IEMAllN8veExecMem.cpp@ 105284

Last change on this file since 105284 was 105282, checked in by vboxsync, 7 months ago
VMM/IEM: Build fix for the recompiler on linux.am64, bugref:10391
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 66.9 KB

Line
1	/* $Id: IEMAllN8veExecMem.cpp 105282 2024-07-11 20:25:22Z vboxsync $ */
2	/** @file
3	* IEM - Native Recompiler
4	*
5	* Logging group IEM_RE_NATIVE assignments:
6	* - Level 1 (Log) : ...
7	* - Flow (LogFlow) : ...
8	* - Level 2 (Log2) : Details calls as they're recompiled.
9	* - Level 3 (Log3) : Disassemble native code after recompiling.
10	* - Level 4 (Log4) : ...
11	* - Level 5 (Log5) : ...
12	* - Level 6 (Log6) : ...
13	* - Level 7 (Log7) : ...
14	* - Level 8 (Log8) : ...
15	* - Level 9 (Log9) : ...
16	* - Level 10 (Log10): ...
17	* - Level 11 (Log11): Variable allocator.
18	* - Level 12 (Log12): Register allocator.
19	*/
20
21	/*
22	* Copyright (C) 2023 Oracle and/or its affiliates.
23	*
24	* This file is part of VirtualBox base platform packages, as
25	* available from https://www.virtualbox.org.
26	*
27	* This program is free software; you can redistribute it and/or
28	* modify it under the terms of the GNU General Public License
29	* as published by the Free Software Foundation, in version 3 of the
30	* License.
31	*
32	* This program is distributed in the hope that it will be useful, but
33	* WITHOUT ANY WARRANTY; without even the implied warranty of
34	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
35	* General Public License for more details.
36	*
37	* You should have received a copy of the GNU General Public License
38	* along with this program; if not, see <https://www.gnu.org/licenses>.
39	*
40	* SPDX-License-Identifier: GPL-3.0-only
41	*/
42
43
44	/*********************************************************************************************************************************
45	* Header Files *
46	*********************************************************************************************************************************/
47	#define LOG_GROUP LOG_GROUP_IEM_RE_NATIVE
48	#define IEM_WITH_OPAQUE_DECODER_STATE
49	#define VMM_INCLUDED_SRC_include_IEMMc_h /* block IEMMc.h inclusion. */
50	#include <VBox/vmm/iem.h>
51	#include <VBox/vmm/cpum.h>
52	#include "IEMInternal.h"
53	#include <VBox/vmm/vmcc.h>
54	#include <VBox/log.h>
55	#include <VBox/err.h>
56	#include <VBox/param.h>
57	#include <iprt/assert.h>
58	#include <iprt/mem.h>
59	#include <iprt/string.h>
60	#if defined(RT_ARCH_AMD64)
61	# include <iprt/x86.h>
62	#elif defined(RT_ARCH_ARM64)
63	# include <iprt/armv8.h>
64	#endif
65
66	#ifdef RT_OS_WINDOWS
67	# include <iprt/formats/pecoff.h> /* this is incomaptible with windows.h, thus: */
68	extern "C" DECLIMPORT(uint8_t) __cdecl RtlAddFunctionTable(void *pvFunctionTable, uint32_t cEntries, uintptr_t uBaseAddress);
69	extern "C" DECLIMPORT(uint8_t) __cdecl RtlDelFunctionTable(void *pvFunctionTable);
70	#else
71	# include <iprt/formats/dwarf.h>
72	# if defined(RT_OS_DARWIN)
73	# include <libkern/OSCacheControl.h>
74	# include <mach/mach.h>
75	# include <mach/mach_vm.h>
76	# define IEMNATIVE_USE_LIBUNWIND
77	extern "C" void __register_frame(const void *pvFde);
78	extern "C" void __deregister_frame(const void *pvFde);
79	# else
80	# ifdef DEBUG_bird /** @todo not thread safe yet */
81	# define IEMNATIVE_USE_GDB_JIT
82	# endif
83	# ifdef IEMNATIVE_USE_GDB_JIT
84	# include <iprt/critsect.h>
85	# include <iprt/once.h>
86	# include <iprt/formats/elf64.h>
87	# endif
88	extern "C" void __register_frame_info(void pvBegin, void pvObj); /* found no header for these two */
89	extern "C" void __deregister_frame_info(void pvBegin); /* (returns pvObj from __register_frame_info call) */
90	# endif
91	#endif
92
93	#include "IEMN8veRecompiler.h"
94
95
96	/*********************************************************************************************************************************
97	* Executable Memory Allocator *
98	*********************************************************************************************************************************/
99	/** The chunk sub-allocation unit size in bytes. */
100	#define IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE 256
101	/** The chunk sub-allocation unit size as a shift factor. */
102	#define IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT 8
103	/** Enables adding a header to the sub-allocator allocations.
104	* This is useful for freeing up executable memory among other things. */
105	#define IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
106	/** Use alternative pruning. */
107	#define IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING
108
109
110	#if defined(IN_RING3) && !defined(RT_OS_WINDOWS)
111	# ifdef IEMNATIVE_USE_GDB_JIT
112	# define IEMNATIVE_USE_GDB_JIT_ET_DYN
113
114	/** GDB JIT: Code entry. */
115	typedef struct GDBJITCODEENTRY
116	{
117	struct GDBJITCODEENTRY *pNext;
118	struct GDBJITCODEENTRY *pPrev;
119	uint8_t *pbSymFile;
120	uint64_t cbSymFile;
121	} GDBJITCODEENTRY;
122
123	/** GDB JIT: Actions. */
124	typedef enum GDBJITACTIONS : uint32_t
125	{
126	kGdbJitaction_NoAction = 0, kGdbJitaction_Register, kGdbJitaction_Unregister
127	} GDBJITACTIONS;
128
129	/** GDB JIT: Descriptor. */
130	typedef struct GDBJITDESCRIPTOR
131	{
132	uint32_t uVersion;
133	GDBJITACTIONS enmAction;
134	GDBJITCODEENTRY *pRelevant;
135	GDBJITCODEENTRY *pHead;
136	/** Our addition: */
137	GDBJITCODEENTRY *pTail;
138	} GDBJITDESCRIPTOR;
139
140	/** GDB JIT: Our simple symbol file data. */
141	typedef struct GDBJITSYMFILE
142	{
143	Elf64_Ehdr EHdr;
144	# ifndef IEMNATIVE_USE_GDB_JIT_ET_DYN
145	Elf64_Shdr aShdrs[5];
146	# else
147	Elf64_Shdr aShdrs[7];
148	Elf64_Phdr aPhdrs[2];
149	# endif
150	/** The dwarf ehframe data for the chunk. */
151	uint8_t abEhFrame[512];
152	char szzStrTab[128];
153	Elf64_Sym aSymbols[3];
154	# ifdef IEMNATIVE_USE_GDB_JIT_ET_DYN
155	Elf64_Sym aDynSyms[2];
156	Elf64_Dyn aDyn[6];
157	# endif
158	} GDBJITSYMFILE;
159
160	extern "C" GDBJITDESCRIPTOR __jit_debug_descriptor;
161	extern "C" DECLEXPORT(void) __jit_debug_register_code(void);
162
163	/** Init once for g_IemNativeGdbJitLock. */
164	static RTONCE g_IemNativeGdbJitOnce = RTONCE_INITIALIZER;
165	/** Init once for the critical section. */
166	static RTCRITSECT g_IemNativeGdbJitLock;
167
168	/** GDB reads the info here. */
169	GDBJITDESCRIPTOR __jit_debug_descriptor = { 1, kGdbJitaction_NoAction, NULL, NULL };
170
171	/** GDB sets a breakpoint on this and checks __jit_debug_descriptor when hit. */
172	DECL_NO_INLINE(RT_NOTHING, DECLEXPORT(void)) __jit_debug_register_code(void)
173	{
174	ASMNopPause();
175	}
176
177	/** @callback_method_impl{FNRTONCE} */
178	static DECLCALLBACK(int32_t) iemNativeGdbJitInitOnce(void *pvUser)
179	{
180	RT_NOREF(pvUser);
181	return RTCritSectInit(&g_IemNativeGdbJitLock);
182	}
183
184
185	# endif /* IEMNATIVE_USE_GDB_JIT */
186
187	/**
188	* Per-chunk unwind info for non-windows hosts.
189	*/
190	typedef struct IEMEXECMEMCHUNKEHFRAME
191	{
192	# ifdef IEMNATIVE_USE_LIBUNWIND
193	/** The offset of the FDA into abEhFrame. */
194	uintptr_t offFda;
195	# else
196	/** 'struct object' storage area. */
197	uint8_t abObject[1024];
198	# endif
199	# ifdef IEMNATIVE_USE_GDB_JIT
200	# if 0
201	/** The GDB JIT 'symbol file' data. */
202	GDBJITSYMFILE GdbJitSymFile;
203	# endif
204	/** The GDB JIT list entry. */
205	GDBJITCODEENTRY GdbJitEntry;
206	# endif
207	/** The dwarf ehframe data for the chunk. */
208	uint8_t abEhFrame[512];
209	} IEMEXECMEMCHUNKEHFRAME;
210	/** Pointer to per-chunk info info for non-windows hosts. */
211	typedef IEMEXECMEMCHUNKEHFRAME *PIEMEXECMEMCHUNKEHFRAME;
212	#endif
213
214
215	/**
216	* An chunk of executable memory.
217	*/
218	typedef struct IEMEXECMEMCHUNK
219	{
220	/** Number of free items in this chunk. */
221	uint32_t cFreeUnits;
222	/** Hint were to start searching for free space in the allocation bitmap. */
223	uint32_t idxFreeHint;
224	/** Pointer to the readable/writeable view of the memory chunk. */
225	void *pvChunkRw;
226	/** Pointer to the readable/executable view of the memory chunk. */
227	void *pvChunkRx;
228	#ifdef IEMNATIVE_WITH_RECOMPILER_PER_CHUNK_TAIL_CODE
229	/** Pointer to the context structure detailing the per chunk common code. */
230	PCIEMNATIVEPERCHUNKCTX pCtx;
231	#endif
232	#ifdef IN_RING3
233	/**
234	* Pointer to the unwind information.
235	*
236	* This is used during C++ throw and longjmp (windows and probably most other
237	* platforms). Some debuggers (windbg) makes use of it as well.
238	*
239	* Windows: This is allocated from hHeap on windows because (at least for
240	* AMD64) the UNWIND_INFO structure address in the
241	* RUNTIME_FUNCTION entry is an RVA and the chunk is the "image".
242	*
243	* Others: Allocated from the regular heap to avoid unnecessary executable data
244	* structures. This points to an IEMEXECMEMCHUNKEHFRAME structure. */
245	void *pvUnwindInfo;
246	#elif defined(IN_RING0)
247	/** Allocation handle. */
248	RTR0MEMOBJ hMemObj;
249	#endif
250	} IEMEXECMEMCHUNK;
251	/** Pointer to a memory chunk. */
252	typedef IEMEXECMEMCHUNK *PIEMEXECMEMCHUNK;
253
254
255	/**
256	* Executable memory allocator for the native recompiler.
