Port mul intrinsics to traits

Also add a few features to the `intrinsics!` macro
This commit is contained in:
Alex Crichton 2017-06-22 22:36:37 -07:00
parent eeb44abacf
commit 275d1032b5
3 changed files with 211 additions and 82 deletions

View File

@ -21,10 +21,18 @@ pub mod udiv;
pub trait Int:
Copy +
PartialEq +
PartialOrd +
ops::AddAssign +
ops::Add<Output = Self> +
ops::Sub<Output = Self> +
ops::Div<Output = Self> +
ops::Shl<u32, Output = Self> +
ops::Shr<u32, Output = Self> +
ops::BitOr<Output = Self> +
// ops::BitAnd<Output = Self> +
ops::BitXor<Output = Self> +
ops::BitAnd<Output = Self> +
ops::BitAndAssign +
ops::Not<Output = Self> +
{
/// Type with the same width but other signedness
type OtherSign: Int;
@ -34,8 +42,8 @@ pub trait Int:
/// Returns the bitwidth of the int type
fn bits() -> u32;
/// Returns the zero representation of this number
fn zero() -> Self;
fn one() -> Self;
/// Extracts the sign from self and returns a tuple.
///
@ -51,6 +59,12 @@ pub trait Int:
/// Convert to a signed representation
fn unsigned(self) -> Self::UnsignedInt;
// copied from primitive integers, but put in a trait
fn max_value() -> Self;
fn min_value() -> Self;
fn wrapping_add(self, other: Self) -> Self;
fn wrapping_mul(self, other: Self) -> Self;
}
macro_rules! int_impl {
@ -63,6 +77,10 @@ macro_rules! int_impl {
0
}
fn one() -> Self {
1
}
fn bits() -> u32 {
$bits
}
@ -74,6 +92,22 @@ macro_rules! int_impl {
fn unsigned(self) -> $uty {
self
}
fn max_value() -> Self {
<Self>::max_value()
}
fn min_value() -> Self {
<Self>::min_value()
}
fn wrapping_add(self, other: Self) -> Self {
<Self>::wrapping_add(self, other)
}
fn wrapping_mul(self, other: Self) -> Self {
<Self>::wrapping_mul(self, other)
}
}
impl Int for $ity {
@ -88,6 +122,10 @@ macro_rules! int_impl {
0
}
fn one() -> Self {
1
}
fn extract_sign(self) -> (bool, $uty) {
if self < 0 {
(true, (!(self as $uty)).wrapping_add(1))
@ -99,6 +137,22 @@ macro_rules! int_impl {
fn unsigned(self) -> $uty {
self as $uty
}
fn max_value() -> Self {
<Self>::max_value()
}
fn min_value() -> Self {
<Self>::min_value()
}
fn wrapping_add(self, other: Self) -> Self {
<Self>::wrapping_add(self, other)
}
fn wrapping_mul(self, other: Self) -> Self {
<Self>::wrapping_mul(self, other)
}
}
}
}
@ -113,7 +167,9 @@ pub trait LargeInt: Int {
type HighHalf: Int;
fn low(self) -> Self::LowHalf;
fn low_as_high(low: Self::LowHalf) -> Self::HighHalf;
fn high(self) -> Self::HighHalf;
fn high_as_low(low: Self::HighHalf) -> Self::LowHalf;
fn from_parts(low: Self::LowHalf, high: Self::HighHalf) -> Self;
}
@ -126,9 +182,15 @@ macro_rules! large_int {
fn low(self) -> $tylow {
self as $tylow
}
fn low_as_high(low: $tylow) -> $tyhigh {
low as $tyhigh
}
fn high(self) -> $tyhigh {
(self >> $halfbits) as $tyhigh
}
fn high_as_low(high: $tyhigh) -> $tylow {
high as $tylow
}
fn from_parts(low: $tylow, high: $tyhigh) -> $ty {
low as $ty | ((high as $ty) << $halfbits)
}

