IEMAllN8veExecMem.cpp@ 106443

Last change on this file since 106443 was 106405, checked in by vboxsync, 5 weeks ago
VMM/IEM: Corrected the idxFreeHint setting at the end of iemExecMemAllocatorPrune, it was set to the end of the purned area rather than the start. 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 106405 2024-10-16 21:24:15Z 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 offPruneStart = offChunk;
419	uint32_t const offPruneEnd = RT_MIN(offChunk + cbToPrune, cbChunk);
420
421	/*
422	* Do the pruning. The current approach is the sever kind.
423	*
424	* This is memory bound, as we must load both the allocation header and the
425	* associated TB and then modify them. So, the CPU isn't all that unitilized
426	* here. Try apply some prefetching to speed it up a tiny bit.
427	*/
428	uint64_t cbPruned = 0;
429	uint64_t const u64MagicAndChunkIdx = RT_MAKE_U64(IEMEXECMEMALLOCHDR_MAGIC, idxChunk);
430	uint8_t * const pbChunk = (uint8_t *)pExecMemAllocator->aChunks[idxChunk].pvChunkRx;
431	while (offChunk < offPruneEnd)
432	{
433	PIEMEXECMEMALLOCHDR pHdr = (PIEMEXECMEMALLOCHDR)&pbChunk[offChunk];
434
435	/* Is this the start of an allocation block for a TB? (We typically
436	have one allocation at the start of each chunk for the unwind info
437	where pTb is NULL.) */
438	PIEMTB pTb;
439	if ( pHdr->u64MagicAndChunkIdx == u64MagicAndChunkIdx
440	&& RT_LIKELY((pTb = pHdr->pTb) != NULL))
441	{
442	AssertPtr(pTb);
443
444	uint32_t const cbBlock = RT_ALIGN_32(pTb->Native.cInstructions * sizeof(IEMNATIVEINSTR) + sizeof(*pHdr),
445	IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE);
446
447	/* Prefetch the next header before freeing the current one and its TB. */
448	/** @todo Iff the block size was part of the header in some way, this could be
449	* a tiny bit faster. */
450	offChunk += cbBlock;
451	#if defined(_MSC_VER) && defined(RT_ARCH_AMD64)
452	_mm_prefetch((char *)&pbChunk[offChunk], _MM_HINT_T0);
453	#elif defined(_MSC_VER) && defined(RT_ARCH_ARM64)
454	__prefetch(&pbChunk[offChunk]);
455	#else
456	__builtin_prefetch(&pbChunk[offChunk], 1 /rw/);
457	#endif
458	/* Some paranoia first, though. */
459	AssertBreakStmt(offChunk <= cbChunk, offChunk -= cbBlock - IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE);
460	cbPruned += cbBlock;
461
462	#if 1
463	iemTbAllocatorFreeBulk(pVCpu, pTbAllocator, pTb);
464	#else
465	iemTbAllocatorFree(pVCpu, pTb);
466	#endif
467	}
468	else
469	offChunk += IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE;
470	}
471	STAM_REL_PROFILE_ADD_PERIOD(&pExecMemAllocator->StatPruneRecovered, cbPruned);
472
473	pVCpu->iem.s.ppTbLookupEntryR3 = &pVCpu->iem.s.pTbLookupEntryDummyR3;
474
475	/*
476	* Save the current pruning point.
477	*/
478	pExecMemAllocator->offChunkPrune = offChunk;
479	pExecMemAllocator->idxChunkPrune = idxChunk;
480
481	/* Set the hint to the start of the pruned region. */
482	pExecMemAllocator->idxChunkHint = idxChunk;
483	pExecMemAllocator->aChunks[idxChunk].idxFreeHint = offPruneStart / IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE;
484
485	STAM_REL_PROFILE_STOP(&pExecMemAllocator->StatPruneProf, a);
486	}
487	#endif /* IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING */
488
489
490	#if defined(VBOX_STRICT) \|\| 0
491	/**
492	* The old bitmap scanner code, for comparison and assertions.
493	*/
494	static uint32_t iemExecMemAllocatorFindReqFreeUnitsOld(uint64_t *pbmAlloc, uint32_t cToScan, uint32_t cReqUnits)
495	{
496	/** @todo This can probably be done more efficiently for non-x86 systems. */
497	int iBit = ASMBitFirstClear(pbmAlloc, cToScan);
498	while (iBit >= 0 && (uint32_t)iBit <= cToScan - cReqUnits)
499	{
500	uint32_t idxAddBit = 1;
501	while (idxAddBit < cReqUnits && !ASMBitTest(pbmAlloc, (uint32_t)iBit + idxAddBit))
502	idxAddBit++;
503	if (idxAddBit >= cReqUnits)
504	return (uint32_t)iBit;
505	iBit = ASMBitNextClear(pbmAlloc, cToScan, iBit + idxAddBit - 1);
506	}
507	return UINT32_MAX;
508	}
509	#endif
510
511
512	/**
513	* Bitmap scanner code that looks for a bunch of @a cReqUnits zero bits.
514	*
515	* Booting win11 with a r165098 release build the average native TB size is
516	* around 9 units (of 256 bytes). So, it is unlikely we need to scan any
517	* subsequent words once we hit a patch of zeros, thus @a a_fBig.
518	*
519	* @todo This needs more tweaking. While it is faster the the old code,
520	* it doens't seem like it's all that much. :/
521	*/
522	template<const bool a_fBig>
523	static uint32_t iemExecMemAllocatorFindReqFreeUnits(uint64_t *pbmAlloc, uint32_t c64WordsToScan, uint32_t cReqUnits)
524	{
525	/*
526	* Scan the (section of the) allocation bitmap in 64-bit words.
527	*/
528	unsigned cPrevLeadingZeros = 0;
529	for (uint32_t off = 0; off < c64WordsToScan; off++)
530	{
531	uint64_t uWord = pbmAlloc[off];
532	if (uWord == UINT64_MAX)
533	{
534	/*
535	* Getting thru patches of UINT64_MAX is a frequent problem when the allocator
536	* fills up, so it's definitely worth optimizing.
537	*
538	* The complicated code below is a bit faster on arm. Reducing the per TB cost
539	* from 4255ns to 4106ns (best run out of 10). On win/amd64 there isn't an
540	* obvious gain here, at least not with the data currently being profiled.
541	*/
542	#if 1
543	off++;
544	uint32_t cQuads = (c64WordsToScan - off) / 4;
545
546	/* Align. */
547	if (cQuads > 1)
548	switch (((uintptr_t)&pbmAlloc[off] / sizeof(uint64_t)) & 3)
549	{
550	case 0:
551	break;
552	case 1:
553	{
554	uWord = pbmAlloc[off];
555	uint64_t uWord1 = pbmAlloc[off + 1];
556	uint64_t uWord2 = pbmAlloc[off + 2];
557	if ((uWord & uWord1 & uWord2) == UINT64_MAX)
558	{
559	off += 3;
560	cQuads = (c64WordsToScan - off) / 4;
561	}
562	else if (uWord == UINT64_MAX)
563	{
564	if (uWord1 != UINT64_MAX)
565	{
566	uWord = uWord1;
567	off += 1;
568	}
569	else
570	{
571	uWord = uWord2;
572	off += 2;
573	}
574	}
575	break;
576	}
577	case 2:
578	{
579	uWord = pbmAlloc[off];
580	uint64_t uWord1 = pbmAlloc[off + 1];
581	if ((uWord & uWord1) == UINT64_MAX)
582	{
583	off += 2;
584	cQuads = (c64WordsToScan - off) / 4;
585	}
586	else if (uWord == UINT64_MAX)
587	{
588	uWord = uWord1;
589	off += 1;
590	}
591	break;
592	}
593	case 3:
594	uWord = pbmAlloc[off];
595	if (uWord == UINT64_MAX)
596	{
597	off++;
598	cQuads = (c64WordsToScan - off) / 4;
599	}
600	break;
601	}
602	if (uWord == UINT64_MAX)
603	{
604	/* Looping over 32 bytes at a time. */
605	for (;;)
606	{
607	if (cQuads-- > 0)
608	{
609	uWord = pbmAlloc[off + 0];
610	uint64_t uWord1 = pbmAlloc[off + 1];
611	uint64_t uWord2 = pbmAlloc[off + 2];
612	uint64_t uWord3 = pbmAlloc[off + 3];
613	if ((uWord & uWord1 & uWord2 & uWord3) == UINT64_MAX)
614	off += 4;
615	else
616	{
617	if (uWord != UINT64_MAX)
618	{ }
619	else if (uWord1 != UINT64_MAX)
620	{
621	uWord = uWord1;
622	off += 1;
623	}
624	else if (uWord2 != UINT64_MAX)
625	{
626	uWord = uWord2;
627	off += 2;
628	}
629	else
630	{
631	uWord = uWord3;
632	off += 3;
633	}
634	break;
635	}
636	}
637	else
638	{
639	if (off < c64WordsToScan)
640	{
641	uWord = pbmAlloc[off];
642	if (uWord != UINT64_MAX)
643	break;
644	off++;
645	if (off < c64WordsToScan)
646	{
647	uWord = pbmAlloc[off];
648	if (uWord != UINT64_MAX)
649	break;
650	off++;
651	if (off < c64WordsToScan)
652	{
653	uWord = pbmAlloc[off];
654	if (uWord != UINT64_MAX)
655	break;
656	Assert(off + 1 == c64WordsToScan);
657	}
658	}
659	}
660	return UINT32_MAX;
661	}
662	}
663	}
664	#else
665	do
666	{
667	off++;
668	if (off < c64WordsToScan)
669	uWord = pbmAlloc[off];
670	else
671	return UINT32_MAX;
672	} while (uWord == UINT64_MAX);
673	#endif
674	cPrevLeadingZeros = 0;
675	}
676
677	/*
678	* If we get down here, we have a word that isn't UINT64_MAX.
679	*/
680	if (uWord != 0)
681	{
682	/*
683	* Fend of large request we cannot satisfy before the first set bit.
684	*/
685	if (!a_fBig \|\| cReqUnits < 64 + cPrevLeadingZeros)
686	{
687	#ifdef __GNUC__
688	unsigned cZerosInWord = __builtin_popcountl(~uWord);
689	#elif defined(_MSC_VER) && defined(RT_ARCH_AMD64)
690	unsigned cZerosInWord = __popcnt64(~uWord);
691	#elif defined(_MSC_VER) && defined(RT_ARCH_ARM64)
692	unsigned cZerosInWord = _CountOneBits64(~uWord);
693	#else
694	# pragma message("need popcount intrinsic or something...")
695	unsigned cZerosInWord = 0;
696	for (uint64_t uTmp = ~uWords; uTmp; cZerosInWord++)
697	uTmp &= uTmp - 1; /* Clears the least significant bit set. */
698	#endif
699	if (cZerosInWord + cPrevLeadingZeros >= cReqUnits)
700	{
701	/* Check if we've got a patch of zeros at the trailing end
702	when joined with the previous word: */
703	#ifdef __GNUC__
704	unsigned cTrailingZeros = __builtin_ctzl(uWord);
705	#else
706	unsigned cTrailingZeros = ASMBitFirstSetU64(uWord) - 1;
707	#endif
708	if (cPrevLeadingZeros + cTrailingZeros >= cReqUnits)
709	return off * 64 - cPrevLeadingZeros;
710
711	/*
712	* Try leading zeros before we get on with the tedious stuff.
713	*/
714	#ifdef __GNUC__
715	cPrevLeadingZeros = __builtin_clzl(uWord);
716	#else
717	cPrevLeadingZeros = 64 - ASMBitLastSetU64(uWord);
718	#endif
719	if (cPrevLeadingZeros >= cReqUnits)
720	return (off + 1) * 64 - cPrevLeadingZeros;
721
722	/*
723	* Check the popcount again sans leading & trailing before looking
724	* inside the word.
