AK: Implement floating-point conversions for big-endian

Author: https://github.com/sideeffect42
Commit: https://github.com/LadybirdBrowser/ladybird/commit/1bc44376c0
Pull-request: https://github.com/LadybirdBrowser/ladybird/pull/415
Reviewed-by: https://github.com/ADKaster
Reviewed-by: https://github.com/DanShaders

AK/FloatingPoint.h

@@ -25,9 +25,15 @@ union FloatExtractor<f128> {
     static constexpr int exponent_bits = 15;
     static constexpr unsigned exponent_max = 32767;
     struct [[gnu::packed]] {
+#    if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
+        ComponentType sign : 1;
+        ComponentType exponent : 15;
+        ComponentType mantissa : 112;
+#    else
         ComponentType mantissa : 112;
         ComponentType exponent : 15;
         ComponentType sign : 1;
+#    endif
     };
     f128 d;
 };
@@ -51,9 +57,15 @@ union FloatExtractor<f80> {
         // However, since all bit-fiddling float code assumes IEEE floats, it cannot handle this properly.
         // If we pretend that 80-bit floats are IEEE floats with 64-bit mantissas, almost everything works correctly
         // and we just need a few special cases.
+#    if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
+        ComponentType sign : 1;
+        ComponentType exponent : 15;
+        ComponentType mantissa : 64;
+#    else
         ComponentType mantissa : 64;
         ComponentType exponent : 15;
         ComponentType sign : 1;
+#    endif
     };
     f80 d;
 };
@@ -76,9 +88,15 @@ union FloatExtractor<f64> {
         //        very intuitive and portable behaviour on windows, but it doesn't
         //        work with the msvc ABI.
         //        See <https://github.com/llvm/llvm-project/issues/24757>
+#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
+        ComponentType sign : 1;
+        ComponentType exponent : 11;
+        ComponentType mantissa : 52;
+#else
         ComponentType mantissa : 52;
         ComponentType exponent : 11;
         ComponentType sign : 1;
+#endif
     };
     f64 d;
 };
@@ -93,9 +111,15 @@ union FloatExtractor<f32> {
     static constexpr int exponent_bits = 8;
     static constexpr ComponentType exponent_max = 255;
     struct [[gnu::packed]] {
+#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
+        ComponentType sign : 1;
+        ComponentType exponent : 8;
+        ComponentType mantissa : 23;
+#else
         ComponentType mantissa : 23;
         ComponentType exponent : 8;
         ComponentType sign : 1;
+#endif
     };
     f32 d;
 };

AK/FloatingPointStringConversions.cpp

@@ -176,8 +176,6 @@ static constexpr auto max_representable_power_of_ten_in_u64 = 19;
 static_assert(1e19 <= static_cast<double>(NumericLimits<u64>::max()));
 static_assert(1e20 >= static_cast<double>(NumericLimits<u64>::max()));
 
-static_assert(HostIsLittleEndian, "Float parsing currently assumes little endian, this fact is only used in fast parsing of 8 digits at a time"
-                                  "\nyou _should_ only need to change read eight_digits to make this big endian compatible.");
 constexpr u64 read_eight_digits(char const* string)
 {
     u64 val;
@@ -206,6 +204,7 @@ constexpr static u32 eight_digits_to_value(u64 value)
 {
     // THIS DOES ABSOLUTELY ASSUME has_eight_digits is true
 
+    if constexpr (AK::HostIsLittleEndian) {
         // This trick is based on https://johnnylee-sde.github.io/Fast-numeric-string-to-int/
         // FIXME: fast_float uses a slightly different version, but that is far harder
         //        to understand and does not seem to improve performance substantially.
@@ -238,8 +237,24 @@ constexpr static u32 eight_digits_to_value(u64 value)
 
         // With the example this gives 0x$$$$$$$$00bc614e
         //                                       12345678
+
         // Now we just truncate to the lower part
         return u32(value);
+    } else {
+        value -= 0x3030303030303030;
+
+        value = (value & 0x0fL)
+            + ((value & (0x0fL << 8)) >> 8) * 10
+            + ((value & (0x0fL << 16)) >> 16) * 100
+            + ((value & (0x0fL << 24)) >> 24) * 1000
+            + ((value & (0x0fL << 32)) >> 32) * 10000
+            + ((value & (0x0fL << 40)) >> 40) * 100000
+            + ((value & (0x0fL << 48)) >> 48) * 1000000
+            + ((value & (0x0fL << 56)) >> 56) * 10000000;
+
+        // Now we just truncate to the lower part
+        return u32(value);
+    }
 }
 
 template<typename IsDoneCallback, typename Has8CharsLeftCallback>

AK/Math.h

@@ -8,6 +8,7 @@
 
 #include <AK/BuiltinWrappers.h>
 #include <AK/Concepts.h>
+#include <AK/Endian.h>
 #include <AK/FloatingPoint.h>
 #include <AK/NumericLimits.h>
 #include <AK/StdLibExtraDetails.h>
@@ -765,14 +766,21 @@ constexpr T log2(T x)
 
     // FIXME: Handle denormalized numbers separately
 
-    FloatExtractor<T> mantissa_ext {
+    // (1 <= mantissa < 2)
+    T m;
+    if constexpr (HostIsLittleEndian) {
+        m = ((FloatExtractor<T>) {
                  .mantissa = ext.mantissa,
                  .exponent = FloatExtractor<T>::exponent_bias,
-        .sign = ext.sign
+                 .sign = ext.sign })
-    };
+                .d;
-
+    } else {
-    // (1 <= mantissa < 2)
+        m = ((FloatExtractor<T>) {
-    T m = mantissa_ext.d;
+                 .sign = ext.sign,
+                 .exponent = FloatExtractor<T>::exponent_bias,
+                 .mantissa = ext.mantissa })
+                .d;
+    }
 
     // This is a reconstruction of one of Sun's algorithms
     // They use a transformation to lower the problem space,