Auto merge of #27110 - arthurprs:optintstr, r=Gankro

I wrote a reasonably optimized version for both functions. Further optimizations are possible but I tried to keep the code size small (which I think is important), it's a road of diminished gains.

The repository used for testing/benchmarks is https://github.com/arthurprs/rust-optimized-inttostr

Benchmarks are ran for 3 different distributions, bellow are string length histograms for the u32 type

* _h (big numbers skew)
[0, 0, 5, 29, 103, 212, 551, 1138, 1887, 3196, 2879]
* _m (slight small number skew):
[0, 2807, 1334, 1057, 905, 821, 772, 707, 627, 605, 365]
* _l (small numbers skew):
[0, 8004, 567, 351, 248, 212, 170, 126, 136, 112, 74]

Tested processors are
* x64 laptop (i7-2670QM)
* x32 server (Digital Ocean E5-2630L-v2)

### Display

It uses a small look up table (200 bytes) and decode up to 4 characters at a time. I also took special precautions to reduce 64bit arithmetic on 32bit architectures and the gains are huge in these cases.

Overall, on modern 64bit CPUs it's pretty much the same speed as the stdlib implementation for very small numbers (0..99), but pulls ahead as the length of the decimal increases. On slight older CPUs (w/ worse ALUs) or 32bit architectures it's pretty much always faster.

x64 benchmarks
```
test bench::display_h_new_u08     ... bench:      71,041 ns/iter (+/- 2,894)
test bench::display_h_new_u16     ... bench:     378,255 ns/iter (+/- 36,547)
test bench::display_h_new_u32     ... bench:   4,232,483 ns/iter (+/- 509,661)
test bench::display_h_new_u64     ... bench:   5,166,740 ns/iter (+/- 421,124)
test bench::display_h_stdlib_u08  ... bench:      73,536 ns/iter (+/- 5,287)
test bench::display_h_stdlib_u16  ... bench:     451,443 ns/iter (+/- 16,879)
test bench::display_h_stdlib_u32  ... bench:   5,551,070 ns/iter (+/- 518,151)
test bench::display_h_stdlib_u64  ... bench:   8,624,374 ns/iter (+/- 643,701)
test bench::display_l_new_u08     ... bench:      71,547 ns/iter (+/- 504)
test bench::display_l_new_u16     ... bench:     399,727 ns/iter (+/- 28,030)
test bench::display_l_new_u32     ... bench:   4,365,303 ns/iter (+/- 414,414)
test bench::display_l_new_u64     ... bench:   5,302,382 ns/iter (+/- 292,324)
test bench::display_l_stdlib_u08  ... bench:      75,445 ns/iter (+/- 2,487)
test bench::display_l_stdlib_u16  ... bench:     444,313 ns/iter (+/- 16,203)
test bench::display_l_stdlib_u32  ... bench:   5,761,801 ns/iter (+/- 387,186)
test bench::display_l_stdlib_u64  ... bench:   8,790,365 ns/iter (+/- 614,846)
test bench::display_m_new_u08     ... bench:      71,820 ns/iter (+/- 2,956)
test bench::display_m_new_u16     ... bench:     399,649 ns/iter (+/- 20,643)
test bench::display_m_new_u32     ... bench:   4,355,561 ns/iter (+/- 179,189)
test bench::display_m_new_u64     ... bench:   5,070,594 ns/iter (+/- 341,950)
test bench::display_m_stdlib_u08  ... bench:      74,900 ns/iter (+/- 1,909)
test bench::display_m_stdlib_u16  ... bench:     448,788 ns/iter (+/- 20,791)
test bench::display_m_stdlib_u32  ... bench:   5,717,939 ns/iter (+/- 316,824)
test bench::display_m_stdlib_u64  ... bench:   8,787,160 ns/iter (+/- 482,864)
```

