[libcxx-commits] [libcxx] [libc++] Optimize `to_chars` for bases 2, 8 and 16 (PR #223222)

[David Bayer via libcxx-commits]

================
@@ -197,21 +224,36 @@ struct _LIBCPP_HIDDEN __integral<16> {
     if (__n > __cap)
       return {__last, errc::value_too_large};
 
-    __last             = __first + __n;
-    char* __p          = __last;
-    unsigned __divisor = 256;
-    while (__value > __divisor) {
-      unsigned __c = __value % __divisor;
+    __last    = __first + __n;
+    char* __p = __last;
+
+    constexpr auto __ncs_per_word   = 8u;
+    constexpr auto __divisor        = uint64_t{1} << (4 * __ncs_per_word);
+    constexpr auto __dep_mask       = uint64_t{0x0f0f'0f0f'0f0f'0f0f};
+    constexpr auto __add_six_mask   = uint64_t{0x0606'0606'0606'0606};
+    constexpr auto __high_bit_mask  = uint64_t{0x1010'1010'1010'1010};
+    constexpr auto __c0_offset_mask = uint64_t{0x3030'3030'3030'3030};
+    while (__value >= __divisor) {
+      const auto __digits = __builtin_elementwise_pdep(static_cast<uint64_t>(__value), __dep_mask);
+    const auto __alpha = ((__digits + __add_six_mask) & __high_bit_mask) >> 4;
+    auto __cs          = __digits + __c0_offset_mask + __alpha * 0x27;
----------------
davebayer wrote:

For base 16, `'0'` and `'a'` have different offsets, so we need to account for that

https://github.com/llvm/llvm-project/pull/223222