[ 2017-03-12, post-Kona ]

Resolved by P0623R0.

There is a surprising sequential operation applying BinaryOp1 in inner_product that may, for example, require additional storage for the parallel algorithms to enable effective distribution of work, and is likely to be a performance bottleneck. GENERALIZED_SUM is probably intended here for the parallel version of the algorithm, with the corresponding strengthening on constraints on BinaryOp1 to allow arbitrary order of evaluation.

Proposed change: For the overloads taking an execution policy, copy the current specification, but replace algorithm in Effects with

GENERALIZED_SUM(plus<>(), init, multiplies<>(*i1, *i2), ...)
GENERALIZED_SUM(binary_op1, init, binary_op2(*i1, *i2), ...)