Remove address taken, add optnone

[dblaikie]

Summary

Address taken function may not always be cold.
optnone functions should not be optimized.

[sebpop]

lgtm

[dblaikie]

Test case?

[hiraditya]

Will add one soon!

[hiraditya]

Adding test

[dblaikie]

test/Transforms/HotColdSplit/split-cold-1.ll

29 ↗

(On Diff #170349)

I'm guessing if this is marked "noinline" then this test wouldn't fail with the patch not applied? (because it'd be caught by the noinline case a few lines later - producing the same behavior)

With the "noinline" removed, does this test fail without the patch applied?

[aeubanks]

test/Transforms/HotColdSplit/split-cold-1.ll

29 ↗

(On Diff #170349)

This patch shouldn't be necessary, optnone requires noinline:

$ cat /tmp/a.ll
define void @test1() optnone {
  ret void
}
$ bin/opt.exe -verify -disable-output /tmp/a.ll
Attribute 'optnone' requires 'noinline'!
void ()* @test1
C:\src\llvm-project\build\rel\bin\opt.exe: C:/src/tmp/a.ll: error: input module is broken!

[dblaikie]

test/Transforms/HotColdSplit/split-cold-1.ll

29 ↗

(On Diff #170349)

Great! Given how old the patch is, I'll just commandeer and abandon it.

lib/Transforms/IPO/HotColdSplitting.cpp

 bool HotColdSplitting::shouldOutlineFrom(const Function &F) const {
   // Do not try to outline again from an already outlined cold function.
   if (OutlinedFunctions.count(&F))
     return false;
 
   if (F.size() <= 2)
     return false;
 
-  if (F.hasAddressTaken())
+  if (F.hasFnAttribute(Attribute::OptimizeNone))
     return false;
 
   if (F.hasFnAttribute(Attribute::AlwaysInline))
     return false;
 
   if (F.hasFnAttribute(Attribute::NoInline))
     return false;