257	*/
258	typedef struct IEMEXECMEMALLOCATOR
259	{
260	/** Magic value (IEMEXECMEMALLOCATOR_MAGIC). */
261	uint32_t uMagic;
262
263	/** The chunk size. */
264	uint32_t cbChunk;
265	/** The maximum number of chunks. */
266	uint32_t cMaxChunks;
267	/** The current number of chunks. */
268	uint32_t cChunks;
269	/** Hint where to start looking for available memory. */
270	uint32_t idxChunkHint;
271	/** Statistics: Current number of allocations. */
272	uint32_t cAllocations;
273
274	/** The total amount of memory available. */
275	uint64_t cbTotal;
276	/** Total amount of free memory. */
277	uint64_t cbFree;
278	/** Total amount of memory allocated. */
279	uint64_t cbAllocated;
280
281	/** Pointer to the allocation bitmaps for all the chunks (follows aChunks).
282	*
283	* Since the chunk size is a power of two and the minimum chunk size is a lot
284	* higher than the IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE, each chunk will always
285	* require a whole number of uint64_t elements in the allocation bitmap. So,
286	* for sake of simplicity, they are allocated as one continous chunk for
287	* simplicity/laziness. */
288	uint64_t *pbmAlloc;
289	/** Number of units (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE) per chunk. */
290	uint32_t cUnitsPerChunk;
291	/** Number of bitmap elements per chunk (for quickly locating the bitmap
292	* portion corresponding to an chunk). */
293	uint32_t cBitmapElementsPerChunk;
294
295	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING
296	/** The next chunk to prune in. */
297	uint32_t idxChunkPrune;
298	/** Where in chunk offset to start pruning at. */
299	uint32_t offChunkPrune;
300	/** Profiling the pruning code. */
301	STAMPROFILE StatPruneProf;
302	/** Number of bytes recovered by the pruning. */
303	STAMPROFILE StatPruneRecovered;
304	#endif
305
306	#ifdef VBOX_WITH_STATISTICS
307	STAMPROFILE StatAlloc;
308	#endif
309
310
311	#if defined(IN_RING3) && !defined(RT_OS_WINDOWS)
312	/** Pointer to the array of unwind info running parallel to aChunks (same
313	* allocation as this structure, located after the bitmaps).
314	* (For Windows, the structures must reside in 32-bit RVA distance to the
315	* actual chunk, so they are allocated off the chunk.) */
316	PIEMEXECMEMCHUNKEHFRAME paEhFrames;
317	#endif
318
319	/** The allocation chunks. */
320	RT_FLEXIBLE_ARRAY_EXTENSION
321	IEMEXECMEMCHUNK aChunks[RT_FLEXIBLE_ARRAY];
322	} IEMEXECMEMALLOCATOR;
323	/** Pointer to an executable memory allocator. */
324	typedef IEMEXECMEMALLOCATOR *PIEMEXECMEMALLOCATOR;
325
326	/** Magic value for IEMEXECMEMALLOCATOR::uMagic (Scott Frederick Turow). */
327	#define IEMEXECMEMALLOCATOR_MAGIC UINT32_C(0x19490412)
328
329
330	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
331	/**
332	* Allocation header.
333	*/
334	typedef struct IEMEXECMEMALLOCHDR
335	{
336	/** Magic value / eyecatcher (IEMEXECMEMALLOCHDR_MAGIC). */
337	uint32_t uMagic;
338	/** The allocation chunk (for speeding up freeing). */
339	uint32_t idxChunk;
340	/** Pointer to the translation block the allocation belongs to.
341	* This is the whole point of the header. */
342	PIEMTB pTb;
343	} IEMEXECMEMALLOCHDR;
344	/** Pointer to an allocation header. */
345	typedef IEMEXECMEMALLOCHDR *PIEMEXECMEMALLOCHDR;
346	/** Magic value for IEMEXECMEMALLOCHDR ('ExeM'). */
347	# define IEMEXECMEMALLOCHDR_MAGIC UINT32_C(0x4d657845)
348	#endif
349
350
351	static int iemExecMemAllocatorGrow(PVMCPUCC pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator);
352
353
354	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING
355	/**
356	* Frees up executable memory when we're out space.
357	*
358	* This is an alternative to iemTbAllocatorFreeupNativeSpace() that frees up
359	* space in a more linear fashion from the allocator's point of view. It may
360	* also defragment if implemented & enabled
361	*/
362	static void iemExecMemAllocatorPrune(PVMCPU pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator)
363	{
364	# ifndef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
365	# error "IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING requires IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER"
366	# endif
367	STAM_REL_PROFILE_START(&pExecMemAllocator->StatPruneProf, a);
368
369	/*
370	* Before we can start, we must process delayed frees.
371	*/
372	iemTbAllocatorProcessDelayedFrees(pVCpu, pVCpu->iem.s.pTbAllocatorR3);
373
374	AssertCompile(RT_IS_POWER_OF_TWO(IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE));
375
376	uint32_t const cbChunk = pExecMemAllocator->cbChunk;
377	AssertReturnVoid(RT_IS_POWER_OF_TWO(cbChunk));
378	AssertReturnVoid(cbChunk >= _1M && cbChunk <= _256M); /* see iemExecMemAllocatorInit */
379
380	uint32_t const cChunks = pExecMemAllocator->cChunks;
381	AssertReturnVoid(cChunks == pExecMemAllocator->cMaxChunks);
382	AssertReturnVoid(cChunks >= 1);
383
384	Assert(!pVCpu->iem.s.pCurTbR3);
385
386	/*
387	* Decide how much to prune. The chunk is is a multiple of two, so we'll be
388	* scanning a multiple of two here as well.
389	*/
390	uint32_t cbToPrune = cbChunk;
391
392	/* Never more than 25%. */
393	if (cChunks < 4)
394	cbToPrune /= cChunks == 1 ? 4 : 2;
395
396	/* Upper limit. In a debug build a 4MB limit averages out at ~0.6ms per call. */
397	if (cbToPrune > _4M)
398	cbToPrune = _4M;
399
400	/*
401	* Adjust the pruning chunk and offset accordingly.
402	*/
403	uint32_t idxChunk = pExecMemAllocator->idxChunkPrune;
404	uint32_t offChunk = pExecMemAllocator->offChunkPrune;
405	offChunk &= ~(uint32_t)(IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1U);
406	if (offChunk >= cbChunk)
407	{
408	offChunk = 0;
409	idxChunk += 1;
410	}
411	if (idxChunk >= cChunks)
412	{
413	offChunk = 0;
414	idxChunk = 0;
415	}
416
417	uint32_t const offPruneEnd = RT_MIN(offChunk + cbToPrune, cbChunk);
418
419	/*
420	* Do the pruning. The current approach is the sever kind.
421	*/
422	uint64_t cbPruned = 0;
423	uint8_t * const pbChunk = (uint8_t *)pExecMemAllocator->aChunks[idxChunk].pvChunkRx;
424	while (offChunk < offPruneEnd)
425	{
426	PIEMEXECMEMALLOCHDR pHdr = (PIEMEXECMEMALLOCHDR)&pbChunk[offChunk];
427
428	/* Is this the start of an allocation block for TB? (We typically have
429	one allocation at the start of each chunk for the unwind info where
430	pTb is NULL.) */
431	if ( pHdr->uMagic == IEMEXECMEMALLOCHDR_MAGIC
432	&& pHdr->pTb != NULL
433	&& pHdr->idxChunk == idxChunk)
434	{
435	PIEMTB const pTb = pHdr->pTb;
436	AssertPtr(pTb);
437
438	uint32_t const cbBlock = RT_ALIGN_32(pTb->Native.cInstructions * sizeof(IEMNATIVEINSTR) + sizeof(*pHdr),
439	IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE);
440	AssertBreakStmt(offChunk + cbBlock <= cbChunk, offChunk += IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE); /* paranoia */
441
442	iemTbAllocatorFree(pVCpu, pTb);
443
444	cbPruned += cbBlock;
445	offChunk += cbBlock;
446	}
447	else
448	offChunk += IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE;
449	}
450	STAM_REL_PROFILE_ADD_PERIOD(&pExecMemAllocator->StatPruneRecovered, cbPruned);
451
452	/*
453	* Save the current pruning point.
454	*/
455	pExecMemAllocator->offChunkPrune = offChunk;
456	pExecMemAllocator->idxChunkPrune = idxChunk;
457
458	STAM_REL_PROFILE_STOP(&pExecMemAllocator->StatPruneProf, a);
459	}
460	#endif /* IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING */
461
462
463	/**
464	* Try allocate a block of @a cReqUnits in the chunk @a idxChunk.
465	*/
466	static void *
467	iemExecMemAllocatorAllocInChunkInt(PIEMEXECMEMALLOCATOR pExecMemAllocator, uint64_t *pbmAlloc, uint32_t idxFirst,
468	uint32_t cToScan, uint32_t cReqUnits, uint32_t idxChunk, PIEMTB pTb,
469	void *ppvExec, PCIEMNATIVEPERCHUNKCTX ppChunkCtx)
470	{
471	/*
472	* Shift the bitmap to the idxFirst bit so we can use ASMBitFirstClear.
473	*/
474	Assert(!(cToScan & 63));
475	Assert(!(idxFirst & 63));
476	Assert(cToScan + idxFirst <= pExecMemAllocator->cUnitsPerChunk);
477	pbmAlloc += idxFirst / 64;
478
479	/*
480	* Scan the bitmap for cReqUnits of consequtive clear bits
481	*/
482	/** @todo This can probably be done more efficiently for non-x86 systems. */
483	int iBit = ASMBitFirstClear(pbmAlloc, cToScan);
484	while (iBit >= 0 && (uint32_t)iBit <= cToScan - cReqUnits)
485	{
486	uint32_t idxAddBit = 1;
487	while (idxAddBit < cReqUnits && !ASMBitTest(pbmAlloc, (uint32_t)iBit + idxAddBit))
488	idxAddBit++;
489	if (idxAddBit >= cReqUnits)
490	{
491	ASMBitSetRange(pbmAlloc, (uint32_t)iBit, (uint32_t)iBit + cReqUnits);
492
493	PIEMEXECMEMCHUNK const pChunk = &pExecMemAllocator->aChunks[idxChunk];
494	pChunk->cFreeUnits -= cReqUnits;
495	pChunk->idxFreeHint = (uint32_t)iBit + cReqUnits;
496
497	pExecMemAllocator->cAllocations += 1;
498	uint32_t const cbReq = cReqUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
499	pExecMemAllocator->cbAllocated += cbReq;
500	pExecMemAllocator->cbFree -= cbReq;
501	pExecMemAllocator->idxChunkHint = idxChunk;
502
503	void * const pvMemRw = (uint8_t *)pChunk->pvChunkRw
504	+ ((idxFirst + (uint32_t)iBit) << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT);
505
506	if (ppChunkCtx)
507	*ppChunkCtx = pChunk->pCtx;
508
509	/*
510	* Initialize the header and return.
