[libc++] Optimize std::find if types are integral and have the same signedness

[philnik777]

Fixes #70238

[ldionne]

Overall I think this makes sense, but it's tricky so we really need some tests.

Also, this is the kind of stuff that I would really expect something like Clang IR to render completely obsolete (along with most of the memchr-related optimizations, I guess). But we're not there yet I guess.

CC @bcardosolopes for awareness -- we have a bunch of these optimizations where we try to lower things like std::find(char*, char*, char) to a call to memchr, and we're thinking that Clang IR might make it easy for the compiler to understand that this optimization is possible without us telling it. In case you have thoughts about this, we'd love to hear them.

[llvmbot]

@llvm/pr-subscribers-libcxx

Author: None (philnik777)

Changes

Fixes #70238

---

Full diff: https://github.com/llvm/llvm-project/pull/70345.diff

1 Files Affected:

• (modified) libcxx/include/__algorithm/find.h (+13)

diff --git a/libcxx/include/__algorithm/find.h b/libcxx/include/__algorithm/find.h
index d7c268bc6b338b0..019848158c174df 100644
--- a/libcxx/include/__algorithm/find.h
+++ b/libcxx/include/__algorithm/find.h
@@ -73,6 +73,19 @@ __find_impl(_Tp* __first, _Tp* __last, const _Up& __value, _Proj&) {
 }
 #endif // _LIBCPP_HAS_NO_WIDE_CHARACTERS
 
+template <class _Tp,
+          class _Up,
+          class _Proj,
+          __enable_if_t<__is_identity<_Proj>::value && !__libcpp_is_trivially_equality_comparable<_Tp, _Up>::value &&
+                            is_integral<_Tp>::value && is_integral<_Up>::value,
+                        int> = 0>
+_LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX14 _Tp*
+__find_impl(_Tp* __first, _Tp* __last, const _Up& __value, _Proj& __proj) {
+  if (__value < numeric_limits<_Tp>::min() || __value > numeric_limits<_Tp>::max())
+    return __last;
+  return std::__find_impl(__first, __last, _Tp(__value), __proj);
+}
+
 // __bit_iterator implementation
 template <bool _ToFind, class _Cp, bool _IsConst>
 _LIBCPP_CONSTEXPR_SINCE_CXX20 _LIBCPP_HIDE_FROM_ABI __bit_iterator<_Cp, _IsConst>

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[ldionne]

IMO it would be reasonable to land this improvement without handling different signedness for now.

[EricWF]

Are there benchmark results for this?

[philnik777]

Are there benchmark results for this?

https://reviews.llvm.org/D144394#4138976

[ldionne]

I think this looks good to me if we have tests with other than 0.

[ldionne]

Actually I am going to approve this but please add the non-zero tests, and ping me again if you need to change the implementation when you add these tests.