Merge feature/simd-strlen into master

Kraionix · Kraionix · commit c5496ee771ee · 2025-03-01T14:42:22.000+07:00
diff --git a/imgui.cpp b/imgui.cpp
@@ -1959,6 +1959,152 @@ ImVec2 ImTriangleClosestPoint(const ImVec2& a, const ImVec2& b, const ImVec2& c,
 // [SECTION] MISC HELPERS/UTILITIES (String, Format, Hash functions)
 //-----------------------------------------------------------------------------
 
+#if defined IMGUI_ENABLE_AVX2_IMSTRLEN
+size_t ImStrlen(const char* str)
+{
+    const size_t SIMD_LENGTH = 32;
+    const size_t SIMD_LENGTH_MASK = SIMD_LENGTH - 1;
+
+    const unsigned char* begin = (unsigned char*)str;
+    const unsigned char* ptr = begin;
+
+    // first page
+    {
+        const size_t PAGE_LENGTH = 4096;
+        const size_t PAGE_LENGTH_MASK = PAGE_LENGTH - 1;
+
+        const unsigned char* page_end = (const unsigned char*)_andn_u64(PAGE_LENGTH_MASK, (uintptr_t)ptr + PAGE_LENGTH_MASK);
+        const unsigned char* align_page_end = (const unsigned char*)(page_end - SIMD_LENGTH);
+
+        // if ptr is far the end of page
+        if (ptr <= align_page_end)
+        {
+            __m256i target = _mm256_setzero_si256();
+
+            // if ptr not aligned, align ptr to SIMD_LENGTH
+            if ((uintptr_t)ptr & SIMD_LENGTH_MASK)
+            {
+                __m256i chunk = _mm256_lddqu_si256((const __m256i*)ptr);
+                int mask = _mm256_movemask_epi8(_mm256_cmpeq_epi8(chunk, target));
+
+                if (mask)
+                    return (uintptr_t)(ptr - begin + _tzcnt_u32(mask));
+
+                ptr = (const unsigned char*)_andn_u64(SIMD_LENGTH_MASK, (uintptr_t)ptr + SIMD_LENGTH_MASK);
+            }
+
+            // main loop of first page
+            for (; ptr <= align_page_end; ptr += SIMD_LENGTH)
+            {
+                __m256i chunk = _mm256_load_si256((const __m256i*)ptr);
+                int mask = _mm256_movemask_epi8(_mm256_cmpeq_epi8(chunk, target));
+
+                if (mask)
+                    return (uintptr_t)(ptr - begin + _tzcnt_u32(mask));
+
+                _mm_prefetch((const char*)ptr + 1024, _MM_HINT_T0);
+            }
+        }
+
+        // if ptr is near the end of page
+        for (; ptr < page_end; ptr++)
+        {
+            if (!(*ptr))
+                return (uintptr_t)(ptr - begin);
+        }
+    }
+
+    __m256i target = _mm256_setzero_si256();
+
+    // main loop
+    for (; ; ptr += SIMD_LENGTH)
+    {
+        __m256i chunk = _mm256_load_si256((const __m256i*)ptr);
+        int mask = _mm256_movemask_epi8(_mm256_cmpeq_epi8(chunk, target));
+
+        if (mask)
+            return (uintptr_t)(ptr - begin + _tzcnt_u32(mask));
+
