)pvChunk);
1889	# else
1890	pEhFrame->GdbJitEntry.cbSymFile = sizeof(GDBJITSYMFILE);
1891	# endif
1892
1893	RTOnce(&g_IemNativeGdbJitOnce, iemNativeGdbJitInitOnce, NULL);
1894	RTCritSectEnter(&g_IemNativeGdbJitLock);
1895	pEhFrame->GdbJitEntry.pNext = NULL;
1896	pEhFrame->GdbJitEntry.pPrev = __jit_debug_descriptor.pTail;
1897	if (__jit_debug_descriptor.pTail)
1898	__jit_debug_descriptor.pTail->pNext = &pEhFrame->GdbJitEntry;
1899	else
1900	__jit_debug_descriptor.pHead = &pEhFrame->GdbJitEntry;
1901	__jit_debug_descriptor.pTail = &pEhFrame->GdbJitEntry;
1902	__jit_debug_descriptor.pRelevant = &pEhFrame->GdbJitEntry;
1903
1904	/* Notify GDB: */
1905	__jit_debug_descriptor.enmAction = kGdbJitaction_Register;
1906	__jit_debug_register_code();
1907	__jit_debug_descriptor.enmAction = kGdbJitaction_NoAction;
1908	RTCritSectLeave(&g_IemNativeGdbJitLock);
1909
1910	# else /* !IEMNATIVE_USE_GDB_JIT */
1911	RT_NOREF(pVCpu);
1912	# endif /* !IEMNATIVE_USE_GDB_JIT */
1913
1914	return VINF_SUCCESS;
1915	}
1916
1917	# endif /* !RT_OS_WINDOWS */
1918	#endif /* IN_RING3 */
1919
1920
1921	/**
1922	* Adds another chunk to the executable memory allocator.
1923	*
1924	* This is used by the init code for the initial allocation and later by the
1925	* regular allocator function when it's out of memory.
1926	*/
1927	static int iemExecMemAllocatorGrow(PVMCPUCC pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator)
1928	{
1929	/* Check that we've room for growth. */
1930	uint32_t const idxChunk = pExecMemAllocator->cChunks;
1931	AssertLogRelReturn(idxChunk < pExecMemAllocator->cMaxChunks, VERR_OUT_OF_RESOURCES);
1932
1933	/* Allocate a chunk. */
1934	#ifdef RT_OS_DARWIN
1935	void *pvChunk = RTMemPageAllocEx(pExecMemAllocator->cbChunk, 0);
1936	#else
1937	void *pvChunk = RTMemPageAllocEx(pExecMemAllocator->cbChunk, RTMEMPAGEALLOC_F_EXECUTABLE);
1938	#endif
1939	AssertLogRelReturn(pvChunk, VERR_NO_EXEC_MEMORY);
1940
1941	#ifdef RT_OS_DARWIN
1942	/*
1943	* Because it is impossible to have a RWX memory allocation on macOS try to remap the memory
1944	* chunk readable/executable somewhere else so we can save us the hassle of switching between
1945	* protections when exeuctable memory is allocated.
