Currently, when emitting an array in the global scope, we have to pessimistically use JSOP_NEWINIT instead of JSOP_NEWARRAY because we do not yet know how many globals will be sitting underneath the stack and whether we will cause a stack depth overflow. This is the source of the slowdown in bug 171262 comment 24. Another instance of the same problem came up with the new JSOP_CONCATN in bug 419743. This pessimism could be avoided if the global-fixup pass done by compileScript took into account the immediate operand of JSOP_NEWARRAY and JSOP_CONCATN to detect the overflow thereby allowing these operands to be safely emitted.

I think this has since been fixed by bhackett's array init changes.