There has been a recent effort in the literature to reconsider
grammar-dependent software development from an engineering point of view.  As
part of that effort, we examine a deficiency in the state of the art of
practical LR parsing.  Specifically, LALR sometimes generates parser tables
that do not accept the full language that the grammar developer expects, but
canonical LR is too inefficient to be practical.  In response, many researchers
have attempted to develop minimal LR algorithms.  In this paper, we demonstrate
that a well known algorithm described by David Pager and implemented in Menhir,
the most robust minimal LR(1) implementation we have discovered, does not
always achieve the power of canonical LR(1) if the given grammar is non-LR(1).
We also present an original minimal LR(1) algorithm, IELR(1), which we have
implemented as an extension of GNU Bison and which does not exhibit this
deficiency.  Finally, using our implementation, we demonstrate the relevance of
this deficiency in several real-world grammars, and we show that our
implementation is feasible for running on typical modern home and office
computing equipment.