Re: [Qemu-devel] [PATCH 0/5] tcg conditional set, round 4

[Richard Henderson]

On 12/20/2009 02:57 PM, Paul Brook wrote:
> On Saturday 19 December 2009, Richard Henderson wrote:
> > Changes from round 3:
> >
> >   * Drop movcond for now.
> >   * Only use movzbl and not xor in setcond.
>
> I'm still catching up on mail backlog from this thread, but I'm concerned that
> we're exposing setcond to the target translation code if we're planning on
> implementing movcond later.

I personally was only planning to use setcond in those cases where the 
target actually wants a 0/1 value.  E.g. i386 setcc, or alpha cmp*, or 
mips slt*.

As for shifts and masks, that wasn't in my plans.  I had a comment in 
there about all the tricks that gcc plays with a conditional move of two 
constants, and the fact that I didn't think it worth the effort.

Indeed, my plan for today was to cut down my existing movcond patch to 
make it simpler, as that seems to be the way Aurelien wants this to go. 
 Life conspired against and I got nothing done, but still.

I *am* convinced that to remove either VTRUE or VFALSE as arguments to 
the movcond primitive (implementing dest = (cond ? vtrue : dest) would 
do too much violence to both the liveness analysis and the register 
allocator within TCG.  The best I can do to remove the complexity is:

static void tcg_out_movcond(TCGContext *s, int cond, TCGArg dest,
                            TCGArg cmp1, TCGArg cmp2, int const_cmp2,
                            TCGArg vtrue, int rexw)
{
    tcg_out_cmp(s, cond, cmp1, cmp2, const_cmp2, rexw);
    /* cmovcc */
    tcg_out_modrm(s, 0x40 | tcg_cond_to_jcc[cond] | P_EXT | rexw,
                  dest, vtrue);
}
...
    { INDEX_op_movcond_i32, { "r", "r", "ri", "r", "0" } },
    { INDEX_op_movcond_i64, { "r", "r", "re", "r", "0" } },

using matching constraints in the target and directly implement the 
conditional move.  This eliminates the code I previously had that 
checked for things like dest=vfalse and inverted the condition.  I had 
planned to simplify the i386 version similarly, even in the case where 
cmovcc is not available.

I would appreciate some direction here, so as to avoid wasting my time 
with a solution that won't be accepted.


r~