On Wed, Dec 2, 2015 at 9:33 AM, Richard Sandiford
<richard.sandif...@arm.com> wrote:
> The problem in the testcase was that tree-complex.c was trying
> to fold ABS_EXPRs of SSA names that didn't yet have a definition
> (because the SSA names were real and imaginary parts of a complex
> SSA name whose definition hadn't yet been visited by the pass).
> tree-complex.c uses a straightforward walk in index order:
>
> /* ??? Ideally we'd traverse the blocks in breadth-first order. */
> old_last_basic_block = last_basic_block_for_fn (cfun);
> FOR_EACH_BB_FN (bb, cfun)
> {
>
> and in the testcase, we have a block A with a single successor B that
> comes before it. B has no other predecessor and has a complex division
> that uses an SSA name X defined in A, so we split the components of X
> before we reach the definition of X. (I imagine cfgcleanup would
> clean this up by joining A and B.)
>
> Tested on x86_64-linux-gnu. OK to install?
I think the new checks are just bogus because all SSA names in the IL ought
to have a def-stmt (if only a GIMPLE_NOP).
So I think what tree-complex.c does it just wrong(TM) and worked by
accident only.
So I suggest instead compute a proper CFG order and process basic blocks in
that order instead. (a domwalk doesn't work as explained in the PR)
Thanks,
Richard.
> Thanks,
> Richard
>
>
> gcc/
> PR tree-optimization/68146
> * fold-const.c (tree_binary_nonnegative_warnv_p): Check for null
> SSA_NAME_DEF_STMTs.
> (integer_valued_real_call_p): Likewise.
>
> gcc/testsuite/
> * gfortran.fortran-torture/compile/pr68146.f90: New test.
>
> diff --git a/gcc/fold-const.c b/gcc/fold-const.c
> index 16bff5f..c99e78e 100644
> --- a/gcc/fold-const.c
> +++ b/gcc/fold-const.c
> @@ -12867,6 +12867,8 @@ tree_binary_nonnegative_warnv_p (enum tree_code code,
> tree type, tree op0,
> bool
> tree_single_nonnegative_warnv_p (tree t, bool *strict_overflow_p, int depth)
> {
> + gimple *stmt;
> +
> if (TYPE_UNSIGNED (TREE_TYPE (t)))
> return true;
>
> @@ -12892,8 +12894,9 @@ tree_single_nonnegative_warnv_p (tree t, bool
> *strict_overflow_p, int depth)
> to provide it through dataflow propagation. */
> return (!name_registered_for_update_p (t)
> && depth < PARAM_VALUE (PARAM_MAX_SSA_NAME_QUERY_DEPTH)
> - && gimple_stmt_nonnegative_warnv_p (SSA_NAME_DEF_STMT (t),
> - strict_overflow_p, depth));
> + && (stmt = SSA_NAME_DEF_STMT (t))
> + && gimple_stmt_nonnegative_warnv_p (stmt, strict_overflow_p,
> + depth));
>
> default:
> return tree_simple_nonnegative_warnv_p (TREE_CODE (t), TREE_TYPE (t));
> @@ -13508,6 +13511,7 @@ integer_valued_real_call_p (combined_fn fn, tree
> arg0, tree arg1, int depth)
> bool
> integer_valued_real_single_p (tree t, int depth)
> {
> + gimple *stmt;
> switch (TREE_CODE (t))
> {
> case REAL_CST:
> @@ -13524,8 +13528,8 @@ integer_valued_real_single_p (tree t, int depth)
> to provide it through dataflow propagation. */
> return (!name_registered_for_update_p (t)
> && depth < PARAM_VALUE (PARAM_MAX_SSA_NAME_QUERY_DEPTH)
> - && gimple_stmt_integer_valued_real_p (SSA_NAME_DEF_STMT (t),
> - depth));
> + && (stmt = SSA_NAME_DEF_STMT (t))
> + && gimple_stmt_integer_valued_real_p (stmt, depth));
>
> default:
> break;
> diff --git a/gcc/testsuite/gfortran.fortran-torture/compile/pr68146.f90
> b/gcc/testsuite/gfortran.fortran-torture/compile/pr68146.f90
> new file mode 100644
> index 0000000..7f75ec0
> --- /dev/null
> +++ b/gcc/testsuite/gfortran.fortran-torture/compile/pr68146.f90
> @@ -0,0 +1,11 @@
> +subroutine foo(a1, a2, s1, s2, n)
> + integer :: n
> + complex(kind=8) :: a1(n), a2(n), s1, s2
> + do i = 1, n
> + a1(i) = i
> + end do
> + s1 = 20.0 / s2
> + do i = 1, n
> + a2(i) = i / s2
> + end do
> +end subroutine foo
>
