rustc -O doesn't work on ARM

[chris-morgan]

Steps to reproduce:

1. Take a build of rustc for ARM (specifically, gnueabi; gnueabihf is untested and may or may not exhibit this problem). Luqman has one you can use at http://luqman.ca/rust-builds/.
2. Attempt to compile something with optimisations, e.g. echo 'fn main(){}' | rustc - -O.

This fails altogether:

Stack dump:
0.      Running pass 'Function Pass Manager' on module 'rust_out.rc'.
1.      Running pass 'Machine Common Subexpression Elimination' on function '@main'
Segmentation fault

A more complex example is building chris-morgan/rust-http@e909dfb; in this case it fails in a different way:

rustc: /scratch/laden/rust/src/llvm/include/llvm/ADT/ScopedHashTable.h:238: llvm::ScopedHashTableScope<K, V, KInfo, AllocatorTy>::~ScopedHashTableScope() [with K = llvm::Value*, V = std::pair<llvm::Value*, unsigned int>, KInfo = llvm::DenseMapInfo<llvm::Value*>, AllocatorTy = llvm::RecyclingAllocator<llvm::BumpPtrAllocator, llvm::ScopedHashTableVal<llvm::Value*, std::pair<llvm::Value*, unsigned int> >, 20u, 4u>]: Assertion `HT.TopLevelMap[ThisEntry->getKey()] == ThisEntry &&"Scope imbalance!"' failed.
Stack dump:
0.      Running pass 'Early CSE' on function '@"_ZN6client7request13RequestWriter3new18url_to_socket_addr19h3ff5b73b6544aal7ag11v0.1$x2dpreE"'
Aborted

I believe these are LLVM bugs rather than rustc bugs, but at the least Rust is triggering them. Should I be reporting them in LLVM?

[thestinger]

@chris-morgan: Use rustc --emit-llvm to output LLVM bytecode, and then use bugpoint -simplify-cfg -early-cse from Rust's LLVM directory (in Release+Asserts/bin) on the bytecode file. It should narrow down small a test case.

[luqmana]

So it turns out it was a problem with the gcc version I was using (Debian 4.4.6-14) miscompiling LLVM in some way. Switched to Debian 4.7.2-4 and it seems to work.

[chris-morgan]

This has turned out to be a false alarm caused by old software; @luqmana provided me with a new build using GCC 4.7 to build LLVM instead of GCC 4.4 and -O is now working. Yes, compilation of the http crate on my tablet can now take more than a minute instead of twenty seconds.