This is the mail archive of the
mailing list for the GCC project.
Re: [tuples] gimple_assign_subcode for GIMPLE_SINGLE_RHS
- From: "Diego Novillo" <dnovillo at google dot com>
- To: "Zdenek Dvorak" <rakdver at kam dot mff dot cuni dot cz>
- Cc: gcc at gcc dot gnu dot org, aldyh at redhat dot com, "William Maddox" <maddox at google dot com>
- Date: Sun, 9 Mar 2008 08:16:12 -0500
- Subject: Re: [tuples] gimple_assign_subcode for GIMPLE_SINGLE_RHS
- References: <20080309002946.GA22003@kam.mff.cuni.cz>
On Sat, Mar 8, 2008 at 19:29, Zdenek Dvorak <email@example.com> wrote:
> The problem of course is that for GIMPLE_SINGLE_RHS, we do not maintain
> the invariant that
> gimple_assign_subcode (stmt) == TREE_CODE (gimple_assign_rhs1 (stmt)),
> so gimple_assign_subcode typically will not be SSA_NAME, but VAR_DECL.
> Enforcing this invariant might be hard and probably asking for more
> trouble than it is worth. However, perhaps it would make sense
> to use some special tree code to indicate GIMPLE_SINGLE_RHS, in order
> to avoid confusion?
Hmm, yeah. I remember talking about this with Bill a few days ago.
He had a similar idea.
At the time I was a bit reluctant to add a new tree code just for
this, but now you've run into it and I was also looking at some code
that would run into the same problem. Now I agree that adding a new
tree code will be the cleanest way. We will be removing several tree
codes when tuple merges, so the net result in tree.def will be a
reduction of tree codes.
In this case, the renamer was at fault. When it replaced the USE
operand with an SSA_NAME it failed to change the subcode.
The other problem with not using a special code for GIMPLE_SINGLE_RHS
a_3 = b_5
is (if we fixed the renamer bug)
GIMPLE_ASSIGN <SSA_NAME, SSA_NAME<a_3>, SSA_NAME <b_5>>
So, the subcode seems to duplicate the code on the RHS. That looks odd.
So, what about adding a GIMPLE_COPY code? The code would have 0
operands and used only for its numeric value.