Toolset update: MSVC Compiler 19.52.36725, large Fadsv7 pool

[Stephan T. Lavavej (StephanTLavavej)]

This makes our CI system simpler and faster now that we have sufficient quota for a large Fadsv7 pool (non-SMT Zen 5, temp storage, can use 80-core VMs).

• Run 'Validate Files' before 'clang-format Files', comment why.
• create-1es-hosted-pool.ps1: Add $DiskType to handle NVMe vs. SCSI.
  • This will be useful in the future when ARM64 SKUs support NVMe.
• create-1es-hosted-pool.ps1: Fadsv7 quota increase from 21 48-core VMs to 23 64-core VMs.
• create-1es-hosted-pool.ps1: Drop Fasv6 and Fasv7.
• create-1es-hosted-pool.ps1: Fadsv7 quota increase to 32 80-core VMs.
• Unify x64 Slow/Medium/Fast pools.
• Cycle locations.
• provision-image.ps1: Detect errors from curl.exe.
  • If downloads fail, this will make it a lot more obvious why.
• provision-image.ps1: Reorder and use long options for curl.exe.

  -f, --fail           Fail fast with no output on HTTP errors
  -s, --silent         Silent mode
  -S, --show-error     Show error even when -s is used
  -L, --location       Follow redirects
  -o, --output <file>  Write to file instead of stdout
  

• create-1es-hosted-pool.ps1: Set $PSNativeCommandUseErrorActionPreference just in case.
  • We don't have any native commands right now.
• PowerShell 7.6.6.
• MSVC Compiler 19.52.36725.
• New pools.
• MSVC-PR-766982 renamed /experimental:mathlib to /Zc:cmath.
• Cite MSVC-PR-777572 which was merged 2026-09-03.
• MSVC-PR-767184/MSVC-PR-772024 (merged 2026-08-19/2026-08-20) fixed LNK2005 for frexp(), before Overhaul constexpr <cmath> #6434 (merged 2026-09-04) used inline namespaces to allow mixing.
• MSVC-PR-767404 (merged 2026-08-07) fixed the comparison functions in constant evaluation and fpclassify for subnormal floats.
• MSVC-PR-767414/MSVC-PR-768260 (merged 2026-08-07/2026-08-10) fixed LNK2019 for ARM64EC.
• Overhaul constexpr <cmath> #6434 (merged 2026-09-04) removed the #pragma function stuff that made /O2 coverage necessary.
• MSVC-PR-771026 (merged 2026-08-19) added /std:c++23 and deprecated /std:c++23preview.
• MSVC-PR-771455 "Mark noreturn intrinsics as cold code" was merged 2026-08-26, so after __fastfail we no longer need _STL_UNREACHABLE to improve codegen.
• Add /Zc:cmath to libcxx, reported [libcxx][test] signaling_NaN() should not be used in constexpr <cmath> tests llvm/llvm-project#222431.
• Azure Pipelines: Adjust timeouts.
  • Remove most of the fine-grained task timeouts; we don't need them.
  • Keep the Code Format Validation job timeout at 5 minutes. It currently takes ~2m so this leaves a 2.5x margin.
  • Reduce the build-and-test.yml job timeout from 30 to 20 minutes. The slowest shards are ARM64EC and they take ~9m, so this leaves a 2.2x margin. Making this arch/stage dependent would be valuable, since the early builds and the x64/x86 runs are faster.
  • Reduce the Run Tests task timeout from 30 to 15 minutes. The slowest ARM64EC shard takes 5m56s so this leaves a 2.5x margin.
• Azure Pipelines: Refine timeouts.
  • Use 6 minutes for early x64/x86 builds (2.4x margin), and 9 minutes for early ARM64/ARM64EC builds (2.5x margin).
  • Use 5 minutes for Configure Tests (3.3x margin), same as Code Format Validation.
  • Use 15 minutes for full x64/x86 tests (2.5x margin) and 20 minutes for full ARM64/ARM64EC tests (2.2x margin).
  • Use 20 minutes for x64/x86 ASan (2.8x margin) and 25 minutes for ARM64 ASan (2.5x margin).
• Reported <cmath>: constexpr fpclassify() rejects signaling NaNs on x86 #6445.

[Stephan T. Lavavej (StephanTLavavej)]

I'm mirroring this to the MSVC-internal repo. Please notify me if any further changes are pushed, otherwise no action is required.