diff --git a/ddmd/func.d b/ddmd/func.d
index 1b88603950..ba5e7b1909 100644
--- a/ddmd/func.d
+++ b/ddmd/func.d
@@ -1583,7 +1583,7 @@ public:
                 if (f.linkage == LINKd || (f.parameters && Parameter.dim(f.parameters)))
                 {
                     // Declare _argptr
-                    Type t = Type.tvalist;
+                    Type t = Type.tvalist.semantic(loc, sc);
                     argptr = new VarDeclaration(Loc(), t, Id._argptr, null);
                     argptr.storage_class |= STCtemp;
                     argptr.semantic(sc2);
diff --git a/driver/linker.cpp b/driver/linker.cpp
index ca5e5fa335..ef9934406c 100644
--- a/driver/linker.cpp
+++ b/driver/linker.cpp
@@ -230,10 +230,23 @@ static int linkObjToBinaryGcc(bool sharedLib, bool fullyStatic) {
         break;
       }
     } else {
-      if (global.params.is64bit) {
-        args.push_back("-m64");
-      } else {
-        args.push_back("-m32");
+      switch (global.params.targetTriple.getArch()) {
+      case llvm::Triple::arm:
+      case llvm::Triple::armeb:
+      case llvm::Triple::aarch64:
+      case llvm::Triple::aarch64_be:
+#if LDC_LLVM_VER == 305
+      case llvm::Triple::arm64:
+      case llvm::Triple::arm64_be:
+#endif
+        break;
+      default:
+        if (global.params.is64bit) {
+          args.push_back("-m64");
+        }
+        else {
+          args.push_back("-m32");
+        }
       }
     }
   }
diff --git a/gen/abi-aarch64.cpp b/gen/abi-aarch64.cpp
index de6b277a53..86feb4dcab 100644
--- a/gen/abi-aarch64.cpp
+++ b/gen/abi-aarch64.cpp
@@ -112,8 +112,7 @@ struct AArch64TargetABI : TargetABI {
     // using TypeIdentifier here is a bit wonky but works, as long as the name
     // is actually available in the scope (this is what DMD does, so if a better
     // solution is found there, this should be adapted).
-    return (createTypeIdentifier(Loc(), Identifier::idPool("__va_list_tag")))
-        ->pointerTo();
+    return (new TypeIdentifier(Loc(), Identifier::idPool("__va_list")));
   }
 };
 
diff --git a/gen/abi-arm.cpp b/gen/abi-arm.cpp
index 3f4626cb22..81ae57651e 100644
--- a/gen/abi-arm.cpp
+++ b/gen/abi-arm.cpp
@@ -67,14 +67,29 @@ struct ArmTargetABI : TargetABI {
   }
 
   bool passByVal(Type *t) override {
-    // AAPCS does not pass byval
-    return false;
+    // AAPCS does not use an indirect arg to pass aggregates, however
+    // clang uses byval for types > 64-bytes, then llvm backend
+    // converts back to non-byval.  Without this special handling the
+    // optimzer generates bad code (e.g. std.random unittest crash).
+    t = t->toBasetype();
+    return ((t->ty == Tsarray || t->ty == Tstruct) && t->size() > 64);
 
-    // Note: the codegen is horrible for Tsarrays passed this way - tries to do
-    // copy without a loop for huge arrays.  Would be better if byval was used
-    // for arrays, but then there is an optimizer problem in the "top-down list
-    // latency scheduler" pass that reorders instructions incorrectly if byval
-    // used.
+    // Note: byval can have a codegen problem with -O1 and higher.
+    // What happens is that load instructions are being incorrectly
+    // reordered before stores.  It is a problem in the LLVM backend.
+    // The outcome is a program with incorrect results or crashes.
+    // It happens in the "top-down list latency scheduler" pass
+    //
+    //   https://forum.dlang.org/post/m2r3u5ac0c.fsf@comcast.net
+    //
+    // Revist and determine if the byval problem is only for small
+    // structs, say 16-bytes or less, that can entirely fit in
+    // registers.
+    
+    // Note: the codegen is horrible for Tsarrays passed this way -
+    // does a copy without a loop for huge arrays.  Could be better if
+    // byval was always used for sarrays, and maybe can if above
+    // problem is better understood.
   }
 
   void rewriteFunctionType(TypeFunction *tf, IrFuncTy &fty) override {
@@ -94,6 +109,11 @@ struct ArmTargetABI : TargetABI {
       if (!arg->byref)
         rewriteArgument(fty, *arg);
     }
+
+    // extern(D): reverse parameter order for non variadics, for DMD-compliance
+    if (tf->linkage == LINKd && tf->varargs != 1 && fty.args.size() > 1) {
+      fty.reverseParams = true;
+    }
   }
 
   void rewriteArgument(IrFuncTy &fty, IrFuncTyArg &arg) override {
diff --git a/gen/asmstmt.cpp b/gen/asmstmt.cpp
index 9fca1a77ca..85071dc68b 100644
--- a/gen/asmstmt.cpp
+++ b/gen/asmstmt.cpp
@@ -114,7 +114,10 @@ Statement *asmSemantic(AsmStatement *s, Scope *sc) {
     err = true;
   }
   if (err) {
-    fatal();
+    if (!global.gag) {
+      fatal();
+    }
+    return s;
   }
 
   // puts(toChars());
diff --git a/gen/functions.cpp b/gen/functions.cpp
index 9af72f71b9..630d739b56 100644
--- a/gen/functions.cpp
+++ b/gen/functions.cpp
@@ -883,7 +883,7 @@ void DtoDefineFunction(FuncDeclaration *fd) {
     // D varargs: prepare _argptr and _arguments
     if (f->linkage == LINKd && f->varargs == 1) {
       // allocate _argptr (of type core.stdc.stdarg.va_list)
-      LLValue *argptrmem = DtoAlloca(Type::tvalist, "_argptr_mem");
+      LLValue *argptrmem = DtoAlloca(Type::tvalist->semantic(fd->loc, fd->scope), "_argptr_mem");
       irFunc->_argptr = argptrmem;
 
       // initialize _argptr with a call to the va_start intrinsic
diff --git a/gen/tollvm.cpp b/gen/tollvm.cpp
index 37454dfcdb..b599b43401 100644
--- a/gen/tollvm.cpp
+++ b/gen/tollvm.cpp
@@ -63,7 +63,9 @@ LLAttribute DtoShouldExtend(Type *type) {
 
     case Tuns8:
     case Tuns16:
-      return LLAttribute::ZExt;
+    case Tchar:
+    case Twchar:
+	return LLAttribute::ZExt;
 
     default:
       // Do not extend.
@@ -410,6 +412,15 @@ LLConstant *DtoConstFP(Type *t, longdouble value) {
     return LLConstantFP::get(gIR->context(), APFloat(APFloat::x87DoubleExtended,
                                                      APInt(80, 2, bits)));
   }
+  if (llty == LLType::getFP128Ty(gIR->context())) {
+    union {
+      longdouble ld;
+      uint64_t bits[2];
+    } t;
+    t.ld = value;
+    return LLConstantFP::get(
+      gIR->context(), APFloat(APFloat::IEEEquad, APInt(128, 2, t.bits)));
+  }
   if (llty == LLType::getPPC_FP128Ty(gIR->context())) {
     uint64_t bits[] = {0, 0};
     bits[0] = *reinterpret_cast<uint64_t *>(&value);
diff --git a/ir/irdsymbol.cpp b/ir/irdsymbol.cpp
index efa772ec16..c821da7d83 100644
--- a/ir/irdsymbol.cpp
+++ b/ir/irdsymbol.cpp
@@ -27,7 +27,9 @@ void IrDsymbol::resetAll() {
   }
 }
 
