Refactor integer to float conversion

tgross35 · tgross35 · commit 84021e790b88 · 2024-10-30T12:25:22.000-05:00
Extract some common routines to separate functions in order to
deduplicate code and remove some of the magic.
diff --git a/src/float/conv.rs b/src/float/conv.rs
@@ -6,73 +6,161 @@ use super::Float;
 
 /// Conversions from integers to floats.
 ///
-/// These are hand-optimized bit twiddling code,
-/// which unfortunately isn't the easiest kind of code to read.
+/// The algorithm is explained here: <https://blog.m-ou.se/floats/>. It roughly does the following:
+/// - Calculate a base mantissa by shifting the integer into mantissa position. This gives us a
+///   mantissa _with the implicit bit set_!
+/// - Figure out if rounding needs to occur by classifying the bits that are to be truncated. Some
+///   patterns are used to simplify this. Adjust the mantissa with the result if needed.
+/// - Calculate the exponent based on the base-2 logarithm of `i` (leading zeros). Subtract one.
+/// - Shift the exponent and add the mantissa to create the final representation. Subtracting one
+///   from the exponent (above) accounts for the explicit bit being set in the mantissa.
 ///
-/// The algorithm is explained here: <https://blog.m-ou.se/floats/>
+/// # Terminology
+///
+/// - `i`: the original integer
+/// - `i_m`: the integer, shifted fully left (no leading zeros)
+/// - `n`: number of leading zeroes
+/// - `e`: the resulting exponent. Usually 1 is subtracted to offset the mantissa implicit bit.
+/// - `m_base`: the mantissa before adjusting for truncated bits. Implicit bit is usually set.
+/// - `adj`: the bits that will be truncated, possibly compressed in some way.
+/// - `m`: the resulting mantissa. Implicit bit is usually set.
 mod int_to_float {
+    use super::*;
+
+    /// Calculate the exponent from the number of leading zeros.
+    ///
+    /// Usually 1 is subtracted from this function's result, so that a mantissa with the implicit
+    /// bit set can be added back later.
+    fn exp<I: Int, F: Float<Int: CastFrom<u32>>>(n: u32) -> F::Int {
+        F::Int::cast_from(F::EXPONENT_BIAS - 1 + I::BITS - n)
+    }
+
+    /// Adjust a mantissa with dropped bits to perform correct rounding.
+    ///
+    /// The dropped bits should be exactly the bits that get truncated (left-aligned), but they
+    /// can be combined or compressed in some way that simplifies operations.
+    fn m_adj<F: Float>(m_base: F::Int, dropped_bits: F::Int) -> F::Int {
+        // Branchlessly extract a `1` if rounding up should happen, 0 otherwise
+        // This accounts for rounding to even.
+        let adj = (dropped_bits - (dropped_bits >> (F::BITS - 1) & !m_base)) >> (F::BITS - 1);
+
+        // Add one when we need to round up. Break ties to even.
+        m_base + adj
+    }
+
+    /// Shift the exponent to its position and add the mantissa.
+    ///
+    /// If the mantissa has the implicit bit set, the exponent should be one less than its actual
+    /// value to cancel it out.
+    fn repr<F: Float>(e: F::Int, m: F::Int) -> F::Int {
+        // + rather than | so the mantissa can overflow into the exponent
+        (e << F::SIGNIFICAND_BITS) + m
+    }
+
+    /// Shift distance from a left-aligned integer to a smaller float.
+    fn shift_f_lt_i<I: Int, F: Float>() -> u32 {
+        (I::BITS - F::BITS) + F::EXPONENT_BITS
+    }
+
+    /// Shift distance from an integer with `n` leading zeros to a smaller float.
+    fn shift_f_gt_i<I: Int, F: Float>(n: u32) -> u32 {
+        F::SIGNIFICAND_BITS - I::BITS + 1 + n
+    }
+
+    /// Perform a signed operation as unsigned, then add the sign back.
+    pub fn signed<I, F, Conv>(i: I, conv: Conv) -> F
+    where
+        F: Float,
+        I: Int,
+        F::Int: CastFrom<I>,
+        Conv: Fn(I::UnsignedInt) -> F::Int,
+    {
+        let sign_bit = F::Int::cast_from(i >> (I::BITS - 1)) << (F::BITS - 1);
+        F::from_bits(conv(i.unsigned_abs()) | sign_bit)
+    }
+
     pub fn u32_to_f32_bits(i: u32) -> u32 {
         if i == 0 {
             return 0;
         }
         let n = i.leading_zeros();
-        let a = (i << n) >> 8; // Significant bits, with bit 24 still in tact.
-        let b = (i << n) << 24; // Insignificant bits, only relevant for rounding.
-        let m = a + ((b - (b >> 31 & !a)) >> 31); // Add one when we need to round up. Break ties to even.
-        let e = 157 - n; // Exponent plus 127, minus one.
-        (e << 23) + m // + not |, so the mantissa can overflow into the exponent.
+        // Mantissa with implicit bit set (significant bits)
+        let m_base = (i << n) >> f32::EXPONENT_BITS;
+        // Bits that will be dropped (insignificant bits)
+        let adj = (i << n) << (f32::SIGNIFICAND_BITS + 1);
+        let m = m_adj::<f32>(m_base, adj);
+        let e = exp::<u32, f32>(n) - 1;
+        repr::<f32>(e, m)
     }
 
