SPARC alignment issue in LTO

[dhduvall]

The function extract_bytecode_format_version() in rustc_trans::back::lto is consistently crashing with a SIGBUS on SPARC when compiling Firefox. I'm seeing pos set to 11, and the code generated for the coercion in the unsafe statement isn't intelligent enough to deal with the lack of alignment.

The reduced testcase is simple:

fn main() {
    let arr: [u8; 8] = [0; 8];
    let byte_data = &arr[1..5];
    let data = unsafe { *(byte_data.as_ptr() as *const u32) };
    u32::from_le(data);
}

I can fix that with a copy_from_slice():

    let mut byte_data: [u8; 4] = [0; 4];
    byte_data.copy_from_slice(&bc[pos..pos + 4]);

but I don't know if there's a more idiomatic way, and I don't know how best to make sure that on architectures that can handle the one-byte alignment, it skips the copy.

I suspect that extract_compressed_bytecode_size_v1() has a similar issue.

[parched]

*(byte_data.as_ptr() as *const u32) is undefined behaviour. I think the best fix would just be to << and | the bytes by hand without any unsafe code. Yes, extract_compressed_bytecode_size_v1() also has the same issue.

[ollie27]

Another option would be to replace *(byte_data.as_ptr() as *const u32) with ptr::read_unaligned(byte_data.as_ptr() as *const u32).

[infinity0]

This also caused test failures on armhf on Ubuntu because they pass -Bsymbolic-functions to ld(1). (Debian does not, and did not see the test failures.) The PR fixes that too.