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

[David Bayer via libcxx-commits]

================
@@ -126,21 +131,32 @@ struct _LIBCPP_HIDDEN __integral<2> {
     int __n         = __width(__value);
     if (__n > __cap)
       return {__last, errc::value_too_large};
-
     __last                   = __first + __n;
     char* __p                = __last;
-    const unsigned __divisor = 16;
-    while (__value > __divisor) {
-      unsigned __c = __value % __divisor;
+
+    constexpr auto __ncs_per_word   = 8u;
+    constexpr auto __divisor        = 1u << __ncs_per_word;
+    constexpr auto __dep_mask       = uint64_t{0x0101'0101'0101'0101};
+    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);
----------------
davebayer wrote:

Absolutely, I just wanted to discuss first whether you want this thing in before spending the time on portability and more performance analysis

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