725	*/
726	cZerosInWord -= cPrevLeadingZeros + cTrailingZeros;
727	if (cZerosInWord >= cReqUnits)
728	{
729	/* 1; 64 - 0 - 1 = 63; */
730	unsigned const iBitLast = 64 - cPrevLeadingZeros - cReqUnits; /** @todo boundrary */
731	unsigned iBit = cTrailingZeros;
732	uWord >>= cTrailingZeros;
733	do
734	{
735	Assert(uWord & 1);
736	#ifdef __GNUC__
737	unsigned iZeroBit = __builtin_ctzl(~uWord);
738	#else
739	unsigned iZeroBit = ASMBitFirstSetU64(~uWord) - 1;
740	#endif
741	iBit += iZeroBit;
742	uWord >>= iZeroBit;
743	Assert(iBit <= iBitLast);
744	Assert((uWord & 1) == 0);
745	#ifdef __GNUC__
746	unsigned cZeros = __builtin_ctzl(uWord);
747	#else
748	unsigned cZeros = ASMBitFirstSetU64(uWord) - 1;
749	#endif
750	if (cZeros >= cReqUnits)
751	return off * 64 + iBit;
752
753	cZerosInWord -= cZeros; /* (may underflow as we will count shifted in zeros) */
754	iBit += cZeros;
755	uWord >>= cZeros;
756	} while ((int)cZerosInWord >= (int)cReqUnits && iBit < iBitLast);
757	}
758	continue; /* we've already calculated cPrevLeadingZeros */
759	}
760	}
761
762	/* Update the leading (MSB) zero count. */
763	#ifdef __GNUC__
764	cPrevLeadingZeros = __builtin_clzl(uWord);
765	#else
766	cPrevLeadingZeros = 64 - ASMBitLastSetU64(uWord);
767	#endif
768	}
769	/*
770	* uWord == 0
771	*/
772	else
773	{
774	if RT_CONSTEXPR_IF(!a_fBig)
775	return off * 64 - cPrevLeadingZeros;
776	else /* keep else */
777	{
778	if (cPrevLeadingZeros + 64 >= cReqUnits)
779	return off * 64 - cPrevLeadingZeros;
780	for (uint32_t off2 = off + 1;; off2++)
781	{
782	if (off2 < c64WordsToScan)
783	{
784	uWord = pbmAlloc[off2];
785	if (uWord == UINT64_MAX)
786	{
787	cPrevLeadingZeros = 0;
788	break;
789	}
790	if (uWord == 0)
791	{
792	if (cPrevLeadingZeros + (off2 - off + 1) * 64 >= cReqUnits)
793	return off * 64 - cPrevLeadingZeros;
794	}
795	else
796	{
797	#ifdef __GNUC__
798	unsigned cTrailingZeros = __builtin_ctzl(uWord);
799	#else
800	unsigned cTrailingZeros = ASMBitFirstSetU64(uWord) - 1;
801	#endif
802	if (cPrevLeadingZeros + (off2 - off) * 64 + cTrailingZeros >= cReqUnits)
803	return off * 64 - cPrevLeadingZeros;
804	#ifdef __GNUC__
805	cPrevLeadingZeros = __builtin_clzl(uWord);
806	#else
807	cPrevLeadingZeros = 64 - ASMBitLastSetU64(uWord);
808	#endif
809	break;
810	}
811	}
812	else
813	return UINT32_MAX;
814	}
815	}
816	}
817	}
818	return UINT32_MAX;
819	}
820
821
822	/**
823	* Try allocate a block of @a cReqUnits in the chunk @a idxChunk.
824	*/
825	static void *
826	iemExecMemAllocatorAllocInChunkInt(PIEMEXECMEMALLOCATOR pExecMemAllocator, uint64_t *pbmAlloc, uint32_t idxFirst,
827	uint32_t cToScan, uint32_t cReqUnits, uint32_t idxChunk, PIEMTB pTb,
828	void *ppvExec, PCIEMNATIVEPERCHUNKCTX ppChunkCtx)
829	{
830	/*
831	* Shift the bitmap to the idxFirst bit so we can use ASMBitFirstClear.
832	*/
833	Assert(!(cToScan & 63));
834	Assert(!(idxFirst & 63));
835	Assert(cToScan + idxFirst <= pExecMemAllocator->cUnitsPerChunk);
836	pbmAlloc += idxFirst / 64;
837	cToScan += idxFirst & 63;
838	Assert(!(cToScan & 63));
839
840	#if 1
841	uint32_t const iBit = cReqUnits < 64
842	? iemExecMemAllocatorFindReqFreeUnits<false>(pbmAlloc, cToScan / 64, cReqUnits)
843	: iemExecMemAllocatorFindReqFreeUnits<true>( pbmAlloc, cToScan / 64, cReqUnits);
844	# ifdef VBOX_STRICT
845	uint32_t const iBitOld = iemExecMemAllocatorFindReqFreeUnitsOld(pbmAlloc, cToScan, cReqUnits);
846	AssertMsg( iBit == iBitOld
847	\|\| (iBit / 64) == (iBitOld / 64), /* New algorithm will return trailing hit before middle. */
848	("iBit=%#x (%#018RX64); iBitOld=%#x (%#018RX64); cReqUnits=%#x\n",
849	iBit, iBit != UINT32_MAX ? pbmAlloc[iBit / 64] : 0,
850	iBitOld, iBitOld != UINT32_MAX ? pbmAlloc[iBitOld / 64] : 0, cReqUnits));
851	# endif
852	#else
853	uint32_t const iBit = iemExecMemAllocatorFindReqFreeUnitsOld(pbmAlloc, cToScan, cReqUnits);
854	#endif
855	if (iBit != UINT32_MAX)
856	{
857	ASMBitSetRange(pbmAlloc, (uint32_t)iBit, (uint32_t)iBit + cReqUnits);
858
859	PIEMEXECMEMCHUNK const pChunk = &pExecMemAllocator->aChunks[idxChunk];
860	pChunk->cFreeUnits -= cReqUnits;
861	pChunk->idxFreeHint = (uint32_t)iBit + cReqUnits;
862
863	pExecMemAllocator->cAllocations += 1;
864	uint32_t const cbReq = cReqUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
865	pExecMemAllocator->cbAllocated += cbReq;
866	pExecMemAllocator->cbFree -= cbReq;
867	pExecMemAllocator->idxChunkHint = idxChunk;
868
869	void * const pvMemRw = (uint8_t *)pChunk->pvChunkRw
870	+ ((idxFirst + (uint32_t)iBit) << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT);
871
872	if (ppChunkCtx)
873	*ppChunkCtx = pChunk->pCtx;
874
875	/*
876	* Initialize the header and return.
877	*/
878	# ifdef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
879	PIEMEXECMEMALLOCHDR const pHdr = (PIEMEXECMEMALLOCHDR)pvMemRw;
880	pHdr->uMagic = IEMEXECMEMALLOCHDR_MAGIC;
881	pHdr->idxChunk = idxChunk;
882	pHdr->pTb = pTb;
883
884	if (ppvExec)
885	ppvExec = (uint8_t )pChunk->pvChunkRx
886	+ ((idxFirst + (uint32_t)iBit) << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT)
887	+ sizeof(*pHdr);
888
889	return pHdr + 1;
890	#else
891	if (ppvExec)
892	ppvExec = (uint8_t )pChunk->pvChunkRx
893	+ ((idxFirst + (uint32_t)iBit) << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT);
894
895	RT_NOREF(pTb);
896	return pvMem;
897	#endif
898	}
899
900	return NULL;
901	}
902
903
904	/**
905	* Converts requested number of bytes into a unit count.
906	*/
907	DECL_FORCE_INLINE(uint32_t) iemExecMemAllocBytesToUnits(uint32_t cbReq)
908	{
909	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
910	return (cbReq + sizeof(IEMEXECMEMALLOCHDR) + IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1)
911	#else
912	return (cbReq + IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1)
913	#endif
914	>> IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
915	}
916
917
918	DECL_FORCE_INLINE(PIEMNATIVEINSTR)
919	iemExecMemAllocatorAllocUnitsInChunkInner(PIEMEXECMEMALLOCATOR pExecMemAllocator, uint32_t idxChunk, uint32_t cReqUnits,
920	PIEMTB pTb, PIEMNATIVEINSTR ppaExec, PCIEMNATIVEPERCHUNKCTX ppChunkCtx)
921	{
922	uint64_t * const pbmAlloc = &pExecMemAllocator->pbmAlloc[pExecMemAllocator->cBitmapElementsPerChunk * idxChunk];
923	uint32_t const idxHint = pExecMemAllocator->aChunks[idxChunk].idxFreeHint & ~(uint32_t)63;
924	if (idxHint + cReqUnits <= pExecMemAllocator->cUnitsPerChunk)
925	{
926	void *pvRet = iemExecMemAllocatorAllocInChunkInt(pExecMemAllocator, pbmAlloc, idxHint,
927	pExecMemAllocator->cUnitsPerChunk - idxHint,
928	cReqUnits, idxChunk, pTb, (void **)ppaExec, ppChunkCtx);
929	if (pvRet)
930	return (PIEMNATIVEINSTR)pvRet;
931	}
932	void *pvRet = iemExecMemAllocatorAllocInChunkInt(pExecMemAllocator, pbmAlloc, 0,
933	RT_MIN(pExecMemAllocator->cUnitsPerChunk,
934	RT_ALIGN_32(idxHint + cReqUnits, 64*4)),
935	cReqUnits, idxChunk, pTb, (void **)ppaExec, ppChunkCtx);
936	if (pvRet)
937	return (PIEMNATIVEINSTR)pvRet;
938
939	pExecMemAllocator->cFruitlessChunkScans += 1;
940	return NULL;
941	}
942
943
944	DECLINLINE(PIEMNATIVEINSTR)
945	iemExecMemAllocatorAllocBytesInChunk(PIEMEXECMEMALLOCATOR pExecMemAllocator, uint32_t idxChunk, uint32_t cbReq,
946	PIEMNATIVEINSTR *ppaExec)
947	{
948	uint32_t const cReqUnits = iemExecMemAllocBytesToUnits(cbReq);
949	if (cReqUnits <= pExecMemAllocator->aChunks[idxChunk].cFreeUnits)
950	return iemExecMemAllocatorAllocUnitsInChunkInner(pExecMemAllocator, idxChunk, cReqUnits, NULL /pTb/,
951	ppaExec, NULL /ppChunkCtx/);
952	return NULL;
953	}
954
955
956	/**
957	* Allocates @a cbReq bytes of executable memory.
958	*
959	* @returns Pointer to the readable/writeable memory, NULL if out of memory or other problem
960	* encountered.
961	* @param pVCpu The cross context virtual CPU structure of the
962	* calling thread.
963	* @param cbReq How many bytes are required.
964	* @param pTb The translation block that will be using the allocation.
965	* @param ppaExec Where to return the pointer to executable view of
966	* the allocated memory, optional.
967	* @param ppChunkCtx Where to return the per chunk attached context
968	* if available, optional.
