On Mon, Nov 22, 2021 at 9:40 AM Andrew Pinski <pins...@gmail.com> wrote:
>
> On Wed, Oct 27, 2021 at 3:42 AM Richard Biener via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
> >
> > On Wed, Oct 27, 2021 at 12:00 PM apinski--- via Gcc-patches
> > <gcc-patches@gcc.gnu.org> wrote:
> > >
> > > From: Andrew Pinski <apin...@marvell.com>
> > >
> > > The problem here is tree-ssa-forwprop.c likes to produce
> > > &MEM <const char *> [(void *)_4 + 152B] which is the same as
> > > _4 p+ 152 which the rest of GCC likes better.
> > > This implements this transformation back to pointer plus to
> > > improve better code generation later on.
> >
> > Why do you think so? Can you pin-point the transform that now
> > fixes the new testcase?
>
> So we had originally:
> language_names_p_9 = &MEM <const char *> [(void *)_4 + 24B];
> ...
> _2 = _4 + 40;
Of course if that would have been
_2 = &MEM [_4 + 40B];
the issue would be fixed as well. That said, I agree that _4 + 40
is better but I think we should canonicalize all &MEM[SSA + CST]
this way. There is a canonicalization phase in fold_stmt_1:
/* First do required canonicalization of [TARGET_]MEM_REF addresses
after propagation.
??? This shouldn't be done in generic folding but in the
propagation helpers which also know whether an address was
propagated.
Also canonicalize operand order. */
switch (gimple_code (stmt))
{
case GIMPLE_ASSIGN:
if (gimple_assign_rhs_class (stmt) == GIMPLE_SINGLE_RHS)
{
tree *rhs = gimple_assign_rhs1_ptr (stmt);
if ((REFERENCE_CLASS_P (*rhs)
|| TREE_CODE (*rhs) == ADDR_EXPR)
&& maybe_canonicalize_mem_ref_addr (rhs))
changed = true;
where this could be done (apart from adding a match.pd pattern for this).
> if (_2 != language_names_p_9)
>
> Forwprop is able to figure out that the above if statement is now
> always false as we have (_4 +p 40) != (_4 +p 24) which gets simplified
> via a match-and-simplify pattern ().
> Does that answer your question?
>
> I will look into the other comments in a new patch.
>
> Thanks,
> Andrew
>
> >
> > Comments below
> >
> > > OK? Bootstrapped and tested on aarch64-linux-gnu.
> > >
> > > Changes from v1:
> > > * v2: Add comments.
> > >
> > > gcc/ChangeLog:
> > >
> > > PR tree-optimization/102216
> > > * tree-ssa-forwprop.c (rewrite_assign_addr): New function.
> > > (forward_propagate_addr_expr_1): Use rewrite_assign_addr
> > > when rewriting into the addr_expr into an assignment.
> > >
> > > gcc/testsuite/ChangeLog:
> > >
> > > PR tree-optimization/102216
> > > * g++.dg/tree-ssa/pr102216.C: New test.
> > > ---
> > > gcc/testsuite/g++.dg/tree-ssa/pr102216.C | 22 +++++++++
> > > gcc/tree-ssa-forwprop.c | 58 ++++++++++++++++++------
> > > 2 files changed, 67 insertions(+), 13 deletions(-)
> > > create mode 100644 gcc/testsuite/g++.dg/tree-ssa/pr102216.C
> > >
> > > diff --git a/gcc/testsuite/g++.dg/tree-ssa/pr102216.C
> > > b/gcc/testsuite/g++.dg/tree-ssa/pr102216.C
> > > new file mode 100644
> > > index 00000000000..b903e4eb57d
> > > --- /dev/null
> > > +++ b/gcc/testsuite/g++.dg/tree-ssa/pr102216.C
> > > @@ -0,0 +1,22 @@
> > > +/* { dg-options "-O2 -fdump-tree-optimized" } */
> > > +void link_error ();
> > > +void g ()
> > > +{
> > > + const char **language_names;
> > > +
> > > + language_names = new const char *[6];
> > > +
> > > + const char **language_names_p = language_names;
> > > +
> > > + language_names_p++;
> > > + language_names_p++;
> > > + language_names_p++;
> > > +
> > > + if ( (language_names_p) - (language_names+3) != 0)
> > > + link_error();
> > > + delete[] language_names;
> > > +}
> > > +/* We should have removed the link_error on the gimple level as GCC
> > > should
> > > + be able to tell that language_names_p is the same as
> > > language_names+3. */
> > > +/* { dg-final { scan-tree-dump-times "link_error" 0 "optimized" } } */
> > > +
> > > diff --git a/gcc/tree-ssa-forwprop.c b/gcc/tree-ssa-forwprop.c
> > > index a830bab78ba..e4331c60525 100644
> > > --- a/gcc/tree-ssa-forwprop.c
> > > +++ b/gcc/tree-ssa-forwprop.c
> > > @@ -637,6 +637,47 @@ forward_propagate_into_cond (gimple_stmt_iterator
> > > *gsi_p)
> > > return 0;
> > > }
> > >
> > > +/* Rewrite the DEF_RHS as needed into the (plain) use statement. */
> > > +
> > > +static void
> > > +rewrite_assign_addr (gimple_stmt_iterator *use_stmt_gsi, tree def_rhs)
> > > +{
> > > + tree def_rhs_base;
> > > + poly_int64 def_rhs_offset;
> > > +
> > > + /* Get the base and offset. */
> > > + if ((def_rhs_base = get_addr_base_and_unit_offset (TREE_OPERAND
> > > (def_rhs, 0),
> > > + &def_rhs_offset)))
> >
> > So this will cause us to rewrite &MEM[p_1].a.b.c; to a pointer-plus,
> > right? Don't
> > we want to preserve that for object-size stuff? So maybe directly pattern
> > match ADDR_EXPR <MEM_REF <SSA_NAME, ..>> only?
> >
> > > + {
> > > + tree new_ptr;
> > > + poly_offset_int off = 0;
> > > +
> > > + /* If the base was a MEM, then add the offset to the other
> > > + offset and adjust the base. */
> > > + if (TREE_CODE (def_rhs_base) == MEM_REF)
> > > + {
> > > + off += mem_ref_offset (def_rhs_base);
> > > + new_ptr = TREE_OPERAND (def_rhs_base, 0);
> > > + }
> > > + else
> > > + new_ptr = build_fold_addr_expr (def_rhs_base);
> > > +
> > > + /* If we have the new base is not an address express, then use a
> > > p+ expression
> > > + as the new expression instead of &MEM[x, offset]. */
> > > + if (TREE_CODE (new_ptr) != ADDR_EXPR)
> > > + {
> > > + tree offset = wide_int_to_tree (sizetype, off);
> > > + def_rhs = build2 (POINTER_PLUS_EXPR, TREE_TYPE (def_rhs),
> > > new_ptr, offset);
> >
> > Ick. You should be able to use gimple_assign_set_rhs_with_ops.
> >
> > > + }
> > > + }
> > > +
> > > + /* Replace the rhs with the new expression. */
> > > + def_rhs = unshare_expr (def_rhs);
> >
> > and definitely no need to unshare anything here?
> >
> > > + gimple_assign_set_rhs_from_tree (use_stmt_gsi, def_rhs);
> > > + gimple *use_stmt = gsi_stmt (*use_stmt_gsi);
> > > + update_stmt (use_stmt);
> > > +}
> > > +
> > > /* We've just substituted an ADDR_EXPR into stmt. Update all the
> > > relevant data structures to match. */
> > >
> > > @@ -696,8 +737,8 @@ forward_propagate_addr_expr_1 (tree name, tree
> > > def_rhs,
> > > if (single_use_p
> > > && useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE
> > > (def_rhs)))
> > > {
> > > - gimple_assign_set_rhs1 (use_stmt, unshare_expr (def_rhs));
> > > - gimple_assign_set_rhs_code (use_stmt, TREE_CODE (def_rhs));
> > > + rewrite_assign_addr (use_stmt_gsi, def_rhs);
> > > + gcc_assert (gsi_stmt (*use_stmt_gsi) == use_stmt);
> > > return true;
> > > }
> > >
> > > @@ -741,14 +782,7 @@ forward_propagate_addr_expr_1 (tree name, tree
> > > def_rhs,
> > > if (forward_propagate_addr_expr (lhs, new_def_rhs, single_use_p))
> > > return true;
> > >
> > > - if (useless_type_conversion_p (TREE_TYPE (lhs),
> > > - TREE_TYPE (new_def_rhs)))
> > > - gimple_assign_set_rhs_with_ops (use_stmt_gsi, TREE_CODE
> > > (new_def_rhs),
> > > - new_def_rhs);
> > > - else if (is_gimple_min_invariant (new_def_rhs))
> > > - gimple_assign_set_rhs_with_ops (use_stmt_gsi, NOP_EXPR,
> > > new_def_rhs);
> > > - else
> > > - return false;
> > > + rewrite_assign_addr (use_stmt_gsi, new_def_rhs);
> > > gcc_assert (gsi_stmt (*use_stmt_gsi) == use_stmt);
> > > update_stmt (use_stmt);
> > > return true;
> > > @@ -951,9 +985,7 @@ forward_propagate_addr_expr_1 (tree name, tree
> > > def_rhs,
> > > unshare_expr (def_rhs),
> > > fold_convert (ptr_type_node,
> > > rhs2)));
> > > - gimple_assign_set_rhs_from_tree (use_stmt_gsi, new_rhs);
> > > - use_stmt = gsi_stmt (*use_stmt_gsi);
> > > - update_stmt (use_stmt);
> > > + rewrite_assign_addr (use_stmt_gsi, new_rhs);
> >
> > so you only do this after addr_expr forwarding but not on stmts in
> > general? You could
> > do it that way in the 2nd loop over the BB.
> >
> > > tidy_after_forward_propagate_addr (use_stmt);
> > > return true;
> > > }
> > > --
> > > 2.17.1
> > >
