IEMAllN8veExecMem.cpp@ 106723

Last change on this file since 106723 was 106465, checked in by vboxsync, 6 weeks ago
VMM/IEM: Added iemNativeEmitLoadGprWithGstReg[Ex]T and iemNativeEmitStoreGprToGstReg[Ex]T as better way of explictly loading & storing standard guest registers. 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 106465 2024-10-18 00:27:52Z 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	RT_GCC_EXTENSION
326	union
327	{
328	struct
329	{
330	/** Magic value / eyecatcher (IEMEXECMEMALLOCHDR_MAGIC). */
331	uint32_t uMagic;
332	/** The allocation chunk (for speeding up freeing). */
333	uint32_t idxChunk;
334	};
335	/** Combined magic and chunk index, for the pruning scanner code. */
336	uint64_t u64MagicAndChunkIdx;
337	};
338	/** Pointer to the translation block the allocation belongs to.
339	* This is the whole point of the header. */
340	PIEMTB pTb;
341	} IEMEXECMEMALLOCHDR;
342	/** Pointer to an allocation header. */
343	typedef IEMEXECMEMALLOCHDR *PIEMEXECMEMALLOCHDR;
344	/** Magic value for IEMEXECMEMALLOCHDR ('ExeM'). */
345	# define IEMEXECMEMALLOCHDR_MAGIC UINT32_C(0x4d657845)
346	#endif
347
348
349	static int iemExecMemAllocatorGrow(PVMCPUCC pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator);
350
351
352	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING
353	/**
354	* Frees up executable memory when we're out space.
355	*
356	* This is an alternative to iemTbAllocatorFreeupNativeSpace() that frees up
357	* space in a more linear fashion from the allocator's point of view. It may
358	* also defragment if implemented & enabled
359	*/
360	static void iemExecMemAllocatorPrune(PVMCPU pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator)
361	{
362	# ifndef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
363	# error "IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING requires IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER"
364	# endif
365	STAM_REL_PROFILE_START(&pExecMemAllocator->StatPruneProf, a);
366
367	/*
368	* Before we can start, we must process delayed frees.
369	*/
370	#if 1
371	PIEMTBALLOCATOR const pTbAllocator = iemTbAllocatorFreeBulkStart(pVCpu);
372	#else
373	iemTbAllocatorProcessDelayedFrees(pVCpu, pVCpu->iem.s.pTbAllocatorR3);
374	#endif
375
376	AssertCompile(RT_IS_POWER_OF_TWO(IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE));
377
378	uint32_t const cbChunk = pExecMemAllocator->cbChunk;
379	AssertReturnVoid(RT_IS_POWER_OF_TWO(cbChunk));
380	AssertReturnVoid(cbChunk >= _1M && cbChunk <= _256M); /* see iemExecMemAllocatorInit */
381
382	uint32_t const cChunks = pExecMemAllocator->cChunks;
383	AssertReturnVoid(cChunks == pExecMemAllocator->cMaxChunks);
384	AssertReturnVoid(cChunks >= 1);
385
386	Assert(!pVCpu->iem.s.pCurTbR3);
387
388	/*
389	* Decide how much to prune. The chunk is is a multiple of two, so we'll be
390	* scanning a multiple of two here as well.
391	*/
392	uint32_t cbToPrune = cbChunk;
393
394	/* Never more than 25%. */
395	if (cChunks < 4)
396	cbToPrune /= cChunks == 1 ? 4 : 2;
397
398	/* Upper limit. In a debug build a 4MB limit averages out at ~0.6ms per call. */
399	if (cbToPrune > _4M)
400	cbToPrune = _4M;
401
402	/*
403	* Adjust the pruning chunk and offset accordingly.
404	*/
405	uint32_t idxChunk = pExecMemAllocator->idxChunkPrune;
406	uint32_t offChunk = pExecMemAllocator->offChunkPrune;
407	offChunk &= ~(uint32_t)(IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1U);
408	if (offChunk >= cbChunk)
409	{
410	offChunk = 0;
411	idxChunk += 1;
412	}
413	if (idxChunk >= cChunks)
414	{
415	offChunk = 0;
416	idxChunk = 0;
417	}
418
419	uint32_t const offPruneStart = offChunk;
420	uint32_t const offPruneEnd = RT_MIN(offChunk + cbToPrune, cbChunk);
421
422	/*
423	* Do the pruning. The current approach is the sever kind.
424	*
425	* This is memory bound, as we must load both the allocation header and the
426	* associated TB and then modify them. So, the CPU isn't all that unitilized
427	* here. Try apply some prefetching to speed it up a tiny bit.
428	*/
429	uint64_t cbPruned = 0;
430	uint64_t const u64MagicAndChunkIdx = RT_MAKE_U64(IEMEXECMEMALLOCHDR_MAGIC, idxChunk);
431	uint8_t * const pbChunk = (uint8_t *)pExecMemAllocator->aChunks[idxChunk].pvChunkRx;
432	while (offChunk < offPruneEnd)
433	{
434	PIEMEXECMEMALLOCHDR pHdr = (PIEMEXECMEMALLOCHDR)&pbChunk[offChunk];
435
436	/* Is this the start of an allocation block for a TB? (We typically
437	have one allocation at the start of each chunk for the unwind info
438	where pTb is NULL.) */
439	PIEMTB pTb;
440	if ( pHdr->u64MagicAndChunkIdx == u64MagicAndChunkIdx
441	&& RT_LIKELY((pTb = pHdr->pTb) != NULL))
442	{
443	AssertPtr(pTb);
444
445	uint32_t const cbBlock = RT_ALIGN_32(pTb->Native.cInstructions * sizeof(IEMNATIVEINSTR) + sizeof(*pHdr),
446	IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE);
447
448	/* Prefetch the next header before freeing the current one and its TB. */
449	/** @todo Iff the block size was part of the header in some way, this could be
450	* a tiny bit faster. */
451	offChunk += cbBlock;
452	#if defined(_MSC_VER) && defined(RT_ARCH_AMD64)
453	_mm_prefetch((char *)&pbChunk[offChunk], _MM_HINT_T0);
454	#elif defined(_MSC_VER) && defined(RT_ARCH_ARM64)
455	__prefetch(&pbChunk[offChunk]);
456	#else
457	__builtin_prefetch(&pbChunk[offChunk], 1 /rw/);
458	#endif
459	/* Some paranoia first, though. */
460	AssertBreakStmt(offChunk <= cbChunk, offChunk -= cbBlock - IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE);
461	cbPruned += cbBlock;
462
463	#if 1
464	iemTbAllocatorFreeBulk(pVCpu, pTbAllocator, pTb);
465	#else
466	iemTbAllocatorFree(pVCpu, pTb);
467	#endif
468	}
469	else
470	offChunk += IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE;
471	}
472	STAM_REL_PROFILE_ADD_PERIOD(&pExecMemAllocator->StatPruneRecovered, cbPruned);
473
474	pVCpu->iem.s.ppTbLookupEntryR3 = &pVCpu->iem.s.pTbLookupEntryDummyR3;
475
476	/*
477	* Save the current pruning point.
478	*/
479	pExecMemAllocator->offChunkPrune = offChunk;
480	pExecMemAllocator->idxChunkPrune = idxChunk;
481
482	/* Set the hint to the start of the pruned region. */
483	pExecMemAllocator->idxChunkHint = idxChunk;
484	pExecMemAllocator->aChunks[idxChunk].idxFreeHint = offPruneStart / IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE;
485
486	STAM_REL_PROFILE_STOP(&pExecMemAllocator->StatPruneProf, a);
487	}
488	#endif /* IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING */
489
490
491	#if defined(VBOX_STRICT) \|\| 0
492	/**
493	* The old bitmap scanner code, for comparison and assertions.
494	*/
495	static uint32_t iemExecMemAllocatorFindReqFreeUnitsOld(uint64_t *pbmAlloc, uint32_t cToScan, uint32_t cReqUnits)
496	{
497	/** @todo This can probably be done more efficiently for non-x86 systems. */
498	int iBit = ASMBitFirstClear(pbmAlloc, cToScan);
499	while (iBit >= 0 && (uint32_t)iBit <= cToScan - cReqUnits)
500	{
501	uint32_t idxAddBit = 1;
502	while (idxAddBit < cReqUnits && !ASMBitTest(pbmAlloc, (uint32_t)iBit + idxAddBit))
503	idxAddBit++;
504	if (idxAddBit >= cReqUnits)
505	return (uint32_t)iBit;
506	iBit = ASMBitNextClear(pbmAlloc, cToScan, iBit + idxAddBit - 1);
507	}
508	return UINT32_MAX;
509	}
510	#endif
511
512
513	/**
514	* Bitmap scanner code that looks for a bunch of @a cReqUnits zero bits.
515	*
516	* Booting win11 with a r165098 release build the average native TB size is
517	* around 9 units (of 256 bytes). So, it is unlikely we need to scan any
518	* subsequent words once we hit a patch of zeros, thus @a a_fBig.
519	*
520	* @todo This needs more tweaking. While it is faster the the old code,
521	* it doens't seem like it's all that much. :/
522	*/
523	template<const bool a_fBig>
524	static uint32_t iemExecMemAllocatorFindReqFreeUnits(uint64_t *pbmAlloc, uint32_t c64WordsToScan, uint32_t cReqUnits)
525	{
526	/*
527	* Scan the (section of the) allocation bitmap in 64-bit words.
528	*/
529	unsigned cPrevLeadingZeros = 0;
530	for (uint32_t off = 0; off < c64WordsToScan; off++)
531	{
532	uint64_t uWord = pbmAlloc[off];
533	if (uWord == UINT64_MAX)
534	{
535	/*
536	* Getting thru patches of UINT64_MAX is a frequent problem when the allocator
537	* fills up, so it's definitely worth optimizing.
538	*
539	* The complicated code below is a bit faster on arm. Reducing the per TB cost
540	* from 4255ns to 4106ns (best run out of 10). On win/amd64 there isn't an
541	* obvious gain here, at least not with the data currently being profiled.
542	*/
543	#if 1
544	off++;
545	uint32_t cQuads = (c64WordsToScan - off) / 4;
546
547	/* Align. */
548	if (cQuads > 1)
549	switch (((uintptr_t)&pbmAlloc[off] / sizeof(uint64_t)) & 3)
550	{
551	case 0:
552	break;
553	case 1:
554	{
555	uWord = pbmAlloc[off];
556	uint64_t uWord1 = pbmAlloc[off + 1];
557	uint64_t uWord2 = pbmAlloc[off + 2];
558	if ((uWord & uWord1 & uWord2) == UINT64_MAX)
559	{
560	off += 3;
561	cQuads = (c64WordsToScan - off) / 4;
562	}
563	else if (uWord == UINT64_MAX)
564	{
565	if (uWord1 != UINT64_MAX)
566	{
567	uWord = uWord1;
568	off += 1;
569	}
570	else
571	{
572	uWord = uWord2;
573	off += 2;
574	}
575	}
576	break;
577	}
578	case 2:
579	{
580	uWord = pbmAlloc[off];
581	uint64_t uWord1 = pbmAlloc[off + 1];
582	if ((uWord & uWord1) == UINT64_MAX)
583	{
584	off += 2;
585	cQuads = (c64WordsToScan - off) / 4;
586	}
587	else if (uWord == UINT64_MAX)
588	{
589	uWord = uWord1;
590	off += 1;
591	}
592	break;
593	}
594	case 3:
595	uWord = pbmAlloc[off];
596	if (uWord == UINT64_MAX)
597	{
598	off++;
599	cQuads = (c64WordsToScan - off) / 4;
600	}
601	break;
602	}
603	if (uWord == UINT64_MAX)
604	{
605	/* Looping over 32 bytes at a time. */
606	for (;;)
607	{
608	if (cQuads-- > 0)
609	{
610	uWord = pbmAlloc[off + 0];
611	uint64_t uWord1 = pbmAlloc[off + 1];
612	uint64_t uWord2 = pbmAlloc[off + 2];
613	uint64_t uWord3 = pbmAlloc[off + 3];
614	if ((uWord & uWord1 & uWord2 & uWord3) == UINT64_MAX)
615	off += 4;
616	else
617	{
618	if (uWord != UINT64_MAX)
619	{ }
620	else if (uWord1 != UINT64_MAX)
621	{
622	uWord = uWord1;
623	off += 1;
624	}
625	else if (uWord2 != UINT64_MAX)
626	{
627	uWord = uWord2;
628	off += 2;
629	}
630	else
631	{
632	uWord = uWord3;
633	off += 3;
634	}
635	break;
636	}
637	}
638	else
639	{
640	if (off < c64WordsToScan)
641	{
642	uWord = pbmAlloc[off];
643	if (uWord != UINT64_MAX)
644	break;
645	off++;
646	if (off < c64WordsToScan)
647	{
648	uWord = pbmAlloc[off];
649	if (uWord != UINT64_MAX)
650	break;
651	off++;
652	if (off < c64WordsToScan)
653	{
654	uWord = pbmAlloc[off];
655	if (uWord != UINT64_MAX)
656	break;
657	Assert(off + 1 == c64WordsToScan);
658	}
659	}
660	}
661	return UINT32_MAX;
662	}
663	}
664	}
665	#else
666	do
667	{
668	off++;
669	if (off < c64WordsToScan)
670	uWord = pbmAlloc[off];
671	else
672	return UINT32_MAX;
673	} while (uWord == UINT64_MAX);
674	#endif
675	cPrevLeadingZeros = 0;
676	}
677
678	/*
679	* If we get down here, we have a word that isn't UINT64_MAX.
