[libc++] Add constant folding for optimized std::find variants

[philnik777]

Currently Clang isn't able to constant fold memchr and wmemchr. This adds some basic constant folding to __constexpr_memchr and __constexpr_wmemchr, mostly alleviating the problem.

While not necessarily representative of the real world, an optimized version of the find.pass.cpp test is reduced from ~13k to ~4.5k lines of assembly.

[llvmbot]

@llvm/pr-subscribers-libcxx

Author: Nikolas Klauser (philnik777)

Changes

Currently Clang isn't able to constant fold memchr and wmemchr. This adds some basic constant folding to __constexpr_memchr and __constexpr_wmemchr, mostly alleviating the problem.

While not necessarily representative of the real world, an optimized version of the find.pass.cpp test is reduced from ~13k to ~4.5k lines of assembly.

---

Full diff: https://github.com/llvm/llvm-project/pull/96491.diff

2 Files Affected:

• (modified) libcxx/include/__string/constexpr_c_functions.h (+13)
• (modified) libcxx/include/cwchar (+13)

diff --git a/libcxx/include/__string/constexpr_c_functions.h b/libcxx/include/__string/constexpr_c_functions.h
index a978f816f1897..cd5b63ae47d1e 100644
--- a/libcxx/include/__string/constexpr_c_functions.h
+++ b/libcxx/include/__string/constexpr_c_functions.h
@@ -132,6 +132,19 @@ _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX14 _Tp* __constexpr_memchr(_Tp*
   static_assert(sizeof(_Tp) == 1 && __libcpp_is_trivially_equality_comparable<_Tp, _Up>::value,
                 "Calling memchr on non-trivially equality comparable types is unsafe.");
 
+  if (__builtin_constant_p(__count) && __builtin_constant_p(__value)) {
+    for (; __count; --__count) {
+      if (!__builtin_constant_p(*__str))
+        break;
+      if (*__str == __value)
+        return __str;
+      ++__str;
+    }
+  }
+
+  if (__builtin_constant_p(__count) && __count == 0)
+    return nullptr;
+
   if (__libcpp_is_constant_evaluated()) {
 // use __builtin_char_memchr to optimize constexpr evaluation if we can
 #if _LIBCPP_STD_VER >= 17 && __has_builtin(__builtin_char_memchr)
diff --git a/libcxx/include/cwchar b/libcxx/include/cwchar
index 08cfac58c846a..76fdf08988dc5 100644
--- a/libcxx/include/cwchar
+++ b/libcxx/include/cwchar
@@ -231,6 +231,19 @@ _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX14 _Tp* __constexpr_wmemchr(_Tp
                     __libcpp_is_trivially_equality_comparable<_Tp, _Tp>::value,
                 "Calling wmemchr on non-trivially equality comparable types is unsafe.");
 
+  if (__builtin_constant_p(__count) && __builtin_constant_p(__value)) {
+    for (; __count; --__count) {
+      if (!__builtin_constant_p(*__str))
+        break;
+      if (*__str == __value)
+        return __str;
+      ++__str;
+    }
+  }
+
+  if (__builtin_constant_p(__count) && __count == 0)
+    return nullptr;
+
 #if __has_builtin(__builtin_wmemchr)
   if (!__libcpp_is_constant_evaluated()) {
     wchar_t __value_buffer = 0;

[ldionne]

I think we can get rid of this branch since __builtin_constant_p is always true inside a constant expression. IOW, the code path above is always taken when __libcpp_is_constant_evaluated() is true, so the branch below is basically dead now.

[ldionne]

Let's add some kind of comment explaining that this is to aid constant-folding and it also handles evaluation inside constant expressions, since that is not super obvious.

[ldionne]

Can we file a bug report against Clang to complain that __builtin_memchr doesn't constant-fold? It seems like an obvious QOI issue since Clang can already constant-evaluate that function.

[ldionne]

From live review: It seems like Clang only fails to constant fold this for a local array. Since that's the case, IMO it would make sense to file a bug against the optimizer instead of working around it in the library.

[philnik777]

I've submitted #114668 and #100929 to track the two issues.