JIT, AOT preinit, and CoreLib: Saturate float/double conversions to small integral types

[Copilot]

Unchecked float/double → small integral type (sbyte, byte, short, ushort, char) conversions did not saturate at the small-type boundary. For example, (short)(double)32768.000000000007 produced -32768 instead of 32767, diverging from the saturating R → int contract introduced in .NET 9.

The JIT expands R → smallT as R → int → smallT. The inner R → int saturates, but the outer int → smallT only truncates low bits. The CoreCLR interpreter VM helper and the ILC type preinitializer had analogous bugs and were updated for parity with runtime behavior. In addition, CoreLib TryConvert paths for float/double to small integral types were updated to use direct casts on CoreCLR (while preserving existing MONO behavior under #if MONO).

Summary

Fixes saturating semantics for unchecked float/double → small integral type casts across all supported JIT targets (xarch, x86, arm64, RISC-V64, LoongArch64, ARM32, and WASM), the CoreCLR interpreter VM helper, and the NativeAOT type preinitializer, and applies corresponding CoreCLR-side TryConvert simplifications that rely on the corrected cast semantics. Mono runtime backends still need analogous updates and are tracked separately; the JIT regression test is [SkipOnMono] in the meantime.

Changes Made

• src/coreclr/jit/morph.cpp — fgMorphExpandCast: Added target-specific handling so R -> small saturates at destination bounds across supported architectures:
  • FEATURE_HW_INTRINSICS targets use float-domain min/max clamping before existing cast expansion.
  • WASM lowers min/max via GT_INTRINSIC.
  • ARM32, RISC-V64, and LoongArch64 use integer-domain saturation (R -> int32 followed by NI_PRIMITIVE_SaturateTo{Int8,Int16,UInt8,UInt16}), preserving NaN→0 semantics from the saturating R → int cast.
• JIT plumbing for the new NI_PRIMITIVE_SaturateTo{Int8,Int16,UInt8,UInt16} intrinsics and the WASM binary GT_INTRINSIC(MaxNative/MinNative) produced by morph:
  • gentree.cpp: added the four SaturateTo* cases to the unary GT_INTRINSIC cost switch in gtSetEvalOrder, added a TARGET_WASM block to the binary GT_INTRINSIC cost switch for MaxNative/MinNative, and added dump labels for the four SaturateTo* IDs (fixes the 'Unknown binary GT_INTRINSIC operator' assert seen on e.g. System.Half:op_Explicit(System.Half):byte).
  • importercalls.cpp IsTargetIntrinsic: added the four SaturateTo* IDs to the ARM, RISCV64, and LoongArch64 switches so rationalize.cpp's IsTargetIntrinsic assertion is satisfied.
  • importercalls.cpp impPrimitiveNamedIntrinsic: split the small-target-type path so NI_PRIMITIVE_ConvertToInteger emits a single saturating R -> small cast while NI_PRIMITIVE_ConvertToIntegerNative keeps the original two-cast (saturating R -> int followed by truncating int -> small) pattern that preserves native truncating semantics.
  • valuenum.cpp fgValueNumberIntrinsic: added an explicit branch for the SaturateTo* IDs that gives the node an opaque unary VN, backed by new VNF_SaturateToInt8/Int16/UInt8/UInt16 entries in valuenumfuncs.h, so VN no longer hits the assert(NI_System_Object_GetType) for these intrinsics.
  • valuenum.cpp EvalMathFuncBinary: extended the MaxNative/MinNative constant-folding (TYP_DOUBLE / TYP_FLOAT) and VNFunc selection (VNF_MaxNumber / VNF_MinNumber) cases to also apply on WASM, not just RISC-V64, so the binary GT_INTRINSIC produced by morph on WASM no longer hits unreached() here.
  • assertionprop.cpp IntegralRange::ForNode: taught range analysis the exact result range for each SaturateTo* intrinsic ([ByteMin, ByteMax], [ShortMin, ShortMax], [0, UByteMax], [0, UShortMax]).
• ARM32 codegen for SaturateTo*: new INS_ssat/INS_usat Thumb-2 instructions in instrsarm.h, encoding/disassembly support in emitarm.cpp, LSRA/codegen wiring in lsraarm.cpp and codegenarmarch.cpp.
• RISC-V64 / LoongArch64 codegen for SaturateTo*: branch-based clamp in codegenriscv64.cpp / codegenloongarch64.cpp, with LSRA support in lsrariscv64.cpp / lsraloongarch64.cpp.
• src/coreclr/vm/interpexec.cpp — ConvFpHelper: clamps to numeric_limits<TResult> (the destination type) instead of TIntermediate, so floating-point → small-integral conversions now saturate at the destination type's range.
• src/coreclr/tools/aot/ILCompiler.Compiler/Compiler/TypePreinit.cs: float-domain conv_i1/conv_i2/conv_u1/conv_u2 cases now explicitly saturate via Math.Clamp (NaN → 0), so AOT-baked statics match runtime behavior regardless of which host JIT runs ILC. Original simple-cast version is kept commented out with a TODO to restore once the fix has propagated through the toolchain.
• CoreLib TryConvert simplifications (Double.cs, Single.cs, SByte.cs, Int16.cs): float/double → small-integral TryConvertTo* / TryConvertFrom* paths now use a direct cast on CoreCLR, with the existing manual clamp preserved under #if MONO.
• Tests:
  • Re-enabled previously-disabled DoubleTests.GenericMath.ConvertToIntegerTest / SingleTests.GenericMath.ConvertToIntegerTest (issue (short)(double)32768.000000000007 produces invalid result #116823 suppressions removed).
  • Added a JitBlue regression test for the float/double → small-integral saturation contract (marked [SkipOnMono] pending Mono backend updates).
  • Trimmed DoubleTests.GenericMath.ConvertToIntegerNativeTest / SingleTests.GenericMath.ConvertToIntegerNativeTest to validate only in-range inputs and removed the [ActiveIssue("#124344", ... IsAppleMobile && IsX64Process && IsCoreCLR)] suppression, since ConvertToIntegerNative is explicitly allowed to use platform-native behavior for out-of-range inputs (which can also evolve over time, e.g. as AVX10.2 saturating conversions get exposed on xarch).
  • Updated the interpreter test comment/coverage.

