Toolset update: MSVC Compiler 19.51.36014

[StephanTLavavej]

Changelog

• Merged C++23 features:
  • P2255R2 Type Traits To Detect References Binding To Temporaries
  • P2674R1 is_implicit_lifetime
    • Initially supported for MSVC and Clang only. (EDG support will be enabled when the necessary compiler builtins become available.)
• Improved test coverage:
  • Added ARM64 ASan test coverage.
• Code cleanups:
  • Removed compiler bug workarounds.
• Infrastructure improvements:
  • Updated dependencies.
    • Updated MSVC Compiler to 19.51.36014 (now required).
    • Updated CMake to 4.2.1 (now required).
    • Updated Python to 3.14.3 (now required).

Commits

• Cycle locations for ARM64.
• Python 3.14.3.
• CMake 4.2.1.
• MSVC Compiler 19.51.36014.
  • Not yet required in yvals_core.h; the internal toolset is still 19.50.35721.
• README.md: Install "(Preview)" Build Tools.
• README.md: Update Command Prompt instructions, adding -vcvars_ver=preview.
• provision-image.ps1: Update to Preview components.
• provision-image.ps1: Patch Launch-VsDevShell.ps1.
• CMakePresets.json: Set "toolset" with a "value" of "version=preview".
  • This teaches the IDE to pick up the Preview toolset. Thanks to Vivy in the STL Discord.
• Use 10 shards for x64/x86 too.
• VSO-2574432 "C1XX incorrectly handles a requires-expression checking for r.operator T();" was fixed on 2025-09-13.
• VSO-1513409 "Bogus warning C4100: '<_Args_0>': unreferenced formal parameter when if constexpr selects another branch" was fixed on 2025-06-27.
• EDG 6.8 implemented multidimensional subscript.
• Enable WG21-P2674R1 is_implicit_lifetime in C++23 mode for MSVC.
  • This was MSVC-PR-675420 on 2025-10-01.
• Enable WG21-P2255R2 reference_meows_from_temporary in C++23 mode for MSVC and EDG.
  • This was MSVC-PR-682362 on 2025-10-27 and VS-PR-688674 on 2025-11-18. CUDA is limited to C++20, so we don't need to worry about it. Reported VSO-2742607 "C1XX __reference_constructs_from_temporary mishandles a couple of cases". Use strict_latest_matrix.lst because there are (predictably) lots of errors in permissive mode due to the Evil Extension.
• MSVC implemented WG21-P2448R2 "Relaxing some constexpr restrictions" with MSVC-PR-686379 on 2025-12-03.
• MSVC fixed a constexpr bug!
• Reported [libc++][test] facet.num.get.members/get_long.pass.cpp uses std::to_string() without including <string> llvm/llvm-project#182390.
• Reported [libc++][test] vector/trivial_relocation.pass.cpp uses implementation-specific attribute [[clang::trivial_abi]] llvm/llvm-project#182392.
• Reported [libc++][test] rand.dist.bern.negbin/eval.pass.cpp constructs with p == 1, which has undefined behavior llvm/llvm-project#182397.
• Template parameters: _Pc => _ParseContext
  • This and the following commits are fixing widespread naming inconsistency, and will make it easier to remove the compiler bug workaround.
• Function parameters: _Pc => _Parse_ctx
• Template parameters: _ParseCtx => _ParseContext
• Function parameters: _ParseCtx => _Parse_ctx
• Function parameters: _FormatCtx => _Format_ctx
• Template parameters: _FmtCtx => _FormatContext
• _Fill_align_and_width_formatter::_Format() shouldn't be constexpr.
  • This was bogus because _Fill_align_and_width_formatter_format() was correctly non-constexpr.
• _Fill_align_and_width_formatter::_Format() return type: auto => _FormatContext::iterator
• Update citations: VSO-1236041 => VSO-2744645
  • VSO-1236041 was reportedly fixed on 2025-09-08, but it didn't entirely work. I've reported VSO-2744645 "Standard Library Header Units: Declaring and defining constexpr member functions in different headers fails to compile" which should be a very easy fix for @xiangfan-ms according to @cdacamar.
• Drop _NODISCARD on workaround functions to match their callers.
  • This will reduce code churn when we unify them.
• <format>: Remove workarounds for declaring and defining member functions in different headers.
  • For pair's piecewise constructor, we used _Tuple_get() as a perma-workaround, and I don't want to change that now. Note that I'm transferring an _STL_INTERNAL_STATIC_ASSERT in _Tuple_formatter_common_base::_Format().
• libcxx/expected_results.txt: Overhaul the "XFAILS WHICH PASS" section.
  • Now that we have full ARM64 test coverage, and we're enabling ARM64 ASan, this needs to be updated. Transfer the tests that actually contain "XFAIL: msvc" to the section for the MSVC-internal test harness. A couple of tests don't contain "XFAIL: msvc" (anymore?) and are simply passing, so drop their mentions completely.
• Partially enable ARM64 ASan, add test coverage.
  • And mention ASan in the ASan pipeline's stage and display names. Add dependsOn: [] to the x64 stage for symmetry, following the example of the Code Format stage in azure-pipelines.yml. Update comment in azure-pipelines.yml to follow our preferred arch order. Not enabling GH_002030_asan_annotate_string and GH_002030_asan_annotate_vector yet due to Clang issues.
• New pools.
• Upgrade all "constexpr step limits" tests to SKIPPED.
  • And remove the few :2 (Clang-specific) configs; if Clang has trouble, C1XX is very likely to also have trouble. vector.bool/flip.pass.cpp:0 was unexpectedly passing for x86. Playing whack-a-mole is just not worth my time.

Testing

STL-ASan-CI passed for x64, x86, and ARM64.