680	*/
681	if (uWord != 0)
682	{
683	/*
684	* Fend of large request we cannot satisfy before the first set bit.
685	*/
686	if (!a_fBig \|\| cReqUnits < 64 + cPrevLeadingZeros)
687	{
688	#ifdef __GNUC__
689	unsigned cZerosInWord = __builtin_popcountl(~uWord);
690	#elif defined(_MSC_VER) && defined(RT_ARCH_AMD64)
691	unsigned cZerosInWord = __popcnt64(~uWord);
692	#elif defined(_MSC_VER) && defined(RT_ARCH_ARM64)
693	unsigned cZerosInWord = _CountOneBits64(~uWord);
694	#else
695	# pragma message("need popcount intrinsic or something...")
696	unsigned cZerosInWord = 0;
697	for (uint64_t uTmp = ~uWords; uTmp; cZerosInWord++)
698	uTmp &= uTmp - 1; /* Clears the least significant bit set. */
699	#endif
700	if (cZerosInWord + cPrevLeadingZeros >= cReqUnits)
701	{
702	/* Check if we've got a patch of zeros at the trailing end
703	when joined with the previous word: */
704	#ifdef __GNUC__
705	unsigned cTrailingZeros = __builtin_ctzl(uWord);
706	#else
707	unsigned cTrailingZeros = ASMBitFirstSetU64(uWord) - 1;
708	#endif
709	if (cPrevLeadingZeros + cTrailingZeros >= cReqUnits)
710	return off * 64 - cPrevLeadingZeros;
711
712	/*
713	* Try leading zeros before we get on with the tedious stuff.
714	*/
715	#ifdef __GNUC__
716	cPrevLeadingZeros = __builtin_clzl(uWord);
717	#else
718	cPrevLeadingZeros = 64 - ASMBitLastSetU64(uWord);
719	#endif
720	if (cPrevLeadingZeros >= cReqUnits)
721	return (off + 1) * 64 - cPrevLeadingZeros;
722
723	/*
724	* Check the popcount again sans leading & trailing before looking
725	* inside the word.
726	*/
727	cZerosInWord -= cPrevLeadingZeros + cTrailingZeros;
728	if (cZerosInWord >= cReqUnits)
729	{
730	/* 1; 64 - 0 - 1 = 63; */
731	unsigned const iBitLast = 64 - cPrevLeadingZeros - cReqUnits; /** @todo boundrary */
732	unsigned iBit = cTrailingZeros;
733	uWord >>= cTrailingZeros;
734	do
735	{
736	Assert(uWord & 1);
737	#ifdef __GNUC__
738	unsigned iZeroBit = __builtin_ctzl(~uWord);
739	#else
740	unsigned iZeroBit = ASMBitFirstSetU64(~uWord) - 1;
741	#endif
742	iBit += iZeroBit;
743	uWord >>= iZeroBit;
744	Assert(iBit <= iBitLast);
745	Assert((uWord & 1) == 0);
746	#ifdef __GNUC__
747	unsigned cZeros = __builtin_ctzl(uWord);
748	#else
749	unsigned cZeros = ASMBitFirstSetU64(uWord) - 1;
750	#endif
751	if (cZeros >= cReqUnits)
752	return off * 64 + iBit;
753
754	cZerosInWord -= cZeros; /* (may underflow as we will count shifted in zeros) */
755	iBit += cZeros;
756	uWord >>= cZeros;
757	} while ((int)cZerosInWord >= (int)cReqUnits && iBit < iBitLast);
758	}
759	continue; /* we've already calculated cPrevLeadingZeros */
760	}
761	}
762
763	/* Update the leading (MSB) zero count. */
764	#ifdef __GNUC__
765	cPrevLeadingZeros = __builtin_clzl(uWord);
766	#else
767	cPrevLeadingZeros = 64 - ASMBitLastSetU64(uWord);
768	#endif
769	}
770	/*
771	* uWord == 0
772	*/
773	else
774	{
775	if RT_CONSTEXPR_IF(!a_fBig)
776	return off * 64 - cPrevLeadingZeros;
777	else /* keep else */
778	{
779	if (cPrevLeadingZeros + 64 >= cReqUnits)
780	return off * 64 - cPrevLeadingZeros;
781	for (uint32_t off2 = off + 1;; off2++)
782	{
783	if (off2 < c64WordsToScan)
784	{
785	uWord = pbmAlloc[off2];
786	if (uWord == UINT64_MAX)
787	{
788	cPrevLeadingZeros = 0;
789	break;
790	}
791	if (uWord == 0)
792	{
793	if (cPrevLeadingZeros + (off2 - off + 1) * 64 >= cReqUnits)
794	return off * 64 - cPrevLeadingZeros;
795	}
796	else
797	{
798	#ifdef __GNUC__
799	unsigned cTrailingZeros = __builtin_ctzl(uWord);
800	#else
801	unsigned cTrailingZeros = ASMBitFirstSetU64(uWord) - 1;
802	#endif
803	if (cPrevLeadingZeros + (off2 - off) * 64 + cTrailingZeros >= cReqUnits)
804	return off * 64 - cPrevLeadingZeros;
805	#ifdef __GNUC__
806	cPrevLeadingZeros = __builtin_clzl(uWord);
807	#else
808	cPrevLeadingZeros = 64 - ASMBitLastSetU64(uWord);
809	#endif
810	break;
811	}
812	}
813	else
814	return UINT32_MAX;
815	}
816	}
817	}
818	}
819	return UINT32_MAX;
820	}
821
822
823	/**
824	* Try allocate a block of @a cReqUnits in the chunk @a idxChunk.
825	*/
826	static void *
827	iemExecMemAllocatorAllocInChunkInt(PIEMEXECMEMALLOCATOR pExecMemAllocator, uint64_t *pbmAlloc, uint32_t idxFirst,
828	uint32_t cToScan, uint32_t cReqUnits, uint32_t idxChunk, PIEMTB pTb,
829	void *ppvExec, PCIEMNATIVEPERCHUNKCTX ppChunkCtx)
830	{
831	/*
832	* Shift the bitmap to the idxFirst bit so we can use ASMBitFirstClear.
833	*/
834	Assert(!(cToScan & 63));
835	Assert(!(idxFirst & 63));
836	Assert(cToScan + idxFirst <= pExecMemAllocator->cUnitsPerChunk);
837	pbmAlloc += idxFirst / 64;
838	cToScan += idxFirst & 63;
839	Assert(!(cToScan & 63));
840
841	#if 1
842	uint32_t const iBit = cReqUnits < 64
843	? iemExecMemAllocatorFindReqFreeUnits<false>(pbmAlloc, cToScan / 64, cReqUnits)
844	: iemExecMemAllocatorFindReqFreeUnits<true>( pbmAlloc, cToScan / 64, cReqUnits);
845	# ifdef VBOX_STRICT
846	uint32_t const iBitOld = iemExecMemAllocatorFindReqFreeUnitsOld(pbmAlloc, cToScan, cReqUnits);
847	AssertMsg( iBit == iBitOld
848	\|\| (iBit / 64) == (iBitOld / 64), /* New algorithm will return trailing hit before middle. */
849	("iBit=%#x (%#018RX64); iBitOld=%#x (%#018RX64); cReqUnits=%#x\n",
850	iBit, iBit != UINT32_MAX ? pbmAlloc[iBit / 64] : 0,
851	iBitOld, iBitOld != UINT32_MAX ? pbmAlloc[iBitOld / 64] : 0, cReqUnits));
852	# endif
853	#else
854	uint32_t const iBit = iemExecMemAllocatorFindReqFreeUnitsOld(pbmAlloc, cToScan, cReqUnits);
855	#endif
856	if (iBit != UINT32_MAX)
857	{
858	ASMBitSetRange(pbmAlloc, (uint32_t)iBit, (uint32_t)iBit + cReqUnits);
859
860	PIEMEXECMEMCHUNK const pChunk = &pExecMemAllocator->aChunks[idxChunk];
861	pChunk->cFreeUnits -= cReqUnits;
862	pChunk->idxFreeHint = (uint32_t)iBit + cReqUnits;
863
864	pExecMemAllocator->cAllocations += 1;
865	uint32_t const cbReq = cReqUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
866	pExecMemAllocator->cbAllocated += cbReq;
867	pExecMemAllocator->cbFree -= cbReq;
868	pExecMemAllocator->idxChunkHint = idxChunk;
869
870	void * const pvMemRw = (uint8_t *)pChunk->pvChunkRw
871	+ ((idxFirst + (uint32_t)iBit) << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT);
872
873	if (ppChunkCtx)
874	*ppChunkCtx = pChunk->pCtx;
875
876	/*
877	* Initialize the header and return.
878	*/
879	# ifdef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
880	PIEMEXECMEMALLOCHDR const pHdr = (PIEMEXECMEMALLOCHDR)pvMemRw;
881	pHdr->uMagic = IEMEXECMEMALLOCHDR_MAGIC;
882	pHdr->idxChunk = idxChunk;
883	pHdr->pTb = pTb;
884
885	if (ppvExec)
886	ppvExec = (uint8_t )pChunk->pvChunkRx
887	+ ((idxFirst + (uint32_t)iBit) << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT)
888	+ sizeof(*pHdr);
889
890	return pHdr + 1;
891	#else
892	if (ppvExec)
893	ppvExec = (uint8_t )pChunk->pvChunkRx
894	+ ((idxFirst + (uint32_t)iBit) << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT);
895
896	RT_NOREF(pTb);
897	return pvMem;
898	#endif
899	}
900
901	return NULL;
902	}
903
904
905	/**
906	* Converts requested number of bytes into a unit count.
