Inefficient codegen with struct consisting of 2 SIMD vectors.

[MichalPetryka]

Description

CoreCLR emits inefficient codegen when working with structs consisting of 2 SIMD vectors, for example custom structs with 2xVector128 or Vector512 on platforms without AVX512.

If you compare the codegen for methods A and B in the reproduction, you can notice the JIT spilling the custom structs instead of promoting them in B. The goal here would be to have B emit the same codegen as A does.
If you look at Memmove there, you'll notice that the JIT does not spill them, but introduces a lot of redundant adds and leas in the codegen.

Reproduction

Configuration

Windows x64 .Net 8 preview 2

Regression?

Probably not, checked just stack spilling and it occurs on .Net 7

[ghost]

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

Issue Details

---

Description

CoreCLR emits inefficient codegen when working with structs consisting of 2 SIMD vectors, for example custom structs with 2xVector128 or Vector512 on platforms without AVX512.

If you compare the codegen for methods A and B in the reproduction, you can notice the JIT spilling the custom structs instead of promoting them.
If you look at Memmove there, you'll notice that the JIT does not spill them, but introduces a lot of redundant adds and leas in the codegen.

Reproduction

Configuration

Windows x64 .Net 8 preview 2

Regression?

Probably not, checked just stack spilling and it occurs on .Net 7

Author:	MichalPetryka
Assignees:	-
Labels:	tenet-performance, area-CodeGen-coreclr, untriaged
Milestone:	-

[jakobbotsch]

The problem is here:

runtime/src/coreclr/jit/lclvars.cpp

Lines 2252 to 2270 in e43ddbf

	#ifdef FEATURE_SIMD
	// Check to see if this is a SIMD type.
	// We will only check this if we have already found a SIMD type, which will be true if
	// we have encountered any SIMD intrinsics.
	if (compiler->usesSIMDTypes() && (fieldInfo.fldSize == 0) && compiler->isSIMDorHWSIMDClass(fieldInfo.fldTypeHnd))
	{
	unsigned simdSize;
	CorInfoType simdBaseJitType = compiler->getBaseJitTypeAndSizeOfSIMDType(fieldInfo.fldTypeHnd, &simdSize);
	// We will only promote fields of SIMD types that fit into a SIMD register.
	if (simdBaseJitType != CORINFO_TYPE_UNDEF)
	{
	if ((simdSize >= compiler->minSIMDStructBytes()) && (simdSize <= compiler->maxSIMDStructBytes()))
	{
	fieldInfo.fldType = compiler->getSIMDTypeForSize(simdSize);
	fieldInfo.fldSize = simdSize;
	}
	}
	}
	#endif // FEATURE_SIMD

Since B uses no SIMD intrinsics the JIT assumes the fields of Vector2x128 are not SIMD types and we end up skipping promotion due to that.