[RFC] optimize x - y cmp 0 with undefined overflow

On Mon, May 26, 2014 at 12:22 PM, Eric Botcazou <ebotcazou@adacore.com> wrote:
> Hi,
>
> the motivation of this work is to get rid of the range check on Z in:
>
> function F (X : Integer; Y : Integer ) return Positive is
>    Z : Integer;
> begin
>    if X >= Y then
>       return 1;
>    end if;
>    Z := Y - X;
>    return Z;
> end;
>
> An equivalent C testcase is:
>
> extern void abort (void);
>
> int foo (int x, int y)
> {
>   int z;
>
>   if (x >= y)
>     return 1;
>
>   z = y - x;
>   if (z <= 0)
>     abort ();
>
>   return z;
> }
>
> for which the call to abort is not optimized at -O2.
>
>
> fold_comparison knows how to optimize X +- C1 CMP C2 to X CMP C2 -+ C1 with
> undefined overflow so optimizing X - Y CMP 0 to X CMP Y is a generalization.
>
> Once this is done, forwprop will immediately optimize:
>
> extern void abort (void);
>
> int foo (int x, int y)
> {
>   int z;
>
>   if (x >= y)
>     return 1;
>
>   z = y - x;
>   if (z <= 0)
>     abort ();
>
>   return 0;
> }
>
> because 'z' has a single use, but the original testcase won't be optimized.
>
>
> The attached patch implements the missing bits in vrp_evaluate_conditional by
> manual propagating the operands of a defining PLUS_EXPR or MINUS_EXPR for an
> SSA name in a condition; an other possibility could be DOM instead of VRP.
>
> This comes with 4 testcases: the original Ada testcase, the C equivalent, a
> testcase for the folding and another one for the -Wstrict-overflow warning.
>
> Tested on x86_64-suse-linux with no regressions.

+      tree new_const
+       = fold_build2_loc (loc, reverse_op, TREE_TYPE (arg1), const2, const1);

       /* If the constant operation overflowed this can be
         simplified as a comparison against INT_MAX/INT_MIN.  */
-      if (TREE_CODE (lhs) == INTEGER_CST
-         && TREE_OVERFLOW (lhs))
+      if (TREE_OVERFLOW (new_const))

well, either use int_const_binop above or retain the check (or use
TREE_OVERFLOW_P).  Bonus points for using wide-ints here
and not relying on TREE_OVERFLOW.

+  /* Transform comparisons of the form X - Y CMP 0 to X CMP Y.  */
+  if (TREE_CODE (arg0) == MINUS_EXPR
+      && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (arg1))

any good reason for using TYPE_OVERFLOW_UNDEFINED on the
type of arg1 instead on the type of the MINUS (yes, they should
match, but it really looks odd ... the overflow of the minus has to be
undefined).  Also for EQ_EXPR and NE_EXPR the transform is
fine even when !TYPE_OVERFLOW_UNDEFINED (and we seem
to perform it already somewhere).  Please look where and try to
add the undefined overflow case to it.

As for the VRP pieces I don't understand what is missing here
(well, compare_range_with_value and/or compare_values might
not handle this case?  then better fix that to improve symbolic
range handling in general?).  Also I have a longstanding patch
in my tree that does

maybe that helps as well.

Thanks,
Richard.

>
> 2014-05-26  Eric Botcazou  <ebotcazou@adacore.com>
>
>         * fold-const.c (fold_comparison): Clean up and simplify X +- C1 CMP C2
>         to X CMP C2 -+ C1 transformation, add X - Y CMP 0 to X CMP Y.
>         * tree-vrp.c (vrp_evaluate_conditional_warnv_with_fold): New function.
>         (vrp_evaluate_conditional): Call it on SSA names with defining PLUS_EXPR
>         or MINUS_EXPR if the evaluation of the condition yielded nothing.
>
>
> 2014-05-26  Eric Botcazou  <ebotcazou@adacore.com>
>
>         * gcc.dg/fold-compare-8.c: New test.
>         * gcc.dg/Wstrict-overflow-25.c: Likewise.
>         * gcc.dg/tree-ssa/vrp92.c: Likewise.
>         * gnat.dg/opt38.adb: Likewise.
>
>
> --
> Eric Botcazou

[RFC] optimize x - y cmp 0 with undefined overflow

Commit Message

Comments

Patch