x86 benchmarks
```
test bench::display_h_new_u08     ... bench:      94,246 ns/iter (+/- 34,872)
test bench::display_h_new_u16     ... bench:     533,805 ns/iter (+/- 22,499)
test bench::display_h_new_u32     ... bench:   6,127,747 ns/iter (+/- 2,192,789)
test bench::display_h_new_u64     ... bench:  14,994,203 ns/iter (+/- 1,609,345)
test bench::display_h_stdlib_u08  ... bench:     107,233 ns/iter (+/- 8,571)
test bench::display_h_stdlib_u16  ... bench:     631,186 ns/iter (+/- 11,332)
test bench::display_h_stdlib_u32  ... bench:   7,696,344 ns/iter (+/- 957,917)
test bench::display_h_stdlib_u64  ... bench:  45,677,401 ns/iter (+/- 4,991,344)
test bench::display_l_new_u08     ... bench:      95,855 ns/iter (+/- 27,735)
test bench::display_l_new_u16     ... bench:     532,084 ns/iter (+/- 40,479)
test bench::display_l_new_u32     ... bench:   5,973,953 ns/iter (+/- 211,676)
test bench::display_l_new_u64     ... bench:  14,773,064 ns/iter (+/- 1,276,579)
test bench::display_l_stdlib_u08  ... bench:     106,350 ns/iter (+/- 63,963)
test bench::display_l_stdlib_u16  ... bench:     637,746 ns/iter (+/- 101,005)
test bench::display_l_stdlib_u32  ... bench:   7,740,640 ns/iter (+/- 848,478)
test bench::display_l_stdlib_u64  ... bench:  44,846,932 ns/iter (+/- 4,514,694)
test bench::display_m_new_u08     ... bench:      94,549 ns/iter (+/- 13,029)
test bench::display_m_new_u16     ... bench:     546,030 ns/iter (+/- 35,804)
test bench::display_m_new_u32     ... bench:   5,983,924 ns/iter (+/- 1,180,559)
test bench::display_m_new_u64     ... bench:  14,817,873 ns/iter (+/- 2,271,464)
test bench::display_m_stdlib_u08  ... bench:     107,806 ns/iter (+/- 8,805)
test bench::display_m_stdlib_u16  ... bench:     630,714 ns/iter (+/- 6,586)
test bench::display_m_stdlib_u32  ... bench:   7,784,210 ns/iter (+/- 358,601)
test bench::display_m_stdlib_u64  ... bench:  46,223,927 ns/iter (+/- 6,553,176)
```

### from_str_radix (FromStr)

All valid digits are ascii so I modified the function to use the underlining bytes instead and simplified the match to avoid wasting cycles.

x64 benchmarks
```
test bench::from_str_h_new_u08    ... bench:      28,153 ns/iter (+/- 624)
test bench::from_str_h_new_u16    ... bench:     223,513 ns/iter (+/- 11,554)
test bench::from_str_h_new_u32    ... bench:   3,098,935 ns/iter (+/- 231,022)
test bench::from_str_h_new_u64    ... bench:   5,009,900 ns/iter (+/- 341,961)
test bench::from_str_h_stdlib_u08 ... bench:      34,033 ns/iter (+/- 2,068)
test bench::from_str_h_stdlib_u16 ... bench:     248,785 ns/iter (+/- 14,208)
test bench::from_str_h_stdlib_u32 ... bench:   4,150,536 ns/iter (+/- 266,070)
test bench::from_str_h_stdlib_u64 ... bench:   6,817,997 ns/iter (+/- 449,838)
test bench::from_str_l_new_u08    ... bench:      27,552 ns/iter (+/- 1,500)
test bench::from_str_l_new_u16    ... bench:     234,360 ns/iter (+/- 13,144)
test bench::from_str_l_new_u32    ... bench:   3,140,261 ns/iter (+/- 248,175)
test bench::from_str_l_new_u64    ... bench:   5,176,583 ns/iter (+/- 350,416)
test bench::from_str_l_stdlib_u08 ... bench:      35,060 ns/iter (+/- 2,154)
test bench::from_str_l_stdlib_u16 ... bench:     252,135 ns/iter (+/- 23,461)
test bench::from_str_l_stdlib_u32 ... bench:   4,154,599 ns/iter (+/- 369,606)
test bench::from_str_l_stdlib_u64 ... bench:   6,892,767 ns/iter (+/- 213,030)
test bench::from_str_m_new_u08    ... bench:      28,252 ns/iter (+/- 1,384)
test bench::from_str_m_new_u16    ... bench:     231,051 ns/iter (+/- 16,540)
test bench::from_str_m_new_u32    ... bench:   3,166,504 ns/iter (+/- 134,418)
test bench::from_str_m_new_u64    ... bench:   5,103,195 ns/iter (+/- 218,912)
test bench::from_str_m_stdlib_u08 ... bench:      35,012 ns/iter (+/- 2,735)
test bench::from_str_m_stdlib_u16 ... bench:     250,967 ns/iter (+/- 14,708)
test bench::from_str_m_stdlib_u32 ... bench:   4,101,845 ns/iter (+/- 205,802)
test bench::from_str_m_stdlib_u64 ... bench:   6,823,001 ns/iter (+/- 267,215)
```

x86 benchmarks
```
test bench::from_str_h_new_u08    ... bench:      23,682 ns/iter (+/- 3,590)
test bench::from_str_h_new_u16    ... bench:     190,916 ns/iter (+/- 29,688)
test bench::from_str_h_new_u32    ... bench:   2,649,952 ns/iter (+/- 308,576)
test bench::from_str_h_new_u64    ... bench:  23,458,434 ns/iter (+/- 2,327,427)
test bench::from_str_h_stdlib_u08 ... bench:      45,551 ns/iter (+/- 6,968)
test bench::from_str_h_stdlib_u16 ... bench:     313,739 ns/iter (+/- 17,175)
test bench::from_str_h_stdlib_u32 ... bench:   4,615,669 ns/iter (+/- 470,766)
test bench::from_str_h_stdlib_u64 ... bench:  30,589,482 ns/iter (+/- 2,278,996)
test bench::from_str_l_new_u08    ... bench:      23,763 ns/iter (+/- 5,545)
test bench::from_str_l_new_u16    ... bench:     185,472 ns/iter (+/- 33,097)
test bench::from_str_l_new_u32    ... bench:   2,691,307 ns/iter (+/- 473,886)
test bench::from_str_l_new_u64    ... bench:  22,952,593 ns/iter (+/- 1,963,742)
test bench::from_str_l_stdlib_u08 ... bench:      45,285 ns/iter (+/- 16,337)
test bench::from_str_l_stdlib_u16 ... bench:     313,624 ns/iter (+/- 6,643)
test bench::from_str_l_stdlib_u32 ... bench:   4,595,679 ns/iter (+/- 1,876,361)
test bench::from_str_l_stdlib_u64 ... bench:  30,434,683 ns/iter (+/- 1,901,996)
test bench::from_str_m_new_u08    ... bench:      23,812 ns/iter (+/- 1,505)
test bench::from_str_m_new_u16    ... bench:     185,553 ns/iter (+/- 19,788)
test bench::from_str_m_new_u32    ... bench:   2,614,920 ns/iter (+/- 66,230)
test bench::from_str_m_new_u64    ... bench:  23,241,778 ns/iter (+/- 3,474,077)
test bench::from_str_m_stdlib_u08 ... bench:      45,634 ns/iter (+/- 1,436)
test bench::from_str_m_stdlib_u16 ... bench:     316,479 ns/iter (+/- 21,212)
test bench::from_str_m_stdlib_u32 ... bench:   4,609,147 ns/iter (+/- 487,068)
test bench::from_str_m_stdlib_u64 ... bench:  30,165,173 ns/iter (+/- 1,601,830)
```

Author: bors

src/libcore/fmt/num.rs

@@ -20,18 +20,25 @@ use fmt;
 use num::Zero;
 use ops::{Div, Rem, Sub};
 use str;
+use slice;
+use ptr;
+use mem;
 
