Support smaller failure in libcore

[bharrisau]

With the reshuffle of libcore (#13851) it is becoming easier to build an embedded project with Rust. I've implemented the new weak failure in Zinc.rs, but as a result the binaries are 5-6kB larger (hackndev/zinc#41).

The two sources of additional size are c_str_to_static_slice pulling in is_utf8 in an assertion (could this be debug_assert?), and the format_args! macro. Trimming these gets the 5-6kB back (i.e. the empty application goes from 7576 bytes to 1888 bytes).

@alexcrichton: I'm not sure the best solution. My limited knowledge of the system suggests having a --cfg flag to compile libcore with stringless failures (maybe --cfg failure_tiny). Just replacing the c_str_to_static_slice with a zero len Slice {data: pointer, len: 0} (we can get the info from the debugger) and passing a zero length string to format_args!.

[alexcrichton]

The specific case you've mentioned needs to stay as an assert (or return failure), but I believe that the real problem here is that the compiler is invoking these intrinsics with *u8 instead of &str. If the compiler invoked the intrinsics with &str, none of the assertion code would be necessary.

[bharrisau]

Removing the conversion of c-string to Rust strings saved some of the size. There is 1.3kB remaining in assertion messages and file names. And 3.7kB in string formatting from secret_show inclusions from format_args!. I managed to track down where the secret_show functions were being included from the asm, a pointer to the function is included in the Arguments<> passed to the begin_unwind. (It didn't come up as a function call because there is no branch instruction)

@alexcrichton Thanks for your help - is there anything more that can be done? Or is this a will-not-fix thing? For my use I may end up patching failure.rs to pass a blank Arguments and a zero length file name just to cut the size down. Zinc isn't anything more than an experiment at this point, so I'm very grateful for your assistance so far shaving a few kB off.

[alexcrichton]

I'm all for shaving off what we can, and in general reducing code size. It sounds like the remaining code size is just code in general, so there may not any more low-hanging fruit in terms of code-size. If rust-lang/rfcs#93 is accepted then the footprint of secret_show will likely be reduced to smaller than it is today.

I wouldn't consider this a will-not-fix thing, but we also don't want to compromise too much on the current functionality of libcore. If you find any size-related issues, please definitely feel free to open an issue, I'd love to trim it down to as small as can be!