Subzero: Run sandboxed cross tests, and do some cleanup.

src/IceELFObjectWriter.cpp

Index: src/IceELFObjectWriter.cpp
diff --git a/src/IceELFObjectWriter.cpp b/src/IceELFObjectWriter.cpp
index ff68ba2582621fe7c6eda64d2a12950063d35120..230848c4f2c0a0c3de441e33a102c886f5af78fa 100644
--- a/src/IceELFObjectWriter.cpp
+++ b/src/IceELFObjectWriter.cpp
@@ -36,7 +36,7 @@ struct {
 #define X(tag, str, is_elf64, e_machine, e_flags)                              \
   { is_elf64, e_machine, e_flags }                                             \
   ,
-      TARGETARCH_TABLE
+    TARGETARCH_TABLE
 #undef X
 };
 
@@ -244,7 +244,7 @@ void ELFObjectWriter::writeFunctionCode(const IceString &FuncName,
   Section->appendData(Str, Asm->getBufferView());
   uint8_t SymbolType;
   uint8_t SymbolBinding;
-  if (IsInternal) {
+  if (IsInternal && !Ctx.getFlags().getDisableInternal()) {
     SymbolType = STT_NOTYPE;
     SymbolBinding = STB_LOCAL;
   } else {
@@ -545,6 +545,12 @@ void ELFObjectWriter::setUndefinedSyms(const ConstantList &UndefSyms) {
   for (const Constant *S : UndefSyms) {
     const auto Sym = llvm::cast<ConstantRelocatable>(S);
     const IceString &Name = Sym->getName();
+    bool BadIntrinsic;
+    const Intrinsics::FullIntrinsicInfo *Info =
+        Ctx.getIntrinsicsInfo().find(Name, BadIntrinsic);
+    if (Info)

[jvoung (off chromium), 2015/03/03 22:58:01]

oops didn't think an intrinsic name would make it this far down, but I guess
it's added to the extern syms list early?

[Jim Stichnoth, 2015/03/03 23:18:51]

Yeah, these llvm.* symbols were making it into the ELF file and causing
undefined symbol errors during the link.  I'm not sure why it wasn't happening
for the spec2k builds - maybe it was something about the gold arguments?  In any
case, it seemed like keeping them out of the ELF file was the right thing to do.

This definitely happens for the IceConverter, but I'm not sure whether the
PNaClTranslator does it as well.  Regardless, it seemed more appropriate to
continue dealing with intrinsic issues during lowering and afterwards.

[jvoung (off chromium), 2015/03/03 23:52:52]

Okay, I agree it shouldn't be in the undefined symbols list, so it's not really
gold's problem.
I thought it would be better that earlier layers like
IceConverter/PNaClTranslator or the target lowering to "fully" deal with
intrinsics and not expose intrinsics to this layer. For example, at this layer
it should have "setjmp" but not "llvm.setjmp". Otherwise, this is fine for now
to get testing going.

+      continue;
+    assert(!BadIntrinsic);
     assert(Sym->getOffset() == 0);
     assert(Sym->getSuppressMangling());
     SymTab->noteUndefinedSym(Name, NullSection);