[backport] Assorted fixes and optimizations

core/string/string_builder.cpp

@@ -68,8 +68,10 @@ String StringBuilder::as_string() const {
 	int godot_string_elem = 0;
 	int c_string_elem = 0;
 
-	for (int i = 0; i < appended_strings.size(); i++) {
-		if (appended_strings[i] == -1) {
+	for (uint32_t i = 0; i < appended_strings.size(); i++) {
+		const int32_t str_len = appended_strings[i];
+
+		if (str_len == -1) {
 			// Godot string
 			const String &s = strings[godot_string_elem];
 
@@ -81,11 +83,11 @@ String StringBuilder::as_string() const {
 		} else {
 			const char *s = c_strings[c_string_elem];
 
-			for (int32_t j = 0; j < appended_strings[i]; j++) {
+			for (int32_t j = 0; j < str_len; j++) {
 				buffer[current_position + j] = s[j];
 			}
 
-			current_position += appended_strings[i];
+			current_position += str_len;
 
 			c_string_elem++;
 		}

core/string/string_builder.h

@@ -32,17 +32,17 @@
 #define STRING_BUILDER_H
 
 #include "core/string/ustring.h"
-#include "core/templates/vector.h"
+#include "core/templates/local_vector.h"
 
 class StringBuilder {
 	uint32_t string_length = 0;
 
-	Vector<String> strings;
-	Vector<const char *> c_strings;
+	LocalVector<String> strings;
+	LocalVector<const char *> c_strings;
 
 	// -1 means it's a Godot String
 	// a natural number means C string.
-	Vector<int32_t> appended_strings;
+	LocalVector<int32_t> appended_strings;
 
 public:
 	StringBuilder &append(const String &p_string);

core/string/string_name.cpp

@@ -186,12 +186,6 @@ bool StringName::operator!=(const char *p_name) const {
 	return !(operator==(p_name));
 }
 
-bool StringName::operator!=(const StringName &p_name) const {
-	// the real magic of all this mess happens here.
-	// this is why path comparisons are very fast
-	return _data != p_name._data;
-}
-
 char32_t StringName::operator[](int p_index) const {
 	if (_data) {
 		if (_data->cname) {

core/string/string_name.h

@@ -131,10 +131,13 @@ class StringName {
 		return _data >= p_name._data;
 	}
 	_FORCE_INLINE_ bool operator==(const StringName &p_name) const {
-		// the real magic of all this mess happens here.
-		// this is why path comparisons are very fast
+		// The real magic of all this mess happens here.
+		// This is why path comparisons are very fast.
 		return _data == p_name._data;
 	}
+	_FORCE_INLINE_ bool operator!=(const StringName &p_name) const {
+		return _data != p_name._data;
+	}
 	_FORCE_INLINE_ uint32_t hash() const {
 		if (_data) {
 			return _data->hash;
@@ -145,7 +148,6 @@ class StringName {
 	_FORCE_INLINE_ const void *data_unique_pointer() const {
 		return (void *)_data;
 	}
-	bool operator!=(const StringName &p_name) const;
 
 	_FORCE_INLINE_ operator String() const {
 		if (_data) {
@@ -185,8 +187,22 @@ class StringName {
 	};
 
 	StringName &operator=(const StringName &p_name);
+	StringName &operator=(StringName &&p_name) {
+		if (_data == p_name._data) {
+			return *this;
+		}
+
+		unref();
+		_data = p_name._data;
+		p_name._data = nullptr;
+		return *this;
+	}
 	StringName(const char *p_name, bool p_static = false);
 	StringName(const StringName &p_name);
+	StringName(StringName &&p_name) {
+		_data = p_name._data;
+		p_name._data = nullptr;
+	}
 	StringName(const String &p_name, bool p_static = false);
 	StringName(const StaticCString &p_static_string, bool p_static = false);
 	StringName() {}

core/string/ustring.cpp

@@ -319,21 +319,14 @@ void String::copy_from(const char *p_cstr) {
 
 	resize(len + 1); // include 0
 
+	const char *end = p_cstr + len;
 	char32_t *dst = ptrw();
 
-	for (size_t i = 0; i <= len; i++) {
-#if CHAR_MIN == 0
-		uint8_t c = p_cstr[i];
-#else
-		uint8_t c = p_cstr[i] >= 0 ? p_cstr[i] : uint8_t(256 + p_cstr[i]);
-#endif
-		if (c == 0 && i < len) {
-			print_unicode_error("NUL character", true);
-			dst[i] = _replacement_char;
-		} else {
-			dst[i] = c;
-		}
+	for (; p_cstr < end; ++p_cstr, ++dst) {
+		// If char is int8_t, a set sign bit will be reinterpreted as 256 - val implicitly.
+		*dst = static_cast<uint8_t>(*p_cstr);
 	}
+	*dst = 0;
 }
 
 void String::copy_from(const char *p_cstr, const int p_clip_to) {
@@ -356,22 +349,14 @@ void String::copy_from(const char *p_cstr, const int p_clip_to) {
 
 	resize(len + 1); // include 0
 
+	const char *end = p_cstr + len;
 	char32_t *dst = ptrw();
 
-	for (int i = 0; i < len; i++) {
-#if CHAR_MIN == 0
-		uint8_t c = p_cstr[i];
-#else
-		uint8_t c = p_cstr[i] >= 0 ? p_cstr[i] : uint8_t(256 + p_cstr[i]);
-#endif
-		if (c == 0) {
-			print_unicode_error("NUL character", true);
-			dst[i] = _replacement_char;
-		} else {
-			dst[i] = c;
-		}
+	for (; p_cstr < end; ++p_cstr, ++dst) {
+		// If char is int8_t, a set sign bit will be reinterpreted as 256 - val implicitly.
+		*dst = static_cast<uint8_t>(*p_cstr);
 	}
-	dst[len] = 0;
+	*dst = 0;
 }
 
 void String::copy_from(const wchar_t *p_cstr) {
@@ -463,27 +448,25 @@ void String::copy_from(const char32_t *p_cstr, const int p_clip_to) {
 // p_length <= p_char strlen
 void String::copy_from_unchecked(const char32_t *p_char, const int p_length) {
 	resize(p_length + 1);
+
+	const char32_t *end = p_char + p_length;
 	char32_t *dst = ptrw();
-	dst[p_length] = 0;
 
-	for (int i = 0; i < p_length; i++) {
-		if (p_char[i] == 0) {
-			print_unicode_error("NUL character", true);
-			dst[i] = _replacement_char;
+	for (; p_char < end; ++p_char, ++dst) {
+		const char32_t chr = *p_char;
+		if ((chr & 0xfffff800) == 0xd800) {
+			print_unicode_error(vformat("Unpaired surrogate (%x)", (uint32_t)chr));
+			*dst = _replacement_char;
 			continue;
 		}
-		if ((p_char[i] & 0xfffff800) == 0xd800) {
-			print_unicode_error(vformat("Unpaired surrogate (%x)", (uint32_t)p_char[i]));
-			dst[i] = _replacement_char;
-			continue;
-		}
-		if (p_char[i] > 0x10ffff) {
-			print_unicode_error(vformat("Invalid unicode codepoint (%x)", (uint32_t)p_char[i]));
-			dst[i] = _replacement_char;
+		if (chr > 0x10ffff) {
+			print_unicode_error(vformat("Invalid unicode codepoint (%x)", (uint32_t)chr));
+			*dst = _replacement_char;
 			continue;
 		}
-		dst[i] = p_char[i];
+		*dst = chr;
 	}
+	*dst = 0;
 }
 
 void String::operator=(const char *p_str) {
@@ -3969,17 +3952,17 @@ float String::similarity(const String &p_string) const {
 		return 0.0f;
 	}
 
-	Vector<String> src_bigrams = bigrams();
-	Vector<String> tgt_bigrams = p_string.bigrams();
-
-	int src_size = src_bigrams.size();
-	int tgt_size = tgt_bigrams.size();
+	const int src_size = length() - 1;
+	const int tgt_size = p_string.length() - 1;
 
-	int sum = src_size + tgt_size;
+	const int sum = src_size + tgt_size;
 	int inter = 0;
 	for (int i = 0; i < src_size; i++) {
+		const char32_t i0 = get(i);
+		const char32_t i1 = get(i + 1);
+
 		for (int j = 0; j < tgt_size; j++) {
-			if (src_bigrams[i] == tgt_bigrams[j]) {
+			if (i0 == p_string.get(j) && i1 == p_string.get(j + 1)) {
 				inter++;
 				break;
 			}

core/templates/list.h

@@ -522,6 +522,15 @@ class List {
 			it = it->next();
 		}
 	}
+	void operator=(List &&p_list) {
+		if (unlikely(this == &p_list)) {
+			return;
+		}
+
+		clear();
+		_data = p_list._data;
+		p_list._data = nullptr;
+	}
 
