------- Additional Comments From law at redhat dot com 2005-01-24 20:46 -------
Subject: Re: computed gotos are not folded
back to regular gotos when it is found that they are not computed gotos
at
all
On Sun, 2005-01-23 at 21:09 +0000, kazu at cs dot umass dot edu wrote:
> ------- Additional Comments From kazu at cs dot umass dot edu 2005-01-23
> 21:09 -------
> Here is another idea that would not compilicate the DOM.
> That is, we can set up things so that the actual threading part will be
> done by DOM.
>
> Suppose we have a factored computed goto block like so:
>
> # p_2 = PHI <&L0(0), &L1(1), p_3(2)>;
> # b_1 = PHI <3(0), 5(1), 4(2)>;
> <L2>:;
> ... possibly some other statements ...
> goto p_2;
>
>
> Let's split this block like so:
>
>
> # p_2 = PHI <&L0(0), &L1(1), p_3(2)>;
> # b_1 = PHI <3(0), 5(1), 4(2)>;
> <L2>:;
> ... possibly some other statements ...
>
> <L3>:; // fall through from L2.
> goto p_2;
>
>
> Note that we are still in SSA form without need for rewriting.
>
> Now let's add a new PHI node and a new SWITCH_EXPR to <L2> like so:
>
>
> # p_2 = PHI <&L0(0), &L1(1), p_3(2)>;
> # b_1 = PHI <3(0), 5(1), 4(2)>;
> # fac_1 = PHI <0(0), 1(1), 2(2)>;
> <L2>:;
> ... possibly some other statements ...
> switch (fac_1)
> {
> case 0: goto L0; <- a threadable, normal edge
> case 1: goto L1; <- a threadable, normal edge
> default: goto <L3>; <- not threadable because p_3 is an SSA_NAME
> }
>
> <L3>:;
> goto p_2;
>
> That is, we create a new PHI node fac_1 so that we can build a
> dispatch table using a SWITCH_EXPR.
>
> Now it's easy for DOM to pick up the jump threading opportunities at
> SWITCH_EXPR.
>
> If you like, you can say that we "reduce" the factored computed goto
> block to a SWITCH_EXPR.
>
> Any thoughts?
Interesting approach. I'm not sure if your approach is easier
than just expanding the threading code to handle threading indirect
jumps.
I don't initially *think* that extending the threading code would be
all that difficult. In fact, all I think we need to do is update
the selection code to detect this case -- I think the code to update
the CFG and SSA graphs will DTRT without any changes.
Jeff
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18133
