Add a setting to opt out of rewriting message sends with visible implementations

The setting is on by default to preserve the current experience. It
should likely be revisited as related changes are merged.

I find this to be confusing as often as it is helpful since it picks the
first implementation of the selector it sees without consideration for
the type of the receiver.

This is particularly annoying if the binary being analyzed implements a
method with the same name as a commonly-used method on a system type
(`-description` or `-path`, for instance), or has methods with generic
names (`-initWithURL:`) on many types.

Explicit cross-references from selectors to method implementations make
it possible to see the potential implementations without rewriting the
call, and an Objective-C pseudo-language
(Vector35/binaryninja-api#6807) provides an even
more natural representation of Objective-C message sends without these
downsides.

Author: bdash

Plugin.cpp

@@ -41,6 +41,15 @@ BINARYNINJAPLUGIN bool CorePluginInit()
 
     BinaryNinja::LogRegistry::CreateLogger(PluginLoggerName);
 
+    auto settings = BinaryNinja::Settings::Instance();
+    settings->RegisterSetting("core.function.objectiveC.assumeMessageSendTarget",
+        R"({
+		"title" : "Rewrite objc_msgSend calls to first visible implementation",
+		"type" : "boolean",
+		"default" : true,
+		"description" : "Message sends of selectors with any visible implementation are replaced with a direct call to the first visible implementation. Note that this can produce false positives if the selector is implemented by more than one class, or shares a name with a method from a system framework."
+		})");
+
     return true;
 }
 }

Workflow.cpp

@@ -115,6 +115,8 @@ bool Workflow::rewriteMethodCall(LLILFunctionRef ssa, size_t insnIndex)
     ssa->GetFunction()->SetAutoCallTypeAdjustment(ssa->GetFunction()->GetArchitecture(), insn.address, {funcType, BN_DEFAULT_CONFIDENCE});
     // --
 
+    if (!BinaryNinja::Settings::Instance()->Get<bool>("core.function.objectiveC.assumeMessageSendTarget"))
+        return false;
 
     // Check the analysis info for a selector reference corresponding to the
     // current selector. It is possible no such selector reference exists, for
@@ -126,6 +128,10 @@ bool Workflow::rewriteMethodCall(LLILFunctionRef ssa, size_t insnIndex)
     const auto info = GlobalState::analysisInfo(bv);
     if (!info)
         return false;
+
+    // Attempt to look up the implementation for the given selector, first by
+    // using the raw selector, then by the address of the selector reference. If
+    // the lookup fails in both cases, abort.
     std::vector<uint64_t> imps;
     if (const auto& it = info->selRefToImp.find(rawSelector); it != info->selRefToImp.end())
         imps = it->second;
@@ -135,10 +141,6 @@ bool Workflow::rewriteMethodCall(LLILFunctionRef ssa, size_t insnIndex)
     if (imps.empty())
         return false;
 
-    // Attempt to look up the implementation for the given selector, first by
-    // using the raw selector, then by the address of the selector reference. If
-    // the lookup fails in both cases, abort.
-
     // k: This is the same behavior as before, however it is more apparent now by implementation
     //      that we are effectively just guessing which method this hits. This has _obvious_ drawbacks,
     //      but until we have more robust typing and objective-c type libraries, fixing this would