1946	*/
1947	int rc = VERR_NO_EXEC_MEMORY;
1948	mach_port_t hPortTask = mach_task_self();
1949	mach_vm_address_t AddrChunk = (mach_vm_address_t)pvChunk;
1950	mach_vm_address_t AddrRemapped = 0;
1951	vm_prot_t ProtCur = 0;
1952	vm_prot_t ProtMax = 0;
1953	kern_return_t krc = mach_vm_remap(hPortTask, &AddrRemapped, pExecMemAllocator->cbChunk, 0,
1954	VM_FLAGS_ANYWHERE \| VM_FLAGS_RETURN_DATA_ADDR,
1955	hPortTask, AddrChunk, FALSE, &ProtCur, &ProtMax,
1956	VM_INHERIT_NONE);
1957	if (krc == KERN_SUCCESS)
1958	{
1959	krc = mach_vm_protect(mach_task_self(), AddrRemapped, pExecMemAllocator->cbChunk, FALSE, VM_PROT_READ \| VM_PROT_EXECUTE);
1960	if (krc == KERN_SUCCESS)
1961	rc = VINF_SUCCESS;
1962	else
1963	{
1964	AssertLogRelMsgFailed(("mach_vm_protect -> %d (%#x)\n", krc, krc));
1965	krc = mach_vm_deallocate(hPortTask, AddrRemapped, pExecMemAllocator->cbChunk);
1966	Assert(krc == KERN_SUCCESS);
1967	}
1968	}
1969	else
1970	AssertLogRelMsgFailed(("mach_vm_remap -> %d (%#x)\n", krc, krc));
1971	if (RT_FAILURE(rc))
1972	{
1973	RTMemPageFree(pvChunk, pExecMemAllocator->cbChunk);
1974	return rc;
1975	}
1976
1977	void pvChunkRx = (void )AddrRemapped;
1978	#else
1979	int rc = VINF_SUCCESS;
1980	void *pvChunkRx = pvChunk;
1981	#endif
1982
1983	/*
1984	* Add the chunk.
1985	*
1986	* This must be done before the unwind init so windows can allocate
1987	* memory from the chunk when using the alternative sub-allocator.
1988	*/
1989	pExecMemAllocator->aChunks[idxChunk].pvChunkRw = pvChunk;
1990	pExecMemAllocator->aChunks[idxChunk].pvChunkRx = pvChunkRx;
1991	#ifdef IN_RING3
1992	pExecMemAllocator->aChunks[idxChunk].pvUnwindInfo = NULL;
1993	#endif
1994	pExecMemAllocator->aChunks[idxChunk].cFreeUnits = pExecMemAllocator->cUnitsPerChunk;
1995	pExecMemAllocator->aChunks[idxChunk].idxFreeHint = 0;
1996	memset(&pExecMemAllocator->pbmAlloc[pExecMemAllocator->cBitmapElementsPerChunk * idxChunk],
1997	0, sizeof(pExecMemAllocator->pbmAlloc[0]) * pExecMemAllocator->cBitmapElementsPerChunk);
1998
1999	pExecMemAllocator->cChunks = idxChunk + 1;
2000	pExecMemAllocator->idxChunkHint = idxChunk;
2001
2002	pExecMemAllocator->cbTotal += pExecMemAllocator->cbChunk;
2003	pExecMemAllocator->cbFree += pExecMemAllocator->cbChunk;
2004
2005	/* If there is a chunk context init callback call it. */
2006	rc = iemNativeRecompileAttachExecMemChunkCtx(pVCpu, idxChunk, &pExecMemAllocator->aChunks[idxChunk].pCtx);
2007	#ifdef IN_RING3
2008	/*
2009	* Initialize the unwind information (this cannot really fail atm).
2010	* (This sets pvUnwindInfo.)
2011	*/
2012	if (RT_SUCCESS(rc))
2013	rc = iemExecMemAllocatorInitAndRegisterUnwindInfoForChunk(pVCpu, pExecMemAllocator, pvChunkRx, idxChunk);
2014	#endif
2015	if (RT_SUCCESS(rc))
2016	{ /* likely */ }
2017	else
2018	{
2019	/* Just in case the impossible happens, undo the above up: */
2020	pExecMemAllocator->cbTotal -= pExecMemAllocator->cbChunk;
2021	pExecMemAllocator->cbFree -= pExecMemAllocator->aChunks[idxChunk].cFreeUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
2022	pExecMemAllocator->cChunks = idxChunk;
2023	memset(&pExecMemAllocator->pbmAlloc[pExecMemAllocator->cBitmapElementsPerChunk * idxChunk],
2024	0xff, sizeof(pExecMemAllocator->pbmAlloc[0]) * pExecMemAllocator->cBitmapElementsPerChunk);
2025	pExecMemAllocator->aChunks[idxChunk].pvChunkRw = NULL;
2026	pExecMemAllocator->aChunks[idxChunk].cFreeUnits = 0;
2027
2028	# ifdef RT_OS_DARWIN
2029	krc = mach_vm_deallocate(mach_task_self(), (mach_vm_address_t)pExecMemAllocator->aChunks[idxChunk].pvChunkRx,
2030	pExecMemAllocator->cbChunk);
2031	Assert(krc == KERN_SUCCESS);
2032	# endif
2033
2034	RTMemPageFree(pvChunk, pExecMemAllocator->cbChunk);
2035	return rc;
2036	}
2037
2038	return VINF_SUCCESS;
2039	}
2040
2041
2042	/**
2043	* Initializes the executable memory allocator for native recompilation on the
2044	* calling EMT.
