[PATCH] PrettyStackTrace guards around ASTReader

[Jordan Rose]

I originally considered putting it into the Deserializing RAII, but decided against it, mainly because Deserializing can stack but I only really wanted one message in the trace. It's more about "this problem might just be the module cache" than actually providing a trace. There are also a few points where I added a trace where we don't bother setting up a Deserializing object, like selectors.

================
Comment at: lib/Serialization/ASTReader.cpp:70
@@ +69,3 @@
+      "\n\n*** Crashed while deserializing from an AST file (e.g. PCH or PCM)."
+      "\n*** Consider clearing your module cache.\n";
+  public:
----------------
rsmith wrote:
> I'm not sure I like this as a user-facing message; the revision number is supposed to be part of the configuration, so the 'clear your module cache' advice really only applies to people actively hacking on Clang. But I'm fine with the change with or without this line.
The revision number thing hasn't worked for a long time, at least not for git. You're still right that it only applies to compiler devs, but this means there are a lot of people building tools on top of Clang (in our case the Swift compiler) that break when things are updated and don't know anything about modules.

I keep meaning to fix the revision thing but not getting around to it. The requirement last time I tried (years ago) was that it has to not do work on null builds.

http://reviews.llvm.org/D6138