Mark and expose additional Vector functions as Intrinsic #77562
Conversation
dotnet-issue-labeler
bot
added
area-CodeGen-coreclr
|
Note regarding the This serves as a reminder for when your PR is modifying a ref *.cs file and adding/modifying public APIs, to please make sure the API implementation in the src *.cs file is documented with triple slash comments, so the PR reviewers can sign off that change. |
ghost
assigned 
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch Issue DetailsThis resolves #76593
|
| @@ -1150,7 +1151,7 @@ internal static Vector256<T> Create<T>(Vector128<T> lower, Vector128<T> upper) | |||
| /// <returns>A new <see cref="Vector256{T}" /> instance with the first element initialized to <paramref name="value" /> and the remaining elements initialized to zero.</returns> | |||
| /// <exception cref="NotSupportedException">The type of <paramref name="value" /> (<typeparamref name="T" />) is not supported.</exception> | |||
| [MethodImpl(MethodImplOptions.AggressiveInlining)] | |||
| internal static Vector256<T> CreateScalar<T>(T value) | |||
| public static Vector256<T> CreateScalar<T>(T value) | |||
There was a problem hiding this comment.
Could we change this to
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static Vector256<T> CreateScalar<T>(T value)
where T : struct
{
if (Avx.IsSupported)
{
return Vector128.CreateScalar(value).ToVector256Unsafe();
}
return Vector128.CreateScalar(value).ToVector256();
}I end up not using this method a lot of times because it currently emits an extra vmovaps to clear the upper lane after the scalar move, even though the scalar move will have already zeroed the upper lane with VEX encoding.
The extra vmovaps should get no-op'ed by the CPU front-end anyway, but this is a common need, so smaller codegen would be better.
There was a problem hiding this comment.
I'm going to update this to be a proper intrinsic in a follow up PR.
|
CC. @dotnet/jit-contrib for the runtime side tweaks |
tannergooding
force-pushed
the
fix-76593
branch
from
2ad678f to
a52829b
Compare
tannergooding
force-pushed
the
fix-76593
branch
from
4ad6072 to
8d9c929
Compare
| @@ -19573,9 +19681,22 @@ GenTree* Compiler::gtNewSimdBinOpNode(genTreeOps op, | |||
| case GT_RSH: | |||
| case GT_RSZ: | |||
| { | |||
| assert(!varTypeIsFloating(simdBaseType)); | |||
There was a problem hiding this comment.
I see that Vector*<T> is going to support shifting so it makes sense that the restriction is lifted, then you normalize the base type.
| @@ -2030,17 +2055,21 @@ GenTree* Compiler::impBaseIntrinsic(NamedIntrinsic intrinsic, | |||
|
|
|||
| if ((simdSize != 32) || compExactlyDependsOn(InstructionSet_AVX2)) | |||
| { | |||
| genTreeOps op = varTypeIsUnsigned(simdBaseType) ? GT_RSZ : GT_RSH; | |||
There was a problem hiding this comment.
If the base type is unsigned, is op_RightShift basically going to be the same thing as op_UnsignedRightShift?
There was a problem hiding this comment.
Correct, that's how op_RightShift is expected to behave for unsigned types
|
/azp run runtime-coreclr jitstress-isas-x86, runtime-coreclr jitstress-isas-arm, runtime-coreclr outerloop |
|
Azure Pipelines successfully started running 3 pipeline(s). |
|
|
This resolves #76593