Re: compiler defects and mission critical failures

Christopher Glaeser (cdg@nullstone.com) wrote:
|> >Is there any information on failures of mission critical applications
|> >which were caused by a defect in a compiler? In particular, is there
|> >any information on compiler defects which resulted in significant loss
|> >of money, damage or loss of equipment, or injury or loss of human life?


|> > "THIS COMPILER SHOULD NOT BE USED FOR MISSION-CRITICAL APPLICATIONS."


metzger@bach.convex.com (Robert Metzger) writes:
|> I found zero references to bugs of the type you are interested in mentioned
|> in the book, and I don't recall any from SEN. But my memory could be
|> faulty, so I suggest you do the market research yourself.
|> Looking for someone to give you grist for the marketing-brochure-mill, eh?


To my knowledge, also "no". Since, to my knowledge, every compiler (except
perhaps one, a BCPL compiler) that I've used has contained bugs, it could
be that people who program up mission critical applications either follow
the warranty advice, or else they test to a fare-thee-well, or both. From
working on a complex optimizing compiler, I found the ability of some bugs
to hide (in the face of continuous testing and development) to be quite
remarkable -- one bug that I introduced lay hidden for 18 months of
development before it was found (basically, in 18 months of testing, if
anyone had ever fed the compiler a particular weird flowgraph, it would
have crashed. A C++ front-end did it.)


And, this is not for lack of testing. In the last year of this particular
compiler's development, a team of 5 engineers performed something like one
CPU-year of testing per engineer, all by themselves, with additional
testing performed by the floating-point accuracy/performance group and a
consultant hired only to do testing. After it shipped, there were still
(at least) a couple dozen bugs remaining to be found by customers (or so I
was told -- I had changed jobs by then). At least two of them (mine)
were head-smacking-stupid -- a little more development discipline (write
a unit test for every enhancement, no matter how trivial) would have
caught them both.


David Chase
CenterLine Software
--

Return to the comp.compilers page.
Search the comp.compilers archives again.