2045	*
2046	* @returns VBox status code.
2047	* @param pVCpu The cross context virtual CPU structure of the calling
2048	* thread.
2049	* @param cbMax The max size of the allocator.
2050	* @param cbInitial The initial allocator size.
2051	* @param cbChunk The chunk size, 0 or UINT32_MAX for default (@a cbMax
2052	* dependent).
2053	*/
2054	int iemExecMemAllocatorInit(PVMCPU pVCpu, uint64_t cbMax, uint64_t cbInitial, uint32_t cbChunk) RT_NOEXCEPT
2055	{
2056	/*
2057	* Validate input.
2058	*/
2059	AssertLogRelMsgReturn(cbMax >= _1M && cbMax <= _4G+_4G, ("cbMax=%RU64 (%RX64)\n", cbMax, cbMax), VERR_OUT_OF_RANGE);
2060	AssertReturn(cbInitial <= cbMax, VERR_OUT_OF_RANGE);
2061	AssertLogRelMsgReturn( cbChunk != UINT32_MAX
2062	\|\| cbChunk == 0
2063	\|\| ( RT_IS_POWER_OF_TWO(cbChunk)
2064	&& cbChunk >= _1M
2065	&& cbChunk <= _256M
2066	&& cbChunk <= cbMax),
2067	("cbChunk=%RU32 (%RX32) cbMax=%RU64\n", cbChunk, cbChunk, cbMax),
2068	VERR_OUT_OF_RANGE);
2069
2070	/*
2071	* Adjust/figure out the chunk size.
2072	*/
2073	if (cbChunk == 0 \|\| cbChunk == UINT32_MAX)
2074	{
2075	if (cbMax >= _256M)
2076	cbChunk = _64M;
2077	else
2078	{
2079	if (cbMax < _16M)
2080	cbChunk = cbMax >= _4M ? _4M : (uint32_t)cbMax;
2081	else
2082	cbChunk = (uint32_t)cbMax / 4;
2083	if (!RT_IS_POWER_OF_TWO(cbChunk))
2084	cbChunk = RT_BIT_32(ASMBitLastSetU32(cbChunk));
2085	}
2086	}
2087	#if defined(RT_OS_AMD64)
2088	Assert(cbChunk <= _2G);
2089	#elif defined(RT_OS_ARM64)
2090	if (cbChunk > _128M)
2091	cbChunk = _128M; /* Max relative branch distance is +/-2^(25+2) = +/-0x8000000 (134 217 728). */
2092	#endif
2093
2094	if (cbChunk > cbMax)
2095	cbMax = cbChunk;
2096	else
2097	cbMax = (cbMax - 1 + cbChunk) / cbChunk * cbChunk;
2098	uint32_t const cMaxChunks = (uint32_t)(cbMax / cbChunk);
2099	AssertLogRelReturn((uint64_t)cMaxChunks * cbChunk == cbMax, VERR_INTERNAL_ERROR_3);
2100
2101	/*
2102	* Allocate and initialize the allocatore instance.
