Untouched open Tcl bug tickets: Not reproduced

Triage of 818 open and Pending Bug tickets of Tcl with no comment or check-in from me on 2026-10-16 (in Tcl's workflow Pending means "resolved, waiting for confirmation"). Reproduced on Linux (main 9.1 tip, 9.0 tip, 8.7b1, 8.6 tip, 8.5, 8.4 and old releases), on FreeBSD, NetBSD, OpenBSD and OpenIndiana VMs for platform tickets, and on a Windows 11 VM. Rows marked patch link to a prototype patch (287 in all, in tcl-triage/prototypes). Updated 2026-11-11: the emoji before each ticket id shows its progress.

Not reproduced 74

⬜bb9ca2fa952026(chan)io-29.26 fails on Ubuntu 26.4NOT-REPRODUCED (main, 9.0, 8.6 on Ubuntu 26.04.1 locally)

Open ticket bb9ca2fa95 · 25. Channel System · filed 2026-08-25

Full verdict

NOT-REPRODUCED (main, 9.0, 8.6 on Ubuntu 26.04.1 locally)

  • Platform unix (CI only)
  • Tested io-29.26 with -constraints notInCIenv on own main build; the test body in a loop (300 iterations, 8 in parallel, with stdin/stdout/stderr closed) on main, 9.0.5, 8.6.18. Local cat is uutils 0.10.0 (rust-coreutils 0.10.0-1ubuntu2~26.04.1), which uses splice(); no SIGPIPE in any run.
  • Cause unknown. The child cat gets SIGPIPE, so it writes after Tcl closed the read end. That points at the cat on the GitHub runner image (maybe an older uutils), not at Tcl.
  • Prototype none (not reproducible here; the test is already skipped in CI with notInCIenv)
  • Suggested ticket comment Cannot reproduce on Ubuntu 26.04.1 with uutils cat 0.10.0 (300 runs, also in parallel and with closed std channels) on main, 9.0 and 8.6. The CI runner's cat may behave differently; the notInCIenv constraint looks sufficient, or the test could use gnucat / a tclsh child when available.
⬜c0573789862026exec-19.1 test failure on Ubuntu (WSL)patchNOT-REPRODUCED (main, 9.0, 8.6 on native Linux; 30/30 runs give 26, even with 30 copies in parallel). Test bug, not a Tcl bug

Open ticket c057378986 · - New Builtin Commands · filed 2026-06-13

🧪 Prototype check (2026-10-24)

c057378986.patch: VERIFIED

  • Base main (tip d6dd6d0f71)
  • Before exec-19.1 passes on native Linux (no WSL here). Simulated slow process creation: work-a5/ex19.tcl, a copy of the test with one child prefixed by sleep 2.5 -> old logic (after 5000): size=24 (expect 26), i.e. the WSL/macOS failure
  • Patch applied cleanly (13 changed lines, test only)
  • After same simulation with the new polling logic: size=26 ... after 6039 ms; make test -file exec.test 98 passed / 0 failed, exec-19.1 PASSED
  • Test the patch changes exec-19.1 itself; the old version fails and the new one passes under the delayed-start simulation (not on WSL itself)
Full verdict

NOT-REPRODUCED (main, 9.0, 8.6 on native Linux; 30/30 runs give 26, even with 30 copies in parallel). Test bug, not a Tcl bug

  • Platform unix (WSL)
  • Tested standalone copy of the test, 10 runs each on main (own build), 9.0.5, 8.6.18 at the same time; strace on main
  • Cause Tcl does its part. strace shows all four >>/2>> redirections opened with O_WRONLY|O_CREAT|O_APPEND (generic/tclPipe.c). The test reads the file size after a fixed after 5000, and 4 children that each run 3 × sleep 1 finish late when process creation is slow (WSL on an NTFS mount, macOS). If one child starts 2.5 s late, the old test gives 24, which is the macOS result in 538ba1e615. I could not check whether drvfs/NTFS appends are really atomic; that needs WSL.
  • Prototype prototypes/c057378986.patch (13 changed lines, test exec-19.1 itself): poll until the file reaches 26 bytes or 30 s pass, then wait another 0.5 s. The test now passes in ~4 s instead of 5 s, and passes with the 2.5 s delayed start. exec.test passes.
  • Suggested ticket comment Tcl does open the file with O_APPEND (checked with strace). The test reads the size after a fixed 5 s, which is too short when process startup is slow. Proposed fix: make exec-19.1 wait until the output arrives (30 s timeout) instead of a fixed after 5000. If it still fails on WSL with that change, then appends to the NTFS mount are really being lost.
⬜081338a2de2024Tcl_SetAssocData with Tcl_InterpDeleteProc segfaults on exitNOT-REPRODUCED (main, 9.0.5, 8.6.18); no reporter script, needs info

Open ticket 081338a2de · 41. Memory Allocation · filed 2024-10-22

Full verdict

NOT-REPRODUCED (main, 9.0.5, 8.6.18); no reporter script, needs info

  • Platform generic
  • Tested loadable extension (work-a6/assoc/assoc.c) setting assoc data with a delProc that frees data and calls Tcl_GetAssocData / Tcl_DeleteAssocData, loaded into main and child interps, then interp delete + exit (also TCL_FINALIZE_ON_EXIT=1), under valgrind: clean on all three. Embedding program (emb.c) doing Tcl_DeleteInterp+Tcl_Finalize: clean. Tcl_Finalize before Tcl_DeleteInterp, or Tcl_Release after Tcl_Finalize, panics, but in 8.6 as well (API misuse; 9.x shows SIGILL because its panic uses a trap where 8.6 calls abort).
  • Cause unknown. The double free from 34870ab575 (Tcl_DeleteAssocData inside a delProc) was fixed before 9.0.0 (0aedd271b2, 2024-05-30), so it is not the cause for 9.0.1.
  • Prototype none
  • Suggested ticket comment Cannot reproduce on 9.0.5 or main: delProcs registered with Tcl_SetAssocData run cleanly (valgrind) on interp delete, on exit and with Tcl_DeleteInterp+Tcl_Finalize in an embedding app. Calling Tcl_DeleteInterp or Tcl_Release after Tcl_Finalize panics, but 8.6 does too. Can you post a minimal program and a backtrace? Otherwise we propose to close this.
⬜718de2132f2023review use of strip on the stubs libraryNOT-REPRODUCED (main, 8.6 on FreeBSD 15.1 with base llvm-strip 19.1.7; Linux LLVM 21 earlier)

Open ticket 718de2132f · 53. Configuration and Build Tools · filed 2023-04-13

Full verdict

NOT-REPRODUCED (main, 8.6 on FreeBSD 15.1 with base llvm-strip 19.1.7; Linux LLVM 21 earlier)

  • Platform build (FreeBSD/LLVM toolchain)
  • Tested stub objects of main (tclStubLib, tclStubLibTbl, tclTomMathStubLib, tclOOStubLib) compiled with clang 21 -O2 -fPIC (they do contain 7 .L local symbols referenced by relocations), archived with llvm-ar-21; llvm-strip-21 -x, llvm-strip-21 --strip-unneeded and GNU strip -x all succeed, and llvm-strip -x keeps the referenced .L symbols. So current LLVM no longer errors out; the failure was LLVM 15 on FreeBSD 14 (2023). Main still has INSTALL_STRIP_LIBRARY = strip -x (unix/Makefile.in:154); the --strip-unneeded change was reverted (82baf7827f).
  • Cause (was) older llvm-strip refusing to strip symbols referenced by relocations with -x.
  • Prototype none
  • Suggested ticket comment With LLVM 21 llvm-strip -x on libtclstub.a works and keeps the relocation-referenced .L symbols, so this looks like an old llvm-strip issue. To confirm on FreeBSD: make install-strip with the base llvm-strip; if it passes, this can be closed.
  • To run on the FreeBSD VM build main with clang, make install-strip INSTALL_ROOT=/tmp/x and check it gets past installing libtclstub.a.
Unix VMs, run 1
  • Platform build (FreeBSD/LLVM toolchain)
  • Verdict NOT-REPRODUCED (main, 8.6 on FreeBSD 15.1 with base llvm-strip 19.1.7; Linux LLVM 21 earlier)
  • Tested FreeBSD 15.1-RELEASE-p1 amd64, clang 19.1.7, /usr/bin/strip = llvm-strip 19.1.7 ("compatible with GNU strip"). make install-strip INSTALL_ROOT=/tmp/... succeeds on main and core-8-6-branch (both --enable-symbols and a default -O2 build); strip -x libtclstub.a exits 0 and keeps the relocation-referenced .L symbols (tclStubLib.o has 8 .L relocations on main, 5 on 8.6).
  • Cause (was) llvm-strip 15 refusing -x on local symbols referenced by relocations; newer llvm-strip keeps them like GNU strip.
  • Prototype none
  • Suggested ticket comment On FreeBSD 15.1 with the base llvm-strip 19.1.7, make install-strip works on main and core-8-6-branch, and strip -x keeps the .L symbols referenced by relocations. So this was an llvm-strip 15 issue; suggest closing.
⬜87986108832023TclIsInfinite(NaN) returns true for MSVCpatchNOT-REPRODUCED at script level (8.6.18 MSVC 2022: all four NaN/bignum comparisons give 0; 9.0/main also 0). The macro is still wrong in 8.6 generic…

Open ticket 8798610883 · 52. Portability Support · filed 2023-04-08

🧪 Prototype check (2026-10-24)

8798610883.patch: VERIFIED (macro fix correct, as a latent-bug cleanup) — but the triage's symptom ("$nan < 2**70 gives 1 on MSVC") is wrong and the added test is ineffective; the suggested ticket comment should be corrected (keep or drop the test as a plain regression test)

  • Base core-8-6-branch (tip 326e1bf923)
  • Before MSVC emulated on Linux by enabling the _MSC_VER branch of tclInt.h with -DEMU_MSC -D_finite=__builtin_isfinite -D_isnan=__builtin_isnan (temporary #if defined(_MSC_VER) || defined(EMU_MSC)). C check (work-a2/v2/isinf.c): unpatched macro -> IsInfinite(NaN)=1. BUT no script-visible effect: a battery of 41 NaN operations (work-a2/v2/nan.tcl: NaN vs 2**70 in all six comparisons, format, int/entier/wide/round/sqrt/isqrt, lsort/lsearch -real, binary format, min/max, arithmetic) gives identical output on the emulated-MSVC unpatched build, the emulated patched build and the native build. The comparison the triage named cannot see a NaN: INST_LT etc. in 8.6 tclExecute.c:6153 return early for TCL_NUMBER_NAN before CompareTwoNumbers (its only caller); the other TclIsInfinite sites check TclIsNaN first or reject NaN earlier.
  • Patch applied cleanly (offset 4; 8 changed lines, content unchanged)
  • After emulated build: IsInfinite(NaN)=0, Inf still 1; expr-bug-8798610883 passes
  • Test existing expr-bug-8798610883 does NOT fail before (passes on the emulated MSVC unpatched build too) -> it does not test the bug; none possible (the wrong macro value is not observable from scripts in 8.6). win/8798610883.tcl will likewise print NOT-REPRODUCED on real MSVC builds.
Full verdict

NOT-REPRODUCED at script level (8.6.18 MSVC 2022: all four NaN/bignum comparisons give 0; 9.0/main also 0). The macro is still wrong in 8.6 generic/tclInt.h, but it is not reachable from scripts

  • Platform Windows (MSVC builds of 8.6 only)
  • Tested code reading of 8.6/9.0/main tclInt.h; Linux 8.6.18/9.0.5 give the correct result for the repro; prototype test on 8.6 tip (Linux)
  • Cause 8.6 generic/tclInt.h defines TclIsInfinite(d) as !_finite(d) for _MSC_VER, which is true for NaN. Of the call sites, only TclCompareTwoNumbers (double vs bignum, both directions) lets a NaN through, so $nan < 270 and 270 > $nan should return 1 on MSVC instead of 0. Other sites check TclIsNaN first or reject NaN earlier.
  • Prototype prototypes/8798610883.patch (core-8-6-branch, 8 changed lines, test expr-bug-8798610883); windows repro win/8798610883.tcl
  • Suggested ticket comment In 8.6 only, !_finite() is true for NaN, so MSVC builds compare NaN with a bignum as ordered (TclCompareTwoNumbers). Proposed fix: add && !_isnan(d) to the MSVC TclIsInfinite, with test expr-bug-8798610883. 9.0 uses isinf() and is not affected.
Windows VM, run 1
  • Platform Windows (MSVC builds of 8.6)
  • Verdict NOT-REPRODUCED at script level (8.6.18 MSVC 2022: all four NaN/bignum comparisons give 0; 9.0/main also 0). The macro is still wrong in 8.6 generic/tclInt.h, but it is not reachable from scripts
  • Tested VM 9.1b1, 9.0.5, 8.6.18 with win/8798610883.tcl; code reading of 8.6
  • Cause the Linux analysis was wrong about the comparison path: in 8.6 TEBC (tclExecute.c ~6152) NaN operands of comparisons are handled before CompareTwoNumbers, so the NaN case never reaches TclIsInfinite there. The other call sites check TclIsNaN first or require d >= 0. The only place left is the public Tcl_InitBignumFromDouble(NaN), where MSVC returns "integer value too large to represent" (arguably better than what the non-MSVC path does).
  • Prototype prototypes/8798610883.patch is still a correct cleanup, but its test expr-bug-8798610883 passes without the fix on Windows, so it does not cover anything; drop the test or keep the one-line header fix only
  • Suggested ticket comment In 8.6 the MSVC TclIsInfinite() is still true for NaN, but no script-level path reaches it with a NaN (comparisons handle NaN first), so there is no visible bug; 9.0 uses isinf(). The macro could be fixed with && !_isnan(d) for cleanliness, or the ticket closed.
⬜1f1b1686262023Tcl 9 StringObj.3 - lengthPtr not documentedNOT-REPRODUCED (all tested)

Open ticket 1f1b168626 · 69. Other · filed 2023-01-27

Full verdict

NOT-REPRODUCED (all tested)

  • Platform docs
  • Tested rendered doc/StringObj.3 with man -l (run in doc/, so .so man.macros resolves) for main, the 9.0.0 tag and the 2023-01-27 revision 4fb77727c5. All show Tcl_Size | int *lengthPtr (out) (size_t | int in 2023) with its description. The source has always had .AP ... *lengthPtr out. The reporter's "missing name" came from their older man/groff (man 2.9.4).
  • Cause n/a (renderer). Unrelated typo found nearby: the synopsis in main doc/StringObj.3:86 has \fBTcl_IsEmpty\fR(\fIfIobjPtr\fR), which should be \fIobjPtr.
  • Prototype none
  • Suggested ticket comment lengthPtr is documented (.AP "Tcl_Size \&| int" *lengthPtr out), and current groff renders it with the name. The missing name was a rendering problem in the old man version. Closing. (Separately, the Tcl_IsEmpty synopsis line has a stray "fI".)
⬜a35d210bfc2022exec:ed process killed with parentNOT-REPRODUCED (main, 9.0, 8.6)

Open ticket a35d210bfc · 16. Commands A-H · filed 2022-11-30

Full verdict

NOT-REPRODUCED (main, 9.0, 8.6)

  • Platform Windows
  • Tested Linux 9.0.5 and 8.6.18 with win/a35d210bfc.tcl. It builds a chain of 4 generations, each doing exec [info nameofexecutable] ... & and then exiting, and the last one must still be alive 3 s later. Both the tclsh and the Tk variant pass. (The Tk variant ran with withdrawn windows on the inherited DISPLAY=:0 rather than Xvfb, by mistake.)
  • Cause unknown. On Windows, look at win/tclWinPipe.c:TclpCreateProcess (creation flags and console handling when the parent is a GUI/no-console process: the second generation is a wish started from wish).
  • Prototype none (Windows-only)
  • Suggested ticket comment Not reproducible on Linux. Repro for Windows (tclsh and Tk variants, prints a verdict line): win/a35d210bfc.tcl. Run it with both tclsh and wish from 9.0 and 8.6.
Windows VM, run 3
  • Platform Windows
  • Verdict NOT-REPRODUCED (main, 9.0, 8.6)
  • Tested VM, console: win/a35d210bfc.tcl tclsh variant (4 generations of exec ... & then exit, the last must be alive 3 s later) OK on 9.1b1, 9.0.5, 8.6.18. Interactive session (scheduled task, wish started hidden by wscript, all windows withdrawn; work-win3/a35w.tcl): chains of 4 wish generations, each started with exec [info nameofexecutable] ... & from a wish that then exits with exit 0 or destroy .: all generations start and the last survives its parent, with wish 9.1b1, 9.0.5 and 8.6.18.
  • Cause unknown; Tcl creates no job object, so the kill-with-parent most likely came from the reporter's environment (e.g. a launcher/terminal job with kill-on-close)
  • Prototype none
  • Suggested ticket comment Not reproducible on Windows 11 with wish 8.6.18, 9.0.5 and 9.1: a chain of four wish processes, each restarting itself with &#91;exec ... &&#93; and then &#91;destroy .&#93;, keeps the newest one alive. How was the first wish started (Explorer, cmd, Windows Terminal, an IDE)? A job object of the launcher could explain it.
⬜c5348d22cf2022font-4.14/font-4.16 fail in Tk 8.7 on WindowsNOT-REPRODUCED on Tk 9.1b1 / 9.0.5 (both tests would pass); REPRODUCED for font-4.16 on Tk 8.6.18 (test expectation, NOT-A-BUG in Tcl)

Open ticket c5348d22cf · 44. UTF-8 Strings · filed 2022-03-25

Full verdict

NOT-REPRODUCED on Tk 9.1b1 / 9.0.5 (both tests would pass); REPRODUCED for font-4.16 on Tk 8.6.18 (test expectation, NOT-A-BUG in Tcl)