511	*/
512	# ifdef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
513	PIEMEXECMEMALLOCHDR const pHdr = (PIEMEXECMEMALLOCHDR)pvMemRw;
514	pHdr->uMagic = IEMEXECMEMALLOCHDR_MAGIC;
515	pHdr->idxChunk = idxChunk;
516	pHdr->pTb = pTb;
517
518	if (ppvExec)
519	ppvExec = (uint8_t )pChunk->pvChunkRx
520	+ ((idxFirst + (uint32_t)iBit) << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT)
521	+ sizeof(*pHdr);
522
523	return pHdr + 1;
524	#else
525	if (ppvExec)
526	ppvExec = (uint8_t )pChunk->pvChunkRx
527	+ ((idxFirst + (uint32_t)iBit) << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT);
528
529	RT_NOREF(pTb);
530	return pvMem;
531	#endif
532	}
533
534	iBit = ASMBitNextClear(pbmAlloc, cToScan, iBit + idxAddBit - 1);
535	}
536	return NULL;
537	}
538
539
540	static PIEMNATIVEINSTR
541	iemExecMemAllocatorAllocInChunk(PIEMEXECMEMALLOCATOR pExecMemAllocator, uint32_t idxChunk, uint32_t cbReq, PIEMTB pTb,
542	PIEMNATIVEINSTR ppaExec, PCIEMNATIVEPERCHUNKCTX ppChunkCtx)
543	{
544	/*
545	* Figure out how much to allocate.
546	*/
547	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
548	uint32_t const cReqUnits = (cbReq + sizeof(IEMEXECMEMALLOCHDR) + IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1)
549	#else
550	uint32_t const cReqUnits = (cbReq + IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1)
551	#endif
552	>> IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
553	if (cReqUnits <= pExecMemAllocator->aChunks[idxChunk].cFreeUnits)
554	{
555	uint64_t * const pbmAlloc = &pExecMemAllocator->pbmAlloc[pExecMemAllocator->cBitmapElementsPerChunk * idxChunk];
556	uint32_t const idxHint = pExecMemAllocator->aChunks[idxChunk].idxFreeHint & ~(uint32_t)63;
557	if (idxHint + cReqUnits <= pExecMemAllocator->cUnitsPerChunk)
558	{
559	void *pvRet = iemExecMemAllocatorAllocInChunkInt(pExecMemAllocator, pbmAlloc, idxHint,
560	pExecMemAllocator->cUnitsPerChunk - idxHint,
561	cReqUnits, idxChunk, pTb, (void **)ppaExec, ppChunkCtx);
562	if (pvRet)
563	return (PIEMNATIVEINSTR)pvRet;
564	}
565	return (PIEMNATIVEINSTR)iemExecMemAllocatorAllocInChunkInt(pExecMemAllocator, pbmAlloc, 0,
566	RT_MIN(pExecMemAllocator->cUnitsPerChunk,
567	RT_ALIGN_32(idxHint + cReqUnits, 64)),
568	cReqUnits, idxChunk, pTb, (void **)ppaExec, ppChunkCtx);
569	}
570	return NULL;
571	}
572
573
574	/**
575	* Allocates @a cbReq bytes of executable memory.
576	*
577	* @returns Pointer to the readable/writeable memory, NULL if out of memory or other problem
578	* encountered.
579	* @param pVCpu The cross context virtual CPU structure of the
580	* calling thread.
581	* @param cbReq How many bytes are required.
582	* @param pTb The translation block that will be using the allocation.
583	* @param ppaExec Where to return the pointer to executable view of
584	* the allocated memory, optional.
585	* @param ppChunkCtx Where to return the per chunk attached context
586	* if available, optional.
587	*/
588	DECLHIDDEN(PIEMNATIVEINSTR) iemExecMemAllocatorAlloc(PVMCPU pVCpu, uint32_t cbReq, PIEMTB pTb,
589	PIEMNATIVEINSTR ppaExec, PCIEMNATIVEPERCHUNKCTX ppChunkCtx) RT_NOEXCEPT
590	{
591	PIEMEXECMEMALLOCATOR pExecMemAllocator = pVCpu->iem.s.pExecMemAllocatorR3;
592	AssertReturn(pExecMemAllocator && pExecMemAllocator->uMagic == IEMEXECMEMALLOCATOR_MAGIC, NULL);
593	AssertMsgReturn(cbReq > 32 && cbReq < _512K, ("%#x\n", cbReq), NULL);
594	STAM_PROFILE_START(&pExecMemAllocator->StatAlloc, a);
595
596	for (unsigned iIteration = 0;; iIteration++)
597	{
598	if (cbReq <= pExecMemAllocator->cbFree)
599	{
600	uint32_t const cChunks = pExecMemAllocator->cChunks;
601	uint32_t const idxChunkHint = pExecMemAllocator->idxChunkHint < cChunks ? pExecMemAllocator->idxChunkHint : 0;
602	for (uint32_t idxChunk = idxChunkHint; idxChunk < cChunks; idxChunk++)
603	{
604	PIEMNATIVEINSTR const pRet = iemExecMemAllocatorAllocInChunk(pExecMemAllocator, idxChunk, cbReq, pTb,
605	ppaExec, ppChunkCtx);
606	if (pRet)
607	{
608	STAM_PROFILE_STOP(&pExecMemAllocator->StatAlloc, a);
609	return pRet;
610	}
611	}
612	for (uint32_t idxChunk = 0; idxChunk < idxChunkHint; idxChunk++)
613	{
614	PIEMNATIVEINSTR const pRet = iemExecMemAllocatorAllocInChunk(pExecMemAllocator, idxChunk, cbReq, pTb,
615	ppaExec, ppChunkCtx);
616	if (pRet)
617	{
618	STAM_PROFILE_STOP(&pExecMemAllocator->StatAlloc, a);
619	return pRet;
620	}
621	}
622	}
623
624	/*
625	* Can we grow it with another chunk?
626	*/
627	if (pExecMemAllocator->cChunks < pExecMemAllocator->cMaxChunks)
628	{
629	int rc = iemExecMemAllocatorGrow(pVCpu, pExecMemAllocator);
630	AssertLogRelRCReturn(rc, NULL);
631
632	uint32_t const idxChunk = pExecMemAllocator->cChunks - 1;
633	PIEMNATIVEINSTR const pRet = iemExecMemAllocatorAllocInChunk(pExecMemAllocator, idxChunk, cbReq, pTb,
634	ppaExec, ppChunkCtx);
635	if (pRet)
636	{
637	STAM_PROFILE_STOP(&pExecMemAllocator->StatAlloc, a);
638	return pRet;
639	}
640	AssertFailed();
641	}
642
643	/*
644	* Try prune native TBs once.
645	*/
646	if (iIteration == 0)
647	{
648	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING
649	iemExecMemAllocatorPrune(pVCpu, pExecMemAllocator);
650	#else
651	/* No header included in the instruction count here. */
652	uint32_t const cNeededInstrs = RT_ALIGN_32(cbReq, IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE) / sizeof(IEMNATIVEINSTR);
653	iemTbAllocatorFreeupNativeSpace(pVCpu, cNeededInstrs);
654	#endif
655	}
656	else
657	{
658	STAM_REL_COUNTER_INC(&pVCpu->iem.s.StatNativeExecMemInstrBufAllocFailed);
659	STAM_PROFILE_STOP(&pExecMemAllocator->StatAlloc, a);
660	return NULL;
661	}
662	}
663	}
664
665
666	/** This is a hook to ensure the instruction cache is properly flushed before the code in the memory
667	* given by @a pv and @a cb is executed */
668	DECLHIDDEN(void) iemExecMemAllocatorReadyForUse(PVMCPUCC pVCpu, void *pv, size_t cb) RT_NOEXCEPT
669	{
670	#ifdef RT_OS_DARWIN
671	/*
672	* Flush the instruction cache:
673	* https://developer.apple.com/documentation/apple-silicon/porting-just-in-time-compilers-to-apple-silicon
674	*/
675	/* sys_dcache_flush(pv, cb); - not necessary */
676	sys_icache_invalidate(pv, cb);
677	RT_NOREF(pVCpu);
678
679	#elif defined(RT_OS_LINUX) && defined(RT_ARCH_ARM64)
680	RT_NOREF(pVCpu);
681
682	/* There is __builtin___clear_cache() but it flushes both the instruction and data cache, so do it manually. */
683	static uint32_t s_u32CtrEl0 = 0;
684	if (!s_u32CtrEl0)
685	asm volatile ("mrs %0, ctr_el0":"=r" (s_u32CtrEl0));
686	uintptr_t cbICacheLine = (uintptr_t)4 << (s_u32CtrEl0 & 0xf);
687
688	uintptr_t pb = (uintptr_t)pv & ~(cbICacheLine - 1);
689	for (; pb < (uintptr_t)pv + cb; pb += cbICacheLine)
690	asm volatile ("ic ivau, %0" : : "r" (pb) : "memory");
691
692	asm volatile ("dsb ish\n\t isb\n\t" : : : "memory");
693
694	#else
695	RT_NOREF(pVCpu, pv, cb);
696	#endif
697	}
698
699
700	/**
701	* Frees executable memory.
702	*/
703	DECLHIDDEN(void) iemExecMemAllocatorFree(PVMCPU pVCpu, void *pv, size_t cb) RT_NOEXCEPT
704	{
705	PIEMEXECMEMALLOCATOR pExecMemAllocator = pVCpu->iem.s.pExecMemAllocatorR3;
706	Assert(pExecMemAllocator && pExecMemAllocator->uMagic == IEMEXECMEMALLOCATOR_MAGIC);
707	AssertPtr(pv);
708	#ifndef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
709	Assert(!((uintptr_t)pv & (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1)));
710
711	/* Align the size as we did when allocating the block. */
712	cb = RT_ALIGN_Z(cb, IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE);
713
714	#else
715	PIEMEXECMEMALLOCHDR pHdr = (PIEMEXECMEMALLOCHDR)pv - 1;
716	Assert(!((uintptr_t)pHdr & (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE - 1)));
717	AssertReturnVoid(pHdr->uMagic == IEMEXECMEMALLOCHDR_MAGIC);
718	uint32_t const idxChunk = pHdr->idxChunk;
719	AssertReturnVoid(idxChunk < pExecMemAllocator->cChunks);
720	pv = pHdr;
721
722	/* Adjust and align the size to cover the whole allocation area. */
723	cb = RT_ALIGN_Z(cb + sizeof(*pHdr), IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SIZE);
724	#endif
725
726	/* Free it / assert sanity. */
727	bool fFound = false;
728	uint32_t const cbChunk = pExecMemAllocator->cbChunk;
729	#ifndef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
730	uint32_t const cChunks = pExecMemAllocator->cChunks;
731	for (uint32_t idxChunk = 0; idxChunk < cChunks; idxChunk++)
732	#endif
733	{
734	uintptr_t const offChunk = (uintptr_t)pv - (uintptr_t)pExecMemAllocator->aChunks[idxChunk].pvChunkRx;
735	fFound = offChunk < cbChunk;
736	if (fFound)
737	{
738	uint32_t const idxFirst = (uint32_t)offChunk >> IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
739	uint32_t const cReqUnits = (uint32_t)cb >> IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
740
741	/* Check that it's valid and free it. */
742	uint64_t * const pbmAlloc = &pExecMemAllocator->pbmAlloc[pExecMemAllocator->cBitmapElementsPerChunk * idxChunk];
743	AssertReturnVoid(ASMBitTest(pbmAlloc, idxFirst));
744	for (uint32_t i = 1; i < cReqUnits; i++)
745	AssertReturnVoid(ASMBitTest(pbmAlloc, idxFirst + i));
746	ASMBitClearRange(pbmAlloc, idxFirst, idxFirst + cReqUnits);
747
748	/* Invalidate the header using the writeable memory view. */
749	pHdr = (PIEMEXECMEMALLOCHDR)((uintptr_t)pExecMemAllocator->aChunks[idxChunk].pvChunkRw + offChunk);
750	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALLOC_HEADER
751	pHdr->uMagic = 0;
752	pHdr->idxChunk = 0;
753	pHdr->pTb = NULL;
754	#endif
755	pExecMemAllocator->aChunks[idxChunk].cFreeUnits += cReqUnits;
756	pExecMemAllocator->aChunks[idxChunk].idxFreeHint = idxFirst;
757
758	/* Update the stats. */
759	pExecMemAllocator->cbAllocated -= cb;
760	pExecMemAllocator->cbFree += cb;
761	pExecMemAllocator->cAllocations -= 1;
762	return;
763	}
764	}
765	AssertFailed();
766	}
767
768
769	/**
770	* Interface used by iemNativeRecompileAttachExecMemChunkCtx and unwind info
771	* generators.
