`std::bigint` contains the following code.
```rust
borrow = *elem << (uint::bits - n_bits);
```
The code above contains a bug that the value of the right operand of the shift operator exceeds the size of the left operand,
because sizeof(*elem) == 32, and 0 <= n_bits < 32 in 64bit architecture.
If `--opt-level` option is not given to rustc, the code above runs as if the right operand is `(uint::bits - n_bits) % 32`,
but if --opt-level is given, `borrow` is always zero.
I wonder why this bug is not catched in the libstd's testsuite (I try the `rustc --test --opt-level=2 bigint.rs` before fixing the bug,
but the unittest passes normally.)
This pull request also removes the implicit vector copies in `bigint.rs`.
This replaces the wrapper around the runtime RNG with a pure Rust implementation of the same algorithm. This is much faster (up to 5x), and is hopefully safer.
There is still (a little) room for optimisation: testing by summing 100,000,000 random `u32`s indicates this is about ~~40-50%~~ 10% slower than the pure C implementation (running as standalone executable, not in the runtime).
(Only 6d50d55 is part of this PR, the first two are from #6058, but are required for the rt rng to be correct to compare against in the tests.)
This replaces the wrapper around the runtime RNG with a pure Rust
implementation of the same algorithm. This is faster (up to 5x), and
is hopefully safer.
There is still much room for optimisation: testing by summing 100,000,000
random `u32`s indicates this is about 40-50% slower than the pure C
implementation (running as standalone executable, not in the runtime).
As discussed on issue #4819, I have created four new traits: `Algebraic`, `Trigonometric`, `Exponential` and `Hyperbolic`, and moved the appropriate methods into them from `Real`.
~~~rust
pub trait Algebraic {
fn pow(&self, n: Self) -> Self;
fn sqrt(&self) -> Self;
fn rsqrt(&self) -> Self;
fn cbrt(&self) -> Self;
fn hypot(&self, other: Self) -> Self;
}
pub trait Trigonometric {
fn sin(&self) -> Self;
fn cos(&self) -> Self;
fn tan(&self) -> Self;
fn asin(&self) -> Self;
fn acos(&self) -> Self;
fn atan(&self) -> Self;
fn atan2(&self, other: Self) -> Self;
}
pub trait Exponential {
fn exp(&self) -> Self;
fn exp2(&self) -> Self;
fn expm1(&self) -> Self;
fn log(&self) -> Self;
fn log2(&self) -> Self;
fn log10(&self) -> Self;
}
pub trait Hyperbolic: Exponential {
fn sinh(&self) -> Self;
fn cosh(&self) -> Self;
fn tanh(&self) -> Self;
}
~~~
There was some discussion over whether we should shorten the names, for example `Trig` and `Exp`. No abbreviations have been agreed on yet, but this could be considered in the future.
Additionally, `Integer::divisible_by` has been renamed to `Integer::is_multiple_of`.
One of the tests seems to have no current equivalent that's similar. Please let me know if that's incorrect, and I'll try fixing it instead of deleting it. I suppose a struct could be used instead of `any` and `match type`, but it seems like the original intent of the test was to exercise `match type`
After discussions on IRC and #4819, we have decided to revert this change. This is due to the traits expressing different ideas and because hyperbolic functions are not trivially implementable from exponential functions for floating-point types.
The Hyperbolic Functions are trivially implemented in terms of `exp`, so it's simpler to group them the Exponential trait. In the future these would have default implementations.