969	*/
970	DECLHIDDEN(PIEMNATIVEINSTR) iemExecMemAllocatorAlloc(PVMCPU pVCpu, uint32_t cbReq, PIEMTB pTb,
971	PIEMNATIVEINSTR ppaExec, PCIEMNATIVEPERCHUNKCTX ppChunkCtx) RT_NOEXCEPT
972	{
973	PIEMEXECMEMALLOCATOR pExecMemAllocator = pVCpu->iem.s.pExecMemAllocatorR3;
974	AssertReturn(pExecMemAllocator && pExecMemAllocator->uMagic == IEMEXECMEMALLOCATOR_MAGIC, NULL);
975	AssertMsgReturn(cbReq > 32 && cbReq < _512K, ("%#x\n", cbReq), NULL);
976	STAM_PROFILE_START(&pExecMemAllocator->StatAlloc, a);
977
978	uint32_t const cReqUnits = iemExecMemAllocBytesToUnits(cbReq);
979	STAM_COUNTER_INC(&pExecMemAllocator->aStatSizes[cReqUnits < RT_ELEMENTS(pExecMemAllocator->aStatSizes) ? cReqUnits : 0]);
980	for (unsigned iIteration = 0;; iIteration++)
981	{
982	if ( cbReq * 2 <= pExecMemAllocator->cbFree
983	\|\| (cReqUnits == 1 \|\| pExecMemAllocator->cbFree >= IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE) )
984	{
985	uint32_t const cChunks = pExecMemAllocator->cChunks;
986	uint32_t const idxChunkHint = pExecMemAllocator->idxChunkHint < cChunks ? pExecMemAllocator->idxChunkHint : 0;
987
988	/*
989	* We do two passes here, the first pass we skip chunks with fewer than cReqUnits * 16,
990	* the 2nd pass we skip chunks. The second pass checks the one skipped in the first pass.
991	*/
992	for (uint32_t cMinFreePass = cReqUnits == 1 ? cReqUnits : cReqUnits * 16, cMaxFreePass = UINT32_MAX;;)
993	{
994	for (uint32_t idxChunk = idxChunkHint; idxChunk < cChunks; idxChunk++)
995	if ( pExecMemAllocator->aChunks[idxChunk].cFreeUnits >= cMinFreePass
996	&& pExecMemAllocator->aChunks[idxChunk].cFreeUnits <= cMaxFreePass)
997	{
998	PIEMNATIVEINSTR const pRet = iemExecMemAllocatorAllocUnitsInChunkInner(pExecMemAllocator, idxChunk,
999	cReqUnits, pTb, ppaExec, ppChunkCtx);
1000	if (pRet)
1001	{
1002	STAM_PROFILE_STOP(&pExecMemAllocator->StatAlloc, a);
1003	#ifdef VBOX_WITH_STATISTICS
1004	pExecMemAllocator->cbUnusable += (cReqUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT) - cbReq;
1005	#endif
1006	return pRet;
1007	}
1008	}
1009	for (uint32_t idxChunk = 0; idxChunk < idxChunkHint; idxChunk++)
1010	if ( pExecMemAllocator->aChunks[idxChunk].cFreeUnits >= cMinFreePass
1011	&& pExecMemAllocator->aChunks[idxChunk].cFreeUnits <= cMaxFreePass)
1012	{
1013	PIEMNATIVEINSTR const pRet = iemExecMemAllocatorAllocUnitsInChunkInner(pExecMemAllocator, idxChunk,
1014	cReqUnits, pTb, ppaExec, ppChunkCtx);
1015	if (pRet)
1016	{
1017	STAM_PROFILE_STOP(&pExecMemAllocator->StatAlloc, a);
1018	#ifdef VBOX_WITH_STATISTICS
1019	pExecMemAllocator->cbUnusable += (cReqUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT) - cbReq;
1020	#endif
1021	return pRet;
1022	}
1023	}
1024	if (cMinFreePass <= cReqUnits * 2)
1025	break;
1026	cMaxFreePass = cMinFreePass - 1;
1027	cMinFreePass = cReqUnits * 2;
1028	}
1029	}
1030
1031	/*
1032	* Can we grow it with another chunk?
1033	*/
1034	if (pExecMemAllocator->cChunks < pExecMemAllocator->cMaxChunks)
1035	{
1036	int rc = iemExecMemAllocatorGrow(pVCpu, pExecMemAllocator);
1037	AssertLogRelRCReturn(rc, NULL);
1038
1039	uint32_t const idxChunk = pExecMemAllocator->cChunks - 1;
1040	PIEMNATIVEINSTR const pRet = iemExecMemAllocatorAllocUnitsInChunkInner(pExecMemAllocator, idxChunk, cReqUnits, pTb,
1041	ppaExec, ppChunkCtx);
1042	if (pRet)
1043	{
1044	STAM_PROFILE_STOP(&pExecMemAllocator->StatAlloc, a);
1045	#ifdef VBOX_WITH_STATISTICS
1046	pExecMemAllocator->cbUnusable += (cReqUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT) - cbReq;
1047	#endif
1048	return pRet;
1049	}
1050	AssertFailed();
1051	}
1052
1053	/*
1054	* Try prune native TBs once.
1055	*/
1056	if (iIteration == 0)
1057	{
1058	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING
1059	iemExecMemAllocatorPrune(pVCpu, pExecMemAllocator);
1060	#else
1061	/* No header included in the instruction count here. */
1062	uint32_t const cNeededInstrs = RT_ALIGN_32(cbReq, IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE) / sizeof(IEMNATIVEINSTR);
1063	iemTbAllocatorFreeupNativeSpace(pVCpu, cNeededInstrs);
1064	#endif
1065	}
1066	else
1067	{
1068	STAM_REL_COUNTER_INC(&pVCpu->iem.s.StatNativeExecMemInstrBufAllocFailed);
1069	STAM_PROFILE_STOP(&pExecMemAllocator->StatAlloc, a);
1070	return NULL;
1071	}
1072	}
1073	}
1074
1075
1076	/** This is a hook to ensure the instruction cache is properly flushed before the code in the memory
1077	* given by @a pv and @a cb is executed */
1078	DECLHIDDEN(void) iemExecMemAllocatorReadyForUse(PVMCPUCC pVCpu, void *pv, size_t cb) RT_NOEXCEPT
1079	{
1080	#ifdef RT_OS_DARWIN
1081	/*
1082	* We need to synchronize the stuff we wrote to the data cache with the
1083	* instruction cache, since these aren't coherent on arm (or at least not
1084	* on Apple Mn CPUs).
1085	*
1086	* Note! Since we don't any share JIT'ed code with the other CPUs, we don't
1087	* really care whether the dcache is fully flushed back to memory. It
1088	* only needs to hit the level 2 cache, which the level 1 instruction
1089	* and data caches seems to be sharing. In ARM terms, we need to reach
1090	* a point of unification (PoU), rather than a point of coherhency (PoC).
1091	*
1092	* https://developer.apple.com/documentation/apple-silicon/porting-just-in-time-compilers-to-apple-silicon
1093	*
1094	* https://developer.arm.com/documentation/den0013/d/Caches/Point-of-coherency-and-unification
1095	*
1096	* Experimenting with the approach used by sys_icache_invalidate() and
1097	* tweaking it a little, could let us shave off a bit of effort. The thing
1098	* that slows the apple code down on an M2 (runing Sonoma 13.4), seems to
1099	* the 'DSB ISH' instructions performed every 20 icache line flushes.
1100	* Skipping these saves ~100ns or more per TB when profiling the native
1101	* recompiler on the TBs from a win11 full boot-desktop-shutdow sequence.
1102	* Thus we will leave DCACHE_ICACHE_SYNC_WITH_WITH_IVAU_DSB undefined if we
1103	* can.
1104	*
1105	* There appears not to be much difference between DSB options 'ISH',
1106	* 'ISHST', 'NSH' and 'NSHST'. The latter is theoretically all we need, so
1107	* we'll use that one.
1108	*
1109	* See https://developer.arm.com/documentation/100941/0101/Barriers for
1110	* details on the barrier options.
1111	*
1112	* Note! The CFG value "/IEM/HostICacheInvalidationViaHostAPI" can be used
1113	* to disabling the experimental code should it misbehave.
1114	*/
1115	uint8_t const fHostICacheInvalidation = pVCpu->iem.s.fHostICacheInvalidation;
1116	if (!(fHostICacheInvalidation & IEMNATIVE_ICACHE_F_USE_HOST_API))
1117	{
1118	# define DCACHE_ICACHE_SYNC_DSB_OPTION "nshst"
1119	/# define DCACHE_ICACHE_SYNC_WITH_WITH_IVAU_DSB/
1120
1121	/* Skipping this is fine, but doesn't impact perf much. */
1122	__asm__ __volatile__("dsb " DCACHE_ICACHE_SYNC_DSB_OPTION);
1123
1124	/* Invalidate the icache for the range [pv,pv+cb). */
1125	# ifdef DCACHE_ICACHE_SYNC_WITH_WITH_IVAU_DSB
1126	size_t const cIvauDsbEvery= 20;
1127	unsigned cDsb = cIvauDsbEvery;
1128	# endif
1129	size_t const cbCacheLine = 64;
1130	size_t cbInvalidate = cb + ((uintptr_t)pv & (cbCacheLine - 1)) ;
1131	size_t cCacheLines = RT_ALIGN_Z(cbInvalidate, cbCacheLine) / cbCacheLine;
1132	uintptr_t uPtr = (uintptr_t)pv & ~(uintptr_t)(cbCacheLine - 1);
1133	for (;; uPtr += cbCacheLine)
1134	{
1135	__asm__ /__volatile__/("ic ivau, %0" : : "r" (uPtr));
1136	cCacheLines -= 1;
1137	if (!cCacheLines)
1138	break;
1139	# ifdef DCACHE_ICACHE_SYNC_WITH_WITH_IVAU_DSB
1140	cDsb -= 1;
1141	if (cDsb != 0)
1142	{ /* likely */ }
1143	else
1144	{
1145	__asm__ __volatile__("dsb " DCACHE_ICACHE_SYNC_DSB_OPTION);
1146	cDsb = cIvauDsbEvery;
1147	}
1148	# endif
1149	}
1150
1151	/*
1152	* The DSB here is non-optional it seems.
1153	*
1154	* The following ISB can be omitted on M2 without any obvious sideeffects,
1155	* it produces better number in the above mention profiling scenario.
1156	* This could be related to the kHasICDSB flag in cpu_capabilities.h,
1157	* but it doesn't look like that flag is set here (M2, Sonoma 13.4).
1158	*
1159	* I've made the inclusion of the ISH barrier as configurable and with
1160	* a default of skipping it.
1161	*/
1162	if (!(fHostICacheInvalidation & IEMNATIVE_ICACHE_F_END_WITH_ISH))
1163	__asm__ __volatile__("dsb " DCACHE_ICACHE_SYNC_DSB_OPTION
1164	::: "memory");
1165	else
1166	__asm__ __volatile__("dsb " DCACHE_ICACHE_SYNC_DSB_OPTION "\n\t"
1167	"isb"
1168	::: "memory");
1169	}
1170	else
1171	sys_icache_invalidate(pv, cb);
1172
1173	#elif defined(RT_OS_LINUX) && defined(RT_ARCH_ARM64)
1174	RT_NOREF(pVCpu);
1175
1176	/* There is __builtin___clear_cache() but it flushes both the instruction and data cache, so do it manually. */
1177	static uint32_t s_u32CtrEl0 = 0;
1178	if (!s_u32CtrEl0)
1179	asm volatile ("mrs %0, ctr_el0":"=r" (s_u32CtrEl0));
1180	uintptr_t cbICacheLine = (uintptr_t)4 << (s_u32CtrEl0 & 0xf);
1181
1182	uintptr_t pb = (uintptr_t)pv & ~(cbICacheLine - 1);
1183	for (; pb < (uintptr_t)pv + cb; pb += cbICacheLine)
1184	asm volatile ("ic ivau, %0" : : "r" (pb) : "memory");
1185
1186	asm volatile ("dsb ish\n\t isb\n\t" : : : "memory");
1187
1188	#else
1189	RT_NOREF(pVCpu, pv, cb);
1190	#endif
1191	}
1192
1193
1194	/**
1195	* Frees executable memory.
