From a2c6059905bb9cb719e58f11f31d52102f2ad49a Mon Sep 17 00:00:00 2001
From: Ole Erik Peistorpet <OleErikPeistorpet@users.noreply.github.com>
Date: Sun, 26 May 2024 19:30:00 +0200
Subject: [PATCH] Guiding the compiler: when to inline inside append,
 emplace_back

---
 auxi/core_util.h | 15 ++++++++++++
 dynarray.h       | 62 ++++++++++++++++++++++++++----------------------
 2 files changed, 49 insertions(+), 28 deletions(-)

diff --git a/auxi/core_util.h b/auxi/core_util.h
index cdbdc5d6..e592f193 100644
--- a/auxi/core_util.h
+++ b/auxi/core_util.h
@@ -28,6 +28,21 @@
 #endif
 
 
+#ifndef OEL_HAS_LIKELY
+	#if (__cplusplus >= 202001 and (!defined __clang__ or __clang_major__ >= 12)) or (defined _MSC_VER and _MSC_VER >= 1929)
+	#define OEL_HAS_LIKELY  1
+	#else
+	#define OEL_HAS_LIKELY  0
+	#endif
+#endif
+
+#if OEL_HAS_LIKELY
+	#define OEL_UNLIKELY  [[unlikely]]
+#else
+	#define OEL_UNLIKELY
+#endif
+
+
 #ifdef __GNUC__
 	#define OEL_ALWAYS_INLINE __attribute__((always_inline))
 #else
diff --git a/dynarray.h b/dynarray.h
index 1838cb7e..56c9dfdb 100644
--- a/dynarray.h
+++ b/dynarray.h
@@ -401,19 +401,9 @@ class dynarray
 		(void) _debugSizeUpdater{_m};
 	}
 
-#ifdef _MSC_VER
-	__declspec(noinline) // to get the compiler to inline calling function
-#endif
-	void _growBy(size_type const count)
-	{
-		auto const s = size();
-		_realloc(_calcCapAdd(count, s), s);
-	}
-
-	void _growByOne()
-	{
-		_realloc(_calcCapAddOne(), size());
-	}
+	// These are not defined inline as a compiler hint
+	void _growBy(size_type const);
+	void _growByOne();
 
 
 	template< typename UninitFiller >
@@ -522,7 +512,7 @@ class dynarray
 	template< typename InputIter >
 	InputIter _doAppend(InputIter src, size_type const count)
 	{
-		if (_spareCapacity() < count)
+		if (_spareCapacity() < count) OEL_UNLIKELY
 			_growBy(count);
 
 		if constexpr (can_memmove_with<T *, InputIter>)
@@ -681,6 +671,23 @@ typename dynarray<T, Alloc>::iterator
 	return _detail::MakeDynarrIter(_m, pPos);
 }
 
+
+template< typename T, typename Alloc >
+#if defined _MSC_VER and !OEL_HAS_LIKELY
+	__declspec(noinline) // to get the compiler to inline calling function
+#endif
+void dynarray<T, Alloc>::_growBy(size_type const count)
+{
+	auto const s = size();
+	_realloc(_calcCapAdd(count, s), s);
+}
+
+template< typename T, typename Alloc >
+void dynarray<T, Alloc>::_growByOne()
+{
+	_realloc(_calcCapAddOne(), size());
+}
+
 template< typename T, typename Alloc >
 template< typename... Args >
 inline T & dynarray<T, Alloc>::emplace_back(Args &&... args) &
@@ -695,6 +702,19 @@ inline T & dynarray<T, Alloc>::emplace_back(Args &&... args) &
 	return *(_m.end++);
 }
 
+template< typename T, typename Alloc >
+inline void dynarray<T, Alloc>::append(size_type count, const T & val)
+{
+	if (_spareCapacity() < count) OEL_UNLIKELY
+		_growBy(count);
+
+	auto const pos = _m.end;
+	_uninitFill::template call< forward_t<const T &> >(pos, pos + count, _m, val);
+
+	_debugSizeUpdater guard{_m};
+	_m.end += count;
+}
+
 
 template< typename T, typename Alloc >
 dynarray<T, Alloc>::dynarray(dynarray && other, Alloc a)
@@ -805,20 +825,6 @@ void dynarray<T, Alloc>::shrink_to_fit()
 }
 
 
-template< typename T, typename Alloc >
-inline void dynarray<T, Alloc>::append(size_type count, const T & val)
-{
-	if (_spareCapacity() < count)
-		_growBy(count);
-
-	auto const pos = _m.end;
-	_uninitFill::template call< forward_t<const T &> >(pos, pos + count, _m, val);
-
-	_debugSizeUpdater guard{_m};
-	_m.end += count;
-}
-
-
 template< typename T, typename Alloc >
 inline void dynarray<T, Alloc>::pop_back() noexcept
 {