From 88ace6e4aeb607d2239d80cc1693140d92f7bccd Mon Sep 17 00:00:00 2001
From: Stenzek <stenzek@gmail.com>
Date: Tue, 14 May 2024 23:50:26 +1000
Subject: [PATCH] CMake: Detect cache line size dynamically on AArch64 Linux

---
 CMakeLists.txt                      |  1 +
 CMakeModules/DuckStationUtils.cmake | 55 +++++++++++++++++++++++++++++
 src/common/CMakeLists.txt           |  3 ++
 src/common/types.h                  |  4 ++-
 4 files changed, 62 insertions(+), 1 deletion(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index d41ed7619..b4e9ba6df 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -29,6 +29,7 @@ detect_operating_system()
 detect_compiler()
 detect_architecture()
 detect_page_size()
+detect_cache_line_size()
 
 # Build options. Depends on system attributes.
 include(DuckStationBuildOptions)
diff --git a/CMakeModules/DuckStationUtils.cmake b/CMakeModules/DuckStationUtils.cmake
index 7088df29e..b2dd91c6a 100644
--- a/CMakeModules/DuckStationUtils.cmake
+++ b/CMakeModules/DuckStationUtils.cmake
@@ -139,3 +139,58 @@ int main() {
     set(HOST_PAGE_SIZE ${detect_page_size_output} CACHE STRING "Reported host page size")
   endif()
 endfunction()
+
+function(detect_cache_line_size)
+  # This is only needed for ARM64, or if the user hasn't overridden it explicitly.
+  if(NOT CPU_ARCH_ARM64 OR HOST_CACHE_LINE_SIZE)
+    return()
+  endif()
+
+  if(NOT LINUX)
+    # For universal Apple builds, we use preprocessor macros to determine page size.
+    # Similar for Windows, except it's always 64 bytes.
+    return()
+  endif()
+
+  if(CMAKE_CROSSCOMPILING)
+    message(WARNING "Cross-compiling and can't determine page size, assuming default.")
+    return()
+  endif()
+
+  message(STATUS "Determining host cache line size")
+  set(detect_cache_line_size_file ${CMAKE_BINARY_DIR}${CMAKE_FILES_DIRECTORY}/CMakeTmp/src.c)
+  file(WRITE ${detect_cache_line_size_file} "
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+int main() {
+  int l1i = sysconf(_SC_LEVEL1_DCACHE_LINESIZE);
+  int l1d = sysconf(_SC_LEVEL1_ICACHE_LINESIZE);
+  int res = (l1i > l1d) ? l1i : l1d;
+  for (int index = 0; index < 16; index++) {
+    char buf[128];
+    snprintf(buf, sizeof(buf), \"/sys/devices/system/cpu/cpu0/cache/index%d/coherency_line_size\", index);
+    FILE* fp = fopen(buf, \"rb\");
+    if (!fp)
+      break;
+    fread(buf, sizeof(buf), 1, fp);
+    fclose(fp);
+    int val = atoi(buf);
+    res = (val > res) ? val : res;
+  }
+  printf(\"%d\", res);
+  return (res > 0) ? EXIT_SUCCESS : EXIT_FAILURE;
+}")
+  try_run(
+    detect_cache_line_size_run_result
+    detect_cache_line_size_compile_result
+    ${CMAKE_BINARY_DIR}${CMAKE_FILES_DIRECTORY}
+    ${detect_cache_line_size_file}
+    RUN_OUTPUT_VARIABLE detect_cache_line_size_output)
+  if(NOT detect_cache_line_size_compile_result OR NOT detect_cache_line_size_run_result EQUAL 0)
+    message(FATAL_ERROR "Could not determine host cache line size.")
+  else()
+    message(STATUS "Host cache line size: ${detect_cache_line_size_output}")
+    set(HOST_CACHE_LINE_SIZE ${detect_cache_line_size_output} CACHE STRING "Reported host cache line size")
+  endif()
+endfunction()
diff --git a/src/common/CMakeLists.txt b/src/common/CMakeLists.txt
index e60015533..d2a1fa52f 100644
--- a/src/common/CMakeLists.txt
+++ b/src/common/CMakeLists.txt
@@ -111,3 +111,6 @@ endif()
 if(HOST_PAGE_SIZE)
   target_compile_definitions(common PUBLIC "-DOVERRIDE_HOST_PAGE_SIZE=${HOST_PAGE_SIZE}")
 endif()
+if(HOST_CACHE_LINE_SIZE)
+	target_compile_definitions(common PUBLIC "-DOVERRIDE_HOST_CACHE_LINE_SIZE=${HOST_CACHE_LINE_SIZE}")
+endif()
diff --git a/src/common/types.h b/src/common/types.h
index 3ea9afe0d..849d51d72 100644
--- a/src/common/types.h
+++ b/src/common/types.h
@@ -192,7 +192,9 @@ static constexpr u32 HOST_PAGE_SHIFT = 12;
 #endif
 
 // Host cache line sizes.
-#if defined(__APPLE__) && defined(__aarch64__)
+#if defined(OVERRIDE_HOST_CACHE_LINE_SIZE)
+static constexpr u32 HOST_CACHE_LINE_SIZE = OVERRIDE_HOST_CACHE_LINE_SIZE;
+#elif defined(__APPLE__) && defined(__aarch64__)
 static constexpr u32 HOST_CACHE_LINE_SIZE = 128; // Apple Silicon uses 128b cache lines.
 #else
 static constexpr u32 HOST_CACHE_LINE_SIZE = 64; // Everything else is 64b.