View File

@ -1,95 +1,96 @@
use core::ops;
use int::LargeInt;
use int::Int;
macro_rules! mul {
($(#[$attr:meta])+ |
$abi:tt, $intrinsic:ident: $ty:ty) => {
/// Returns `a * b`
$(#[$attr])+
pub extern $abi fn $intrinsic(a: $ty, b: $ty) -> $ty {
let half_bits = <$ty>::bits() / 4;
let lower_mask = !0 >> half_bits;
let mut low = (a.low() & lower_mask).wrapping_mul(b.low() & lower_mask);
trait Mul: LargeInt {
fn mul(self, other: Self) -> Self {
let half_bits = Self::bits() / 4;
let lower_mask = !<<Self as LargeInt>::LowHalf>::zero() >> half_bits;
let mut low = (self.low() & lower_mask).wrapping_mul(other.low() & lower_mask);
let mut t = low >> half_bits;
low &= lower_mask;
t += (a.low() >> half_bits).wrapping_mul(b.low() & lower_mask);
t += (self.low() >> half_bits).wrapping_mul(other.low() & lower_mask);
low += (t & lower_mask) << half_bits;
let mut high = (t >> half_bits) as hty!($ty);
let mut high = Self::low_as_high(t >> half_bits);
t = low >> half_bits;
low &= lower_mask;
t += (b.low() >> half_bits).wrapping_mul(a.low() & lower_mask);
t += (other.low() >> half_bits).wrapping_mul(self.low() & lower_mask);
low += (t & lower_mask) << half_bits;
high += (t >> half_bits) as hty!($ty);
high += (a.low() >> half_bits).wrapping_mul(b.low() >> half_bits) as hty!($ty);
high = high.wrapping_add(a.high().wrapping_mul(b.low() as hty!($ty)))
.wrapping_add((a.low() as hty!($ty)).wrapping_mul(b.high()));
<$ty>::from_parts(low, high)
}
high += Self::low_as_high(t >> half_bits);
high += Self::low_as_high((self.low() >> half_bits).wrapping_mul(other.low() >> half_bits));
high = high.wrapping_add(self.high().wrapping_mul(Self::low_as_high(other.low())))
.wrapping_add(Self::low_as_high(self.low()).wrapping_mul(other.high()));
Self::from_parts(low, high)
}
}
macro_rules! mulo {
($intrinsic:ident: $ty:ty) => {
// Default is "C" ABI
mulo!($intrinsic: $ty, "C");
};
($intrinsic:ident: $ty:ty, $abi:tt) => {
/// Returns `a * b` and sets `*overflow = 1` if `a * b` overflows
#[cfg_attr(not(test), no_mangle)]
pub extern $abi fn $intrinsic(a: $ty, b: $ty, overflow: &mut i32) -> $ty {
impl Mul for u64 {}
impl Mul for i128 {}
trait Mulo: Int + ops::Neg<Output = Self> {
fn mulo(self, other: Self, overflow: &mut i32) -> Self {
*overflow = 0;
let result = a.wrapping_mul(b);
if a == <$ty>::min_value() {
if b != 0 && b != 1 {
let result = self.wrapping_mul(other);
if self == Self::min_value() {
if other != Self::zero() && other != Self::one() {
*overflow = 1;
}
return result;
}
if b == <$ty>::min_value() {
if a != 0 && a != 1 {
if other == Self::min_value() {
if self != Self::zero() && self != Self::one() {
*overflow = 1;
}
return result;
}
let sa = a >> (<$ty>::bits() - 1);
let abs_a = (a ^ sa) - sa;
let sb = b >> (<$ty>::bits() - 1);
let abs_b = (b ^ sb) - sb;
if abs_a < 2 || abs_b < 2 {
let sa = self >> (Self::bits() - 1);
let abs_a = (self ^ sa) - sa;
let sb = other >> (Self::bits() - 1);
let abs_b = (other ^ sb) - sb;
let two = Self::one() + Self::one();
if abs_a < two || abs_b < two {
return result;
}
if sa == sb {
if abs_a > <$ty>::max_value() / abs_b {
if abs_a > Self::max_value() / abs_b {
*overflow = 1;
}
} else {
if abs_a > <$ty>::min_value() / -abs_b {
if abs_a > Self::min_value() / -abs_b {
*overflow = 1;
}
}
result
}
}
}
#[cfg(not(all(feature = "c", target_arch = "x86")))]
mul!(#[cfg_attr(all(not(test), not(target_arch = "arm")), no_mangle)]
#[cfg_attr(all(not(test), target_arch = "arm"), inline(always))]
| "C", __muldi3: u64);
impl Mulo for i32 {}
impl Mulo for i64 {}
impl Mulo for i128 {}
#[cfg(not(target_arch = "arm"))]
mul!(#[cfg_attr(not(test), no_mangle)]
| "C", __multi3: i128);
intrinsics! {
#[cfg(not(all(feature = "c", target_arch = "x86")))]
pub extern "C" fn __muldi3(a: u64, b: u64) -> u64 {
a.mul(b)
}
#[cfg(target_arch = "arm")]
mul!(#[cfg_attr(not(test), no_mangle)]
| "aapcs", __multi3: i128);
#[aapcs_on_arm]
pub extern "C" fn __multi3(a: i128, b: i128) -> i128 {
a.mul(b)
}
mulo!(__mulosi4: i32);
mulo!(__mulodi4: i64);
pub extern "C" fn __mulosi4(a: i32, b: i32, oflow: &mut i32) -> i32 {
a.mulo(b, oflow)
}
#[cfg(all(windows, target_pointer_width="64"))]
mulo!(__muloti4: i128, "unadjusted");
#[cfg(not(all(windows, target_pointer_width="64")))]
mulo!(__muloti4: i128);
pub extern "C" fn __mulodi4(a: i64, b: i64, oflow: &mut i32) -> i64 {
a.mulo(b, oflow)
}
#[unadjusted_on_win64]
pub extern "C" fn __muloti4(a: i128, b: i128, oflow: &mut i32) -> i128 {
a.mulo(b, oflow)
}
}

View File

@ -1,5 +1,10 @@
macro_rules! intrinsics {
() => ();
// Anything which has a `not(feature = "c")` we'll generate a shim function
// which calls out to the C function if the `c` feature is enabled.
// Otherwise if the `c` feature isn't enabled then we'll just have a normal
// intrinsic.
(
#[cfg(not(all(feature = "c", $($cfg_clause:tt)*)))]
$(#[$attr:meta])*
@ -32,6 +37,67 @@ macro_rules! intrinsics {
intrinsics!($($rest)*);
);
// We recognize the `#[aapcs_only_on_arm]` attribute here and generate the
// same intrinsic but force it to have the `"aapcs"` calling convention on
// ARM and `"C"` elsewhere.
(
#[aapcs_on_arm]
$(#[$attr:meta])*
pub extern $abi:tt fn $name:ident( $($argname:ident: $ty:ty),* ) -> $ret:ty {
$($body:tt)*
}
$($rest:tt)*
) => (
#[cfg(target_arch = "arm")]
intrinsics! {
$(#[$attr])*
pub extern "aapcs" fn $name( $($argname: $ty),* ) -> $ret {
$($body)*
}
}
#[cfg(not(target_arch = "arm"))]
intrinsics! {
$(#[$attr])*
pub extern $abi fn $name( $($argname: $ty),* ) -> $ret {
$($body)*
}
}
intrinsics!($($rest)*);
);
// Like aapcs above we recognize an attribute for the "unadjusted" abi on
// win64 for some methods.
(
#[unadjusted_on_win64]
$(#[$attr:meta])*
pub extern $abi:tt fn $name:ident( $($argname:ident: $ty:ty),* ) -> $ret:ty {
$($body:tt)*
}
$($rest:tt)*
) => (
#[cfg(all(windows, target_pointer_width = "64"))]
intrinsics! {
$(#[$attr])*
pub extern "unadjusted" fn $name( $($argname: $ty),* ) -> $ret {
$($body)*
}
}
#[cfg(not(all(windows, target_pointer_width = "64")))]
intrinsics! {
$(#[$attr])*
pub extern $abi fn $name( $($argname: $ty),* ) -> $ret {
$($body)*
}
}
intrinsics!($($rest)*);
);
(
$(#[$attr:meta])*
pub extern $abi:tt fn $name:ident( $($argname:ident: $ty:ty),* ) -> $ret:ty {