[libc-commits] [libc] [libc] Add wcscasecmp and wcsncasecmp implementations. (PR #222154)

[Michael Jones via libc-commits]

================
@@ -796,6 +797,15 @@ is_char_or_wchar(wchar_t ch, [[maybe_unused]] char, wchar_t wc_value) {
   return (ch == wc_value);
 }
 
+// Returns a three-way comparison between two wide character code points.
+LIBC_INLINE int threeway_cmp_single(wchar_t left, wchar_t right) {
+  // Valid UTF-32 code points are in [0, 0x10FFFF]. If invalid code points were
+  // treated as UB, the comparison below could instead be done in 32 bits.
+  uint64_t diff = static_cast<uint64_t>(left) - static_cast<uint64_t>(right);
+  return cpp::bit_cast<int32_t>(static_cast<uint32_t>(diff >> 32) |
+                                static_cast<uint32_t>(diff & 0xFFFF));
+}
+
----------------
michaelrj-google wrote:

I was looking into if this optimization was effective on godbolt: https://godbolt.org/z/6exE3qnsj

It looks like the naive approach gets pretty much optimized away by the compiler. Given that I think the naive approach is probably a better option.

https://github.com/llvm/llvm-project/pull/222154