Often, the papers would get the general idea right, but if implemented as described would not work at all or fail on edge cases. The best example I have is an algorithm to find the lexicographically-minimal string rotation[0]. The simplest and fastest algorithm to do this is based on the KMP string search algo, and is tribal knowledge among ACM ICPC competitors. I thought it was pretty neat and wanted to cement this algorithm in popular knowledge, so I set about researching and writing the Wikipedia article.
I found the KMP-based algorithm in a 1980 paper[1] by Kellogg S. Booth. The paper has very detailed pseudocode which does not work. At all. The tribal knowledge version I inherited had similarities in the general idea of the algorithm (use of the KMP preprocessing step) but everything else was different. I scoured the internet for a retraction or correction, but all I found was a paper written in 1995[2] which mentioned in passing errors in the 1980 paper.
I do wonder exactly how common this is. I emailed a professor who co-wrote one of the papers, and he replied that "it seems to me that all the algorithms (including our own) turned out to have errors in them!" Has anyone done studies into errors in computer science papers?
[0] https://en.wikipedia.org/wiki/Lexicographically_minimal_stri...
[1] http://www.sciencedirect.com/science/article/pii/00200190809...
[2] http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.55.9...