1196	*/
1197	DECLHIDDEN(void) iemExecMemAllocatorFree(PVMCPU pVCpu, void *pv, size_t cb) RT_NOEXCEPT
1198	{
1199	PIEMEXECMEMALLOCATOR pExecMemAllocator = pVCpu->iem.s.pExecMemAllocatorR3;
1200	Assert(pExecMemAllocator && pExecMemAllocator->uMagic == IEMEXECMEMALLOCATOR_MAGIC);
1201	AssertPtr(pv);
1202	#ifdef VBOX_WITH_STATISTICS
1203	size_t const cbOrig = cb;
1204	#endif
1205	#ifndef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
1206	Assert(!((uintptr_t)pv & (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1)));
1207
1208	/* Align the size as we did when allocating the block. */
1209	cb = RT_ALIGN_Z(cb, IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE);
1210
1211	#else
1212	PIEMEXECMEMALLOCHDR pHdr = (PIEMEXECMEMALLOCHDR)pv - 1;
1213	Assert(!((uintptr_t)pHdr & (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1)));
1214	AssertReturnVoid(pHdr->uMagic == IEMEXECMEMALLOCHDR_MAGIC);
1215	uint32_t const idxChunk = pHdr->idxChunk;
1216	AssertReturnVoid(idxChunk < pExecMemAllocator->cChunks);
1217	pv = pHdr;
1218
1219	/* Adjust and align the size to cover the whole allocation area. */
1220	cb = RT_ALIGN_Z(cb + sizeof(*pHdr), IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE);
1221	#endif
1222
1223	/* Free it / assert sanity. */
1224	bool fFound = false;
1225	uint32_t const cbChunk = pExecMemAllocator->cbChunk;
1226	#ifndef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
1227	uint32_t const cChunks = pExecMemAllocator->cChunks;
1228	for (uint32_t idxChunk = 0; idxChunk < cChunks; idxChunk++)
1229	#endif
1230	{
1231	uintptr_t const offChunk = (uintptr_t)pv - (uintptr_t)pExecMemAllocator->aChunks[idxChunk].pvChunkRx;
1232	fFound = offChunk < cbChunk;
1233	if (fFound)
1234	{
1235	uint32_t const idxFirst = (uint32_t)offChunk >> IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
1236	uint32_t const cReqUnits = (uint32_t)cb >> IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
1237
1238	/* Check that it's valid and free it. */
1239	uint64_t * const pbmAlloc = &pExecMemAllocator->pbmAlloc[pExecMemAllocator->cBitmapElementsPerChunk * idxChunk];
1240	AssertReturnVoid(ASMBitTest(pbmAlloc, idxFirst));
1241	for (uint32_t i = 1; i < cReqUnits; i++)
1242	AssertReturnVoid(ASMBitTest(pbmAlloc, idxFirst + i));
1243	ASMBitClearRange(pbmAlloc, idxFirst, idxFirst + cReqUnits);
1244
1245	/* Invalidate the header using the writeable memory view. */
1246	pHdr = (PIEMEXECMEMALLOCHDR)((uintptr_t)pExecMemAllocator->aChunks[idxChunk].pvChunkRw + offChunk);
1247	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
1248	pHdr->uMagic = 0;
1249	pHdr->idxChunk = 0;
1250	pHdr->pTb = NULL;
1251	#endif
1252	pExecMemAllocator->aChunks[idxChunk].cFreeUnits += cReqUnits;
1253	pExecMemAllocator->aChunks[idxChunk].idxFreeHint = idxFirst;
1254
1255	/* Update the stats. */
1256	pExecMemAllocator->cbAllocated -= cb;
1257	pExecMemAllocator->cbFree += cb;
1258	pExecMemAllocator->cAllocations -= 1;
1259	#ifdef VBOX_WITH_STATISTICS
1260	pExecMemAllocator->cbUnusable -= (cReqUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT) - cbOrig;
1261	#endif
1262	return;
1263	}
1264	}
1265	AssertFailed();
1266	}
1267
1268
1269	/**
1270	* Interface used by iemNativeRecompileAttachExecMemChunkCtx and unwind info
1271	* generators.
1272	*/
1273	DECLHIDDEN(PIEMNATIVEINSTR)
1274	iemExecMemAllocatorAllocFromChunk(PVMCPU pVCpu, uint32_t idxChunk, uint32_t cbReq, PIEMNATIVEINSTR *ppaExec)
1275	{
1276	PIEMEXECMEMALLOCATOR pExecMemAllocator = pVCpu->iem.s.pExecMemAllocatorR3;
1277	AssertReturn(idxChunk < pExecMemAllocator->cChunks, NULL);
1278	Assert(cbReq < _1M);
1279	return iemExecMemAllocatorAllocBytesInChunk(pExecMemAllocator, idxChunk, cbReq, ppaExec);
1280	}
1281
1282
1283	/**
1284	* For getting the per-chunk context detailing common code for a TB.
1285	*
1286	* This is for use by the disassembler.
1287	*/
1288	DECLHIDDEN(PCIEMNATIVEPERCHUNKCTX) iemExecMemGetTbChunkCtx(PVMCPU pVCpu, PCIEMTB pTb)
1289	{
1290	PIEMEXECMEMALLOCATOR pExecMemAllocator = pVCpu->iem.s.pExecMemAllocatorR3;
1291	if ((pTb->fFlags & IEMTB_F_TYPE_MASK) == IEMTB_F_TYPE_NATIVE)
1292	{
1293	uintptr_t const uAddress = (uintptr_t)pTb->Native.paInstructions;
1294	uint32_t const cbChunk = pExecMemAllocator->cbChunk;
1295	uint32_t idxChunk = pExecMemAllocator->cChunks;
1296	while (idxChunk-- > 0)
1297	if (uAddress - (uintptr_t)pExecMemAllocator->aChunks[idxChunk].pvChunkRx < cbChunk)
1298	return pExecMemAllocator->aChunks[idxChunk].pCtx;
1299	}
1300	return NULL;
1301	}
1302
1303
1304	#ifdef IN_RING3
1305	# ifdef RT_OS_WINDOWS
1306
1307	/**
1308	* Initializes the unwind info structures for windows hosts.
1309	*/
1310	static int
1311	iemExecMemAllocatorInitAndRegisterUnwindInfoForChunk(PVMCPUCC pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator,
1312	void *pvChunk, uint32_t idxChunk)
1313	{
1314	RT_NOREF(pVCpu);
1315
1316	/*
1317	* The AMD64 unwind opcodes.
1318	*
1319	* This is a program that starts with RSP after a RET instruction that
1320	* ends up in recompiled code, and the operations we describe here will
1321	* restore all non-volatile registers and bring RSP back to where our
1322	* RET address is. This means it's reverse order from what happens in
1323	* the prologue.
1324	*
1325	* Note! Using a frame register approach here both because we have one
1326	* and but mainly because the UWOP_ALLOC_LARGE argument values
1327	* would be a pain to write initializers for. On the positive
1328	* side, we're impervious to changes in the the stack variable
1329	* area can can deal with dynamic stack allocations if necessary.
1330	*/
1331	static const IMAGE_UNWIND_CODE s_aOpcodes[] =
1332	{
1333	{ { 16, IMAGE_AMD64_UWOP_SET_FPREG, 0 } }, /* RSP = RBP - FrameOffset * 10 (0x60) */
1334	{ { 16, IMAGE_AMD64_UWOP_ALLOC_SMALL, 0 } }, /* RSP += 8; */
1335	{ { 14, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_x15 } }, /* R15 = [RSP]; RSP += 8; */
1336	{ { 12, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_x14 } }, /* R14 = [RSP]; RSP += 8; */
1337	{ { 10, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_x13 } }, /* R13 = [RSP]; RSP += 8; */
1338	{ { 8, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_x12 } }, /* R12 = [RSP]; RSP += 8; */
1339	{ { 7, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_xDI } }, /* RDI = [RSP]; RSP += 8; */
1340	{ { 6, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_xSI } }, /* RSI = [RSP]; RSP += 8; */
1341	{ { 5, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_xBX } }, /* RBX = [RSP]; RSP += 8; */
1342	{ { 4, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_xBP } }, /* RBP = [RSP]; RSP += 8; */
1343	};
1344	union
1345	{
1346	IMAGE_UNWIND_INFO Info;
1347	uint8_t abPadding[RT_UOFFSETOF(IMAGE_UNWIND_INFO, aOpcodes) + 16];
1348	} s_UnwindInfo =
1349	{
1350	{
1351	/* .Version = */ 1,
1352	/* .Flags = */ 0,
1353	/* .SizeOfProlog = / 16, / whatever */
1354	/* .CountOfCodes = */ RT_ELEMENTS(s_aOpcodes),
1355	/* .FrameRegister = */ X86_GREG_xBP,
1356	/* .FrameOffset = / (-IEMNATIVE_FP_OFF_LAST_PUSH + 8) / 16 / we're off by one slot. sigh. */,
1357	}
1358	};
1359	AssertCompile(-IEMNATIVE_FP_OFF_LAST_PUSH < 240 && -IEMNATIVE_FP_OFF_LAST_PUSH > 0);
1360	AssertCompile((-IEMNATIVE_FP_OFF_LAST_PUSH & 0xf) == 8);
1361
1362	/*
1363	* Calc how much space we need and allocate it off the exec heap.
1364	*/
1365	unsigned const cFunctionEntries = 1;
1366	unsigned const cbUnwindInfo = sizeof(s_aOpcodes) + RT_UOFFSETOF(IMAGE_UNWIND_INFO, aOpcodes);
1367	unsigned const cbNeeded = sizeof(IMAGE_RUNTIME_FUNCTION_ENTRY) * cFunctionEntries + cbUnwindInfo;
1368	PIMAGE_RUNTIME_FUNCTION_ENTRY const paFunctions
1369	= (PIMAGE_RUNTIME_FUNCTION_ENTRY)iemExecMemAllocatorAllocBytesInChunk(pExecMemAllocator, idxChunk, cbNeeded, NULL);
1370	AssertReturn(paFunctions, VERR_INTERNAL_ERROR_5);
1371	pExecMemAllocator->aChunks[idxChunk].pvUnwindInfo = paFunctions;
1372
1373	/*
1374	* Initialize the structures.
1375	*/
1376	PIMAGE_UNWIND_INFO const pInfo = (PIMAGE_UNWIND_INFO)&paFunctions[cFunctionEntries];
1377
1378	paFunctions[0].BeginAddress = 0;
1379	paFunctions[0].EndAddress = pExecMemAllocator->cbChunk;
1380	paFunctions[0].UnwindInfoAddress = (uint32_t)((uintptr_t)pInfo - (uintptr_t)pvChunk);
1381
1382	memcpy(pInfo, &s_UnwindInfo, RT_UOFFSETOF(IMAGE_UNWIND_INFO, aOpcodes));
1383	memcpy(&pInfo->aOpcodes[0], s_aOpcodes, sizeof(s_aOpcodes));
1384
1385	/*
1386	* Register it.
1387	*/
1388	uint8_t fRet = RtlAddFunctionTable(paFunctions, cFunctionEntries, (uintptr_t)pvChunk);
1389	AssertReturn(fRet, VERR_INTERNAL_ERROR_3); /* Nothing to clean up on failure, since its within the chunk itself. */
1390
1391	return VINF_SUCCESS;
1392	}
1393
1394
1395	# else /* !RT_OS_WINDOWS */
1396
1397	/**
1398	* Emits a LEB128 encoded value between -0x2000 and 0x2000 (both exclusive).