 #[doc(hidden)]
 trait Int: Zero + PartialEq + PartialOrd + Div<Output=Self> + Rem<Output=Self> +
            Sub<Output=Self> + Copy {
     fn from_u8(u: u8) -> Self;
     fn to_u8(&self) -> u8;
+    fn to_u32(&self) -> u32;
+    fn to_u64(&self) -> u64;
 }
 
 macro_rules! doit {
     ($($t:ident)*) => ($(impl Int for $t {
         fn from_u8(u: u8) -> $t { u as $t }
         fn to_u8(&self) -> u8 { *self as u8 }
+        fn to_u32(&self) -> u32 { *self as u32 }
+        fn to_u64(&self) -> u64 { *self as u64 }
     })*)
 }
 doit! { i8 i16 i32 i64 isize u8 u16 u32 u64 usize }
@@ -188,6 +195,7 @@ macro_rules! radix_fmt {
         }
     }
 }
+
 macro_rules! int_base {
     ($Trait:ident for $T:ident as $U:ident -> $Radix:ident) => {
         #[stable(feature = "rust1", since = "1.0.0")]
@@ -209,17 +217,16 @@ macro_rules! debug {
         }
     }
 }
+
 macro_rules! integer {
     ($Int:ident, $Uint:ident) => {
-        int_base! { Display  for $Int as $Int   -> Decimal }
         int_base! { Binary   for $Int as $Uint  -> Binary }
         int_base! { Octal    for $Int as $Uint  -> Octal }
         int_base! { LowerHex for $Int as $Uint  -> LowerHex }
         int_base! { UpperHex for $Int as $Uint  -> UpperHex }
         radix_fmt! { $Int as $Int, fmt_int }
         debug! { $Int }
 
-        int_base! { Display  for $Uint as $Uint -> Decimal }
         int_base! { Binary   for $Uint as $Uint -> Binary }
         int_base! { Octal    for $Uint as $Uint -> Octal }
         int_base! { LowerHex for $Uint as $Uint -> LowerHex }
@@ -233,3 +240,80 @@ integer! { i8, u8 }
 integer! { i16, u16 }
 integer! { i32, u32 }
 integer! { i64, u64 }
+
+const DEC_DIGITS_LUT: &'static[u8] =
+    b"0001020304050607080910111213141516171819\
+      2021222324252627282930313233343536373839\
+      4041424344454647484950515253545556575859\
+      6061626364656667686970717273747576777879\
+      8081828384858687888990919293949596979899";
+
+macro_rules! impl_Display {
+    ($($t:ident),*: $conv_fn:ident) => ($(
+    impl fmt::Display for $t {
+        #[allow(unused_comparisons)]
+        fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+            let is_positive = *self >= 0;
+            let mut n = if is_positive {
+                self.$conv_fn()
+            } else {
+                // convert the negative num to positive by summing 1 to it's 2 complement
+                (!self.$conv_fn()).wrapping_add(1)
+            };
+            let mut buf: [u8; 20] = unsafe { mem::uninitialized() };
+            let mut curr = buf.len() as isize;
+            let buf_ptr = buf.as_mut_ptr();
+            let lut_ptr = DEC_DIGITS_LUT.as_ptr();
+
+            unsafe {
+                // eagerly decode 4 characters at a time
+                if <$t>::max_value() as u64 >= 10000 {
+                    while n >= 10000 {
+                        let rem = (n % 10000) as isize;
+                        n /= 10000;
+
+                        let d1 = (rem / 100) << 1;
+                        let d2 = (rem % 100) << 1;
+                        curr -= 4;
+                        ptr::copy_nonoverlapping(lut_ptr.offset(d1), buf_ptr.offset(curr), 2);
+                        ptr::copy_nonoverlapping(lut_ptr.offset(d2), buf_ptr.offset(curr + 2), 2);
+                    }
+                }
+
+                // if we reach here numbers are <= 9999, so at most 4 chars long
+                let mut n = n as isize; // possibly reduce 64bit math
+
+                // decode 2 more chars, if > 2 chars
+                if n >= 100 {
+                    let d1 = (n % 100) << 1;
+                    n /= 100;
+                    curr -= 2;
+                    ptr::copy_nonoverlapping(lut_ptr.offset(d1), buf_ptr.offset(curr), 2);
+                }
+
+                // decode last 1 or 2 chars
+                if n < 10 {
+                    curr -= 1;
+                    *buf_ptr.offset(curr) = (n as u8) + 48;
+                } else {
+                    let d1 = n << 1;
+                    curr -= 2;
+                    ptr::copy_nonoverlapping(lut_ptr.offset(d1), buf_ptr.offset(curr), 2);
+                }
+            }
+
+            let buf_slice = unsafe {
+                str::from_utf8_unchecked(
+                    slice::from_raw_parts(buf_ptr.offset(curr), buf.len() - curr as usize))
+            };
+            f.pad_integral(is_positive, "", buf_slice)
+        }
+    })*);
+}
+
+impl_Display!(i8, u8, i16, u16, i32, u32: to_u32);
+impl_Display!(i64, u64: to_u64);
+#[cfg(target_pointer_width = "32")]
+impl_Display!(isize, usize: to_u32);
+#[cfg(target_pointer_width = "64")]
+impl_Display!(isize, usize: to_u64);

src/libcore/num/mod.rs

@@ -24,6 +24,7 @@ use mem::size_of;
 use option::Option::{self, Some, None};
 use result::Result::{self, Ok, Err};
 use str::{FromStr, StrExt};
+use slice::SliceExt;
 
 /// Provides intentionally-wrapped arithmetic on `T`.
 ///
@@ -1448,19 +1449,30 @@ fn from_str_radix<T: FromStrRadixHelper>(src: &str, radix: u32)
                                          -> Result<T, ParseIntError> {
     use self::IntErrorKind::*;
     use self::ParseIntError as PIE;
+
     assert!(radix >= 2 && radix <= 36,
            "from_str_radix_int: must lie in the range `[2, 36]` - found {}",
            radix);
 
+    if src.is_empty() {
+        return Err(PIE { kind: Empty });
+    }
+
     let is_signed_ty = T::from_u32(0) > T::min_value();
 
-    match src.slice_shift_char() {
-        Some(('-', "")) => Err(PIE { kind: Empty }),
-        Some(('-', src)) if is_signed_ty => {
+    // all valid digits are ascii, so we will just iterate over the utf8 bytes
+    // and cast them to chars. .to_digit() will safely return None for anything
+    // other than a valid ascii digit for a the given radix, including the first-byte
+    // of multi-byte sequences
+    let src = src.as_bytes();
+
+    match (src[0], &src[1..])  {
+        (b'-', digits) if digits.is_empty() => Err(PIE { kind: Empty }),
+        (b'-', digits) if is_signed_ty => {
             // The number is negative
             let mut result = T::from_u32(0);
-            for c in src.chars() {
-                let x = match c.to_digit(radix) {
+            for &c in digits {
+                let x = match (c as char).to_digit(radix) {
                     Some(x) => x,
                     None => return Err(PIE { kind: InvalidDigit }),
                 };
@@ -1475,11 +1487,14 @@ fn from_str_radix<T: FromStrRadixHelper>(src: &str, radix: u32)
             }
             Ok(result)
         },
-        Some((_, _)) => {
+        (c, digits) => {
             // The number is signed
-            let mut result = T::from_u32(0);
-            for c in src.chars() {
-                let x = match c.to_digit(radix) {
+            let mut result = match (c as char).to_digit(radix) {
+                Some(x) => T::from_u32(x),
+                None => return Err(PIE { kind: InvalidDigit }),
+            };
+            for &c in digits {
+                let x = match (c as char).to_digit(radix) {
                     Some(x) => x,
                     None => return Err(PIE { kind: InvalidDigit }),
                 };
@@ -1493,8 +1508,7 @@ fn from_str_radix<T: FromStrRadixHelper>(src: &str, radix: u32)
                 };
             }
             Ok(result)
-        },
-        None => Err(ParseIntError { kind: Empty }),
+        }
     }
 }
 

src/libcoretest/num/mod.rs

@@ -117,7 +117,14 @@ mod tests {
     }
 
     #[test]
-    fn test_int_from_minus_sign() {
-        assert_eq!("-".parse::<i32>().ok(), None);
+    fn test_invalid() {
+        assert_eq!("--129".parse::<i8>().ok(), None);
+        assert_eq!("Съешь".parse::<u8>().ok(), None);
+    }
+
+    #[test]
+    fn test_empty() {
+        assert_eq!("-".parse::<i8>().ok(), None);
+        assert_eq!("".parse::<u8>().ok(), None);
     }
 }