Ensure that we don't try to constant fold BSF(0) or BSR(0)

[tannergooding]

This resolves #81460

Even with the qmark (x == 0) ? 31 : BSR(x), we'd end up getting something like:

lcl = 0;
if (lcl == 0)
   return 31;
return BSR(lcl);

So even though the BSF(0)/BSR(0) is dead code, we don't know it yet and it leads to VN trying to fold and therefore triggering the assert.

The fix is to simply bail out if the constant is zero.

[ghost]

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

Issue Details

---

This resolves #81460

Even with the qmark, we'd end up getting something like:

lcl = 0;
if (lcl == 0)
   return 31;
return BSR(lcl);

So even though the BSF(0)/BSR(0) is dead code, we don't know it yet and it leads to VN trying to fold and therefore triggering the assert.

The fix is to simply bail out if the constant is zero.

Author:	tannergooding
Assignees:	-
Labels:	area-CodeGen-coreclr
Milestone:	-

[tannergooding]

There is potentially something here to be handled in the future in that we are doing VN on things which are "dead".

If we had some minimal early constant propagation or similar, we could avoid doing VN/CSE/folding or other potentially expensive things for code that will just be dropped in one of the next few phases.

[tannergooding]

Brought in latest main to pickup the CI fix