[RyuJIT] Eliminate unecessary copies when passing structs

[CarolEidt]

Currently the JIT introduces unnecessary copies when passing structs by value.
This impacts both the by-ref passing as in the x86/64 windows ABI, as well as the x86/64 unix ABI - though the impact is worse in the latter case.
This is especially problematic for small structs on x86/64 unix, which are passed in registers but needlessly forced to memory. Span<T> and similar types are specific examples of this.
This issue started as https://github.com/dotnet/corefx/issues/27238, but is fundamentally a coreclr JIT issue.
The issues linked in this issue have largely been addressed, but this test case still shows unneessary copies: tests/src/JIT/Regression/JitBlue/DevDiv_590358/DevDiv_590358.cs

category:cq
theme:structs
skill-level:expert
cost:large

[fiigii]

Maybe related to https://github.com/dotnet/coreclr/issues/11030 (dotnet/coreclr#11133 solved part of the problem).

[CarolEidt]

@fiigii thanks for adding those links (especially pointing to the existing issues that @RussKeldorph had linked to).
#7897 had to do with struct returns, which have similar issues, but somewhat different causes.
I'm going to keep this open, as there are ineffiences on Windows as well as Unix, whereas the existng issues are described as Unix-specific. However, I'm going to explicitly link those issues here for cross-reference purposes:
#2908
#5785
#6316
#6265
#6266
It may be that some of the specific issues can be addressed individually, while this overall issue remains open.

[CarolEidt]

A simple example illustrating the excessive copies can be seen with the test case introduced in PR dotnet/coreclr#17418 - tests/src/JIT/Regression/JitBlue/DevDiv_590358/DevDiv_590358.cs

[jakobbotsch]

The main sysV problem was fixed with multireg work, I believe. The implicit byrefs case should be fixed by #79346, will leave it open for that.