[PATCH] clang-cl: implement /fallback mode

[Reid Kleckner]

================
Comment at: include/clang/Driver/CLCompatOptions.td:114
@@ -113,1 +113,3 @@
+def _SLASH_fallback : CLCompileFlag<"fallback">,
+  HelpText<"Fall back to cl.exe if clang-cl fails due to internal error">;
 def _SLASH_Fe : CLJoined<"Fe">,
----------------
Do we want to support multiple fallback modes like fallback for internal errors or fallback for any error?  I guess we can just use this and see what comes up in practice, and eventually get to a place where we only fallback on "could not compile 'foo' yet" diagnostics.

================
Comment at: lib/Driver/Tools.cpp:6721
@@ +6720,3 @@
+
+  // FIXME: If we've put clang-cl as cl.exe on the path, then we might have a problem now :/
+  const char *Exec =
----------------
That seems problematic.  You could use sys::fs::equivalent() to skip a path if it's the current executable, but we'd have to sink that into GetProgramPath().

================
Comment at: test/Driver/cl-fallback.c:8
@@ +7,3 @@
+// RUN: %clang_cl /fallback /Dfoo=bar /Ubaz /Ifoo /Ox /O0 /GR /GR- /LD /LDd \
+// RUN:     /MD /MDd /MTd /MT -### -- %s 2>&1 | FileCheck %s
+// CHECK: ||
----------------
I guess that's the most testing we can do here.  Oh well.

================
Comment at: lib/Driver/Tools.cpp:6687
@@ +6686,3 @@
+  // Flags for which clang-cl have an alias.
+  // FIXME: Could we avoid this manual reverse translation?
+
----------------
Can you iterate over Args.getArgString(I) up to getNumInputArgStrings(), or does that fail?

================
Comment at: lib/Driver/Tools.cpp:3566
@@ +3565,3 @@
+  if (Args.hasArg(options::OPT__SLASH_fallback)) {
+    // FIXME: This is where it gets crazy :)
+    tools::visualstudio::Compile CL(getToolChain());
----------------
I don't see any better way to do it, so this is probably OK.  No need for FIXME.  :)


http://llvm-reviews.chandlerc.com/D1711