| Message ID | CAO2gOZXeHsy1hTaF88gsn=SuVnrQ2aV6XfEkRigmU2YGH9frcw@mail.gmail.com |
|---|---|
| State | New |
| Headers | show |
On Sat, Sep 22, 2012 at 4:08 AM, Dehao Chen <dehao@google.com> wrote: > Hi, > > The problem is due to a bug when move_block_to_fn. edge->goto_block > should be updated even when its locus is unknown. This patch also > fixes the way to reset block for expr. I think the machinery is much more fragile (which is probably why when in SSA form the whole thing is invoked with a NULL_TREE BLOCK). From looking at how the block tree is mangled: /* Rewire BLOCK_SUBBLOCKS of orig_block. */ if (orig_block) { tree block; gcc_assert (BLOCK_SUBBLOCKS (DECL_INITIAL (dest_cfun->decl)) == NULL_TREE); BLOCK_SUBBLOCKS (DECL_INITIAL (dest_cfun->decl)) = BLOCK_SUBBLOCKS (orig_block); for (block = BLOCK_SUBBLOCKS (orig_block); block; block = BLOCK_CHAIN (block)) BLOCK_SUPERCONTEXT (block) = DECL_INITIAL (dest_cfun->decl); BLOCK_SUBBLOCKS (orig_block) = NULL_TREE; } I think we want a simpler version of your fix: > Bootstrapped and pass all gcc regression tests. > > Is it okay for trunk? > > Thanks, > Dehao > > gcc/ChangeLog: > tree-cfg.c (move_stmt_op): Reset the expr block only > when necessary. > (move_block_to_fn): Reset the edge's goto block even > when the goto locus is unknown. > > Index: tree-cfg.c > =================================================================== > --- tree-cfg.c (revision 191614) > +++ tree-cfg.c (working copy) > @@ -6013,7 +6013,9 @@ move_stmt_op (tree *tp, int *walk_subtrees, void * > > if (EXPR_P (t)) > { > - if (TREE_BLOCK (t)) > + if (p->orig_block == NULL_TREE > + || TREE_BLOCK (t) == p->orig_block > + || TREE_BLOCK (t) == NULL_TREE) > TREE_SET_BLOCK (t, p->new_block); Only replace TREE_BLOCK of t if it was equal to p->orig_block or if p->orig_block is NULL_TREE _and_ TREE_BLOCK of t is not NULL_TREE. > } > else if (DECL_P (t) || TREE_CODE (t) == SSA_NAME) > @@ -6315,7 +6317,7 @@ move_block_to_fn (struct function *dest_cfun, basi > } > > FOR_EACH_EDGE (e, ei, bb->succs) > - if (!IS_UNKNOWN_LOCATION (e->goto_locus)) > + if (e->goto_locus) Please compare against UNKNOWN_LOCATION (which is zero, right)? Ok with that changes. Thanks, Richard. > { > tree block = LOCATION_BLOCK (e->goto_locus); > if (d->orig_block == NULL_TREE
Index: tree-cfg.c =================================================================== --- tree-cfg.c (revision 191614) +++ tree-cfg.c (working copy) @@ -6013,7 +6013,9 @@ move_stmt_op (tree *tp, int *walk_subtrees, void * if (EXPR_P (t)) { - if (TREE_BLOCK (t)) + if (p->orig_block == NULL_TREE + || TREE_BLOCK (t) == p->orig_block + || TREE_BLOCK (t) == NULL_TREE) TREE_SET_BLOCK (t, p->new_block); } else if (DECL_P (t) || TREE_CODE (t) == SSA_NAME) @@ -6315,7 +6317,7 @@ move_block_to_fn (struct function *dest_cfun, basi } FOR_EACH_EDGE (e, ei, bb->succs) - if (!IS_UNKNOWN_LOCATION (e->goto_locus)) + if (e->goto_locus) { tree block = LOCATION_BLOCK (e->goto_locus); if (d->orig_block == NULL_TREE
Hi, The problem is due to a bug when move_block_to_fn. edge->goto_block should be updated even when its locus is unknown. This patch also fixes the way to reset block for expr. Bootstrapped and pass all gcc regression tests. Is it okay for trunk? Thanks, Dehao gcc/ChangeLog: tree-cfg.c (move_stmt_op): Reset the expr block only when necessary. (move_block_to_fn): Reset the edge's goto block even when the goto locus is unknown.