Something I noticed while working on that deposit patch for ia64
was how annoying it is to bundle instructions by hand.  And the
fact that it's been done incorrectly at least once.  E.g.

static inline void tcg_out_bswap64(TCGContext *s, TCGArg ret, TCGArg arg)
{
    tcg_out_bundle(s, mII,
                   tcg_opc_m48(TCG_REG_P0, OPC_NOP_M48, 0),
                   tcg_opc_i18(TCG_REG_P0, OPC_NOP_I18, 0),
                   tcg_opc_i3 (TCG_REG_P0, OPC_MUX1_I3, ret, arg, 0xb));

Notice that there's an unnecessary stop bit in there.

This patch has does too much all at once, I'll admit that.  But
before I bother going back to split it into smaller pieces, I
wanted to get some feedback -- including whether it would be
considered at all.

Some statistics that I gleaned from grepping -d out_asm.  Both before
and after the rewrite, the only two linux-user targets that work are
i386 and alpha.  Clearly there are still bugs to fix, but it's not
necessarily a regression.

		Code size			Stop bits
		    old	    new	change		   old	   new	change
i386/ls		2309712	1602736	-31%		215520	 98698	-54%
alpha/ls	1088352	 817888	-25%		106215	 48440	-54%

On the system side, I am able to boot the arm-test kernel.  But
it's a bit harder to get repeatable statistics there.  I assume
that's all timing issues, and what code gets run when, and how
that affects the TB cache.

Anyway, comments greatly appreciated.


r~