In llvm.asan.globals, allow entries to be non-GlobalVariable and skip over them

[kubamracek]

Swift from master currently crashes when building even very simple programs with -O -whole-module-optimization -sanitize=address. The actual crash happens in the AddressSanitizer instrumentation pass because it finds entries inside the llvm.asan.globals list that are not GlobalVariable objects. In short, what happens is that when IRGen emits type descriptors, we add them to the llvm.asan.globals list to avoid instrumenting them, but in WMO, we emit lazy type descriptors and use forward declarations and when those are later replaced with an actual definition, we use replaceAllUsesWith to replace them with a BitCast object. There are valid reasons why we need the BitCast when replacing the type descriptor, so we should instead allow them in the llvm.asan.globals list.

[kubamracek]

@swift-ci please test

[adrian-prantl]

Could you add a .ll test for this that highlights the kind of IR this is supposed to accept?

[rjmccall]

I'll take your word for it that ignoring them is the right thing to do in this case.

This should be upstreamed, though, right?

[kubamracek]

So far, this looks a Swift-only problem, so I want to start here to get a fix out soon.

[rjmccall]

I'm in no position to lecture anyone about preferring to commit things upstream, but this is easily reproducible from valid inputs:

harza:/tmp$ cat a.ll 
@a = common global i64 0
!llvm.asan.globals = !{!0}
!0 = !{i64* @a, !{!"a.ll", i32 0, i32 0}, !"a", i1 false, i1 false}

harza:/tmp$ cat b.ll 
@a = global [10 x i8] zeroinitializer
!llvm.asan.globals = !{!0}
!0 = !{[10 x i8]* @a, !{!"a.ll", i32 0, i32 0}, !"a", i1 false, i1 false}

harza:/tmp$ llvm-link a.ll b.ll -S > c.ll

harza:/tmp$ opt -S -asan c.ll
Assertion failed: (isa<X>(Val) && "cast<Ty>() argument of incompatible type!"), function cast, file /Volumes/Data/swift/llvm/include/llvm/Support/Casting.h, line 255.

[rjmccall]

(Exercise for the reader on how to produce those inputs from valid C.)

[kubamracek]

I see! Moved to https://reviews.llvm.org/D55794 and added a test.

[kubamracek]

Changes this PR to just be a cherry-pick of r349544. @rjmccall looks good for 5.0?

[rjmccall]

Yep!

[kubamracek]

@swift-ci please test

[kubamracek]

@swift-ci please test