Stop zeroing out the scratch buffer in float_to_decimal_common

[bzbarsky]

float_to_decimal_common in src/libcore/fmt/float.rs does this:

let mut buf = [0; 1024];

which zeroes out the big buffer. Given that it's a scratch buffer, there shouldn't be a reason to do that, I would think.

[eddyb]

Link to the code in question. IMO we should start using ArrayVec and variations of it (including ring buffers and even arbitrary number of ranges or other "partial initialization patterns") throughout the libs.

cc @rust-lang/libs

[sfackler]

Is there a concrete performance issue that this is proposing to tackle? I would like to avoid moving from use of initialized to uninitialized memory "just because".

[eddyb]

@sfackler It came up in a Servo profile. But regardless, I would like to use safe abstractions over partial initialization of stack arrays, to better represent the semantics involved.

[bzbarsky]

Is there a concrete performance issue that this is proposing to tackle?

Yes. The zeroing out was about 5% of the time taken in a "get the specified value of top" microbenchmark for stylo.

Or another way to think of it is that it's 15% of the time Safari takes to execute the entire operation. ;)

[sfackler]

Cool, seems worth fixing then 👍