907	*/
908	DECL_FORCE_INLINE(uint32_t) iemExecMemAllocBytesToUnits(uint32_t cbReq)
909	{
910	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
911	return (cbReq + sizeof(IEMEXECMEMALLOCHDR) + IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1)
912	#else
913	return (cbReq + IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1)
914	#endif
915	>> IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
916	}
917
918
919	DECL_FORCE_INLINE(PIEMNATIVEINSTR)
920	iemExecMemAllocatorAllocUnitsInChunkInner(PIEMEXECMEMALLOCATOR pExecMemAllocator, uint32_t idxChunk, uint32_t cReqUnits,
921	PIEMTB pTb, PIEMNATIVEINSTR ppaExec, PCIEMNATIVEPERCHUNKCTX ppChunkCtx)
922	{
923	uint64_t * const pbmAlloc = &pExecMemAllocator->pbmAlloc[pExecMemAllocator->cBitmapElementsPerChunk * idxChunk];
924	uint32_t const idxHint = pExecMemAllocator->aChunks[idxChunk].idxFreeHint & ~(uint32_t)63;
925	if (idxHint + cReqUnits <= pExecMemAllocator->cUnitsPerChunk)
926	{
927	void *pvRet = iemExecMemAllocatorAllocInChunkInt(pExecMemAllocator, pbmAlloc, idxHint,
928	pExecMemAllocator->cUnitsPerChunk - idxHint,
929	cReqUnits, idxChunk, pTb, (void **)ppaExec, ppChunkCtx);
930	if (pvRet)
931	return (PIEMNATIVEINSTR)pvRet;
932	}
933	void *pvRet = iemExecMemAllocatorAllocInChunkInt(pExecMemAllocator, pbmAlloc, 0,
934	RT_MIN(pExecMemAllocator->cUnitsPerChunk,
935	RT_ALIGN_32(idxHint + cReqUnits, 64*4)),
936	cReqUnits, idxChunk, pTb, (void **)ppaExec, ppChunkCtx);
937	if (pvRet)
938	return (PIEMNATIVEINSTR)pvRet;
939
940	pExecMemAllocator->cFruitlessChunkScans += 1;
941	return NULL;
942	}
943
944
945	DECLINLINE(PIEMNATIVEINSTR)
946	iemExecMemAllocatorAllocBytesInChunk(PIEMEXECMEMALLOCATOR pExecMemAllocator, uint32_t idxChunk, uint32_t cbReq,
947	PIEMNATIVEINSTR *ppaExec)
948	{
949	uint32_t const cReqUnits = iemExecMemAllocBytesToUnits(cbReq);
950	if (cReqUnits <= pExecMemAllocator->aChunks[idxChunk].cFreeUnits)
951	return iemExecMemAllocatorAllocUnitsInChunkInner(pExecMemAllocator, idxChunk, cReqUnits, NULL /pTb/,
952	ppaExec, NULL /ppChunkCtx/);
953	return NULL;
954	}
955
956
957	/**
958	* Allocates @a cbReq bytes of executable memory.
959	*
960	* @returns Pointer to the readable/writeable memory, NULL if out of memory or other problem
961	* encountered.
962	* @param pVCpu The cross context virtual CPU structure of the
963	* calling thread.
964	* @param cbReq How many bytes are required.
965	* @param pTb The translation block that will be using the allocation.
966	* @param ppaExec Where to return the pointer to executable view of
967	* the allocated memory, optional.
968	* @param ppChunkCtx Where to return the per chunk attached context
969	* if available, optional.
970	*/
971	DECLHIDDEN(PIEMNATIVEINSTR) iemExecMemAllocatorAlloc(PVMCPU pVCpu, uint32_t cbReq, PIEMTB pTb,
972	PIEMNATIVEINSTR ppaExec, PCIEMNATIVEPERCHUNKCTX ppChunkCtx) RT_NOEXCEPT
973	{
974	PIEMEXECMEMALLOCATOR pExecMemAllocator = pVCpu->iem.s.pExecMemAllocatorR3;
975	AssertReturn(pExecMemAllocator && pExecMemAllocator->uMagic == IEMEXECMEMALLOCATOR_MAGIC, NULL);
976	AssertMsgReturn(cbReq > 32 && cbReq < _512K, ("%#x\n", cbReq), NULL);
977	STAM_PROFILE_START(&pExecMemAllocator->StatAlloc, a);
978
979	uint32_t const cReqUnits = iemExecMemAllocBytesToUnits(cbReq);
980	STAM_COUNTER_INC(&pExecMemAllocator->aStatSizes[cReqUnits < RT_ELEMENTS(pExecMemAllocator->aStatSizes) ? cReqUnits : 0]);
981	for (unsigned iIteration = 0;; iIteration++)
982	{
983	if ( cbReq * 2 <= pExecMemAllocator->cbFree
984	\|\| (cReqUnits == 1 \|\| pExecMemAllocator->cbFree >= IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE) )
985	{
986	uint32_t const cChunks = pExecMemAllocator->cChunks;
987	uint32_t const idxChunkHint = pExecMemAllocator->idxChunkHint < cChunks ? pExecMemAllocator->idxChunkHint : 0;
988
989	/*
990	* We do two passes here, the first pass we skip chunks with fewer than cReqUnits * 16,
991	* the 2nd pass we skip chunks. The second pass checks the one skipped in the first pass.
992	*/
993	for (uint32_t cMinFreePass = cReqUnits == 1 ? cReqUnits : cReqUnits * 16, cMaxFreePass = UINT32_MAX;;)
994	{
995	for (uint32_t idxChunk = idxChunkHint; idxChunk < cChunks; idxChunk++)
996	if ( pExecMemAllocator->aChunks[idxChunk].cFreeUnits >= cMinFreePass
997	&& pExecMemAllocator->aChunks[idxChunk].cFreeUnits <= cMaxFreePass)
998	{
999	PIEMNATIVEINSTR const pRet = iemExecMemAllocatorAllocUnitsInChunkInner(pExecMemAllocator, idxChunk,
1000	cReqUnits, pTb, ppaExec, ppChunkCtx);
1001	if (pRet)
1002	{
1003	STAM_PROFILE_STOP(&pExecMemAllocator->StatAlloc, a);
1004	#ifdef VBOX_WITH_STATISTICS
1005	pExecMemAllocator->cbUnusable += (cReqUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT) - cbReq;
1006	#endif
1007	return pRet;
1008	}
1009	}
1010	for (uint32_t idxChunk = 0; idxChunk < idxChunkHint; idxChunk++)
1011	if ( pExecMemAllocator->aChunks[idxChunk].cFreeUnits >= cMinFreePass
1012	&& pExecMemAllocator->aChunks[idxChunk].cFreeUnits <= cMaxFreePass)
1013	{
1014	PIEMNATIVEINSTR const pRet = iemExecMemAllocatorAllocUnitsInChunkInner(pExecMemAllocator, idxChunk,
1015	cReqUnits, pTb, ppaExec, ppChunkCtx);
1016	if (pRet)
1017	{
1018	STAM_PROFILE_STOP(&pExecMemAllocator->StatAlloc, a);
1019	#ifdef VBOX_WITH_STATISTICS
1020	pExecMemAllocator->cbUnusable += (cReqUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT) - cbReq;
1021	#endif
1022	return pRet;
1023	}
1024	}
1025	if (cMinFreePass <= cReqUnits * 2)
1026	break;
1027	cMaxFreePass = cMinFreePass - 1;
1028	cMinFreePass = cReqUnits * 2;
1029	}
1030	}
1031
1032	/*
1033	* Can we grow it with another chunk?
1034	*/
1035	if (pExecMemAllocator->cChunks < pExecMemAllocator->cMaxChunks)
1036	{
1037	int rc = iemExecMemAllocatorGrow(pVCpu, pExecMemAllocator);
1038	AssertLogRelRCReturn(rc, NULL);
1039
1040	uint32_t const idxChunk = pExecMemAllocator->cChunks - 1;
1041	PIEMNATIVEINSTR const pRet = iemExecMemAllocatorAllocUnitsInChunkInner(pExecMemAllocator, idxChunk, cReqUnits, pTb,
1042	ppaExec, ppChunkCtx);
1043	if (pRet)
1044	{
1045	STAM_PROFILE_STOP(&pExecMemAllocator->StatAlloc, a);
1046	#ifdef VBOX_WITH_STATISTICS
1047	pExecMemAllocator->cbUnusable += (cReqUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT) - cbReq;
1048	#endif
1049	return pRet;
1050	}
1051	AssertFailed();
1052	}
1053
1054	/*
1055	* Try prune native TBs once.
1056	*/
1057	if (iIteration == 0)
1058	{
1059	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING
1060	iemExecMemAllocatorPrune(pVCpu, pExecMemAllocator);
1061	#else
1062	/* No header included in the instruction count here. */
1063	uint32_t const cNeededInstrs = RT_ALIGN_32(cbReq, IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE) / sizeof(IEMNATIVEINSTR);
1064	iemTbAllocatorFreeupNativeSpace(pVCpu, cNeededInstrs);
1065	#endif
1066	}
1067	else
1068	{
1069	STAM_REL_COUNTER_INC(&pVCpu->iem.s.StatNativeExecMemInstrBufAllocFailed);
1070	STAM_PROFILE_STOP(&pExecMemAllocator->StatAlloc, a);
1071	return NULL;
1072	}
1073	}
1074	}
1075
1076
1077	/** This is a hook to ensure the instruction cache is properly flushed before the code in the memory
1078	* given by @a pv and @a cb is executed */
1079	DECLHIDDEN(void) iemExecMemAllocatorReadyForUse(PVMCPUCC pVCpu, void *pv, size_t cb) RT_NOEXCEPT
1080	{
1081	#ifdef RT_OS_DARWIN
1082	/*
1083	* We need to synchronize the stuff we wrote to the data cache with the
1084	* instruction cache, since these aren't coherent on arm (or at least not
1085	* on Apple Mn CPUs).
1086	*
1087	* Note! Since we don't any share JIT'ed code with the other CPUs, we don't
1088	* really care whether the dcache is fully flushed back to memory. It
1089	* only needs to hit the level 2 cache, which the level 1 instruction
1090	* and data caches seems to be sharing. In ARM terms, we need to reach
1091	* a point of unification (PoU), rather than a point of coherhency (PoC).
1092	*
1093	* https://developer.apple.com/documentation/apple-silicon/porting-just-in-time-compilers-to-apple-silicon
1094	*
1095	* https://developer.arm.com/documentation/den0013/d/Caches/Point-of-coherency-and-unification
1096	*
1097	* Experimenting with the approach used by sys_icache_invalidate() and
1098	* tweaking it a little, could let us shave off a bit of effort. The thing
1099	* that slows the apple code down on an M2 (runing Sonoma 13.4), seems to
1100	* the 'DSB ISH' instructions performed every 20 icache line flushes.
1101	* Skipping these saves ~100ns or more per TB when profiling the native
1102	* recompiler on the TBs from a win11 full boot-desktop-shutdow sequence.
1103	* Thus we will leave DCACHE_ICACHE_SYNC_WITH_WITH_IVAU_DSB undefined if we
1104	* can.
1105	*
1106	* There appears not to be much difference between DSB options 'ISH',
1107	* 'ISHST', 'NSH' and 'NSHST'. The latter is theoretically all we need, so
1108	* we'll use that one.
1109	*
1110	* See https://developer.arm.com/documentation/100941/0101/Barriers for
1111	* details on the barrier options.
1112	*
1113	* Note! The CFG value "/IEM/HostICacheInvalidationViaHostAPI" can be used
1114	* to disabling the experimental code should it misbehave.
1115	*/
1116	uint8_t const fHostICacheInvalidation = pVCpu->iem.s.fHostICacheInvalidation;
1117	if (!(fHostICacheInvalidation & IEMNATIVE_ICACHE_F_USE_HOST_API))
1118	{
1119	# define DCACHE_ICACHE_SYNC_DSB_OPTION "nshst"
1120	/# define DCACHE_ICACHE_SYNC_WITH_WITH_IVAU_DSB/
1121
1122	/* Skipping this is fine, but doesn't impact perf much. */
1123	__asm__ __volatile__("dsb " DCACHE_ICACHE_SYNC_DSB_OPTION);
1124
1125	/* Invalidate the icache for the range [pv,pv+cb). */
1126	# ifdef DCACHE_ICACHE_SYNC_WITH_WITH_IVAU_DSB
1127	size_t const cIvauDsbEvery= 20;
1128	unsigned cDsb = cIvauDsbEvery;
1129	# endif
1130	size_t const cbCacheLine = 64;
1131	size_t cbInvalidate = cb + ((uintptr_t)pv & (cbCacheLine - 1)) ;
1132	size_t cCacheLines = RT_ALIGN_Z(cbInvalidate, cbCacheLine) / cbCacheLine;
1133	uintptr_t uPtr = (uintptr_t)pv & ~(uintptr_t)(cbCacheLine - 1);
1134	for (;; uPtr += cbCacheLine)
1135	{
1136	__asm__ /__volatile__/("ic ivau, %0" : : "r" (uPtr));
1137	cCacheLines -= 1;
1138	if (!cCacheLines)
1139	break;
1140	# ifdef DCACHE_ICACHE_SYNC_WITH_WITH_IVAU_DSB
1141	cDsb -= 1;
1142	if (cDsb != 0)
1143	{ /* likely */ }
1144	else
1145	{
1146	__asm__ __volatile__("dsb " DCACHE_ICACHE_SYNC_DSB_OPTION);
1147	cDsb = cIvauDsbEvery;
1148	}
1149	# endif
1150	}
1151
1152	/*
1153	* The DSB here is non-optional it seems.