2103	*/
2104	size_t const offBitmaps = RT_ALIGN_Z(RT_UOFFSETOF_DYN(IEMEXECMEMALLOCATOR, aChunks[cMaxChunks]), RT_CACHELINE_SIZE);
2105	size_t const cbBitmaps = (size_t)(cbChunk >> (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT + 3)) * cMaxChunks;
2106	size_t cbNeeded = offBitmaps + cbBitmaps;
2107	AssertCompile(IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT <= 10);
2108	Assert(cbChunk > RT_BIT_32(IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT + 3));
2109	#if defined(IN_RING3) && !defined(RT_OS_WINDOWS)
2110	size_t const offEhFrames = RT_ALIGN_Z(cbNeeded, RT_CACHELINE_SIZE);
2111	cbNeeded += sizeof(IEMEXECMEMCHUNKEHFRAME) * cMaxChunks;
2112	#endif
2113	PIEMEXECMEMALLOCATOR pExecMemAllocator = (PIEMEXECMEMALLOCATOR)RTMemAllocZ(cbNeeded);
2114	AssertLogRelMsgReturn(pExecMemAllocator, ("cbNeeded=%zx cMaxChunks=%#x cbChunk=%#x\n", cbNeeded, cMaxChunks, cbChunk),
2115	VERR_NO_MEMORY);
2116	pExecMemAllocator->uMagic = IEMEXECMEMALLOCATOR_MAGIC;
2117	pExecMemAllocator->cbChunk = cbChunk;
2118	pExecMemAllocator->cMaxChunks = cMaxChunks;
2119	pExecMemAllocator->cChunks = 0;
2120	pExecMemAllocator->idxChunkHint = 0;
2121	pExecMemAllocator->cAllocations = 0;
2122	pExecMemAllocator->cbTotal = 0;
2123	pExecMemAllocator->cbFree = 0;
2124	pExecMemAllocator->cbAllocated = 0;
2125	#ifdef VBOX_WITH_STATISTICS
2126	pExecMemAllocator->cbUnusable = 0;
2127	#endif
2128	pExecMemAllocator->pbmAlloc = (uint64_t *)((uintptr_t)pExecMemAllocator + offBitmaps);
2129	pExecMemAllocator->cUnitsPerChunk = cbChunk >> IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
2130	pExecMemAllocator->cBitmapElementsPerChunk = cbChunk >> (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT + 6);
2131	memset(pExecMemAllocator->pbmAlloc, 0xff, cbBitmaps); /* Mark everything as allocated. Clear when chunks are added. */
2132	#if defined(IN_RING3) && !defined(RT_OS_WINDOWS)
2133	pExecMemAllocator->paEhFrames = (PIEMEXECMEMCHUNKEHFRAME)((uintptr_t)pExecMemAllocator + offEhFrames);
2134	#endif
2135	for (uint32_t i = 0; i < cMaxChunks; i++)
2136	{
2137	pExecMemAllocator->aChunks[i].cFreeUnits = 0;
2138	pExecMemAllocator->aChunks[i].idxFreeHint = 0;
2139	pExecMemAllocator->aChunks[i].pvChunkRw = NULL;
2140	#ifdef IN_RING0
2141	pExecMemAllocator->aChunks[i].hMemObj = NIL_RTR0MEMOBJ;
2142	#else
2143	pExecMemAllocator->aChunks[i].pvUnwindInfo = NULL;
2144	#endif
2145	}
2146	pVCpu->iem.s.pExecMemAllocatorR3 = pExecMemAllocator;
2147
2148	/*
2149	* Do the initial allocations.
2150	*/
2151	while (cbInitial < (uint64_t)pExecMemAllocator->cChunks * pExecMemAllocator->cbChunk)
2152	{
2153	int rc = iemExecMemAllocatorGrow(pVCpu, pExecMemAllocator);
2154	AssertLogRelRCReturn(rc, rc);
2155	}
2156
2157	pExecMemAllocator->idxChunkHint = 0;
2158
2159	/*
2160	* Register statistics.
