On Mon, 7 Jan 2013, Steven Bosscher wrote:
> Hello,
>
> As-is, tree-ssa-sink.c can only sink GIMPLE_ASSIGN statements. This
> patch lets tree-ssa-sink.c sink pure calls also.
>
> This allows the pass to sink the call to use in the following test
> case (new test case, ssa-sink-10.c to be):
>
> ---------------------- 8< ----------------
> /* { dg-do compile } */
> /* { dg-options "-O -fdump-tree-sink" } */
>
> __attribute__((__noinline__,__noclone__,__pure__,__const__))
> int use (int b)
> {
> return b * 2 + 4;
> }
>
> int foo (int b, int c, int d)
> {
> int res, r = 0;
> res = use (b);
> if (c)
> r = res;
> return r;
> }
>
> /* use(b) should be sunk below the if(c). */
> /* { dg-final { scan-tree-dump-times "Sinking.*use" 1 "sink" } } */
> /* { dg-final { cleanup-tree-dump "sink" } } */
> ---------------------- 8< ----------------
>
> so that the optimized code (.092t.sink dump) looks like this:
>
> ;; Function foo (foo, funcdef_no=1, decl_uid=2008, cgraph_uid=1)
> ...
> Sinking r_3 = use (b_2(D));
> from bb 2 to bb 3
> foo (int b, int c, int d)
> {
> int r;
> <bb 2>:
> if (c_4(D) != 0)
> goto <bb 3>;
> else
> goto <bb 5>;
> <bb 5>:
> goto <bb 4>;
> <bb 3>:
> r_3 = use (b_2(D));
> <bb 4>:
> # r_1 = PHI <0(5), r_3(3)>
> return r_1;
>
> }
>
> The patch allows the sinking of a number of functions in GCC itself
> (mostly simple predicates e.g. from gimple.h) to be sunk in an LTO
> bootstrap.
>
> The patch is a bit larger than necessary because I cleaned up the
> comments a bit. The only real changes are these three bits:
>
> /* We only can sink assignments. */
> - if (!is_gimple_assign (stmt))
> + stmt_lhs = gimple_get_lhs (stmt);
> + if (stmt_lhs == NULL_TREE)
> return false;
>
> and
>
> - && TYPE_MODE (TREE_TYPE (gimple_assign_lhs (stmt))) == BLKmode))
> + && TYPE_MODE (TREE_TYPE (stmt_lhs)) == BLKmode))
>
> and
>
> - && operand_equal_p (gimple_assign_lhs (stmt),
> - gimple_assign_lhs (use_stmt), 0))
> + && operand_equal_p (stmt_lhs, gimple_assign_lhs (use_stmt), 0))
> continue;
>
> Bootstrapped&tested on powerpc64-unknown-linux-gnu.
> OK for GCC 4.9 stage1?
Comments below
> Ciao!
> Steven
>
>
> gcc/
> * tree-ssa-sink.c (statement_sink_location): Handle pure calls.
>
> testsuite/
> * gcc.dg/tree-ssa/ssa-sink-10.c: New test.
>
> Index: tree-ssa-sink.c
> ===================================================================
> --- tree-ssa-sink.c (revision 194924)
> +++ tree-ssa-sink.c (working copy)
> @@ -264,56 +264,48 @@ statement_sink_location (gimple stmt, basic_block
> def_operand_p def_p;
> ssa_op_iter iter;
> imm_use_iterator imm_iter;
> + tree stmt_lhs;
>
> /* We only can sink assignments. */
> - if (!is_gimple_assign (stmt))
> + stmt_lhs = gimple_get_lhs (stmt);
> + if (stmt_lhs == NULL_TREE)
> return false;
Instead of this ...
> - /* We only can sink stmts with a single definition. */
> + /* We only can sink stmts with a single definition.
> + We don't want to sink dead code, so anything with 0 immediate uses
> + is not sunk. */
> def_p = single_ssa_def_operand (stmt, SSA_OP_ALL_DEFS);
> - if (def_p == NULL_DEF_OPERAND_P)
> + if (def_p == NULL_DEF_OPERAND_P
> + || has_zero_uses (DEF_FROM_PTR (def_p)))
> return false;
... do
stmt_lhs = DEF_FROM_PTR (def_p);
which then also handles sinking single-def GIMPLE_ASMs ;)
> - /* Return if there are no immediate uses of this stmt. */
> - if (has_zero_uses (DEF_FROM_PTR (def_p)))
> - return false;
> -
> /* There are a few classes of things we can't or don't move, some because
> we
> don't have code to handle it, some because it's not profitable and some
> - because it's not legal.
> + because it's not legal. */
>
> - We can't sink things that may be global stores, at least not without
> - calculating a lot more information, because we may cause it to no longer
> - be seen by an external routine that needs it depending on where it gets
> - moved to.
> -
> - We don't want to sink loads from memory.
> -
> - We can't sink statements that end basic blocks without splitting the
> - incoming edge for the sink location to place it there.
> -
> - We can't sink statements that have volatile operands.
> -
> - We don't want to sink dead code, so anything with 0 immediate uses is
> not
> - sunk.
> -
> - Don't sink BLKmode assignments if current function has any local
> explicit
> - register variables, as BLKmode assignments may involve memcpy or memset
> - calls or, on some targets, inline expansion thereof that sometimes need
> - to use specific hard registers.
> -
> - */
> - if (stmt_ends_bb_p (stmt)
> + if (/* We can't sink statements that end basic blocks without splitting
> + the incoming edge for the sink location to place it there. */
> + stmt_ends_bb_p (stmt)
> + /* We can't sink statements with side effects, e.g. statements that
> + have volatile operands, or non-pure calls. */
> || gimple_has_side_effects (stmt)
> - || gimple_has_volatile_ops (stmt)
> + /* We don't want to sink loads from memory.
> + We can't sink things that may be global stores, at least not
> + without calculating a lot more information, because we may cause
> + it to no longer be seen by an external routine that needs it
> + depending on where it gets moved to. */
> || (gimple_vuse (stmt) && !gimple_vdef (stmt))
> + /* Don't sink BLKmode assignments if current function has any local
> + explicit register variables, as BLKmode assignments may involve
> + memcpy or memset calls or, on some targets, inline expansion
> + thereof that sometimes need to use specific hard registers. */
> || (cfun->has_local_explicit_reg_vars
> - && TYPE_MODE (TREE_TYPE (gimple_assign_lhs (stmt))) == BLKmode))
> + && TYPE_MODE (TREE_TYPE (stmt_lhs)) == BLKmode))
> return false;
>
> + /* We can't sink across abnormal PHIs. */
> if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (DEF_FROM_PTR (def_p)))
> return false;
> -
> FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_ALL_USES)
> {
> tree use = USE_FROM_PTR (use_p);
> @@ -334,8 +326,7 @@ statement_sink_location (gimple stmt, basic_block
> /* A killing definition is not a use. */
> if (gimple_assign_single_p (use_stmt)
> && gimple_vdef (use_stmt)
> - && operand_equal_p (gimple_assign_lhs (stmt),
> - gimple_assign_lhs (use_stmt), 0))
> + && operand_equal_p (stmt_lhs, gimple_assign_lhs (use_stmt), 0))
I suppose this also needs to handle
stmt_lhs = call ();
or the equivalent asm. Without handling asms thus adjust it to
gimple_get_lhs as you did above, with handling asms instead use
walk_stmt_load_store_addr_ops to walk all stores in the stmt.
Thanks,
Richard.