1399	*/
1400	DECLINLINE(RTPTRUNION) iemDwarfPutLeb128(RTPTRUNION Ptr, int32_t iValue)
1401	{
1402	if (iValue >= 64)
1403	{
1404	Assert(iValue < 0x2000);
1405	*Ptr.pb++ = ((uint8_t)iValue & 0x7f) \| 0x80;
1406	*Ptr.pb++ = (uint8_t)(iValue >> 7) & 0x3f;
1407	}
1408	else if (iValue >= 0)
1409	*Ptr.pb++ = (uint8_t)iValue;
1410	else if (iValue > -64)
1411	*Ptr.pb++ = ((uint8_t)iValue & 0x3f) \| 0x40;
1412	else
1413	{
1414	Assert(iValue > -0x2000);
1415	*Ptr.pb++ = ((uint8_t)iValue & 0x7f) \| 0x80;
1416	*Ptr.pb++ = ((uint8_t)(iValue >> 7) & 0x3f) \| 0x40;
1417	}
1418	return Ptr;
1419	}
1420
1421
1422	/**
1423	* Emits an ULEB128 encoded value (up to 64-bit wide).
1424	*/
1425	DECLINLINE(RTPTRUNION) iemDwarfPutUleb128(RTPTRUNION Ptr, uint64_t uValue)
1426	{
1427	while (uValue >= 0x80)
1428	{
1429	*Ptr.pb++ = ((uint8_t)uValue & 0x7f) \| 0x80;
1430	uValue >>= 7;
1431	}
1432	*Ptr.pb++ = (uint8_t)uValue;
1433	return Ptr;
1434	}
1435
1436
1437	/**
1438	* Emits a CFA rule as register @a uReg + offset @a off.
1439	*/
1440	DECLINLINE(RTPTRUNION) iemDwarfPutCfaDefCfa(RTPTRUNION Ptr, uint32_t uReg, uint32_t off)
1441	{
1442	*Ptr.pb++ = DW_CFA_def_cfa;
1443	Ptr = iemDwarfPutUleb128(Ptr, uReg);
1444	Ptr = iemDwarfPutUleb128(Ptr, off);
1445	return Ptr;
1446	}
1447
1448
1449	/**
1450	* Emits a register (@a uReg) save location:
1451	* CFA + @a off * data_alignment_factor
1452	*/
1453	DECLINLINE(RTPTRUNION) iemDwarfPutCfaOffset(RTPTRUNION Ptr, uint32_t uReg, uint32_t off)
1454	{
1455	if (uReg < 0x40)
1456	*Ptr.pb++ = DW_CFA_offset \| uReg;
1457	else
1458	{
1459	*Ptr.pb++ = DW_CFA_offset_extended;
1460	Ptr = iemDwarfPutUleb128(Ptr, uReg);
1461	}
1462	Ptr = iemDwarfPutUleb128(Ptr, off);
1463	return Ptr;
1464	}
1465
1466
1467	# if 0 /* unused */
1468	/**
1469	* Emits a register (@a uReg) save location, using signed offset:
1470	* CFA + @a offSigned * data_alignment_factor
1471	*/
1472	DECLINLINE(RTPTRUNION) iemDwarfPutCfaSignedOffset(RTPTRUNION Ptr, uint32_t uReg, int32_t offSigned)
1473	{
1474	*Ptr.pb++ = DW_CFA_offset_extended_sf;
1475	Ptr = iemDwarfPutUleb128(Ptr, uReg);
1476	Ptr = iemDwarfPutLeb128(Ptr, offSigned);
1477	return Ptr;
1478	}
1479	# endif
1480
1481
1482	/**
1483	* Initializes the unwind info section for non-windows hosts.
1484	*/
1485	static int
1486	iemExecMemAllocatorInitAndRegisterUnwindInfoForChunk(PVMCPUCC pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator,
1487	void *pvChunk, uint32_t idxChunk)
1488	{
1489	PIEMEXECMEMCHUNKEHFRAME const pEhFrame = &pExecMemAllocator->paEhFrames[idxChunk];
1490	pExecMemAllocator->aChunks[idxChunk].pvUnwindInfo = pEhFrame; /* not necessary, but whatever */
1491
1492	RTPTRUNION Ptr = { pEhFrame->abEhFrame };
1493
1494	/*
1495	* Generate the CIE first.
1496	*/
1497	# ifdef IEMNATIVE_USE_LIBUNWIND /* libunwind (llvm, darwin) only supports v1 and v3. */
1498	uint8_t const iDwarfVer = 3;
1499	# else
1500	uint8_t const iDwarfVer = 4;
1501	# endif
1502	RTPTRUNION const PtrCie = Ptr;
1503	Ptr.pu32++ = 123; / The CIE length will be determined later. */
1504	Ptr.pu32++ = 0 /UINT32_MAX/; / I'm a CIE in .eh_frame speak. */
1505	Ptr.pb++ = iDwarfVer; / DwARF version */
1506	Ptr.pb++ = 0; / Augmentation. */
1507	if (iDwarfVer >= 4)
1508	{
1509	Ptr.pb++ = sizeof(uintptr_t); / Address size. */
1510	Ptr.pb++ = 0; / Segment selector size. */
1511	}
1512	# ifdef RT_ARCH_AMD64
1513	Ptr = iemDwarfPutLeb128(Ptr, 1); /* Code alignment factor (LEB128 = 1). */
1514	# else
1515	Ptr = iemDwarfPutLeb128(Ptr, 4); /* Code alignment factor (LEB128 = 4). */
1516	# endif
1517	Ptr = iemDwarfPutLeb128(Ptr, -8); /* Data alignment factor (LEB128 = -8). */
1518	# ifdef RT_ARCH_AMD64
1519	Ptr = iemDwarfPutUleb128(Ptr, DWREG_AMD64_RA); /* Return address column (ULEB128) */
1520	# elif defined(RT_ARCH_ARM64)
1521	Ptr = iemDwarfPutUleb128(Ptr, DWREG_ARM64_LR); /* Return address column (ULEB128) */
1522	# else
1523	# error "port me"
1524	# endif
1525	/* Initial instructions: */
1526	# ifdef RT_ARCH_AMD64
1527	Ptr = iemDwarfPutCfaDefCfa(Ptr, DWREG_AMD64_RBP, 16); /* CFA = RBP + 0x10 - first stack parameter */
1528	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_RA, 1); /* Ret RIP = [CFA + 1-8] /
1529	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_RBP, 2); /* RBP = [CFA + 2-8] /
1530	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_RBX, 3); /* RBX = [CFA + 3-8] /
1531	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_R12, 4); /* R12 = [CFA + 4-8] /
1532	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_R13, 5); /* R13 = [CFA + 5-8] /
1533	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_R14, 6); /* R14 = [CFA + 6-8] /
1534	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_R15, 7); /* R15 = [CFA + 7-8] /
1535	# elif defined(RT_ARCH_ARM64)
1536	# if 1
1537	Ptr = iemDwarfPutCfaDefCfa(Ptr, DWREG_ARM64_BP, 16); /* CFA = BP + 0x10 - first stack parameter */
1538	# else
1539	Ptr = iemDwarfPutCfaDefCfa(Ptr, DWREG_ARM64_SP, IEMNATIVE_FRAME_VAR_SIZE + IEMNATIVE_FRAME_SAVE_REG_SIZE);
1540	# endif
1541	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_LR, 1); /* Ret PC = [CFA + 1-8] /
1542	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_BP, 2); /* Ret BP = [CFA + 2-8] /
1543	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X28, 3); /* X28 = [CFA + 3-8] /
1544	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X27, 4); /* X27 = [CFA + 4-8] /
1545	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X26, 5); /* X26 = [CFA + 5-8] /
1546	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X25, 6); /* X25 = [CFA + 6-8] /
1547	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X24, 7); /* X24 = [CFA + 7-8] /
1548	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X23, 8); /* X23 = [CFA + 8-8] /
1549	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X22, 9); /* X22 = [CFA + 9-8] /
1550	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X21, 10); /* X21 = [CFA +10-8] /
1551	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X20, 11); /* X20 = [CFA +11-8] /
1552	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X19, 12); /* X19 = [CFA +12-8] /
1553	AssertCompile(IEMNATIVE_FRAME_SAVE_REG_SIZE / 8 == 12);
1554	/** @todo we we need to do something about clearing DWREG_ARM64_RA_SIGN_STATE or something? */
1555	# else
1556	# error "port me"
1557	# endif
1558	while ((Ptr.u - PtrCie.u) & 3)
1559	*Ptr.pb++ = DW_CFA_nop;
1560	/* Finalize the CIE size. */
1561	*PtrCie.pu32 = Ptr.u - PtrCie.u - sizeof(uint32_t);
1562
1563	/*
1564	* Generate an FDE for the whole chunk area.
1565	*/
1566	# ifdef IEMNATIVE_USE_LIBUNWIND
1567	pEhFrame->offFda = Ptr.u - (uintptr_t)&pEhFrame->abEhFrame[0];
1568	# endif
1569	RTPTRUNION const PtrFde = Ptr;
1570	Ptr.pu32++ = 123; / The CIE length will be determined later. */
1571	Ptr.pu32 = Ptr.u - PtrCie.u; / Negated self relative CIE address. */
1572	Ptr.pu32++;
1573	Ptr.pu64++ = (uintptr_t)pvChunk; / Absolute start PC of this FDE. */
1574	Ptr.pu64++ = pExecMemAllocator->cbChunk; / PC range length for this PDE. */
1575	# if 0 /* not requried for recent libunwind.dylib nor recent libgcc/glib. */
1576	*Ptr.pb++ = DW_CFA_nop;
1577	# endif
1578	while ((Ptr.u - PtrFde.u) & 3)
1579	*Ptr.pb++ = DW_CFA_nop;
1580	/* Finalize the FDE size. */
1581	*PtrFde.pu32 = Ptr.u - PtrFde.u - sizeof(uint32_t);
1582
1583	/* Terminator entry. */
1584	*Ptr.pu32++ = 0;
1585	Ptr.pu32++ = 0; / just to be sure... */
1586	Assert(Ptr.u - (uintptr_t)&pEhFrame->abEhFrame[0] <= sizeof(pEhFrame->abEhFrame));
1587
1588	/*
1589	* Register it.
1590	*/
1591	# ifdef IEMNATIVE_USE_LIBUNWIND
1592	__register_frame(&pEhFrame->abEhFrame[pEhFrame->offFda]);
1593	# else
1594	memset(pEhFrame->abObject, 0xf6, sizeof(pEhFrame->abObject)); /* color the memory to better spot usage */
1595	__register_frame_info(pEhFrame->abEhFrame, pEhFrame->abObject);
1596	# endif
1597
1598	# ifdef IEMNATIVE_USE_GDB_JIT
1599	/*
1600	* Now for telling GDB about this (experimental).
1601	*
1602	* This seems to work best with ET_DYN.
