fix linker error due to personality localization

Author: ruccho

.github/scripts/localize-apple-staticlib.sh

@@ -1,7 +1,12 @@
 #!/usr/bin/env bash
-# Localize all non-exported symbols in an Apple static library so that its
-# bundled mimalloc does not coalesce with Unity 6.5's built-in mimalloc when the
-# library is statically linked into UnityFramework (iOS `__Internal`).
+# Localize the bundled mimalloc symbols in an Apple static library so they do
+# not coalesce with Unity 6.5's built-in mimalloc when the library is statically
+# linked into UnityFramework (iOS `__Internal`).
+#
+# Only mimalloc's own symbols are localized. Rust runtime symbols (including
+# personality routines) are intentionally left global: localizing them would
+# give every linked static lib its own personality routine, and compact-unwind
+# can encode at most 3 per image ("ld: Too many personality routines ...").
 #
 # Requires the library to have been compiled with `-fno-common` (see the CI
 # workflow) so that mimalloc's tentative definitions (e.g. `_mi_page_map`) are
@@ -15,9 +20,9 @@ LIB="$1"                 # path to the .a (e.g. libunienc_c.a)
 ARCH="$2"                # ld arch name (e.g. arm64)
 SDK="${3:-iphoneos}"     # xcrun SDK (e.g. iphoneos)
 PLATFORM="${4:-ios}"     # -platform_version platform name (e.g. ios)
-MIN_OS="${5:-13.0}"      # -platform_version minimum OS version
+MIN_OS="${5:-10.0}"      # -platform_version minimum OS version
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-EXPORTS="$SCRIPT_DIR/../../InstantReplay.Externals/unienc/crates/unienc_c/apple-exports.txt"
+UNEXPORTS="$SCRIPT_DIR/../../InstantReplay.Externals/unienc/crates/unienc_c/apple-mimalloc-unexports.txt"
 
 if [ ! -f "$LIB" ]; then
   echo "ERROR: static library not found: $LIB" >&2
@@ -32,7 +37,7 @@ trap 'rm -rf "$WORK"' EXIT
 # Extract every archive member as a loose object. We deliberately do NOT use
 # `ld -r -force_load <archive>`: the modern linker (ld-prime) localizes *all*
 # symbols of a force-loaded archive, which would strip the `_unienc*` exports
-# too. Passing loose objects makes `-exported_symbols_list` behave as intended.
+# and the Rust runtime symbols too. Loose objects make the symbol lists behave.
 ( cd "$WORK" && ar x "$LIB_ABS" )
 member_count="$(ar t "$LIB_ABS" | grep -Ec '\.o$' || true)"
 extracted_count="$(find "$WORK" -maxdepth 1 -name '*.o' | wc -l | tr -d ' ')"
@@ -42,35 +47,36 @@ if [ "$member_count" != "$extracted_count" ]; then
 fi
 
 # Partial link into a single object.
-#   -exported_symbols_list : demote every global NOT matching `_unienc*` to a
-#                            private-extern symbol.
+#   -unexported_symbols_list : demote ONLY the listed mimalloc symbols to
+#                              private-extern (leaves everything else global).
 #   -r (without -keep_private_externs) : convert those private-externs (and the
-#                            ones already produced by `-fvisibility=hidden`) into
-#                            true local symbols.
+#                              ones already produced by `-fvisibility=hidden`)
+#                              into true local symbols.
 #   -platform_version : required by ld-prime even for a `-r` link.
 ( cd "$WORK" && xcrun --sdk "$SDK" ld -r -arch "$ARCH" \
     -platform_version "$PLATFORM" "$MIN_OS" "$SDK_VER" \
-    -exported_symbols_list "$EXPORTS" \
+    -unexported_symbols_list "$UNEXPORTS" \
     *.o -o merged.o )
 
 rm -f "$LIB_ABS"
 xcrun libtool -static -o "$LIB_ABS" "$WORK/merged.o"
 
 # --- Verify (CI gate; also guards against future linker behaviour changes) ---
-echo "=== exported (global) symbols after localization ==="
-nm -gU "$LIB_ABS" | sort -u
-
 # (a) No mimalloc symbol may remain external or private-external. `nm -gU` lists
 #     both (private-externs still coalesce), so it must contain zero of them.
-if nm -gU "$LIB_ABS" | grep -iE '_mi_|_heap_main|_theap_main|mi_unity_'; then
+#     The `_mi_` substring also matches `__mi_*` / `___mi_*`.
+if xcrun nm -gU "$LIB_ABS" | grep -iE '_mi_|_heap_main|_theap'; then
   echo "FAIL: mimalloc symbols are still (private-)external -> would coalesce with Unity" >&2
+  echo "      (if _mi_page_map is listed, the build is missing -fno-common)" >&2
   exit 1
 fi
 
-# (b) FFI entry points must remain global.
-if ! nm -gU "$LIB_ABS" | grep -q '_unienc_UnityPluginLoad'; then
+# (b) FFI entry points must remain global. Use `grep -c` (not `grep -q`): under
+# `set -o pipefail`, `grep -q` exits on first match and SIGPIPEs `nm`, whose
+# non-zero status would then be reported as a (spurious) failure.
+if [ "$(xcrun nm -gU "$LIB_ABS" | grep -c '_unienc_UnityPluginLoad')" -eq 0 ]; then
   echo "FAIL: _unienc_UnityPluginLoad is no longer exported" >&2
   exit 1
 fi
 
-echo "Localization OK: $(nm -gU "$LIB_ABS" | grep -c '_unienc') unienc symbols exported, 0 mimalloc symbols external"
+echo "Localization OK: $(xcrun nm -gU "$LIB_ABS" | grep -c '_unienc') unienc symbols exported, 0 mimalloc symbols external"

.github/workflows/build-unienc.yml

@@ -135,6 +135,7 @@ jobs:
           # mimalloc. Injected via cc-rs's target CFLAGS so no submodule change
           # is needed. Only aarch64 is built for iOS (see matrix).
           CFLAGS_aarch64-apple-ios: -fno-common
+          IPHONEOS_DEPLOYMENT_TARGET: 10.0
       # Localize the bundled mimalloc so it does not coalesce with Unity 6.5's
       # built-in mimalloc when statically linked into UnityFramework.
       - name: Localize non-exported symbols (Unity variant only)

InstantReplay.Externals/unienc/crates/unienc_c/apple-exports.txt

@@ -0,0 +1,18 @@
+# mimalloc symbols to localize when libunienc_c.a is statically linked into
+# UnityFramework (iOS). Consumed via `ld -r -unexported_symbols_list`.
+#
+# We localize ONLY mimalloc's own symbols so they no longer coalesce with Unity
+# 6.5's built-in mimalloc. Everything else (the `_unienc*` FFI entry points AND
+# the Rust runtime symbols such as `rust_eh_personality`) is left global on
+# purpose: localizing personality routines would give each linked static lib its
+# own copy, and the compact-unwind format can encode at most 3 personality
+# routines per image -> "ld: Too many personality routines for compact unwind".
+#
+# The three underscore depths cover the C sources `mi_*`, `_mi_*` and `__mi_*`
+# (Mach-O prepends one leading underscore). Verified to match all 432 mimalloc
+# symbols in the current unity-fork build with zero false positives.
+_mi_*
+__mi_*
+___mi_*
+_heap_main
+_theap_main