[PATCH] Fix folding of X - X to 0

[PATCH] Fix folding of X - X to 0

[Richard Biener]

The following fixes a bug in fold-const.c which happily drops
side-effects when simplifying X - X to 0.  I ran into this because
on match-and-simplify I cannot simplify expressions with
TREE_SIDE_EFFECTS and it happens that TYPE_SIZE for VLAs has
side-effects because it usually contains a SAVE_EXPR which are
forced to have side-effects (even if its arg does not have
side-effects).  This is appearantly because you can use SAVE_EXPRs
to force creation of initialized temporaries at defined place
in the instruction stream.  For example fold-const.c currently
would miscompile

 (save_expr<global> - save_expr<global>, (global = 1, save_expr<global>))

Of course all this makes my life hard on match-and-simplify
branch and I wonder why we need to use SAVE_EXPRs for these cases
where the initialization of a new temporary could be done
explicitely (using INIT_EXPR or TARGET_EXPR).

Anyway - fixing the fold-const.c bug requires fixing the bug
I run into on match-and-simplify branch - here by simply
"stripping" the side-effects to get to the actual difference
of the two DECL_FIELD_OFFSETs (I wonder if I missed an obvious
tool to do such stripping).

Bootstrap & regtest running on x86_64-unknown-linux-gnu.

Of course this doesn't help me on match-and-simplify branch
(example testcase that ICEs there with the assert in stor-layout.c
visible below is gcc.c-torture/execute/20040308-1.c).  In
theory it should be ok to simplify these late in GIMPLE and
for the stor-layout.c case unset TREE_SIDE_EFFECTS temporarily
to make generic_simplify happy ... (ugh).

Thanks,
Richard.

2014-10-22  Richard Biener  <rguenther@suse.de>

	* fold-const.c (fold_binary_loc): Preserve side-effects of
	X - X when simplifying to 0.
	* stor-layout.c (finish_bitfield_representative): Strip
	side-effects of evaluating the difference of two DECL_FIELD_OFFSET.

Index: gcc/fold-const.c
===================================================================
--- gcc/fold-const.c	(revision 216543)
+++ gcc/fold-const.c	(working copy)
@@ -10749,7 +10768,7 @@ fold_binary_loc (location_t loc,
 
       if ((!FLOAT_TYPE_P (type) || !HONOR_NANS (TYPE_MODE (type)))
 	  && operand_equal_p (arg0, arg1, 0))
-	return build_zero_cst (type);
+	return omit_one_operand_loc (loc, type, build_zero_cst (type), arg0);
 
       /* A - B -> A + (-B) if B is easily negatable.  */
       if (negate_expr_p (arg1)
Index: gcc/stor-layout.c
===================================================================
--- gcc/stor-layout.c	(revision 216543)
+++ gcc/stor-layout.c	(working copy)
@@ -1860,6 +1860,8 @@ finish_bitfield_representative (tree rep
 
   size = size_diffop (DECL_FIELD_OFFSET (field),
 		      DECL_FIELD_OFFSET (repr));
+  while (TREE_CODE (size) == COMPOUND_EXPR)
+    size = TREE_OPERAND (size, 1);
   gcc_assert (tree_fits_uhwi_p (size));
   bitsize = (tree_to_uhwi (size) * BITS_PER_UNIT
 	     + tree_to_uhwi (DECL_FIELD_BIT_OFFSET (field))

Re: [PATCH] Fix folding of X - X to 0

[Jeff Law]

On 10/22/14 06:58, Richard Biener wrote:
>
> The following fixes a bug in fold-const.c which happily drops
> side-effects when simplifying X - X to 0.  I ran into this because
> on match-and-simplify I cannot simplify expressions with
> TREE_SIDE_EFFECTS and it happens that TYPE_SIZE for VLAs has
> side-effects because it usually contains a SAVE_EXPR which are
> forced to have side-effects (even if its arg does not have
> side-effects).  This is appearantly because you can use SAVE_EXPRs
> to force creation of initialized temporaries at defined place
> in the instruction stream.  For example fold-const.c currently
> would miscompile
>
>   (save_expr<global> - save_expr<global>, (global = 1, save_expr<global>))
>
> Of course all this makes my life hard on match-and-simplify
> branch and I wonder why we need to use SAVE_EXPRs for these cases
> where the initialization of a new temporary could be done
> explicitely (using INIT_EXPR or TARGET_EXPR).
>
> Anyway - fixing the fold-const.c bug requires fixing the bug
> I run into on match-and-simplify branch - here by simply
> "stripping" the side-effects to get to the actual difference
> of the two DECL_FIELD_OFFSETs (I wonder if I missed an obvious
> tool to do such stripping).
>
> Bootstrap & regtest running on x86_64-unknown-linux-gnu.
>
> Of course this doesn't help me on match-and-simplify branch
> (example testcase that ICEs there with the assert in stor-layout.c
> visible below is gcc.c-torture/execute/20040308-1.c).  In
> theory it should be ok to simplify these late in GIMPLE and
> for the stor-layout.c case unset TREE_SIDE_EFFECTS temporarily
> to make generic_simplify happy ... (ugh).
>
> Thanks,
> Richard.
>
> 2014-10-22  Richard Biener  <rguenther@suse.de>
>
> 	* fold-const.c (fold_binary_loc): Preserve side-effects of
> 	X - X when simplifying to 0.
> 	* stor-layout.c (finish_bitfield_representative): Strip
> 	side-effects of evaluating the difference of two DECL_FIELD_OFFSET.
OK.
jeff