Add the new @llvm.nacl.atomic.fence.all intrinsic

lib/IR/InlineAsm.cpp

 //===-- InlineAsm.cpp - Implement the InlineAsm class ---------------------===//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 // This file implements the InlineAsm class.
@@ skipping 275 matching lines @@
     StructType *STy = dyn_cast<StructType>(Ty->getReturnType());
     if (STy == 0 || STy->getNumElements() != NumOutputs)
       return false;
     break;
   }      
   
   if (Ty->getNumParams() != NumInputs) return false;
   return true;
 }
 
+// @LOCALMOD-START
+bool InlineAsm::isAsmMemory() const {
+  bool retVoid = getFunctionType()->getReturnType()->isVoidTy();
+  bool noArgs = getFunctionType()->getNumParams() == 0 &&
+      !getFunctionType()->isVarArg();
+  bool isEmptyAsm = AsmString.empty();
+  // Different triples will encode "touch everything" differently, e.g.:
+  //  - le32-unknown-nacl has "~{memory}".
+  //  - x86 "~{memory},~{dirflag},~{fpsr},~{flags}".

[jvoung (off chromium), 2013/08/07 22:32:11]

Does that mean that in the backend, we need to convert the fence_all back into
the target-specific "touch everything" asm ?

[JF, 2013/08/07 22:47:41]

Kind of. It comes from Clang's TargetInfo getClobbers, which each target
specifies in Targets.cpp. AFAIU LLVM can't use that code, and
implementation-wise only the x86 target sets any clobbers, the others all have
FIXMEs which sound utterly confused.

My conclusion when I saw this was that what we really care about anyways is the
memory side-effects, and things like flags can be assumed to stay live.

+  // The following code therefore only searches for memory.
+  bool touchesMemory = Constraints.find("~{memory}") != std::string::npos;
+
+  return retVoid && noArgs && hasSideEffects() && isEmptyAsm && touchesMemory;
+}
+// @LOCALMOD-END