[llvm] [GISel][RISCV]Implement indirect parameter passing for large scalars (PR #95429)

[Gábor Spaits via llvm-commits]

spaits wrote:

> > Why don't just do the lowering in the frontend? I.e. return getNaturalAlignIndirect(...) if argument's size exceeds 2xXLEN
> 
> I think we do.

In case of this code https://godbolt.org/z/bqY5Kd5n8 we don't do that in the frontend. This is the IR after the final step in before the irtranslator:
```llvm
define dso_local noundef i128 @fun(__int128, int)(i128 noundef %i, i32 noundef %a) {
entry:
  %i.addr = alloca i128, align 16
  %a.addr = alloca i32, align 4
  %r = alloca i128, align 16
  store i128 %i, ptr %i.addr, align 16
  store i32 %a, ptr %a.addr, align 4
  %0 = load i128, ptr %i.addr, align 16
  %1 = load i32, ptr %a.addr, align 4
  %sh_prom = zext i32 %1 to i128
  %shl = shl i128 %0, %sh_prom
  store i128 %shl, ptr %r, align 16
  %2 = load i128, ptr %r, align 16
  ret i128 %2
}
```

In case of large struct, yes, the frontend wants to pass them as a reference. Code: https://godbolt.org/z/4M1cEa5hh
This is the IR after the final step in before the irtranslator:
```llvm
define dso_local i128 @fun(LargeStruct, int)(%struct.LargeStruct* byval(%struct.LargeStruct) align 8 %0, i32 %a) {
entry:
  %i = alloca %struct.LargeStruct, align 16
  %a.addr = alloca i32, align 4
  %r = alloca i128, align 16
  %1 = bitcast %struct.LargeStruct* %i to i8*
  %2 = bitcast %struct.LargeStruct* %0 to i8*
  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 16 %1, i8* align 8 %2, i32 80, i1 false)
  store i32 %a, i32* %a.addr, align 4
  %a1 = getelementptr inbounds %struct.LargeStruct, %struct.LargeStruct* %i, i32 0, i32 0
  %3 = load i128, i128* %a1, align 16
  %4 = load i32, i32* %a.addr, align 4
  %sh_prom = zext i32 %4 to i128
  %shl = shl i128 %3, %sh_prom
  store i128 %shl, i128* %r, align 16
  %5 = load i128, i128* %r, align 16
  ret i128 %5
}
```

https://github.com/llvm/llvm-project/pull/95429