On 17 January 2014 21:16, Stéphane Ducasse <stephane.ducasse@inria.fr>wrote:
For normal humans you do not convey the pain behind â¦. here is the answer I got from the two amigos igor and ronie that looked at the problem
about the pain behind..
Stef, i'm not sure, maybe for those who not familiar with low-level debugging, the procedures look very scary and traumatizing, but for not for me nor Ronie.. i don't remember but i think it took us less than an hour to determine what's wrong there. And to fix that, it took few seconds. So, it was done very quickly, alas... then i faced integration wall.. and it took me much more time & effort to verify & wait till it reverify things in many places.. because for some strange reason integration validation says there's failing test.. which is not if you run things by yourself. All this pain, of course, doesn't means, that i am against validation. It really helps, but sure thing it is hard to cover all aspects to make sure things run under same conditions as in reality.. and it always be like that. (and speaking about conveying the pain behind... imagine the amount of pain setting up & automating validation procedure and integration, so that it works and we can use it.. )
We debugged in both machines at the same time. Installed some debug symbols, looking at the libpixman source code. Disassembling and looking some instructions.
When I saw the use of intrinsics in the source code, I started to suspect a bit of the alignment of some stuff and when Igor, found the segmentation fault causing instruction movdqa (Move double-quadword aligned), it was clear that it was a stack alignment issue. Checking the Intel manual confirmed this and I asked about whose responsible for aligning the stack, suspecting it was a problem with the library not conforming to the ABI. Then, Igor remembered how to change the stack alignment in native boost and it worked.
This is why having people with such knowledge around is key to let us (the smart cosy smalltalkers) having fun with our cool language.
Stef
Fix integrated.
in 3.0 704
-- Best regards, Igor Stasenko.
-- Best regards, Igor Stasenko.