     pub fn u32_to_f64_bits(i: u32) -> u64 {
         if i == 0 {
             return 0;
         }
         let n = i.leading_zeros();
-        let m = (i as u64) << (21 + n); // Significant bits, with bit 53 still in tact.
-        let e = 1053 - n as u64; // Exponent plus 1023, minus one.
-        (e << 52) + m // Bit 53 of m will overflow into e.
+        // Mantissa with implicit bit set
+        let m = (i as u64) << shift_f_gt_i::<u32, f64>(n);
+        let e = exp::<u32, f64>(n) - 1;
+        repr::<f64>(e, m)
     }
 
     pub fn u64_to_f32_bits(i: u64) -> u32 {
         let n = i.leading_zeros();
-        let y = i.wrapping_shl(n);
-        let a = (y >> 40) as u32; // Significant bits, with bit 24 still in tact.
-        let b = (y >> 8 | y & 0xFFFF) as u32; // Insignificant bits, only relevant for rounding.
-        let m = a + ((b - (b >> 31 & !a)) >> 31); // Add one when we need to round up. Break ties to even.
-        let e = if i == 0 { 0 } else { 189 - n }; // Exponent plus 127, minus one, except for zero.
-        (e << 23) + m // + not |, so the mantissa can overflow into the exponent.
+        let i_m = i.wrapping_shl(n);
+        // Mantissa with implicit bit set
+        let m_base: u32 = (i_m >> shift_f_lt_i::<u64, f32>()) as u32;
+        // The entire lower half of `i` will be truncated (masked portion), plus the
+        // next `EXPONENT_BITS` bits.
+        let adj = (i_m >> f32::EXPONENT_BITS | i_m & 0xFFFF) as u32;
+        let m = m_adj::<f32>(m_base, adj);
+        let e = if i == 0 { 0 } else { exp::<u64, f32>(n) - 1 };
+        repr::<f32>(e, m)
     }
 
     pub fn u64_to_f64_bits(i: u64) -> u64 {
         if i == 0 {
             return 0;
         }
         let n = i.leading_zeros();
-        let a = (i << n) >> 11; // Significant bits, with bit 53 still in tact.
-        let b = (i << n) << 53; // Insignificant bits, only relevant for rounding.
-        let m = a + ((b - (b >> 63 & !a)) >> 63); // Add one when we need to round up. Break ties to even.
-        let e = 1085 - n as u64; // Exponent plus 1023, minus one.
-        (e << 52) + m // + not |, so the mantissa can overflow into the exponent.
+        // Mantissa with implicit bit set
+        let m_base = (i << n) >> f64::EXPONENT_BITS;
+        let adj = (i << n) << (f64::SIGNIFICAND_BITS + 1);
+        let m = m_adj::<f64>(m_base, adj);
+        let e = exp::<u64, f64>(n) - 1;
+        repr::<f64>(e, m)
     }
 
     pub fn u128_to_f32_bits(i: u128) -> u32 {
         let n = i.leading_zeros();
-        let y = i.wrapping_shl(n);
-        let a = (y >> 104) as u32; // Significant bits, with bit 24 still in tact.
-        let b = (y >> 72) as u32 | ((y << 32) >> 32 != 0) as u32; // Insignificant bits, only relevant for rounding.
-        let m = a + ((b - (b >> 31 & !a)) >> 31); // Add one when we need to round up. Break ties to even.
-        let e = if i == 0 { 0 } else { 253 - n }; // Exponent plus 127, minus one, except for zero.
-        (e << 23) + m // + not |, so the mantissa can overflow into the exponent.
+        let i_m = i.wrapping_shl(n); // Mantissa, shifted so the first bit is nonzero
+        let m_base: u32 = (i_m >> shift_f_lt_i::<u128, f32>()) as u32;
+
+        // Within the upper `F::BITS`, everything except for the signifcand
+        // gets truncated
+        let d1: u32 = (i_m >> (u128::BITS - f32::BITS - f32::SIGNIFICAND_BITS - 1)).cast();
+
+        // The entire rest of `i_m` gets truncated. Zero the upper `F::BITS` then just
+        // check if it is nonzero.
+        let d2: u32 = (i_m << f32::BITS >> f32::BITS != 0).into();
+        let adj = d1 | d2;
+
+        // Mantissa with implicit bit set
+        let m = m_adj::<f32>(m_base, adj);
+        let e = if i == 0 { 0 } else { exp::<u128, f32>(n) - 1 };
+        repr::<f32>(e, m)
     }
 
     pub fn u128_to_f64_bits(i: u128) -> u64 {
         let n = i.leading_zeros();
-        let y = i.wrapping_shl(n);
-        let a = (y >> 75) as u64; // Significant bits, with bit 53 still in tact.
-        let b = (y >> 11 | y & 0xFFFF_FFFF) as u64; // Insignificant bits, only relevant for rounding.
-        let m = a + ((b - (b >> 63 & !a)) >> 63); // Add one when we need to round up. Break ties to even.
-        let e = if i == 0 { 0 } else { 1149 - n as u64 }; // Exponent plus 1023, minus one, except for zero.
-        (e << 52) + m // + not |, so the mantissa can overflow into the exponent.
+        let i_m = i.wrapping_shl(n);
+        // Mantissa with implicit bit set
+        let m_base: u64 = (i_m >> shift_f_lt_i::<u128, f64>()) as u64;
+        // The entire lower half of `i` will be truncated (masked portion), plus the
+        // next `EXPONENT_BITS` bits.
+        let adj = (i_m >> f64::EXPONENT_BITS | i_m & 0xFFFF_FFFF) as u64;
+        let m = m_adj::<f64>(m_base, adj);
+        let e = if i == 0 { 0 } else { exp::<u128, f64>(n) - 1 };
+        repr::<f64>(e, m)
     }
 }
 
@@ -113,38 +201,32 @@ intrinsics! {
 intrinsics! {
     #[arm_aeabi_alias = __aeabi_i2f]
     pub extern "C" fn __floatsisf(i: i32) -> f32 {
-        let sign_bit = ((i >> 31) as u32) << 31;
-        f32::from_bits(int_to_float::u32_to_f32_bits(i.unsigned_abs()) | sign_bit)
+        int_to_float::signed(i, int_to_float::u32_to_f32_bits)
     }
 
