------- Comment #4 from amodra at bigpond dot net dot au 2005-11-16 11:29 ------- The problem here is late forcing of fp constants to memory. The first scheduling pass has merrily moved an insn loading a fp constant in amongst the function pointer load sequence, after r2 has been loaded. There is nothing to tell the scheduler that this is wrong, because at this point there is no dependency on r2. Later, in greg pass we decide the constant needs to go to memory, and suddenly you have an r2 dependecy. David, you fixed this on mainline by forcing fp constants to mem early. Any chance of a backport to 4.0?
-- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24779
