There is a series of mistakes in the presentation of the GVNPRE algorithm, both as it appears in the CC paper and in my disseration, specifically in the discussion of the following example which appears as Figure 3(a) in the CC paper (pg 174) and Figure 4.2(a) in my disseration (pg 66).
In the paragraph beginning with "For an antileader set..." (pg 175 in the CC paper) or the paragraph beginning with "Although our methd..." (pg 68-69 in the dissertation), each ocassion where t4 or t6 are mentioned should be replaced with t11; likewise v4 should be v11:
Consider basic block 6 in Figure [3(a)/4.2(a)], alternately with and without the instruction t11 <- o. In the case wehre we exclude that instruction, assume t11 to be global. Without the definition of t11, ANTIC_IN can be represented by the dag in Figure [3(c)/4.2(c)]. The nodes of the dag are pairs of values and the antileaders representing them; edges are determined by the operands of the antileader. If we suppose the block contains t11 <- o as it does in the program, then v11 : t11 is killed, along with all expression that depend on v11, in cascading effect. See Figure [3(d)/4.2(d)].
Thanks to Tom de Vries for identifying this.