-IrDsymbol::IrDsymbol() { list.push_back(this); }
+IrDsymbol::IrDsymbol() : irData(nullptr) {
+  list.push_back(this);
+}
 
 IrDsymbol::IrDsymbol(const IrDsymbol &s) {
   list.push_back(this);
diff --git a/ir/irdsymbol.h b/ir/irdsymbol.h
index b277210f2b..82f38f02ab 100644
--- a/ir/irdsymbol.h
+++ b/ir/irdsymbol.h
@@ -83,7 +83,7 @@ private:
   friend IrField *getIrField(VarDeclaration *decl, bool create);
 
   union {
-    void *irData = nullptr;
+    void *irData;
     IrModule *irModule;
     IrAggr *irAggr;
     IrFunction *irFunc;
diff --git a/ir/irtype.cpp b/ir/irtype.cpp
index 211fd65183..51697be847 100644
--- a/ir/irtype.cpp
+++ b/ir/irtype.cpp
@@ -48,12 +48,21 @@ namespace {
 llvm::Type *getReal80Type(llvm::LLVMContext &ctx) {
   llvm::Triple::ArchType const a = global.params.targetTriple->getArch();
   bool const anyX86 = (a == llvm::Triple::x86) || (a == llvm::Triple::x86_64);
+  bool const anyAarch64 = (a == llvm::Triple::aarch64) || (a == llvm::Triple::aarch64_be)
+#if LDC_LLVM_VER == 305
+                       || (a == llvm::Triple::arm64) || (a == llvm::Triple::arm64_be)
+#endif
+    ;
 
   // only x86 has 80bit float - but no support with MS C Runtime!
   if (anyX86 && !global.params.targetTriple->isWindowsMSVCEnvironment()) {
     return llvm::Type::getX86_FP80Ty(ctx);
   }
 
+  if (anyAarch64) {
+    return llvm::Type::getFP128Ty(ctx);
+  }
+
   return llvm::Type::getDoubleTy(ctx);
 }
 }
diff --git a/runtime/CMakeLists.txt b/runtime/CMakeLists.txt
index 91ff134ce0..e557d895a5 100644
--- a/runtime/CMakeLists.txt
+++ b/runtime/CMakeLists.txt
@@ -116,6 +116,7 @@ if(UNIX)
         message(WARNING "Please consider updating CMake to at least 2.8.5.")
     else()
         list(APPEND DCRT_ASM ${RUNTIME_DIR}/src/core/threadasm.S)
+        list(APPEND DCRT_ASM ${RUNTIME_DIR}/src/ldc/eh_asm.S)
     endif()
     if(APPLE)
         list(APPEND CORE_D_SYS ${CORE_D_OSX})
diff --git a/runtime/druntime b/runtime/druntime
index ced8af41ff..fa0c558126 160000
--- a/runtime/druntime
+++ b/runtime/druntime
@@ -1 +1 @@
-Subproject commit ced8af41ffaf887ef534279289fd857eaa45648a
+Subproject commit fa0c558126f108a4a991d1d206925427dd4bfaae
diff --git a/runtime/phobos b/runtime/phobos
index 32c89fac4e..1ada8d5571 160000
--- a/runtime/phobos
+++ b/runtime/phobos
@@ -1 +1 @@
-Subproject commit 32c89fac4ec8cf58b7489ff4437a2577a9a7dba9
+Subproject commit 1ada8d5571bbfefc71ab45bb102a5455ff54dcb5
diff --git a/tests/d2/dmd-testsuite b/tests/d2/dmd-testsuite
index 331edbc588..729534b5cc 160000
--- a/tests/d2/dmd-testsuite
+++ b/tests/d2/dmd-testsuite
@@ -1 +1 @@
-Subproject commit 331edbc588b60f6adee8eddb78bed97415bc95da
+Subproject commit 729534b5cc3d2f4e058c54796988efb6c3ed3151