Deal with strict aliasing issues (-fstrict-aliasing vs -fno-strict-aliasing)

[briansmith]

Recently we started building with -fno-strict-aliasing because GCC was warning about strict aliasing, e.g. in https://travis-ci.org/briansmith/ring/jobs/103464242:

crypto/modes/gcm.c:517:5: error: dereferencing type-punned pointer will break strict-aliasing rules [-Werror=strict-aliasing]
     PUTU32(ctx->Yi.c + 12, ctr);
     ^

Building with -fstrict-aliasing is really dangerous unless we're really sure the code is correct with respect to these issues. We can probably avoid turning it off as most of the C code for which it should matter will eventually be written in Rust.

However, we need to at least (a) measure the performance difference, and (b) come up with a plan to address any performance issues we find.

[agauniyal]

Hi, did you ever perform any measurements on the above issue, or changed something else? I'm in a similar boat, was wondering how to solve this problem..

[briansmith]

I'm in a similar boat, was wondering how to solve this problem..

What is the exact problem you are having? Is it a problem you're having with ring, or a problem with another library?

[agauniyal]

some code of mine, but it includes passing -fno-strict-aliasing in order to be safe. Same warning and everything..

[briansmith]

I see. I've already removed most pointer casts and most uses of unions from this project, ring, so it's unlikely that we will have to do anything more regarding strict aliasing and we can probably remove -fno-strict-aliasing. OTOH, because we're replacing most of the C code with Rust code, we probably won't benefit from removing the -fno-strict-aliasing anyway, so we might just leave it.

[briansmith]

Update:

• Issue Remove all uses of unions #375 tracks the removal of all unions from ring's C code. There are no more unions in the C code!
• Issue Remove all pointer casts in C code #151 tracks the removal of all pointer casts from ring's C code. That has been done, AFAICT. Just waiting for a second verification.
• Issue Remove integer arithmetic on values converted from pointers #101 tracks reading and using pointers as integers, e.g. tricks using uintptr_t and intptr_t, including abuse of size_t. AFAICT there are no more such casts being done. Again, waiting for a second verification of this.