772	*/
773	DECLHIDDEN(PIEMNATIVEINSTR)
774	iemExecMemAllocatorAllocFromChunk(PVMCPU pVCpu, uint32_t idxChunk, uint32_t cbReq, PIEMNATIVEINSTR *ppaExec)
775	{
776	PIEMEXECMEMALLOCATOR pExecMemAllocator = pVCpu->iem.s.pExecMemAllocatorR3;
777	AssertReturn(idxChunk < pExecMemAllocator->cChunks, NULL);
778	Assert(cbReq < _1M);
779	return iemExecMemAllocatorAllocInChunk(pExecMemAllocator, idxChunk, cbReq, NULL /pTb/, ppaExec, NULL /ppChunkCtx/);
780	}
781
782
783	#ifdef IEMNATIVE_WITH_RECOMPILER_PER_CHUNK_TAIL_CODE
784	/**
785	* For getting the per-chunk context detailing common code for a TB.
786	*
787	* This is for use by the disassembler.
788	*/
789	DECLHIDDEN(PCIEMNATIVEPERCHUNKCTX) iemExecMemGetTbChunkCtx(PVMCPU pVCpu, PCIEMTB pTb)
790	{
791	PIEMEXECMEMALLOCATOR pExecMemAllocator = pVCpu->iem.s.pExecMemAllocatorR3;
792	if ((pTb->fFlags & IEMTB_F_TYPE_MASK) == IEMTB_F_TYPE_NATIVE)
793	{
794	uintptr_t const uAddress = (uintptr_t)pTb->Native.paInstructions;
795	uint32_t const cbChunk = pExecMemAllocator->cbChunk;
796	uint32_t idxChunk = pExecMemAllocator->cChunks;
797	while (idxChunk-- > 0)
798	if (uAddress - (uintptr_t)pExecMemAllocator->aChunks[idxChunk].pvChunkRx < cbChunk)
799	return pExecMemAllocator->aChunks[idxChunk].pCtx;
800	}
801	return NULL;
802	}
803	#endif
804
805
806	#ifdef IN_RING3
807	# ifdef RT_OS_WINDOWS
808
809	/**
810	* Initializes the unwind info structures for windows hosts.
811	*/
812	static int
813	iemExecMemAllocatorInitAndRegisterUnwindInfoForChunk(PVMCPUCC pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator,
814	void *pvChunk, uint32_t idxChunk)
815	{
816	RT_NOREF(pVCpu);
817
818	/*
819	* The AMD64 unwind opcodes.
820	*
821	* This is a program that starts with RSP after a RET instruction that
822	* ends up in recompiled code, and the operations we describe here will
823	* restore all non-volatile registers and bring RSP back to where our
824	* RET address is. This means it's reverse order from what happens in
825	* the prologue.
826	*
827	* Note! Using a frame register approach here both because we have one
828	* and but mainly because the UWOP_ALLOC_LARGE argument values
829	* would be a pain to write initializers for. On the positive
830	* side, we're impervious to changes in the the stack variable
831	* area can can deal with dynamic stack allocations if necessary.
832	*/
833	static const IMAGE_UNWIND_CODE s_aOpcodes[] =
834	{
835	{ { 16, IMAGE_AMD64_UWOP_SET_FPREG, 0 } }, /* RSP = RBP - FrameOffset * 10 (0x60) */
836	{ { 16, IMAGE_AMD64_UWOP_ALLOC_SMALL, 0 } }, /* RSP += 8; */
837	{ { 14, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_x15 } }, /* R15 = [RSP]; RSP += 8; */
838	{ { 12, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_x14 } }, /* R14 = [RSP]; RSP += 8; */
839	{ { 10, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_x13 } }, /* R13 = [RSP]; RSP += 8; */
840	{ { 8, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_x12 } }, /* R12 = [RSP]; RSP += 8; */
841	{ { 7, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_xDI } }, /* RDI = [RSP]; RSP += 8; */
842	{ { 6, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_xSI } }, /* RSI = [RSP]; RSP += 8; */
843	{ { 5, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_xBX } }, /* RBX = [RSP]; RSP += 8; */
844	{ { 4, IMAGE_AMD64_UWOP_PUSH_NONVOL, X86_GREG_xBP } }, /* RBP = [RSP]; RSP += 8; */
845	};
846	union
847	{
848	IMAGE_UNWIND_INFO Info;
849	uint8_t abPadding[RT_UOFFSETOF(IMAGE_UNWIND_INFO, aOpcodes) + 16];
850	} s_UnwindInfo =
851	{
852	{
853	/* .Version = */ 1,
854	/* .Flags = */ 0,
855	/* .SizeOfProlog = / 16, / whatever */
856	/* .CountOfCodes = */ RT_ELEMENTS(s_aOpcodes),
857	/* .FrameRegister = */ X86_GREG_xBP,
858	/* .FrameOffset = / (-IEMNATIVE_FP_OFF_LAST_PUSH + 8) / 16 / we're off by one slot. sigh. */,
859	}
860	};
861	AssertCompile(-IEMNATIVE_FP_OFF_LAST_PUSH < 240 && -IEMNATIVE_FP_OFF_LAST_PUSH > 0);
862	AssertCompile((-IEMNATIVE_FP_OFF_LAST_PUSH & 0xf) == 8);
863
864	/*
865	* Calc how much space we need and allocate it off the exec heap.
866	*/
867	unsigned const cFunctionEntries = 1;
868	unsigned const cbUnwindInfo = sizeof(s_aOpcodes) + RT_UOFFSETOF(IMAGE_UNWIND_INFO, aOpcodes);
869	unsigned const cbNeeded = sizeof(IMAGE_RUNTIME_FUNCTION_ENTRY) * cFunctionEntries + cbUnwindInfo;
870	PIMAGE_RUNTIME_FUNCTION_ENTRY const paFunctions
871	= (PIMAGE_RUNTIME_FUNCTION_ENTRY)iemExecMemAllocatorAllocInChunk(pExecMemAllocator, idxChunk, cbNeeded, NULL, NULL, NULL);
872	AssertReturn(paFunctions, VERR_INTERNAL_ERROR_5);
873	pExecMemAllocator->aChunks[idxChunk].pvUnwindInfo = paFunctions;
874
875	/*
876	* Initialize the structures.
877	*/
878	PIMAGE_UNWIND_INFO const pInfo = (PIMAGE_UNWIND_INFO)&paFunctions[cFunctionEntries];
879
880	paFunctions[0].BeginAddress = 0;
881	paFunctions[0].EndAddress = pExecMemAllocator->cbChunk;
882	paFunctions[0].UnwindInfoAddress = (uint32_t)((uintptr_t)pInfo - (uintptr_t)pvChunk);
883
884	memcpy(pInfo, &s_UnwindInfo, RT_UOFFSETOF(IMAGE_UNWIND_INFO, aOpcodes));
885	memcpy(&pInfo->aOpcodes[0], s_aOpcodes, sizeof(s_aOpcodes));
886
887	/*
888	* Register it.
889	*/
890	uint8_t fRet = RtlAddFunctionTable(paFunctions, cFunctionEntries, (uintptr_t)pvChunk);
891	AssertReturn(fRet, VERR_INTERNAL_ERROR_3); /* Nothing to clean up on failure, since its within the chunk itself. */
892
893	return VINF_SUCCESS;
894	}
895
896
897	# else /* !RT_OS_WINDOWS */
898
899	/**
900	* Emits a LEB128 encoded value between -0x2000 and 0x2000 (both exclusive).
901	*/
902	DECLINLINE(RTPTRUNION) iemDwarfPutLeb128(RTPTRUNION Ptr, int32_t iValue)
903	{
904	if (iValue >= 64)
905	{
906	Assert(iValue < 0x2000);
907	*Ptr.pb++ = ((uint8_t)iValue & 0x7f) \| 0x80;
908	*Ptr.pb++ = (uint8_t)(iValue >> 7) & 0x3f;
909	}
910	else if (iValue >= 0)
911	*Ptr.pb++ = (uint8_t)iValue;
912	else if (iValue > -64)
913	*Ptr.pb++ = ((uint8_t)iValue & 0x3f) \| 0x40;
914	else
915	{
916	Assert(iValue > -0x2000);
917	*Ptr.pb++ = ((uint8_t)iValue & 0x7f) \| 0x80;
918	*Ptr.pb++ = ((uint8_t)(iValue >> 7) & 0x3f) \| 0x40;
919	}
920	return Ptr;
921	}
922
923
924	/**
925	* Emits an ULEB128 encoded value (up to 64-bit wide).
926	*/
927	DECLINLINE(RTPTRUNION) iemDwarfPutUleb128(RTPTRUNION Ptr, uint64_t uValue)
928	{
929	while (uValue >= 0x80)
930	{
931	*Ptr.pb++ = ((uint8_t)uValue & 0x7f) \| 0x80;
932	uValue >>= 7;
933	}
934	*Ptr.pb++ = (uint8_t)uValue;
935	return Ptr;
936	}
937
938
939	/**
940	* Emits a CFA rule as register @a uReg + offset @a off.
941	*/
942	DECLINLINE(RTPTRUNION) iemDwarfPutCfaDefCfa(RTPTRUNION Ptr, uint32_t uReg, uint32_t off)
943	{
944	*Ptr.pb++ = DW_CFA_def_cfa;
945	Ptr = iemDwarfPutUleb128(Ptr, uReg);
946	Ptr = iemDwarfPutUleb128(Ptr, off);
947	return Ptr;
948	}
949
950
951	/**
952	* Emits a register (@a uReg) save location:
953	* CFA + @a off * data_alignment_factor
954	*/
955	DECLINLINE(RTPTRUNION) iemDwarfPutCfaOffset(RTPTRUNION Ptr, uint32_t uReg, uint32_t off)
956	{
957	if (uReg < 0x40)
958	*Ptr.pb++ = DW_CFA_offset \| uReg;
959	else
960	{
961	*Ptr.pb++ = DW_CFA_offset_extended;
962	Ptr = iemDwarfPutUleb128(Ptr, uReg);
963	}
964	Ptr = iemDwarfPutUleb128(Ptr, off);
965	return Ptr;
966	}
967
968
969	# if 0 /* unused */
970	/**
971	* Emits a register (@a uReg) save location, using signed offset:
972	* CFA + @a offSigned * data_alignment_factor
973	*/
974	DECLINLINE(RTPTRUNION) iemDwarfPutCfaSignedOffset(RTPTRUNION Ptr, uint32_t uReg, int32_t offSigned)
975	{
976	*Ptr.pb++ = DW_CFA_offset_extended_sf;
977	Ptr = iemDwarfPutUleb128(Ptr, uReg);
978	Ptr = iemDwarfPutLeb128(Ptr, offSigned);
979	return Ptr;
980	}
981	# endif
982
983
984	/**
985	* Initializes the unwind info section for non-windows hosts.
986	*/
987	static int
988	iemExecMemAllocatorInitAndRegisterUnwindInfoForChunk(PVMCPUCC pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator,
989	void *pvChunk, uint32_t idxChunk)
990	{
991	PIEMEXECMEMCHUNKEHFRAME const pEhFrame = &pExecMemAllocator->paEhFrames[idxChunk];
992	pExecMemAllocator->aChunks[idxChunk].pvUnwindInfo = pEhFrame; /* not necessary, but whatever */
993
994	RTPTRUNION Ptr = { pEhFrame->abEhFrame };
995
996	/*
997	* Generate the CIE first.
998	*/
999	# ifdef IEMNATIVE_USE_LIBUNWIND /* libunwind (llvm, darwin) only supports v1 and v3. */
1000	uint8_t const iDwarfVer = 3;
1001	# else
1002	uint8_t const iDwarfVer = 4;
1003	# endif
1004	RTPTRUNION const PtrCie = Ptr;
1005	Ptr.pu32++ = 123; / The CIE length will be determined later. */
1006	Ptr.pu32++ = 0 /UINT32_MAX/; / I'm a CIE in .eh_frame speak. */
1007	Ptr.pb++ = iDwarfVer; / DwARF version */
1008	Ptr.pb++ = 0; / Augmentation. */
1009	if (iDwarfVer >= 4)
1010	{
1011	Ptr.pb++ = sizeof(uintptr_t); / Address size. */
1012	Ptr.pb++ = 0; / Segment selector size. */
1013	}
1014	# ifdef RT_ARCH_AMD64
1015	Ptr = iemDwarfPutLeb128(Ptr, 1); /* Code alignment factor (LEB128 = 1). */
1016	# else
1017	Ptr = iemDwarfPutLeb128(Ptr, 4); /* Code alignment factor (LEB128 = 4). */
1018	# endif
1019	Ptr = iemDwarfPutLeb128(Ptr, -8); /* Data alignment factor (LEB128 = -8). */
1020	# ifdef RT_ARCH_AMD64
1021	Ptr = iemDwarfPutUleb128(Ptr, DWREG_AMD64_RA); /* Return address column (ULEB128) */
1022	# elif defined(RT_ARCH_ARM64)
1023	Ptr = iemDwarfPutUleb128(Ptr, DWREG_ARM64_LR); /* Return address column (ULEB128) */
1024	# else
1025	# error "port me"
1026	# endif
1027	/* Initial instructions: */
1028	# ifdef RT_ARCH_AMD64
1029	Ptr = iemDwarfPutCfaDefCfa(Ptr, DWREG_AMD64_RBP, 16); /* CFA = RBP + 0x10 - first stack parameter */
1030	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_RA, 1); /* Ret RIP = [CFA + 1-8] /
1031	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_RBP, 2); /* RBP = [CFA + 2-8] /
1032	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_RBX, 3); /* RBX = [CFA + 3-8] /
1033	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_R12, 4); /* R12 = [CFA + 4-8] /
1034	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_R13, 5); /* R13 = [CFA + 5-8] /
1035	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_R14, 6); /* R14 = [CFA + 6-8] /
1036	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_AMD64_R15, 7); /* R15 = [CFA + 7-8] /
1037	# elif defined(RT_ARCH_ARM64)
1038	# if 1
1039	Ptr = iemDwarfPutCfaDefCfa(Ptr, DWREG_ARM64_BP, 16); /* CFA = BP + 0x10 - first stack parameter */
1040	# else
1041	Ptr = iemDwarfPutCfaDefCfa(Ptr, DWREG_ARM64_SP, IEMNATIVE_FRAME_VAR_SIZE + IEMNATIVE_FRAME_SAVE_REG_SIZE);
1042	# endif
1043	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_LR, 1); /* Ret PC = [CFA + 1-8] /
1044	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_BP, 2); /* Ret BP = [CFA + 2-8] /
1045	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X28, 3); /* X28 = [CFA + 3-8] /
1046	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X27, 4); /* X27 = [CFA + 4-8] /
1047	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X26, 5); /* X26 = [CFA + 5-8] /
1048	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X25, 6); /* X25 = [CFA + 6-8] /
1049	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X24, 7); /* X24 = [CFA + 7-8] /
1050	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X23, 8); /* X23 = [CFA + 8-8] /
1051	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X22, 9); /* X22 = [CFA + 9-8] /
1052	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X21, 10); /* X21 = [CFA +10-8] /
1053	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X20, 11); /* X20 = [CFA +11-8] /
1054	Ptr = iemDwarfPutCfaOffset(Ptr, DWREG_ARM64_X19, 12); /* X19 = [CFA +12-8] /
1055	AssertCompile(IEMNATIVE_FRAME_SAVE_REG_SIZE / 8 == 12);
1056	/** @todo we we need to do something about clearing DWREG_ARM64_RA_SIGN_STATE or something? */
1057	# else
1058	# error "port me"
1059	# endif
1060	while ((Ptr.u - PtrCie.u) & 3)
1061	*Ptr.pb++ = DW_CFA_nop;
1062	/* Finalize the CIE size. */
1063	*PtrCie.pu32 = Ptr.u - PtrCie.u - sizeof(uint32_t);
1064
1065	/*
1066	* Generate an FDE for the whole chunk area.
1067	*/
1068	# ifdef IEMNATIVE_USE_LIBUNWIND
1069	pEhFrame->offFda = Ptr.u - (uintptr_t)&pEhFrame->abEhFrame[0];
1070	# endif
1071	RTPTRUNION const PtrFde = Ptr;
1072	Ptr.pu32++ = 123; / The CIE length will be determined later. */
1073	Ptr.pu32 = Ptr.u - PtrCie.u; / Negated self relative CIE address. */
1074	Ptr.pu32++;
1075	Ptr.pu64++ = (uintptr_t)pvChunk; / Absolute start PC of this FDE. */
1076	Ptr.pu64++ = pExecMemAllocator->cbChunk; / PC range length for this PDE. */
1077	# if 0 /* not requried for recent libunwind.dylib nor recent libgcc/glib. */
1078	*Ptr.pb++ = DW_CFA_nop;
1079	# endif
1080	while ((Ptr.u - PtrFde.u) & 3)
1081	*Ptr.pb++ = DW_CFA_nop;
1082	/* Finalize the FDE size. */
1083	*PtrFde.pu32 = Ptr.u - PtrFde.u - sizeof(uint32_t);
1084
1085	/* Terminator entry. */
1086	*Ptr.pu32++ = 0;
1087	Ptr.pu32++ = 0; / just to be sure... */
1088	Assert(Ptr.u - (uintptr_t)&pEhFrame->abEhFrame[0] <= sizeof(pEhFrame->abEhFrame));
1089
1090	/*
1091	* Register it.
1092	*/
1093	# ifdef IEMNATIVE_USE_LIBUNWIND
1094	__register_frame(&pEhFrame->abEhFrame[pEhFrame->offFda]);
1095	# else
1096	memset(pEhFrame->abObject, 0xf6, sizeof(pEhFrame->abObject)); /* color the memory to better spot usage */
1097	__register_frame_info(pEhFrame->abEhFrame, pEhFrame->abObject);
1098	# endif
1099
1100	# ifdef IEMNATIVE_USE_GDB_JIT
1101	/*
1102	* Now for telling GDB about this (experimental).
1103	*
1104	* This seems to work best with ET_DYN.
1105	*/
1106	GDBJITSYMFILE * const pSymFile = (GDBJITSYMFILE *)iemExecMemAllocatorAllocInChunk(pExecMemAllocator, idxChunk,
1107	sizeof(GDBJITSYMFILE), NULL, NULL, NULL);
1108	AssertReturn(pSymFile, VERR_INTERNAL_ERROR_5);
1109	unsigned const offSymFileInChunk = (uintptr_t)pSymFile - (uintptr_t)pvChunk;
1110
1111	RT_ZERO(*pSymFile);
1112
1113	/*
1114	* The ELF header:
1115	*/
1116	pSymFile->EHdr.e_ident[0] = ELFMAG0;
1117	pSymFile->EHdr.e_ident[1] = ELFMAG1;
1118	pSymFile->EHdr.e_ident[2] = ELFMAG2;
1119	pSymFile->EHdr.e_ident[3] = ELFMAG3;
1120	pSymFile->EHdr.e_ident[EI_VERSION] = EV_CURRENT;
1121	pSymFile->EHdr.e_ident[EI_CLASS] = ELFCLASS64;
1122	pSymFile->EHdr.e_ident[EI_DATA] = ELFDATA2LSB;
1123	pSymFile->EHdr.e_ident[EI_OSABI] = ELFOSABI_NONE;
1124	# ifdef IEMNATIVE_USE_GDB_JIT_ET_DYN
1125	pSymFile->EHdr.e_type = ET_DYN;
1126	# else
1127	pSymFile->EHdr.e_type = ET_REL;
1128	# endif
1129	# ifdef RT_ARCH_AMD64
1130	pSymFile->EHdr.e_machine = EM_AMD64;
1131	# elif defined(RT_ARCH_ARM64)
1132	pSymFile->EHdr.e_machine = EM_AARCH64;
1133	# else
1134	# error "port me"
1135	# endif
1136	pSymFile->EHdr.e_version = 1; /?/
1137	pSymFile->EHdr.e_entry = 0;
1138	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1139	pSymFile->EHdr.e_phoff = RT_UOFFSETOF(GDBJITSYMFILE, aPhdrs);
1140	# else
1141	pSymFile->EHdr.e_phoff = 0;
1142	# endif
1143	pSymFile->EHdr.e_shoff = sizeof(pSymFile->EHdr);
1144	pSymFile->EHdr.e_flags = 0;
1145	pSymFile->EHdr.e_ehsize = sizeof(pSymFile->EHdr);
1146	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1147	pSymFile->EHdr.e_phentsize = sizeof(pSymFile->aPhdrs[0]);
1148	pSymFile->EHdr.e_phnum = RT_ELEMENTS(pSymFile->aPhdrs);
1149	# else
1150	pSymFile->EHdr.e_phentsize = 0;
1151	pSymFile->EHdr.e_phnum = 0;
1152	# endif
1153	pSymFile->EHdr.e_shentsize = sizeof(pSymFile->aShdrs[0]);
1154	pSymFile->EHdr.e_shnum = RT_ELEMENTS(pSymFile->aShdrs);
1155	pSymFile->EHdr.e_shstrndx = 0; /* set later */
1156
1157	uint32_t offStrTab = 0;
1158	#define APPEND_STR(a_szStr) do { \
1159	memcpy(&pSymFile->szzStrTab[offStrTab], a_szStr, sizeof(a_szStr)); \
1160	offStrTab += sizeof(a_szStr); \
1161	Assert(offStrTab < sizeof(pSymFile->szzStrTab)); \
1162	} while (0)
1163	#define APPEND_STR_FMT(a_szStr, ...) do { \
1164	offStrTab += RTStrPrintf(&pSymFile->szzStrTab[offStrTab], sizeof(pSymFile->szzStrTab) - offStrTab, a_szStr, __VA_ARGS__); \
1165	offStrTab++; \
1166	Assert(offStrTab < sizeof(pSymFile->szzStrTab)); \
1167	} while (0)
1168
1169	/*
1170	* Section headers.
1171	*/
1172	/* Section header #0: NULL */
1173	unsigned i = 0;
1174	APPEND_STR("");
1175	RT_ZERO(pSymFile->aShdrs[i]);
1176	i++;
1177
1178	/* Section header: .eh_frame */
1179	pSymFile->aShdrs[i].sh_name = offStrTab;
1180	APPEND_STR(".eh_frame");
1181	pSymFile->aShdrs[i].sh_type = SHT_PROGBITS;
1182	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC \| SHF_EXECINSTR;
1183	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN) \|\| defined(IEMNATIVE_USE_GDB_JIT_ELF_RVAS)
1184	pSymFile->aShdrs[i].sh_offset
1185	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, abEhFrame);
1186	# else
1187	pSymFile->aShdrs[i].sh_addr = (uintptr_t)&pSymFile->abEhFrame[0];
1188	pSymFile->aShdrs[i].sh_offset = 0;
1189	# endif
1190
1191	pSymFile->aShdrs[i].sh_size = sizeof(pEhFrame->abEhFrame);
1192	pSymFile->aShdrs[i].sh_link = 0;
1193	pSymFile->aShdrs[i].sh_info = 0;
1194	pSymFile->aShdrs[i].sh_addralign = 1;
1195	pSymFile->aShdrs[i].sh_entsize = 0;
1196	memcpy(pSymFile->abEhFrame, pEhFrame->abEhFrame, sizeof(pEhFrame->abEhFrame));
1197	i++;
1198
1199	/* Section header: .shstrtab */
1200	unsigned const iShStrTab = i;
1201	pSymFile->EHdr.e_shstrndx = iShStrTab;
1202	pSymFile->aShdrs[i].sh_name = offStrTab;
1203	APPEND_STR(".shstrtab");
1204	pSymFile->aShdrs[i].sh_type = SHT_STRTAB;
1205	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC;
1206	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN) \|\| defined(IEMNATIVE_USE_GDB_JIT_ELF_RVAS)
1207	pSymFile->aShdrs[i].sh_offset
1208	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, szzStrTab);
1209	# else
1210	pSymFile->aShdrs[i].sh_addr = (uintptr_t)&pSymFile->szzStrTab[0];
1211	pSymFile->aShdrs[i].sh_offset = 0;
1212	# endif
1213	pSymFile->aShdrs[i].sh_size = sizeof(pSymFile->szzStrTab);
1214	pSymFile->aShdrs[i].sh_link = 0;
1215	pSymFile->aShdrs[i].sh_info = 0;
1216	pSymFile->aShdrs[i].sh_addralign = 1;
1217	pSymFile->aShdrs[i].sh_entsize = 0;
1218	i++;
1219
1220	/* Section header: .symbols */
1221	pSymFile->aShdrs[i].sh_name = offStrTab;
1222	APPEND_STR(".symtab");
1223	pSymFile->aShdrs[i].sh_type = SHT_SYMTAB;
1224	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC;
1225	pSymFile->aShdrs[i].sh_offset
1226	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, aSymbols);
1227	pSymFile->aShdrs[i].sh_size = sizeof(pSymFile->aSymbols);
1228	pSymFile->aShdrs[i].sh_link = iShStrTab;
1229	pSymFile->aShdrs[i].sh_info = RT_ELEMENTS(pSymFile->aSymbols);
1230	pSymFile->aShdrs[i].sh_addralign = sizeof(pSymFile->aSymbols[0].st_value);
1231	pSymFile->aShdrs[i].sh_entsize = sizeof(pSymFile->aSymbols[0]);
1232	i++;
1233
1234	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1235	/* Section header: .symbols */
1236	pSymFile->aShdrs[i].sh_name = offStrTab;
1237	APPEND_STR(".dynsym");
1238	pSymFile->aShdrs[i].sh_type = SHT_DYNSYM;
1239	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC;
1240	pSymFile->aShdrs[i].sh_offset
1241	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, aDynSyms);
1242	pSymFile->aShdrs[i].sh_size = sizeof(pSymFile->aDynSyms);
1243	pSymFile->aShdrs[i].sh_link = iShStrTab;
1244	pSymFile->aShdrs[i].sh_info = RT_ELEMENTS(pSymFile->aDynSyms);
1245	pSymFile->aShdrs[i].sh_addralign = sizeof(pSymFile->aDynSyms[0].st_value);
1246	pSymFile->aShdrs[i].sh_entsize = sizeof(pSymFile->aDynSyms[0]);
1247	i++;
1248	# endif
1249
1250	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1251	/* Section header: .dynamic */
1252	pSymFile->aShdrs[i].sh_name = offStrTab;
1253	APPEND_STR(".dynamic");
1254	pSymFile->aShdrs[i].sh_type = SHT_DYNAMIC;
1255	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC;
1256	pSymFile->aShdrs[i].sh_offset
1257	= pSymFile->aShdrs[i].sh_addr = RT_UOFFSETOF(GDBJITSYMFILE, aDyn);
1258	pSymFile->aShdrs[i].sh_size = sizeof(pSymFile->aDyn);
1259	pSymFile->aShdrs[i].sh_link = iShStrTab;
1260	pSymFile->aShdrs[i].sh_info = 0;
1261	pSymFile->aShdrs[i].sh_addralign = 1;
1262	pSymFile->aShdrs[i].sh_entsize = sizeof(pSymFile->aDyn[0]);
1263	i++;
1264	# endif
1265
1266	/* Section header: .text */
1267	unsigned const iShText = i;
1268	pSymFile->aShdrs[i].sh_name = offStrTab;
1269	APPEND_STR(".text");
1270	pSymFile->aShdrs[i].sh_type = SHT_PROGBITS;
1271	pSymFile->aShdrs[i].sh_flags = SHF_ALLOC \| SHF_EXECINSTR;
1272	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN) \|\| defined(IEMNATIVE_USE_GDB_JIT_ELF_RVAS)
1273	pSymFile->aShdrs[i].sh_offset
1274	= pSymFile->aShdrs[i].sh_addr = sizeof(GDBJITSYMFILE);
1275	# else
1276	pSymFile->aShdrs[i].sh_addr = (uintptr_t)(pSymFile + 1);
1277	pSymFile->aShdrs[i].sh_offset = 0;
1278	# endif
1279	pSymFile->aShdrs[i].sh_size = pExecMemAllocator->cbChunk - offSymFileInChunk - sizeof(GDBJITSYMFILE);
1280	pSymFile->aShdrs[i].sh_link = 0;
1281	pSymFile->aShdrs[i].sh_info = 0;
1282	pSymFile->aShdrs[i].sh_addralign = 1;
1283	pSymFile->aShdrs[i].sh_entsize = 0;
1284	i++;
1285
1286	Assert(i == RT_ELEMENTS(pSymFile->aShdrs));
1287
1288	# if defined(IEMNATIVE_USE_GDB_JIT_ET_DYN)
1289	/*
1290	* The program headers:
1291	*/
1292	/* Everything in a single LOAD segment: */
1293	i = 0;
1294	pSymFile->aPhdrs[i].p_type = PT_LOAD;
1295	pSymFile->aPhdrs[i].p_flags = PF_X \| PF_R;
1296	pSymFile->aPhdrs[i].p_offset
1297	= pSymFile->aPhdrs[i].p_vaddr
1298	= pSymFile->aPhdrs[i].p_paddr = 0;
1299	pSymFile->aPhdrs[i].p_filesz /* Size of segment in file. */
1300	= pSymFile->aPhdrs[i].p_memsz = pExecMemAllocator->cbChunk - offSymFileInChunk;
1301	pSymFile->aPhdrs[i].p_align = HOST_PAGE_SIZE;
1302	i++;
1303	/* The .dynamic segment. */
1304	pSymFile->aPhdrs[i].p_type = PT_DYNAMIC;
1305	pSymFile->aPhdrs[i].p_flags = PF_R;
1306	pSymFile->aPhdrs[i].p_offset
1307	= pSymFile->aPhdrs[i].p_vaddr
1308	= pSymFile->aPhdrs[i].p_paddr = RT_UOFFSETOF(GDBJITSYMFILE, aDyn);
1309	pSymFile->aPhdrs[i].p_filesz /* Size of segment in file. */
1310	= pSymFile->aPhdrs[i].p_memsz = sizeof(pSymFile->aDyn);
1311	pSymFile->aPhdrs[i].p_align = sizeof(pSymFile->aDyn[0].d_tag);
1312	i++;
1313
1314	Assert(i == RT_ELEMENTS(pSymFile->aPhdrs));
1315
1316	/*
1317	* The dynamic section:
1318	*/
1319	i = 0;
1320	pSymFile->aDyn[i].d_tag = DT_SONAME;
1321	pSymFile->aDyn[i].d_un.d_val = offStrTab;
1322	APPEND_STR_FMT("iem-exec-chunk-%u-%u", pVCpu->idCpu, idxChunk);
1323	i++;
1324	pSymFile->aDyn[i].d_tag = DT_STRTAB;
1325	pSymFile->aDyn[i].d_un.d_ptr = RT_UOFFSETOF(GDBJITSYMFILE, szzStrTab);
1326	i++;
1327	pSymFile->aDyn[i].d_tag = DT_STRSZ;
1328	pSymFile->aDyn[i].d_un.d_val = sizeof(pSymFile->szzStrTab);
1329	i++;
1330	pSymFile->aDyn[i].d_tag = DT_SYMTAB;
1331	pSymFile->aDyn[i].d_un.d_ptr = RT_UOFFSETOF(GDBJITSYMFILE, aDynSyms);
1332	i++;
1333	pSymFile->aDyn[i].d_tag = DT_SYMENT;
1334	pSymFile->aDyn[i].d_un.d_val = sizeof(pSymFile->aDynSyms[0]);
1335	i++;
1336	pSymFile->aDyn[i].d_tag = DT_NULL;
1337	i++;
1338	Assert(i == RT_ELEMENTS(pSymFile->aDyn));
1339	# endif /* IEMNATIVE_USE_GDB_JIT_ET_DYN */
1340
1341	/*
1342	* Symbol tables:
1343	*/
1344	/** @todo gdb doesn't seem to really like this ... */
1345	i = 0;
1346	pSymFile->aSymbols[i].st_name = 0;
1347	pSymFile->aSymbols[i].st_shndx = SHN_UNDEF;
1348	pSymFile->aSymbols[i].st_value = 0;
1349	pSymFile->aSymbols[i].st_size = 0;
1350	pSymFile->aSymbols[i].st_info = ELF64_ST_INFO(STB_LOCAL, STT_NOTYPE);
1351	pSymFile->aSymbols[i].st_other = 0 /* STV_DEFAULT */;
1352	# ifdef IEMNATIVE_USE_GDB_JIT_ET_DYN
1353	pSymFile->aDynSyms[0] = pSymFile->aSymbols[i];
1354	# endif
1355	i++;
1356
1357	pSymFile->aSymbols[i].st_name = 0;
1358	pSymFile->aSymbols[i].st_shndx = SHN_ABS;
1359	pSymFile->aSymbols[i].st_value = 0;
1360	pSymFile->aSymbols[i].st_size = 0;
1361	pSymFile->aSymbols[i].st_info = ELF64_ST_INFO(STB_LOCAL, STT_FILE);
1362	pSymFile->aSymbols[i].st_other = 0 /* STV_DEFAULT */;
1363	i++;
1364
1365	pSymFile->aSymbols[i].st_name = offStrTab;
1366	APPEND_STR_FMT("iem_exec_chunk_%u_%u", pVCpu->idCpu, idxChunk);
1367	# if 0
1368	pSymFile->aSymbols[i].st_shndx = iShText;
1369	pSymFile->aSymbols[i].st_value = 0;
1370	# else
1371	pSymFile->aSymbols[i].st_shndx = SHN_ABS;
1372	pSymFile->aSymbols[i].st_value = (uintptr_t)(pSymFile + 1);
1373	# endif
1374	pSymFile->aSymbols[i].st_size = pSymFile->aShdrs[iShText].sh_size;
1375	pSymFile->aSymbols[i].st_info = ELF64_ST_INFO(STB_GLOBAL, STT_FUNC);
1376	pSymFile->aSymbols[i].st_other = 0 /* STV_DEFAULT */;
1377	# ifdef IEMNATIVE_USE_GDB_JIT_ET_DYN
1378	pSymFile->aDynSyms[1] = pSymFile->aSymbols[i];
1379	pSymFile->aDynSyms[1].st_value = (uintptr_t)(pSymFile + 1);
1380	# endif
1381	i++;
1382
1383	Assert(i == RT_ELEMENTS(pSymFile->aSymbols));
1384	Assert(offStrTab < sizeof(pSymFile->szzStrTab));
1385
1386	/*
1387	* The GDB JIT entry and informing GDB.
1388	*/
1389	pEhFrame->GdbJitEntry.pbSymFile = (uint8_t *)pSymFile;
1390	# if 1
1391	pEhFrame->GdbJitEntry.cbSymFile = pExecMemAllocator->cbChunk - ((uintptr_t)pSymFile - (uintptr_t)pvChunk);
1392	# else
1393	pEhFrame->GdbJitEntry.cbSymFile = sizeof(GDBJITSYMFILE);
1394	# endif
1395
1396	RTOnce(&g_IemNativeGdbJitOnce, iemNativeGdbJitInitOnce, NULL);
1397	RTCritSectEnter(&g_IemNativeGdbJitLock);
1398	pEhFrame->GdbJitEntry.pNext = NULL;
1399	pEhFrame->GdbJitEntry.pPrev = __jit_debug_descriptor.pTail;
1400	if (__jit_debug_descriptor.pTail)
1401	__jit_debug_descriptor.pTail->pNext = &pEhFrame->GdbJitEntry;
1402	else
1403	__jit_debug_descriptor.pHead = &pEhFrame->GdbJitEntry;
1404	__jit_debug_descriptor.pTail = &pEhFrame->GdbJitEntry;
1405	__jit_debug_descriptor.pRelevant = &pEhFrame->GdbJitEntry;
1406
1407	/* Notify GDB: */
1408	__jit_debug_descriptor.enmAction = kGdbJitaction_Register;
1409	__jit_debug_register_code();
1410	__jit_debug_descriptor.enmAction = kGdbJitaction_NoAction;
1411	RTCritSectLeave(&g_IemNativeGdbJitLock);
1412
1413	# else /* !IEMNATIVE_USE_GDB_JIT */
1414	RT_NOREF(pVCpu);
1415	# endif /* !IEMNATIVE_USE_GDB_JIT */
1416
1417	return VINF_SUCCESS;
1418	}
1419
1420	# endif /* !RT_OS_WINDOWS */
1421	#endif /* IN_RING3 */
1422
1423
1424	/**
1425	* Adds another chunk to the executable memory allocator.
1426	*
1427	* This is used by the init code for the initial allocation and later by the
1428	* regular allocator function when it's out of memory.
1429	*/
1430	static int iemExecMemAllocatorGrow(PVMCPUCC pVCpu, PIEMEXECMEMALLOCATOR pExecMemAllocator)
1431	{
1432	/* Check that we've room for growth. */
1433	uint32_t const idxChunk = pExecMemAllocator->cChunks;
1434	AssertLogRelReturn(idxChunk < pExecMemAllocator->cMaxChunks, VERR_OUT_OF_RESOURCES);
1435
1436	/* Allocate a chunk. */
1437	#ifdef RT_OS_DARWIN
1438	void *pvChunk = RTMemPageAllocEx(pExecMemAllocator->cbChunk, 0);
1439	#else
1440	void *pvChunk = RTMemPageAllocEx(pExecMemAllocator->cbChunk, RTMEMPAGEALLOC_F_EXECUTABLE);
1441	#endif
1442	AssertLogRelReturn(pvChunk, VERR_NO_EXEC_MEMORY);
1443
1444	#ifdef RT_OS_DARWIN
1445	/*
1446	* Because it is impossible to have a RWX memory allocation on macOS try to remap the memory
1447	* chunk readable/executable somewhere else so we can save us the hassle of switching between
1448	* protections when exeuctable memory is allocated.
1449	*/
1450	int rc = VERR_NO_EXEC_MEMORY;
1451	mach_port_t hPortTask = mach_task_self();
1452	mach_vm_address_t AddrChunk = (mach_vm_address_t)pvChunk;
1453	mach_vm_address_t AddrRemapped = 0;
1454	vm_prot_t ProtCur = 0;
1455	vm_prot_t ProtMax = 0;
1456	kern_return_t krc = mach_vm_remap(hPortTask, &AddrRemapped, pExecMemAllocator->cbChunk, 0,
1457	VM_FLAGS_ANYWHERE \| VM_FLAGS_RETURN_DATA_ADDR,
1458	hPortTask, AddrChunk, FALSE, &ProtCur, &ProtMax,
1459	VM_INHERIT_NONE);
1460	if (krc == KERN_SUCCESS)
1461	{
1462	krc = mach_vm_protect(mach_task_self(), AddrRemapped, pExecMemAllocator->cbChunk, FALSE, VM_PROT_READ \| VM_PROT_EXECUTE);
1463	if (krc == KERN_SUCCESS)
1464	rc = VINF_SUCCESS;
1465	else
1466	{
1467	AssertLogRelMsgFailed(("mach_vm_protect -> %d (%#x)\n", krc, krc));
1468	krc = mach_vm_deallocate(hPortTask, AddrRemapped, pExecMemAllocator->cbChunk);
1469	Assert(krc == KERN_SUCCESS);
1470	}
1471	}
1472	else
1473	AssertLogRelMsgFailed(("mach_vm_remap -> %d (%#x)\n", krc, krc));
1474	if (RT_FAILURE(rc))
1475	{
1476	RTMemPageFree(pvChunk, pExecMemAllocator->cbChunk);
1477	return rc;
1478	}
1479
1480	void pvChunkRx = (void )AddrRemapped;
1481	#else
1482	# if defined(IN_RING3) \|\| defined(IEMNATIVE_WITH_RECOMPILER_PER_CHUNK_TAIL_CODE)
1483	int rc = VINF_SUCCESS;
1484	# endif
1485	void *pvChunkRx = pvChunk;
1486	#endif
1487
1488	/*
1489	* Add the chunk.
1490	*
1491	* This must be done before the unwind init so windows can allocate
1492	* memory from the chunk when using the alternative sub-allocator.
1493	*/
1494	pExecMemAllocator->aChunks[idxChunk].pvChunkRw = pvChunk;
1495	pExecMemAllocator->aChunks[idxChunk].pvChunkRx = pvChunkRx;
1496	#ifdef IN_RING3
1497	pExecMemAllocator->aChunks[idxChunk].pvUnwindInfo = NULL;
1498	#endif
1499	pExecMemAllocator->aChunks[idxChunk].cFreeUnits = pExecMemAllocator->cUnitsPerChunk;
1500	pExecMemAllocator->aChunks[idxChunk].idxFreeHint = 0;
1501	memset(&pExecMemAllocator->pbmAlloc[pExecMemAllocator->cBitmapElementsPerChunk * idxChunk],
1502	0, sizeof(pExecMemAllocator->pbmAlloc[0]) * pExecMemAllocator->cBitmapElementsPerChunk);
1503
1504	pExecMemAllocator->cChunks = idxChunk + 1;
1505	pExecMemAllocator->idxChunkHint = idxChunk;
1506
1507	pExecMemAllocator->cbTotal += pExecMemAllocator->cbChunk;
1508	pExecMemAllocator->cbFree += pExecMemAllocator->cbChunk;
1509
1510	/* If there is a chunk context init callback call it. */
1511	#ifdef IEMNATIVE_WITH_RECOMPILER_PER_CHUNK_TAIL_CODE
1512	pExecMemAllocator->aChunks[idxChunk].pCtx = iemNativeRecompileAttachExecMemChunkCtx(pVCpu, idxChunk);
1513	if (pExecMemAllocator->aChunks[idxChunk].pCtx)
1514	#endif
1515	{
1516	#ifdef IN_RING3
1517	/*
1518	* Initialize the unwind information (this cannot really fail atm).
1519	* (This sets pvUnwindInfo.)
1520	*/
1521	rc = iemExecMemAllocatorInitAndRegisterUnwindInfoForChunk(pVCpu, pExecMemAllocator, pvChunkRx, idxChunk);
1522	#endif
1523	}
1524	#if defined(IN_RING3) \|\| defined(IEMNATIVE_WITH_RECOMPILER_PER_CHUNK_TAIL_CODE)
1525	if (RT_SUCCESS(rc))
1526	{ /* likely */ }
1527	else
1528	{
1529	/* Just in case the impossible happens, undo the above up: */
1530	pExecMemAllocator->cbTotal -= pExecMemAllocator->cbChunk;
1531	pExecMemAllocator->cbFree -= pExecMemAllocator->aChunks[idxChunk].cFreeUnits << IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
1532	pExecMemAllocator->cChunks = idxChunk;
1533	memset(&pExecMemAllocator->pbmAlloc[pExecMemAllocator->cBitmapElementsPerChunk * idxChunk],
1534	0xff, sizeof(pExecMemAllocator->pbmAlloc[0]) * pExecMemAllocator->cBitmapElementsPerChunk);
1535	pExecMemAllocator->aChunks[idxChunk].pvChunkRw = NULL;
1536	pExecMemAllocator->aChunks[idxChunk].cFreeUnits = 0;
1537
1538	# ifdef RT_OS_DARWIN
1539	krc = mach_vm_deallocate(mach_task_self(), (mach_vm_address_t)pExecMemAllocator->aChunks[idxChunk].pvChunkRx,
1540	pExecMemAllocator->cbChunk);
1541	Assert(krc == KERN_SUCCESS);
1542	# endif
1543
1544	RTMemPageFree(pvChunk, pExecMemAllocator->cbChunk);
1545	return rc;
1546	}
1547	#endif
1548
1549	return VINF_SUCCESS;
1550	}
1551
1552
1553	/**
1554	* Initializes the executable memory allocator for native recompilation on the
1555	* calling EMT.
1556	*
1557	* @returns VBox status code.
1558	* @param pVCpu The cross context virtual CPU structure of the calling
1559	* thread.
1560	* @param cbMax The max size of the allocator.
1561	* @param cbInitial The initial allocator size.
1562	* @param cbChunk The chunk size, 0 or UINT32_MAX for default (@a cbMax
1563	* dependent).
1564	*/
1565	int iemExecMemAllocatorInit(PVMCPU pVCpu, uint64_t cbMax, uint64_t cbInitial, uint32_t cbChunk) RT_NOEXCEPT
1566	{
1567	/*
1568	* Validate input.
1569	*/
1570	AssertLogRelMsgReturn(cbMax >= _1M && cbMax <= _4G+_4G, ("cbMax=%RU64 (%RX64)\n", cbMax, cbMax), VERR_OUT_OF_RANGE);
1571	AssertReturn(cbInitial <= cbMax, VERR_OUT_OF_RANGE);
1572	AssertLogRelMsgReturn( cbChunk != UINT32_MAX
1573	\|\| cbChunk == 0
1574	\|\| ( RT_IS_POWER_OF_TWO(cbChunk)
1575	&& cbChunk >= _1M
1576	&& cbChunk <= _256M
1577	&& cbChunk <= cbMax),
1578	("cbChunk=%RU32 (%RX32) cbMax=%RU64\n", cbChunk, cbChunk, cbMax),
1579	VERR_OUT_OF_RANGE);
1580
1581	/*
1582	* Adjust/figure out the chunk size.
1583	*/
1584	if (cbChunk == 0 \|\| cbChunk == UINT32_MAX)
1585	{
1586	if (cbMax >= _256M)
1587	cbChunk = _64M;
1588	else
1589	{
1590	if (cbMax < _16M)
1591	cbChunk = cbMax >= _4M ? _4M : (uint32_t)cbMax;
1592	else
1593	cbChunk = (uint32_t)cbMax / 4;
1594	if (!RT_IS_POWER_OF_TWO(cbChunk))
1595	cbChunk = RT_BIT_32(ASMBitLastSetU32(cbChunk));
1596	}
1597	}
1598	#ifdef IEMNATIVE_WITH_RECOMPILER_PER_CHUNK_TAIL_CODE
1599	# if defined(RT_OS_AMD64)
1600	Assert(cbChunk <= _2G);
1601	# elif defined(RT_OS_ARM64)
1602	if (cbChunk > _128M)
1603	cbChunk = _128M; /* Max relative branch distance is +/-2^(25+2) = +/-0x8000000 (134 217 728). */
1604	# endif
1605	#endif
1606
1607	if (cbChunk > cbMax)
1608	cbMax = cbChunk;
1609	else
1610	cbMax = (cbMax - 1 + cbChunk) / cbChunk * cbChunk;
1611	uint32_t const cMaxChunks = (uint32_t)(cbMax / cbChunk);
1612	AssertLogRelReturn((uint64_t)cMaxChunks * cbChunk == cbMax, VERR_INTERNAL_ERROR_3);
1613
1614	/*
1615	* Allocate and initialize the allocatore instance.
1616	*/
1617	size_t const offBitmaps = RT_ALIGN_Z(RT_UOFFSETOF_DYN(IEMEXECMEMALLOCATOR, aChunks[cMaxChunks]), RT_CACHELINE_SIZE);
1618	size_t const cbBitmaps = (size_t)(cbChunk >> (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT + 3)) * cMaxChunks;
1619	size_t cbNeeded = offBitmaps + cbBitmaps;
1620	AssertCompile(IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT <= 10);
1621	Assert(cbChunk > RT_BIT_32(IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT + 3));
1622	#if defined(IN_RING3) && !defined(RT_OS_WINDOWS)
1623	size_t const offEhFrames = RT_ALIGN_Z(cbNeeded, RT_CACHELINE_SIZE);
1624	cbNeeded += sizeof(IEMEXECMEMCHUNKEHFRAME) * cMaxChunks;
1625	#endif
1626	PIEMEXECMEMALLOCATOR pExecMemAllocator = (PIEMEXECMEMALLOCATOR)RTMemAllocZ(cbNeeded);
1627	AssertLogRelMsgReturn(pExecMemAllocator, ("cbNeeded=%zx cMaxChunks=%#x cbChunk=%#x\n", cbNeeded, cMaxChunks, cbChunk),
1628	VERR_NO_MEMORY);
1629	pExecMemAllocator->uMagic = IEMEXECMEMALLOCATOR_MAGIC;
1630	pExecMemAllocator->cbChunk = cbChunk;
1631	pExecMemAllocator->cMaxChunks = cMaxChunks;
1632	pExecMemAllocator->cChunks = 0;
1633	pExecMemAllocator->idxChunkHint = 0;
1634	pExecMemAllocator->cAllocations = 0;
1635	pExecMemAllocator->cbTotal = 0;
1636	pExecMemAllocator->cbFree = 0;
1637	pExecMemAllocator->cbAllocated = 0;
1638	pExecMemAllocator->pbmAlloc = (uint64_t *)((uintptr_t)pExecMemAllocator + offBitmaps);
1639	pExecMemAllocator->cUnitsPerChunk = cbChunk >> IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT;
1640	pExecMemAllocator->cBitmapElementsPerChunk = cbChunk >> (IEMEXECMEM_ALT_SUB_ALLOC_UNIT_SHIFT + 6);
1641	memset(pExecMemAllocator->pbmAlloc, 0xff, cbBitmaps); /* Mark everything as allocated. Clear when chunks are added. */
1642	#if defined(IN_RING3) && !defined(RT_OS_WINDOWS)
1643	pExecMemAllocator->paEhFrames = (PIEMEXECMEMCHUNKEHFRAME)((uintptr_t)pExecMemAllocator + offEhFrames);
1644	#endif
1645	for (uint32_t i = 0; i < cMaxChunks; i++)
1646	{
1647	pExecMemAllocator->aChunks[i].cFreeUnits = 0;
1648	pExecMemAllocator->aChunks[i].idxFreeHint = 0;
1649	pExecMemAllocator->aChunks[i].pvChunkRw = NULL;
1650	#ifdef IN_RING0
1651	pExecMemAllocator->aChunks[i].hMemObj = NIL_RTR0MEMOBJ;
1652	#else
1653	pExecMemAllocator->aChunks[i].pvUnwindInfo = NULL;
1654	#endif
1655	}
1656	pVCpu->iem.s.pExecMemAllocatorR3 = pExecMemAllocator;
1657
1658	/*
1659	* Do the initial allocations.
1660	*/
1661	while (cbInitial < (uint64_t)pExecMemAllocator->cChunks * pExecMemAllocator->cbChunk)
1662	{
1663	int rc = iemExecMemAllocatorGrow(pVCpu, pExecMemAllocator);
1664	AssertLogRelRCReturn(rc, rc);
1665	}
1666
1667	pExecMemAllocator->idxChunkHint = 0;
1668
1669	/*
1670	* Register statistics.
1671	*/
1672	PUVM const pUVM = pVCpu->pUVCpu->pUVM;
1673	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cAllocations, STAMTYPE_U32, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
1674	"Current number of allocations", "/IEM/CPU%u/re/ExecMem/cAllocations", pVCpu->idCpu);
1675	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cChunks, STAMTYPE_U32, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,
1676	"Currently allocated chunks", "/IEM/CPU%u/re/ExecMem/cChunks", pVCpu->idCpu);
1677	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cMaxChunks, STAMTYPE_U32, STAMVISIBILITY_ALWAYS, STAMUNIT_COUNT,
1678	"Maximum number of chunks", "/IEM/CPU%u/re/ExecMem/cMaxChunks", pVCpu->idCpu);
1679	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbChunk, STAMTYPE_U32, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
1680	"Allocation chunk size", "/IEM/CPU%u/re/ExecMem/cbChunk", pVCpu->idCpu);
1681	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbAllocated, STAMTYPE_U64, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
1682	"Number of bytes current allocated", "/IEM/CPU%u/re/ExecMem/cbAllocated", pVCpu->idCpu);
1683	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbFree, STAMTYPE_U64, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
1684	"Number of bytes current free", "/IEM/CPU%u/re/ExecMem/cbFree", pVCpu->idCpu);
1685	STAMR3RegisterFU(pUVM, &pExecMemAllocator->cbTotal, STAMTYPE_U64, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES,
1686	"Total number of byte", "/IEM/CPU%u/re/ExecMem/cbTotal", pVCpu->idCpu);
1687	#ifdef VBOX_WITH_STATISTICS
1688	STAMR3RegisterFU(pUVM, &pExecMemAllocator->StatAlloc, STAMTYPE_PROFILE, STAMVISIBILITY_ALWAYS, STAMUNIT_TICKS_PER_CALL,
1689	"Profiling the allocator", "/IEM/CPU%u/re/ExecMem/ProfAlloc", pVCpu->idCpu);
1690	#endif
1691	#ifdef IEMEXECMEM_ALT_SUB_WITH_ALT_PRUNING
1692	STAMR3RegisterFU(pUVM, &pExecMemAllocator->StatPruneProf, STAMTYPE_PROFILE, STAMVISIBILITY_ALWAYS, STAMUNIT_TICKS_PER_CALL,
1693	"Pruning executable memory (alt)", "/IEM/CPU%u/re/ExecMem/Pruning", pVCpu->idCpu);
1694	STAMR3RegisterFU(pUVM, &pExecMemAllocator->StatPruneRecovered, STAMTYPE_PROFILE, STAMVISIBILITY_ALWAYS, STAMUNIT_BYTES_PER_CALL,
1695	"Bytes recovered while pruning", "/IEM/CPU%u/re/ExecMem/PruningRecovered", pVCpu->idCpu);
1696	#endif
1697
1698	return VINF_SUCCESS;
1699	}
1700

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

source: vbox/trunk/src/VBox/VMM/VMMAll/IEMAllN8veExecMem.cpp@ 105284

Download in other formats: