Invalid build results (doesn't invalidate) when headers change (method names, enum order)

[nyanpasu64]

I'll fill in details later... but it seems nobody reads this issue tracker anymore

[frerich]

Don't worry, I've followed this issue tracker for the last couple of years (and some other people seem to do so, too). :-)

[nyanpasu64]

1. If I change a method's parameter types in .cpp and .h, I get link errors. The method's users were never recompiled, so the new mangled-name in the implementation != the old mangled-name in the method's users. This happens in CLion, and resolves once I delete homedir/clcache/ and rebuild. (BUG)

2. I edit a header file holding a class { enum {} }. I comment out some enum fields. The project errors out as expected under CLion (no bug). Nothing happens under Visual Studio (BUG).

---

Seems like clcache is completely drunk when I'm invoking through Visual Studio 2017. I edit the header file, comment out an entire enum used in a .cpp file, and Visual Studio thinks the project's unmodified. I do a full clean and rebuild, and no syntax errors pop up.

• Should I blame ReSharper and VAssistX? Probably not.
  • I've deleted homedir/clcache/ several times to fix stale headers. But the behavior keeps recurring, the next time I change headers.

INSTALLATION: For Visual Studio, I copied clcache.exe to cl.exe. I added the directory to PATH by copying the entire PATH prepend from "miniconda3 activate". I had to modify clcache's "cl.exe wrapper detector" so cl.exe wrapper wouldn't call itself recursively. Could my setup break high-level caching and header inclusion detection?

• I removed all folders except C:\Users\jimbo1qaz\Miniconda3\Scripts from clcache.props, but that didn't fix this. https://gist.github.com/jimbo1qaz/693f3ab4fa12b5e2fe465b224f896526

[frerich]

It sounds like you're using the VS IDE and/or MSBuild to do your builds? Can you please try this again, but just by executing the compiler on the command line directly -- are you still able to reproduce the linker error, i.e. clcache seemingly incorrectly reusing a previously generated object file?

[nyanpasu64]

I have reliably reproduced this in my own project, with CLion build system. Clone https://github.com/jimbo1qaz/0CC-FamiTracker/tree/clion and checkout the clion branch.

• CLion jom/ccache setup: jimbo1qaz/0CC-FamiTracker/wiki

To reproduce the bug:

• First build the project using clcache. (I configured CLion to use jom instead of nmake.) Then navigate to Source/SoundGen.{cpp,h}, and change void CSoundGen::GenerateVibratoTable(vibrato_t Type) to int Type, and perform a rebuild.

• Source/CommandLineExport.cpp should get rebuilt, but isn't.

  • Line 100: theApp.GetSoundGenerator()->GenerateVibratoTable(pExportDoc->GetVibratoStyle());

---

Minimal test case: attempt 2.zip

• The args variable is modeled after my CLion CMake arguments, with a few commented out.
• Change call aconda to whatever command places the cl.exe proxy in PATH.
• I call cl.exe once per .cpp file, like CLion CMake's Makefile, executed by nmake/jom.

First launch build.cmd (4294967295).

Then change foo.h and foo.cpp to void Foo::bar(signed x), then launch build.cmd again. (It usually works fine for me, -1).

Then undo both foo.h and foo.cpp to void Foo::bar(unsigned x). Now, (1) either the end result remains -1 despite foo.cpp being std::cout << x, or (2) I get a linking error.

[frerich]

It would be great if you could reduce this setup such that CLion and CMake wouldn't be part of the setup. Having a minimal test case would allow integrating a unit test into our CI system right away.

[nyanpasu64]

I posted a minimal test case, below the divider

Also all Appveyor builds are broken. Should I make a separate issue?

[frerich]

Ah, sorry - I thought that too contained a CMake system. I'll try to find some time to investigate this evening when I'm out of the office.

[nyanpasu64]

Did you look into the issue yet? I suspect because of the extern object in the header, its class isn't treated as a dependency when rebuilding a .cpp using that object.

[frerich]

I only looked at it briefly when creating a unit test to reproduce the issue, but I didn't yet investigate why this happens (and I probably won't get around to it this weekend either). If you would like to give it a shot, feel free to!

[drussel]

We are definitely seeing this too. Invoking clcache via cmake and ninja.

[dioss-Machiel]

Am I correct in that this means clcache is pretty much unusable since certain changes in source code may not be picked up?

[frerich]

I would be very wary when it comes to using clcache, yes. It seems there are some conditions under which it fails to notice that a cache entry cannot be reused.

I haven't used clcache myself in about five years, so I cannot say that I noticed the same symptoms. This project is mostly in maintenance-only mode (when I get around tp it), so I hope that -- as was the case in the past -- someone steps up and 'scratches his own itch' by fixing this problem. The most important thing here would be an extension of the test suite such that we can reliably reproduce the problem in the CI builds.

[Disminder]

+1 to link errors. At least two teams on my work affected via this issue. I tested performance on our project for 100% cache hits: ~8 min without server, ~6 with it. But we can't use it :(

[Disminder]

UPD: in our case it could have been caused by using the '--disable_watching' flag. Probably I need more research. And documentation about what this flag actually do and how it can affect.

[Artalus]

Just stumbled upon this issue and freaked out (obviously), so I decided to get @jimbo1qaz's sample a test run.

It seems that the bug appears only if CLCACHE_HARDLINK is set. If I replace the line in build.cmd with set CLCACHE_HARDLINK= and remove the clcache directory, the rebuilds happen as expected. CLCACHE_SERVER presence doesn't seem to affect the bug.

The sample can also be simplified down to:

//foo.cpp
#include <iostream>
void foo(signed x) {
	std::cout << x << std::endl;
}
//foo.h
#pragma once
void foo(signed x);
//main.cpp
#include "foo.h"
int main() {
  foo(-1);
}