KWSys: use allocator_traits::rebind_alloc for C++20 (hashtable.hxx)

[gdevenyi]

Problem

std::allocator::rebind was deprecated in C++17 and removed in C++20. On newer toolchains (e.g. GCC 14/15 shipping in Fedora 44) libstdc++ no longer provides std::allocator<T>::rebind, so the vendored KWSys hashtable.hxx template fails to compile:

error: no type named 'other' in 'std::allocator<...>::rebind<...>'

This breaks any consumer that pulls in itksys/hash_map.hxx / hash_set.hxx when built in C++20 mode.

Fix

Replace the three typename _Alloc::template rebind<X>::other typedefs in Modules/ThirdParty/KWSys/src/KWSys/hashtable.hxx.in with the standard C++11 replacement:

typename std::allocator_traits<_Alloc>::template rebind_alloc<X>

std::allocator_traits has been the portable way to rebind since C++11 and works across C++11 → C++20. <memory> is already included by this header, so no new include is needed. No behavior change on older standards.

Only hashtable.hxx.in uses rebind; hash_map.hxx.in / hash_set.hxx.in are unaffected.

Notes

• This file is vendored from upstream KWSys (Kitware/KWSys); the same fix likely belongs there too, but this targets release-4.14 directly since that branch is pinned by downstream consumers (e.g. minc-toolkit-v2) and currently fails on Fedora 44.
• Not yet verified by a full local Fedora 44 build; this is the canonical, mechanical C++20 migration of the construct.

🤖 Generated with Claude Code

[greptile-apps]

Greptile Summary

This PR fixes a C++20 compilation failure in the vendored KWSys hashtable.hxx.in by replacing the deprecated (C++17) and removed (C++20) typename _Alloc::template rebind<X>::other pattern with the portable std::allocator_traits<_Alloc>::template rebind_alloc<X> form introduced in C++11.

• All three rebind sites in hashtable.hxx.in are updated; <memory> is already included so no new dependency is needed.
• hash_map.hxx.in and hash_set.hxx.in are unaffected by this change as they do not use rebind directly.

Confidence Score: 5/5

Safe to merge — the change is a mechanical, standard-compliant substitution with no behavioral difference on any supported allocator.

All three rebind usages are updated, <memory> is already included providing std::allocator_traits, and rebind_alloc is defined to consult _Alloc::rebind<X>::other first (when present) so custom allocators remain unaffected. No logic, data layout, or ABI change is introduced.

No files require special attention. The single changed template file is a clean, complete fix.

Important Files Changed

Filename	Overview
Modules/ThirdParty/KWSys/src/KWSys/hashtable.hxx.in	Replaces all three _Alloc::rebind<X>::other typedefs with the C++11-standard std::allocator_traits<_Alloc>::rebind_alloc<X>; fix is complete and correct — no remaining old-style rebind usages, <memory> already included.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["_Alloc (template param)"] --> B{"Standard used"}
    B -->|"C++11 - C++17 (old code)"| C["_Alloc::rebind<X>::other\n(deprecated C++17, removed C++20)"]
    B -->|"C++11 - C++20 (new code)"| D["std::allocator_traits<_Alloc>::rebind_alloc<X>\n(portable since C++11)"]
    D --> E["allocator_type = rebind_alloc<_Val>"]
    D --> F["_M_node_allocator_type = rebind_alloc<_Node>"]
    D --> G["_M_node_ptr_allocator_type = rebind_alloc<_Node*>"]
    C -->|"GCC 14+ / C++20 mode"| H["Compile error: no type named 'other'"]
    D -->|"All compilers / C++11-C++20"| I["Compiles successfully"]

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    A["_Alloc (template param)"] --> B{"Standard used"}
    B -->|"C++11 - C++17 (old code)"| C["_Alloc::rebind<X>::other\n(deprecated C++17, removed C++20)"]
    B -->|"C++11 - C++20 (new code)"| D["std::allocator_traits<_Alloc>::rebind_alloc<X>\n(portable since C++11)"]
    D --> E["allocator_type = rebind_alloc<_Val>"]
    D --> F["_M_node_allocator_type = rebind_alloc<_Node>"]
    D --> G["_M_node_ptr_allocator_type = rebind_alloc<_Node*>"]
    C -->|"GCC 14+ / C++20 mode"| H["Compile error: no type named 'other'"]
    D -->|"All compilers / C++11-C++20"| I["Compiles successfully"]

Loading

_{Reviews (2): Last reviewed commit: "KWSys: use allocator_traits::rebind_allo..." | Re-trigger Greptile}

[dzenanz]

Has something similar already been applied upstream? Should we update KWSys in 4.14 branch, instead of patching it like this?

[gdevenyi]

Has something similar already been applied upstream? Should we update KWSys in 4.14 branch, instead of patching it like this?

I'm happy to do that instead, I'm not sure how to make it happen

[blowekamp]

Has something similar already been applied upstream? Should we update KWSys in 4.14 branch, instead of patching it like this?

I'm happy to do that instead, I'm not sure how to make it happen

You could try using the skill: https://github.com/InsightSoftwareConsortium/ITK/tree/main/.github/skills/update-third-party

[gdevenyi]

Has something similar already been applied upstream? Should we update KWSys in 4.14 branch, instead of patching it like this?

I'm happy to do that instead, I'm not sure how to make it happen

You could try using the skill: main/.github/skills/update-third-party

Thanks I'll take a look

[gdevenyi]

Verified locally: the failure reproduces on GCC 15 / libstdc++ in C++20 mode (CachyOS), not only Fedora 44 — itksys/hashtable.hxx:252/256/258: error: no class template named 'rebind' in 'class std::allocator<char>' when a downstream target (here ITK's KWSys consumed by Convert3D) instantiates the hashtable. After applying this allocator_traits<_Alloc>::rebind_alloc change to the generated header, the compile passes. So this is a general C++20 toolchain issue, not distro-specific.

[gdevenyi]

Closing pending #6515

[gdevenyi]

#6515 isn't a simple solution sadly.

[dzenanz]

Looks good on a glance.

[blowekamp]

O, I missed this was targeting ITK 4.14!

[gdevenyi]

O, I missed this was targeting ITK 4.14!

Yup. I have zombie projects I'm performing necromancy on to keep going.

[hjmjohnson]

Thanks for keeping these old branches alive, @gdevenyi — the necromancy is appreciated.

I'm fine with the change in principle: std::allocator_traits<_Alloc>::rebind_alloc<X> is the standard-blessed replacement and is backward compatible across C++11 → C++20, so nothing built with a C++11/14/17 toolchain regresses. The one caveat worth naming is that std::allocator_traits is itself a C++11 construct — it doesn't exist in C++03 — so strictly speaking this trades the old C++03-compatible rebind<>::other for something that assumes C++11. Almost certainly academic (nobody is building release-4.14 with a pre-C++11 compiler today), but it matters given the 4.x line nominally kept a C++03 baseline.

Which is my larger question: rather than retrofitting C++03-era KWSys for C++20, why not build 4.14 with a C++ standard appropriate to its era? A downstream pinning release-4.14 (minc-toolkit-v2, Convert3D, …) can just pin -DCMAKE_CXX_STANDARD=14 (or gnu++14), which sidesteps this and the rest of the C++17/20 removals you'll otherwise keep hitting one at a time in vendored third-party code. That feels more sustainable than patching each construct as a new toolchain removes it.

If the goal really is "compile clean in C++20 mode," then I agree with @dzenanz that refreshing KWSys from upstream (#6515) is the right long-term fix rather than a local patch — though I see that turned out to be non-trivial.

[gdevenyi]

A downstream pinning release-4.14 (minc-toolkit-v2, Convert3D, …) can just pin -DCMAKE_CXX_STANDARD=14 (or gnu++14),

Thanks @hjmjohnson I was able to implement this and get things working that way.