2161	*/
2162	PUVM const pUVM = pVCpu->pUVCpu->pUVM;
2163	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cAllocations, STAMTYPE_U32, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
2164	"Current number of allocations", "/IEM/CPU%u/re/ExecMem/cAllocations", pVCpu->idCpu);
2165	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cChunks, STAMTYPE_U32, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,
2166	"Currently allocated chunks", "/IEM/CPU%u/re/ExecMem/cChunks", pVCpu->idCpu);
2167	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cMaxChunks, STAMTYPE_U32, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,
2168	"Maximum number of chunks", "/IEM/CPU%u/re/ExecMem/cMaxChunks", pVCpu->idCpu);
2169	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbChunk, STAMTYPE_U32, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
2170	"Allocation chunk size", "/IEM/CPU%u/re/ExecMem/cbChunk", pVCpu->idCpu);
2171	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbAllocated, STAMTYPE_U64, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
2172	"Number of bytes current allocated", "/IEM/CPU%u/re/ExecMem/cbAllocated", pVCpu->idCpu);
2173	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbFree, STAMTYPE_U64, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
2174	"Number of bytes current free", "/IEM/CPU%u/re/ExecMem/cbFree", pVCpu->idCpu);
2175	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbTotal, STAMTYPE_U64, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
2176	"Total number of byte", "/IEM/CPU%u/re/ExecMem/cbTotal", pVCpu->idCpu);
2177	#ifdef VBOX_WITH_STATISTICS
2178	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbUnusable, STAMTYPE_U64, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
2179	"Total number of bytes being unusable", "/IEM/CPU%u/re/ExecMem/cbUnusable", pVCpu->idCpu);
2180	STAMR3RegisterFU(pUVM, &pExecMemAllocator->StatAlloc, STAMTYPE_PROFILE, STAMVISIBILITY_ALWAYS, STAMUNIT_TICKS_PER_CALL,
2181	"Profiling the allocator", "/IEM/CPU%u/re/ExecMem/ProfAlloc", pVCpu->idCpu);
2182	for (unsigned i = 1; i < RT_ELEMENTS(pExecMemAllocator->aStatSizes); i++)
2183	STAMR3RegisterFU(pUVM, &pExecMemAllocator->aStatSizes[i], STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,
2184	"Number of allocations of this number of allocation units",
2185	"/IEM/CPU%u/re/ExecMem/aSize%02u", pVCpu->idCpu, i);
2186	STAMR3RegisterFU(pUVM, &pExecMemAllocator->aStatSizes[0], STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,
2187	"Number of allocations 16 units or larger", "/IEM/CPU%u/re/ExecMem/aSize16OrLarger", pVCpu->idCpu);
2188	#endif
2189	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING
2190	STAMR3RegisterFU(pUVM, &pExecMemAllocator->StatPruneProf, STAMTYPE_PROFILE, STAMVISIBILITY_ALWAYS, STAMUNIT_TICKS_PER_CALL,
2191	"Pruning executable memory (alt)", "/IEM/CPU%u/re/ExecMem/Pruning", pVCpu->idCpu);
2192	STAMR3RegisterFU(pUVM, &pExecMemAllocator->StatPruneRecovered, STAMTYPE_PROFILE, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES_PER_CALL,
2193	"Bytes recovered while pruning", "/IEM/CPU%u/re/ExecMem/PruningRecovered", pVCpu->idCpu);
2194	#endif
2195	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cFruitlessChunkScans, STAMTYPE_U64_RESET, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,
2196	"Chunks fruitlessly scanned for free space", "/IEM/CPU%u/re/ExecMem/FruitlessChunkScans", pVCpu->idCpu);
2197
2198	return VINF_SUCCESS;
2199	}
2200

Note: See TracBrowser for help on using the repository browser.

source: vbox/trunk/src/VBox/VMM/VMMAll/IEMAllN8veExecMem.cpp@ 106378

Download in other formats: