MSVC Backend Updates in Visual Studio 2022 version 17.2

Chris Pulido

In Visual Studio 2022 version 17.2 we have continued to improve the C++ backend with new features, new and improved optimizations, build throughput improvements, and better security. Here is a list of improvements for you to review.

• OpenMP: The task directive as defined by OpenMP 3.1 is supported for -openmp:llvm, including all the clauses. Note that the compiler does not yet support `task` clauses added in later versions of OpenMP. See more details in OpenMP Task Support for C++ in Visual Studio.
• Implemented Intel intrinsic functions for the AVX512-FP16 instruction set extension. More information about these functions can be found on the Intel Intrinsics Guide.
• Implemented Intel intrinsic functions _castf32_u32, _castf64_u64, _castu32_f32, and _castu64_f64 to cast between floating point values and integer values without conversion on x64 and x86. More information about these functions can be found on the Intel Intrinsics Guide.
• New ARM64 compiler flags: /Zc:arm64-aliased-neon-types- and /Zc:arm64-aliased-neon-types. When you pass /Zc:arm64-aliased-neon-types- to cl.exe, the compiler will treat NEON intrinsic types as distinct types for ARM64 as defined by the Procedure Call Standard for the Arm 64-bit Architecture, which is consistent with Clang and GCC. This flag is opt-in, so ARM64 NEON intrinsic code that compiled with previous versions of MSVC will still compile when you upgrade. /Zc:arm64-aliased-neon-types (without the minus sign at the end) is the default behavior.
  • For example, consider two function declarations, void foo(float32x4_t) and void foo(int32x4_t). By default, MSVC considers these two the same declaration, and attempting to define them both would lead to a multiple definition error. With /Zc:arm64-aliased-neon-types-, MSVC will treat them as Clang and GCC would.
• New ARM64 compiler flags: /arch:armv8.0 and /arch:armv8.1. These new flags allow the compiler to generate instructions that were introduced and required by the specified architecture extension. `/arch:armv8.0` is the current default behavior and is the same as if you didn’t specify it. In 17.2, /arch:armv8.1 allows the _Interlocked* intrinsic functions to use the appropriate atomic instruction that was introduced with the ARMv8.1 extension, FEAT_LSE.
• New and improved optimizations
  • The C standard library functions log2 and log2f have been implemented as compiler intrinsic functions on x64 and ARM64. This allows the compiler to perform optimizations with log2 and log2f under /fp:fast x64 and ARM64.
  • Improved auto-vectorizer loop recognition. The auto-vectorizer now recognizes the average pattern and more cases of decrementing induction variables.
  • More peephole optimizations for multiple targets.
  • Improved load/store pairing on ARM64.
• ARM64EC
  • Compiler flags incompatible with the /arm64EC flag are now rejected. This includes all CLR flags, /Gy-, and /Gw-.
  • Added the /MACHINE:ARM64EC flag to link.exe, and removed it from lib.exe. For lib.exe, you should specify /MACHINE:ARM64X.
  • When /arm64EC is passed to cl.exe and cl.exe also invokes link.exe, /MACHINE:ARM64EC will be passed by default to link.exe.

Do you want to experience the new improvements of the C++ backend? Please download the latest Visual Studio 2022 and give it a try! Any feedback is welcome. We can be reached via the comments below, Developer Community, and Twitter (@VisualC)