1154	*
1155	* The following ISB can be omitted on M2 without any obvious sideeffects,
1156	* it produces better number in the above mention profiling scenario.
1157	* This could be related to the kHasICDSB flag in cpu_capabilities.h,
1158	* but it doesn't look like that flag is set here (M2, Sonoma 13.4).
1159	*
1160	* I've made the inclusion of the ISH barrier as configurable and with
1161	* a default of skipping it.
1162	*/
1163	if (!(fHostICacheInvalidation & IEMNATIVE_ICACHE_F_END_WITH_ISH))
1164	__asm__ __volatile__("dsb " DCACHE_ICACHE_SYNC_DSB_OPTION
1165	::: "memory");
1166	else
1167	__asm__ __volatile__("dsb " DCACHE_ICACHE_SYNC_DSB_OPTION "\n\t"
1168	"isb"
1169	::: "memory");
1170	}
1171	else
1172	sys_icache_invalidate(pv, cb);
1173
1174	#elif defined(RT_OS_LINUX) && defined(RT_ARCH_ARM64)
1175	RT_NOREF(pVCpu);
1176
1177	/* There is __builtin___clear_cache() but it flushes both the instruction and data cache, so do it manually. */
1178	static uint32_t s_u32CtrEl0 = 0;
1179	if (!s_u32CtrEl0)
1180	asm volatile ("mrs %0, ctr_el0":"=r" (s_u32CtrEl0));
1181	uintptr_t cbICacheLine = (uintptr_t)4 << (s_u32CtrEl0 & 0xf);
1182
1183	uintptr_t pb = (uintptr_t)pv & ~(cbICacheLine - 1);
1184	for (; pb < (uintptr_t)pv + cb; pb += cbICacheLine)
1185	asm volatile ("ic ivau, %0" : : "r" (pb) : "memory");
1186
1187	asm volatile ("dsb ish\n\t isb\n\t" : : : "memory");
1188
1189	#else
1190	RT_NOREF(pVCpu, pv, cb);
1191	#endif
1192	}
1193
1194
1195	/**
1196	* Frees executable memory.
1197	*/
1198	DECLHIDDEN(void) iemExecMemAllocatorFree(PVMCPU pVCpu, void *pv, size_t cb) RT_NOEXCEPT
1199	{
1200	PIEMEXECMEMALLOCATOR pExecMemAllocator = pVCpu->iem.s.pExecMemAllocatorR3;
1201	Assert(pExecMemAllocator && pExecMemAllocator->uMagic == IEMEXECMEMALLOCATOR_MAGIC);
1202	AssertPtr(pv);
1203	#ifdef VBOX_WITH_STATISTICS
1204	size_t const cbOrig = cb;
1205	#endif
1206	#ifndef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
1207	Assert(!((uintptr_t)pv & (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1)));
1208
1209	/* Align the size as we did when allocating the block. */
1210	cb = RT_ALIGN_Z(cb, IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE);
1211
1212	#else
1213	PIEMEXECMEMALLOCHDR pHdr = (PIEMEXECMEMALLOCHDR)pv - 1;
1214	Assert(!((uintptr_t)pHdr & (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1)));
1215	AssertReturnVoid(pHdr->uMagic == IEMEXECMEMALLOCHDR_MAGIC);
1216	uint32_t const idxChunk = pHdr->idxChunk;
1217	AssertReturnVoid(idxChunk < pExecMemAllocator->cChunks);
1218	pv = pHdr;
1219
1220	/* Adjust and align the size to cover the whole allocation area. */
1221	cb = RT_ALIGN_Z(cb + sizeof(*pHdr), IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE);
1222	#endif
1223
1224	/* Free it / assert sanity. */
1225	bool fFound = false;
1226	uint32_t const cbChunk = pExecMemAllocator->cbChunk;
1227	#ifndef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
1228	uint32_t const cChunks = pExecMemAllocator->cChunks;
1229	for (uint32_t idxChunk = 0; idxChunk < cChunks; idxChunk++)
1230	#endif
1231	{
1232	uintptr_t const offChunk = (uintptr_t)pv - (uintptr_t)pExecMemAllocator->aChunks[idxChunk].pvChunkRx;
1233	fFound = offChunk < cbChunk;
1234	if (fFound)
1235	{
1236	uint32_t const idxFirst = (uint32_t)offChunk >> IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
1237	uint32_t const cReqUnits = (uint32_t)cb >> IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
1238
1239	/* Check that it's valid and free it. */
1240	uint64_t * const pbmAlloc = &pExecMemAllocator->pbmAlloc[pExecMemAllocator->cBitmapElementsPerChunk * idxChunk];
1241	AssertReturnVoid(ASMBitTest(pbmAlloc, idxFirst));
1242	for (uint32_t i = 1; i < cReqUnits; i++)
1243	AssertReturnVoid(ASMBitTest(pbmAlloc, idxFirst + i));
1244	ASMBitClearRange(pbmAlloc, idxFirst, idxFirst + cReqUnits);
1245
1246	/* Invalidate the header using the writeable memory view. */
1247	pHdr = (PIEMEXECMEMALLOCHDR)((uintptr_t)pExecMemAllocator->aChunks[idxChunk].pvChunkRw + offChunk);
1248	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
1249	pHdr->uMagic = 0;
1250	pHdr->idxChunk = 0;
1251	pHdr->pTb = NULL;
1252	#endif
1253	pExecMemAllocator->aChunks[idxChunk].cFreeUnits += cReqUnits;
1254	pExecMemAllocator->aChunks[idxChunk].idxFreeHint = idxFirst;
1255
1256	/* Update the stats. */
1257	pExecMemAllocator->cbAllocated -= cb;
1258	pExecMemAllocator->cbFree += cb;
1259	pExecMemAllocator->cAllocations -= 1;
1260	#ifdef VBOX_WITH_STATISTICS
1261	pExecMemAllocator->cbUnusable -= (cReqUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT) - cbOrig;
1262	#endif
1263	return;
1264	}
1265	}
1266	AssertFailed();
1267	}
1268
1269
1270	/**
1271	* Interface used by iemNativeRecompileAttachExecMemChunkCtx and unwind info
1272	* generators.
1273	*/
1274	DECLHIDDEN(PIEMNATIVEINSTR)
1275	iemExecMemAllocatorAllocFromChunk(PVMCPU pVCpu, uint32_t idxChunk, uint32_t cbReq, PIEMNATIVEINSTR *ppaExec)
1276	{
1277	PIEMEXECMEMALLOCATOR pExecMemAllocator = pVCpu->iem.s.pExecMemAllocatorR3;
1278	AssertReturn(idxChunk < pExecMemAllocator->cChunks, NULL);
1279	Assert(cbReq < _1M);
1280	return iemExecMemAllocatorAllocBytesInChunk(pExecMemAllocator, idxChunk, cbReq, ppaExec);
1281	}
1282
1283
1284	/**
1285	* For getting the per-chunk context detailing common code for a TB.
1286	*
1287	* This is for use by the disassembler.
1288	*/
1289	DECLHIDDEN(PCIEMNATIVEPERCHUNKCTX) iemExecMemGetTbChunkCtx(PVMCPU pVCpu, PCIEMTB pTb)
1290	{
1291	PIEMEXECMEMALLOCATOR pExecMemAllocator = pVCpu->iem.s.pExecMemAllocatorR3;
1292	if ((pTb->fFlags & IEMTB_F_TYPE_MASK) == IEMTB_F_TYPE_NATIVE)
1293	{
1294	uintptr_t const uAddress = (uintptr_t)pTb->Native.paInstructions;
1295	uint32_t const cbChunk = pExecMemAllocator->cbChunk;
1296	uint32_t idxChunk = pExecMemAllocator->cChunks;
1297	while (idxChunk-- > 0)
1298	if (uAddress - (uintptr_t)pExecMemAllocator->aChunks[idxChunk].pvChunkRx < cbChunk)
1299	return pExecMemAllocator->aChunks[idxChunk].pCtx;
1300	}
1301	return NULL;
1302	}
1303
1304
1305	#ifdef IN_RING3
1306	# ifdef RT_OS_WINDOWS
1307
1308	/**
1309	* Initializes the unwind info structures for windows hosts.
1310	*/
1311	static int
1312	iemExecMemAllocatorInitAndRegisterUnwindInfoForChunk(PVMCPUCC pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator,
1313	void *pvChunk, uint32_t idxChunk)
1314	{
1315	RT_NOREF(pVCpu);
1316
1317	/*
1318	* The AMD64 unwind opcodes.
1319	*
1320	* This is a program that starts with RSP after a RET instruction that
1321	* ends up in recompiled code, and the operations we describe here will
1322	* restore all non-volatile registers and bring RSP back to where our
1323	* RET address is. This means it's reverse order from what happens in
1324	* the prologue.
1325	*
1326	* Note! Using a frame register approach here both because we have one
1327	* and but mainly because the UWOP_ALLOC_LARGE argument values
1328	* would be a pain to write initializers for. On the positive
1329	* side, we're impervious to changes in the the stack variable
1330	* area can can deal with dynamic stack allocations if necessary.
1331	*/
1332	static const IMAGE_UNWIND_CODE s_aOpcodes[] =
1333	{
1334	{ { 16, IMAGE_AMD64_UWOP_SET_FPREG, 0 } }, /* RSP = RBP - FrameOffset * 10 (0x60) */
1335	{ { 16, IMAGE_AMD64_UWOP_ALLOC_SMALL, 0 } }, /* RSP += 8; */
1336	{ { 14, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_x15 } }, /* R15 = [RSP]; RSP += 8; */
1337	{ { 12, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_x14 } }, /* R14 = [RSP]; RSP += 8; */
1338	{ { 10, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_x13 } }, /* R13 = [RSP]; RSP += 8; */
1339	{ { 8, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_x12 } }, /* R12 = [RSP]; RSP += 8; */
1340	{ { 7, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_xDI } }, /* RDI = [RSP]; RSP += 8; */
1341	{ { 6, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_xSI } }, /* RSI = [RSP]; RSP += 8; */
1342	{ { 5, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_xBX } }, /* RBX = [RSP]; RSP += 8; */
1343	{ { 4, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_xBP } }, /* RBP = [RSP]; RSP += 8; */
1344	};
1345	union
1346	{
1347	IMAGE_UNWIND_INFO Info;
1348	uint8_t abPadding[RT_UOFFSETOF(IMAGE_UNWIND_INFO, aOpcodes) + 16];
1349	} s_UnwindInfo =
1350	{
1351	{
1352	/* .Version = */ 1,
1353	/* .Flags = */ 0,
1354	/* .SizeOfProlog = / 16, / whatever */
1355	/* .CountOfCodes = */ RT_ELEMENTS(s_aOpcodes),
1356	/* .FrameRegister = */ X86_GREG_xBP,
1357	/* .FrameOffset = / (-IEMNATIVE_FP_OFF_LAST_PUSH + 8) / 16 / we're off by one slot. sigh. */,
1358	}
1359	};
1360	AssertCompile(-IEMNATIVE_FP_OFF_LAST_PUSH < 240 && -IEMNATIVE_FP_OFF_LAST_PUSH > 0);
1361	AssertCompile((-IEMNATIVE_FP_OFF_LAST_PUSH & 0xf) == 8);
1362
1363	/*
1364	* Calc how much space we need and allocate it off the exec heap.
