dealing with unary minus applied to a constant.  The removed code was
mutating the ST, causing a second compilation to fail.  (The peephole
optimizer already takes care of optimizing this case, so there's no
lost optimization opportunity here.)