[LLVMdev] [cfe-dev] !!! 3.2 Release RC2 deadline November 29th

John McCall rjmccall at apple.com
Wed Nov 28 15:09:23 PST 2012


On Nov 28, 2012, at 11:49 AM, David Chisnall <David.Chisnall at cl.cam.ac.uk> wrote:
> On 28 Nov 2012, at 19:41, Tanya Lattner wrote:
> 
>> So why isn't this caught by any of the regression testing (testsuite) if it nothing above -O1 can be compiled? I assume you mean for your project?
> 
> Any Objective-C on non-Apple platforms that uses cleanups (including those Clang inserts for blocks) breaks in certain cases with inlining.  There are also some cases in C++, and cases in LTO (or using C headers in C++) when C++ code and C code using __attribute__((cleanup)) are mixed.  We have an ugly hack in clang that avoids it in the most common case, but there are almost certainly cases that we are not catching.  
> 
>> I'm trying to understand how large of a problem this is. Ideally, we want everything we can to get fixed.. but since there is some confusion and debate on this bug, I'm thinking its not a quick or easy fix.
> 
> There is no debate in the bug.  The consensus (both in the bug report and in-person discussions at the DevMeeting) is that the inliner is doing an obviously wrong thing when it takes a function containing a catchall and one containing a cleanup, inlines one into the other, and ends up with one with just a cleanup.  The discussion in the bug report is due to the fact that there are probably also some corner cases where the wrong thing happens, but the inliner bug is the serious one.

I absolutely agree that the inliner seems to be doing the wrong thing on your testcases, but IIRC, your testcases were all hand-written IR, and the inliner seemed to be doing the right thing for your original source-code testcase.  I don't know of a reason why that should matter, but I'm hesitant to call this a release blocker if you can't duplicate this starting from source code, because it's possible that the inliner is only broken on "unrealistic" IR — which would still definitely be a bug, but not one that should hold up the release.

Just to be clear, anything that a major frontend like clang would emit for actual source code is ipso facto "realistic".  I'm just asking that you find a way to duplicate with ObjC source, not with hand-written IR.

John.



More information about the llvm-dev mailing list