Align function starts to target-specific bundle alignment.

src/IceCfg.cpp

Index: src/IceCfg.cpp
diff --git a/src/IceCfg.cpp b/src/IceCfg.cpp
index 6fe7eb05830b435bedb9cc7fd98caf2fc5b80eb3..7d31845a66b5c1975c223179eee12019d64dcbd0 100644
--- a/src/IceCfg.cpp
+++ b/src/IceCfg.cpp
@@ -314,6 +314,7 @@ void Cfg::emit() {
     Str << "\t.globl\t" << MangledName << "\n";
     Str << "\t.type\t" << MangledName << ",@function\n";
   }
+  Str << "\t.p2align " << getTarget()->getBundleP2Align() << "\n";

[Jim Stichnoth, 2014/08/28 20:26:12]

Should we consider padding with a halt sled, just for defense-in-depth?

[jvoung (off chromium), 2014/08/28 21:02:22]

We could do that. Added a hook for now.

I'm not sure if there's a performance difference, if the processor does
speculative execution, or what not.

E.g., one study showed that different choices of non-executed nops could make a
difference. The guess was if it touched different processor units. Section 5.4
of
http://static.googleusercontent.com/media/research.google.com/en/us/pubs/arch...

[Jim Stichnoth, 2014/08/28 21:22:21]

I'd kind of hope that if this performance difference really is real, it's due to
speculative execution of the never-actually-executed instructions that's
clogging up the functional units, and that speculatively executing a halt
instruction would hopefully not have that effect.

   for (NodeList::const_iterator I = Nodes.begin(), E = Nodes.end(); I != E;
        ++I) {
     (*I)->emit(this);