Note: the previously included change to src/coreclr/interpreter/compiler.cpp (adding a two-step expansion for NI_PRIMITIVE_ConvertToIntegerNative on small target types) was reverted. The saturating behavior is allowed as the native interpreter behavior, and when R2R is present the R2R version of the intrinsic is used, so consistency with R2R is achieved that way.

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

[tannergooding]

This should be ready for review now. Handles the scenario for everything except Mono, which requires a significantly larger and more complex PR covering their multiple backends.

[Copilot]

Pull request overview

Copilot reviewed 26 out of 26 changed files in this pull request and generated no new comments.

[Copilot]

Copilot's findings

• Files reviewed: 26/26 changed files
• Comments generated: 3

[Copilot]

Copilot's findings

• Files reviewed: 26/26 changed files
• Comments generated: 0 new

[jkotas]

Anything left here? It looks good to me overall, the JIT changes can use a review from somebody on codegen.

[Copilot]

Pull request overview

Copilot reviewed 26 out of 26 changed files in this pull request and generated no new comments.

[tannergooding]

Anything left here? It looks good to me overall, the JIT changes can use a review from somebody on codegen.

It should be ready and was just pending review.

CC. @dotnet/jit-contrib for review

[Copilot]

Copilot's findings

• Files reviewed: 26/26 changed files
• Comments generated: 3

[Copilot]

Copilot's findings

• Files reviewed: 26/26 changed files
• Comments generated: 1

[tannergooding]

CI is passing, so this is just pending sign-off to be merged now.