1365	*/
1366	unsigned const cFunctionEntries = 1;
1367	unsigned const cbUnwindInfo = sizeof(s_aOpcodes) + RT_UOFFSETOF(IMAGE_UNWIND_INFO, aOpcodes);
1368	unsigned const cbNeeded = sizeof(IMAGE_RUNTIME_FUNCTION_ENTRY) * cFunctionEntries + cbUnwindInfo;
1369	PIMAGE_RUNTIME_FUNCTION_ENTRY const paFunctions
1370	= (PIMAGE_RUNTIME_FUNCTION_ENTRY)iemExecMemAllocatorAllocBytesInChunk(pExecMemAllocator, idxChunk, cbNeeded, NULL);
1371	AssertReturn(paFunctions, VERR_INTERNAL_ERROR_5);
1372	pExecMemAllocator->aChunks[idxChunk].pvUnwindInfo = paFunctions;
1373
1374	/*
1375	* Initialize the structures.
1376	*/
1377	PIMAGE_UNWIND_INFO const pInfo = (PIMAGE_UNWIND_INFO)&paFunctions[cFunctionEntries];
1378
1379	paFunctions[0].BeginAddress = 0;
1380	paFunctions[0].EndAddress = pExecMemAllocator->cbChunk;
1381	paFunctions[0].UnwindInfoAddress = (uint32_t)((uintptr_t)pInfo - (uintptr_t)pvChunk);
1382
1383	memcpy(pInfo, &s_UnwindInfo, RT_UOFFSETOF(IMAGE_UNWIND_INFO, aOpcodes));
1384	memcpy(&pInfo->aOpcodes[0], s_aOpcodes, sizeof(s_aOpcodes));
1385
1386	/*
1387	* Register it.
1388	*/
1389	uint8_t fRet = RtlAddFunctionTable(paFunctions, cFunctionEntries, (uintptr_t)pvChunk);
1390	AssertReturn(fRet, VERR_INTERNAL_ERROR_3); /* Nothing to clean up on failure, since its within the chunk itself. */
1391
1392	return VINF_SUCCESS;
1393	}
1394
1395
1396	# else /* !RT_OS_WINDOWS */
1397
1398	/**
1399	* Emits a LEB128 encoded value between -0x2000 and 0x2000 (both exclusive).
1400	*/
1401	DECLINLINE(RTPTRUNION) iemDwarfPutLeb128(RTPTRUNION Ptr, int32_t iValue)
1402	{
1403	if (iValue >= 64)
1404	{
1405	Assert(iValue < 0x2000);
1406	*Ptr.pb++ = ((uint8_t)iValue & 0x7f) \| 0x80;
1407	*Ptr.pb++ = (uint8_t)(iValue >> 7) & 0x3f;
1408	}
1409	else if (iValue >= 0)
1410	*Ptr.pb++ = (uint8_t)iValue;
1411	else if (iValue > -64)
1412	*Ptr.pb++ = ((uint8_t)iValue & 0x3f) \| 0x40;
1413	else
1414	{
1415	Assert(iValue > -0x2000);
1416	*Ptr.pb++ = ((uint8_t)iValue & 0x7f) \| 0x80;
1417	*Ptr.pb++ = ((uint8_t)(iValue >> 7) & 0x3f) \| 0x40;
1418	}
1419	return Ptr;
1420	}
1421
1422
1423	/**
1424	* Emits an ULEB128 encoded value (up to 64-bit wide).
1425	*/
1426	DECLINLINE(RTPTRUNION) iemDwarfPutUleb128(RTPTRUNION Ptr, uint64_t uValue)
1427	{
1428	while (uValue >= 0x80)
1429	{
1430	*Ptr.pb++ = ((uint8_t)uValue & 0x7f) \| 0x80;
1431	uValue >>= 7;
1432	}
1433	*Ptr.pb++ = (uint8_t)uValue;
1434	return Ptr;
1435	}
1436
1437
1438	/**
1439	* Emits a CFA rule as register @a uReg + offset @a off.
1440	*/
1441	DECLINLINE(RTPTRUNION) iemDwarfPutCfaDefCfa(RTPTRUNION Ptr, uint32_t uReg, uint32_t off)
1442	{
1443	*Ptr.pb++ = DW_CFA_def_cfa;
1444	Ptr = iemDwarfPutUleb128(Ptr, uReg);
1445	Ptr = iemDwarfPutUleb128(Ptr, off);
1446	return Ptr;
1447	}
1448
1449
1450	/**
1451	* Emits a register (@a uReg) save location:
1452	* CFA + @a off * data_alignment_factor
1453	*/
1454	DECLINLINE(RTPTRUNION) iemDwarfPutCfaOffset(RTPTRUNION Ptr, uint32_t uReg, uint32_t off)
1455	{
1456	if (uReg < 0x40)
1457	*Ptr.pb++ = DW_CFA_offset \| uReg;
1458	else
1459	{
1460	*Ptr.pb++ = DW_CFA_offset_extended;
1461	Ptr = iemDwarfPutUleb128(Ptr, uReg);
1462	}
1463	Ptr = iemDwarfPutUleb128(Ptr, off);
1464	return Ptr;
1465	}
1466
1467
1468	# if 0 /* unused */
1469	/**
1470	* Emits a register (@a uReg) save location, using signed offset:
1471	* CFA + @a offSigned * data_alignment_factor
1472	*/
1473	DECLINLINE(RTPTRUNION) iemDwarfPutCfaSignedOffset(RTPTRUNION Ptr, uint32_t uReg, int32_t offSigned)
1474	{
1475	*Ptr.pb++ = DW_CFA_offset_extended_sf;
1476	Ptr = iemDwarfPutUleb128(Ptr, uReg);
1477	Ptr = iemDwarfPutLeb128(Ptr, offSigned);
1478	return Ptr;
1479	}
1480	# endif
1481
1482
1483	/**
1484	* Initializes the unwind info section for non-windows hosts.
1485	*/
1486	static int
1487	iemExecMemAllocatorInitAndRegisterUnwindInfoForChunk(PVMCPUCC pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator,
1488	void *pvChunk, uint32_t idxChunk)
1489	{
1490	PIEMEXECMEMCHUNKEHFRAME const pEhFrame = &pExecMemAllocator->paEhFrames[idxChunk];
1491	pExecMemAllocator->aChunks[idxChunk].pvUnwindInfo = pEhFrame; /* not necessary, but whatever */
1492
1493	RTPTRUNION Ptr = { pEhFrame->abEhFrame };
1494
1495	/*
1496	* Generate the CIE first.
1497	*/
1498	# ifdef IEMNATIVE_USE_LIBUNWIND /* libunwind (llvm, darwin) only supports v1 and v3. */
1499	uint8_t const iDwarfVer = 3;
1500	# else
1501	uint8_t const iDwarfVer = 4;
1502	# endif
1503	RTPTRUNION const PtrCie = Ptr;
1504	Ptr.pu32++ = 123; / The CIE length will be determined later. */
1505	Ptr.pu32++ = 0 /UINT32_MAX/; / I'm a CIE in .eh_frame speak. */
1506	Ptr.pb++ = iDwarfVer; / DwARF version */
1507	Ptr.pb++ = 0; / Augmentation. */
1508	if (iDwarfVer >= 4)
1509	{
1510	Ptr.pb++ = sizeof(uintptr_t); / Address size. */
1511	Ptr.pb++ = 0; / Segment selector size. */
1512	}
1513	# ifdef RT_ARCH_AMD64
1514	Ptr = iemDwarfPutLeb128(Ptr, 1); /* Code alignment factor (LEB128 = 1). */
1515	# else
1516	Ptr = iemDwarfPutLeb128(Ptr, 4); /* Code alignment factor (LEB128 = 4). */
1517	# endif
1518	Ptr = iemDwarfPutLeb128(Ptr, -8); /* Data alignment factor (LEB128 = -8). */
1519	# ifdef RT_ARCH_AMD64
1520	Ptr = iemDwarfPutUleb128(Ptr, DWREG_AMD64_RA); /* Return address column (ULEB128) */
1521	# elif defined(RT_ARCH_ARM64)
1522	Ptr = iemDwarfPutUleb128(Ptr, DWREG_ARM64_LR); /* Return address column (ULEB128) */
1523	# else
1524	# error "port me"
1525	# endif
1526	/* Initial instructions: */
1527	# ifdef RT_ARCH_AMD64
1528	Ptr = iemDwarfPutCfaDefCfa(Ptr, DWREG_AMD64_RBP, 16); /* CFA = RBP + 0x10 - first stack parameter */
1529	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_RA, 1); /* Ret RIP = [CFA + 1-8] /
1530	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_RBP, 2); /* RBP = [CFA + 2-8] /
1531	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_RBX, 3); /* RBX = [CFA + 3-8] /
1532	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_R12, 4); /* R12 = [CFA + 4-8] /
1533	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_R13, 5); /* R13 = [CFA + 5-8] /
1534	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_R14, 6); /* R14 = [CFA + 6-8] /
1535	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_R15, 7); /* R15 = [CFA + 7-8] /
1536	# elif defined(RT_ARCH_ARM64)
1537	# if 1
1538	Ptr = iemDwarfPutCfaDefCfa(Ptr, DWREG_ARM64_BP, 16); /* CFA = BP + 0x10 - first stack parameter */
1539	# else
1540	Ptr = iemDwarfPutCfaDefCfa(Ptr, DWREG_ARM64_SP, IEMNATIVE_FRAME_VAR_SIZE + IEMNATIVE_FRAME_SAVE_REG_SIZE);
1541	# endif
1542	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_LR, 1); /* Ret PC = [CFA + 1-8] /
1543	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_BP, 2); /* Ret BP = [CFA + 2-8] /
1544	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X28, 3); /* X28 = [CFA + 3-8] /
1545	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X27, 4); /* X27 = [CFA + 4-8] /
1546	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X26, 5); /* X26 = [CFA + 5-8] /
1547	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X25, 6); /* X25 = [CFA + 6-8] /
1548	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X24, 7); /* X24 = [CFA + 7-8] /
1549	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X23, 8); /* X23 = [CFA + 8-8] /
1550	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X22, 9); /* X22 = [CFA + 9-8] /
1551	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X21, 10); /* X21 = [CFA +10-8] /
1552	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X20, 11); /* X20 = [CFA +11-8] /
1553	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X19, 12); /* X19 = [CFA +12-8] /
1554	AssertCompile(IEMNATIVE_FRAME_SAVE_REG_SIZE / 8 == 12);
1555	/** @todo we we need to do something about clearing DWREG_ARM64_RA_SIGN_STATE or something? */
1556	# else
1557	# error "port me"
1558	# endif
1559	while ((Ptr.u - PtrCie.u) & 3)
1560	*Ptr.pb++ = DW_CFA_nop;
1561	/* Finalize the CIE size. */
1562	*PtrCie.pu32 = Ptr.u - PtrCie.u - sizeof(uint32_t);
1563
1564	/*
1565	* Generate an FDE for the whole chunk area.
1566	*/
1567	# ifdef IEMNATIVE_USE_LIBUNWIND
1568	pEhFrame->offFda = Ptr.u - (uintptr_t)&pEhFrame->abEhFrame[0];
1569	# endif
1570	RTPTRUNION const PtrFde = Ptr;
1571	Ptr.pu32++ = 123; / The CIE length will be determined later. */
1572	Ptr.pu32 = Ptr.u - PtrCie.u; / Negated self relative CIE address. */
1573	Ptr.pu32++;
1574	Ptr.pu64++ = (uintptr_t)pvChunk; / Absolute start PC of this FDE. */
1575	Ptr.pu64++ = pExecMemAllocator->cbChunk; / PC range length for this PDE. */
1576	# if 0 /* not requried for recent libunwind.dylib nor recent libgcc/glib. */
1577	*Ptr.pb++ = DW_CFA_nop;
1578	# endif
1579	while ((Ptr.u - PtrFde.u) & 3)
1580	*Ptr.pb++ = DW_CFA_nop;
1581	/* Finalize the FDE size. */
1582	*PtrFde.pu32 = Ptr.u - PtrFde.u - sizeof(uint32_t);
1583
1584	/* Terminator entry. */
1585	*Ptr.pu32++ = 0;
1586	Ptr.pu32++ = 0; / just to be sure... */
1587	Assert(Ptr.u - (uintptr_t)&pEhFrame->abEhFrame[0] <= sizeof(pEhFrame->abEhFrame));
1588
1589	/*
1590	* Register it.
1591	*/
1592	# ifdef IEMNATIVE_USE_LIBUNWIND
1593	__register_frame(&pEhFrame->abEhFrame[pEhFrame->offFda]);
1594	# else
1595	memset(pEhFrame->abObject, 0xf6, sizeof(pEhFrame->abObject)); /* color the memory to better spot usage */
1596	__register_frame_info(pEhFrame->abEhFrame, pEhFrame->abObject);
1597	# endif
1598
1599	# ifdef IEMNATIVE_USE_GDB_JIT
1600	/*
1601	* Now for telling GDB about this (experimental).
1602	*
1603	* This seems to work best with ET_DYN.
1604	*/
1605	GDBJITSYMFILE * const pSymFile = (GDBJITSYMFILE *)iemExecMemAllocatorAllocBytesInChunk(pExecMemAllocator, idxChunk,
1606	sizeof(GDBJITSYMFILE), NULL);
1607	AssertReturn(pSymFile, VERR_INTERNAL_ERROR_5);
1608	unsigned const offSymFileInChunk = (uintptr_t)pSymFile - (uintptr_t)pvChunk;
1609
1610	RT_ZERO(*pSymFile);
1611
1612	/*
1613	* The ELF header:
1614	*/
1615	pSymFile->EHdr.e_ident[0] = ELFMAG0;
1616	pSymFile->EHdr.e_ident[1] = ELFMAG1;
1617	pSymFile->EHdr.e_ident[2] = ELFMAG2;
1618	pSymFile->EHdr.e_ident[3] = ELFMAG3;
1619	pSymFile->EHdr.e_ident[EI_VERSION] = EV_CURRENT;
1620	pSymFile->EHdr.e_ident[EI_CLASS] = ELFCLASS64;
1621	pSymFile->EHdr.e_ident[EI_DATA] = ELFDATA2LSB;
1622	pSymFile->EHdr.e_ident[EI_OSABI] = ELFOSABI_NONE;
1623	# ifdef IEMNATIVE_USE_GDB_JIT_ET_DYN
1624	pSymFile->EHdr.e_type = ET_DYN;
1625	# else
1626	pSymFile->EHdr.e_type = ET_REL;
1627	# endif
1628	# ifdef RT_ARCH_AMD64
1629	pSymFile->EHdr.e_machine = EM_AMD64;
1630	# elif defined(RT_ARCH_ARM64)
1631	pSymFile->EHdr.e_machine = EM_AARCH64;
1632	# else
1633	# error "port me"
1634	# endif
1635	pSymFile->EHdr.e_version = 1; /?/
1636	pSymFile->EHdr.e_entry = 0;
1637	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1638	pSymFile->EHdr.e_phoff = RT_UOFFSETOF(GDBJITSYMFILE, aPhdrs);
1639	# else
1640	pSymFile->EHdr.e_phoff = 0;
1641	# endif
1642	pSymFile->EHdr.e_shoff = sizeof(pSymFile->EHdr);
1643	pSymFile->EHdr.e_flags = 0;
1644	pSymFile->EHdr.e_ehsize = sizeof(pSymFile->EHdr);
1645	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1646	pSymFile->EHdr.e_phentsize = sizeof(pSymFile->aPhdrs[0]);
1647	pSymFile->EHdr.e_phnum = RT_ELEMENTS(pSymFile->aPhdrs);
1648	# else
1649	pSymFile->EHdr.e_phentsize = 0;
1650	pSymFile->EHdr.e_phnum = 0;
1651	# endif
1652	pSymFile->EHdr.e_shentsize = sizeof(pSymFile->aShdrs[0]);
1653	pSymFile->EHdr.e_shnum = RT_ELEMENTS(pSymFile->aShdrs);
1654	pSymFile->EHdr.e_shstrndx = 0; /* set later */
1655
1656	uint32_t offStrTab = 0;
1657	#define APPEND_STR(a_szStr) do { \
1658	memcpy(&pSymFile->szzStrTab[offStrTab], a_szStr, sizeof(a_szStr)); \
1659	offStrTab += sizeof(a_szStr); \
1660	Assert(offStrTab < sizeof(pSymFile->szzStrTab)); \
1661	} while (0)
1662	#define APPEND_STR_FMT(a_szStr, ...) do { \
1663	offStrTab += RTStrPrintf(&pSymFile->szzStrTab[offStrTab], sizeof(pSymFile->szzStrTab) - offStrTab, a_szStr, __VA_ARGS__); \
1664	offStrTab++; \
1665	Assert(offStrTab < sizeof(pSymFile->szzStrTab)); \
1666	} while (0)
1667
1668	/*
1669	* Section headers.
1670	*/
1671	/* Section header #0: NULL */
1672	unsigned i = 0;
1673	APPEND_STR("");
1674	RT_ZERO(pSymFile->aShdrs[i]);
1675	i++;
1676
1677	/* Section header: .eh_frame */
1678	pSymFile->aShdrs[i].sh_name = offStrTab;
1679	APPEND_STR(".eh_frame");
1680	pSymFile->aShdrs[i].sh_type = SHT_PROGBITS;
1681	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC \| SHF_EXECINSTR;
1682	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN) \|\| defined(IEMNATIVE_USE_GDB_JIT_ELF_RVAS)
1683	pSymFile->aShdrs[i].sh_offset
1684	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, abEhFrame);
1685	# else
1686	pSymFile->aShdrs[i].sh_addr = (uintptr_t)&pSymFile->abEhFrame[0];
1687	pSymFile->aShdrs[i].sh_offset = 0;
1688	# endif
1689
1690	pSymFile->aShdrs[i].sh_size = sizeof(pEhFrame->abEhFrame);
1691	pSymFile->aShdrs[i].sh_link = 0;
1692	pSymFile->aShdrs[i].sh_info = 0;
1693	pSymFile->aShdrs[i].sh_addralign = 1;
1694	pSymFile->aShdrs[i].sh_entsize = 0;
1695	memcpy(pSymFile->abEhFrame, pEhFrame->abEhFrame, sizeof(pEhFrame->abEhFrame));
1696	i++;
1697
1698	/* Section header: .shstrtab */
1699	unsigned const iShStrTab = i;
1700	pSymFile->EHdr.e_shstrndx = iShStrTab;
1701	pSymFile->aShdrs[i].sh_name = offStrTab;
1702	APPEND_STR(".shstrtab");
1703	pSymFile->aShdrs[i].sh_type = SHT_STRTAB;
1704	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC;
1705	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN) \|\| defined(IEMNATIVE_USE_GDB_JIT_ELF_RVAS)
1706	pSymFile->aShdrs[i].sh_offset
1707	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, szzStrTab);
1708	# else
1709	pSymFile->aShdrs[i].sh_addr = (uintptr_t)&pSymFile->szzStrTab[0];
1710	pSymFile->aShdrs[i].sh_offset = 0;
1711	# endif
1712	pSymFile->aShdrs[i].sh_size = sizeof(pSymFile->szzStrTab);
1713	pSymFile->aShdrs[i].sh_link = 0;
1714	pSymFile->aShdrs[i].sh_info = 0;
1715	pSymFile->aShdrs[i].sh_addralign = 1;
1716	pSymFile->aShdrs[i].sh_entsize = 0;
1717	i++;
1718
1719	/* Section header: .symbols */
1720	pSymFile->aShdrs[i].sh_name = offStrTab;
1721	APPEND_STR(".symtab");
1722	pSymFile->aShdrs[i].sh_type = SHT_SYMTAB;
1723	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC;
1724	pSymFile->aShdrs[i].sh_offset
1725	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, aSymbols);
1726	pSymFile->aShdrs[i].sh_size = sizeof(pSymFile->aSymbols);
1727	pSymFile->aShdrs[i].sh_link = iShStrTab;
1728	pSymFile->aShdrs[i].sh_info = RT_ELEMENTS(pSymFile->aSymbols);
1729	pSymFile->aShdrs[i].sh_addralign = sizeof(pSymFile->aSymbols[0].st_value);
1730	pSymFile->aShdrs[i].sh_entsize = sizeof(pSymFile->aSymbols[0]);
1731	i++;
1732
1733	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1734	/* Section header: .symbols */
1735	pSymFile->aShdrs[i].sh_name = offStrTab;
1736	APPEND_STR(".dynsym");
1737	pSymFile->aShdrs[i].sh_type = SHT_DYNSYM;
1738	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC;
1739	pSymFile->aShdrs[i].sh_offset
1740	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, aDynSyms);
1741	pSymFile->aShdrs[i].sh_size = sizeof(pSymFile->aDynSyms);
1742	pSymFile->aShdrs[i].sh_link = iShStrTab;
1743	pSymFile->aShdrs[i].sh_info = RT_ELEMENTS(pSymFile->aDynSyms);
1744	pSymFile->aShdrs[i].sh_addralign = sizeof(pSymFile->aDynSyms[0].st_value);
1745	pSymFile->aShdrs[i].sh_entsize = sizeof(pSymFile->aDynSyms[0]);
1746	i++;
1747	# endif
1748
1749	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1750	/* Section header: .dynamic */
1751	pSymFile->aShdrs[i].sh_name = offStrTab;
1752	APPEND_STR(".dynamic");
1753	pSymFile->aShdrs[i].sh_type = SHT_DYNAMIC;
1754	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC;
1755	pSymFile->aShdrs[i].sh_offset
1756	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, aDyn);
1757	pSymFile->aShdrs[i].sh_size = sizeof(pSymFile->aDyn);
1758	pSymFile->aShdrs[i].sh_link = iShStrTab;
1759	pSymFile->aShdrs[i].sh_info = 0;
1760	pSymFile->aShdrs[i].sh_addralign = 1;
1761	pSymFile->aShdrs[i].sh_entsize = sizeof(pSymFile->aDyn[0]);
1762	i++;
1763	# endif
1764
1765	/* Section header: .text */
1766	unsigned const iShText = i;
1767	pSymFile->aShdrs[i].sh_name = offStrTab;
1768	APPEND_STR(".text");
1769	pSymFile->aShdrs[i].sh_type = SHT_PROGBITS;
1770	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC \| SHF_EXECINSTR;
1771	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN) \|\| defined(IEMNATIVE_USE_GDB_JIT_ELF_RVAS)
1772	pSymFile->aShdrs[i].sh_offset
1773	= pSymFile->aShdrs[i].sh_addr = sizeof(GDBJITSYMFILE);
1774	# else
1775	pSymFile->aShdrs[i].sh_addr = (uintptr_t)(pSymFile + 1);
1776	pSymFile->aShdrs[i].sh_offset = 0;
1777	# endif
1778	pSymFile->aShdrs[i].sh_size = pExecMemAllocator->cbChunk - offSymFileInChunk - sizeof(GDBJITSYMFILE);
1779	pSymFile->aShdrs[i].sh_link = 0;
1780	pSymFile->aShdrs[i].sh_info = 0;
1781	pSymFile->aShdrs[i].sh_addralign = 1;
1782	pSymFile->aShdrs[i].sh_entsize = 0;
1783	i++;
1784
1785	Assert(i == RT_ELEMENTS(pSymFile->aShdrs));
1786
1787	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1788	/*
1789	* The program headers:
1790	*/
1791	/* Everything in a single LOAD segment: */
1792	i = 0;
1793	pSymFile->aPhdrs[i].p_type = PT_LOAD;
1794	pSymFile->aPhdrs[i].p_flags = PF_X \| PF_R;
1795	pSymFile->aPhdrs[i].p_offset
1796	= pSymFile->aPhdrs[i].p_vaddr
1797	= pSymFile->aPhdrs[i].p_paddr = 0;
