Repository navigation
JIT: (bug) negating an AVX-512 multiply changes its explicit rounding mode #133519
Description
Repro
using System;
using System.Runtime.CompilerServices;
using System.Runtime.Intrinsics;
using System.Runtime.Intrinsics.X86;
class Program {
[MethodImpl(MethodImplOptions.NoInlining | MethodImplOptions.AggressiveOptimization)]
static Vector512<float> Test(Vector512<float> x) =>
-Avx512F.Multiply(x, Vector512.Create(1.5f), FloatRoundingMode.ToPositiveInfinity);
static void Main() {
if (!Avx512F.IsSupported)
throw new PlatformNotSupportedException("AVX-512F is required.");
float x = BitConverter.Int32BitsToSingle(0x3F800001);
Console.WriteLine(Test(Vector512.Create(x)).AsUInt32().GetElement(0).ToString("X8"));
}
}Expected
BFC00002
The exact product is halfway between two floats. Rounding up gives 0x3FC00002; negating gives 0xBFC00002.
Actual
BFC00001
The negation moved inside the rounded multiply, so rounding went the wrong way.
Platform
Windows x64 with AVX-512F, .NET 11.0.0-rc.1.26425.128.
Activity
dotnet-policy-service commentedon Sep 9, 2026
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.
[-]JIT: negating an AVX-512 multiply changes its explicit rounding mode[/-][+]JIT: (bug) negating an AVX-512 multiply changes its explicit rounding mode[/+]Fix optimizations of explicitly rounded hardware intrinsics (dotnet#1…
EgorBo commentedon Sep 18, 2026
Affected SDK versions: