[lld] [lld-macho] Relax safe ICF's keepUnique for ld64-coalesced data sections (PR #193125)

[Zhaoxuan Jiang via llvm-commits]

================
@@ -583,19 +583,29 @@ void macho::foldIdenticalSections(bool onlyCfStrings) {
 
     bool isCodeSec = isCodeSection(isec);
 
-    // When keepUnique is true, the section is not foldable. Unless we are at
-    // icf level safe_thunks, in which case we still want to fold code sections.
-    // When using safe_thunks we'll apply the safe_thunks logic at merge time
-    // based on the 'keepUnique' flag.
-    bool noUniqueRequirement =
-        !isec->keepUnique ||
-        ((config->icfLevel == ICFLevel::safe_thunks) && isCodeSec);
+    // Determine whether keepUnique forbids folding this section.
+    //   - __cfstring / __objc_classrefs / __objc_selrefs always fold
+    //     regardless of keepUnique. Compilers currently emit over-broad
+    //     __llvm_addrsig entries that can cover non-address-significant data
+    //     symbols in these sections; ld64 coalesces them unconditionally, and
+    //     we match that behavior.
+    //   - Under safe_thunks, keepUnique code sections still fold; the
+    //     safe_thunks logic is applied later at merge time based on the
+    //     keepUnique flag.
+    //   - Otherwise, keepUnique sections are not foldable.
+    bool isUnconditionallyCoalescedData = isCfStringSection(isec) ||
+                                          isClassRefsSection(isec) ||
+                                          isSelRefsSection(isec);
----------------
nocchijiang wrote:

Sorry, I jumped to a conclusion too early. On a closer look `__objc_protolist` / `__objc_protorefs` aren't in ICF's scope, so the two lists don't diverge after all. Applied your suggestion with a short comment noting that the two may diverge in the future so readers know it's safe to change.

https://github.com/llvm/llvm-project/pull/193125