[llvm] [GlobalISel] An optimizing MIR builder (PR #132282)

[Amara Emerson via llvm-commits]

aemerson wrote:

> I removed G_ADD, G_SUB, and G_MUL from the constant folding list in CSEMIRBuilder. It caused a **lot** of regressions. People depend on the constant folding in the builder.

Perhaps, that to me indicates we have some missing optimizations in the combiner in order to make it functionally complete. Not every usage of MIRBuilder in a pipeline can be assumed to be CSE'ing.
>
> GConstant can only provide helpers for scalars. How do you constant fold vectors with out much burden for the user?
> 
> Edit: Exploiting undef and poison might make the program one epsilon smaller. Combiners also rely on builders.

As I said the GIConstant issue isn't the point here.

And I'd recommend taking a look at our GitHub issues to find other more important fixes to do as your first contribution.

https://github.com/llvm/llvm-project/pull/132282