+        _mm_prefetch((const char*)ptr + 1024, _MM_HINT_T0);
+    }
+}
+#elif defined IMGUI_ENABLE_SSE_IMSTRLEN
+size_t ImStrlen(const char* str)
+{
+    const size_t SIMD_LENGTH = 16;
+    const size_t SIMD_LENGTH_MASK = SIMD_LENGTH - 1;
+
+    const unsigned char* begin = (unsigned char*)str;
+    const unsigned char* ptr = begin;
+    const unsigned char ch = '\0';
+
+    // first page
+    {
+        const size_t PAGE_LENGTH = 4096;
+        const size_t PAGE_LENGTH_MASK = PAGE_LENGTH - 1;
+
+        const unsigned char* page_end = (const unsigned char*)(((uintptr_t)ptr + PAGE_LENGTH_MASK) & ~PAGE_LENGTH_MASK);
+        const unsigned char* align_page_end = (const unsigned char*)(page_end - SIMD_LENGTH);
+
+        // if ptr is far the end of page
+        if (ptr <= align_page_end)
+        {
+            __m128i target = _mm_set1_epi8(ch);
+
+            // if ptr not aligned, align ptr to SIMD_LENGTH
+            if ((uintptr_t)ptr & SIMD_LENGTH_MASK)
+            {
+                __m128i chunk = _mm_lddqu_si128((const __m128i*)ptr);
+                int mask = _mm_movemask_epi8(_mm_cmpeq_epi8(chunk, target));
+
+                if (mask)
+                    return (uintptr_t)(ptr + _tzcnt_u32(mask) - begin);
+
+                ptr = (const unsigned char*)(((uintptr_t)ptr + SIMD_LENGTH_MASK) & ~SIMD_LENGTH_MASK);
+            }
+
+            // main loop of first page
+            for (; ptr <= align_page_end; ptr += SIMD_LENGTH)
+            {
+                __m128i chunk = _mm_load_si128((const __m128i*)ptr);
+                int mask = _mm_movemask_epi8(_mm_cmpeq_epi8(chunk, target));
+
+                if (mask)
+                    return (uintptr_t)(ptr + _tzcnt_u32(mask) - begin);
+
+                _mm_prefetch((const char*)ptr + 1024, _MM_HINT_T0);
+            }
+        }
+
+        // if ptr is near the end of page
+        for (; ptr < page_end; ptr++)
+        {
+            if (*ptr == ch)
+                return (uintptr_t)(ptr - begin);
+        }
+    }
+
+    __m128i target = _mm_set1_epi8(ch);
+
+    // main loop
+    for (; ; ptr += SIMD_LENGTH)
+    {
+        __m128i chunk = _mm_load_si128((const __m128i*)ptr);
+        int mask = _mm_movemask_epi8(_mm_cmpeq_epi8(chunk, target));
+
+        if (mask)
+            return (uintptr_t)(ptr + _tzcnt_u32(mask) - begin);
+
+        _mm_prefetch((const char*)ptr + 1024, _MM_HINT_T0);
+    }
+}
+#else
+size_t ImStrlen(const char* str)
+{
+    return strlen(str);
+}
+#endif
+
 #if defined IMGUI_ENABLE_AVX2_IMMEMCHR
 const void* ImMemchr(const void* buf, int val, size_t count)
 {
@@ -2086,15 +2232,15 @@ void ImStrncpy(char* dst, const char* src, size_t count)
 
