Have callees set up their shadow frame.

Before this change we'd have caller's allocate shadow space. Besides
being an unconventional strategy, this also means that we allocate
shadow space for things which will never require shadow space. We even
had to conservatively allocate shadow space before external calls just
in case they call-back into functions that do need shadow space.

This change makes each callee requiring a shadow frame allocate it
itself.

Read the comment at the top of ShadowStack.cpp for implementation
details.

Measuring about a 22% speedup on bigloop.

llvm/include/llvm/InitializePasses.h

@@ -344,6 +344,7 @@ void initializeXRayInstrumentationPass(PassRegistry&);
 void initializeYkStackmapsPass(PassRegistry&);
 void initializeYkSplitBlocksAfterCallsPass(PassRegistry&);
 void initializeYkBasicBlockTracerPass(PassRegistry&);
+void initializeYkShadowStackPass(PassRegistry&);
 void initializeYkModuleClonePass(PassRegistry&);
 } // end namespace llvm
 

llvm/lib/CodeGen/CodeGen.cpp

@@ -153,5 +153,6 @@ void llvm::initializeCodeGen(PassRegistry &Registry) {
   initializeYkStackmapsPass(Registry);
   initializeYkSplitBlocksAfterCallsPass(Registry);
   initializeYkBasicBlockTracerPass(Registry);
+  initializeYkShadowStackPass(Registry);
   initializeYkModuleClonePass(Registry);
 }