| Ticket UUID: | b5777d3d3290348abc980027a96607ccfe4bf4da | ||
| Title: | Crash corner-case in TIP #502 implementation | ||
| Type: | Bug | Created on: | 2020-08-25 16:03:24 |
| Submitter: | jan.nijtmans | Assigned to: | dgp |
| Subsystem: | 13. Index Object | Severity: | Minor |
| Priority: | 5 Medium | Last modified: | 2026-10-22 17:19:39 |
| Status: | Closed | Closed by: | serhiy.storchaka |
| Resolution: | Fixed | Closed on: | 2026-10-21 11:09:22 |
| Version: | 8.7 | ||
| Description: | ||||
|
Try the following: $tclsh8.7 % string index abcd 0-0x10000000000000000 % tailcall cannot find the right splicing spot: should not happen! Abort trap: 6 (press "enter" twice after typing the command). I could re-produce this both on Windows 10 and on MacOS. Bisection shows that this crash is already happening since commit [bae05b9e6f], which is the TIP #502 implementation. Before that, the result was the same as in Tcl 8.6: $ tclsh8.6 % string index abcd 0-10000000000000000 bad index "0-10000000000000000": must be integer?[+-]integer? or end?[+-]integer? | ||||
| User Comments: | ||||
jan.nijtmans added on 2020-08-25 16:12:01:
Branch [bug-b5777d3d32] contains a test-case for this bug. Even though it's a corner-case very unlikely to be encounter by anyone (but ... I did ... by pure chance ...), it's a crash waiting to happen. dgp added on 2020-09-01 22:28:23:
Fix committed to the bug fix branch to solve the reported problem, but not the deeper brokenness that underlies it. dgp added on 2020-09-02 16:25:10:
More direct fix is in dgp-nested-execution. Something like this may need to be done in many other places in the bytecode execution engine for full safety against re-entry into the execution engine. dgp added on 2020-09-02 17:03:52:
Similar issue was addressed in https://core.tcl-lang.org/tcl/tktview/3138178 jan.nijtmans added on 2020-09-16 06:20:53:
I'm wondering if the stack protection isn't really overkill here. How much performance does it cost in the normal case? Therefore keeping open for now. The bug-b5777d3d32 branch undoes the stack protection changes and simply uses a separate interpreter for this case always. Or .. maybe .. just parse the two parts as bigintegers and add/subtract them using libtommath's functions. dgp added on 2020-09-16 14:24:10:
Any of those approaches will work. I don't object to any of them. All have their pros and cons. The broader issue is that we have this point of fragility in the execution engine. Any function call out of TEBCresume that could possibly lead to nested bytecode execution is at risk without this protection. That set of routines isn't stable as development continues. Anything that permits extension callbacks should get protection, and with customizable traces and resolvers, there are many paths that can potentially do that. A "no compromise on safety" approach would protect every call, or at least every call that passes interp as an argument. The (DE)CACHE_STACK_INFO macros themselves are tiny. Maybe I am missing something, but the main effect of them seems to be that they force the more expensive implementation of the INST_START_CMD instruction. serhiy.storchaka added on 2026-10-21 11:09:22:
The crash was fixed in 9.0.0 by [fca1a2a297], with test util-9.59 (passes on trunk and 9.0). 8.6 is not affected, since TIP #502 is 8.7+. The broader question of protecting other re-entry points of the bytecode engine (see branch dgp-nested-execution) can get a ticket of its own if someone wants to pursue it. | ||||