1603	*/
1604	GDBJITSYMFILE * const pSymFile = (GDBJITSYMFILE *)iemExecMemAllocatorAllocBytesInChunk(pExecMemAllocator, idxChunk,
1605	sizeof(GDBJITSYMFILE), NULL);
1606	AssertReturn(pSymFile, VERR_INTERNAL_ERROR_5);
1607	unsigned const offSymFileInChunk = (uintptr_t)pSymFile - (uintptr_t)pvChunk;
1608
1609	RT_ZERO(*pSymFile);
1610
1611	/*
1612	* The ELF header:
1613	*/
1614	pSymFile->EHdr.e_ident[0] = ELFMAG0;
1615	pSymFile->EHdr.e_ident[1] = ELFMAG1;
1616	pSymFile->EHdr.e_ident[2] = ELFMAG2;
1617	pSymFile->EHdr.e_ident[3] = ELFMAG3;
1618	pSymFile->EHdr.e_ident[EI_VERSION] = EV_CURRENT;
1619	pSymFile->EHdr.e_ident[EI_CLASS] = ELFCLASS64;
1620	pSymFile->EHdr.e_ident[EI_DATA] = ELFDATA2LSB;
1621	pSymFile->EHdr.e_ident[EI_OSABI] = ELFOSABI_NONE;
1622	# ifdef IEMNATIVE_USE_GDB_JIT_ET_DYN
1623	pSymFile->EHdr.e_type = ET_DYN;
1624	# else
1625	pSymFile->EHdr.e_type = ET_REL;
1626	# endif
1627	# ifdef RT_ARCH_AMD64
1628	pSymFile->EHdr.e_machine = EM_AMD64;
1629	# elif defined(RT_ARCH_ARM64)
1630	pSymFile->EHdr.e_machine = EM_AARCH64;
1631	# else
1632	# error "port me"
1633	# endif
1634	pSymFile->EHdr.e_version = 1; /?/
1635	pSymFile->EHdr.e_entry = 0;
1636	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1637	pSymFile->EHdr.e_phoff = RT_UOFFSETOF(GDBJITSYMFILE, aPhdrs);
1638	# else
1639	pSymFile->EHdr.e_phoff = 0;
1640	# endif
1641	pSymFile->EHdr.e_shoff = sizeof(pSymFile->EHdr);
1642	pSymFile->EHdr.e_flags = 0;
1643	pSymFile->EHdr.e_ehsize = sizeof(pSymFile->EHdr);
1644	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1645	pSymFile->EHdr.e_phentsize = sizeof(pSymFile->aPhdrs[0]);
1646	pSymFile->EHdr.e_phnum = RT_ELEMENTS(pSymFile->aPhdrs);
1647	# else
1648	pSymFile->EHdr.e_phentsize = 0;
1649	pSymFile->EHdr.e_phnum = 0;
1650	# endif
1651	pSymFile->EHdr.e_shentsize = sizeof(pSymFile->aShdrs[0]);
1652	pSymFile->EHdr.e_shnum = RT_ELEMENTS(pSymFile->aShdrs);
1653	pSymFile->EHdr.e_shstrndx = 0; /* set later */
1654
1655	uint32_t offStrTab = 0;
1656	#define APPEND_STR(a_szStr) do { \
1657	memcpy(&pSymFile->szzStrTab[offStrTab], a_szStr, sizeof(a_szStr)); \
1658	offStrTab += sizeof(a_szStr); \
1659	Assert(offStrTab < sizeof(pSymFile->szzStrTab)); \
1660	} while (0)
1661	#define APPEND_STR_FMT(a_szStr, ...) do { \
1662	offStrTab += RTStrPrintf(&pSymFile->szzStrTab[offStrTab], sizeof(pSymFile->szzStrTab) - offStrTab, a_szStr, __VA_ARGS__); \
1663	offStrTab++; \
1664	Assert(offStrTab < sizeof(pSymFile->szzStrTab)); \
1665	} while (0)
1666
1667	/*
1668	* Section headers.
1669	*/
1670	/* Section header #0: NULL */
1671	unsigned i = 0;
1672	APPEND_STR("");
1673	RT_ZERO(pSymFile->aShdrs[i]);
1674	i++;
1675
1676	/* Section header: .eh_frame */
1677	pSymFile->aShdrs[i].sh_name = offStrTab;
1678	APPEND_STR(".eh_frame");
1679	pSymFile->aShdrs[i].sh_type = SHT_PROGBITS;
1680	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC \| SHF_EXECINSTR;
1681	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN) \|\| defined(IEMNATIVE_USE_GDB_JIT_ELF_RVAS)
1682	pSymFile->aShdrs[i].sh_offset
1683	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, abEhFrame);
1684	# else
1685	pSymFile->aShdrs[i].sh_addr = (uintptr_t)&pSymFile->abEhFrame[0];
1686	pSymFile->aShdrs[i].sh_offset = 0;
1687	# endif
1688
1689	pSymFile->aShdrs[i].sh_size = sizeof(pEhFrame->abEhFrame);
1690	pSymFile->aShdrs[i].sh_link = 0;
1691	pSymFile->aShdrs[i].sh_info = 0;
1692	pSymFile->aShdrs[i].sh_addralign = 1;
1693	pSymFile->aShdrs[i].sh_entsize = 0;
1694	memcpy(pSymFile->abEhFrame, pEhFrame->abEhFrame, sizeof(pEhFrame->abEhFrame));
1695	i++;
1696
1697	/* Section header: .shstrtab */
1698	unsigned const iShStrTab = i;
1699	pSymFile->EHdr.e_shstrndx = iShStrTab;
1700	pSymFile->aShdrs[i].sh_name = offStrTab;
1701	APPEND_STR(".shstrtab");
1702	pSymFile->aShdrs[i].sh_type = SHT_STRTAB;
1703	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC;
1704	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN) \|\| defined(IEMNATIVE_USE_GDB_JIT_ELF_RVAS)
1705	pSymFile->aShdrs[i].sh_offset
1706	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, szzStrTab);
1707	# else
1708	pSymFile->aShdrs[i].sh_addr = (uintptr_t)&pSymFile->szzStrTab[0];
1709	pSymFile->aShdrs[i].sh_offset = 0;
1710	# endif
1711	pSymFile->aShdrs[i].sh_size = sizeof(pSymFile->szzStrTab);
1712	pSymFile->aShdrs[i].sh_link = 0;
1713	pSymFile->aShdrs[i].sh_info = 0;
1714	pSymFile->aShdrs[i].sh_addralign = 1;
1715	pSymFile->aShdrs[i].sh_entsize = 0;
1716	i++;
1717
1718	/* Section header: .symbols */
1719	pSymFile->aShdrs[i].sh_name = offStrTab;
1720	APPEND_STR(".symtab");
1721	pSymFile->aShdrs[i].sh_type = SHT_SYMTAB;
1722	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC;
1723	pSymFile->aShdrs[i].sh_offset
1724	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, aSymbols);
1725	pSymFile->aShdrs[i].sh_size = sizeof(pSymFile->aSymbols);
1726	pSymFile->aShdrs[i].sh_link = iShStrTab;
1727	pSymFile->aShdrs[i].sh_info = RT_ELEMENTS(pSymFile->aSymbols);
1728	pSymFile->aShdrs[i].sh_addralign = sizeof(pSymFile->aSymbols[0].st_value);
1729	pSymFile->aShdrs[i].sh_entsize = sizeof(pSymFile->aSymbols[0]);
1730	i++;
1731
1732	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1733	/* Section header: .symbols */
1734	pSymFile->aShdrs[i].sh_name = offStrTab;
1735	APPEND_STR(".dynsym");
1736	pSymFile->aShdrs[i].sh_type = SHT_DYNSYM;
1737	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC;
1738	pSymFile->aShdrs[i].sh_offset
1739	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, aDynSyms);
1740	pSymFile->aShdrs[i].sh_size = sizeof(pSymFile->aDynSyms);
1741	pSymFile->aShdrs[i].sh_link = iShStrTab;
1742	pSymFile->aShdrs[i].sh_info = RT_ELEMENTS(pSymFile->aDynSyms);
1743	pSymFile->aShdrs[i].sh_addralign = sizeof(pSymFile->aDynSyms[0].st_value);
1744	pSymFile->aShdrs[i].sh_entsize = sizeof(pSymFile->aDynSyms[0]);
1745	i++;
1746	# endif
1747
1748	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1749	/* Section header: .dynamic */
1750	pSymFile->aShdrs[i].sh_name = offStrTab;
1751	APPEND_STR(".dynamic");
1752	pSymFile->aShdrs[i].sh_type = SHT_DYNAMIC;
1753	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC;
1754	pSymFile->aShdrs[i].sh_offset
1755	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, aDyn);
1756	pSymFile->aShdrs[i].sh_size = sizeof(pSymFile->aDyn);
1757	pSymFile->aShdrs[i].sh_link = iShStrTab;
1758	pSymFile->aShdrs[i].sh_info = 0;
1759	pSymFile->aShdrs[i].sh_addralign = 1;
1760	pSymFile->aShdrs[i].sh_entsize = sizeof(pSymFile->aDyn[0]);
1761	i++;
1762	# endif
1763
1764	/* Section header: .text */
1765	unsigned const iShText = i;
1766	pSymFile->aShdrs[i].sh_name = offStrTab;
1767	APPEND_STR(".text");
1768	pSymFile->aShdrs[i].sh_type = SHT_PROGBITS;
1769	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC \| SHF_EXECINSTR;
1770	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN) \|\| defined(IEMNATIVE_USE_GDB_JIT_ELF_RVAS)
1771	pSymFile->aShdrs[i].sh_offset
1772	= pSymFile->aShdrs[i].sh_addr = sizeof(GDBJITSYMFILE);
1773	# else
1774	pSymFile->aShdrs[i].sh_addr = (uintptr_t)(pSymFile + 1);
1775	pSymFile->aShdrs[i].sh_offset = 0;
1776	# endif
1777	pSymFile->aShdrs[i].sh_size = pExecMemAllocator->cbChunk - offSymFileInChunk - sizeof(GDBJITSYMFILE);
1778	pSymFile->aShdrs[i].sh_link = 0;
1779	pSymFile->aShdrs[i].sh_info = 0;
1780	pSymFile->aShdrs[i].sh_addralign = 1;
1781	pSymFile->aShdrs[i].sh_entsize = 0;
1782	i++;
1783
1784	Assert(i == RT_ELEMENTS(pSymFile->aShdrs));
1785
1786	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1787	/*
1788	* The program headers:
1789	*/
1790	/* Everything in a single LOAD segment: */
1791	i = 0;
1792	pSymFile->aPhdrs[i].p_type = PT_LOAD;
1793	pSymFile->aPhdrs[i].p_flags = PF_X \| PF_R;
1794	pSymFile->aPhdrs[i].p_offset
1795	= pSymFile->aPhdrs[i].p_vaddr
1796	= pSymFile->aPhdrs[i].p_paddr = 0;
1797	pSymFile->aPhdrs[i].p_filesz /* Size of segment in file. */
1798	= pSymFile->aPhdrs[i].p_memsz = pExecMemAllocator->cbChunk - offSymFileInChunk;
1799	pSymFile->aPhdrs[i].p_align = HOST_PAGE_SIZE;
1800	i++;
1801	/* The .dynamic segment. */