 char* ImStrdup(const char* str)
 {
-    size_t len = strlen(str);
+    size_t len = ImStrlen(str);
     void* buf = IM_ALLOC(len + 1);
     return (char*)memcpy(buf, (const void*)str, len + 1);
 }
 
 char* ImStrdupcpy(char* dst, size_t* p_dst_size, const char* src)
 {
-    size_t dst_buf_size = p_dst_size ? *p_dst_size : strlen(dst) + 1;
-    size_t src_size = strlen(src) + 1;
+    size_t dst_buf_size = p_dst_size ? *p_dst_size : ImStrlen(dst) + 1;
+    size_t src_size = ImStrlen(src) + 1;
     if (dst_buf_size < src_size)
     {
         IM_FREE(dst);
@@ -2128,7 +2274,7 @@ const char* ImStreolRange(const char* str, const char* str_end)
 
 const char* ImStrbol(const char* buf_mid_line, const char* buf_begin) // find beginning-of-line
 {
-    IM_ASSERT_PARANOID(buf_mid_line >= buf_begin && buf_mid_line <= buf_begin + strlen(buf_begin));
+    IM_ASSERT_PARANOID(buf_mid_line >= buf_begin && buf_mid_line <= buf_begin + ImStrlen(buf_begin));
     while (buf_mid_line > buf_begin && buf_mid_line[-1] != '\n')
         buf_mid_line--;
     return buf_mid_line;
@@ -2137,7 +2283,7 @@ const char* ImStrbol(const char* buf_mid_line, const char* buf_begin) // find be
 const char* ImStristr(const char* haystack, const char* haystack_end, const char* needle, const char* needle_end)
 {
     if (!needle_end)
-        needle_end = needle + strlen(needle);
+        needle_end = needle + ImStrlen(needle);
 
     const char un0 = (char)ImToUpper(*needle);
     while ((!haystack_end && *haystack) || (haystack_end && haystack < haystack_end))
@@ -2258,7 +2404,7 @@ void ImFormatStringToTempBufferV(const char** out_buf, const char** out_buf_end,
         if (buf == NULL)
             buf = "(null)";
         *out_buf = buf;
-        if (out_buf_end) { *out_buf_end = buf + strlen(buf); }
+        if (out_buf_end) { *out_buf_end = buf + ImStrlen(buf); }
     }
     else if (fmt[0] == '%' && fmt[1] == '.' && fmt[2] == '*' && fmt[3] == 's' && fmt[4] == 0)
     {
@@ -2667,7 +2813,7 @@ const char* ImTextFindPreviousUtf8Codepoint(const char* in_text_start, const cha
 int ImTextCountLines(const char* in_text, const char* in_text_end)
 {
     if (in_text_end == NULL)
-        in_text_end = in_text + strlen(in_text); // FIXME-OPT: Not optimal approach, discourage use for now.
+        in_text_end = in_text + ImStrlen(in_text); // FIXME-OPT: Not optimal approach, discourage use for now.
     int count = 0;
     while (in_text < in_text_end)
     {
@@ -2952,7 +3098,7 @@ void ImGuiTextFilter::ImGuiTextRange::split(char separator, ImVector<ImGuiTextRa
 void ImGuiTextFilter::Build()
 {
     Filters.resize(0);
-    ImGuiTextRange input_range(InputBuf, InputBuf + strlen(InputBuf));
+    ImGuiTextRange input_range(InputBuf, InputBuf + ImStrlen(InputBuf));
     input_range.split(',', &Filters);
 
     CountGrep = 0;
@@ -3020,7 +3166,7 @@ char ImGuiTextBuffer::EmptyString[1] = { 0 };
 
 void ImGuiTextBuffer::append(const char* str, const char* str_end)
 {
-    int len = str_end ? (int)(str_end - str) : (int)strlen(str);
+    int len = str_end ? (int)(str_end - str) : (int)ImStrlen(str);
 
     // Add zero-terminator the first time
     const int write_off = (Buf.Size != 0) ? Buf.Size : 1;
@@ -3703,7 +3849,7 @@ void ImGui::RenderText(ImVec2 pos, const char* text, const char* text_end, bool
     else
     {
         if (!text_end)
-            text_end = text + strlen(text); // FIXME-OPT
+            text_end = text + ImStrlen(text); // FIXME-OPT
         text_display_end = text_end;
     }
 
@@ -3721,7 +3867,7 @@ void ImGui::RenderTextWrapped(ImVec2 pos, const char* text, const char* text_end
     ImGuiWindow* window = g.CurrentWindow;
 
     if (!text_end)
-        text_end = text + strlen(text); // FIXME-OPT
+        text_end = text + ImStrlen(text); // FIXME-OPT
 
