(I work on GPU compilers in LLVM.) I'd say that the fact that compilers are designed with a "stupid" IR-generation pass (e.g. the C++ AST to LLVM IR pass in cl

(I work on GPU compilers in LLVM.) I'd say that the fact that compilers are designed with a "stupid" IR-generation pass (e.g. the C++ AST to LLVM IR pass in clang) is part of a broader design strategy, namely, designing compilers to have many simple components that work together to do a complex thing. There are always trade-offs, as you point out, but one of the reasons that it's beneficial to design compilers this way stems from the fact that we generally apply a very high quality bar to compilers, because bugs in the compiler are expensive. Having simple components with strict interfaces allows us to reason precisely about what each of our transformations is supposed to do. That makes it easier to write unit tests and to do code reviews, and ultimately, in my experience, helps us put correctness first. I also don't think that this approach of dumb components necessarily leads to worse code. Indeed, for each of the LLVM missed optimizations in the list, I have a pretty good idea of which pass ought to be responsible for fixing the issue. And because the thing that each of these passes does is simple, I can have some confidence that my change won't break other passes.