  • Platform Windows (Tk test suite)
  • Tested Linux/X11 (Xvfb :71) with wish 9.1b1, 9.0.5, 8.6.18: font actual {-family times} -family -- \U10000 returns the fallback font used for that char ("Noto Sans Linear B"), not the base family ("Nimbus Roman"). In Tk main and 9.0 tests/font.test, font-4.14 was rewritten (now expects an error for -- 𐀀, no constraint) and font-4.16 is still {win knownBug} expecting "times"; in Tk 8.6 both are win tests expecting "Times New Roman".
  • Cause likely the test expectation: "font actual ... char" reports the font that actually renders the char, which for U+10000 is a fallback font on Windows too (Tcl 8.7+/9 handle U+10000 as one char, so the fallback search now happens).
  • Prototype none; repro script win/c5348d22cf.tcl (prints the actual family and a VERDICT line).
  • Suggested ticket comment On X11, font actual for U+10000 returns the fallback font ("Noto Sans Linear B"), which is correct behaviour; font-4.16 probably fails on Windows for the same reason (expectation, not Tcl). font-4.14 has been rewritten in the meantime. To be confirmed on Windows.
Windows VM, run 1
  • Platform Windows (Tk test suite)
  • Verdict NOT-REPRODUCED on Tk 9.1b1 / 9.0.5 (both tests would pass); REPRODUCED for font-4.16 on Tk 8.6.18 (test expectation, NOT-A-BUG in Tcl)
  • Tested VM interactive desktop, wish of tk-ref91/tk-ref90/tk-ref86 running the test bodies directly (the tktest runs of font.test showed error dialogs on the desktop and were killed). 9.x: font actual {-family times} -- 𐀀 gives the expected error "expected a single character" (font-4.14 passes); font actual {-family times} -family -- \U10000 returns "times", which is font-4.16's expected result, but the test is still constrained {win knownBug}. 8.6: font-4.14 (-family -- 𐀀) gives "Times New Roman" (passes); font-4.16 gives "Malgun Gothic", not "Times New Roman" (fails).
  • Cause 8.6 font-4.16: font actual ... char reports the fallback font that renders U+10000, so the expectation is wrong. On 9.x Windows, font actual reports the requested name ("times", not "Times New Roman") and gives no fallback for U+10000.
  • Prototype none (Tk test changes: drop knownBug from font-4.16 in Tk 9.x; in 8.6 font-4.16 change the expected result or mark it knownBug)
  • Suggested ticket comment With current Tk 9.0/9.1 on Windows 11, font-4.14 has been rewritten and passes, and the font-4.16 body returns "times" as expected, so its knownBug constraint can go. In 8.6, font-4.16 fails because font actual reports the fallback font for U+10000 ("Malgun Gothic"), so the expected result is wrong, not Tk.
⬜95061fc7462021Line number discrepencyNOT-REPRODUCED (all tested)

Open ticket 95061fc746 · 17. Commands I-L · filed 2021-06-02

Full verdict

NOT-REPRODUCED (all tested)

  • Platform generic
  • Tested main (own build), 8.6.18, 8.6.11 — info frame -1 line for the command after a braced argument with backslash-newlines. Contexts tried: top-level if, if in a proc, foreach, $cmd (non-compiled if), switch, catch, if/else, if in while, namespace eval, uplevel #0, interp alias, and an execution trace on the first command. The second command always gets the right line (3 lines after the first). This matches sebres' 2021 result. The reporter said it is probably caused by their application's modified Tcl package.
  • Cause unknown; the reporter could not give a plain-Tcl repro
  • Prototype none
  • Suggested ticket comment No plain-Tcl repro in 2021, and none now: in all the contexts tried on 8.6.18 and 9.1, the line after a braced argument with continuation lines is correct. Suggest closing unless a self-contained script turns up.
⬜a8507260182020epoll detection in cross-compile environmentsNOT-REPRODUCED (all tested)

Open ticket a850726018 · 01. Notifier · filed 2020-06-19

Full verdict

NOT-REPRODUCED (all tested)

  • Platform build
  • Tested code reading of unix/configure.ac on main, 9.0, 8.6-era and the 8.7a3 tag (identical logic); out-of-tree configure of main with ac_cv_header_sys_epoll_h=no.
  • Cause configure.ac picks the notifier family with uname -s (build machine), but NOTIFIER_EPOLL is only defined if AC_CHECK_HEADERS([sys/epoll.h]) succeeds with the configured (cross) compiler, so a sysroot without sys/epoll.h gets no NOTIFIER_EPOLL (also true in 8.7a3); the reported failure needs a sysroot that has the header but not the functions, or a stale config cache. The remaining real issue, uname -s instead of the host OS, applies to all of Tcl's unix configure (tcl.m4 SC_CONFIG_SYSTEM too) and has no small fix. A clean override exists: ac_cv_header_sys_epoll_h=no ./configure -> "checking for sys/epoll.h... (cached) no", select notifier used.
  • Prototype none (no failure without a mismatched sysroot; host detection is a configure-wide redesign)
  • Suggested ticket comment NOTIFIER_EPOLL is only defined when sys/epoll.h is found by the configured compiler, so a sysroot lacking it should not get epoll. To force the select notifier, run configure with ac_cv_header_sys_epoll_h=no (no need to edit configure). Can you show config.log if it still happens?
⬜60725611c92019MS VC: error during install of thread packageNOT-REPRODUCED (8.6 tip + Thread 2.8.13, MSVC 2022, x86 and x64)

Open ticket 60725611c9 · - New Builtin Commands · filed 2019-12-23

Full verdict

NOT-REPRODUCED (8.6 tip + Thread 2.8.13, MSVC 2022, x86 and x64)

  • Platform Windows
  • Tested not on Linux (nmake/rc only). win/makefile.vc install-pkgs (8.6 tip) re-runs nmake -f makefile.vc install in each pkgs\*\win, which rebuilds thread.res; RC1109 means rc could not create the output file (e.g. intermediate dir missing or file locked). Nothing in the Tcl makefiles points to it.
  • Cause unknown
  • Prototype none
  • Repro win/60725611c9.cmd (release + install with a forward-slash INSTALLDIR like the reporter; needs a source tree with pkgs\thread*)
  • Suggested ticket comment Can you still reproduce this with 8.6.18 / 9.0.4? RC1109 means rc.exe could not create the .res file; please check whether the Release_VC* intermediate directory exists at that point.
Windows VM, run 3
  • Platform build (Windows, nmake)
  • Verdict NOT-REPRODUCED (8.6 tip + Thread 2.8.13, MSVC 2022, x86 and x64)
  • Tested VM: copies of tcl-wt86 (core-8-6-branch 5d910c9635) with thread-2-8-branch exported as pkgs\thread2.8.13, then the reporter's steps nmake -f makefile.vc INSTALLDIR=%USERPROFILE%/tmp/tri3/inst6072-<arch> release and ... install (drive-less forward-slash INSTALLDIR) from vcvarsall x86 and x64 (work-win3/r6072.cmd): both succeed; install of pkgs\thread2.8.13 does not even re-run rc (thread.res was built by release). Copies deleted afterwards. 9.0 not tried.
  • Cause unknown; RC1109 means rc.exe could not create the .res file (in 8.6.10 the install pass recompiled the resource, so a missing or locked intermediate directory, e.g. after the reporter's distclean/clean, is the likely trigger)
  • Prototype none
  • Suggested ticket comment Not reproducible with the current 8.6 branch and Thread 2.8.13 with MSVC 2022 (x86 and x64): release followed by install with a drive-less INSTALLDIR works and the install pass no longer rebuilds thread.res. Please reopen if it still happens with 8.6.18.
✅016be131a22019Build test socket.test failedNOT-REPRODUCED (main, 9.0, 8.6)

Open ticket 016be131a2 · 53. Configuration and Build Tools · filed 2019-11-27

✅ Since triage (updated 2026-11-11)

Status now: Closed (Out of Date)

  • 💬 2026-10-20 serhiy.storchaka: Not reproducible any more. socket-14.11.0 and socket-14.11.1 passed 40 out of 40 runs each with MSVC 2022 and MSYS2 UCRT64 builds of main (9.1) on Windows 11 outside CI (in CI, 14.11.0 is skipped by the notWinCI constraint), run from an MSYS2 bash; 9.0.5 and 8.6.18 passed too, and close always reports the connect error. Closing as out of date; please reopen if it still fails in your setup.
Full verdict

NOT-REPRODUCED (main, 9.0, 8.6)

  • Platform Windows (MSVC build run from a Cygwin console, 9.0a1)
  • Tested Linux main socket-14.11.* pass (5 runs); repro script on Linux 9.0.5/8.6.18: close always reports the connect error (10/10)
  • Cause unknown; on that Windows setup the failed async connect was not reported by [close]. Since 7e401c5f6b (2020-11-17, dkf, "Test fails with Windows in CI, not MSVC") socket-14.11.0 has constraint notWinCI, so Windows CI never runs it — the failure may still exist. Related: de232b49f2, and 1b9bd55b7a/c8daed39ac (2024, pooryorick, "sporadic failure of 14.11.0", backed out to mistake).
  • Prototype none (needs Windows); repro win/016be131a2.tcl
  • Suggested ticket comment socket-14.11.0 is skipped in Windows CI (notWinCI) since 2020, so this is unverified there. Needs a recheck with current 9.0/9.1 on Windows outside CI.
Windows VM, run 1
  • Platform Windows
  • Verdict NOT-REPRODUCED (main, 9.0, 8.6)
  • Tested VM repro script 10 runs each on 9.1b1, 9.0.5, 8.6.18 (close always reports "transport endpoint is not connected" / 8.6 "socket is not connected"); socket.test -match socket-14.11* with tcltest91 (14.11.0 and 14.11.1 pass; notWinCI is true outside CI) and tcltest86
  • Cause unknown (9.0a1 era, Cygwin console)
  • Prototype none
  • Suggested ticket comment socket-14.11.0 passes with current 9.1, 9.0 and 8.6 builds on Windows 11 outside CI, and close always reports the connect error. Closing as out of date; reopen if it still fails in your setup.
✅00c62f67772019Build test env-4.3, env-4.4 and env-4.5 failedNOT-REPRODUCED (main, 9.0, 8.6)

Open ticket 00c62f6777 · 53. Configuration and Build Tools · filed 2019-11-27

✅ Since triage (updated 2026-11-11)

Status now: Closed (Out of Date)

  • 💬 2026-10-20 serhiy.storchaka: Not reproducible any more. env-4.3 to env-4.5 (and all of env.test) pass on Windows 11 with MSVC 2022 and MSYS2 UCRT64 builds of main (9.1), run from an MSYS2 bash with and without LANG=C.UTF-8; with 9.0.5 and 8.6.18 too. OPTS=utfmax no longer exists: Tcl 9 always uses 4-byte UTF-8. Closing as out of date; please reopen if it still fails in your setup.
Full verdict

NOT-REPRODUCED (main, 9.0, 8.6)

  • Platform Windows (MSVC 2019 x64, 9.0a1, from a Cygwin console with LANG=C.UTF-8; also with OPTS=utfmax)
  • Tested Linux main env.test passes; repro on Linux 9.0.5, 8.6.18 (LANG=C and C.UTF-8) passes
  • Cause unknown; child tclsh printed nothing (exec failed or printenv script errored) in that setup
  • Prototype none; repro win/00c62f6777.tcl (run also with LANG=C.UTF-8 set)
  • Suggested ticket comment Cannot reproduce on Linux; env.test passes in Windows CI with 9.0 (utfmax is the default now). Please retest with 9.0.x; otherwise suggest closing.
Windows VM, run 1
  • Platform Windows
  • Verdict NOT-REPRODUCED (main, 9.0, 8.6)
  • Tested VM repro with and without LANG=C.UTF-8 on 9.1b1, 9.0.5, 8.6.18; env.test passes completely with tcltest91 and tcltest86
  • Cause unknown (9.0a1, utfmax option that no longer exists)
  • Prototype none
  • Suggested ticket comment env-4.3 to 4.5 pass with current 9.1, 9.0 and 8.6 on Windows, also with LANG=C.UTF-8 set. Closing as out of date.
🟢e484a8407c2019pipe: time-related sporadic hangs in io/chan-io testsNOT-REPRODUCED (hangs: not on Win11 per Serhiy, 1400+ runs; test fixes merged into main only)

Open ticket e484a8407c · 25. Channel System · filed 2019-08-04

🟢 Since triage (updated 2026-11-11)

Status now: Open

  • 🔧 2026-10-14 check-in 4042ca23fc on tests-async-pipe-close, in main: Rebase to trunk
Full verdict

NOT-REPRODUCED (hangs: not on Win11 per Serhiy, 1400+ runs; test fixes merged into main only)

  • Platform Windows (tests)
  • Tested Linux main (own build, c092dae4af): io-27.6, io-29.32, chan-io-27.6, chan-io-29.32, 20 runs each, all pass (chan-io ones now actually run). On core-9-0-branch and core-8-6-branch chanio.test still has the mangled constraint "asyncPipeChan Close" (3 places), so those chan-io tests never run there.
  • Cause the tests returned before the child closed the output file (sporadic "permission denied" on Windows with shortened sleeps); chanio.test constraint asyncPipeClose was mangled when the file was created.
  • Existing work tests-async-pipe-close (9b3b167cb5 + 1 more) merged into main by jan.nijtmans (de9d1dcb7c "Fix [e484a8]", 2026-10-16), branch closed; NOT merged into core-9-0-branch or core-8-6-branch.
  • Prototype none (branch exists)
  • Suggested ticket comment (if backported) "Merged into main de9d1dcb7c; 9.0/8.6 still have the mangled asyncPipeClose constraint in chanio.test." Otherwise the ticket can be closed once the backport is done (Jan merged it to main but left the ticket Open).
⬜2d2a3212442019info frame does not correct recognize current type of execution scopeNOT-REPRODUCED (all tested)

Open ticket 2d2a321244 · 16. Commands A-H · filed 2019-03-15

Full verdict

NOT-REPRODUCED (all tested)

  • Platform generic
  • Tested the four snippets fed to tclsh on stdin and in a pty (interactive) on main build, 9.0.5, 8.7b1, 8.6.18 and 8.5.19, 8.6.0/5/8/9/10/11/12: every form (info frame [info frame], info frame 0, inside dict get, inside set) gives type proc both in the proc and in its namespace eval. From a file all give type source. The type eval result in the ticket never appears.
  • Cause none found. type describes where the command's text comes from (the literal ns body sits inside the proc body), not the kind of scope, as info.n says.
  • Prototype none
  • Suggested ticket comment Can't reproduce with 8.5.19 to 9.1: all variants consistently report type proc (or source when run from a file); type describes where the script text comes from, not the kind of scope. The eval result was probably from a development branch. Suggest closing unless there is a script that still shows it.
⬜2d34167e972019sporadic segfault in stack test-case (heisenbug)NOT-REPRODUCED (all tested)

Open ticket 2d34167e97 · 45. Parsing and Eval · filed 2019-03-06

Full verdict

NOT-REPRODUCED (all tested)

  • Platform generic
  • Tested the stack-3.1 script fed via stdin, 100 runs each on main (gcc), main (clang 22 build), 9.0 tip and 8.6 tip: 0 crashes. stack.test passes with the clang build.
  • Cause unknown. The only reports are two Travis clang jobs from 2019, likely from C-stack exhaustion near the 10000 recursion limit on a small CI stack.
  • Prototype none
  • Suggested ticket comment Not reproduced in 100 runs each on main (gcc and clang), 9.0 and 8.6, and nothing new since 2019. Suggest closing as "Works for me" unless it shows up in CI again.
⬜0d3963254c2018leaktest failureNOT-REPRODUCED (all tested)

Open ticket 0d3963254c · None · filed 2018-03-14

Full verdict

NOT-REPRODUCED (all tested)

  • Platform generic (seen only on macOS 10.12 in 2018)
  • Tested --enable-symbols=mem builds on Linux of main (full suite, all 522078 tests: only the 25 known zipfs.test failures caused by TCL_LIBRARY=<wt>/library) and of core-8-6-branch tip (full suite: 45088 tests, 0 failures, stringComp-14.2 passes; on main the test is string.test stringComp-14.22.*). Not tested on macOS. pointsman could not reproduce it on macOS 10.13 either.
  • Cause unknown
  • Prototype none
  • Suggested ticket comment With --enable-symbols=mem builds of the current 8.6 and main, the whole test suite (including stringComp-14.2 / stringComp-14.22.*) passes on Linux, and it did not reproduce on macOS 10.13 in 2018. Suggest closing as works-for-me.
⬜538ba1e6152017Test Failure - exec-19.1 - On MacOSNOT-REPRODUCED (main, 9.0, 8.6 on Linux); CANNOT-TEST (macOS). The test is skipped on the macOS CI with noosxCI (74eb0415b0, 2020-11-16), not fixed

Open ticket 538ba1e615 · - New Builtin Commands · filed 2017-11-07

Full verdict

NOT-REPRODUCED (main, 9.0, 8.6 on Linux); CANNOT-TEST (macOS). The test is skipped on the macOS CI with noosxCI (74eb0415b0, 2020-11-16), not fixed

  • Platform macOS (generic test timing)
  • Tested same runs as c057378986. The result "24" (one line missing) is reproduced exactly on Linux by starting one of the four children 2.5 s late.
  • Cause same as c057378986: the fixed after 5000 in exec-19.1 is too short on a slow or busy machine. Tcl does set O_APPEND.
  • Prototype prototypes/c057378986.patch (shared with c057378986; 13 changed lines). If it works, the noosxCI constraint could be dropped.
  • Suggested ticket comment Same cause as c057378986: exec-19.1 waits a fixed 5 s, and a child that starts late loses its last line in the count, not in the file (a 2.5 s delayed start gives exactly 24 on Linux). Proposed fix: poll for the output with a 30 s timeout, see the patch on c057378986.
✅58053c93b32017Test Failure - tcltest-8.4 - On MSYS builds of core-8-branchNOT-REPRODUCED (main, 8.6; MSVC builds)

Open ticket 58053c93b3 · 34. tcltest Package · filed 2017-11-07

✅ Since triage (updated 2026-11-11)

Status now: Closed (Works For Me)

  • 💬 2026-10-20 serhiy.storchaka: Not reproducible any more. tcltest-8.4 passes on Windows 11 with MSYS2 UCRT64 builds of main (9.1) and core-8-6-branch, as well as with MSVC builds, also for an elevated administrator and on an SMB path. The test was reworked since 2017 to use testchmod (which really denies writes) and the notFAT/notWsl constraints.
Full verdict