     if (text != text_end)
     {
@@ -4394,7 +4540,7 @@ ImGuiWindow::ImGuiWindow(ImGuiContext* ctx, const char* name) : DrawListInst(NUL
     memset(this, 0, sizeof(*this));
     Ctx = ctx;
     Name = ImStrdup(name);
-    NameBufLen = (int)strlen(name) + 1;
+    NameBufLen = (int)ImStrlen(name) + 1;
     ID = ImHashStr(name);
     IDStack.push_back(ID);
     MoveId = GetID("#MOVE");
@@ -8925,7 +9071,7 @@ const char* ImGui::GetKeyChordName(ImGuiKeyChord key_chord)
         (key != ImGuiKey_None || key_chord == ImGuiKey_None) ? GetKeyName(key) : "");
     size_t len;
     if (key == ImGuiKey_None && key_chord != 0)
-        if ((len = strlen(g.TempKeychordName)) != 0) // Remove trailing '+'
+        if ((len = ImStrlen(g.TempKeychordName)) != 0) // Remove trailing '+'
             g.TempKeychordName[len - 1] = 0;
     return g.TempKeychordName;
 }
@@ -14215,7 +14361,7 @@ bool ImGui::SetDragDropPayload(const char* type, const void* data, size_t data_s
         cond = ImGuiCond_Always;
 
     IM_ASSERT(type != NULL);
-    IM_ASSERT(strlen(type) < IM_ARRAYSIZE(payload.DataType) && "Payload type can be at most 32 characters long");
+    IM_ASSERT(ImStrlen(type) < IM_ARRAYSIZE(payload.DataType) && "Payload type can be at most 32 characters long");
     IM_ASSERT((data != NULL && data_size > 0) || (data == NULL && data_size == 0));
     IM_ASSERT(cond == ImGuiCond_Always || cond == ImGuiCond_Once);
     IM_ASSERT(payload.SourceId != 0); // Not called between BeginDragDropSource() and EndDragDropSource()
@@ -14459,7 +14605,7 @@ void ImGui::LogRenderedText(const ImVec2* ref_pos, const char* text, const char*
     }
 
     if (prefix)
-        LogRenderedText(ref_pos, prefix, prefix + strlen(prefix)); // Calculate end ourself to ensure "##" are included here.
+        LogRenderedText(ref_pos, prefix, prefix + ImStrlen(prefix)); // Calculate end ourself to ensure "##" are included here.
 
     // Re-adjust padding if we have popped out of our starting depth
     if (g.LogDepthRef > window->DC.TreeDepth)
@@ -14492,7 +14638,7 @@ void ImGui::LogRenderedText(const ImVec2* ref_pos, const char* text, const char*
     }
 
     if (suffix)
-        LogRenderedText(ref_pos, suffix, suffix + strlen(suffix));
+        LogRenderedText(ref_pos, suffix, suffix + ImStrlen(suffix));
 }
 
 // Start logging/capturing text output
@@ -14758,7 +14904,7 @@ void ImGui::LoadIniSettingsFromMemory(const char* ini_data, size_t ini_size)
     // For user convenience, we allow passing a non zero-terminated string (hence the ini_size parameter).
     // For our convenience and to make the code simpler, we'll also write zero-terminators within the buffer. So let's create a writable copy..
     if (ini_size == 0)
-        ini_size = strlen(ini_data);
+        ini_size = ImStrlen(ini_data);
     g.SettingsIniData.Buf.resize((int)ini_size + 1);
     char* const buf = g.SettingsIniData.Buf.Data;
     char* const buf_end = buf + ini_size;
@@ -14859,7 +15005,7 @@ ImGuiWindowSettings* ImGui::CreateNewWindowSettings(const char* name)
         if (const char* p = strstr(name, "###"))
             name = p;
     }
-    const size_t name_len = strlen(name);
+    const size_t name_len = ImStrlen(name);
 