     #[arm_aeabi_alias = __aeabi_i2d]
     pub extern "C" fn __floatsidf(i: i32) -> f64 {
-        let sign_bit = ((i >> 31) as u64) << 63;
-        f64::from_bits(int_to_float::u32_to_f64_bits(i.unsigned_abs()) | sign_bit)
+        int_to_float::signed(i, int_to_float::u32_to_f64_bits)
     }
 
     #[arm_aeabi_alias = __aeabi_l2f]
     pub extern "C" fn __floatdisf(i: i64) -> f32 {
-        let sign_bit = ((i >> 63) as u32) << 31;
-        f32::from_bits(int_to_float::u64_to_f32_bits(i.unsigned_abs()) | sign_bit)
+        int_to_float::signed(i, int_to_float::u64_to_f32_bits)
     }
 
     #[arm_aeabi_alias = __aeabi_l2d]
     pub extern "C" fn __floatdidf(i: i64) -> f64 {
-        let sign_bit = ((i >> 63) as u64) << 63;
-        f64::from_bits(int_to_float::u64_to_f64_bits(i.unsigned_abs()) | sign_bit)
+        int_to_float::signed(i, int_to_float::u64_to_f64_bits)
     }
 
     #[cfg_attr(target_os = "uefi", unadjusted_on_win64)]
     pub extern "C" fn __floattisf(i: i128) -> f32 {
-        let sign_bit = ((i >> 127) as u32) << 31;
-        f32::from_bits(int_to_float::u128_to_f32_bits(i.unsigned_abs()) | sign_bit)
+        int_to_float::signed(i, int_to_float::u128_to_f32_bits)
     }
 
     #[cfg_attr(target_os = "uefi", unadjusted_on_win64)]
     pub extern "C" fn __floattidf(i: i128) -> f64 {
-        let sign_bit = ((i >> 127) as u64) << 63;
-        f64::from_bits(int_to_float::u128_to_f64_bits(i.unsigned_abs()) | sign_bit)
+        int_to_float::signed(i, int_to_float::u128_to_f64_bits)
     }
 }
 
diff --git a/src/int/mod.rs b/src/int/mod.rs
@@ -83,6 +83,7 @@ pub(crate) trait Int: MinInt
 
     fn unsigned(self) -> Self::UnsignedInt;
     fn from_unsigned(unsigned: Self::UnsignedInt) -> Self;
+    fn unsigned_abs(self) -> Self::UnsignedInt;
 
     fn from_bool(b: bool) -> Self;
 
@@ -178,7 +179,6 @@ macro_rules! int_impl_common {
         fn wrapping_mul(self, other: Self) -> Self {
             <Self>::wrapping_mul(self, other)
         }
-
         fn wrapping_sub(self, other: Self) -> Self {
             <Self>::wrapping_sub(self, other)
         }
@@ -235,6 +235,10 @@ macro_rules! int_impl {
                 me
             }
 
+            fn unsigned_abs(self) -> Self {
+                self
+            }
+
             fn abs_diff(self, other: Self) -> Self {
                 if self < other {
                     other.wrapping_sub(self)
@@ -268,6 +272,10 @@ macro_rules! int_impl {
                 me as $ity
             }
 
+            fn unsigned_abs(self) -> Self::UnsignedInt {
+                self.unsigned_abs()
+            }
+
             fn abs_diff(self, other: Self) -> $uty {
                 self.wrapping_sub(other).wrapping_abs() as $uty
             }
diff --git a/testcrate/tests/conv.rs b/testcrate/tests/conv.rs
@@ -8,7 +8,7 @@ use compiler_builtins::float::Float;
 use rustc_apfloat::{Float as _, FloatConvert as _};
 use testcrate::*;
 
-mod int_to_float {
+mod i_to_f {
     use super::*;
 
     macro_rules! i_to_f {