1802	pSymFile->aPhdrs[i].p_type = PT_DYNAMIC;
1803	pSymFile->aPhdrs[i].p_flags = PF_R;
1804	pSymFile->aPhdrs[i].p_offset
1805	= pSymFile->aPhdrs[i].p_vaddr
1806	= pSymFile->aPhdrs[i].p_paddr = RT_UOFFSETOF(GDBJITSYMFILE, aDyn);
1807	pSymFile->aPhdrs[i].p_filesz /* Size of segment in file. */
1808	= pSymFile->aPhdrs[i].p_memsz = sizeof(pSymFile->aDyn);
1809	pSymFile->aPhdrs[i].p_align = sizeof(pSymFile->aDyn[0].d_tag);
1810	i++;
1811
1812	Assert(i == RT_ELEMENTS(pSymFile->aPhdrs));
1813
1814	/*
1815	* The dynamic section:
1816	*/
1817	i = 0;
1818	pSymFile->aDyn[i].d_tag = DT_SONAME;
1819	pSymFile->aDyn[i].d_un.d_val = offStrTab;
1820	APPEND_STR_FMT("iem-exec-chunk-%u-%u", pVCpu->idCpu, idxChunk);
1821	i++;
1822	pSymFile->aDyn[i].d_tag = DT_STRTAB;
1823	pSymFile->aDyn[i].d_un.d_ptr = RT_UOFFSETOF(GDBJITSYMFILE, szzStrTab);
1824	i++;
1825	pSymFile->aDyn[i].d_tag = DT_STRSZ;
1826	pSymFile->aDyn[i].d_un.d_val = sizeof(pSymFile->szzStrTab);
1827	i++;
1828	pSymFile->aDyn[i].d_tag = DT_SYMTAB;
1829	pSymFile->aDyn[i].d_un.d_ptr = RT_UOFFSETOF(GDBJITSYMFILE, aDynSyms);
1830	i++;
1831	pSymFile->aDyn[i].d_tag = DT_SYMENT;
1832	pSymFile->aDyn[i].d_un.d_val = sizeof(pSymFile->aDynSyms[0]);
1833	i++;
1834	pSymFile->aDyn[i].d_tag = DT_NULL;
1835	i++;
1836	Assert(i == RT_ELEMENTS(pSymFile->aDyn));
1837	# endif /* IEMNATIVE_USE_GDB_JIT_ET_DYN */
1838
1839	/*
1840	* Symbol tables:
1841	*/
1842	/** @todo gdb doesn't seem to really like this ... */
1843	i = 0;
1844	pSymFile->aSymbols[i].st_name = 0;
1845	pSymFile->aSymbols[i].st_shndx = SHN_UNDEF;
1846	pSymFile->aSymbols[i].st_value = 0;
1847	pSymFile->aSymbols[i].st_size = 0;
1848	pSymFile->aSymbols[i].st_info = ELF64_ST_INFO(STB_LOCAL, STT_NOTYPE);
1849	pSymFile->aSymbols[i].st_other = 0 /* STV_DEFAULT */;
1850	# ifdef IEMNATIVE_USE_GDB_JIT_ET_DYN
1851	pSymFile->aDynSyms[0] = pSymFile->aSymbols[i];
1852	# endif
1853	i++;
1854
1855	pSymFile->aSymbols[i].st_name = 0;
1856	pSymFile->aSymbols[i].st_shndx = SHN_ABS;
1857	pSymFile->aSymbols[i].st_value = 0;
1858	pSymFile->aSymbols[i].st_size = 0;
1859	pSymFile->aSymbols[i].st_info = ELF64_ST_INFO(STB_LOCAL, STT_FILE);
1860	pSymFile->aSymbols[i].st_other = 0 /* STV_DEFAULT */;
1861	i++;
1862
1863	pSymFile->aSymbols[i].st_name = offStrTab;
1864	APPEND_STR_FMT("iem_exec_chunk_%u_%u", pVCpu->idCpu, idxChunk);
1865	# if 0
1866	pSymFile->aSymbols[i].st_shndx = iShText;
1867	pSymFile->aSymbols[i].st_value = 0;
1868	# else
1869	pSymFile->aSymbols[i].st_shndx = SHN_ABS;
1870	pSymFile->aSymbols[i].st_value = (uintptr_t)(pSymFile + 1);
1871	# endif
1872	pSymFile->aSymbols[i].st_size = pSymFile->aShdrs[iShText].sh_size;
1873	pSymFile->aSymbols[i].st_info = ELF64_ST_INFO(STB_GLOBAL, STT_FUNC);
1874	pSymFile->aSymbols[i].st_other = 0 /* STV_DEFAULT */;
1875	# ifdef IEMNATIVE_USE_GDB_JIT_ET_DYN
1876	pSymFile->aDynSyms[1] = pSymFile->aSymbols[i];
1877	pSymFile->aDynSyms[1].st_value = (uintptr_t)(pSymFile + 1);
1878	# endif
1879	i++;
1880
1881	Assert(i == RT_ELEMENTS(pSymFile->aSymbols));
1882	Assert(offStrTab < sizeof(pSymFile->szzStrTab));
1883
1884	/*
1885	* The GDB JIT entry and informing GDB.
1886	*/
1887	pEhFrame->GdbJitEntry.pbSymFile = (uint8_t *)pSymFile;
1888	# if 1
1889	pEhFrame->GdbJitEntry.cbSymFile = pExecMemAllocator->cbChunk - ((uintptr_t)pSymFile - (uintptr_t)pvChunk);
1890	# else
1891	pEhFrame->GdbJitEntry.cbSymFile = sizeof(GDBJITSYMFILE);
1892	# endif
1893
1894	RTOnce(&g_IemNativeGdbJitOnce, iemNativeGdbJitInitOnce, NULL);
1895	RTCritSectEnter(&g_IemNativeGdbJitLock);
1896	pEhFrame->GdbJitEntry.pNext = NULL;
1897	pEhFrame->GdbJitEntry.pPrev = __jit_debug_descriptor.pTail;
1898	if (__jit_debug_descriptor.pTail)
1899	__jit_debug_descriptor.pTail->pNext = &pEhFrame->GdbJitEntry;
1900	else
1901	__jit_debug_descriptor.pHead = &pEhFrame->GdbJitEntry;
1902	__jit_debug_descriptor.pTail = &pEhFrame->GdbJitEntry;
1903	__jit_debug_descriptor.pRelevant = &pEhFrame->GdbJitEntry;
1904
1905	/* Notify GDB: */
1906	__jit_debug_descriptor.enmAction = kGdbJitaction_Register;
1907	__jit_debug_register_code();
1908	__jit_debug_descriptor.enmAction = kGdbJitaction_NoAction;
1909	RTCritSectLeave(&g_IemNativeGdbJitLock);
1910
1911	# else /* !IEMNATIVE_USE_GDB_JIT */
1912	RT_NOREF(pVCpu);
1913	# endif /* !IEMNATIVE_USE_GDB_JIT */
1914
1915	return VINF_SUCCESS;
1916	}
1917
1918	# endif /* !RT_OS_WINDOWS */
1919	#endif /* IN_RING3 */
1920
1921
1922	/**
1923	* Adds another chunk to the executable memory allocator.
1924	*
1925	* This is used by the init code for the initial allocation and later by the
1926	* regular allocator function when it's out of memory.
1927	*/
1928	static int iemExecMemAllocatorGrow(PVMCPUCC pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator)
1929	{
1930	/* Check that we've room for growth. */
1931	uint32_t const idxChunk = pExecMemAllocator->cChunks;
1932	AssertLogRelReturn(idxChunk < pExecMemAllocator->cMaxChunks, VERR_OUT_OF_RESOURCES);
1933
1934	/* Allocate a chunk. */
1935	#ifdef RT_OS_DARWIN
1936	void *pvChunk = RTMemPageAllocEx(pExecMemAllocator->cbChunk, 0);
1937	#else
1938	void *pvChunk = RTMemPageAllocEx(pExecMemAllocator->cbChunk, RTMEMPAGEALLOC_F_EXECUTABLE);
1939	#endif
1940	AssertLogRelReturn(pvChunk, VERR_NO_EXEC_MEMORY);
1941
1942	#ifdef RT_OS_DARWIN
1943	/*
1944	* Because it is impossible to have a RWX memory allocation on macOS try to remap the memory
1945	* chunk readable/executable somewhere else so we can save us the hassle of switching between
1946	* protections when exeuctable memory is allocated.
1947	*/
1948	int rc = VERR_NO_EXEC_MEMORY;
1949	mach_port_t hPortTask = mach_task_self();
1950	mach_vm_address_t AddrChunk = (mach_vm_address_t)pvChunk;
1951	mach_vm_address_t AddrRemapped = 0;
1952	vm_prot_t ProtCur = 0;
1953	vm_prot_t ProtMax = 0;
1954	kern_return_t krc = mach_vm_remap(hPortTask, &AddrRemapped, pExecMemAllocator->cbChunk, 0,
1955	VM_FLAGS_ANYWHERE \| VM_FLAGS_RETURN_DATA_ADDR,
1956	hPortTask, AddrChunk, FALSE, &ProtCur, &ProtMax,
1957	VM_INHERIT_NONE);
1958	if (krc == KERN_SUCCESS)
1959	{
1960	krc = mach_vm_protect(mach_task_self(), AddrRemapped, pExecMemAllocator->cbChunk, FALSE, VM_PROT_READ \| VM_PROT_EXECUTE);
1961	if (krc == KERN_SUCCESS)
1962	rc = VINF_SUCCESS;
1963	else
1964	{
1965	AssertLogRelMsgFailed(("mach_vm_protect -> %d (%#x)\n", krc, krc));
1966	krc = mach_vm_deallocate(hPortTask, AddrRemapped, pExecMemAllocator->cbChunk);
1967	Assert(krc == KERN_SUCCESS);
1968	}
1969	}
1970	else
1971	AssertLogRelMsgFailed(("mach_vm_remap -> %d (%#x)\n", krc, krc));
1972	if (RT_FAILURE(rc))
1973	{
1974	RTMemPageFree(pvChunk, pExecMemAllocator->cbChunk);
1975	return rc;
1976	}
1977
1978	void pvChunkRx = (void )AddrRemapped;
1979	#else
1980	int rc = VINF_SUCCESS;
1981	void *pvChunkRx = pvChunk;
1982	#endif
1983
1984	/*
1985	* Add the chunk.
1986	*
1987	* This must be done before the unwind init so windows can allocate
1988	* memory from the chunk when using the alternative sub-allocator.
1989	*/
1990	pExecMemAllocator->aChunks[idxChunk].pvChunkRw = pvChunk;
1991	pExecMemAllocator->aChunks[idxChunk].pvChunkRx = pvChunkRx;
1992	#ifdef IN_RING3
1993	pExecMemAllocator->aChunks[idxChunk].pvUnwindInfo = NULL;
1994	#endif
1995	pExecMemAllocator->aChunks[idxChunk].cFreeUnits = pExecMemAllocator->cUnitsPerChunk;
1996	pExecMemAllocator->aChunks[idxChunk].idxFreeHint = 0;
1997	memset(&pExecMemAllocator->pbmAlloc[pExecMemAllocator->cBitmapElementsPerChunk * idxChunk],
1998	0, sizeof(pExecMemAllocator->pbmAlloc[0]) * pExecMemAllocator->cBitmapElementsPerChunk);
1999
2000	pExecMemAllocator->cChunks = idxChunk + 1;
2001	pExecMemAllocator->idxChunkHint = idxChunk;
2002
2003	pExecMemAllocator->cbTotal += pExecMemAllocator->cbChunk;
2004	pExecMemAllocator->cbFree += pExecMemAllocator->cbChunk;
2005
2006	/* If there is a chunk context init callback call it. */
2007	rc = iemNativeRecompileAttachExecMemChunkCtx(pVCpu, idxChunk, &pExecMemAllocator->aChunks[idxChunk].pCtx);
2008	#ifdef IN_RING3
2009	/*
2010	* Initialize the unwind information (this cannot really fail atm).
2011	* (This sets pvUnwindInfo.)