     // Allocate chunk
     const size_t chunk_size = sizeof(ImGuiWindowSettings) + name_len + 1;
@@ -15151,7 +15297,7 @@ static void Platform_SetClipboardTextFn_DefaultImpl(ImGuiContext*, const char* t
     if (!main_clipboard)
         PasteboardCreate(kPasteboardClipboard, &main_clipboard);
     PasteboardClear(main_clipboard);
-    CFDataRef cf_data = CFDataCreate(kCFAllocatorDefault, (const UInt8*)text, strlen(text));
+    CFDataRef cf_data = CFDataCreate(kCFAllocatorDefault, (const UInt8*)text, ImStrlen(text));
     if (cf_data)
     {
         PasteboardPutItemFlavor(main_clipboard, (PasteboardItemID)1, CFSTR("public.utf8-plain-text"), cf_data, 0);
@@ -15205,7 +15351,7 @@ static void Platform_SetClipboardTextFn_DefaultImpl(ImGuiContext* ctx, const cha
 {
     ImGuiContext& g = *ctx;
     g.ClipboardHandlerData.clear();
-    const char* text_end = text + strlen(text);
+    const char* text_end = text + ImStrlen(text);
     g.ClipboardHandlerData.resize((int)(text_end - text) + 1);
     memcpy(&g.ClipboardHandlerData[0], text, (size_t)(text_end - text));
     g.ClipboardHandlerData[(int)(text_end - text)] = 0;
@@ -16975,7 +17121,7 @@ void ImGui::DebugHookIdInfo(ImGuiID id, ImGuiDataType data_type, const void* dat
         ImFormatString(info->Desc, IM_ARRAYSIZE(info->Desc), "%d", (int)(intptr_t)data_id);
         break;
     case ImGuiDataType_String:
-        ImFormatString(info->Desc, IM_ARRAYSIZE(info->Desc), "%.*s", data_id_end ? (int)((const char*)data_id_end - (const char*)data_id) : (int)strlen((const char*)data_id), (const char*)data_id);
+        ImFormatString(info->Desc, IM_ARRAYSIZE(info->Desc), "%.*s", data_id_end ? (int)((const char*)data_id_end - (const char*)data_id) : (int)ImStrlen((const char*)data_id), (const char*)data_id);
         break;
     case ImGuiDataType_Pointer:
         ImFormatString(info->Desc, IM_ARRAYSIZE(info->Desc), "(void*)0x%p", data_id);
diff --git a/imgui_draw.cpp b/imgui_draw.cpp
@@ -2672,7 +2672,7 @@ ImFont* ImFontAtlas::AddFontFromFileTTF(const char* filename, float size_pixels,
     {
         // Store a short copy of filename into into the font name for convenience
         const char* p;
-        for (p = filename + strlen(filename); p > filename && p[-1] != '/' && p[-1] != '\\'; p--) {}
+        for (p = filename + ImStrlen(filename); p > filename && p[-1] != '/' && p[-1] != '\\'; p--) {}
         ImFormatString(font_cfg.Name, IM_ARRAYSIZE(font_cfg.Name), "%s, %.0fpx", p, size_pixels);
     }
     return AddFontFromMemoryTTF(data, (int)data_size, size_pixels, &font_cfg, glyph_ranges);
@@ -2707,7 +2707,7 @@ ImFont* ImFontAtlas::AddFontFromMemoryCompressedTTF(const void* compressed_ttf_d
 
 ImFont* ImFontAtlas::AddFontFromMemoryCompressedBase85TTF(const char* compressed_ttf_data_base85, float size_pixels, const ImFontConfig* font_cfg, const ImWchar* glyph_ranges)
 {
-    int compressed_ttf_size = (((int)strlen(compressed_ttf_data_base85) + 4) / 5) * 4;
+    int compressed_ttf_size = (((int)ImStrlen(compressed_ttf_data_base85) + 4) / 5) * 4;
     void* compressed_ttf = IM_ALLOC((size_t)compressed_ttf_size);
     Decode85((const unsigned char*)compressed_ttf_data_base85, (unsigned char*)compressed_ttf);
     ImFont* font = AddFontFromMemoryCompressedTTF(compressed_ttf, compressed_ttf_size, size_pixels, font_cfg, glyph_ranges);
@@ -4029,7 +4029,7 @@ const char* ImFont::CalcWordWrapPositionA(float scale, const char* text, const c
 ImVec2 ImFont::CalcTextSizeA(float size, float max_width, float wrap_width, const char* text_begin, const char* text_end, const char** remaining)
 {
     if (!text_end)
-        text_end = text_begin + strlen(text_begin); // FIXME-OPT: Need to avoid this.
+        text_end = text_begin + ImStrlen(text_begin); // FIXME-OPT: Need to avoid this.
 
