Question re tree-ssa-ccp.c:substitute_and_fold

Question re tree-ssa-ccp.c:substitute_and_fold

[Richard Kenner]

This is PR18662, the failure of ACATS test c97114a.

This ICE's with

../../test/c97114a.adb: In function 'C97114A.T':
../../test/c97114a.adb:132: error: Statement marked for throw, but doesn't.
#   VUSE <S0D.926_251>;
D.1121_112 = S0D.926.ARRAYD.924[1]{lb: 1 sz: 8}.sD.818;

Before replace_uses_in, the subscript 1 was a variable (D.1120_111).
Since flag_non_call_exceptions is true, the former statement can trap
but the latter cannot.

But because replaced_address is still zero, we don't call
maybe_clean_eh_stmt and we need to.

The patch below calls it unconditionally.  That may not be correct,
though I'm not not at all sure what the correct condition is.

Suggestions?

*** tree-ssa-ccp.c	23 Nov 2004 01:27:41 -0000	2.51
--- tree-ssa-ccp.c	24 Nov 2004 14:55:10 -0000
*************** substitute_and_fold (void)
*** 583,593 ****
  		 need to rename VDEFs.  */
  	      if (replaced_address || changed)
! 		{
! 		  mark_new_vars_to_rename (stmt, vars_to_rename);
! 		  if (maybe_clean_eh_stmt (stmt))
! 		    tree_purge_dead_eh_edges (bb);
! 		}
  	      else
  		modify_stmt (stmt);
  	    }
  
--- 583,591 ----
  		 need to rename VDEFs.  */
  	      if (replaced_address || changed)
! 		mark_new_vars_to_rename (stmt, vars_to_rename);
  	      else
  		modify_stmt (stmt);
+ 	      if (maybe_clean_eh_stmt (stmt))
+ 		tree_purge_dead_eh_edges (bb);
  	    }
  


Note that a lot of the functions in tree-eh.c don't have the required
documentation in front of them.

Re: Question re tree-ssa-ccp.c:substitute_and_fold

[Diego Novillo]

On Fri, 2004-11-26 at 20:53 -0500, Richard Kenner wrote:
> This is PR18662, the failure of ACATS test c97114a.
> 
> This ICE's with
> 
> ../../test/c97114a.adb: In function 'C97114A.T':
> ../../test/c97114a.adb:132: error: Statement marked for throw, but doesn't.
> #   VUSE <S0D.926_251>;
> D.1121_112 = S0D.926.ARRAYD.924[1]{lb: 1 sz: 8}.sD.818;
> 
> Before replace_uses_in, the subscript 1 was a variable (D.1120_111).
> Since flag_non_call_exceptions is true, the former statement can trap
> but the latter cannot.
> 
> But because replaced_address is still zero, we don't call
> maybe_clean_eh_stmt and we need to.
> 
> The patch below calls it unconditionally.  That may not be correct,
> though I'm not not at all sure what the correct condition is.
> 
> Suggestions?
> 
I had to do the same thing on tree-cleanup-branch.  Similar reasons.
You also want to unconditionally call modify_stmt in here.  So you end
up with:

          if (did_replace)
            {
              bool changed = fold_stmt (bsi_stmt_ptr (i));
              stmt = bsi_stmt(i);

              /* If we folded a builtin function, we'll likely
                 need to rename VDEFs.  */
              if (replaced_address || changed)
                mark_new_vars_to_rename (stmt, vars_to_rename);

              /* If we cleaned up EH information from the statement,
                 remove EH edges.  */
              if (maybe_clean_eh_stmt (stmt))
                tree_purge_dead_eh_edges (bb);

              modify_stmt (stmt);
            }


Diego.

Re: Question re tree-ssa-ccp.c:substitute_and_fold

[Richard Kenner]

    I had to do the same thing on tree-cleanup-branch.  Similar reasons.
    You also want to unconditionally call modify_stmt in here.  So you end
    up with:

Can you to import it in from that branch?  I don't have that branch
checked out anyplace yet.  Or else just send me the ChangeLog
for it and I'll do it.

Thanks.

Re: Question re tree-ssa-ccp.c:substitute_and_fold

[Diego Novillo]

On Sun, 2004-11-28 at 18:18 -0500, Richard Kenner wrote:
>     I had to do the same thing on tree-cleanup-branch.  Similar reasons.
>     You also want to unconditionally call modify_stmt in here.  So you end
>     up with:
> 
> Can you to import it in from that branch?  I don't have that branch
> checked out anyplace yet.  Or else just send me the ChangeLog
> for it and I'll do it.
> 
Not really.  The diff contains many other changes.  That's inside a
heavily re-written portion of the compiler.  It's easier if you just c-
n-p the snippet I sent.  The only difference with your patch is that it
unconditionally calls modify_stmt.  The rest is identical.


Diego.