1798	pSymFile->aPhdrs[i].p_filesz /* Size of segment in file. */
1799	= pSymFile->aPhdrs[i].p_memsz = pExecMemAllocator->cbChunk - offSymFileInChunk;
1800	pSymFile->aPhdrs[i].p_align = HOST_PAGE_SIZE;
1801	i++;
1802	/* The .dynamic segment. */
1803	pSymFile->aPhdrs[i].p_type = PT_DYNAMIC;
1804	pSymFile->aPhdrs[i].p_flags = PF_R;
1805	pSymFile->aPhdrs[i].p_offset
1806	= pSymFile->aPhdrs[i].p_vaddr
1807	= pSymFile->aPhdrs[i].p_paddr = RT_UOFFSETOF(GDBJITSYMFILE, aDyn);
1808	pSymFile->aPhdrs[i].p_filesz /* Size of segment in file. */
1809	= pSymFile->aPhdrs[i].p_memsz = sizeof(pSymFile->aDyn);
1810	pSymFile->aPhdrs[i].p_align = sizeof(pSymFile->aDyn[0].d_tag);
1811	i++;
1812
1813	Assert(i == RT_ELEMENTS(pSymFile->aPhdrs));
1814
1815	/*
1816	* The dynamic section:
1817	*/
1818	i = 0;
1819	pSymFile->aDyn[i].d_tag = DT_SONAME;
1820	pSymFile->aDyn[i].d_un.d_val = offStrTab;
1821	APPEND_STR_FMT("iem-exec-chunk-%u-%u", pVCpu->idCpu, idxChunk);
1822	i++;
1823	pSymFile->aDyn[i].d_tag = DT_STRTAB;
1824	pSymFile->aDyn[i].d_un.d_ptr = RT_UOFFSETOF(GDBJITSYMFILE, szzStrTab);
1825	i++;
1826	pSymFile->aDyn[i].d_tag = DT_STRSZ;
1827	pSymFile->aDyn[i].d_un.d_val = sizeof(pSymFile->szzStrTab);
1828	i++;
1829	pSymFile->aDyn[i].d_tag = DT_SYMTAB;
1830	pSymFile->aDyn[i].d_un.d_ptr = RT_UOFFSETOF(GDBJITSYMFILE, aDynSyms);
1831	i++;
1832	pSymFile->aDyn[i].d_tag = DT_SYMENT;
1833	pSymFile->aDyn[i].d_un.d_val = sizeof(pSymFile->aDynSyms[0]);
1834	i++;
1835	pSymFile->aDyn[i].d_tag = DT_NULL;
1836	i++;
1837	Assert(i == RT_ELEMENTS(pSymFile->aDyn));
1838	# endif /* IEMNATIVE_USE_GDB_JIT_ET_DYN */
1839
1840	/*
1841	* Symbol tables:
1842	*/
1843	/** @todo gdb doesn't seem to really like this ... */
1844	i = 0;
1845	pSymFile->aSymbols[i].st_name = 0;
1846	pSymFile->aSymbols[i].st_shndx = SHN_UNDEF;
1847	pSymFile->aSymbols[i].st_value = 0;
1848	pSymFile->aSymbols[i].st_size = 0;
1849	pSymFile->aSymbols[i].st_info = ELF64_ST_INFO(STB_LOCAL, STT_NOTYPE);
1850	pSymFile->aSymbols[i].st_other = 0 /* STV_DEFAULT */;
1851	# ifdef IEMNATIVE_USE_GDB_JIT_ET_DYN
1852	pSymFile->aDynSyms[0] = pSymFile->aSymbols[i];
1853	# endif
1854	i++;
1855
1856	pSymFile->aSymbols[i].st_name = 0;
1857	pSymFile->aSymbols[i].st_shndx = SHN_ABS;
1858	pSymFile->aSymbols[i].st_value = 0;
1859	pSymFile->aSymbols[i].st_size = 0;
1860	pSymFile->aSymbols[i].st_info = ELF64_ST_INFO(STB_LOCAL, STT_FILE);
1861	pSymFile->aSymbols[i].st_other = 0 /* STV_DEFAULT */;
1862	i++;
1863
1864	pSymFile->aSymbols[i].st_name = offStrTab;
1865	APPEND_STR_FMT("iem_exec_chunk_%u_%u", pVCpu->idCpu, idxChunk);
1866	# if 0
1867	pSymFile->aSymbols[i].st_shndx = iShText;
1868	pSymFile->aSymbols[i].st_value = 0;
1869	# else
1870	pSymFile->aSymbols[i].st_shndx = SHN_ABS;
1871	pSymFile->aSymbols[i].st_value = (uintptr_t)(pSymFile + 1);
1872	# endif
1873	pSymFile->aSymbols[i].st_size = pSymFile->aShdrs[iShText].sh_size;
1874	pSymFile->aSymbols[i].st_info = ELF64_ST_INFO(STB_GLOBAL, STT_FUNC);
1875	pSymFile->aSymbols[i].st_other = 0 /* STV_DEFAULT */;
1876	# ifdef IEMNATIVE_USE_GDB_JIT_ET_DYN
1877	pSymFile->aDynSyms[1] = pSymFile->aSymbols[i];
1878	pSymFile->aDynSyms[1].st_value = (uintptr_t)(pSymFile + 1);
1879	# endif
1880	i++;
1881
1882	Assert(i == RT_ELEMENTS(pSymFile->aSymbols));
1883	Assert(offStrTab < sizeof(pSymFile->szzStrTab));
1884
1885	/*
1886	* The GDB JIT entry and informing GDB.
1887	*/
1888	pEhFrame->GdbJitEntry.pbSymFile = (uint8_t *)pSymFile;
1889	# if 1
1890	pEhFrame->GdbJitEntry.cbSymFile = pExecMemAllocator->cbChunk - ((uintptr_t)pSymFile - (uintptr_t)pvChunk);
1891	# else
1892	pEhFrame->GdbJitEntry.cbSymFile = sizeof(GDBJITSYMFILE);
1893	# endif
1894
1895	RTOnce(&g_IemNativeGdbJitOnce, iemNativeGdbJitInitOnce, NULL);
1896	RTCritSectEnter(&g_IemNativeGdbJitLock);
1897	pEhFrame->GdbJitEntry.pNext = NULL;
1898	pEhFrame->GdbJitEntry.pPrev = __jit_debug_descriptor.pTail;
1899	if (__jit_debug_descriptor.pTail)
1900	__jit_debug_descriptor.pTail->pNext = &pEhFrame->GdbJitEntry;
1901	else
1902	__jit_debug_descriptor.pHead = &pEhFrame->GdbJitEntry;
1903	__jit_debug_descriptor.pTail = &pEhFrame->GdbJitEntry;
1904	__jit_debug_descriptor.pRelevant = &pEhFrame->GdbJitEntry;
1905
1906	/* Notify GDB: */
1907	__jit_debug_descriptor.enmAction = kGdbJitaction_Register;
1908	__jit_debug_register_code();
1909	__jit_debug_descriptor.enmAction = kGdbJitaction_NoAction;
1910	RTCritSectLeave(&g_IemNativeGdbJitLock);
1911
1912	# else /* !IEMNATIVE_USE_GDB_JIT */
1913	RT_NOREF(pVCpu);
1914	# endif /* !IEMNATIVE_USE_GDB_JIT */
1915
1916	return VINF_SUCCESS;
1917	}
1918
1919	# endif /* !RT_OS_WINDOWS */
1920	#endif /* IN_RING3 */
1921
1922
1923	/**
1924	* Adds another chunk to the executable memory allocator.
1925	*
1926	* This is used by the init code for the initial allocation and later by the
1927	* regular allocator function when it's out of memory.
1928	*/
1929	static int iemExecMemAllocatorGrow(PVMCPUCC pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator)
1930	{
1931	/* Check that we've room for growth. */
1932	uint32_t const idxChunk = pExecMemAllocator->cChunks;
1933	AssertLogRelReturn(idxChunk < pExecMemAllocator->cMaxChunks, VERR_OUT_OF_RESOURCES);
1934
1935	/* Allocate a chunk. */
1936	#ifdef RT_OS_DARWIN
1937	void *pvChunk = RTMemPageAllocEx(pExecMemAllocator->cbChunk, 0);
1938	#else
1939	void *pvChunk = RTMemPageAllocEx(pExecMemAllocator->cbChunk, RTMEMPAGEALLOC_F_EXECUTABLE);
1940	#endif
1941	AssertLogRelReturn(pvChunk, VERR_NO_EXEC_MEMORY);
1942
1943	#ifdef RT_OS_DARWIN
1944	/*
1945	* Because it is impossible to have a RWX memory allocation on macOS try to remap the memory
1946	* chunk readable/executable somewhere else so we can save us the hassle of switching between
1947	* protections when exeuctable memory is allocated.
1948	*/
1949	int rc = VERR_NO_EXEC_MEMORY;
1950	mach_port_t hPortTask = mach_task_self();
1951	mach_vm_address_t AddrChunk = (mach_vm_address_t)pvChunk;
1952	mach_vm_address_t AddrRemapped = 0;
1953	vm_prot_t ProtCur = 0;
1954	vm_prot_t ProtMax = 0;
1955	kern_return_t krc = mach_vm_remap(hPortTask, &AddrRemapped, pExecMemAllocator->cbChunk, 0,
1956	VM_FLAGS_ANYWHERE \| VM_FLAGS_RETURN_DATA_ADDR,
1957	hPortTask, AddrChunk, FALSE, &ProtCur, &ProtMax,
1958	VM_INHERIT_NONE);
1959	if (krc == KERN_SUCCESS)
1960	{
1961	krc = mach_vm_protect(mach_task_self(), AddrRemapped, pExecMemAllocator->cbChunk, FALSE, VM_PROT_READ \| VM_PROT_EXECUTE);
1962	if (krc == KERN_SUCCESS)
1963	rc = VINF_SUCCESS;
1964	else
1965	{
1966	AssertLogRelMsgFailed(("mach_vm_protect -> %d (%#x)\n", krc, krc));
1967	krc = mach_vm_deallocate(hPortTask, AddrRemapped, pExecMemAllocator->cbChunk);
1968	Assert(krc == KERN_SUCCESS);
1969	}
1970	}
1971	else
1972	AssertLogRelMsgFailed(("mach_vm_remap -> %d (%#x)\n", krc, krc));
1973	if (RT_FAILURE(rc))
1974	{
1975	RTMemPageFree(pvChunk, pExecMemAllocator->cbChunk);
1976	return rc;
1977	}
1978
1979	void pvChunkRx = (void )AddrRemapped;
1980	#else
1981	int rc = VINF_SUCCESS;
1982	void *pvChunkRx = pvChunk;
1983	#endif
1984
1985	/*
1986	* Add the chunk.
1987	*
1988	* This must be done before the unwind init so windows can allocate
1989	* memory from the chunk when using the alternative sub-allocator.
1990	*/
1991	pExecMemAllocator->aChunks[idxChunk].pvChunkRw = pvChunk;
1992	pExecMemAllocator->aChunks[idxChunk].pvChunkRx = pvChunkRx;
1993	#ifdef IN_RING3
1994	pExecMemAllocator->aChunks[idxChunk].pvUnwindInfo = NULL;
1995	#endif
1996	pExecMemAllocator->aChunks[idxChunk].cFreeUnits = pExecMemAllocator->cUnitsPerChunk;
1997	pExecMemAllocator->aChunks[idxChunk].idxFreeHint = 0;
1998	memset(&pExecMemAllocator->pbmAlloc[pExecMemAllocator->cBitmapElementsPerChunk * idxChunk],
1999	0, sizeof(pExecMemAllocator->pbmAlloc[0]) * pExecMemAllocator->cBitmapElementsPerChunk);
2000
2001	pExecMemAllocator->cChunks = idxChunk + 1;
2002	pExecMemAllocator->idxChunkHint = idxChunk;
2003
2004	pExecMemAllocator->cbTotal += pExecMemAllocator->cbChunk;
2005	pExecMemAllocator->cbFree += pExecMemAllocator->cbChunk;
2006
2007	/* If there is a chunk context init callback call it. */
2008	rc = iemNativeRecompileAttachExecMemChunkCtx(pVCpu, idxChunk, &pExecMemAllocator->aChunks[idxChunk].pCtx);
2009	#ifdef IN_RING3
2010	/*
2011	* Initialize the unwind information (this cannot really fail atm).
2012	* (This sets pvUnwindInfo.)