     const float line_height = size;
     const float scale = size / FontSize;
@@ -4129,7 +4129,7 @@ void ImFont::RenderText(ImDrawList* draw_list, float size, const ImVec2& pos, Im
         return;
 
     if (!text_end)
-        text_end = text_begin + strlen(text_begin); // ImGui:: functions generally already provides a valid text_end, so this is merely to handle direct calls.
+        text_end = text_begin + ImStrlen(text_begin); // ImGui:: functions generally already provides a valid text_end, so this is merely to handle direct calls.
 
     const float scale = size / FontSize;
     const float line_height = FontSize * scale;
diff --git a/imgui_internal.h b/imgui_internal.h
@@ -90,8 +90,10 @@ Index of this file:
 
 // Only AVX2 supports integer and byte instructions for 256-bit registers. Implementation this on AVX1 is not possible.
 #if defined(IMGUI_ENABLE_AVX2)
+#define IMGUI_ENABLE_AVX2_IMSTRLEN
 #define IMGUI_ENABLE_AVX2_IMMEMCHR
 #elif defined(IMGUI_ENABLE_AVX) || defined(IMGUI_ENABLE_SSE)
+#define IMGUI_ENABLE_SSE_IMSTRLEN
 #define IMGUI_ENABLE_SSE_IMMEMCHR
 #endif
 
@@ -394,6 +396,7 @@ static inline bool      ImIsPowerOfTwo(int v)           { return v != 0 && (v &
 static inline bool      ImIsPowerOfTwo(ImU64 v)         { return v != 0 && (v & (v - 1)) == 0; }
 static inline int       ImUpperPowerOfTwo(int v)        { v--; v |= v >> 1; v |= v >> 2; v |= v >> 4; v |= v >> 8; v |= v >> 16; v++; return v; }
 
+IMGUI_API size_t        ImStrlen(const char* str);                                          // Compute the length of a null-terminated string.
 IMGUI_API const void*   ImMemchr(const void* buf, int val, size_t count);                   // Find first occurrence of 'val' in buffer given length.
 IMGUI_API int           ImStricmp(const char* str1, const char* str2);                      // Case insensitive compare.
 IMGUI_API int           ImStrnicmp(const char* str1, const char* str2, size_t count);       // Case insensitive compare to a certain count.
diff --git a/imgui_tables.cpp b/imgui_tables.cpp
@@ -1648,7 +1648,7 @@ void ImGui::TableSetupColumn(const char* label, ImGuiTableColumnFlags flags, flo
     if (label != NULL && label[0] != 0)
     {
         column->NameOffset = (ImS16)table->ColumnsNames.size();
-        table->ColumnsNames.append(label, label + strlen(label) + 1);
+        table->ColumnsNames.append(label, label + ImStrlen(label) + 1);
     }
 }
 
diff --git a/imgui_widgets.cpp b/imgui_widgets.cpp

Original file line number	Diff line number	Diff line change
`@@ -1648,7 +1648,7 @@ void ImGui::TableSetupColumn(const char* label, ImGuiTableColumnFlags flags, flo`
`1648`	`1648`	`if (label != NULL && label[0] != 0)`
`1649`	`1649`	`{`
`1650`	`1650`	`column->NameOffset = (ImS16)table->ColumnsNames.size();`
`1651`		`- table->ColumnsNames.append(label, label + strlen(label) + 1);`
	`1651`	`+ table->ColumnsNames.append(label, label + ImStrlen(label) + 1);`
`1652`	`1652`	`}`
`1653`	`1653`	`}`
`1654`	`1654`