2012	*/
2013	if (RT_SUCCESS(rc))
2014	rc = iemExecMemAllocatorInitAndRegisterUnwindInfoForChunk(pVCpu, pExecMemAllocator, pvChunkRx, idxChunk);
2015	#endif
2016	if (RT_SUCCESS(rc))
2017	{ /* likely */ }
2018	else
2019	{
2020	/* Just in case the impossible happens, undo the above up: */
2021	pExecMemAllocator->cbTotal -= pExecMemAllocator->cbChunk;
2022	pExecMemAllocator->cbFree -= pExecMemAllocator->aChunks[idxChunk].cFreeUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
2023	pExecMemAllocator->cChunks = idxChunk;
2024	memset(&pExecMemAllocator->pbmAlloc[pExecMemAllocator->cBitmapElementsPerChunk * idxChunk],
2025	0xff, sizeof(pExecMemAllocator->pbmAlloc[0]) * pExecMemAllocator->cBitmapElementsPerChunk);
2026	pExecMemAllocator->aChunks[idxChunk].pvChunkRw = NULL;
2027	pExecMemAllocator->aChunks[idxChunk].cFreeUnits = 0;
2028
2029	# ifdef RT_OS_DARWIN
2030	krc = mach_vm_deallocate(mach_task_self(), (mach_vm_address_t)pExecMemAllocator->aChunks[idxChunk].pvChunkRx,
2031	pExecMemAllocator->cbChunk);
2032	Assert(krc == KERN_SUCCESS);
2033	# endif
2034
2035	RTMemPageFree(pvChunk, pExecMemAllocator->cbChunk);
2036	return rc;
2037	}
2038
2039	return VINF_SUCCESS;
2040	}
2041
2042
2043	/**
2044	* Initializes the executable memory allocator for native recompilation on the
2045	* calling EMT.
2046	*
2047	* @returns VBox status code.
2048	* @param pVCpu The cross context virtual CPU structure of the calling
2049	* thread.
2050	* @param cbMax The max size of the allocator.
2051	* @param cbInitial The initial allocator size.
2052	* @param cbChunk The chunk size, 0 or UINT32_MAX for default (@a cbMax
2053	* dependent).
2054	*/
2055	int iemExecMemAllocatorInit(PVMCPU pVCpu, uint64_t cbMax, uint64_t cbInitial, uint32_t cbChunk) RT_NOEXCEPT
2056	{
2057	/*
2058	* Validate input.
2059	*/
2060	AssertLogRelMsgReturn(cbMax >= _1M && cbMax <= _4G+_4G, ("cbMax=%RU64 (%RX64)\n", cbMax, cbMax), VERR_OUT_OF_RANGE);
2061	AssertReturn(cbInitial <= cbMax, VERR_OUT_OF_RANGE);
2062	AssertLogRelMsgReturn( cbChunk != UINT32_MAX
2063	\|\| cbChunk == 0
2064	\|\| ( RT_IS_POWER_OF_TWO(cbChunk)
2065	&& cbChunk >= _1M
2066	&& cbChunk <= _256M
2067	&& cbChunk <= cbMax),
2068	("cbChunk=%RU32 (%RX32) cbMax=%RU64\n", cbChunk, cbChunk, cbMax),
2069	VERR_OUT_OF_RANGE);
2070
2071	/*
2072	* Adjust/figure out the chunk size.
2073	*/
2074	if (cbChunk == 0 \|\| cbChunk == UINT32_MAX)
2075	{
2076	if (cbMax >= _256M)
2077	cbChunk = _64M;
2078	else
2079	{
2080	if (cbMax < _16M)
2081	cbChunk = cbMax >= _4M ? _4M : (uint32_t)cbMax;
2082	else
2083	cbChunk = (uint32_t)cbMax / 4;
2084	if (!RT_IS_POWER_OF_TWO(cbChunk))
2085	cbChunk = RT_BIT_32(ASMBitLastSetU32(cbChunk));
2086	}
2087	}
2088	#if defined(RT_OS_AMD64)
2089	Assert(cbChunk <= _2G);
2090	#elif defined(RT_OS_ARM64)
2091	if (cbChunk > _128M)
2092	cbChunk = _128M; /* Max relative branch distance is +/-2^(25+2) = +/-0x8000000 (134 217 728). */
2093	#endif
2094
2095	if (cbChunk > cbMax)
2096	cbMax = cbChunk;
2097	else
2098	cbMax = (cbMax - 1 + cbChunk) / cbChunk * cbChunk;
2099	uint32_t const cMaxChunks = (uint32_t)(cbMax / cbChunk);
2100	AssertLogRelReturn((uint64_t)cMaxChunks * cbChunk == cbMax, VERR_INTERNAL_ERROR_3);
2101
2102	/*
2103	* Allocate and initialize the allocatore instance.
2104	*/
2105	size_t const offBitmaps = RT_ALIGN_Z(RT_UOFFSETOF_DYN(IEMEXECMEMALLOCATOR, aChunks[cMaxChunks]), RT_CACHELINE_SIZE);
2106	size_t const cbBitmaps = (size_t)(cbChunk >> (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT + 3)) * cMaxChunks;
2107	size_t cbNeeded = offBitmaps + cbBitmaps;
2108	AssertCompile(IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT <= 10);
2109	Assert(cbChunk > RT_BIT_32(IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT + 3));
2110	#if defined(IN_RING3) && !defined(RT_OS_WINDOWS)
2111	size_t const offEhFrames = RT_ALIGN_Z(cbNeeded, RT_CACHELINE_SIZE);
2112	cbNeeded += sizeof(IEMEXECMEMCHUNKEHFRAME) * cMaxChunks;
2113	#endif
2114	PIEMEXECMEMALLOCATOR pExecMemAllocator = (PIEMEXECMEMALLOCATOR)RTMemAllocZ(cbNeeded);
2115	AssertLogRelMsgReturn(pExecMemAllocator, ("cbNeeded=%zx cMaxChunks=%#x cbChunk=%#x\n", cbNeeded, cMaxChunks, cbChunk),
2116	VERR_NO_MEMORY);
2117	pExecMemAllocator->uMagic = IEMEXECMEMALLOCATOR_MAGIC;
2118	pExecMemAllocator->cbChunk = cbChunk;
2119	pExecMemAllocator->cMaxChunks = cMaxChunks;
2120	pExecMemAllocator->cChunks = 0;
2121	pExecMemAllocator->idxChunkHint = 0;
2122	pExecMemAllocator->cAllocations = 0;
2123	pExecMemAllocator->cbTotal = 0;
2124	pExecMemAllocator->cbFree = 0;
2125	pExecMemAllocator->cbAllocated = 0;
2126	#ifdef VBOX_WITH_STATISTICS
2127	pExecMemAllocator->cbUnusable = 0;
2128	#endif
2129	pExecMemAllocator->pbmAlloc = (uint64_t *)((uintptr_t)pExecMemAllocator + offBitmaps);
2130	pExecMemAllocator->cUnitsPerChunk = cbChunk >> IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
2131	pExecMemAllocator->cBitmapElementsPerChunk = cbChunk >> (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT + 6);
2132	memset(pExecMemAllocator->pbmAlloc, 0xff, cbBitmaps); /* Mark everything as allocated. Clear when chunks are added. */
2133	#if defined(IN_RING3) && !defined(RT_OS_WINDOWS)
2134	pExecMemAllocator->paEhFrames = (PIEMEXECMEMCHUNKEHFRAME)((uintptr_t)pExecMemAllocator + offEhFrames);
2135	#endif
2136	for (uint32_t i = 0; i < cMaxChunks; i++)
2137	{
2138	pExecMemAllocator->aChunks[i].cFreeUnits = 0;
2139	pExecMemAllocator->aChunks[i].idxFreeHint = 0;
2140	pExecMemAllocator->aChunks[i].pvChunkRw = NULL;
2141	#ifdef IN_RING0
2142	pExecMemAllocator->aChunks[i].hMemObj = NIL_RTR0MEMOBJ;
2143	#else
2144	pExecMemAllocator->aChunks[i].pvUnwindInfo = NULL;
2145	#endif
2146	}
2147	pVCpu->iem.s.pExecMemAllocatorR3 = pExecMemAllocator;
2148
2149	/*
2150	* Do the initial allocations.
2151	*/
2152	while (cbInitial < (uint64_t)pExecMemAllocator->cChunks * pExecMemAllocator->cbChunk)
2153	{
2154	int rc = iemExecMemAllocatorGrow(pVCpu, pExecMemAllocator);
2155	AssertLogRelRCReturn(rc, rc);
2156	}
2157
2158	pExecMemAllocator->idxChunkHint = 0;
2159
2160	/*
2161	* Register statistics.
2162	*/
2163	PUVM const pUVM = pVCpu->pUVCpu->pUVM;
2164	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cAllocations, STAMTYPE_U32, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
2165	"Current number of allocations", "/IEM/CPU%u/re/ExecMem/cAllocations", pVCpu->idCpu);
2166	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cChunks, STAMTYPE_U32, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,
2167	"Currently allocated chunks", "/IEM/CPU%u/re/ExecMem/cChunks", pVCpu->idCpu);
2168	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cMaxChunks, STAMTYPE_U32, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,
2169	"Maximum number of chunks", "/IEM/CPU%u/re/ExecMem/cMaxChunks", pVCpu->idCpu);
2170	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbChunk, STAMTYPE_U32, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
2171	"Allocation chunk size", "/IEM/CPU%u/re/ExecMem/cbChunk", pVCpu->idCpu);
2172	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbAllocated, STAMTYPE_U64, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
2173	"Number of bytes current allocated", "/IEM/CPU%u/re/ExecMem/cbAllocated", pVCpu->idCpu);
2174	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbFree, STAMTYPE_U64, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
2175	"Number of bytes current free", "/IEM/CPU%u/re/ExecMem/cbFree", pVCpu->idCpu);
2176	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbTotal, STAMTYPE_U64, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
2177	"Total number of byte", "/IEM/CPU%u/re/ExecMem/cbTotal", pVCpu->idCpu);
2178	#ifdef VBOX_WITH_STATISTICS
2179	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbUnusable, STAMTYPE_U64, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
2180	"Total number of bytes being unusable", "/IEM/CPU%u/re/ExecMem/cbUnusable", pVCpu->idCpu);
2181	STAMR3RegisterFU(pUVM, &pExecMemAllocator->StatAlloc, STAMTYPE_PROFILE, STAMVISIBILITY_ALWAYS, STAMUNIT_TICKS_PER_CALL,
2182	"Profiling the allocator", "/IEM/CPU%u/re/ExecMem/ProfAlloc", pVCpu->idCpu);
2183	for (unsigned i = 1; i < RT_ELEMENTS(pExecMemAllocator->aStatSizes); i++)
2184	STAMR3RegisterFU(pUVM, &pExecMemAllocator->aStatSizes[i], STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,
2185	"Number of allocations of this number of allocation units",
2186	"/IEM/CPU%u/re/ExecMem/aSize%02u", pVCpu->idCpu, i);
2187	STAMR3RegisterFU(pUVM, &pExecMemAllocator->aStatSizes[0], STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,
2188	"Number of allocations 16 units or larger", "/IEM/CPU%u/re/ExecMem/aSize16OrLarger", pVCpu->idCpu);
2189	#endif
2190	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING
2191	STAMR3RegisterFU(pUVM, &pExecMemAllocator->StatPruneProf, STAMTYPE_PROFILE, STAMVISIBILITY_ALWAYS, STAMUNIT_TICKS_PER_CALL,
2192	"Pruning executable memory (alt)", "/IEM/CPU%u/re/ExecMem/Pruning", pVCpu->idCpu);
2193	STAMR3RegisterFU(pUVM, &pExecMemAllocator->StatPruneRecovered, STAMTYPE_PROFILE, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES_PER_CALL,
2194	"Bytes recovered while pruning", "/IEM/CPU%u/re/ExecMem/PruningRecovered", pVCpu->idCpu);
2195	#endif
2196	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cFruitlessChunkScans, STAMTYPE_U64_RESET, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,
2197	"Chunks fruitlessly scanned for free space", "/IEM/CPU%u/re/ExecMem/FruitlessChunkScans", pVCpu->idCpu);
2198
2199	return VINF_SUCCESS;
2200	}
2201

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

source: vbox/trunk/src/VBox/VMM/VMMAll/IEMAllN8veExecMem.cpp@ 106443

Download in other formats: