Debian Codegraph @

extra/gperftools/gperftools-2.15/NEWS

raw
1 == 5 Jan 2024
2 gperftools 2.15 is out!
3
4 This release has the following bug fixes:
5
6 * Xiaowei Wang has pointed out the pthread linking issue on cmake on
7 older glibcs (where -pthread is not implicit). See
8 https://github.com/gperftools/gperftools/pull/1473 for more details.
9
10 * Mikael Simberg and Tom "spot" Callaway have pointed out the missing
11 symbols issue when linking PPC or i386
12 builds. https://github.com/gperftools/gperftools/issues/1474 has all
13 the details.
14
15 Huge thanks to all contributors!
16
17 == 31 Dec 2023
18 gperftools 2.14 is out!
19
20 This release has the following set of notable changes:
21
22 * Roman Geissler has contributed a fix to nasty initialization bug
23 introduced in 2.13 (see github issue #1452 for one example where it
24 fails).
25
26 * spinlock delay support now has proper windows support. Instead of
27 simply sleeping, it uses WaitOnAddress (which is basically windows
28 equivalent of futexes). This improvement was contributed by Lennox
29 Ho.
30
31 * we now have basic QNX support (basic malloc + heap profiler)
32 championed by Xiang.Lin. Thanks! Do note, however, that QNX doesn't
33 provide SIGPROF ticks, so there will be no cpu profiler support on
34 this OS.
35
36 * Yikai Zhao has contributed several fixes to important corner cases
37 of generic_fp stacktrace method.
38
39 * several people have contributed various improvements to our cmake
40 build: Lennox Ho, Sergey Fedorov, Mateusz Jakub Fila. But do note
41 that cmake build is still incomplete and best-effort.
42
43 * Julian Schroeder have fixed generic_fp incompatibility with ARM
44 pointer auth.
45
46 * Mateusz Jakub Fila has contributed implementation of mallocinfo2
47 function (64-bit version of mallinfo).
48
49 * Lennox Ho has updated C malloc extension shims to include
50 {Set,Get}MemoryReleaseRate.
51
52 * Lennox Ho has contributed the ability to disable malloc functions
53 patching on windows when TCMALLOC_DISABLE_REPLACEMENT=1 environment
54 variable is set.
55
56 * User poljak181 has contributed a fix to infinite recursion in some
57 cases of malloc hooks (or user-replaced operator new) and
58 MallocExtension::instance().
59
60 * Sergey Fedorov has contributed a fix to use MAP_ANON on some older
61 OSes without MAP_ANONYMOUS.
62
63 * the way we detect working ucontext->pc extraction method was
64 reworked and is now fully compile-time as opposed to
65 config-time. This means no more duplication and mismatches between
66 autoconf and cmake bits in this area.
67
68 List of relevant tickets can be seen online at:
69 https://github.com/gperftools/gperftools/issues?q=label%3Afixed-in-2.14+
70
71 == 11 Sep 2023
72 gperftools 2.13 is out!
73
74 This release includes a few minor fixes:
75
76 * Ivan Dlugos has fixed some issues with cmake and config.h defines.
77
78 * 32-bit builds no longer require 64-bit atomics (which we wrongly
79 introduced in 2.11 and which broke builds on some 32-bit
80 architectures).
81
82 * generic_fp backtracing method now uses robust address probing
83 method. The previous approach had occasional false positives, which
84 caused occasional rare crashes.
85
86 * In some cases, MSVC generated TrivialOnce machine code that
87 deadlocked programs on startup. The issue is now fixed.
88
89 == 24 Aug 2023
90 gperftools 2.12 is out!
91
92 Brett T. Warden contributed one significant fix. After a change in the
93 previous release, we installed broken pkg-config files. Brett noticed
94 and fixed that. Huge thanks!
95
96 == 14 Aug 2023
97 gperftools 2.11 is out!
98
99 Few minor fixes since rc couple weeks ago. Plus couple notable
100 contributions:
101
102 * Artem Polyakov has contributed auto-detection of several MPI systems
103 w.r.t. filenames used by HEAPPROFILE and CPUPROFILE environment
104 variables. Also, we now support HEAPPROFILE_USE_PID and
105 CPUPROFILE_USE_PID environment variables that force profile
106 filenames to have pid appended. Which will be useful for some
107 programs that fork for parallelism. See
108 https://github.com/gperftools/gperftools/pull/1263 for details.
109
110 * Ken Raffenetti has extended MPI detection mentioned above with
111 detection of MPICH system.
112
113 Thanks a lot!
114
115 == 31 July 2023
116 gperftools 2.11rc is out!
117
118 Most notable change is that Linux/aarch64 and Linux/riscv are now
119 fully supported. That is, all unit tests pass on those architectures
120 (previously the heap leak checker was broken).
121
122 Also notable is that heap leak checker support is officially
123 deprecated as of this release. All bug fixes from now are on a best
124 effort basis. For clarity we also declare that it is only expected to
125 work (for some definition of work) on Linux/x86 (all kinds),
126 Linux/aarch64, Linux/arm, Linux/ppc (untested as of this writing) and
127 Linux/mips (untested as well). While some functionality worked in the
128 past on BSDs, it was never fully functional; and will never be. We
129 strongly recommend everyone to switch to asan and friends.
130
131 For major internal changes it is also worth mentioning that we now
132 fully switched to C++-11 std::atomic. All custom OS- and arch-specific
133 atomic bits have been removed at last.
134
135 Another notable change is that mmap and sbrk hooks facility is now
136 no-op. We keep API and ABI for formal compatibility, but the calls to
137 add mmap/sbrk hooks do nothing and return an error (whenever possible
138 as part of API). There seem to be no users of it anyways, and mmap
139 replacement API that is part of that facility really screwed up 64-bit
140 offsets on (some/most) 32-bit systems. Internally for heap profiler
141 and heap checker we have a new, but non-public API (see mmap_hook.h).
142
143 Most tests now pass on NetBSD x86-64 (I tested on version 9.2). And
144 only one that fails is new stacktrace test for stacktraces from signal
145 handler (so there could be some imperfections for cpu profiles).
146
147 We don't warn people away from the libgcc stacktrace capturing method
148 anymore. In fact users on most recent glibc-s are advised to use it
149 (pass --enable-libgcc-unwinder-by-default). This is thanks to the
150 dl_find_object API offered by glibc which allows this implementation
151 to be fully async-signal-safe. Modern Linux distros should from now on
152 build their gperftools package with this enabled (other than those
153 built on top of musl).
154
155 generic_fp and generic_fp_unsafe stacktrace capturing methods have
156 been expanded for more architectures and even some basic non-Linux
157 support. We have completely removed old x86-specific frame pointer
158 stacktrace implementation in favor of those 2. _unsafe one should be
159 roughly equivalent to the old x86 method. And 'safe' one is
160 recommended as a new default for those who want FP-based
161 stacktracing. Safe implementation robustly checks memory before
162 accessing it, preventing unlikely, but not impossible crashes when
163 frame pointers are bogus.
164
165 On platforms that support it, we now build gperftools with
166 "-fno-omit-frame-pointer -momit-leaf-frame-pointer". This makes
167 gperftools mostly frame-pointer-ful, but without performance hit in
168 places that matter (this is how Google builds their binaries
169 BTW). That should cover gcc (at least) on x86, aarch64 and
170 riscv. Intention for this change is to make distro-shipped
171 libtcmalloc.so compatible with frame-pointer stacktrace capturing (for
172 those who still do heap profiling, for example). Of course, passing
173 --enable-frame-pointers still gives you full frame pointers (i.e. even
174 for leaf functions).
175
176 There is now support for detecting actual page size at
177 runtime. tcmalloc will now allocate memory in units of this page
178 size. It particularly helps on arms with 64k pages to return memory
179 back to the kernel. But it is somewhat controversial, because it
180 effectively bumps tcmalloc logical page size on those machines
181 potentially increasing fragmentation. In any case, there is now a new
182 environment variable TCMALLOC_OVERRIDE_PAGESIZE allowing people to
183 override this check. I.e. to either reduce effective page size down to
184 tcmalloc's logical page size or to increase it.
185
186 MallocExtension::MarkThreadTemporarilyIdle has been changed to be
187 identical to MarkThreadIdle. MarkThreadTemporarilyIdle is believed to
188 be unused, anyways. See issue #880 for details.
189
190 There are a whole bunch of smaller fixes. Many of those smaller fixes
191 had no associated ticket, but some had. People are advised to see here
192 for list of notable tickets closed in this release:
193 https://github.com/gperftools/gperftools/issues?q=label%3Afixed-in-2.11+
194
195 Some of those tickets are quite notable (fixes for rare deadlocks in
196 cpu profiler ProfilerStop or while capturing heap growth stacktraces
197 (aka growthz)).
198
199 Here is list of notable contributions:
200
201 * Chris Cambly has contributed initial support for AIX
202
203 * Ali Saidi has contributed SpinlockPause implementation for aarch64
204
205 * Henrik Reinstädtler has contributed fix for cpuprofiler on aarch64
206 OSX
207
208 * Gabriel Marin has backported Chromium's commit for always sanity
209 checking large frees
210
211 * User zhangyiru has contributed a fix to report the number of leaked
212 bytes as size_t instead of (usually 32-bit) int.
213
214 * Sergey Fedorov has contributed some fix for building on older
215 ppc-based OSX-es
216
217 * User tigeran has removed unused using declaration
218
219 Huge thanks to all contributors.
220
221 == 30 May 2022 ==
222 gperftools 2.10 is out!
223
224 Here are notable changes:
225
226 * Matt T. Proud contributed documentation fix to call Go programming
227 language by it's true name instead of golang.
228
229 * Robert Scott contributed debugallocator feature to use readable
230 (PROT_READ) fence pages. This is activated by
231 TCMALLOC_PAGE_FENCE_READABLE environment veriable.
232
233 * User stdpain contributed fix for cmake detection of libunwind.
234
235 * Natale Patriciello contributed fix for OSX Monterey support.
236
237 * Volodymyr Nikolaichuk contributed support for returning memory back
238 to OS by using mmap with MAP_FIXED and PROT_NONE. It is off by
239 default and enabled by preprocessor define:
240 FREE_MMAP_PROT_NONE. This should help OSes that don't support
241 Linux-style madvise MADV_DONTNEED or BSD-style MADV_FREE.
242
243 * Jingyun Hua has contributed basic support for LoongArch.
244
245 * Github issue #1338 of failing to build on some recent musl versions
246 has been fixed.
247
248 * Github issue #1321 of failing to ship cmake bits with .tar.gz
249 archive has been fixed.
250
251 == 2 March 2021 ==
252 gperftools 2.9.1 is out!
253
254 Minor fixes landed since previous release:
255
256 * OSX builds new prefer backtrace() and have somewhat working heap
257 sampling.
258
259 * Incorrect assertion failure was fixed that crashed tcmalloc if
260 assertions were on and sized delete was used. More details in github
261 issue #1254.
262
263 == 21 February 2021 ==
264 gperftools 2.9 is out!
265
266 Few more changes landed compared to rc:
267
268 * Venkatesh Srinivas has contributed thread-safety annotations
269 support.
270
271 * couple more unit test bugs that caused tcmalloc_unittest to fail on
272 recent clang has been fixed.
273
274 * usage of unsupportable linux_syscall_support.h has been removed from
275 few places. Building with --disable-heap-checker now completely
276 avoids it. Expect complete death of this header in next major
277 release.
278
279 == 14 February 2021 ==
280 gperftools 2.9rc is out!
281
282 Here are notable changes:
283
284 * Jarno Rajahalme has contributed fix for crashing bug in syscalls
285 support for aarch64.
286
287 * User SSE4 has contributed basic support for Elbrus 2000 architecture
288 (!)
289
290 * Venkatesh Srinivas has contributed cleanup to atomic ops.
291
292 * Đoàn Trần Công Danh has fixed cpu profiler compilation on musl.
293
294 * there is now better backtracing support for aarch64 and
295 riscv. x86-64 with frame pointers now also defaults to this new
296 "generic" frame pointer backtracer.
297
298 * emergency malloc is now enabled by default. Fixes hang on musl when
299 libgcc backtracer is enabled.
300
301 * bunch of legacy config tests has been removed
302
303 == 20 December 2020 ==
304 gperftools 2.8.1 is out!
305
306 Here are notable changes:
307
308 * previous release contained change to release memory without page
309 heap lock, but this change had at least one bug that caused to
310 crashes and corruption when running under aggressive decommit mode
311 (this is not default). While we check for other bugs, this feature
312 was reverted. See github issue #1204 and issue #1227.
313
314 * stack traces depth captured by gperftools is now up to 254 levels
315 deep. Thanks to Kerrick Staley for this small but useful tweak.
316
317 * Levon Ter-Grigoryan has contributed small fix for compiler warning.
318
319 * Grant Henke has contributed updated detection of program counter
320 register for OS X on arm64.
321
322 * Tim Gates has contributed small typo fix.
323
324 * Steve Langasek has contributed basic build fixes for riscv64 (!).
325
326 * Isaac Hier and okhowang have contributed premiliminary port of build
327 infrastructure to cmake. This works, but it is very premiliminary.
328 Autotools-based build is the only officially supported build for
329 now.
330
331 == 6 July 2020 ==
332 gperftools 2.8 is out!
333
334 Here are notable changes:
335
336 * ProfilerGetStackTrace is now officially supported API for
337 libprofiler. Contributed by Kirill Müller.
338
339 * Build failures on mingw were fixed. This fixed issue #1108.
340
341 * Build failure of page_heap_test on MSVC was fixed.
342
343 * Ryan Macnak contributed fix for compiling linux syscall support on
344 i386 and recent GCCs. This fixed issue #1076.
345
346 * test failures caused by new gcc 10 optimizations were fixed. Same
347 change also fixed tests on clang.
348
349 == 8 Mar 2020 ==
350 gperftools 2.8rc is out!
351
352 Here are notable changes:
353
354 * building code now requires c++11 or later. Bundled MSVC project was
355 converted to Visual Studio 2015.
356
357 * User obones contributed fix for windows x64 TLS callbacks. This
358 fixed leak of thread caches on thread exists in 64-bit windows.
359
360 * releasing memory back to kernel is now made with page heap lock
361 dropped.
362
363 * HoluWu contributed fix for correct malloc patching on debug builds
364 on windows. This configuration previously crashed.
365
366 * Romain Geissler contributed fix for tls access during early tls
367 initialization on dlopen.
368
369 * large allocation reports are now silenced by default. Since not all
370 programs want their stderr polluted by those messages. Contributed
371 by Junhao Li.
372
373 * HolyWu contributed improvements to MSVC project files. Notably,
374 there is now project for "overriding" version of tcmalloc.
375
376 * MS-specific _recalloc is now correctly zeroing only malloced
377 part. This fix was contributed by HolyWu.
378
379 * Brian Silverman contributed correctness fix to sampler_test.
380
381 * Gabriel Marin ported few fixes from chromium's fork. As part of
382 those fixes, we reduced number of static initializers (forbidden in
383 chromium). Also we now syscalls via syscall function instead of
384 reimplementing direct way to make syscalls on each platform.
385
386 * Brian Silverman fixed flakiness in page heap test.
387
388 * There is now configure flag to skip installing perl pprof, since
389 external golang pprof is much superior. --disable-deprecated-pprof
390 is the flag.
391
392 * Fabric Fontaine contributed fixes to drop use of nonstandard
393 __off64_t type.
394
395 * Fabrice Fontaine contributed build fix to check for presence of
396 nonstandard __sbrk functions. It is only used by mmap hooks code and
397 (rightfully) not available on musl.
398
399 * Fabrice Fontaine contributed build fix around mmap64 macro and
400 function conflict in same cases.
401
402 * there is now configure time option to enable aggressive decommit by
403 default. Contributed by Laurent
404 Stacul. --enable-aggressive-decommit-by-default is the flag.
405
406 * Tulio Magno Quites Machado Filho contributed build fixes for ppc
407 around ucontext access.
408
409 * User pkubaj contributed couple build fixes for FreeBSD/ppc.
410
411 * configure now always assumes we have mmap. This fixes configure
412 failures on some linux guests inside virtualbox. This fixed issue
413 #1008.
414
415 * User shipujin contributed syscall support fixes for mips64 (big and
416 little endian).
417
418 * Henrik Edin contributed configurable support for wide range of
419 malloc page sizes. 4K, 8K, 16K, 32K, 64K, 128K and 256K are now
420 supported via existing --with-tcmalloc-pagesize flag to configure.
421
422 * Jon Kohler added overheads fields to per-size-class textual
423 stats. Stats that are available via
424 MallocExtension::instance()->GetStats().
425
426 * tcmalloc can now avoid fallback from memfs to default sys
427 allocator. TCMALLOC_MEMFS_DISABLE_FALLBACK switches this on. This
428 was contributed by Jon Kohler.
429
430 * Ilya Leoshkevich fixed mmap syscall support on s390.
431
432 * Todd Lipcon contributed small build warning fix.
433
434 * User prehistoricpenguin contributed misc source file mode fixes (we
435 still had few few c++ files marked executable).
436
437 * User invalid_ms_user contributed fix for typo.
438
439 * Jakub Wilk contributed typos fixes.
440
441 == 29 Apr 2018 ==
442 gperftools 2.7 is out!
443
444 Few people contributed minor, but important fixes since rc.
445
446 Changes:
447
448 * bug in span stats printing introduced by new scalable page heap
449 change was fixed.
450
451 * Christoph Müllner has contributed couple warnings fixes and initial
452 support for aarch64_ilp32 architecture.
453
454 * Ben Dang contributed documentation fix for heap checker.
455
456 * Fabrice Fontaine contributed fixed for linking benchmarks with
457 --disable-static.
458
459 * Holy Wu has added sized deallocation unit tests.
460
461 * Holy Wu has enabled support of sized deallocation (c++14) on recent
462 MSVC.
463
464 * Holy Wu has fixed MSVC build in WIN32_OVERRIDE_ALLOCATORS mode. This
465 closed issue #716.
466
467 * Holy Wu has contributed cleanup of config.h used on windows.
468
469 * Mao Huang has contributed couple simple tcmalloc changes from
470 chromium code base. Making our tcmalloc forks a tiny bit closer.
471
472 * issue #946 that caused compilation failures on some Linux clang
473 installations has been fixed. Much thanks to github user htuch for
474 helping to diagnose issue and proposing a fix.
475
476 * Tulio Magno Quites Machado Filho has contributed build-time fix for
477 PPC (for problem introduced in one of commits since RC).
478
479 == 18 Mar 2018 ==
480 gperftools 2.7rc is out!
481
482 Changes:
483
484 * Most notable change in this release is that very large allocations
485 (>1MiB) are now handled be O(log n) implementation. This is
486 contributed by Todd Lipcon based on earlier work by Aliaksei
487 Kandratsenka and James Golick. Special thanks to Alexey Serbin for
488 contributing OSX fix for that commit.
489
490 * detection of sized deallocation support is improved. Which should
491 fix another set of issues building on OSX. Much thanks to Alexey
492 Serbin for reporting the issue, suggesting a fix and verifying it.
493
494 * Todd Lipcon made a change to extend page heaps freelists to 1 MiB
495 (up from 1MiB - 8KiB). This may help a little for some workloads.
496
497 * Ishan Arora contributed typo fix to docs
498
499 == 9 Dec 2017 ==
500 gperftools 2.6.3 is out!
501
502 Just two fixes were made in this release:
503
504 * Stephan Zuercher has contributed a build fix for some recent XCode
505 versions. See issue #942 for more details.
506
507 * assertion failure on some windows builds introduced by 2.6.2 was
508 fixed. Thanks to github user nkeemik for reporting it and testing
509 fix. See issue #944 for more details.
510
511 == 30 Nov 2017 ==
512 gperftools 2.6.2 is out!
513
514 Most notable change is recently added support for C++17 over-aligned
515 allocation operators contributed by Andrey Semashev. I've extended his
516 implemention to have roughly same performance as malloc/new. This
517 release also has native support for C11 aligned_alloc.
518
519 Rest is mostly bug fixes:
520
521 * Jianbo Yang has contributed a fix for potentially severe data race
522 introduced by malloc fast-path work in gperftools 2.6. This race
523 could cause occasional violation of total thread cache size
524 constraint. See issue #929 for more details.
525
526 * Correct behavior in out-of-memory condition in fast-path cases was
527 restored. This was another bug introduced by fast-path optimization
528 in gperftools 2.6 which caused operator new to silently return NULL
529 instead of doing correct C++ OOM handling (calling new_handler and
530 throwing bad_alloc).
531
532 * Khem Raj has contributed couple build fixes for newer glibcs (ucontext_t vs
533 struct ucontext and loff_t definition)
534
535 * Piotr Sikora has contributed build fix for OSX (not building unwind
536 benchmark). This was issue #910 (thanks to Yuriy Solovyov for
537 reporting it).
538
539 * Dorin Lazăr has contributed fix for compiler warning
540
541 * issue #912 (occasional deadlocking calling getenv too early on
542 windows) was fixed. Thanks to github user shangcangriluo for
543 reporting it.
544
545 * Couple earlier lsan-related commits still causing occasional issues
546 linking on OSX has been reverted. See issue #901.
547
548 * Volodimir Krylov has contributed GetProgramInvocationName for FreeBSD
549
550 * changsu lee has contributed couple minor correctness fixes (missing
551 va_end() and missing free() call in rarely executed Symbolize path)
552
553 * Andrew C. Morrow has contributed some more page heap stats. See issue
554 #935.
555
556 * some cases of built-time warnings from various gcc/clang versions
557 about throw() declarations have been fixes.
558
559 == 9 July 2017 ==
560
561 gperftools 2.6.1 is out! This is mostly bug-fixes release.
562
563 * issue #901: build issue on OSX introduced in last-time commit in 2.6
564 was fixed (contributed by Francis Ricci)
565
566 * tcmalloc_minimal now works on 32-bit ABI of mips64. This is issue
567 #845. Much thanks to Adhemerval Zanella and github user mtone.
568
569 * Romain Geissler contributed build fix for -std=c++17. This is pull
570 request #897.
571
572 * As part of fixing issue #904, tcmalloc atfork handler is now
573 installed early. This should fix slight chance of hitting deadlocks
574 at fork in some cases.
575
576 == 4 July 2017 ==
577
578 gperftools 2.6 is out!
579
580 * Kim Gräsman contributed documentation update for HEAPPROFILESIGNAL
581 environment variable
582
583 * KernelMaker contributed fix for population of min_object_size field
584 returned by MallocExtension::GetFreeListSizes
585
586 * commit 8c3dc52fcfe0 "issue-654: [pprof] handle split text segments"
587 was reverted. Some OSX users reported issues with this commit. Given
588 our pprof implementation is strongly deprecated it is best to drop
589 recently introduced features rather than breaking it badly.
590
591 * Francis Ricci contributed improvement for interaction with leak
592 sanitizer.
593
594 == 22 May 2017 ==
595
596 gperftools 2.6rc4 is out!
597
598 Dynamic sized delete is disabled by default again. There is no hope of
599 it working with eager dynamic symbols resolution (-z now linker
600 flag). More details in
601 https://bugzilla.redhat.com/show_bug.cgi?id=1452813
602
603 == 21 May 2017 ==
604
605 gperftools 2.6rc3 is out!
606
607 gperftools compilation on older systems (e.g. rhel 5) was fixed. This
608 was originally reported in github issue #888.
609
610 == 14 May 2017 ==
611
612 gperftools 2.6rc2 is out!
613
614 Just 2 small fixes on top of 2.6rc. Particularly, Rajalakshmi
615 Srinivasaraghavan contributed build fix for ppc32.
616
617 == 14 May 2017 ==
618
619 gperftools 2.6rc is out!
620
621 Highlights of this release are performance work on malloc fast-path
622 and support for more modern visual studio runtimes, and deprecation of
623 bundled pprof. Another significant performance-affecting changes are
624 reverting central free list transfer batch size back to 32 and
625 disabling of aggressive decommit mode by default.
626
627 Note, while we still ship perl implementation of pprof, everyone is
628 strongly advised to use golang reimplementation of pprof from
629 https://github.com/google/pprof.
630
631 Here are notable changes in more details (and see ChangeLog for full
632 details):
633
634 * a bunch of performance tweaks to tcmalloc fast-path were
635 merged. This speeds up critical path of tcmalloc by few tens of
636 %. Well tuned and allocation-heavy programs should see substantial
637 performance boost (should apply to all modern elf platforms). This
638 is based on Google-internal tcmalloc changes for fast-path (with
639 obvious exception of lacking per-cpu mode, of course). Original
640 changes were made by Aliaksei Kandratsenka. And Andrew Hunter,
641 Dmitry Vyukov and Sanjay Ghemawat contributed with reviews and
642 discussions.
643
644 * Architectures with 48 bits address space (x86-64 and aarch64) now
645 use faster 2 level page map. This was ported from Google-internal
646 change by Sanjay Ghemawat.
647
648 * Default value of TCMALLOC_TRANSFER_NUM_OBJ was returned back to
649 32. Larger values have been found to hurt certain programs (but help
650 some other benchmarks). Value can still be tweaked at run time via
651 environment variable.
652
653 * tcmalloc aggressive decommit mode is now disabled by default
654 again. It was found to degrade performance of certain tensorflow
655 benchmarks. Users who prefer smaller heap over small performance win
656 can still set environment variable TCMALLOC_AGGRESSIVE_DECOMMIT=t.
657
658 * runtime switchable sized delete support has be fixed and re-enabled
659 (on GNU/Linux). Programs that use C++ 14 or later that use sized
660 delete can again be sped up by setting environment variable
661 TCMALLOC_ENABLE_SIZED_DELETE=t. Support for enabling sized
662 deallication support at compile-time is still present, of course.
663
664 * tcmalloc now explicitly avoids use of MADV_FREE on Linux, unless
665 TCMALLOC_USE_MADV_FREE is defined at compile time. This is because
666 performance impact of MADV_FREE is not well known. Original issue
667 #780 raised by Mathias Stearn.
668
669 * issue #786 with occasional deadlocks in stack trace capturing via
670 libunwind was fixed. It was originally reported as Ceph issue:
671 http://tracker.ceph.com/issues/13522
672
673 * ChangeLog is now automatically generated from git log. Old ChangeLog
674 is now ChangeLog.old.
675
676 * tcmalloc now provides implementation of nallocx. Function was
677 originally introduced by jemalloc and can be used to return real
678 allocation size given allocation request size. This is ported from
679 Google-internal tcmalloc change contributed by Dmitry Vyukov.
680
681 * issue #843 which made tcmalloc crash when used with erlang runtime
682 was fixed.
683
684 * issue #839 which caused tcmalloc's aggressive decommit mode to
685 degrade performance in some corner cases was fixed.
686
687 * Bryan Chan contributed support for 31-bit s390.
688
689 * Brian Silverman contributed compilation fix for 32-bit ARMs
690
691 * Issue #817 that was causing tcmalloc to fail on windows 10 and
692 later, as well as on recent msvc was fixed. We now patch _free_base
693 as well.
694
695 * a bunch of minor documentaion/typos fixes by: Mike Gaffney
696 <[email protected]>, iivlev <[email protected]>, savefromgoogle
697 <[email protected]>, John McDole
698 <[email protected]>, zmertens <[email protected]>, Kirill Müller
699 <[email protected]>, Eugene <[email protected]>, Ola Olsson
700 <[email protected]>, Mostyn Bramley-Moore <[email protected]>
701
702 * Tulio Magno Quites Machado Filho has contributed removal of
703 deprecated glibc malloc hooks.
704
705 * Issue #827 that caused intercepting malloc on osx 10.12 to fail was
706 fixed, by copying fix made by Mike Hommey to jemalloc. Much thanks
707 to Koichi Shiraishi and David Ribeiro Alves for reporting it and
708 testing fix.
709
710 * Aman Gupta and Kenton Varda contributed minor fixes to pprof (but
711 note again that pprof is deprecated)
712
713 * Ryan Macnak contributed compilation fix for aarch64
714
715 * Francis Ricci has fixed unaligned memory access in debug allocator
716
717 * TCMALLOC_PAGE_FENCE_NEVER_RECLAIM now actually works thanks to
718 contribution by Andrew Morrow.
719
720 == 12 Mar 2016 ==
721
722 gperftools 2.5 is out!
723
724 Just single bugfix was merged after rc2. Which was fix for issue #777.
725
726 == 5 Mar 2016 ==
727
728 gperftools 2.5rc2 is out!
729
730 New release contains just few commits on top of first release
731 candidate. One of them is build fix for Visual Studio. Another
732 significant change is that dynamic sized delete is now disabled by
733 default. It turned out that IFUNC relocations are not supporting our
734 advanced use case on all platforms and in all cases.
735
736 == 21 Feb 2016 ==
737
738 gperftools 2.5rc is out!
739
740 Here are major changes since 2.4:
741
742 * we've moved to github!
743
744 * Bryan Chan has contributed s390x support
745
746 * stacktrace capturing via libgcc's _Unwind_Backtrace was implemented
747 (for architectures with missing or broken libunwind).
748
749 * "emergency malloc" was implemented. Which unbreaks recursive calls
750 to malloc/free from stacktrace capturing functions (such us glib'c
751 backtrace() or libunwind on arm). It is enabled by
752 --enable-emergency-malloc configure flag or by default on arm when
753 --enable-stacktrace-via-backtrace is given. It is another fix for a
754 number common issues people had on platforms with missing or broken
755 libunwind.
756
757 * C++14 sized-deallocation is now supported (on gcc 5 and recent
758 clangs). It is off by default and can be enabled at configure time
759 via --enable-sized-delete. On GNU/Linux it can also be enabled at
760 run-time by either TCMALLOC_ENABLE_SIZED_DELETE environment variable
761 or by defining tcmalloc_sized_delete_enabled function which should
762 return 1 to enable it.
763
764 * we've lowered default value of transfer batch size to 512. Previous
765 value (bumped up in 2.1) was too high and caused performance
766 regression for some users. 512 should still give us performance
767 boost for workloads that need higher transfer batch size while not
768 penalizing other workloads too much.
769
770 * Brian Silverman's patch finally stopped arming profiling timer
771 unless profiling is started.
772
773 * Andrew Morrow has contributed support for obtaining cache size of the
774 current thread and softer idling (for use in MongoDB).
775
776 * we've implemented few minor performance improvements, particularly
777 on malloc fast-path.
778
779 A number of smaller fixes were made. Many of them were contributed:
780
781 * issue that caused spurious profiler_unittest.sh failures was fixed.
782
783 * Jonathan Lambrechts contributed improved callgrind format support to
784 pprof.
785
786 * Matt Cross contributed better support for debug symbols in separate
787 files to pprof.
788
789 * Matt Cross contributed support for printing collapsed stack frame
790 from pprof aimed at producing flame graphs.
791
792 * Angus Gratton has contributed documentation fix mentioning that on
793 windows only tcmalloc_minimal is supported.
794
795 * Anton Samokhvalov has made tcmalloc use mi_force_{un,}lock on OSX
796 instead of pthread_atfork. Which apparently fixes forking
797 issues tcmalloc had on OSX.
798
799 * Milton Chiang has contributed support for building 32-bit gperftools
800 on arm8.
801
802 * Patrick LoPresti has contributed support for specifying alternative
803 profiling signal via CPUPROFILE_TIMER_SIGNAL environment variable.
804
805 * Paolo Bonzini has contributed support configuring filename for
806 sending malloc tracing output via TCMALLOC_TRACE_FILE environment
807 variable.
808
809 * user spotrh has enabled use of futex on arm.
810
811 * user mitchblank has contributed better declaration for arg-less
812 profiler functions.
813
814 * Tom Conerly contributed proper freeing of memory allocated in
815 HeapProfileTable::FillOrderedProfile on error paths.
816
817 * user fdeweerdt has contributed curl arguments handling fix in pprof
818
819 * Frederik Mellbin fixed tcmalloc's idea of mangled new and delete
820 symbols on windows x64
821
822 * Dair Grant has contributed cacheline alignment for ThreadCache
823 objects
824
825 * Fredrik Mellbin has contributed updated windows/config.h for Visual
826 Studio 2015 and other windows fixes.
827
828 * we're not linking libpthread to libtcmalloc_minimal anymore. Instead
829 libtcmalloc_minimal links to pthread symbols weakly. As a result
830 single-threaded programs remain single-threaded when linking to or
831 preloading libtcmalloc_minimal.so.
832
833 * Boris Sazonov has contributed mips compilation fix and printf misue
834 in pprof.
835
836 * Adhemerval Zanella has contributed alignment fixes for statically
837 allocated variables.
838
839 * Jens Rosenboom has contributed fixes for heap-profiler_unittest.sh
840
841 * gshirishfree has contributed better description for GetStats method.
842
843 * cyshi has contributed spinlock pause fix.
844
845 * Chris Mayo has contributed --docdir argument support for configure.
846
847 * Duncan Sands has contributed fix for function aliases.
848
849 * Simon Que contributed better include for malloc_hook_c.h
850
851 * user wmamrak contributed struct timespec fix for Visual Studio 2015.
852
853 * user ssubotin contributed typo in PrintAvailability code.
854
855
856 == 10 Jan 2015 ==
857
858 gperftools 2.4 is out! The code is exactly same as 2.4rc.
859
860 == 28 Dec 2014 ==
861
862 gperftools 2.4rc is out!
863
864 Here are changes since 2.3:
865
866 * enabled aggressive decommit option by default. It was found to
867 significantly improve memory fragmentation with negligible impact on
868 performance. (Thanks to investigation work performed by Adhemerval
869 Zanella)
870
871 * added ./configure flags for tcmalloc pagesize and tcmalloc
872 allocation alignment. Larger page sizes have been reported to
873 improve performance occasionally. (Patch by Raphael Moreira Zinsly)
874
875 * sped-up hot-path of malloc/free. By about 5% on static library and
876 about 10% on shared library. Mainly due to more efficient checking
877 of malloc hooks.
878
879 * improved stacktrace capturing in cpu profiler (due to issue found by
880 Arun Sharma). As part of that issue pprof's handling of cpu profiles
881 was also improved.
882
883 == 7 Dec 2014 ==
884
885 gperftools 2.3 is out!
886
887 Here are changes since 2.3rc:
888
889 * (issue 658) correctly close socketpair fds on failure (patch by glider)
890
891 * libunwind integration can be disabled at configure time (patch by
892 Raphael Moreira Zinsly)
893
894 * libunwind integration is disabled by default for ppc64 (patch by
895 Raphael Moreira Zinsly)
896
897 * libunwind integration is force-disabled for OSX. It was not used by
898 default anyways. Fixes compilation issue I saw.
899
900 == 2 Nov 2014 ==
901
902 gperftools 2.3rc is out!
903
904 Most small improvements in this release were made to pprof tool.
905
906 New experimental Linux-only (for now) cpu profiling mode is a notable
907 big improvement.
908
909 Here are notable changes since 2.2.1:
910
911 * (issue-631) fixed debugallocation miscompilation on mmap-less
912 platforms (courtesy of user iamxujian)
913
914 * (issue-630) reference to wrong PROFILE (vs. correct CPUPROFILE)
915 environment variable was fixed (courtesy of WenSheng He)
916
917 * pprof now has option to display stack traces in output for heap
918 checker (courtesy of Michael Pasieka)
919
920 * (issue-636) pprof web command now works on mingw
921
922 * (issue-635) pprof now handles library paths that contain spaces
923 (courtesy of user [email protected])
924
925 * (issue-637) pprof now has an option to not strip template arguments
926 (patch by jiakai)
927
928 * (issue-644) possible out-of-bounds access in GetenvBeforeMain was
929 fixed (thanks to user abyss.7)
930
931 * (issue-641) pprof now has an option --show_addresses (thanks to user
932 yurivict). New option prints instruction address in addition to
933 function name in stack traces
934
935 * (issue-646) pprof now works around some issues of addr2line
936 reportedly when DWARF v4 format is used (patch by Adam McNeeney)
937
938 * (issue-645) heap profiler exit message now includes remaining memory
939 allocated info (patch by user yurivict)
940
941 * pprof code that finds location of /proc/<pid>/maps in cpu profile
942 files is now fixed (patch by Ricardo M. Correia)
943
944 * (issue-654) pprof now handles "split text segments" feature of
945 Chromium for Android. (patch by simonb)
946
947 * (issue-655) potential deadlock on windows caused by early call to
948 getenv in malloc initialization code was fixed (bug reported and fix
949 proposed by user zndmitry)
950
951 * incorrect detection of arm 6zk instruction set support
952 (-mcpu=arm1176jzf-s) was fixed. (Reported by pedronavf on old
953 issue-493)
954
955 * new cpu profiling mode on Linux is now implemented. It sets up
956 separate profiling timers for separate threads. Which improves
957 accuracy of profiling on Linux a lot. It is off by default. And is
958 enabled if both librt.f is loaded and CPUPROFILE_PER_THREAD_TIMERS
959 environment variable is set. But note that all threads need to be
960 registered via ProfilerRegisterThread.
961
962 == 21 Jun 2014 ==
963
964 gperftools 2.2.1 is out!
965
966 Here's list of fixes:
967
968 * issue-626 was closed. Which fixes initialization statically linked
969 tcmalloc.
970
971 * issue 628 was closed. It adds missing header file into source
972 tarball. This fixes for compilation on PPC Linux.
973
974 == 3 May 2014 ==
975
976 gperftools 2.2 is out!
977
978 Here are notable changes since 2.2rc:
979
980 * issue 620 (crash on windows when c runtime dll is reloaded) was
981 fixed
982
983 == 19 Apr 2014 ==
984
985 gperftools 2.2rc is out!
986
987 Here are notable changes since 2.1:
988
989 * a number of fixes for a number compilers and platforms. Notably
990 Visual Studio 2013, recent mingw with c++ threads and some OSX
991 fixes.
992
993 * we now have mips and mips64 support! (courtesy of Jovan Zelincevic,
994 Jean Lee, user xiaoyur347 and others)
995
996 * we now have aarch64 (aka arm64) support! (contributed by Riku
997 Voipio)
998
999 * there's now support for ppc64-le (by Raphael Moreira Zinsly and
1000 Adhemerval Zanella)
1001
1002 * there's now some support of uclibc (contributed by user xiaoyur347)
1003
1004 * google/ headers will now give you deprecation warning. They are
1005 deprecated since 2.0
1006
1007 * there's now new api: tc_malloc_skip_new_handler (ported from chromium
1008 fork)
1009
1010 * issue-557: added support for dumping heap profile via signal (by
1011 Jean Lee)
1012
1013 * issue-567: Petr Hosek contributed SysAllocator support for windows
1014
1015 * Joonsoo Kim contributed several speedups for central freelist code
1016
1017 * TCMALLOC_MAX_TOTAL_THREAD_CACHE_BYTES environment variable now works
1018
1019 * configure scripts are now using AM_MAINTAINER_MODE. It'll only
1020 affect folks who modify source from .tar.gz and want automake to
1021 automatically rebuild Makefile-s. See automake documentation for
1022 that.
1023
1024 * issue-586: detect main executable even if PIE is active (based on
1025 patch by user themastermind1). Notably, it fixes profiler use with
1026 ruby.
1027
1028 * there is now support for switching backtrace capturing method at
1029 runtime (via TCMALLOC_STACKTRACE_METHOD and
1030 TCMALLOC_STACKTRACE_METHOD_VERBOSE environment variables)
1031
1032 * there is new backtrace capturing method using -finstrument-functions
1033 prologues contributed by user xiaoyur347
1034
1035 * few cases of crashes/deadlocks in profiler were addressed. See
1036 (famous) issue-66, issue-547 and issue-579.
1037
1038 * issue-464 (memory corruption in debugalloc's realloc after
1039 memallign) is now fixed
1040
1041 * tcmalloc is now able to release memory back to OS on windows
1042 (issue-489). The code was ported from chromium fork (by a number of
1043 authors).
1044
1045 * Together with issue-489 we ported chromium's "aggressive decommit"
1046 mode. In this mode (settable via malloc extension and via
1047 environment variable TCMALLOC_AGGRESSIVE_DECOMMIT), free pages are
1048 returned back to OS immediately.
1049
1050 * MallocExtension::instance() is now faster (based on patch by
1051 Adhemerval Zanella)
1052
1053 * issue-610 (hangs on windows in multibyte locales) is now fixed
1054
1055 The following people helped with ideas or patches (based on git log,
1056 some contributions purely in bugtracker might be missing): Andrew
1057 C. Morrow, yurivict, Wang YanQing, Thomas Klausner,
1058 [email protected], Dai MIKURUBE, Joon-Sung Um, Jovan
1059 Zelincevic, Jean Lee, Petr Hosek, Ben Avison, drussel, Joonsoo Kim,
1060 Hannes Weisbach, xiaoyur347, Riku Voipio, Adhemerval Zanella, Raphael
1061 Moreira Zinsly
1062
1063 == 30 July 2013 ==
1064
1065 gperftools 2.1 is out!
1066
1067 Just few fixes where merged after rc. Most notably:
1068
1069 * Some fixes for debug allocation on POWER/Linux
1070
1071 == 20 July 2013 ==
1072
1073 gperftools 2.1rc is out!
1074
1075 As a result of more than a year of contributions we're ready for 2.1
1076 release.
1077
1078 But before making that step I'd like to create RC and make sure people
1079 have chance to test it.
1080
1081 Here are notable changes since 2.0:
1082
1083 * fixes for building on newer platforms. Notably, there's now initial
1084 support for x32 ABI (--enable-minimal only at this time))
1085
1086 * new getNumericProperty stats for cache sizes
1087
1088 * added HEAP_PROFILER_TIME_INTERVAL variable (see documentation)
1089
1090 * added environment variable to control heap size (TCMALLOC_HEAP_LIMIT_MB)
1091
1092 * added environment variable to disable release of memory back to OS
1093 (TCMALLOC_DISABLE_MEMORY_RELEASE)
1094
1095 * cpu profiler can now be switched on and off by sending it a signal
1096 (specified in CPUPROFILESIGNAL)
1097
1098 * (issue 491) fixed race-ful spinlock wake-ups
1099
1100 * (issue 496) added some support for fork-ing of process that is using
1101 tcmalloc
1102
1103 * (issue 368) improved memory fragmentation when large chunks of
1104 memory are allocated/freed
1105
1106 == 03 February 2012 ==
1107
1108 I've just released gperftools 2.0
1109
1110 The `google-perftools` project has been renamed to `gperftools`. I
1111 (csilvers) am stepping down as maintainer, to be replaced by
1112 David Chappelle. Welcome to the team, David! David has been an
1113 an active contributor to perftools in the past -- in fact, he's the
1114 only person other than me that already has commit status. I am
1115 pleased to have him take over as maintainer.
1116
1117 I have both renamed the project (the Google Code site renamed a few
1118 weeks ago), and bumped the major version number up to 2, to reflect
1119 the new community ownership of the project. Almost all the
1120 [http://gperftools.googlecode.com/svn/tags/gperftools-2.0/ChangeLog changes]
1121 are related to the renaming.
1122
1123 The main functional change from google-perftools 1.10 is that
1124 I've renamed the `google/` include-directory to be `gperftools/`
1125 instead. New code should `#include <gperftools/tcmalloc.h>`/etc.
1126 (Most users of perftools don't need any perftools-specific includes at
1127 all, so this is mostly directed to "power users.") I've kept the old
1128 names around as forwarding headers to the new, so `#include
1129 <google/tcmalloc.h>` will continue to work.
1130
1131 (The other functional change which I snuck in is getting rid of some
1132 bash-isms in one of the unittest driver scripts, so it could run on
1133 Solaris.)
1134
1135 Note that some internal names still contain the text `google`, such as
1136 the `google_malloc` internal linker section. I think that's a
1137 trickier transition, and can happen in a future release (if at all).
1138
1139
1140 === 31 January 2012 ===
1141
1142 I've just released perftools 1.10
1143
1144 There is an API-incompatible change: several of the methods in the
1145 `MallocExtension` class have changed from taking a `void*` to taking a
1146 `const void*`. You should not be affected by this API change
1147 unless you've written your own custom malloc extension that derives
1148 from `MallocExtension`, but since it is a user-visible change, I have
1149 upped the `.so` version number for this release.
1150
1151 This release focuses on improvements to linux-syscall-support.h,
1152 including ARM and PPC fixups and general cleanups. I hope this will
1153 magically fix an array of bugs people have been seeing.
1154
1155 There is also exciting news on the porting front, with support for
1156 patching win64 assembly contributed by IBM Canada! This is an
1157 important step -- perhaps the most difficult -- to getting perftools
1158 to work on 64-bit windows using the patching technique (it doesn't
1159 affect the libc-modification technique). `premable_patcher_test` has
1160 been added to help test these changes; it is meant to compile under
1161 x86_64, and won't work under win32.
1162
1163 For the full list of changes, including improved `HEAP_PROFILE_MMAP`
1164 support, see the
1165 [http://gperftools.googlecode.com/svn/tags/google-perftools-1.10/ChangeLog ChangeLog].
1166
1167
1168 === 24 January 2011 ===
1169
1170 The `google-perftools` Google Code page has been renamed to
1171 `gperftools`, in preparation for the project being renamed to
1172 `gperftools`. In the coming weeks, I'll be stepping down as
1173 maintainer for the perftools project, and as part of that Google is
1174 relinquishing ownership of the project; it will now be entirely
1175 community run. The name change reflects that shift. The 'g' in
1176 'gperftools' stands for 'great'. :-)
1177
1178 === 23 December 2011 ===
1179
1180 I've just released perftools 1.9.1
1181
1182 I missed including a file in the tarball, that is needed to compile on
1183 ARM. If you are not compiling on ARM, or have successfully compiled
1184 perftools 1.9, there is no need to upgrade.
1185
1186
1187 === 22 December 2011 ===
1188
1189 I've just released perftools 1.9
1190
1191 This change has a slew of improvements, from better ARM and freebsd
1192 support, to improved performance by moving some code outside of locks,
1193 to better pprof reporting of code with overloaded functions.
1194
1195 The full list of changes is in the
1196 [http://google-perftools.googlecode.com/svn/tags/google-perftools-1.9/ChangeLog ChangeLog].
1197
1198
1199 === 26 August 2011 ===
1200
1201 I've just released perftools 1.8.3
1202
1203 The star-crossed 1.8 series continues; in 1.8.1, I had accidentally
1204 removed some code that was needed for FreeBSD. (Without this code
1205 many apps would crash at startup.) This release re-adds that code.
1206 If you are not on FreeBSD, or are using FreeBSD with perftools 1.8 or
1207 earlier, there is no need to upgrade.
1208
1209 === 11 August 2011 ===
1210
1211 I've just released perftools 1.8.2
1212
1213 I was incorrectly calculating the patch-level in the configuration
1214 step, meaning the TC_VERSION_PATCH #define in tcmalloc.h was wrong.
1215 Since the testing framework checks for this, it was failing. Now it
1216 should work again. This time, I was careful to re-run my tests after
1217 upping the version number. :-)
1218
1219 If you don't care about the TC_VERSION_PATCH #define, there's no
1220 reason to upgrae.
1221
1222 === 26 July 2011 ===
1223
1224 I've just released perftools 1.8.1
1225
1226 I was missing an #include that caused the build to break under some
1227 compilers, especially newer gcc's, that wanted it. This only affects
1228 people who build from source, so only the .tar.gz file is updated from
1229 perftools 1.8. If you didn't have any problems compiling perftools
1230 1.8, there's no reason to upgrade.
1231
1232 === 15 July 2011 ===
1233
1234 I've just released perftools 1.8
1235
1236 Of the many changes in this release, a good number pertain to porting.
1237 I've revamped OS X support to use the malloc-zone framework; it should
1238 now Just Work to link in tcmalloc, without needing
1239 `DYLD_FORCE_FLAT_NAMESPACE` or the like. (This is a pretty major
1240 change, so please feel free to report feedback at
1241 [email protected].) 64-bit Windows support is also
1242 improved, as is ARM support, and the hooks are in place to improve
1243 FreeBSD support as well.
1244
1245 On the other hand, I'm seeing hanging tests on Cygwin. I see the same
1246 hanging even with (the old) perftools 1.7, so I'm guessing this is
1247 either a problem specific to my Cygwin installation, or nobody is
1248 trying to use perftools under Cygwin. If you can reproduce the
1249 problem, and even better have a solution, you can report it at
1250 [email protected].
1251
1252 Internal changes include several performance and space-saving tweaks.
1253 One is user-visible (but in "stealth mode", and otherwise
1254 undocumented): you can compile with `-DTCMALLOC_SMALL_BUT_SLOW`. In
1255 this mode, tcmalloc will use less memory overhead, at the cost of
1256 running (likely not noticeably) slower.
1257
1258 There are many other changes as well, too numerous to recount here,
1259 but present in the
1260 [http://google-perftools.googlecode.com/svn/tags/google-perftools-1.8/ChangeLog ChangeLog].
1261
1262
1263 === 7 February 2011 ===
1264
1265 Thanks to endlessr..., who
1266 [http://code.google.com/p/google-perftools/issues/detail?id=307 identified]
1267 why some tests were failing under MSVC 10 in release mode. It does not look
1268 like these failures point toward any problem with tcmalloc itself; rather, the
1269 problem is with the test, which made some assumptions that broke under the
1270 some aggressive optimizations used in MSVC 10. I'll fix the test, but in
1271 the meantime, feel free to use perftools even when compiled under MSVC
1272 10.
1273
1274 === 4 February 2011 ===
1275
1276 I've just released perftools 1.7
1277
1278 I apologize for the delay since the last release; so many great new
1279 patches and bugfixes kept coming in (and are still coming in; I also
1280 apologize to those folks who have to slip until the next release). I
1281 picked this arbitrary time to make a cut.
1282
1283 Among the many new features in this release is a multi-megabyte
1284 reduction in the amount of tcmalloc overhead uder x86_64, improved
1285 performance in the case of contention, and many many bugfixes,
1286 especially architecture-specific bugfixes. See the
1287 [http://google-perftools.googlecode.com/svn/tags/google-perftools-1.7/ChangeLog ChangeLog]
1288 for full details.
1289
1290 One architecture-specific change of note is added comments in the
1291 [http://google-perftools.googlecode.com/svn/tags/perftools-1.7/README README]
1292 for using tcmalloc under OS X. I'm trying to get my head around the
1293 exact behavior of the OS X linker, and hope to have more improvements
1294 for the next release, but I hope these notes help folks who have been
1295 having trouble with tcmalloc on OS X.
1296
1297 *Windows users*: I've heard reports that some unittests fail on
1298 Windows when compiled with MSVC 10 in Release mode. All tests pass in
1299 Debug mode. I've not heard of any problems with earlier versions of
1300 MSVC. I don't know if this is a problem with the runtime patching (so
1301 the static patching discussed in README_windows.txt will still work),
1302 a problem with perftools more generally, or a bug in MSVC 10. Anyone
1303 with windows expertise that can debug this, I'd be glad to hear from!
1304
1305
1306 === 5 August 2010 ===
1307
1308 I've just released perftools 1.6
1309
1310 This version also has a large number of minor changes, including
1311 support for `malloc_usable_size()` as a glibc-compatible alias to
1312 `malloc_size()`, the addition of SVG-based output to `pprof`, and
1313 experimental support for tcmalloc large pages, which may speed up
1314 tcmalloc at the cost of greater memory use. To use tcmalloc large
1315 pages, see the
1316 [http://google-perftools.googlecode.com/svn/tags/perftools-1.6/INSTALL
1317 INSTALL file]; for all changes, see the
1318 [http://google-perftools.googlecode.com/svn/tags/perftools-1.6/ChangeLog
1319 ChangeLog].
1320
1321 OS X NOTE: improvements in the profiler unittest have turned up an OS
1322 X issue: in multithreaded programs, it seems that OS X often delivers
1323 the profiling signal (from sigitimer()) to the main thread, even when
1324 it's sleeping, rather than spawned threads that are doing actual work.
1325 If anyone knows details of how OS X handles SIGPROF events (from
1326 setitimer) in threaded programs, and has insight into this problem,
1327 please send mail to [email protected].
1328
1329 To see if you're affected by this, look for profiling time that pprof
1330 attributes to `___semwait_signal`. This is work being done in other
1331 threads, that is being attributed to sleeping-time in the main thread.
1332
1333
1334 === 20 January 2010 ===
1335
1336 I've just released perftools 1.5
1337
1338 This version has a slew of changes, leading to somewhat faster
1339 performance and improvements in portability. It adds features like
1340 `ITIMER_REAL` support to the cpu profiler, and `tc_set_new_mode` to
1341 mimic the windows function of the same name. Full details are in the
1342 [http://google-perftools.googlecode.com/svn/tags/perftools-1.5/ChangeLog
1343 ChangeLog].
1344
1345
1346 === 11 September 2009 ===
1347
1348 I've just released perftools 1.4
1349
1350 The major change this release is the addition of a debugging malloc
1351 library! If you link with `libtcmalloc_debug.so` instead of
1352 `libtcmalloc.so` (and likewise for the `minimal` variants) you'll get
1353 a debugging malloc, which will catch double-frees, writes to freed
1354 data, `free`/`delete` and `delete`/`delete[]` mismatches, and even
1355 (optionally) writes past the end of an allocated block.
1356
1357 We plan to do more with this library in the future, including
1358 supporting it on Windows, and adding the ability to use the debugging
1359 library with your default malloc in addition to using it with
1360 tcmalloc.
1361
1362 There are also the usual complement of bug fixes, documented in the
1363 ChangeLog, and a few minor user-tunable knobs added to components like
1364 the system allocator.
1365
1366
1367 === 9 June 2009 ===
1368
1369 I've just released perftools 1.3
1370
1371 Like 1.2, this has a variety of bug fixes, especially related to the
1372 Windows build. One of my bugfixes is to undo the weird `ld -r` fix to
1373 `.a` files that I introduced in perftools 1.2: it caused problems on
1374 too many platforms. I've reverted back to normal `.a` files. To work
1375 around the original problem that prompted the `ld -r` fix, I now
1376 provide `libtcmalloc_and_profiler.a`, for folks who want to link in
1377 both.
1378
1379 The most interesting API change is that I now not only override
1380 `malloc`/`free`/etc, I also expose them via a unique set of symbols:
1381 `tc_malloc`/`tc_free`/etc. This enables clients to write their own
1382 memory wrappers that use tcmalloc:
1383 {{{
1384 void* malloc(size_t size) { void* r = tc_malloc(size); Log(r); return r; }
1385 }}}
1386
1387
1388 === 17 April 2009 ===
1389
1390 I've just released perftools 1.2.
1391
1392 This is mostly a bugfix release. The major change is internal: I have
1393 a new system for creating packages, which allows me to create 64-bit
1394 packages. (I still don't do that for perftools, because there is
1395 still no great 64-bit solution, with libunwind still giving problems
1396 and --disable-frame-pointers not practical in every environment.)
1397
1398 Another interesting change involves Windows: a
1399 [http://code.google.com/p/google-perftools/issues/detail?id=126 new
1400 patch] allows users to choose to override malloc/free/etc on Windows
1401 rather than patching, as is done now. This can be used to create
1402 custom CRTs.
1403
1404 My fix for this
1405 [http://groups.google.com/group/google-perftools/browse_thread/thread/1ff9b50043090d9d/a59210c4206f2060?lnk=gst&q=dynamic#a59210c4206f2060
1406 bug involving static linking] ended up being to make libtcmalloc.a and
1407 libperftools.a a big .o file, rather than a true `ar` archive. This
1408 should not yield any problems in practice -- in fact, it should be
1409 better, since the heap profiler, leak checker, and cpu profiler will
1410 now all work even with the static libraries -- but if you find it
1411 does, please file a bug report.
1412
1413 Finally, the profile_handler_unittest provided in the perftools
1414 testsuite (new in this release) is failing on FreeBSD. The end-to-end
1415 test that uses the profile-handler is passing, so I suspect the
1416 problem may be with the test, not the perftools code itself. However,
1417 I do not know enough about how itimers work on FreeBSD to be able to
1418 debug it. If you can figure it out, please let me know!
1419
1420 === 11 March 2009 ===
1421
1422 I've just released perftools 1.1!
1423
1424 It has many changes since perftools 1.0 including
1425
1426 * Faster performance due to dynamically sized thread caches
1427 * Better heap-sampling for more realistic profiles
1428 * Improved support on Windows (MSVC 7.1 and cygwin)
1429 * Better stacktraces in linux (using VDSO)
1430 * Many bug fixes and feature requests
1431
1432 Note: if you use the CPU-profiler with applications that fork without
1433 doing an exec right afterwards, please see the README. Recent testing
1434 has shown that profiles are unreliable in that case. The problem has
1435 existed since the first release of perftools. We expect to have a fix
1436 for perftools 1.2. For more details, see
1437 [http://code.google.com/p/google-perftools/issues/detail?id=105 issue 105].
1438
1439 Everyone who uses perftools 1.0 is encouraged to upgrade to perftools
1440 1.1. If you see any problems with the new release, please file a bug
1441 report at http://code.google.com/p/google-perftools/issues/list.
1442
1443 Enjoy!