Trying to repro locally.

I'm not sure if this is a compiler/linker bug, or a code bug. The problem is that there are two extern string constants with the same (empty) value in Chromium builds. clang/lld is collapsing them together so that they have the same pointer value.

// App GUIDs for Google Chrome and the Google Chrome binaries.
#if defined(GOOGLE_CHROME_BUILD)
const wchar_t kChromeAppGuid[] = L"{8A69D345-D564-463c-AFF1-A69D9E530F96}";
const wchar_t kBinariesAppGuid[] = L"{4DC8B4CA-1BDA-483e-B5FA-D3C12E15B62D}";
#else
const wchar_t kChromeAppGuid[] = L"";
const wchar_t kBinariesAppGuid[] = L"";
#endif

I can easily put gibberish in there to prevent this, but I would not expect these to be folded together like this. Am I relying on undefined behavior (two constants having distinct pointer values), or is clang/lld doing something wrong?

I think your code is correct from a C++ language perspective, objects are supposed to have identity and this is supposed to work:
static char a, b;
void f() { assert(&a != &b); }

I don't think there's anything special about wide character arrays that means their addresses are not important. String literals can definitely get folded, so the result of '"a" == "a"' is unpredictable, but that's not what's going on here.

So this is an lld icf bug then?

ICF of only code is also non-conforming, so it's hard to say whether this is a bug or a feature. You can call it a bug on the basis that it introduces compatibility issues with MSVC, but some users might appreciate the more aggressive deduplication of constants.

Is this something you want to fix in lld?

I inclined to say no, because the unsafe ICF optimization is on by default on Windows anyways. We are a little bit more aggressive than what MSVC does, but in spirit we are doing what they intend to do.

I have a fix in https://codereview.chromium.org/2236843002/. I'd appreciate it if one of y'all could take a look and LGTM+CQ if you like it. It avoids the pointer comparison altogether rather than hacking around the linker's behavior. Thanks.

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/e5c331861fe09a14c01bd70f70087c7b99c1df24

commit e5c331861fe09a14c01bd70f70087c7b99c1df24
Author: grt <grt@chromium.org>
Date: Fri Aug 12 21:26:02 2016

Do not rely on identical string constants having distinct pointers.

BUG= 635943 
R=rnk@chromium.org

Review-Url: https://codereview.chromium.org/2236843002
Cr-Commit-Position: refs/heads/master@{#411775}

[modify] https://crrev.com/e5c331861fe09a14c01bd70f70087c7b99c1df24/chrome/tools/disable_outdated_build_detector/constants.cc
[modify] https://crrev.com/e5c331861fe09a14c01bd70f70087c7b99c1df24/chrome/tools/disable_outdated_build_detector/constants.h
[modify] https://crrev.com/e5c331861fe09a14c01bd70f70087c7b99c1df24/chrome/tools/disable_outdated_build_detector/disable_outdated_build_detector.cc
[modify] https://crrev.com/e5c331861fe09a14c01bd70f70087c7b99c1df24/chrome/tools/disable_outdated_build_detector/google_update_integration.cc
[modify] https://crrev.com/e5c331861fe09a14c01bd70f70087c7b99c1df24/chrome/tools/disable_outdated_build_detector/google_update_integration.h

(This kind of feels like an lld bug to me tbh. I don't remember seeing this type of problem with gold's heuristic. I suppose it doesn't trigger very often though.)