GVN hoists loads over intrinsics

[jdoerfert]

Originally reported here: https://discourse.llvm.org/t/bug-gvn-memdep-bug-in-the-presence-of-intrinsics/59402

GVN, or probably memdep analysis, seem to think intrinsics are somewhat special
and consequently assume they cannot "effectively clobber" a memory location. It
might also be globals-mod-ref or something else that provides AA information.

The following example (https://godbolt.org/z/TG6dYvP6M 1) shows how an unknown function
prevents an invalid load hoist while an unknown intrinsic does not. Again, the preload
is not sound in a multi-thread environment.

The simple reproducer can be found here: https://godbolt.org/z/cKMEezbjY
Compiled with -globals-aa -gvn the below code will hoist the load above the barrier but not the unknown function.

@G = internal global i32 undef

define void @test_barrier(i1 %c) {
  br i1 %c, label %init, label %check
init:
  store i32 0, ptr @G
  br label %check
check:
  call void @llvm.amdgcn.s.barrier()
  %v = load i32, ptr @G
  %cmp = icmp eq i32 %v, 0
  call void @llvm.assume(i1 %cmp)
  ret void
}

define void @test_unknown(i1 %c) {
  br i1 %c, label %init, label %check
init:
  store i32 0, ptr @G
  br label %check
check:
  call void @unknown()
  %v = load i32, ptr @G
  %cmp = icmp eq i32 %v, 0
  call void @llvm.assume(i1 %cmp)
  ret void
}

declare void @unknown()
declare void @llvm.amdgcn.s.barrier()
declare void @llvm.assume(i1 noundef)

[jdoerfert]

@efriedma-quic @arsenm Any thoughts?

[efriedma-quic]

Is this distinct from https://reviews.llvm.org/D115302 ?

[jdoerfert]

Is this distinct from https://reviews.llvm.org/D115302 ?

I think it is. Or at least the fix I'm preparing is. I think the reason for this problem is d4133ac which survived one way or another until today.

[jdoerfert]

Proposed fix: https://reviews.llvm.org/D123531