2013	*/
2014	if (RT_SUCCESS(rc))
2015	rc = iemExecMemAllocatorInitAndRegisterUnwindInfoForChunk(pVCpu, pExecMemAllocator, pvChunkRx, idxChunk);
2016	#endif
2017	if (RT_SUCCESS(rc))
2018	{ /* likely */ }
2019	else
2020	{
2021	/* Just in case the impossible happens, undo the above up: */
2022	pExecMemAllocator->cbTotal -= pExecMemAllocator->cbChunk;
2023	pExecMemAllocator->cbFree -= pExecMemAllocator->aChunks[idxChunk].cFreeUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
2024	pExecMemAllocator->cChunks = idxChunk;
2025	memset(&pExecMemAllocator->pbmAlloc[pExecMemAllocator->cBitmapElementsPerChunk * idxChunk],
2026	0xff, sizeof(pExecMemAllocator->pbmAlloc[0]) * pExecMemAllocator->cBitmapElementsPerChunk);
2027	pExecMemAllocator->aChunks[idxChunk].pvChunkRw = NULL;
2028	pExecMemAllocator->aChunks[idxChunk].cFreeUnits = 0;
2029
2030	# ifdef RT_OS_DARWIN
2031	krc = mach_vm_deallocate(mach_task_self(), (mach_vm_address_t)pExecMemAllocator->aChunks[idxChunk].pvChunkRx,
2032	pExecMemAllocator->cbChunk);
2033	Assert(krc == KERN_SUCCESS);
2034	# endif
2035
2036	RTMemPageFree(pvChunk, pExecMemAllocator->cbChunk);
2037	return rc;
2038	}
2039
2040	return VINF_SUCCESS;
2041	}
2042
2043
2044	/**
2045	* Initializes the executable memory allocator for native recompilation on the
2046	* calling EMT.
2047	*
2048	* @returns VBox status code.
2049	* @param pVCpu The cross context virtual CPU structure of the calling
2050	* thread.
2051	* @param cbMax The max size of the allocator.
2052	* @param cbInitial The initial allocator size.
2053	* @param cbChunk The chunk size, 0 or UINT32_MAX for default (@a cbMax
2054	* dependent).
2055	*/
2056	int iemExecMemAllocatorInit(PVMCPU pVCpu, uint64_t cbMax, uint64_t cbInitial, uint32_t cbChunk) RT_NOEXCEPT
2057	{
2058	/*
2059	* Validate input.
2060	*/
2061	AssertLogRelMsgReturn(cbMax >= _1M && cbMax <= _4G+_4G, ("cbMax=%RU64 (%RX64)\n", cbMax, cbMax), VERR_OUT_OF_RANGE);
2062	AssertReturn(cbInitial <= cbMax, VERR_OUT_OF_RANGE);
2063	AssertLogRelMsgReturn( cbChunk != UINT32_MAX
2064	\|\| cbChunk == 0
2065	\|\| ( RT_IS_POWER_OF_TWO(cbChunk)
2066	&& cbChunk >= _1M
2067	&& cbChunk <= _256M
2068	&& cbChunk <= cbMax),
2069	("cbChunk=%RU32 (%RX32) cbMax=%RU64\n", cbChunk, cbChunk, cbMax),
2070	VERR_OUT_OF_RANGE);
2071
2072	/*
2073	* Adjust/figure out the chunk size.
2074	*/
2075	if (cbChunk == 0 \|\| cbChunk == UINT32_MAX)
2076	{
2077	if (cbMax >= _256M)
2078	cbChunk = _64M;
2079	else
2080	{
2081	if (cbMax < _16M)
2082	cbChunk = cbMax >= _4M ? _4M : (uint32_t)cbMax;
2083	else
2084	cbChunk = (uint32_t)cbMax / 4;
2085	if (!RT_IS_POWER_OF_TWO(cbChunk))
2086	cbChunk = RT_BIT_32(ASMBitLastSetU32(cbChunk));
2087	}
2088	}
2089	#if defined(RT_OS_AMD64)
2090	Assert(cbChunk <= _2G);
2091	#elif defined(RT_OS_ARM64)
2092	if (cbChunk > _128M)
2093	cbChunk = _128M; /* Max relative branch distance is +/-2^(25+2) = +/-0x8000000 (134 217 728). */
2094	#endif
2095
2096	if (cbChunk > cbMax)
2097	cbMax = cbChunk;
2098	else
2099	cbMax = (cbMax - 1 + cbChunk) / cbChunk * cbChunk;
2100	uint32_t const cMaxChunks = (uint32_t)(cbMax / cbChunk);
2101	AssertLogRelReturn((uint64_t)cMaxChunks * cbChunk == cbMax, VERR_INTERNAL_ERROR_3);
2102
2103	/*
2104	* Allocate and initialize the allocatore instance.
2105	*/
2106	size_t const offBitmaps = RT_ALIGN_Z(RT_UOFFSETOF_DYN(IEMEXECMEMALLOCATOR, aChunks[cMaxChunks]), RT_CACHELINE_SIZE);
2107	size_t const cbBitmaps = (size_t)(cbChunk >> (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT + 3)) * cMaxChunks;
2108	size_t cbNeeded = offBitmaps + cbBitmaps;
2109	AssertCompile(IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT <= 10);
2110	Assert(cbChunk > RT_BIT_32(IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT + 3));
2111	#if defined(IN_RING3) && !defined(RT_OS_WINDOWS)
2112	size_t const offEhFrames = RT_ALIGN_Z(cbNeeded, RT_CACHELINE_SIZE);
2113	cbNeeded += sizeof(IEMEXECMEMCHUNKEHFRAME) * cMaxChunks;
2114	#endif
2115	PIEMEXECMEMALLOCATOR pExecMemAllocator = (PIEMEXECMEMALLOCATOR)RTMemAllocZ(cbNeeded);
2116	AssertLogRelMsgReturn(pExecMemAllocator, ("cbNeeded=%zx cMaxChunks=%#x cbChunk=%#x\n", cbNeeded, cMaxChunks, cbChunk),
2117	VERR_NO_MEMORY);
2118	pExecMemAllocator->uMagic = IEMEXECMEMALLOCATOR_MAGIC;
2119	pExecMemAllocator->cbChunk = cbChunk;
2120	pExecMemAllocator->cMaxChunks = cMaxChunks;
2121	pExecMemAllocator->cChunks = 0;
2122	pExecMemAllocator->idxChunkHint = 0;
2123	pExecMemAllocator->cAllocations = 0;
2124	pExecMemAllocator->cbTotal = 0;
2125	pExecMemAllocator->cbFree = 0;
2126	pExecMemAllocator->cbAllocated = 0;
2127	#ifdef VBOX_WITH_STATISTICS
2128	pExecMemAllocator->cbUnusable = 0;
2129	#endif
2130	pExecMemAllocator->pbmAlloc = (uint64_t *)((uintptr_t)pExecMemAllocator + offBitmaps);
2131	pExecMemAllocator->cUnitsPerChunk = cbChunk >> IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
2132	pExecMemAllocator->cBitmapElementsPerChunk = cbChunk >> (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT + 6);
2133	memset(pExecMemAllocator->pbmAlloc, 0xff, cbBitmaps); /* Mark everything as allocated. Clear when chunks are added. */
2134	#if defined(IN_RING3) && !defined(RT_OS_WINDOWS)
2135	pExecMemAllocator->paEhFrames = (PIEMEXECMEMCHUNKEHFRAME)((uintptr_t)pExecMemAllocator + offEhFrames);
2136	#endif
2137	for (uint32_t i = 0; i < cMaxChunks; i++)
2138	{
2139	pExecMemAllocator->aChunks[i].cFreeUnits = 0;
2140	pExecMemAllocator->aChunks[i].idxFreeHint = 0;
2141	pExecMemAllocator->aChunks[i].pvChunkRw = NULL;
2142	#ifdef IN_RING0
2143	pExecMemAllocator->aChunks[i].hMemObj = NIL_RTR0MEMOBJ;
2144	#else
2145	pExecMemAllocator->aChunks[i].pvUnwindInfo = NULL;
2146	#endif
2147	}
2148	pVCpu->iem.s.pExecMemAllocatorR3 = pExecMemAllocator;
2149
2150	/*
2151	* Do the initial allocations.
2152	*/
2153	while (cbInitial < (uint64_t)pExecMemAllocator->cChunks * pExecMemAllocator->cbChunk)
2154	{
2155	int rc = iemExecMemAllocatorGrow(pVCpu, pExecMemAllocator);
2156	AssertLogRelRCReturn(rc, rc);
2157	}
2158
2159	pExecMemAllocator->idxChunkHint = 0;
2160
2161	/*
2162	* Register statistics.
2163	*/
2164	PUVM const pUVM = pVCpu->pUVCpu->pUVM;
2165	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cAllocations, STAMTYPE_U32, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
2166	"Current number of allocations", "/IEM/CPU%u/re/ExecMem/cAllocations", pVCpu->idCpu);
2167	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cChunks, STAMTYPE_U32, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,
2168	"Currently allocated chunks", "/IEM/CPU%u/re/ExecMem/cChunks", pVCpu->idCpu);
2169	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cMaxChunks, STAMTYPE_U32, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,
2170	"Maximum number of chunks", "/IEM/CPU%u/re/ExecMem/cMaxChunks", pVCpu->idCpu);
2171	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbChunk, STAMTYPE_U32, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
2172	"Allocation chunk size", "/IEM/CPU%u/re/ExecMem/cbChunk", pVCpu->idCpu);
2173	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbAllocated, STAMTYPE_U64, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
2174	"Number of bytes current allocated", "/IEM/CPU%u/re/ExecMem/cbAllocated", pVCpu->idCpu);
2175	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbFree, STAMTYPE_U64, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
2176	"Number of bytes current free", "/IEM/CPU%u/re/ExecMem/cbFree", pVCpu->idCpu);
2177	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbTotal, STAMTYPE_U64, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
2178	"Total number of byte", "/IEM/CPU%u/re/ExecMem/cbTotal", pVCpu->idCpu);
2179	#ifdef VBOX_WITH_STATISTICS
2180	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbUnusable, STAMTYPE_U64, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
2181	"Total number of bytes being unusable", "/IEM/CPU%u/re/ExecMem/cbUnusable", pVCpu->idCpu);
2182	STAMR3RegisterFU(pUVM, &pExecMemAllocator->StatAlloc, STAMTYPE_PROFILE, STAMVISIBILITY_ALWAYS, STAMUNIT_TICKS_PER_CALL,
2183	"Profiling the allocator", "/IEM/CPU%u/re/ExecMem/ProfAlloc", pVCpu->idCpu);
2184	for (unsigned i = 1; i < RT_ELEMENTS(pExecMemAllocator->aStatSizes); i++)
2185	STAMR3RegisterFU(pUVM, &pExecMemAllocator->aStatSizes[i], STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,
2186	"Number of allocations of this number of allocation units",
2187	"/IEM/CPU%u/re/ExecMem/aSize%02u", pVCpu->idCpu, i);
2188	STAMR3RegisterFU(pUVM, &pExecMemAllocator->aStatSizes[0], STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,
2189	"Number of allocations 16 units or larger", "/IEM/CPU%u/re/ExecMem/aSize16OrLarger", pVCpu->idCpu);
2190	#endif
2191	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING
2192	STAMR3RegisterFU(pUVM, &pExecMemAllocator->StatPruneProf, STAMTYPE_PROFILE, STAMVISIBILITY_ALWAYS, STAMUNIT_TICKS_PER_CALL,
2193	"Pruning executable memory (alt)", "/IEM/CPU%u/re/ExecMem/Pruning", pVCpu->idCpu);
2194	STAMR3RegisterFU(pUVM, &pExecMemAllocator->StatPruneRecovered, STAMTYPE_PROFILE, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES_PER_CALL,
2195	"Bytes recovered while pruning", "/IEM/CPU%u/re/ExecMem/PruningRecovered", pVCpu->idCpu);
2196	#endif
2197	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cFruitlessChunkScans, STAMTYPE_U64_RESET, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,
2198	"Chunks fruitlessly scanned for free space", "/IEM/CPU%u/re/ExecMem/FruitlessChunkScans", pVCpu->idCpu);
2199
2200	return VINF_SUCCESS;
2201	}
2202

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

source: vbox/trunk/src/VBox/VMM/VMMAll/IEMAllN8veExecMem.cpp@ 106723

Download in other formats: