VMOV silently drop the size annotation

[eponier]

When compiling the following program

export fn main () -> reg u256 {
  reg u256 xmm;
  reg u64 x;
  x = 2;
  xmm = #VMOV_256(x);
  return xmm;
}

we get a typing error because xmm is of type u256 instead of u128. Indeed, the annotation _256 is silently dropped. It would be better to have an error in this case. Maybe it was done like this to avoid another constructor PrimXXX?

[eponier]

I discovered that for VMOV, but there are certainly other instructions with the same problem.

[vbgl]

The instructions descriptors should be slightly redesigned to explicitly document what sizes they support: currently this information is spread in the instruction syntax, semantics, parsing rules etc. Would it also fix this issue?

[eponier]

I guess so. But I think this could also be solved without such a refactoring (probably a tedious work though).

As usual, I am used to creating an issue as soon as I discover a strange behaviour, but this does not mean I care much about the issue being solved. I want to be sure that we have a trace of the issue.

[eponier]

Btw, at least from a user perspective, this issue is similar to https://github.com/jasmin-lang/jasmin/issues/69 where another annotation is silently ignored.

[vbgl]

In this case, the annotation is not ignored. It qualifies the argument to the intrinsic (here expression “x”). The result is always a u128, hence the error message.

If the failing line is changed to:

xmm = (256u)#VMOV_256(x);

then we get a type-error about the argument:

typing error: can not implicitly cast u64 into u256

I do not deny that there is some UX issue (it is difficult to properly use the MOVD / VMOV intrinsics), but I think that this issue is invalid and should be closed.

[eponier]

Indeed, the annotation has an impact only the type of the argument. It's probably better to close this.