> which I believe is also somewhat at odds with advanced optimizations, like those used in LLVM

You can configure the optimization level to 0, then only constant folding is performed when the IR is built, and nothing else is done.