Copy an object passed by value a word at a time on i386

[tamnd]

On i386 a call that passed a large structure by value copied it into the outgoing area with memcpy. The copy's own three arguments go on the stack too, at the bottom of the area the call is being written into, so they overwrote the arguments already stored there. 00204 of c-testsuite passes two large structures and printed garbage for the first, which is why both i686 rung 0 jobs fail on main.

A convention with fewer than three argument registers, which is every i386 one, now copies a word at a time however large the object is. GCC makes no library call while it sets up a call's arguments either. This also takes away the nested memcpy call that ld.lld refused to link as "cannot be preempted" when the callee was a local function.

On the Linux box 00204 now matches the expected output at -O0 and -O2, with and without PIE, rung 0 passes 440 of 440 on both i686-linux-gnu and x86_64-linux-gnu, and the rucc-codegen tests and clippy pass. A new test checks that a 4096 byte object passed on i386 is copied with no call to memcpy.