NOT-REPRODUCED (main, 8.6; MSVC builds)

  • Platform Windows
  • Tested main on Linux: tcltest-8.4 passes; win/58053c93b3.tcl gives "is not writable" as expected. Since 2017 the test was changed: it now uses testchmod 0o444 plus file attributes -readonly, and constraints {unixOrWin notRoot notFAT notWsl} (1cb4cb2fe2, 2023-03-02, for the constraint line)
  • Cause on Windows a read-only attribute on a directory does not stop writes; whether testchmod's ACL change works depends on the user (administrator) and on the file system (the report used a network drive O:)
  • Prototype none
  • Repro win/58053c93b3.tcl (run with tcltest.exe from the build dir, on a local NTFS drive and if possible on a mapped network drive); prints the FS, the method used, whether the directory is still writable, and a VERDICT line
  • Suggested ticket comment tcltest-8.4 has since been changed (testchmod plus the notFAT/notWsl constraints). Does it still fail on a current MSYS2 build? On a network drive Windows may ignore the permission change, in which case the test should be skipped there.
Windows VM, run 3
  • Platform Windows (test suite)
  • Verdict NOT-REPRODUCED (main, 8.6; MSVC builds)
  • Tested VM, elevated administrator account: tcltest.test -match tcltest-8.4 PASSED with tcltest91.exe (9.1b1) and tcltest86.exe (8.6.18). win/58053c93b3.tcl: with testchmod the directory is rejected as "not writable" and really cannot be written, on local NTFS and on an SMB path (\\localhost\C$\..., no drive mapping), 9.1b1 and 8.6.18; with plain tclsh (only file attributes -readonly) the directory stays writable, which is why the test now relies on testchmod. No MinGW/MSYS2 build of Tcl was made (testchmod in win/tclWinTest.c is the same code for both compilers).
  • Cause the 2017 failure came from the old test setup (read-only attribute only); the test was reworked to testchmod + notFAT/notWsl constraints (1cb4cb2fe2)
  • Prototype none
  • Suggested ticket comment tcltest-8.4 passes on Windows 11 with current 8.6 and 9.1 test builds, also for an elevated administrator and on an SMB share: the test now uses testchmod, which really denies writes. Closing unless it still fails on a current MSYS2 build.
⬜93da2e42b92017Test Failure - iortrans-11.1 + iortrans-11.2 - On MSYS builds of core-8-branchNOT-REPRODUCED (main with MSYS2 UCRT64 gcc and MSVC; 8.6 MSVC)

Open ticket 93da2e42b9 · 20. [interp] · filed 2017-11-07

Full verdict

NOT-REPRODUCED (main with MSYS2 UCRT64 gcc and MSVC; 8.6 MSVC)

  • Platform Windows
  • Tested Linux main and 8.6: iortrans-11.1/11.2/11.3 pass, and a /proc/PID/fd check at every removeFile call shows tempchanfile is already closed by cleanup time, so there is no fd leak on Unix.
  • Cause (suspected): the tests move the channel to another interp with testchannel cut/splice, so tempdone's catch {close $tempchan} in the main interp does nothing. The file is closed only when the child interp/transform is torn down; if that happens later on Windows (or leaks), removeFile fails with EACCES because Windows cannot delete an open file.
  • Repro win/93da2e42b9.tcl. Run it with the build's tcltest.exe: tcltest.exe 93da2e42b9.tcl C:/path/to/tcl/tests. It runs ioTrans.test -match iortrans-11.* in %TEMP% and prints a VERDICT line.
  • Prototype none (depends on the Windows result)
  • Suggested ticket comment (after the Windows run) Passes on Linux; on Windows the cleanup fails because tempchanfile is still open after the channel was spliced into another interp.
Windows VM, run 1
  • Platform Windows
  • Verdict NOT-REPRODUCED (main with MSYS2 UCRT64 gcc and MSVC; 8.6 MSVC)
  • Tested win/93da2e42b9.tcl 3× with a MinGW (UCRT64) tcltest.exe of main, once each with MSVC tcltest91/tcltest86; full ioTrans.test with all three: 72 passed, 34 skipped (thread), 0 failed, no leftover tempchanfile
  • Cause unknown; possibly fixed by later channel/transform cleanup changes
  • Prototype none
  • Suggested ticket comment iortrans-11.1/11.2 pass with current main built with MSYS2 gcc (and with MSVC, also 8.6), and the temp file is deleted. Closing as out of date.
⏳6b3a427c252017Crash with fconfigureNOT-REPRODUCED (main, 8.7b1; 8.6 uses select and is not affected by this code)

Open ticket 6b3a427c25 · 24. Channel Commands · filed 2017-08-11 · Pending (resolved, waiting for confirmation)

⏳ Since triage (updated 2026-11-11)

Status now: Pending (Works For Me)

    Full verdict

    NOT-REPRODUCED (main, 8.7b1; 8.6 uses select and is not affected by this code)

    • Platform unix (Linux epoll notifier, 8.7+)
    • Tested Linux, work-a2/6b3.tcl under valgrind: main (-DPURIFY build, epoll), 8.7b1, 8.6.18; 100 rounds of async connects to localhost (::1 refused, retry on 127.0.0.1; half to a closed port) with readable/writable fileevents, interleaved with opening regular files and fconfigure -encoding with and without fileevents -> no valgrind errors.
    • Cause unknown. The invalid write in the report is the LIST_REMOVE(filePtr, readyNode) for regular files in PlatformEventsControl(EPOLL_CTL_DEL) (unix/tclEpollNotfy.c; still the same logic, now on EPERM). It writes through an uninitialised node only when a FileHandler for that fd was never put on the regular-file list, e.g. a handler created for a socket whose fd was closed without Tcl_DeleteFileHandler and then reused by the regular file (an EPOLL_CTL_MOD that fails with EPERM is ignored). I found no Tcl code path that leaks a handler like that (TcpConnect deletes the handler before retrying); the reporter's own application (tkscidb, C++) may. The uninitialised epoll_ctl(event) bytes are just the unused half of the data union on 32-bit, harmless.
    • Existing work none (Pending, question to reporter 2026-10-12)
    • Prototype none (not reproducible). A hardening (zero filePtr->readyNode at creation and skip LIST_REMOVE when it was never inserted) would be ~5 lines if wanted.
    • Suggested ticket comment none needed; still waiting for a script. Can be closed if the reporter does not answer.
    ⬜32c5740a4d2015Undesirable [lappend] performance as LIST_MAX is reachedNOT-REPRODUCED (all tested) — the demo itself was fixed by dgp's 2015-11-30 trunk commit (realloc in Tcl_ListObjReplace, roadmap item 1); the ticke…

    Open ticket 32c5740a4d · 14. List Object · filed 2015-11-17

    Full verdict

    NOT-REPRODUCED (all tested) — the demo itself was fixed by dgp's 2015-11-30 trunk commit (realloc in Tcl_ListObjReplace, roadmap item 1); the ticket stayed open only for the cosmetic items 2–3

    • Platform generic
    • Tested work-a1/32c5.tcl (dgp's global-level lappend my_list A variant, i.e. the INST_LAPPEND_LIST_STK path). 8.6.18: a steady ~25 s per 50M elements all the way to LIST_MAX (536870909, then a clean "max length of a Tcl list" error), 0.06 s per 100k at 402.6M–403.4M where the slowdown used to start. Main (debug build): linear, ~84 s per 50M up to 200M; with 64-bit Tcl_Size its LIST_MAX is out of reach.
    • Cause (historical) Tcl_ListObjReplace copied the list on every growth instead of trying realloc. The 9.0 list rewrite (ListStore with shared growth code) also covers dgp's item 3, the refactoring.
    • Prototype none (nothing left to fix)
    • Suggested ticket comment Rechecked: growth stays linear up to LIST_MAX on 8.6.18 and on main (where LIST_MAX is 64-bit). The list code refactoring came with 9.0's new list implementation. Suggest closing.
    ⬜60c12e14992015Tests socket-1[45]\* need replacing.NOT-REPRODUCED (all tested)

    Open ticket 60c12e1499 · 27. Channel Types · filed 2015-08-02

    Full verdict

    NOT-REPRODUCED (all tested)

    • Platform generic (test suite)
    • Tested -match socket-1[45].* of socket.test on main (own build), 9.0 tip and 8.6 tip tcltest, Linux: all pass (main 28+9, 9.0 37, 8.6 35 tests run) in ~2 s each; localhost only (socket-14.13 with 255.255.255.255 is Windows-only).
    • Cause n/a — the tests have since been reworked (tcltests::localIPv4Address/localDualStackAddress helpers, constraints localhost_v4/v6, generated socket-15.1.N matrix).
    • Prototype none
    • Suggested ticket comment socket-14.* and socket-15.* now pass on Linux in about 2 seconds on 8.6, 9.0 and main after the test rework. Proposing to close; please reopen with the failing environment if they still misbehave somewhere.
    ✅40f315bc232014[format %g] does not portably produce a Tcl numberpatch 86patch ANOT-REPRODUCED for the original report (main, 9.0, 8.6 with MSVC 2022 UCRT: %g/%e/%f/%G/%E/%a and padded/signed forms of ±Inf are all valid Tcl num…

    Open ticket 40f315bc23 · 48. Number Handling · filed 2014-10-22

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-23 check-in 326e1bf923 on core-8-6-branch, in main, 9.0, 8.6: Fix 40f315bc23: [format %g] does not portably produce a Tcl number
    • 💬 2026-10-23 serhiy.storchaka: With msvcrt.dll (e.g. MinGW builds), 8.6 still gave 1.#INF. Fixed on main and 9.0 by 185825858e, backported to 8.6 in 326e1bf923.
    Full verdict

    NOT-REPRODUCED for the original report (main, 9.0, 8.6 with MSVC 2022 UCRT: %g/%e/%f/%G/%E/%a and padded/signed forms of ±Inf are all valid Tcl numbers). The side bug REPRODUCED on Windows as on Linux (main, 9.0): format %A Inf → IxF, %A -Inf → -xNF

    • Platform Windows
    • Tested Linux main build, 9.0.5, 8.7b1, 8.6.18 (8.6 has no %a). Windows repro: win/40f315bc23.tcl (checks every float conversion of ±Inf with string is double; prints VERDICT). It is likely fixed in VC2015+ UCRT builds ("inf"); old msvcrt.dll (MinGW without __USE_MINGW_ANSI_STDIO) printed "1.#INF".
    • Cause Windows: the C runtime's printf (via TclFormatDouble → snprintf). %A: generic/tclStringObj.c:Tcl_AppendFormatToObj always overwrites byte 1 with 'x' (assumes the output starts with "0X"), from 5cfe15b5d6 (2018-02-14).
    • Prototype prototypes/40f315bc23-A.patch (18 changed lines, test format-19.5): fixes only the %A garbling (finds "0X" with strstr; leaves INF alone); format.test passes. Windows Inf: none until the VM run shows whether current builds still print 1.#INF.
    • Suggested ticket comment Needs re-checking on current Windows builds (UCRT prints "inf"). While testing I found that &#91;format %A&#93; garbles negative, signed, padded and infinite values (e.g. <code>-xX1.8p+0</code>, <code>IxF</code>); a proposed fix with test format-19.5 is available.
    Windows VM, run 1
    • Platform Windows (original); the %A part is generic
    • Verdict NOT-REPRODUCED for the original report (main, 9.0, 8.6 with MSVC 2022 UCRT: %g/%e/%f/%G/%E/%a and padded/signed forms of ±Inf are all valid Tcl numbers). The side bug REPRODUCED on Windows as on Linux (main, 9.0): format %A Inf → IxF, %A -Inf → -xNF
    • Tested VM tclsh 9.1b1, 9.0.5, 8.6.18 (8.6 without %a/%A, which it lacks)
    • Cause original: old msvcrt printf ("1.#INF"), gone with UCRT. %A: generic/tclStringObj.c always overwrites byte 1 with 'x'.
    • Prototype prototypes/40f315bc23-A.patch (%A fix unchanged; test format-19.5 rewritten to use explicit precisions, because MSVC prints all 13 hex digits for %A 1.5 → 0x1.8000000000000p+0, so the original test failed on Windows only). format.test passes on Windows (wt build) and Linux (wt-win1 build). 20 changed lines.
    • Suggested ticket comment Current MSVC (UCRT) builds print "inf", so &#91;format&#93; of Inf gives valid Tcl numbers on 8.6, 9.0 and 9.1; the original report can be closed as out of date. Separately, &#91;format %A&#93; garbles signed, padded and infinite values on all platforms; proposed fix with test format-19.5.
    ⬜f8e4918f792014self test registry-6.21 fails on Win8 64bitNOT-REPRODUCED (main, 9.0, 8.6)

    Open ticket f8e4918f79 · 32. registry Package · filed 2014-09-23

    Full verdict

    NOT-REPRODUCED (main, 9.0, 8.6)

    • Platform Windows
    • Tested Windows only; not tested here. registry-6.21 has since been rewritten to use 16383-char names/values.
    • Cause unknown (reported for 8.5.16, 32-bit VC12: a 199-char multi_sz value came back with garbage at the end)
    • Prototype none. Windows repro: win/f8e4918f79.tcl (sz/multi_sz/expand_sz round trip for lengths 199..16383)
    • Suggested ticket comment Not reproducible on 8.6.5 (apnadkarni), and registry-6.21 now uses 16383-char names and values. If win/f8e4918f79-style round trips pass on 9.0/8.6, close as Works For Me.
    Windows VM, run 4
    • Platform Windows
    • Verdict NOT-REPRODUCED (main, 9.0, 8.6)
    • Tested VM win/f8e4918f79.tcl (sz/multi_sz/expand_sz round trips for names and values of 199, 200, 255, 256, 1000, 16383 chars in an HKCU test key, deleted afterwards): no mismatch on 9.1b1 (registry 1.4, built into tclsh91), 9.0.5 (1.3.7), 8.6.18 (1.3.5). registry.test with the DLL loaded explicitly (work-win4/regt.tcl): 121 passed, 4 skipped, 0 failed with tcltest86 and tcltest90 (incl. registry-6.21). 32-bit builds not tried.
    • Cause n/a (8.5.16-era report; apnadkarni could not reproduce it on 8.6.5 either)
    • Prototype none
    • Suggested ticket comment registry-6.21 and round trips of sz/multi_sz/expand_sz values from 199 to 16383 chars pass on Windows 11 with 8.6.18, 9.0.5 and 9.1. Closing as Works For Me.
    ⬜714db57d202014[string match] performance regressionNOT-REPRODUCED (all tested)

    Open ticket 714db57d20 · 18. Commands M-Z · filed 2014-07-28

    Full verdict

    NOT-REPRODUCED (all tested)

    • Platform generic
    • Tested string match *a*a*a*a*b [string repeat a 30] and a long *a*b*c*d match, optimized builds 8.6.1, 8.6.2, 8.6.18, 9.0.5, 9.1b1, two to three runs each: 8.6.1 ~285 us, 8.6.2 ~305 us, 8.6.18 ~265 us, 9.0.5 ~230-250 us, 9.1b1 ~245 us. The long-pattern case is equal (~1.2 ms) everywhere.
    • Cause the blamed check-in 7abbaed029 only touches TclOO stub/build files and load.test, nothing on the string match path. dgp saw no loss with -O0 and blamed optimizer and code layout effects. Current versions are faster than 8.6.2.
    • Prototype none
    • Suggested ticket comment 7abbaed029 does not touch the string match code, and current builds (8.6.18, 9.0, 9.1) are as fast as or faster than 8.6.1 on recursive *a*a*... patterns. This looks like an optimizer/code-layout artifact of that time; close.
    ⬜6c49da8a192014Edge case in [file normalize]?NOT-REPRODUCED (main, 9.0, 8.6, 8.6.0-8.6.4, 8.5.0-8.5.19 on Linux); reproduced only with 8.4.20

    Open ticket 6c49da8a19 · 36. Pathname Management · filed 2014-01-31

    Full verdict

    NOT-REPRODUCED (main, 9.0, 8.6, 8.6.0-8.6.4, 8.5.0-8.5.19 on Linux); reproduced only with 8.4.20

    • Platform generic
    • Tested exact layout from the report (subdir/sidedir_link -> absolute sidedir), relative and absolute arguments, cwd reached through a symlinked parent (as /tmp on macOS); all 8.5+ give .../fileA.txt like realpath; 8.4.20 gives .../subdir/fileA.txt. The 8.6.1 claim could not be reproduced (8.6.1 build gives the right answer).
    • Cause 8.4's normalization collapsed ".." lexically; 8.5 TclFSNormalizeAbsolutePath reads the link before backing up over it (trunk check-ins "fix to file normalization with relative links/with links" b69743eda9 2003-12-17, ea7a46acae 2004-01-29). The remaining link + ".." bugs (link targets containing "..") are 3577801 below.
    • Prototype none (not reproduced; the related fix is prototypes/3577801fff.patch)
    • Suggested ticket comment Cannot reproduce with 8.5.0 or later on Linux (8.4.20 shows it): [file normalize sidedir_link/../fileA.txt] gives the same result as realpath. The remaining link/".." cases are tracked in 3577801 (proposed fix there). Can be closed unless someone still sees it.
    ⬜3613497fff2013tclThreadAlloc crashes under heavy loadNOT-REPRODUCED (main, 8.6.18); no reproducer in the ticket

    Open ticket 3613497fff · 41. Memory Allocation · filed 2013-05-17

    Full verdict

    NOT-REPRODUCED (main, 8.6.18); no reproducer in the ticket

    • Platform generic
    • Tested stress script (work-a6/t9.tcl): 24 threads for 0.4 s each doing dict/list churn, plus nested short-lived threads so caches keep moving to the shared pool. 40/40 clean runs on main -O2 + Thread 3.0.7 and 40/40 on 8.6.18 + Thread 2.8.13. No 8.5.12 baseline: none of the Thread builds here loads into 8.5.
    • Cause unknown. Since 8.5.12, tclThreadAlloc.c got the zippy FIFO rework (b285882061, 2015-04), the fix for duplicate TclpFreeAllocCache calls at thread exit (3cc1d91345, 7a3c8a4b0d, 2016-10-11) and an allocator init fix that came up with NaviServer (7837a4a2ff, 2017-01-16). Any of these may have been the crash. The crashing frame is in NaviServer code (Ns_SetFindCmp), so heap corruption by the application is also possible.
    • Prototype none
    • Suggested ticket comment Several crash fixes went into tclThreadAlloc.c after 8.5.12 (&#91;7a3c8a4b0d&#93; duplicate TclpFreeAllocCache at thread exit, &#91;7837a4a2ff&#93; allocator init). A 24-thread stress test runs clean 40/40 on 8.6.18 and 9.1. Without a reproducer for a current version, we propose to close this.
    ⬜3584573fff2012cannot set thread stack size on linuxNOT-REPRODUCED (main, 9.0, 8.6): HAVE_PTHREAD_ATTR_SETSTACKSIZE=1 is defined on Linux in all current builds (unix/tcl.m4 has run AC_CHECK_FUNCS(pth…

    Open ticket 3584573fff · 49. Threading · filed 2012-11-06

    Full verdict

    NOT-REPRODUCED (main, 9.0, 8.6): HAVE_PTHREAD_ATTR_SETSTACKSIZE=1 is defined on Linux in all current builds (unix/tcl.m4 has run AC_CHECK_FUNCS(pthread_attr_setstacksize) since 2000), so TclpThreadCreate does honour a non-default stackSize.

    • Platform unix
    • Tested configure output of the wt-a6 main, tcl9.0 and tcl8.6 trees; thread count of an idle tclsh (1 on main, 9.0.5 and 8.6.18)
    • Cause n/a. On the second point, main and 9.0 use the epoll notifier on Linux, which has no notifier thread. 8.6's select notifier starts its thread lazily with TCL_THREAD_STACK_DEFAULT, which costs address space only, not RSS (as ferrieux noted).
    • Prototype none
    • Suggested ticket comment Current builds on Linux define HAVE_PTHREAD_ATTR_SETSTACKSIZE, so the stackSize given to Tcl_CreateThread is applied. Tcl 9 uses the epoll notifier on Linux and starts no notifier thread at all. Proposed to close.
    ⬜3561500fff2012[format] "-0.000" renders differently 32 vs 64 bit hardwareNOT-REPRODUCED (main, 9.0, 8.6 on Windows 11, MSVC 2022)

    Open ticket 3561500fff · 48. Number Handling · filed 2012-08-25

    Full verdict

    NOT-REPRODUCED (main, 9.0, 8.6 on Windows 11, MSVC 2022)

    • Platform Windows
    • Tested Linux main build, 9.0.5, 8.6.18, 8.5.19; attachments read (all 3 present)
    • Cause the CRT's printf drops the sign of -0.0 on XP msvcrt. Current builds use VC2015+ UCRT, which keeps it.
    • Prototype none (CRT behaviour on an unsupported OS); Windows repro win/3561500fff.tcl
    • Suggested ticket comment Both reports came from the same 32-bit tclkit, on XP vs Windows 7, so it was the XP C runtime. If current Windows builds print -0.000 (win/3561500fff.tcl), close as out of date.
    Windows VM, run 1
    • Platform Windows
    • Verdict NOT-REPRODUCED (main, 9.0, 8.6 on Windows 11, MSVC 2022)
    • Tested VM 9.1b1, 9.0.5, 8.6.18: %9.3f -0.000, %.3f -0.0, %g -0.0, %e -0.0, %.3f -0.0001 all keep the minus sign
    • Cause XP msvcrt.dll printf (as the Linux analysis found); UCRT is correct
    • Prototype none
    • Suggested ticket comment Current Windows builds (UCRT) print -0.000 the same as Linux; the difference was the XP C runtime. Closing as out of date.
    ✅3544685fff2012Threaded build failures on OpenBSD-currentpatchNOT-REPRODUCED (main, 8.6 on OpenBSD 7.9, 10 runs each). Side finding: on 8.7/9.x async-4.3 is skipped on every non-MSVC build because its constrai…

    Open ticket 3544685fff · 49. Threading · filed 2012-07-16

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-17 check-in fdf3d4acf3 on async-test-msvc-constraint, in main: Fix the knownMsvcBug constraint in async.test, inverted since c0c61fa7cf: async-4.3 ran only on MSVC builds, where it is known to fail, and was skipped everyw
    • 💬 2026-10-17 serhiy.storchaka: async-4.3 has not been run outside MSVC builds since c0c61fa7cf (2021): the knownMsvcBug constraint became true on MSVC instead of false. It passes on OpenBSD 7.9 (main, 10 of 10 runs) and on Linux (20 of 20 runs) with the constraint fixed in branch async-test-msvc-constraint.
    Full verdict

    NOT-REPRODUCED (main, 8.6 on OpenBSD 7.9, 10 runs each). Side finding: on 8.7/9.x async-4.3 is skipped on every non-MSVC build because its constraint was inverted in 9ddb36e0bd (TIP #599, 2021-11-26).

    • Platform unix (OpenBSD)
    • Tested none here (no OpenBSD); ticket history only
    • Cause OS timer behaviour (pthread_cond_timedwait waking up ~400 ms late), per ferrieux's ktrace analysis
    • Prototype none
    • To run on the OpenBSD VM main and 8.6 builds, make test TESTFLAGS="-file {event.test interp.test async.test}" repeated about 10 times. If clean, close.
    • Suggested ticket comment The Tcl-side bug (async-4.3) was fixed in 2012. The other failures came from late pthread_cond_timedwait wakeups on OpenBSD of that time. If event/interp/async tests pass repeatedly on a current OpenBSD, this can be closed.
    Unix VMs, run 1
    • Platform unix (OpenBSD)
    • Verdict NOT-REPRODUCED (main, 8.6 on OpenBSD 7.9, 10 runs each). Side finding: on 8.7/9.x async-4.3 is skipped on every non-MSVC build because its constraint was inverted in 9ddb36e0bd (TIP #599, 2021-11-26).
    • Tested OpenBSD 7.9 amd64 (8 CPUs), cc (clang), threaded --enable-symbols builds of main (+ Thread 3.0.7) and core-8-6-branch. event.test, interp.test and async.test, 10 runs each on both branches (main and 8.6 loops ran at the same time): 0 failures. interp-36.7, event-11.4 and event-12.4 ran and passed. On 8.6 async-4.3 ran and passed every time. On main async-4.3 was SKIPPED (knownMsvcBug). With the constraint fixed it ran and passed 10/10 on main.
    • Cause the original failures were OpenBSD 5.x pthread_cond_timedwait oversleeping (async-4.3 itself was fixed in 2012). Separate test bug: tests/async.test sets testConstraint knownMsvcBug [expr {[tcl::build-info msvc]>0}]. Before 9ddb36e0bd it was ![info exists ::env(CI_BUILD_WITH_MSVC)], i.e. true everywhere *except* MSVC. Now async-4.3 runs only on MSVC, the one place where it is known to fail.
    • Prototype prototypes/3544685fff.patch (2 changed lines: testConstraint knownMsvcBug [expr {![tcl::build-info msvc]}]; verified on OpenBSD main: async-4.3 runs and passes 10/10). It applies to core-9-0-branch as well (same line).
    • Suggested ticket comment On OpenBSD 7.9, event.test, interp.test and async.test pass 10 out of 10 runs on main and 8.6, so this can be closed. Unrelated finding: since 9ddb36e0bd the knownMsvcBug constraint in async.test is inverted, so async-4.3 runs only on MSVC. Proposed one-line fix (test passes on OpenBSD with it).
    ⏳3461051fff2011Crash on "exit" when resources are lowNOT-REPRODUCED (main on Linux/Xvfb; Windows 8.6.18 + Thread 2.8.13 and 9.1 + Thread 3.0.7 already tested by Serhiy, 10/10 clean)

    Open ticket 3461051fff · 49. Threading · filed 2011-12-16 · Pending (resolved, waiting for confirmation)

    ⏳ Since triage (updated 2026-11-11)

    Status now: Pending (Works For Me)

      Full verdict

      NOT-REPRODUCED (main on Linux/Xvfb; Windows 8.6.18 + Thread 2.8.13 and 9.1 + Thread 3.0.7 already tested by Serhiy, 10/10 clean)

      • Platform Windows (resource exhaustion), generic exit path
      • Tested Linux, Xvfb, wish 9.1 (main install) + Thread 3.0.7 built in work-a2: attached main.tcl with the network ping replaced by after 10 (no outside network), 20 threads, 200 toplevels x 50 entries, then exit: 10/10 runs exit 0. X11 has no window-handle quota, so the "low resources" part is Windows-only.
      • Cause unknown; dgp suspected Thread 2.7b1's re-enabled finalization; Thread 2.8/3.0 finalization was reworked since.
      • Existing work none (Pending / Works For Me at Serhiy's request, 2026-10-12)
      • Prototype none (not reproducible)
      • Suggested ticket comment none needed; can be closed.
      ⬜3400826fff2011Test iocmd.tf-24.16 fails on netbsdNOT-REPRODUCED (main + Thread 3.0.7 on NetBSD 10.1)

      Open ticket 3400826fff · 25. Channel System · filed 2011-08-30

      Full verdict

      NOT-REPRODUCED (main + Thread 3.0.7 on NetBSD 10.1)

      • Platform unix (NetBSD)
      • Tested test has the thread constraint, so it is skipped in our tclsh builds (no Thread package). Code reading only.
      • Cause The reported failure had an extra BG {finalize rc0} because the test drained the event queue with update, which let a second background flush run. Before 8.6.0 the test was rewritten to stop at the first BG write (vwait endbody-24.16, cc4f894e4a 2011-09-19, plus an after 1000 safety timeout in 29535daab1 2012-04-30), so this result is likely gone.
      • Prototype none
      • Suggested ticket comment The test was changed after this report (cc4f894e4a, 29535daab1) to wait for the first background write instead of running update. Can someone with NetBSD confirm that it now passes (it needs the Thread package)?
      • To run on the NetBSD VM build main + Thread, then ./tcltest ../tests/ioCmd.test -match 'iocmd.tf-24.*' (repeat ~20 times).
      Unix VMs, run 1
      • Platform unix (NetBSD)
      • Verdict NOT-REPRODUCED (main + Thread 3.0.7 on NetBSD 10.1)
      • Tested NetBSD 10.1 amd64 (8 CPUs), gcc, main built with --enable-symbols plus Thread trunk 3.0.7. Full ioCmd.test with TCLLIBPATH set to Thread: 377 total, 0 failed (the thread tests run). -match 'iocmd.tf-24.*' repeated 20 times: 0 failing runs; iocmd.tf-24.16 PASSED (checked with -verbose). 8.6 not run with Thread (it needs Thread 2.8); the 8.6 test is the same rewritten one.
      • Cause the test drained the queue with update; it was rewritten before 8.6.0 (cc4f894e4a, 29535daab1).
      • Prototype none
      • Suggested ticket comment On NetBSD 10.1 with Thread 3.0.7, iocmd.tf-24.16 and the rest of ioCmd.test pass on main (20 out of 20 runs of iocmd.tf-24.*). The test was rewritten after this report (cc4f894e4a, 29535daab1); suggest closing.
      ⬜3104471fff2010"command returned bad code: 2" in bgerrorNOT-REPRODUCED (main, 9.0.5, 8.6.18, 8.6.0)

      Open ticket 3104471fff · 01. Notifier · filed 2010-11-07

      Full verdict

      NOT-REPRODUCED (main, 9.0.5, 8.6.18, 8.6.0)

      • Platform generic
      • Tested coroutine used directly as a chan event handler, where a helper does return -level [info level] inside try/finally on EOF (the Wibble pattern), with and without the apply {{} {tailcall co}} wrapper: finally runs, no bgerror on any version.
      • Cause In 2010 trunk [info level] inside a coroutine apparently counted the frames outside the coroutine, so return -level [info level] escaped the coroutine and reached the event loop as TCL_RETURN. In 8.6.0+ [info level] is relative to the coroutine. A real return -level 2 out of a coroutine still gives "command returned bad code: 2", which is correct. Side note: for non-error codes TclDefaultBgErrorHandlerObjCmd (generic/tclEvent.c) prints "command returned bad code: 2x" (the -errorinfo, which starts with the result "x", is appended without a newline); the stale errorInfo from the report is gone.
      • Prototype none
      • Suggested ticket comment Not reproducible in 8.6.0 and later: return -level [info level] inside a coroutine called from [chan event] now stays inside the coroutine. Proposing to close.
      ⬜3103740fff2010socket_inet6-12.2 inconsistent failuresNOT-REPRODUCED (all tested)

      Open ticket 3103740fff · 27. Channel Types · filed 2010-11-05

      Full verdict

      NOT-REPRODUCED (all tested)

      • Platform unix
      • Tested socket.test -match 'socket_*-12.*' 40 runs on main (own build) and 40 on the 8.6 tip tcltest: 0 failures (inet, inet6 and any variants).
      • Cause unknown (2010 timing problem, probably the timeout tightening dgp suspected); the test has not failed here.
      • Prototype none
      • Suggested ticket comment socket_*-12.2 passed 40 of 40 runs on 9.1 and 8.6 (Linux). Proposing to close unless someone still sees it.
      ⬜3103656fff2010Case insensitive test result for test winFCmd-19.7NOT-REPRODUCED (main, 8.6)

      Open ticket 3103656fff · 37. File System · filed 2010-11-05

      Full verdict

      NOT-REPRODUCED (main, 8.6)

      • Platform Windows
      • Tested Linux only (n/a). winFCmd-19.7 is unchanged on main (constraint now win), still expects the lower-case tcl[pid].tmp from glob. Original failure was intermittent (Jan saw it once in 144443 runs; Jos often, with TEMP as a short 8.3 name).
      • Cause unknown; suspected win/tclWinFile.c:TclpObjNormalizePath returning a differently cased leaf for a nonexistent file (when TEMP has short names, or a scanner holds the file).
      • Prototype none (Windows only; a robust test would compare with string equal -nocase)
      • Suggested ticket comment The case change was rare even in 2010 (once in 144443 runs for Jan) and has not been reported since. Does winFCmd-19.7 still fail for anyone on 8.6.18 or 9.0? If not, close as works-for-me.
      Windows VM, run 5
      • Platform Windows
      • Verdict NOT-REPRODUCED (main, 8.6)
      • Tested VM win/3103656fff.tcl (20000 in-process + 100 subprocess normalizations) on 9.1b1 and 8.6.18, with the default TEMP and with TEMP/TMP set to an 8.3 short name (...\TEMPOR~1, 8dot3 enabled on C:); plus winFCmd-19.7 itself 20 times per version under the short-name TEMP: 0 mismatches, 0 failures (Defender real-time protection on).
      • Cause unknown (see result-44.md)
      • Prototype none
      • Suggested ticket comment Not reproducible on Windows 11 with 8.6.18 and 9.1 (40000 normalizations and 40 runs of winFCmd-19.7, also with TEMP as an 8.3 short name). Closing as works-for-me; please reopen if it fails again.
      ✅3006571fff2010error deleting: file already exists on VistaNOT-REPRODUCED (main, 8.6)

      Open ticket 3006571fff · 37. File System · filed 2010-05-24

      ✅ Since triage (updated 2026-11-11)

      Status now: Closed (Fixed)

      • 🔧 2026-10-23 check-in cfac6a1e07 on win-delete-retry-not-empty, in main: Fix 3006571fff: [file delete -force] of a directory failed with "file already exists" when another process (e.g. a virus scanner or an indexer) kept a file in
      • 🔧 2026-10-29 check-in 82967ff42a on trunk, in main, 9.0: Fix [3006571]: error deleting: file already exists on Vista
      • 🔧 2026-10-29 check-in 102cd04c1f on core-9-0-branch, in main, 9.0, 8.6: Fix [3006571]: error deleting: file already exists on Vista
      • 🔧 2026-10-29 check-in 4c56d701ab on core-8-6-branch, in main, 9.0, 8.6: Fix [3006571]: error deleting: file already exists on Vista
      • 💬 2026-10-23 serhiy.storchaka: Reproduced on 8.6, 9.0 and main under Wine, and on Windows 11 on an SMB share, when another process keeps a file in the tree open with FILE_SHARE_DELETE (as virus scanners and indexers do): the file is only marked for deletion, so removing its directory fails with ERROR_DIR_NOT_EMPTY, reported as "file already exists". Local NTFS on Windows 10 1903+ is not affected. Proposed fix in branch win-delete-retry-not-empt…
      • 💬 2026-10-29 jan.nijtmans: Fixed in 82967ff42a, core-9-0-branch and core.8.6.branch Many thanks! One closing '}' was placed wrong in your patch. Not a serious deal, but some compilers might still produce a warning on that .... spotted in time.
      Full verdict

      NOT-REPRODUCED (main, 8.6)

      • Platform Windows
      • Tested Linux main (n/a, no failures). Code unchanged on main/9.0/8.6. Windows repro: win/3006571fff.tcl (300 create/delete cycles, prints how many deletes failed).
      • Cause win/tclWinFCmd.c:TraverseWinTree/DoRemoveJustDirectory. DeleteFileW/RemoveDirectoryW only mark an entry for deletion while another process (indexer, antivirus, CIFS client) holds a handle, so removing the parent fails with ERROR_DIR_NOT_EMPTY (mapped to EEXIST, "file already exists") or ERROR_ACCESS_DENIED. There is no retry. Windows 10 1903+ with NTFS uses POSIX delete semantics by default, so it may not reproduce on Windows 11 local disks.
      • Prototype none (a fix means a retry loop or FILE_DISPOSITION_POSIX_SEMANTICS deletion, a behaviour design choice; needs Windows)
      • Suggested ticket comment Caused by Windows' delayed deletion when another process holds a handle; file delete -force has no retry. Does this still happen on Windows 10/11 (POSIX delete semantics on NTFS)? If it only affects network shares, a short retry in TraverseWinTree is one option.
      Windows VM, run 5
      • Platform Windows
      • Verdict NOT-REPRODUCED (main, 8.6)
      • Tested VM win/3006571fff.tcl: 1000 create/delete cycles each on 9.1b1 and 8.6.18 on local NTFS, and 500 each with the cwd on the SMB loopback share //localhost/C$/...: 0 failures (Defender real-time protection on).
      • Cause (see result-44.md) delayed deletion while another process holds a handle; Windows 10 1903+ uses POSIX delete semantics on NTFS
      • Prototype none
      • Suggested ticket comment Not reproducible on Windows 11 with 8.6.18 and 9.1 (1000 iterations on NTFS, 500 over SMB): newer Windows deletes with POSIX semantics, so a file held open by a scanner no longer blocks removing its directory. Closing as works-for-me.
      ⬜3002022fff2010clock microseconds broken on WindowsNOT-REPRODUCED (main 32- and 64-bit, 9.0 and 8.6 64-bit) on Windows 11; the heuristic is still in the code but cannot trigger on current Windows

      Open ticket 3002022fff · 06. Time Measurement · filed 2010-05-15

      Full verdict

      NOT-REPRODUCED (main 32- and 64-bit, 9.0 and 8.6 64-bit) on Windows 11; the heuristic is still in the code but cannot trigger on current Windows

      • Platform Windows
      • Tested read win/tclWinTime.c on main, 9.0, 8.6 tips; Linux 9.0.5 has microsecond resolution (control run of the script)
      • Cause win/tclWinTime.c IsPerfCounterAvailable (8.6: inline in NativeGetTime): when the QPC frequency is above 15 MHz, the performance counter is disabled unless CPUID says GenuineIntel P4/HT with all CPUs on one package. Then NativeGetMicroseconds falls back to the millisecond system time. The whole block is #if !defined(_WIN64), so 64-bit builds always use QPC, and only 32-bit builds on AMD or newer Intel CPUs get the millisecond resolution. TclpGetWideClicks uses QPC directly in all builds.
      • Prototype none (whether to drop the CPUID heuristic on Win32 is a policy decision; modern Windows guarantees an invariant QPC)
      • Repro win/3002022fff.tcl (run with a 32-bit and a 64-bit tclsh)
      • Suggested ticket comment 64-bit builds skip the CPUID check and have microsecond resolution. 32-bit builds still disable QPC on non-Intel or post-P4 CPUs. Since Vista/Win7, QPC is reliable across cores, so the Win32 heuristic could be removed.
      Windows VM, run 3
      • Platform Windows (32-bit builds only)
      • Verdict NOT-REPRODUCED (main 32- and 64-bit, 9.0 and 8.6 64-bit) on Windows 11; the heuristic is still in the code but cannot trigger on current Windows
      • Tested VM (Intel i9-14900HX) win/3002022fff.tcl: 9.1b1, 9.0.5, 8.6.18 x64 and a 32-bit 9.1b1 tclsh built from tcl-wt91 (MACHINE=IX86, removed afterwards): 1998–1999 of 2000 samples have sub-ms digits, 2000 distinct values. QueryPerformanceFrequency is 10 MHz.
      • Cause win/tclWinTime.c IsPerfCounterAvailable only distrusts QPC (Win32 builds) when its frequency is above 15 MHz; Windows 10 and later report a fixed 10 MHz QPC (and Vista/7 used HPET/PM timer or TSC/1024 rates), so the CPUID check is dead code on supported systems
      • Prototype none (optional cleanup: drop the #if !defined(_WIN64) CPUID block)
      • Suggested ticket comment On Windows 11 the 32-bit and 64-bit builds of 9.1 (and 64-bit 8.6.18/9.0.5) have microsecond resolution: QPC now runs at a fixed 10 MHz, below the 15 MHz threshold of the Win32 heuristic in IsPerfCounterAvailable, so the heuristic no longer fires. Closing as out of date; the CPUID block could be removed as a cleanup.
      ⬜2966493fff2010chan-io-53.9 test is timing sensitiveNOT-REPRODUCED (all tested)

      Open ticket 2966493fff · 25. Channel System · filed 2010-03-09

      Full verdict

      NOT-REPRODUCED (all tested)

      • Platform generic
      • Tested chan-io-53.9 and io-53.9, 40 runs each on main (own build), 8 at a time in parallel to load the machine; chan-io-53.9 40 runs on the 8.6 tip tcltest: 0 failures.
      • Cause unknown; dgp could not repeat it himself in 2010 (single failure on a busy machine).
      • Prototype none
      • Suggested ticket comment chan-io-53.9/io-53.9 passed 40 of 40 runs under parallel load on 9.1 and 8.6. The reporter never saw it again either; proposing to close.
      ⬜2912794fff2009TCL install after compile win7 vs2008NOT-REPRODUCED (main; 8.6 and 9.0 have the same buildall.vc.bat sequence, not run)

      Open ticket 2912794fff · None · filed 2009-12-11

      Full verdict

      NOT-REPRODUCED (main; 8.6 and 9.0 have the same buildall.vc.bat sequence, not run)

      • Platform build (Windows)
      • Tested read win/buildall.vc.bat on main: it now runs nmake ... release htmlhelp OPTS=none %1 and then nmake ... shell OPTS=static %1, so with install the static shell is built before installing in the same nmake call; probably fixed, not run. Repro: win/2912794fff.cmd (clean tree, VS prompt).
      • Cause in 2009 the install step needed tclsh86sx.exe, which only buildall.vc.bat all built.
      • Prototype none
      • Suggested ticket comment buildall.vc.bat now builds the static shell before installing in the same nmake call (shell install); if the Windows check passes, close as out of date. A related remaining issue (static dde lib in install) is 2b874804b5.
      Windows VM, run 5
      • Platform build (Windows)
      • Verdict NOT-REPRODUCED (main; 8.6 and 9.0 have the same buildall.vc.bat sequence, not run)
      • Tested VM, fresh fossil open of main in a new dir, MSVC 2022 x64: win/2912794fff.cmd (buildall.vc.bat install) built release and failed only in htmlhelp (tcltk-man2html.tcl picked an unrelated ../../../tcl* dir as source tree in my scratch layout; not related to the ticket) and then stopped at a pause. Re-run of buildall's two nmake steps without htmlhelp on a clean tree: nmake release OPTS=none install and nmake shell OPTS=static install both succeed; INSTALLDIR/bin gets tcl91.dll, tclsh91.exe and tclsh91s.exe.
      • Cause (2009) install needed the static shell before it was built; buildall.vc.bat now builds shell in the same nmake call as install.
      • Prototype none
      • Suggested ticket comment On Windows 11 with VS 2022, buildall.vc.bat's steps (release ... install, then shell OPTS=static install) succeed on a clean main checkout and install both shells. Closing as out of date.
      ⬜2896758fff2009fcopy swamp (bytes vs chars), and http packagepatchNOT-REPRODUCED (main, 9.0; the three-way rule is gone) — 8.6 keeps the old rule, which is what its fcopy.n documents; small DOCS nit on main/9.0

      Open ticket 2896758fff · 24. Channel Commands · filed 2009-11-12

      🧪 Prototype check (2026-10-24)

      2896758fff.patch: VERIFIED (not part of the patch: Lossless() tests outStatePtr->inputEncodingFlags where outputEncodingFlags looks intended, as the triage noted)

      • Base main (tip d6dd6d0f71); docs only
      • Before doc/fcopy.n says that in the deprecated tcl8 mode "an eventually given size is in bytes". On the tip tcltest 2896.tcl (8 x U+00E9 in utf-8, both channels utf-8, -translation lf): profile tcl8 -size 4 -> fcopy returns 4, output 8 bytes, so -size counts characters; only without -size does the tcl8 path copy raw bytes (returns 16). generic/tclIO.c Lossless() requires toRead == -1 for that path. So the doc sentence is wrong
      • Patch applied cleanly (8 changed lines; also fixes "Depreciated")
      • After man -l fcopy.n (from doc/) renders "...set to “tcl8”, the input encoding matches the output encoding and no -size is given."
      • Test none possible (docs only)
      Full verdict

      NOT-REPRODUCED (main, 9.0; the three-way rule is gone) — 8.6 keeps the old rule, which is what its fcopy.n documents; small DOCS nit on main/9.0

      • Platform generic
      • Tested main (own build), 9.0.5, 8.6.18: utf-8->utf-8, utf-8->unicode, binary->utf-8, iso8859-5->iso8859-5, -profile tcl8 same encoding, all with -size 4
      • Cause In 9.x generic/tclIO.c:Lossless() moves raw bytes only for binary->binary, or for same encoding + tcl8 profile *without* -size (toRead == -1). So -size is in bytes iff the input is binary, otherwise characters, as fcopy.n says. 8.6: same encoding -> bytes (documented there). fcopy.n on main/9.0 still says "An eventually given size is in bytes in this case" for the tcl8 deprecated mode, which is wrong since that mode is never used with -size. (Side note: Lossless() checks outStatePtr->inputEncodingFlags where outputEncodingFlags looks intended.) http.tcl copies from a binary socket, so -blocksize is bytes: no issue there.
      • Prototype prototypes/2896758fff.patch (6 changed lines, doc only: fcopy.n deprecated-mode paragraph)
      • Suggested ticket comment Since 9.0 the rule is simple: fcopy -size is in bytes when the input channel is binary, and in characters otherwise (the same-encoding byte mode is only used without -size). 8.6 behaves as its documentation says. Proposed to close after a small fcopy.n fix.
      ⬜2815500fff2009enabling dtrace breaks error handlingNOT-REPRODUCED (main, 8.6 with --enable-dtrace on OpenIndiana x86-64, gcc 14.3, illumos dtrace); SPARC CANNOT-TEST (no SPARC VM)

      Open ticket 2815500fff · 55. Other Tools · filed 2009-07-01

      Full verdict

      NOT-REPRODUCED (main, 8.6 with --enable-dtrace on OpenIndiana x86-64, gcc 14.3, illumos dtrace); SPARC CANNOT-TEST (no SPARC VM)

      • Platform unix (Solaris 10 SPARC, gcc 3.4.6, Sun D 1.3)
      • Tested nothing could be run. das could not reproduce with Mac OS X DTrace in 2009. The DTrace probe code has been reworked a lot since then (Dispatch() single probe point cf9917a942 2013, 9.0 Tcl_Size work 2023–24).
      • Cause unknown. Most likely an old gcc / Sun D 1.3 interaction with the USDT probe macros around command return codes, so not Tcl logic.
      • Prototype none
      • Suggested ticket comment Cannot reproduce without a Solaris DTrace setup. The probe code changed substantially since 8.5.7. If still relevant, please retest with 8.6.18/9.0 (./configure --enable-dtrace && make, then proc bad {} {error bad}; catch bad must return 1); otherwise we will close this as out of date.

      (For the main session: try on the OpenIndiana VM with --enable-dtrace and run make test.)

      Unix VMs, run 1
      • Platform unix (Solaris 10 SPARC, gcc 3.4.6, Sun D 1.3)
      • Verdict NOT-REPRODUCED (main, 8.6 with --enable-dtrace on OpenIndiana x86-64, gcc 14.3, illumos dtrace); SPARC CANNOT-TEST (no SPARC VM)
      • Tested OpenIndiana (illumos-4648b9b8c3) i86pc. Out-of-tree --enable-symbols --enable-dtrace builds of main and core-8-6-branch (configure finds /usr/sbin/dtrace; tclDTrace.o linked). proc bad {} {error bad}; catch bad returns 1 on both, with and without probes enabled. Full suite (-notfile http*.test) with DTrace: main 522078 tests, 37 failed; 8.6 45088 tests, 10 failed. The failures are the same without DTrace (scan-15.1, socket-14.x async connect; zipfs only with TCL_LIBRARY set, gmake test -file zipfs.test passes). With probes enabled (pfexec dtrace -Z on proc-return/cmd-return/inst-done, 1.3 M inst-done firings), error/proc/basic/append/compile.test pass on both.
      • Cause unknown; most likely gcc 3.4.6 / Sun D 1.3 on SPARC around the USDT probe code of 2009. The probe code has been reworked since.
      • Prototype none
      • Suggested ticket comment Cannot reproduce on OpenIndiana (x86-64, gcc 14, illumos dtrace): with --enable-dtrace, main and 8.6 return 1 from catch bad and pass the test suite the same as without DTrace, also with probes enabled. SPARC was not available. Suggest closing as out of date.
      ⬜2750491fff2009Tcl exit can misbehave on WindowsNOT-REPRODUCED (main, 9.0, 8.6)

      Open ticket 2750491fff · 52. Portability Support · filed 2009-04-10

      Full verdict

      NOT-REPRODUCED (main, 9.0, 8.6)

      • Platform Windows
      • Tested wrote win/2750491fff.tcl: 2×N child tclsh runs (with and without TCL_FINALIZE_ON_EXIT=1) that open a localhost socket, use the event loop and exit 7. It checks every exit code and detects hangs with a 20 s timeout. On Linux 9.0.5, 40/40 runs returned 7.
      • Cause unknown and not reproducible on demand (the reporter saw a low-frequency wrong exit code or hang, and exit code 0 in a TCL_MEM_DEBUG shared build). The exit sequence was reworked after this report (bug 2001201): since 8.6, Tcl_Exit by default does not run full Tcl_Finalize unless TCL_FINALIZE_ON_EXIT is set. The proposed TerminateProcess() patch was not applied, and main still ends with plain exit(status) in generic/tclEvent.c.
      • Prototype none
      • Suggested ticket comment The exit sequence was reworked in 8.6 (finalization on exit is now opt-in), and nobody has reproduced this since 2010. Unless the repro script shows wrong exit codes on current Windows builds, close as out of date.

      (For the main session: run tclsh90 win\2750491fff.tcl 500 on the VM, also with a TCL_MEM_DEBUG / --enable-symbols=mem shared build if one is available.)

      Windows VM, run 1
      • Platform Windows
      • Verdict NOT-REPRODUCED (main, 9.0, 8.6)
      • Tested win/2750491fff.tcl with 500 child runs each (250 with and 250 without TCL_FINALIZE_ON_EXIT) on 9.1b1, 9.0.5, 8.6.18, plus 200 runs with a shared TCL_MEM_DEBUG build of main (OPTS=symbols STATS=memdbg): every exit code was 7, no hangs
      • Cause unknown; the exit sequence was reworked in 8.6 (bug 2001201)
      • Prototype none
      • Suggested ticket comment 2200 test runs on current 8.6, 9.0 and 9.1 Windows builds, including a shared TCL_MEM_DEBUG build, gave the right exit code every time and never hung. Closing as out of date, as suggested in 2010.
      ⬜2666369fff2009inconsistent [format %e] outputNOT-REPRODUCED (main, 9.0, 8.6 with MSVC 2022 / UCRT)

      Open ticket 2666369fff · 10. Objects · filed 2009-03-05

      Full verdict

      NOT-REPRODUCED (main, 9.0, 8.6 with MSVC 2022 / UCRT)

      • Platform Windows
      • Tested Linux main, 9.0.5, 8.6.18: 3.113412e+00 (correct). Repro script win/2666369fff.tcl (prints VERDICT).
      • Cause [format] in generic/tclStringObj.c passes %e/%g to the C runtime's snprintf. The old MSVCRT printed 3-digit exponents; the UCRT (VS2015+) prints 2 by default. Builds that use msvcrt.dll (older MinGW without UCRT) may still show it.
      • Prototype none (needs a Windows check first; a fix would be _set_output_format(_TWO_DIGIT_EXPONENT) for msvcrt-based builds, or Tcl's own exponent formatting)
      • Suggested ticket comment Cannot be reproduced on Unix. Run win/2666369fff.tcl on MSVC (UCRT) and MinGW builds. With the UCRT the exponent should have 2 digits, as on Unix.
      Windows VM, run 2
      • Platform Windows
      • Verdict NOT-REPRODUCED (main, 9.0, 8.6 with MSVC 2022 / UCRT)
      • Tested VM 9.1b1, 9.0.5, 8.6.18 with win/2666369fff.tcl: %e 3.11341234 → 3.113412e+00, %E → E+00, %.3e 12345.0 → 1.234e+04, %e 1e100 → 1.000000e+100, same as Unix
      • Cause the old msvcrt.dll printf printed 3-digit exponents; the UCRT (VS2015+, also MSYS2 UCRT64) prints 2
      • Prototype none
      • Suggested ticket comment With current Windows builds (MSVC 2022 / UCRT) &#91;format %e&#93; gives 2-digit exponents (3.113412e+00) on 8.6, 9.0 and 9.1, as on Unix; the 3-digit form came from the old msvcrt.dll C runtime. Closing as out of date.
      ⬜2035893fff2008can NRE break limits and script cancellation?NOT-REPRODUCED (all tested)

      Open ticket 2035893fff · 45. Parsing and Eval · filed 2008-08-02

      Full verdict

      NOT-REPRODUCED (all tested)

      • Platform generic
      • Tested main (own build), 8.6.18: command and time limits stop infinite loops built from tailcall, coroutine resume/yield, yield+tailcall, deep uplevel recursion, apply+tailcall and a self-recursive ensemble (the last two hit the nesting limit first). Cancellation keeps working afterwards (see 2114165fff).
      • Cause msofer's remaining point (2009) stands by design: an NRE callback that runs long C code postpones cancellation until it returns, like any C code. That is not a hole in the limit/cancel checks.
      • Prototype none
      • Suggested ticket comment Command and time limits still work with tailcall, coroutines, uplevel and apply loops in 8.6 and 9.1. The only remaining case, a C callback that runs for a long time, is the same as any long-running C command. Suggest closing.
      ⬜1963074fff2008fCmd.test 27.5 and 27.6 failure on SPARC SolarispatchfCmd-27.5/27.6 NOT-REPRODUCED (main, 8.6 on OpenIndiana, also in a setgid directory of a group the user is not in). The underlying chown() of an un…

      Open ticket 1963074fff · 37. File System · filed 2008-05-13

      🧪 Prototype check (2026-10-24)

      1963074fff.patch: VERIFIED (real Solaris still untested)

      • Base main (tip d6dd6d0f71)
      • Before tcltest tests/fCmd.test -match 'fCmd-27.*' -> fCmd-27.7 FAILED (chown of / to its own owner/group as non-root: operation not permitted); with the EPERM chown shim (LD_PRELOAD=work-a5/shim/eperm.so, simulating _POSIX_CHOWN_RESTRICTED) fCmd-27.5, 27.6, 27.7 FAILED, as reported on Solaris
      • Patch applied cleanly (test hunk offset -9 lines; 19 changed lines)
      • After fCmd 261/0 failed (also 0 failed with the EPERM shim), unixFCmd 30/0, cmdAH 16971/0, fileSystem 141/0
      • Test existing fCmd-27.7 (unix notRoot) fails before / passes after
      Full verdict

      fCmd-27.5/27.6 NOT-REPRODUCED (main, 8.6 on OpenIndiana, also in a setgid directory of a group the user is not in). The underlying chown() of an unchanged owner is REPRODUCED (main, 8.6).

      • Platform unix
      • Tested main (shim work-a5/shim/eperm.so), 9.0 tip (new test fails), code reading
      • Cause unix/tclUnixFCmd.c:SetGroupAttribute/SetOwnerAttribute always call chown(), even when setting the file's current group/owner. With _POSIX_CHOWN_RESTRICTED a non-member/non-root gets EPERM. That happens when the file inherits a group the user isn't in (setgid dir, NFS).
      • Prototype prototypes/1963074fff.patch (19 changed lines, test fCmd-27.7: unchanged owner+group on / as non-root. It fails on 9.0 tip with "operation not permitted" and passes with the patch. fCmd.test also passes with the eperm shim.)
      • Suggested ticket comment The tests set the file's current group, but chown() is still called and fails under restricted chown when the group isn't one of the user's. Proposed fix: skip chown() when the owner/group is unchanged (new test fCmd-27.7). Also covers duplicate 1840211.
      Unix VMs, run 1
      • Platform unix (Solaris/illumos)
      • Verdict fCmd-27.5/27.6 NOT-REPRODUCED (main, 8.6 on OpenIndiana, also in a setgid directory of a group the user is not in). The underlying chown() of an unchanged owner is REPRODUCED (main, 8.6).
      • Tested OpenIndiana as a normal user (uid serhiy, gid staff). fCmd.test passes on main (306/0 failed), 8.6 (256/0) and with the prototype (307/0). As root I made a setgid directory with group other (removed afterwards); files created there get group other. fCmd-27.* run there still passes, and chgrp other f works too: illumos lets the owner set the file's current group under rstchown, even as a non-member, so this case is no longer hit. Setting the unchanged owner of a file one does not own still fails: file attributes / -owner root -group root → main "could not set owner for file "/": operation not permitted", 8.6 "not owner". With the prototype it returns "" (new fCmd-27.7 passes).
      • Cause unix/tclUnixFCmd.c:SetGroupAttribute/SetOwnerAttribute always call chown(); old Solaris (and NFS, per the ticket) refuse even a no-op change.
      • Prototype prototypes/1963074fff.patch (19 changed lines, test fCmd-27.7). Verified on OpenIndiana: fCmd.test passes with it, fCmd-27.7 fails without it. Behaviour change to note in review: setting the current owner/group of someone else's file now succeeds instead of EPERM.
      • Suggested ticket comment On current illumos fCmd-27.5/27.6 pass, also for files in a foreign-group setgid directory, because chown() to the current group is allowed there. A no-op owner change on a file the user doesn't own still fails (EPERM). Proposed fix (skip chown() when nothing changes, test fCmd-27.7) is verified on OpenIndiana.
      ⬜1958930fff2008exec with pipes and fileevents leaks ~2Mb everytimeNOT-REPRODUCED (main, 9.0, 8.6)

      Open ticket 1958930fff · 24. Channel Commands · filed 2008-05-06

      Full verdict

      NOT-REPRODUCED (main, 9.0, 8.6)

      • Platform Windows
      • Tested code reading main/9.0/8.6; repro script win/1958930fff.tcl (core-only variant from ferrieux's 2008-05-08 analysis: close a non-blocking open |tclsh block.tcl while the child is still running, 15 times; reports VM growth per close and slow closes)
      • Cause win/tclWinPipe.c:TclPipeThreadStop terminates the reader thread blocked in ReadFile with TerminateThread, whose stack is never freed. On main/9.0 it first calls CancelSynchronousIo(hThread) (arrived on main via 5295f674c1, 2021-01-08 merge from 8.7), so ReadFile should return and the thread exit cleanly; core-8-6-branch has no CancelSynchronousIo and still waits 20 ms + 5 s and then calls TerminateThread. The reporter's own case (TclX pipes) was a script problem (unused pipe ends left open), as he confirmed.
      • Prototype none (if 8.6 reproduces, backporting the CancelSynchronousIo call is the fix; needs Windows to verify)
      • Suggested ticket comment Since 8.7/9.0 TclPipeThreadStop cancels the blocked ReadFile with CancelSynchronousIo before falling back to TerminateThread, which should remove the leak; 8.6 still terminates the thread. To be verified on Windows with win/1958930fff.tcl.
      Windows VM, run 2
      • Platform Windows
      • Verdict NOT-REPRODUCED (main, 9.0, 8.6)
      • Tested VM win/1958930fff.tcl (15 non-blocking closes of open |tclsh block.tcl while the child runs) and a variant with 40 closes that also counts threads and private bytes 9 s later (work-win2/1958b.tcl), 3 runs: 9.1b1 and 9.0.5 grow by 0 bytes / 0 threads; 8.6.18 grows by one thread (+2.3 MB virtual) per run, the same for 15 and 40 closes, so it is a one-time helper thread, not a per-close leak. Rarely a close takes > 1 s (the 5 s wait before the TerminateThread fallback): 1 of 120 on 8.6, 1 of 80 on 9.1, 0 of 55 on 9.0 — without memory growth. (The second 9.0.5 run hung in after for good; that is the separate Tcl_Sleep bug at the end of this file.)
      • Cause the Linux analysis was wrong about 8.6: core-8-6-branch TclPipeThreadStop also calls CancelSynchronousIo (through tclWinProcs.cancelSynchronousIo), so the blocked ReadFile is cancelled and the reader thread exits cleanly; TerminateThread remains only as a rare fallback
      • Prototype none
      • Suggested ticket comment No leak on Windows 11 with 8.6.18, 9.0.5 or 9.1: closing pipes to running children 40 times does not grow memory or the thread count, because the blocked reader thread is now cancelled with CancelSynchronousIo instead of being terminated. The reporter's own case was a script issue. Closing as fixed.
      ⬜1906935fff2008Output File Size is incorrectNOT-REPRODUCED (main, 8.6, 8.5)

      Open ticket 1906935fff · 37. File System · filed 2008-03-04

      Full verdict

      NOT-REPRODUCED (main, 8.6, 8.5)

      • Platform Windows
      • Tested win/1906935fff.tcl on Linux: sparse 4 GB + 3 MB file, the reporter's seek 16 current/read/tell loop across the 4 GB boundary, output size checked. OK on all versions.
      • Cause unknown (reported for 8.5.1 on 32-bit Windows XP. Possibly a wide-offset bug in the Windows channel driver that has long been fixed. Reporter never answered the 8.5.2 rc question.)
      • Prototype none
      • Suggested ticket comment Doesn't reproduce on Linux with a 4 GB+ file (8.5.1 to 9.1). Windows repro script attached: win/1906935fff.tcl (writes ~4 GB to %TEMP%). If it passes on Windows too, close as Out of Date.
      Windows VM, run 5
      • Platform Windows
      • Verdict NOT-REPRODUCED (main, 8.6, 8.5)
      • Tested VM work-win5/lf.tcl (win/1906935fff.tcl with a pre-created sparse 4 GB + 3 MB source, fsutil sparse setflag + seteof, because the VM disk has < 3 GB free): the reporter's seek 16 current/read/tell loop across the 4 GB boundary, output size checked. OK on 64-bit 9.1b1, 8.6.18, 8.6.0, 8.5.19 and on 32-bit (MACHINE=IX86) builds of main and core-8-6-branch.
      • Cause unknown (8.5.1 on 32-bit XP; long gone)
      • Prototype none
      • Suggested ticket comment Not reproducible on Windows 11 with 32- and 64-bit builds of 8.6.18 and 9.1 (and 8.5.19): seek/read/tell across 4 GB and the output size are correct. Closing as out of date.
      ⬜1904907fff2008Fix [chan create]/return dict key finalization order issuesNOT-REPRODUCED (main, 9.0, 8.6, 8.5) — the crash was worked around in 2008 (dgp's keys[i] = NULL in ReleaseKeys, committed by andreas_kupries 200…

      Open ticket 1904907fff · 24. Channel Commands · filed 2008-02-29

      Full verdict

      NOT-REPRODUCED (main, 9.0, 8.6, 8.5) — the crash was worked around in 2008 (dgp's keys[i] = NULL in ReleaseKeys, committed by andreas_kupries 2008-03-08); ticket was kept open for a "proper" finalization-order fix

      • Platform generic
      • Tested attached tst_chan.tcl and variants (read caught; no read so the data is left for exit-time flush) on main (own build), 9.0.5, 8.6.18, 8.5.19, and valgrind on a -DPURIFY main build: no crash, no valgrind errors. The refchan write method is not called at exit at all now (the buffered "HELLO THERE" is silently dropped), so the late Tcl_GetReturnOptions path is not even reached with this script.
      • Cause generic/tclResult.c:GetKeys/ReleaseKeys keep per-thread return-option key objects released by a thread exit handler; ReleaseKeys NULLs them, so a later GetKeys re-creates them (the "leak" is that the second set may never be freed).
      • Prototype none (no reproducer left; the remaining work is a cleanup, not a bug fix)
      • Suggested ticket comment No crash with the attached script on 9.1, 9.0 or 8.6 (also clean under valgrind); the 2008 workaround in ReleaseKeys is still in place. The remaining concern is a possible small leak at finalization. Suggest closing unless someone wants to pursue moving the key cleanup to interp deletion.
      ⏳1888596fff2008Segfault with tclhttpd and TLSNOT-REPRODUCED (main, 9.0, 8.6)

      Open ticket 1888596fff · 25. Channel System · filed 2008-02-07 · Pending (resolved, waiting for confirmation)

      ⏳ Since triage (updated 2026-11-11)

      Status now: Pending (Works For Me)

        Full verdict

        NOT-REPRODUCED (main, 9.0, 8.6)

        • Platform generic
        • Tested Linux, work-a2/188.tcl under valgrind: main (-DPURIFY build), 9.0.5, 8.6.18; writable fileevent scripts on async sockets that fail after closing the channel, replacing/deleting their own handler, closing and opening a new channel with its own failing handler, and popping/pushing a stacked transform (stand-in for TLS) before failing, 50 rounds each -> no valgrind errors, no crash. tclhttpd/TLS 1.5 not tested.
        • Cause (historic) the attached 8.4.17 backtrace shows DeleteScriptRecord() walking a freed script-record list after the script failed (stacked channel: the notified channel differs from esPtr->chanPtr). TclChannelEventScriptInvoker() now saves the mask first, preserves the channel and skips DeleteScriptRecord() when the channel was closed (typePtr == NULL).
        • Existing work none (Pending / Works For Me at Serhiy's request, 2026-10-12)
        • Prototype none (not reproducible)
        • Suggested ticket comment none needed; resolution holds (also with stacked channels popped/pushed in the handler), can be closed.
        ⬜1840211fff2007tcl 8.5.0 rc0 compile and test error reportduplicate of 1963074fff for the fCmd-27.5/27.6 part (NOT-REPRODUCED on OpenIndiana, see there); the tclEnv.c part was fixed in 2007

        Open ticket 1840211fff · 37. File System · filed 2007-11-28

        Full verdict

        duplicate of 1963074fff for the fCmd-27.5/27.6 part (NOT-REPRODUCED on OpenIndiana, see there); the tclEnv.c part was fixed in 2007

        • Platform unix
        • Tested see 1963074fff
        • Cause see 1963074fff
        • Prototype prototypes/1963074fff.patch (shared)
        • Suggested ticket comment The tclEnv.c warning was fixed in 2007. The fCmd-27.5/27.6 failures are the same as 1963074fff; proposed fix there. Closing as Duplicate.
        Unix VMs, run 1
        • Platform unix
        • Verdict duplicate of 1963074fff for the fCmd-27.5/27.6 part (NOT-REPRODUCED on OpenIndiana, see there); the tclEnv.c part was fixed in 2007
        • Tested see 1963074fff
        • Cause see 1963074fff
        • Prototype prototypes/1963074fff.patch (shared)
        • Suggested ticket comment The tclEnv.c warning was fixed in 2007. fCmd-27.5/27.6 pass on current OpenIndiana; the remaining no-op chown() issue is tracked in 1963074fff (proposed fix there). Closing as Duplicate.
        ⬜1829708fff2007http.test triggers segfault (sometimes)NOT-REPRODUCED (all tested)

        Open ticket 1829708fff · 49. Threading · filed 2007-11-10

        Full verdict

        NOT-REPRODUCED (all tested)

        • Platform unix
        • Tested http.test 40x each on main (own build + Thread 3.0.7), 9.0 tip tcltest + Thread 3.0.7, 8.6 tip tcltest + Thread 2.8.13 (http.test now runs its server via thread::create, testthread is gone): 40/40 pass, no crash on all three. George's variant (thread with a server socket in thread::wait, main exits immediately, with/without async thread::exit; work-a6/ht1.tcl, ht2.tcl) 40x each on main, 9.0.5, 8.6.18: all exit 0.
        • Cause teardown race when the main thread exits while other threads still run (2007 notifier); the unix notifier/finalization has been reworked since (epoll notifier, TclFinalizeNotifier, [Bug 1618235] fix).
        • Prototype none
        • Suggested ticket comment Not reproducible any more: http.test (now using the Thread package) and the exit-with-running-thread variant pass 40/40 on 9.1, 9.0 and 8.6. Closing; please reopen with a current reproducer.
        ⬜1811495fff2007clock scan (possible?) discrepancyNOT-REPRODUCED (all tested)

        Open ticket 1811495fff · 06. Time Measurement · filed 2007-10-11

        Full verdict

        NOT-REPRODUCED (all tested)

        • Platform unix
        • Tested main build, 9.0.5, 8.6.18, 8.5.19 with TZ=Europe/Vienna: clock scan 01/01/1902 = -2145920400 (+0100), the same as -timezone :Europe/Vienna, and the 1916 transition is correct. TZ=CET gives +0000 in 1902, which is correct: current tzdata links CET to Europe/Brussels, and glibc agrees.
        • Cause the reporter's own data pointed at old 8.5b1 builds (8.5.2 was right)
        • Prototype none
        • Suggested ticket comment Not reproducible with 8.5.19, 8.6.18 or 9.0.5: TZ=Europe/Vienna gives -2145920400 for 01/01/1902, matching :Europe/Vienna. Suggest closing.
        ⬜1802202fff2007Enabling DTrace support on Solaris 10 causes tkbiff failureNOT-REPRODUCED (main, 8.6 with --enable-dtrace on OpenIndiana x86-64); SPARC and 8.4 CANNOT-TEST

        Open ticket 1802202fff · 55. Other Tools · filed 2007-09-25

        Full verdict

        NOT-REPRODUCED (main, 8.6 with --enable-dtrace on OpenIndiana x86-64); SPARC and 8.4 CANNOT-TEST

        • Platform unix (Solaris 10 SPARC, --enable-dtrace)
        • Tested none; read the ticket. The reporter's later test-suite run with --enable-dtrace on 8.4.16 showed catch returning 0 for errors (append-4.9 etc.), i.e. a mis-built/miscompiled 8.4.16 dtrace build on SPARC, not a tkbiff/regexp issue.
        • Cause unknown; most likely the 8.4-era DTrace probe build (tclDTrace.o link step) on SPARC. The DTrace code has been reworked many times since 8.4.16 (EOL).
        • Prototype none (cannot build with DTrace here)
        • Suggested ticket comment 8.4 is no longer maintained. If anyone still builds with --enable-dtrace on Solaris/illumos, please run make test on 8.6.18 or 9.0.4 with and without --enable-dtrace and report the difference; otherwise this can be closed as out of date. (Main session: could try on the OpenIndiana VM — x86 only — with ./configure --enable-dtrace && make test.)
        Unix VMs, run 1
        • Platform unix (Solaris 10 SPARC, --enable-dtrace)
        • Verdict NOT-REPRODUCED (main, 8.6 with --enable-dtrace on OpenIndiana x86-64); SPARC and 8.4 CANNOT-TEST
        • Tested same DTrace builds and runs as 2815500fff. The reporter's later finding (catch returning 0 for errors in an 8.4.16 dtrace build, append-4.9 etc.) does not occur: the full suite including append.test, regexp.test and error.test gives the same results as without DTrace, also with probes enabled for error/proc/basic/append/compile.test. tkbiff itself (IMAP to an outside server) not tried.
        • Cause unknown; 8.4-era DTrace build on SPARC (same symptom as 2815500fff).
        • Prototype none
        • Suggested ticket comment Not reproducible with current sources: on OpenIndiana x86-64, main and 8.6 built with --enable-dtrace pass the test suite the same as without it, also with probes enabled. 8.4 is no longer maintained and SPARC was not available. Suggest closing as out of date.
        ⬜1771754fff2007fCmd-6.26 successNOT-REPRODUCED (all tested). The reporter had an automounter/LDAP home where $HOME differs from the canonical path; can't recreate it here.

        Open ticket 1771754fff · 36. Pathname Management · filed 2007-08-10

        Full verdict

        NOT-REPRODUCED (all tested). The reporter had an automounter/LDAP home where $HOME differs from the canonical path; can't recreate it here.

        • Platform unix
        • Tested fCmd-6.26 on main and 8.6 tip with HOME set to a symlink to the real dir (passes); old test body on 8.4.15, 8.5.0, 8.6.18 (error message uses the $HOME spelling)
        • Cause unknown, probably Tcl 8.4's [file dirname ~]/[file tail ~] vs $HOME mismatch on automounted homes. The test was rewritten for 9.0 (uses [file home] and $env(HOME)).
        • Prototype none
        • Suggested ticket comment Can't reproduce with a symlinked $HOME on 8.4.15 to 9.1, and the test has been rewritten since (no tilde in 9.0). Closing as Out of Date; reopen if it still fails with a current release.
        ⬜1752124fff2007Tcl_FSGetNormalizedPath is hairier monsterNOT-REPRODUCED (main, valgrind with a -DPURIFY build)

        Open ticket 1752124fff · 37. File System · filed 2007-07-11

        Full verdict

        NOT-REPRODUCED (main, valgrind with a -DPURIFY build)

        • Platform generic
        • Tested C program calling Tcl_FSGetNormalizedPath(NULL, Tcl_FSJoinToPath(parent, 1, &elem)) on a zero-refcount result, for absolute, ".." and relative parents. Valgrind is clean on main (work-a5/fsgnp.c, purify/ build).
        • Cause the 2007 refcount issue in tclPathObj.c no longer shows up. doc/FileSystem.3 now says Tcl_FSJoinToPath returns a zero-reference object. It doesn't explicitly warn about passing that to Tcl_FSGetNormalizedPath.
        • Prototype none
        • Suggested ticket comment Doesn't reproduce on 9.1 (valgrind with -DPURIFY, zero-refcount Tcl_FSJoinToPath result passed to Tcl_FSGetNormalizedPath). trofs was worked around in 2007. Closing as Out of Date.
        ⬜1745928fff2007thread's memory isn't being releasedNOT-REPRODUCED (main, 8.6) as a Tcl/Thread leak; a small working-set creep outside Tcl's allocator remains (see Tested)

        Open ticket 1745928fff · 80. Thread Package · filed 2007-07-01

        Full verdict

        NOT-REPRODUCED (main, 8.6) as a Tcl/Thread leak; a small working-set creep outside Tcl's allocator remains (see Tested)

        • Platform Windows (original report); generic check on Linux
        • Tested Linux main PURIFY build + Thread 3.0.7: 5x200 create/join threads, VmRSS 8364 -> 8464 K and then flat; 8.6.18 + Thread 2.8.13: 6964 -> 7040 K then flat; valgrind --leak-check=full on PURIFY main, 20 and 100 threads: 0 bytes definitely/indirectly/possibly lost. Windows script: win/1745928fff.tcl (working set via tasklist, needs Thread).
        • Cause unknown; the original numbers (Task Manager working set, dropping to 84K later) look like normal working set behaviour; vasiljevic's 2010 "huge leak in HEAD" was never pinned down.
        • Prototype none
        • Suggested ticket comment No leak on Linux with Tcl 9.1/Thread 3.0.7 or 8.6.18/Thread 2.8.13 (RSS flat over 1000 threads, valgrind clean). To be rechecked on Windows; otherwise close as out of date.
        Windows VM, run 5
        • Platform Windows
        • Verdict NOT-REPRODUCED (main, 8.6) as a Tcl/Thread leak; a small working-set creep outside Tcl's allocator remains (see Tested)
        • Tested VM win/1745928fff.tcl (5 x 200 create/join): 9.1b1 + Thread 3.0.7 +136 K, 8.6.18 + Thread 2.8.13 +556 K after warm-up. Longer run (work-win5/tm.tcl, 25 x 200): 9.1 +524 K, 8.6 +4.2 MB (~0.9 KB/thread, linear); with core testthread instead of Thread (tm2.tcl, tcltest): 9.1 +2.2 MB, 8.6 +4.9 MB over 4800 threads. Memory-debug (TCL_MEM_DEBUG) main build + Thread 3.0.7 (tm3.tcl): memory info current bytes allocated stays flat (967549) from round 1 to 4, so the creep is not in ckalloc'ed memory (OS/CRT heap or per-thread system resources; not pinned down). The ticket's numbers (one thread, ~800 K working set, dropping later) are normal working-set behaviour.
        • Cause none in Tcl found
        • Prototype none
        • Suggested ticket comment The reported numbers are normal Windows working-set behaviour. On Windows 11 a memory-debug build of 9.1 with Thread 3.0.7 shows no growth of Tcl allocations over 1000 threads; the working set creeps by well under 1 KB per thread outside Tcl's allocator. Closing as out of date.
        ⬜1706593fff2007autoloading causes (itcl) 'class already exists' errorNOT-REPRODUCED (main with Itcl 4.3.8 trunk)

        Open ticket 1706593fff · 38. Init - Library - Autoload · filed 2007-04-24

        Full verdict

        NOT-REPRODUCED (main with Itcl 4.3.8 trunk)

        • Platform generic (Itcl + auto_mkindex)
        • Tested Itcl trunk built against main; class A with a declared but unimplemented method foo, auto_mkindex, then A a; a foo: tclIndex contains only auto_index(A) (Itcl 4's auto_mkindex parser no longer indexes declared-only members) and the call fails cleanly with 'member function "::A::foo" is not defined and cannot be autoloaded', with no re-source and no "class already exists". Attachment tclbug.zip is truncated in the repository (1218 bytes, "missing 44265497 bytes"), so the exact original script was not run. Itcl 3.3 not tested.
        • Cause Itcl 3 auto_mkindex parser indexed declared-only members; Itcl's member autoload (itclMethod.c calls ::auto_load) then re-sourced the class file.
        • Prototype none
        • Suggested ticket comment Not reproducible with Itcl 4 on 9.1: declared-only methods are no longer put in tclIndex, and calling one gives a clear "not defined and cannot be autoloaded" error. The attachment is truncated. Suggest closing (it was an Itcl 3 issue anyway).
        ⬜1675258fff2007mem alloc performance on LinuxNOT-REPRODUCED (all tested)

        Open ticket 1675258fff · 41. Memory Allocation · filed 2007-03-06

        Full verdict

        NOT-REPRODUCED (all tested)

        • Platform unix (Linux)
        • Tested FOREACH/LAPPEND and EVAL/LAPPEND micro-benchmark (work-a6/lap2.tcl, source list built outside the timing) on Linux x86_64 with release builds 8.4.20, 8.5.19, 8.6.18, 9.0.5, 9.1b1: e.g. FOREACH 10000: 189 / 241 / 110 / 174 / 164 us; EVAL 10000: 39 / 14 / 14 / 29 / 31 us. No 2-3x loss vs 8.4 any more; 8.6 is fastest, 9.x somewhat slower than 8.6 (64-bit Tcl_Size lists) but not slower than 8.4.
        • Cause 2007 observation (first allocation-heavy benchmark run slow on Linux 2.6 i686, likely malloc/page-fault warm-up); not attributable to current code.
        • Prototype none
        • Suggested ticket comment On current Linux the 2007 slowdown does not show: lappend loops in 8.6/9.0/9.1 are as fast as or faster than 8.4.20. Closing as out of date.
        ⬜1593707fff2006win32 error during simultaneous file open for readNOT-REPRODUCED (main, 8.6)

        Open ticket 1593707fff · 49. Threading · filed 2006-11-09

        Full verdict

        NOT-REPRODUCED (main, 8.6)

        • Platform Windows
        • Tested Linux main + Thread 3.0.7: win/1593707fff.tcl (8 threads x 500 open/read/close + file stat/readable on one file) -> no errors (as expected on POSIX). Windows script: win/1593707fff.tcl (needs Thread; prints VERDICT).
        • Cause by reading win/tclWinChan.c TclpOpenFileChannel (main) opens with FILE_SHARE_READ|FILE_SHARE_WRITE and NativeStat (tclWinFile.c) with READ|WRITE|DELETE, so concurrent read-only opens should not conflict; transient ERROR_SHARING_VIOLATION would then come from third parties (antivirus/indexer). Delete-while-open fails by design (no FILE_SHARE_DELETE); the script reports it only for information.
        • Prototype none (reporter's retry loop is a workaround, not a fix)
        • Suggested ticket comment Concurrent read-only opens use FILE_SHARE_READ|FILE_SHARE_WRITE, and stat uses all share flags, so Tcl itself should not cause sharing violations here. To be rechecked on Windows 11 with a self-contained script; if it does not reproduce, close.
        Windows VM, run 5
        • Platform Windows
        • Verdict NOT-REPRODUCED (main, 8.6)
        • Tested VM win/1593707fff.tcl (8 threads x 500 open/read/close + file stat + file readable on one file, plus 200 delete-while-reading races) 41 times each on 9.1b1 + Thread 3.0.7 and 8.6.18 + Thread 2.8.13 (Defender on): 0 open/stat errors, 0 delete failures in all 82 runs.
        • Cause none (open uses FILE_SHARE_READ|FILE_SHARE_WRITE; see result-51.md)
        • Prototype none
        • Suggested ticket comment Not reproducible on Windows 11 with 8.6.18 and 9.1 (41 runs each of 8 threads opening, reading and stat-ing the same file 500 times): no sharing violations. Closing as works-for-me.
        ✅1567838fff2006Clock Creep using [after] on WindowsNOT-REPRODUCED (main, 9.0, 8.6; no cumulative creep over 600 s on the VM)

        Open ticket 1567838fff · 03. Timer Events · filed 2006-09-29

        ✅ Since triage (updated 2026-11-11)

        Status now: Closed (Fixed)

        • 💬 2026-10-21 serhiy.storchaka: This was a calibration problem of the Windows high-resolution clock. It has been reworked since: b6fc234ef3 (ticket b87ad7e914, in 8.6.7) fixed time drifts and jumps caused by Tcl_GetTime() calls disturbing the calibration data, and f6637d3dd8 (in 8.6.10) reworked the calibration cycle for smoother drifting toward the system time. Both are in all maintained versions. Not reproduced by me; please reopen if it…
        Full verdict

        NOT-REPRODUCED (main, 9.0, 8.6; no cumulative creep over 600 s on the VM)

        • Platform Windows
        • Tested Linux: not applicable (unix/tclUnixTime.c reads clock_gettime/gettimeofday directly, no calibration). Code reading: win/tclWinTime.c on main, 9.0 and 8.6 still derives Tcl_GetTime from QueryPerformanceCounter calibrated against the system clock by CalibrationThread, which is where a creep would come from.
        • Cause unknown; suspected drift of the calibrated performance-counter clock in win/tclWinTime.c (NativeGetTime/UpdateTimeEachSecond) relative to the system clock.
        • Prototype none
        • Repro win/1567838fff.tcl (runs the reporter's [after 500] loop and compares [clock milliseconds] with PowerShell's UtcNow every 30 s for 600 s by default; prints a VERDICT line with the drift in ms/day). Run with 9.0.5, 8.6.18 and main; a longer run (e.g. tclsh 1567838fff.tcl 3600) gives a clearer number.
        • Suggested ticket comment (after the Windows run) Tcl's clock on Windows is still the calibrated performance counter of win/tclWinTime.c; measured drift vs. the system clock over N s on Windows 11 with 9.0/8.6/main: ...
        Windows VM, run 2
        • Platform Windows
        • Verdict NOT-REPRODUCED (main, 9.0, 8.6; no cumulative creep over 600 s on the VM)
        • Tested VM win/1567838fff.tcl, 600 s, 9.1b1, 9.0.5 and 8.6.18 run concurrently: Tcl's clock minus the system clock (read through PowerShell every 30 s) wanders between −80 and +162 ms with no trend (final values +30, −43, −8 ms); the three versions move together, so the scatter is the VM clock / PowerShell latency, not Tcl. The script's "BUG REPRODUCED" line for 9.0/8.6 is its noise limit being exceeded once (162 vs 153 ms), not a creep. Caveat: in this VM QPC and the system clock drift apart by up to 2 s per minute (see b7b707a310), so a creep of a few ms/day on real hardware cannot be measured here.
        • Cause none found; the calibration in win/tclWinTime.c keeps Tcl's clock locked to the system clock over minutes
        • Prototype none
        • Suggested ticket comment On Windows 11 (VM) with 8.6.18, 9.0.5 and 9.1 an &#91;after 500&#93; loop does not make &#91;clock&#93; creep: over 10 minutes the offset to the system clock stays within ±0.16 s without a trend. The calibration code has been rewritten since 8.4; closing as out of date unless someone still sees it on real hardware.
        ⬜1502071fff2006Tcl_GetMaster can access freed interp (interp-16.1)NOT-REPRODUCED (all tested)

        Open ticket 1502071fff · 20. [interp] · filed 2006-06-07

        Full verdict

        NOT-REPRODUCED (all tested)

        • Platform generic
        • Tested ASan + -DPURIFY build of main: interp.test (391 tests) and the interp-16.1/16.2 deletion-order cases repeated with more work in the child after the parent is gone: no ASan reports. In 9.x the only core caller of Tcl_GetParent (tclCompile.c) only compares it with NULL. Not checked separately on 9.0/8.6 (same code).
        • Cause by code reading, child.parentInterp is never reset when the parent is freed, so a C extension that keeps a deleted child alive with Tcl_Preserve and calls Tcl_GetParent on it can get a dangling pointer. No script-level path dereferences it: a deleted child refuses evaluation ("attempt to call eval in deleted interpreter").
        • Prototype none (no reproducer; as msofer wrote, a C API caveat rather than a script-visible bug)
        • Suggested ticket comment No use-after-free with an ASan build in 9.1 (interp.test and variants of interp-16.x). The core only compares Tcl_GetParent() with NULL. Suggest closing unless someone has a C-level reproducer.
        ⏳1437595fff2006WinSock cores/hangs on thread or app exitNOT-REPRODUCED (9.1 main, 9.0 tip) — but REPRODUCED-like on 8.6 tip only under ephemeral TCP port exhaustion: crashes (SIGSEGV) and process-wide "s…

        Open ticket 1437595fff · 27. Channel Types · filed 2006-02-23 · Pending (resolved, waiting for confirmation)

        ⏳ Since triage (updated 2026-11-11)

        Status now: Pending (Works For Me)

          Full verdict

          NOT-REPRODUCED (9.1 main, 9.0 tip) — but REPRODUCED-like on 8.6 tip only under ephemeral TCP port exhaustion: crashes (SIGSEGV) and process-wide "sockets are not available on this system" / "winsock is not initialized"

          • Platform Windows
          • Tested Linux main build + Thread 3.0.7 with win/1437595fff.tcl (500 sequential + 10x50 concurrent threads leaving a server and client socket open, then exit with sockets open in main and 10 live threads): 5/5 clean. The Windows-specific socket thread/window teardown is not exercised on Linux.
          • Cause (historic) TclpFinalizeSockets waited on readyEvent, which was also signalled by socket activity, and InitSockets marked init done on failure; the socket subsystem was reworked since (late exit handlers, per-thread TSD), per ferrieux 2009.
          • Existing work none (Pending / Works For Me at Serhiy's request, 2026-10-12)
          • Prototype none (not reproducible)
          • Repro win/1437595fff.tcl (needs Thread; runs the stress in a child 5 times, prints VERDICT line; kills a hung child after 120 s)
          • Suggested ticket comment none needed; Serhiy's Windows test already covers it. Can be closed (8.4 is unmaintained).
          Windows VM, run 6
          • Platform Windows
          • Verdict NOT-REPRODUCED (9.1 main, 9.0 tip) — but REPRODUCED-like on 8.6 tip only under ephemeral TCP port exhaustion: crashes (SIGSEGV) and process-wide "sockets are not available on this system" / "winsock is not initialized"
          • Tested win/1437595fff.tcl (+ work-win6/1437595fff-v.tcl = same with a run count argument and printing the child output on failure), Thread 3.0.7 (9.x) / 2.8.13 (8.6). 5 runs each: 9.1 5/5 clean, 9.0 5/5 clean, 8.6 tip 1/5 failed. Then 40 runs each: 8.6 tip 5/40 failed (2 SIGSEGV, 3 exit 1), 9.1 40/40 clean, 9.0 40/40 clean; 8.6 tip + backport of a36f76eb6b (40b1814b93: wait on the socket thread handle instead of readyEvent, sock = INVALID_SOCKET after closesocket in the bind/listen loop): still 4/40 failed (2 SIGSEGV). No hang anywhere.
          • What the failures are the stress itself exhausts the ephemeral port range (Get-NetTCPConnection: ~16300 TIME_WAIT of 16384 dynamic ports). A probe (work-win6/1437-phase.tcl) shows the first error is then couldn't open socket: address already in use on socket -server ... 0, on all versions. 9.1 then simply reports that error (1 of 10 runs failed with exit 1 under full exhaustion, clean error text, no crash). 8.6 instead turns it into sockets are not available on this system (TclpHasSockets: initialized != 1, which only InitSockets' initFailure path sets, process-wide, to -1, so all later socket commands in every thread fail, also in the main thread) and winsock is not initialized, and sometimes a SIGSEGV. Without port exhaustion 8.6 runs clean (11/11 probe runs, 10/10 interleaved runs).
          • Cause 9.x: none found. 8.6: not pinned down. win/tclWinSock.c InitSockets in 8.6 still sets the global initialized = -1 when one thread's per-thread setup fails (window/thread creation), disabling sockets for the whole process; 8.7/9.0 reworked this (InitSocketWindowClass + TclInitSockets, which panics instead) and have a36f76eb6b (2025-06-25, 9.0 only, never backported to 8.6). The a36f76eb6b backport alone does not fix the 8.6 symptoms, so something else in the 8.6 per-thread init/finalize path fails under this load (likely a stale TSD/handle reuse between exiting and starting threads).
          • Prototype none (the obvious backport does not help; cause not isolated within the time box; backport diff kept at work-win6/p86.patch for reference)
          • Suggested ticket comment Not reproducible with current 9.0 and 9.1 on Windows 11 (Thread 3.0.7, 40 runs each of win/1437595fff.tcl). Can be closed. (Separately: 8.6 tip misbehaves under ephemeral port exhaustion, "sockets are not available on this system" and occasional crashes; 9.x reports "address already in use" correctly. Worth a new 8.6 ticket if 8.6 still gets such fixes.)
          ⏳1160057fff2005Seg fault from usleepNOT-REPRODUCED (main, 8.6, 8.4)

          Open ticket 1160057fff · 49. Threading · filed 2005-03-09 · Pending (resolved, waiting for confirmation)

          ⏳ Since triage (updated 2026-11-11)

          Status now: Pending (Works For Me)

            Full verdict

            NOT-REPRODUCED (main, 8.6, 8.4)

            • Platform unix
            • Tested Linux, glibc current, g++: the reporter's C++ extension (stubs-enabled, otherwise unchanged) loaded into my main build, 8.6.18 and the non-threaded 8.4.20; 20 cancellations of a thread sleeping in usleep() with a C++ destructor on its stack per run, 40 runs each -> no crash.
            • Cause not Tcl; most likely the 2005 NPTL/glibc forced-unwind of pthread_cancel in a process where libpthread/libgcc_s were not set up at startup (non-threaded Tcl loading a threaded C++ library). glibc >= 2.34 has libpthread in libc.
            • Existing work none (Pending / Works For Me at Serhiy's request, 2026-10-12)
            • Prototype none (not reproducible)
            • Suggested ticket comment none needed; the resolution holds (also with a non-threaded 8.4.20 on current glibc), can be closed.
            ⬜1080650fff2004file normalize + env(HOME)=C:NOT-REPRODUCED (main, 9.0, 8.6)

            Open ticket 1080650fff · 08. Environment Variables · filed 2004-12-07

            Full verdict

            NOT-REPRODUCED (main, 9.0, 8.6)

            • Platform Windows
            • Tested Linux only (not applicable: no drive-relative paths). vincentdarley reported it fixed in 8.5a2. In 9.x ~ is no longer expanded by file normalize (TIP 602); the repro tries file join $env(HOME) tclshrc.tcl, c:tclshrc.tcl, and ~/tclshrc.tcl (8.x) / file tildeexpand (9.x), and checks that normalize is idempotent and absolute. Repro: win/1080650fff.tcl.
            • Cause unknown (8.4 tilde substitution produced a drive-relative path that normalize did not resolve)
            • Prototype none (needs Windows)
            • Suggested ticket comment Needs a test on Windows with 9.0/8.6; reported fixed in 8.5 already, and in 9.x &#91;file normalize&#93; no longer expands ~.
            Windows VM, run 2
            • Platform Windows
            • Verdict NOT-REPRODUCED (main, 9.0, 8.6)
            • Tested VM win/1080650fff.tcl with env(HOME)=c:, cwd %USERPROFILE%\tmp\tri2: file join $env(HOME) tclshrc.tcl, c:tclshrc.tcl, and ~/tclshrc.tcl (8.6) / file tildeexpand ~ (9.x) all normalize to the absolute %USERPROFILE%/tmp/tri2/tclshrc.tcl, and normalizing again gives the same path, on 9.1b1, 9.0.5 and 8.6.18
            • Cause fixed long ago (reported fixed in 8.5a2)
            • Prototype none
            • Suggested ticket comment Works on 8.6.18, 9.0.5 and 9.1 on Windows 11: with env(HOME)=c: the drive-relative path normalizes to an absolute path under the current directory of C:, and normalize is idempotent. Closing as fixed.
            ⏳994748ffff2004Tcl_RegisterChannel: duplicate channel namesNOT-REPRODUCED (main, 9.0, 8.6)

            Open ticket 994748ffff · 25. Channel System · filed 2004-07-20 · Pending (resolved, waiting for confirmation)

            ⏳ Since triage (updated 2026-11-11)

            Status now: Pending (Works For Me)

              Full verdict

              NOT-REPRODUCED (main, 9.0, 8.6)

              • Platform generic
              • Tested Linux, Xvfb: wish 9.1 (main install), 9.0.5, 8.6.18; 300 open client sockets, then 300 more created from a <Destroy> binding of . during exit -> all created, clean exit (rc 0) on all three.
              • Cause socket channel names are now sock<address of state> on Unix (SOCK_TEMPLATE in unix/tclUnixSock.c) and Windows (TclWinGenerateChannelName(..., "sock", statePtr)), so a live channel name cannot be reused; the old counter-based names could wrap/collide.
              • Existing work none (Pending / Works For Me at Serhiy's request, 2026-10-12)
              • Prototype none (not reproducible)
              • Suggested ticket comment none needed; the resolution holds (also confirmed on Linux with 9.0.5), can be closed.
              ⬜821499ffff2003Tcl_AppendToObj disobeys lengthNOT-REPRODUCED (all tested) — Tcl_AppendToObj(obj, "\302\242", 1) appends exactly one byte, both to a pure-Unicode object (reporter's case) and to …

              Open ticket 821499ffff · 10. Objects · filed 2003-10-10

              Full verdict

              NOT-REPRODUCED (all tested) — Tcl_AppendToObj(obj, "\302\242", 1) appends exactly one byte, both to a pure-Unicode object (reporter's case) and to a plain string object

              • Platform generic
              • Tested C program (work-a3/821499.c) against main build, 9.0.5, 8.6.18, and old 8.4.2/8.4.11/8.4.15/8.4.20, 8.5.0/8.5.5/8.5.10/8.5.19, 8.6.0–8.6.14. Up to 8.6.12 the lone byte decodes as 0xFFC2 (sign-extended), from 8.6.13 on as 0xC2; never as 0xA2, so the scan never runs past the given length
              • Cause the 2003 over-scan in the Unicode-rep append path is gone (it did not show even in 8.4.2)
              • Prototype none (not reproducible)
              • Suggested ticket comment Cannot reproduce with 8.4.2 through 9.1: the length is honoured and the lone lead byte becomes U+00C2 (U+FFC2 before 8.6.13). Proposing to close as out of date.
              ⬜803370ffff2003marked async handler not getting invokedNOT-REPRODUCED (Linux: 8.6 unthreaded, 8.6 threaded, main); HP-UX CANNOT-TEST

              Open ticket 803370ffff · 01. Notifier · filed 2003-09-09

              Full verdict

              NOT-REPRODUCED (Linux: 8.6 unthreaded, 8.6 threaded, main); HP-UX CANNOT-TEST

              • Platform unix (HP-UX 10.20 report)
              • Tested small extension (work-a4/sigext/sigasync.c) whose SIGALRM handler calls Tcl_AsyncMark(), with the interp idle in vwait: the handler runs after 100 ms (the alarm time) on an unthreaded 8.6 tip build (--disable-threads, select() returns EINTR), threaded 8.6.18 and main. Also testasync marklater from another thread wakes an idle vwait in 0.1 s on main/9.0/8.6. Tcl 9 has no unthreaded build any more.
              • Cause on the reporter's system the signal apparently didn't interrupt select() (another thread got it, or the select restarts). Unthreaded Tcl_AlertNotifier still does nothing, but on Linux EINTR covers that.
              • Prototype none. Side finding: tclTest.c AsyncHandlerProc calls Tcl_GetStringResult(interp) before its interp != NULL check, so a testasync handler that fires while idle in vwait (Tcl_ServiceEvent -> Tcl_AsyncInvoke(NULL, 0)) segfaults the tcltest binary (main, 9.0, 8.6). Test-only, a one-line fix.
              • Suggested ticket comment Can't reproduce on Linux: a signal handler calling Tcl_AsyncMark wakes an idle event loop in 8.6 (threaded and unthreaded) and 9.x (always threaded), because select() returns EINTR. HP-UX 10.20 is long gone; closing as out of date.
              ⬜717848ffff2003Notifier lock in sub-process on windowsNOT-REPRODUCED (main, 9.0, 8.6)

              Open ticket 717848ffff · 01. Notifier · filed 2003-04-08

              Full verdict

              NOT-REPRODUCED (main, 9.0, 8.6)

              • Platform Windows
              • Tested the attached main.tcl/sub.tcl (original sub.tcl, artifact d6fb4fa87a, with fileevent stdin readable on the never-written pipe) plus a retrying client: the subprocess accepts, prints its args and [info hostname] on main, 9.0, 8.6 (Linux). Self-contained repro win/717848ffff.tcl (the child also checks that a later after fires) gives VERDICT: OK on all three on Linux.
              • Cause unknown. On Windows the stdin pipe reader thread of the child plus the [info hostname] winsock init in the accept callback hung the notifier in 8.4.
              • Prototype none
              • Suggested ticket comment Not reproducible on Unix. Please run win/717848ffff.tcl-style repro (child with fileevent stdin readable, [info hostname] and after in the accept callback) with 8.6.18/9.0 on Windows; close if it passes.
              Windows VM, run 3
              • Platform Windows
              • Verdict NOT-REPRODUCED (main, 9.0, 8.6)
              • Tested VM win/717848ffff.tcl (child tclsh with stdin as a never-written pipe and fileevent stdin readable; accept callback runs [info hostname] and an after): 5 of 5 runs OK on each of 9.1b1, 9.0.5, 8.6.18
              • Cause the 8.4-era hang between the stdin pipe reader thread and winsock init no longer happens
              • Prototype none
              • Suggested ticket comment Not reproducible on Windows 11 with 8.6.18, 9.0.5 and 9.1: a child with &#91;fileevent stdin readable&#93; on a pipe accepts a connection, runs &#91;info hostname&#93; and later timers fine. Closing as out of date.
              ⬜708166ffff2003dde servers can hang on exitNOT-REPRODUCED (main, 9.0, 8.6) for the hang on exit; registering a server name while another DDE server is blocked takes about 2 s

              Open ticket 708166ffff · 28. dde Package · filed 2003-03-22

              Full verdict

              NOT-REPRODUCED (main, 9.0, 8.6) for the hang on exit; registering a server name while another DDE server is blocked takes about 2 s

              • Platform Windows
              • Tested code reading only (main win/tclWinDde.c): DdeExitProc still calls DdeNameService(ddeInstance, NULL, 0, DNS_UNREGISTER), which DDEML implements by sending (not posting) messages to every DDE window, so it blocks on a server that does not pump messages. The dde package does not exist on Unix.
              • Cause win/tclWinDde.c DdeExitProc -> DdeNameService(DNS_UNREGISTER) (and probably the DNS_REGISTER in DdeSetServerName) synchronously broadcasting to a non-responsive DDE server.
              • Prototype none (needs Windows to confirm; possible approach: skip the global unregister in the exit handler, or do it on a helper thread with a timeout)
              • Repro win/708166ffff.tcl (starts server A that blocks 20 s without pumping messages, then times a second server B registering and exiting; VERDICT line, >5 s = reproduced; it reports the registration time separately).
              • Suggested ticket comment DdeExitProc still calls DdeNameService(DNS_UNREGISTER) unconditionally; needs a recheck on current Windows with the attached-style repro (blocking server A, server B exits).
              Windows VM, run 4
              • Platform Windows
              • Verdict NOT-REPRODUCED (main, 9.0, 8.6) for the hang on exit; registering a server name while another DDE server is blocked takes about 2 s
              • Tested VM win/708166ffff.tcl over ssh (server A sets a servername and blocks 20 s without pumping messages; server B registers and exits): B's dde servername takes 2028–2032 ms, B's exit 13–16 ms, on 9.1b1, 9.0.5, 8.6.18. Without a blocked server, dde servername takes 0 ms (work-win4/ddeb.tcl).
              • Cause exit: DdeNameService(DNS_UNREGISTER) no longer blocks on a non-responsive DDE window on Windows 11. Registration: DdeSetServerName -> DdeGetServicesList broadcasts WM_DDE_INITIATE with SendMessageTimeoutW(SMTO_ABORTIFHUNG, 1000 ms), so a blocked server delays it by the timeout but cannot hang it.
              • Prototype none
              • Suggested ticket comment Not reproducible on Windows 11 with 8.6.18, 9.0.5 and 9.1: with a DDE server that does not pump messages, a second server exits at once; only &#91;dde servername&#93; is delayed by about 2 s (the SendMessageTimeout in the service enumeration). Closing as out of date.
              ⬜636161ffff2002Tcl_FS* problems on windows&threadsNOT-REPRODUCED (main, 9.0, 8.6)

              Open ticket 636161ffff · 25. Channel System · filed 2002-11-10

              Full verdict

              NOT-REPRODUCED (main, 9.0, 8.6)

              • Platform Windows (tclkit + tclvfs + Thread, 8.4.1)
              • Tested main + Thread 3.0.7 on Linux: 40 threads each source a file (40/40), and the same from a zipfs mount, which is Tcl 9's built-in counterpart of a tclkit VFS (40/40). Repro win/636161ffff.tcl does the same, prints VERDICT and exits; a crash shows up as no verdict line.
              • Cause unknown. The comments point to tclvfs/tclkit (plain shells were fine, according to hobbs).
              • Prototype none
              • Suggested ticket comment source in many threads works on 9.x, also from a zipfs mount. The report is about 8.4 tclkit+tclvfs; unless it still fails with a current Windows build (win/636161ffff.tcl-style test), close as out of date.
              Windows VM, run 3
              • Platform Windows (originally tclkit + tclvfs, 8.4.1)
              • Verdict NOT-REPRODUCED (main, 9.0, 8.6)
              • Tested VM win/636161ffff.tcl, 40 thread::create threads each source a file: 40/40 on 9.1b1, 9.0.5, 8.6.18 (3 runs each) and the 9.1 memdebug Debug build; also from a zipfs mount (Tcl 9's built-in VFS) 40/40 on 9.1b1 and 9.0.5
              • Cause n/a (the report involved tclkit/tclvfs; plain shells were already fine in 2002)
              • Prototype none
              • Suggested ticket comment &#91;source&#93; in many threads works on Windows 11 with 8.6.18, 9.0.5 and 9.1, also from a zipfs mount (memdebug build clean). The report is about 8.4 tclkit + tclvfs; closing as out of date.
              ⬜576158ffff2002Unkown floating-pointer errno=4NOT-REPRODUCED (all tested)

              Open ticket 576158ffff · 07. Variables · filed 2002-07-01

              Full verdict

              NOT-REPRODUCED (all tested)

              • Platform unix
              • Tested the attachment (bug_test.tgz, stored double-UTF-8-encoded in the repo; recoverable by decoding to Latin-1) needs wish + a SWIG module with a SIGALRM thread. Equivalent tclsh test (work-a3/576158/run.tcl + alarm.c: 1 kHz SIGALRM without SA_RESTART, 200000 iterations of the reporter's unbraced [expr $count * 2.5] plus sin() from event callbacks): 0 errors on main, 9.0, 8.6, 8.4.20, 8.4.2. 8.3.3 not available
              • Cause in 8.3, expr parsed numbers with strtod() and checked a stale errno (EINTR from an interrupted syscall). Since 8.5 numbers are parsed by Tcl's own TclParseNumber; errno is only checked after libm calls, and generic/tclBasic.c clears it (errno = 0) right before each call
              • Prototype none
              • Suggested ticket comment Cannot reproduce on 8.4–9.1 with a SIGALRM stress test equivalent to the attachment. Since 8.5 expr no longer uses strtod/errno to parse numbers, and math functions clear errno before each call. Proposing to close as out of date.
              ⬜502022ffff2002Tcl_AsyncMark doesn't awaken the NotifieNOT-REPRODUCED (main, 9.0, 8.6 threaded builds); the non-threaded case is gone in 9.x

              Open ticket 502022ffff · 01. Notifier · filed 2002-01-10

              Full verdict

              NOT-REPRODUCED (main, 9.0, 8.6 threaded builds); the non-threaded case is gone in 9.x

              • Platform unix
              • Tested testasync marklater (marks from another thread after 1 s) during vwait with a 3 s timeout, using the tcltest of main, 9.0 tip and 8.6 tip: the handler runs after about 1 s inside vwait (Tcl_DoOneEvent -> Tcl_ServiceEvent -> Tcl_AsyncInvoke). Non-threaded 8.6 judged by reading the code only.
              • Cause Tcl_AsyncMark calls Tcl_ThreadAlert -> Tcl_AlertNotifier, which wakes the threaded notifier. Tcl 9 has no non-threaded build any more. In an 8.6 --disable-threads build Tcl_AlertNotifier is still a no-op, but a mark made from a signal handler interrupts select() (EINTR) and Tcl_ServiceEvent then runs the async handlers. Only a mark from a foreign OS thread in a non-threaded build stays unnoticed, and that case is outside the non-threaded model anyway.
              • Side finding the test run crashes in generic/tclTest.c:AsyncHandlerProc, which calls Tcl_GetStringResult(interp) before its interp != NULL check. When the event loop invokes async handlers, interp is NULL (segfault in Tcl_GetObjResult, all of main/9.0/8.6). This is test code only, but async handlers invoked from the event loop are never tested.
              • Prototype none (no core bug; the tclTest.c NULL-interp crash could be a separate 2-line fix)
              • Suggested ticket comment Tcl 9 is always threaded, so Tcl_AsyncMark always wakes the notifier through Tcl_ThreadAlert. In threaded 8.6/9.0/9.1 an async mark from another thread is handled inside &#91;vwait&#93; right away. Can be closed.
              ⬜219307ffff2000Tcl's fileevent callback on a socket does not work the sameNOT-REPRODUCED (main, 9.0, 8.6)

              Open ticket 219307ffff · 27. Channel Types · filed 2000-10-26

              Full verdict

              NOT-REPRODUCED (main, 9.0, 8.6)

              • Platform Windows
              • Tested main, 9.0.5, 8.6.18 on Linux with win/219307ffff.tcl (localhost, ephemeral port). The server closes at once; the client's writable callback fires, the first write succeeds, the second fails with broken pipe.
              • Cause on Unix the writable event fires for a socket in an error/closed state, as documented (fileevent: "error conditions" make the channel readable/writable). The ticket says Windows didn't fire at all after the peer closed; win/tclWinSock.c would have to be checked for FD_CLOSE turning into a writable event.
              • Prototype none
              • Repro win/219307ffff.tcl (prints a VERDICT line, exits within 3 s)
              • Suggested ticket comment On Unix the behaviour is as documented: writable fires and the write reports "broken pipe" (main, 9.0, 8.6). Checking whether Windows still fails to fire the writable callback after the peer closes.
              Windows VM, run 3
              • Platform Windows
              • Verdict NOT-REPRODUCED (main, 9.0, 8.6)
              • Tested VM win/219307ffff.tcl (server closes the accepted socket at once; client has a nonblocking writable handler): on 9.1b1, 9.0.5 and 8.6.18 the writable callback fires, the first write succeeds, the second fails with software caused connection abort — the same sequence as on Linux (there broken pipe)
              • Cause n/a (the 1999 report; only the OS error text differs)
              • Prototype none
              • Suggested ticket comment Windows 11 now behaves like Unix (8.6.18, 9.0.5, 9.1): the writable handler fires on a socket whose peer closed it and the write reports the error ("software caused connection abort" vs. "broken pipe"). Closing as out of date.