Unix: Unnecessary struct copy while passsing struct of size <=16

[erozenfeld]

This is a copy of #6316. For some reason it wasn't possible to update the description of that issue.

Struct arguments whose size is <= 16 bytes generate unnecessary copies. @sivarv categorized the cases below:

The following are the struct passing cases on Amd64 Unix that are of interest:

struct of size 1,2,4 or 8 in memory/register is getting passed in an arg reg or on stack
struct of size 3,5,6,7 in memory/register is getting passed in an arg reg or on stack
struct of size 9-16 bytes in memory is getting passed in two arg registers
struct of size 9-16 bytes (e.g. SIMD structs Vector3 and Vector4) in a register is getting passed in two arg registers
struct of size 9-16 bytes in memory/register is getting passed on stack.
In fgMorphArgs, eeGetSystemVAmd64PassStructInRegisterDescriptor classifies the cases above as register argument except simd from vector and sets structDesc.passedInRegisters to true, this always calls fgMakeOutgoingStructArgCopy() and generates copies.

category:cq
theme:calling-convention
skill-level:intermediate
cost:medium
impact:small

[erozenfeld]

Comment from @CarolEidt on Jun 12 2018:

dotnet/coreclr#18358 eliminated the unnecessary copy when passed on stack.
There are still cases where structs are copied unnecessarily when passed in registers.

[jakobbotsch]

@SingleAccretion is this still relevant?

[SingleAccretion]

A little bit, yes. I believe there is still some copying of Vector3s we can avoid.