Win8 baseline: Improve STL Hardening codegen size

[StephanTLavavej]

🗺️ Overview

• Will be merged after Win8 baseline: Drop support for Win7 / Server 2008 R2 #5432.
• Followup to STL Hardening #5274.
• For MSVC, this uses __fastfail:
  • "Support for the native fast fail mechanism began in Windows 8."
  • I am manually declaring the intrinsic because it's currently declared by the large <intrin.h>. We generally try to avoid duplicating intrinsic declarations (due to calling convention madness) but in this one case I think it's fine. Because <yvals.h> is very central and special, moving the declaration into <intrin0.h> might still be overly impactful.
  • I also need to hardcode FAST_FAIL_INVALID_ARG because there's no way we can drag in <winnt.h>.
• For Clang, this uses __builtin_verbose_trap:
  • This is the same mechanism used by libc++ Hardening.
  • It's cool in that it emits a single instruction, but when debugging info is enabled, it records the reason without bloating the codegen itself. Conveniently, we're already passing messages down to this layer.
• With STL Hardening enabled, I measured v[idx] on x64 with /O2:
  • This reduced codegen size from 56 bytes to 29 bytes by avoiding the _invoke_watson() function call.
  • Clang codegen was similarly improved.
• In <ranges>, we need to avoid a Clang error: "argument to __builtin_verbose_trap must be a pointer to a constant string".
• We need to work around VSO-2457624 "/clr /std:c++20 silent bad codegen with the __fastfail intrinsic".

📜 Changelog

• The STL no longer supports targeting Windows 7 and Server 2008 R2:
  • Improved the performance of STL Hardening by using the MSVC __fastfail intrinsic that was added in Windows 8, and the Clang __builtin_verbose_trap intrinsic.

[StephanTLavavej]

I'm mirroring this to the MSVC-internal repo - please notify me if any further changes are pushed.

[StephanTLavavej]

I pushed a commit to work around /clr silent bad codegen.