On Tue, Jun 7, 2011 at 8:31 AM, Jakub Jelinek <ja...@redhat.com> wrote:
> Hi!
>
> While for the trunk I hope Michael will finalize a much better fix,
> this patch provides a quick workaround for 4.6 branch.
>
> In particular, I'd like to avoid reverting the
> http://gcc.gnu.org/ml/gcc-patches/2011-01/msg01442.html
> patch, because if GIMPLE passes don't do any significant code motion,
> sharing the stack slots e.g. for Fortran I/O parameter block helps quite a
> bit. So, what this patch attempts to do is (only for the cases which
> changed by the above mentioned patch) make the stack slot sharing
> conditional on no significant code motion (loop {,complete} unrolling and
> other
> kind of bb duplication).
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for 4.6 branch?
> Or is it way too ugly and should I rather revert there my patch?
Ugh ...
> 2011-06-07 Jakub Jelinek <ja...@redhat.com>
>
> PR fortran/49103
> * tree.h (DECL_NONSHAREABLE): Define.
> (struct tree_decl_common): Change decl_common_unused to
> decl_nonshareable_flag.
> * cfgexpand.c (expand_used_vars_for_block, clear_tree_used):
> Ignore vars with DECL_NONSHAREABLE bit set.
> * tree-cfg.c (gimple_duplicate_bb): Set DECL_NONSHAREABLE
> on stores to automatic aggregate vars.
>
> * gfortran.dg/pr49103.f90: New test.
>
> --- gcc/tree.h.jj 2011-03-14 14:12:15.000000000 +0100
> +++ gcc/tree.h 2011-05-31 14:05:34.000000000 +0200
> @@ -1330,6 +1330,10 @@ extern void omp_clause_range_check_faile
> #define DECL_READ_P(NODE) \
> (TREE_CHECK2 (NODE, VAR_DECL, PARM_DECL)->decl_common.decl_read_flag)
>
> +#define DECL_NONSHAREABLE(NODE) \
> + (TREE_CHECK2 (NODE, VAR_DECL, \
> + RESULT_DECL)->decl_common.decl_nonshareable_flag)
> +
> /* In a CALL_EXPR, means that the call is the jump from a thunk to the
> thunked-to function. */
> #define CALL_FROM_THUNK_P(NODE) (CALL_EXPR_CHECK (NODE)->base.protected_flag)
> @@ -2787,8 +2791,9 @@ struct GTY(()) tree_decl_common {
> being set. */
> unsigned decl_read_flag : 1;
>
> - /* Padding so that 'off_align' can be on a 32-bit boundary. */
> - unsigned decl_common_unused : 1;
> + /* In VAR_DECL or RESULT_DECL set when significant code movement precludes
> + attempting to share the stack slot with some other variable. */
> + unsigned decl_nonshareable_flag : 1;
>
> /* DECL_OFFSET_ALIGN, used only for FIELD_DECLs. */
> unsigned int off_align : 8;
> --- gcc/cfgexpand.c.jj 2011-05-04 10:46:52.000000000 +0200
> +++ gcc/cfgexpand.c 2011-05-31 14:08:36.000000000 +0200
> @@ -1134,7 +1134,9 @@ expand_used_vars_for_block (tree block,
>
> /* Expand all variables at this level. */
> for (t = BLOCK_VARS (block); t ; t = DECL_CHAIN (t))
> - if (TREE_USED (t))
> + if (TREE_USED (t)
> + && ((TREE_CODE (t) != VAR_DECL && TREE_CODE (t) != RESULT_DECL)
> + || !DECL_NONSHAREABLE (t)))
> expand_one_var (t, toplevel, true);
>
> this_sv_num = stack_vars_num;
> @@ -1167,6 +1169,8 @@ clear_tree_used (tree block)
>
> for (t = BLOCK_VARS (block); t ; t = DECL_CHAIN (t))
> /* if (!TREE_STATIC (t) && !DECL_EXTERNAL (t)) */
> + if ((TREE_CODE (t) != VAR_DECL && TREE_CODE (t) != RESULT_DECL)
> + || !DECL_NONSHAREABLE (t))
> TREE_USED (t) = 0;
>
> for (t = BLOCK_SUBBLOCKS (block); t ; t = BLOCK_CHAIN (t))
> --- gcc/tree-cfg.c.jj 2011-03-14 14:12:15.000000000 +0100
> +++ gcc/tree-cfg.c 2011-05-31 15:22:14.000000000 +0200
> @@ -5117,6 +5117,7 @@ gimple_duplicate_bb (basic_block bb)
> {
> def_operand_p def_p;
> ssa_op_iter op_iter;
> + tree lhs;
>
> stmt = gsi_stmt (gsi);
> if (gimple_code (stmt) == GIMPLE_LABEL)
> @@ -5130,6 +5131,29 @@ gimple_duplicate_bb (basic_block bb)
> maybe_duplicate_eh_stmt (copy, stmt);
> gimple_duplicate_stmt_histograms (cfun, copy, cfun, stmt);
>
> + /* When copying around a stmt writing into a local non-user
> + aggregate, make sure it won't share stack slot with other
> + vars. */
> + lhs = gimple_get_lhs (stmt);
> + if (lhs
> + && (handled_component_p (lhs)
> + || TREE_CODE (lhs) == MEM_REF
> + || TREE_CODE (lhs) == TARGET_MEM_REF
> + || TREE_CODE (lhs) == VAR_DECL
> + || TREE_CODE (lhs) == RESULT_DECL))
> + {
> + tree base = get_base_address (lhs);
Probably easier and more complete to do
if (lhs && TREE_CODE (lhs) != SSA_NAME)
{
tree base = get_base_address (lhs);
I don't like the patch too much, but it looks reasonable. At least reverting
your patch doesn't really fix anything.
Any opinions from others?
Thanks,
Richard.
> + if (base
> + && (TREE_CODE (base) == VAR_DECL
> + || TREE_CODE (base) == RESULT_DECL)
> + && DECL_IGNORED_P (base)
> + && !TREE_STATIC (base)
> + && !DECL_EXTERNAL (base)
> + && (TREE_CODE (base) != VAR_DECL
> + || !DECL_HAS_VALUE_EXPR_P (base)))
> + DECL_NONSHAREABLE (base) = 1;
> + }
> +
> /* Create new names for all the definitions created by COPY and
> add replacement mappings for each new name. */
> FOR_EACH_SSA_DEF_OPERAND (def_p, copy, op_iter, SSA_OP_ALL_DEFS)
> --- gcc/testsuite/gfortran.dg/pr49103.f90.jj 2011-05-31 13:52:43.000000000
> +0200
> +++ gcc/testsuite/gfortran.dg/pr49103.f90 2011-05-31 13:57:16.000000000
> +0200
> @@ -0,0 +1,19 @@
> +! PR fortran/49103
> +! { dg-do run }
> + integer :: a(2), b(2), i, j
> + open (10, status='scratch')
> + do j = 1, 2
> + a = (/ 0, 0 /)
> + b = (/ 1, 1 /)
> + do i = 1, 2
> + write (10, *) a
> + write (10, *) b
> + end do
> + end do
> + rewind (10)
> + do i = 0, 7
> + read (10, *) a
> + if (any (a .ne. mod (i, 2))) call abort
> + end do
> + close (10)
> +end
>
> Jakub
>
