Quest: optimize compilation rules, adjust stack protection, and increase indirect dispatch cache size

android/QuestGameKit.psm1

@@ -505,7 +505,9 @@ function Invoke-QuestGameBuild {
     $escape = { param([string]$p) $p.Replace('$', '$$').Replace(':', '$:').Replace(' ', '$ ') }
     [void]$ninjaText.AppendLine('ninja_required_version = 1.10')
     [void]$ninjaText.AppendLine("pool translated`n  depth = $TranslatedJobs")
-    [void]$ninjaText.AppendLine("rule cxx`n  command = $(ConvertTo-QuotedArgument $ClangCxx) @`$flags -c `$in -o `$out`n  description = Compiling `$in")
+    # Depfiles, so a changed runtime header recompiles the shards that include it instead of
+    # leaving objects built against the previous kit.
+    [void]$ninjaText.AppendLine("rule cxx`n  command = $(ConvertTo-QuotedArgument $ClangCxx) @`$flags -MD -MF `$out.d -c `$in -o `$out`n  depfile = `$out.d`n  deps = gcc`n  description = Compiling `$in")
     [void]$ninjaText.AppendLine("rule asm`n  command = $(ConvertTo-QuotedArgument $ClangC) @`$flags -c `$in -o `$out`n  description = Assembling `$in")
     [void]$ninjaText.AppendLine("rule link`n  command = $(ConvertTo-QuotedArgument $ClangCxx) @`$flags -o `$out @`$out.rsp`n  rspfile = `$out.rsp`n  rspfile_content = `$inputs`n  description = Linking `$out")
     $objectsOf = @{}

docs/quest-port.md

@@ -419,6 +419,16 @@ Android facts this design rests on, all measured on a Quest 3:
   SDL3 built shared (or `-DAURORA_SDL3_PROVIDER=system` for the AAR prefab),
   tests off, products built as `libmain.so` / `libmain_retro_rewind.so`,
   `-mcpu=cortex-a77` (Quest 2's XR2 Gen 1; Quest 3/Pro are supersets).
+  The products link with `-Wl,-Bsymbolic` and the translated shards compile
+  with `-fno-stack-protector`: without them every one of the 29,000 translated
+  functions called its neighbours and the runtime through a PLT stub (26,973
+  of them, 702 after), which was 4% of the game thread on a Quest 3, and the
+  NDK's default canaries cost cycles in code whose state lives in guest memory.
+  With those and a larger indirect-dispatch memo (`kIndirectDispatchCacheEntries`),
+  a twelve-kart race start went from a 30 fps retrace lock to a steady 60.
+  The initial-exec TLS model is not an option: Bionic refuses it in a library
+  loaded with `dlopen`, which is how SDL loads the game
+  (`dlopen failed: TLS symbol ... using IE access model`).
 - `runtime/cmake/PublicProducts.cmake` gains `MKW_GENERATED_DIR` so a build
   configured from a checkout can name the translator output tree, and on
   Android rewrites the PE/COFF `.section .rdata,"dr"` of a Windows-generated

runtime/cmake/PublicProducts.cmake

@@ -58,8 +58,11 @@ function(mkw_apply_common_compile_options target)
 endfunction()
 
 function(mkw_apply_translated_compile_options target)
+    # The NDK toolchain adds -fstack-protector-strong to every object. Translated guest code keeps
+    # its state in guest memory and the CpuContext, so the canaries only cost cycles there.
     target_compile_options(${target} PRIVATE
-        -O2 ${MKW_TRANSLATED_PPC_FP_OPTIONS} -fno-slp-vectorize -w -pipe)
+        -O2 ${MKW_TRANSLATED_PPC_FP_OPTIONS} -fno-slp-vectorize -w -pipe
+        $<$<PLATFORM_ID:Android>:-fno-stack-protector>)
 endfunction()
 
 function(mkw_configure_object_target target)
@@ -319,6 +322,11 @@ function(mkw_configure_product target)
         target_link_libraries(${target} PRIVATE mkw::libco ${CMAKE_DL_LIBS})
     elseif(MKW_PLATFORM_ANDROID)
         target_link_libraries(${target} PRIVATE mkw::libco android log vulkan ${CMAKE_DL_LIBS})
+        # A shared library's own calls and data references go through the PLT and GOT unless the
+        # symbols are bound at link time; with 29,000 translated functions calling each other and
+        # the runtime, those stubs were 4% of the game thread on the Quest. Nothing interposes
+        # symbols of the game library, and the JNI and SDL_main exports stay exported.
+        target_link_options(${target} PRIVATE "-Wl,-Bsymbolic")
     endif()
     if(MKW_PLATFORM_WINDOWS)
         foreach(runtime_dll libc++.dll libunwind.dll)

runtime/include/abi_bridge.h

@@ -303,7 +303,10 @@ struct KnownTypedNativeCpuCall {
 // a miss just re-runs the sorted lookup. 512 entries (12 KiB) covers the per-frame indirect
 // working set while staying L1/L2 resident, unlike 4096 which would thrash L2. Must stay a
 // power of two, the index masks with (size - 1).
-inline constexpr size_t kIndirectDispatchCacheEntries = 512;
+// Direct-mapped by a hash of the target address. A race keeps a few thousand distinct indirect
+// targets live, so 512 entries missed about one call in three on the Quest and fell into the
+// table's binary search; 8192 entries are 128 KB per memo and fit the big cores' L2.
+inline constexpr size_t kIndirectDispatchCacheEntries = 8192;
 
 // Namespace-scope `inline thread_local`, not function-local `static thread_local`, to avoid a
 // thread-static init epoch check on every bctrl (same as g_currentCpuContext in ppc_runtime.h).