Stack overflow in upstream GCC when compiling huge functions

[antoyo]

Stack overflow happens in verify_vssa.

Upstream issue: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86422

@FractalFir: were you able to create a reproducer in C for this bug?

[FractalFir]

Not yet, but it should not be too hard. Will let you know once I have done so.

[FractalFir]

Additionally, I think we maybe should open an upstream issue in rustlang/rust for this.

The codegen for string matches is abysmal, both for compile and runtime performance.

This is not just a GCC-specifc problem. The MIR generated for this is just bad all around. LLVM also chokes on this, it just does not crash.

Even if this is a low priority, I still think this should be documented somewhere.

[antoyo]

So, you mentioned that rustc compiles matches on strings as an if-else ladder?
If so, it would indeed be a good idea to open an issue for this if it's not already the case.

I know it's possible a backend compiles a switch using different strategies (like a jump table if the values are dense or even a binary search if the values are sparse). I don't know if LLVM or GCC does this, but if we don't even use a switch in this case, they won't be able to do that optimization anyway.

[ptrca]

This might be it: rust-lang/rust#39525 (comment)

[FractalFir]

Yeah, I think that is it. Sad to see that the issue is 8 years old, but I suppose this is not an easy problem.