 	// Random access to elements, use with care,
 	// do not use for iteration.
@@ -760,6 +769,10 @@ class List {
 			it = it->next();
 		}
 	}
+	List(List &&p_list) {
+		_data = p_list._data;
+		p_list._data = nullptr;
+	}
 
 	List() {}
 

core/templates/local_vector.h

@@ -322,6 +322,16 @@ class LocalVector {
 			data[i] = p_from.data[i];
 		}
 	}
+	_FORCE_INLINE_ LocalVector(LocalVector &&p_from) {
+		data = p_from.data;
+		count = p_from.count;
+		capacity = p_from.capacity;
+
+		p_from.data = nullptr;
+		p_from.count = 0;
+		p_from.capacity = 0;
+	}
+
 	inline void operator=(const LocalVector &p_from) {
 		resize(p_from.size());
 		for (U i = 0; i < p_from.count; i++) {
@@ -334,6 +344,26 @@ class LocalVector {
 			data[i] = p_from[i];
 		}
 	}
+	inline void operator=(LocalVector &&p_from) {
+		if (unlikely(this == &p_from)) {
+			return;
+		}
+		reset();
+
+		data = p_from.data;
+		count = p_from.count;
+		capacity = p_from.capacity;
+
+		p_from.data = nullptr;
+		p_from.count = 0;
+		p_from.capacity = 0;
+	}
+	inline void operator=(Vector<T> &&p_from) {
+		resize(p_from.size());
+		for (U i = 0; i < count; i++) {
+			data[i] = std::move(p_from[i]);
+		}
+	}
 
 	_FORCE_INLINE_ ~LocalVector() {
 		if (data) {

core/typedefs.h

@@ -43,6 +43,7 @@
 // Should be available everywhere.
 #include "core/error/error_list.h"
 #include <cstdint>
+#include <utility>
 
 // Turn argument to string constant:
 // https://gcc.gnu.org/onlinedocs/cpp/Stringizing.html#Stringizing
@@ -126,13 +127,7 @@ constexpr auto CLAMP(const T m_a, const T2 m_min, const T3 m_max) {
 
 // Generic swap template.
 #ifndef SWAP
-#define SWAP(m_x, m_y) __swap_tmpl((m_x), (m_y))
-template <typename T>
-inline void __swap_tmpl(T &x, T &y) {
-	T aux = x;
-	x = y;
-	y = aux;
-}
+#define SWAP(m_x, m_y) std::swap((m_x), (m_y))
 #endif // SWAP
 
 /* Functions to handle powers of 2 and shifting. */

drivers/d3d12/rendering_device_driver_d3d12.cpp

@@ -2225,6 +2225,7 @@ RDD::FenceID RenderingDeviceDriverD3D12::fence_create() {
 
 Error RenderingDeviceDriverD3D12::fence_wait(FenceID p_fence) {
 	FenceInfo *fence = (FenceInfo *)(p_fence.id);
+	fence->d3d_fence->SetEventOnCompletion(fence->fence_value, fence->event_handle);
 	DWORD res = WaitForSingleObjectEx(fence->event_handle, INFINITE, FALSE);
 #ifdef PIX_ENABLED
 	PIXNotifyWakeFromFenceSignal(fence->event_handle);
@@ -2319,7 +2320,6 @@ Error RenderingDeviceDriverD3D12::command_queue_execute_and_present(CommandQueue
 			FenceInfo *fence = (FenceInfo *)(p_cmd_fence.id);
 			fence->fence_value++;
 			command_queue->d3d_queue->Signal(fence->d3d_fence.Get(), fence->fence_value);
-			fence->d3d_fence->SetEventOnCompletion(fence->fence_value, fence->event_handle);
 		}
 	}
 
@@ -2526,8 +2526,6 @@ Error RenderingDeviceDriverD3D12::swap_chain_resize(CommandQueueID p_cmd_queue,
 			break;
 	}
 
-	print_verbose("Using swap chain flags: " + itos(creation_flags) + ", sync interval: " + itos(sync_interval) + ", present flags: " + itos(present_flags));
-
 	if (swap_chain->d3d_swap_chain != nullptr && creation_flags != swap_chain->creation_flags) {
 		// The swap chain must be recreated if the creation flags are different.
 		_swap_chain_release(swap_chain);