intrinsics-testing

Working with hardware intrinsics

Always go through the BC wrappers for ISA detection

Rule: when gating an intrinsic codepath on the availability of a hardware ISA, use the wrappers under Org.BouncyCastle.Runtime.Intrinsics — never call System.Runtime.Intrinsics.X86.*.IsSupported (or ArmBase.*.IsSupported, etc.) directly from engine / digest / math code.

Use:

if (Org.BouncyCastle.Runtime.Intrinsics.X86.Aes.IsEnabled)
{
    // AES-NI implementation
}

Not:

#if NETCOREAPP3_0_OR_GREATER
if (System.Runtime.Intrinsics.X86.Aes.IsSupported)   // ❌ direct check
{
    // ...
}
#endif

Why it matters:

• Single point of indirection. All gates can be flipped globally — for testing, for opt-out scenarios, or for the profile-based gating described below — without touching every callsite.
• TFM uniformity. The wrappers compile cleanly on TFMs that don't have System.Runtime.Intrinsics at all (net461, net472, netstandard2.0) by hard-coding IsEnabled => false. Callers don't need their own #if NETCOREAPP3_0_OR_GREATER guard around the check itself — only around the body that uses the intrinsics types.
• Forward compatibility. See "Limited ISA profiles" below.

The wrapper set is intentionally narrow

The wrappers don't aim to mirror the whole BCL surface — they only cover ISAs we actually use somewhere in the codebase. Today that's just Aes, Avx2, Bmi1, Bmi2, Pclmulqdq, Sse2, Sse41, Ssse3 (under crypto/src/runtime/intrinsics/x86/). Anything else (e.g. Avx, Avx512F, Sse42, ARM AdvSimd, ARM Crypto) has no wrapper because nothing in BC has needed one yet.

When introducing a new intrinsic codepath that depends on an ISA without an existing wrapper, add the wrapper first. Pattern, mirroring the existing files:

namespace Org.BouncyCastle.Runtime.Intrinsics.X86
{
    internal static class Avx512F
    {
#if NETCOREAPP3_0_OR_GREATER
        internal static bool IsEnabled => System.Runtime.Intrinsics.X86.Avx512F.IsSupported;
#else
        internal static bool IsEnabled => false;
#endif
    }
}

Then have the new engine/digest/math code consume Avx512F.IsEnabled rather than the BCL IsSupported directly. This keeps every gate routable through the same future profile mechanism.

Don't add wrappers speculatively for ISAs that no code uses yet — they'd just be dead surface area and would muddy the "what does BC actually exercise" picture. Add them on demand, in the same PR as the first consumer.

Limited ISA profiles (planned direction)

The cross-product of individual ISA toggles is too large to test exhaustively. The planned evolution is to support only a small, enumerated set of ISA profiles (combinations of available ISAs that ship together — e.g. "scalar only", "SSE2+AES+PCLMULQDQ", "AVX2+AES+PCLMULQDQ+BMI2", etc.) so that complete coverage is actually achievable.

Implications for code being written today:

• Going through the wrappers (rule above) is what makes the eventual profile gating possible — the wrappers will be where profile selection plugs in.
• Don't introduce new ISA combinations that wouldn't fit a small profile set. If you find yourself wanting a check like "AVX2 but no BMI2," step back: that combination probably isn't on any plausible profile list and you're multiplying the test matrix for a configuration that won't actually ship.
• New intrinsic codepaths should be tractable to fully exercise via the testing approach below; if a path is reachable only under an exotic ISA mix, that's a smell.

Testing fallback paths

When a change touches intrinsics code, the default dotnet test run only exercises one ISA path — whichever the host CPU supports. To verify the fallbacks, re-run the suite with specific instruction sets disabled via the .NET runtime's Enable* toggles. The wrappers honor these toggles automatically because their IsEnabled properties delegate to IsSupported, which is what the runtime flag affects.

When this applies

The toggles only affect TFMs with System.Runtime.Intrinsics available — i.e. net6.0 (and netcoreapp3.1 via the COMPlus_ prefix). The .NET Framework targets (net461, net472) don't compile the intrinsics paths at all (#if NETCOREAPP3_0_OR_GREATER guards them out), so there's nothing to toggle there.

Code locations likely affected by ISA toggles:

• crypto/src/runtime/intrinsics/ — the IsEnabled wrappers themselves.
• *_X86.cs files (e.g. AesEngine_X86.cs, Blake2b_X86.cs, Haraka512_X86.cs).
• Engines/modes/math files with inline #if NETCOREAPP3_0_OR_GREATER blocks that go through the wrappers (e.g. ChaCha7539Engine.cs, GCMBlockCipher.cs, Nat256.cs). Any callsite that bypasses the wrappers (see the rule above) is itself a bug — fix the wrapper indirection rather than working around it.

Env-var prefix

• .NET 5+ (which includes net6.0): DOTNET_Enable*=0
• netcoreapp3.1: COMPlus_Enable*=0

Setting the value to 0 disables that instruction set; the runtime will then report IsSupported = false for it and any others that depend on it. Because BC's IsEnabled wrappers delegate to IsSupported, the toggles propagate transparently to all gated codepaths.

PowerShell variants

Default (best-available) baseline:

dotnet test --framework net6.0 crypto/test/BouncyCastle.Crypto.Tests.csproj

Disable AVX2 (forces the SSE2/SSSE3/SSE41 paths):

$env:DOTNET_EnableAVX2=0
dotnet test --framework net6.0 crypto/test/BouncyCastle.Crypto.Tests.csproj
Remove-Item env:DOTNET_EnableAVX2

Disable AES-NI + PCLMULQDQ (forces software AES and software GHASH):

$env:DOTNET_EnableAES=0
$env:DOTNET_EnablePCLMULQDQ=0
dotnet test --framework net6.0 crypto/test/BouncyCastle.Crypto.Tests.csproj
Remove-Item env:DOTNET_EnableAES
Remove-Item env:DOTNET_EnablePCLMULQDQ

Disable all hardware intrinsics in one shot (forces the pure-managed scalar paths):

$env:DOTNET_EnableHWIntrinsic=0
dotnet test --framework net6.0 crypto/test/BouncyCastle.Crypto.Tests.csproj
Remove-Item env:DOTNET_EnableHWIntrinsic

For netcoreapp3.1, swap DOTNET_ for COMPlus_:

$env:COMPlus_EnableAVX2=0
dotnet test --framework netcoreapp3.1 crypto/test/BouncyCastle.Crypto.Tests.csproj
Remove-Item env:COMPlus_EnableAVX2

Full list of toggles used in this codebase

DOTNET_EnableAES, DOTNET_EnableAVX, DOTNET_EnableAVX2, DOTNET_EnableBMI1, DOTNET_EnableBMI2, DOTNET_EnablePCLMULQDQ, DOTNET_EnableSSE2, DOTNET_EnableSSE41, DOTNET_EnableSSSE3, plus the catch-all DOTNET_EnableHWIntrinsic.

Caveats

• DOTNET_EnableSSE2=0 is ignored on x64. The .NET runtime treats SSE2 as part of the x64 ABI, so the toggle is honored only on 32-bit hosts. To actually test without SSE2, run with dotnet test -a x86 ... (and pick a TFM that supports x86, e.g. netcoreapp3.1 or net6.0).
• Pick targeted combinations rather than running the full cross-product — a few well-chosen disables (highest ISA off, AES-NI off, all-off) usually cover the meaningful fallback shapes for a given change. Once the limited-profiles model lands, the targeted set will be the profile list itself.
• Failed assertions under a disabled ISA usually mean the fallback path diverged from the intrinsic path; the failing test name plus the disabled toggle identifies the regression area.

When not to run these

Don't run the variant suites for changes that don't touch intrinsics code or anything they call into. The default test run is sufficient for non-intrinsics work, and the variant runs add real time to the loop.