On Thu, Jul 5, 2012 at 11:53 PM, Steven Bosscher <stevenb....@gmail.com> wrote:
> Hello,
>
> This is a first stab at making the core CFG code independent of the
> expression intermediate representation datatypes. This patch deals
> with cfg.c and cfganal.c but it shouldn't be a surprise that the patch
> touches a lot of other files.
>
> I removed all IR headers (tree.h and rtl.h) from function.h and
> basic-block.h, and that caused some fall-out left and right. Who would
> have expected that ira-conflicts.c depends on tree.h?! Or that cfg.c
> did not include basic-block.h directly, but only inherited from
> cfgloop.h?
>
> There were also a couple of places where an RTL pass was ported to
> GIMPLE but still used RTL CFG dump functions.
>
> The only bit I'm not very happy with, is the
> set_loop_copy/get_loop_copy change. I did not want cfg.c to be
> dependent on cfgloop.h ony for struct loop, but I also did not want to
> duplicate the original_table machinery from cfg.c, or export it from
> there. This was the least ugly solution I could come up with. Looking
> back at the change now, I wonder whether I can use a pointer-set
> instead and move set_loop_copy and get_loop_copy to cfgloopmanip.c.
>
> Otherwise, I think this shows the direction I'm trying to take things.
> My plan for the CFG code is to end up with one common CFG header and a
> cfgrtl.h and tree-cfg.h. It is unfortunate that this involves moving a
> lot of functions from one file to another, but that is IMHO an
> inevitable cost of trying for a more towards a more modular build-up
> of the compiler.
>
> Bootstrapped and tested on powerpc64-unknown-linux-gnu. I have a bug
> somewhere in the profiling code that gives a bunch of test suite
> failures, but I'm looking for comments/suggestions/... already, so
> that I don't do too much useless work :-)
> (Without IBM's gcc110 donation, I would not have been able to do all this
> work!)
Wow ;) I agree about the set_loop_copy/get_loop_copy change - I'd simply
postpone it with a comment why cfgloop.h is needed and investigate the
pointer-map idea.
Generally I think dump_* routines should work even when cfun is not the
function we dump an edge for. Thus,
- /* both ENTRY_BLOCK_PTR & EXIT_BLOCK_PTR depend upon cfun. */
- if (cfun && side == ENTRY_BLOCK_PTR)
+ if (side == ENTRY_BLOCK_PTR)
fputs (" ENTRY", file);
- else if (cfun && side == EXIT_BLOCK_PTR)
+ else if (side == EXIT_BLOCK_PTR)
fputs (" EXIT", file);
should instead test side->index == ENTRY_BLOCK / EXIT_BLOCK.
Otherwise the patch looks really good - best if teamed with updating
doc/cfg.texi ;)
Thanks,
Richard.
> Ciao!
> Steven
