On Thu, 21 Jun 2018, Richard Biener wrote:
>
> PR86223 points out that we currently gimplify the testcase inconsistently.
> For the incomplete CTORs we use block-clearing while for the complete
> one we emit initializations of the individual elements (up to the
> limits imposed in following checks).
>
> So the following makes us always use = {}; form for all-zero CTORs
> which is most compact for GIMPLE IL and should only result in better
> code (fingers crossing...) since but not only beacause SRA got the
> ability to handle a = .LC0; style inits as well.
>
> Bootstrap & regtest running on x86_64-unknown-linux-gnu.
Fine apart from
FAIL: g++.dg/tm/pr45940-3.C -std=gnu++11 (internal compiler error)
FAIL: g++.dg/tm/pr45940-3.C -std=gnu++11 (test for excess errors)
FAIL: g++.dg/tm/pr45940-3.C -std=gnu++14 (internal compiler error)
FAIL: g++.dg/tm/pr45940-3.C -std=gnu++14 (test for excess errors)
FAIL: g++.dg/tm/pr45940-3.C -std=gnu++98 (internal compiler error)
FAIL: g++.dg/tm/pr45940-3.C -std=gnu++98 (test for excess errors)
FAIL: g++.dg/tm/pr45940-4.C -std=gnu++11 (internal compiler error)
FAIL: g++.dg/tm/pr45940-4.C -std=gnu++11 (test for excess errors)
FAIL: g++.dg/tm/pr45940-4.C -std=gnu++14 (internal compiler error)
FAIL: g++.dg/tm/pr45940-4.C -std=gnu++14 (test for excess errors)
FAIL: g++.dg/tm/pr45940-4.C -std=gnu++98 (internal compiler error)
FAIL: g++.dg/tm/pr45940-4.C -std=gnu++98 (test for excess errors)
during GIMPLE pass: tmmark^M
In constructor 'shared_count::shared_count() transaction_safe':^M
cc1plus: internal compiler error: in create_tmp_var, at
gimple-expr.c:479^M
0x6c8dc7 create_tmp_var(tree_node*, char const*)^M
/space/rguenther/src/svn/trunk/gcc/gimple-expr.c:479^M
0xc5375b create_tmp_from_val^M
/space/rguenther/src/svn/trunk/gcc/gimplify.c:516^M
0xc5375b lookup_tmp_var^M
/space/rguenther/src/svn/trunk/gcc/gimplify.c:537^M
0xc5375b internal_get_tmp_var^M
/space/rguenther/src/svn/trunk/gcc/gimplify.c:590^M
0xc4d481 gimplify_expr(tree_node**, gimple**, gimple**, bool
(*)(tree_node*), int)^M
/space/rguenther/src/svn/trunk/gcc/gimplify.c:12386^M
0xc572d4 gimplify_addr_expr^M
/space/rguenther/src/svn/trunk/gcc/gimplify.c:5994^M
0xc4f82d gimplify_expr(tree_node**, gimple**, gimple**, bool
(*)(tree_node*), int)^M
/space/rguenther/src/svn/trunk/gcc/gimplify.c:11485^M
0xc5f4cc force_gimple_operand_1(tree_node*, gimple**, bool
(*)(tree_node*), tree_node*)^M
/space/rguenther/src/svn/trunk/gcc/gimplify-me.c:78^M
0xc5f57f force_gimple_operand_gsi_1(gimple_stmt_iterator*, tree_node*,
bool (*)(tree_node*), tree_node*, bool, gsi_iterator_update)^M
/space/rguenther/src/svn/trunk/gcc/gimplify-me.c:115^M
0xe9ac6a expand_assign_tm^M
/space/rguenther/src/svn/trunk/gcc/trans-mem.c:2446^M
0xe9df02 expand_block_tm^M
/space/rguenther/src/svn/trunk/gcc/trans-mem.c:2646^M
0xe9df02 execute_tm_mark^M
/space/rguenther/src/svn/trunk/gcc/trans-mem.c:3130^M
0xe9df02 execute^M
/space/rguenther/src/svn/trunk/gcc/trans-mem.c:3175^M
we're doing gimplify_addr (gsi, rhs); on a {} RHS. Looks like TM
doesn't handle stores from {} at all... is there a TM-safe
memset()?
and
FAIL: gnat.dg/opt34.adb scan-tree-dump esra "Created a replacement for
result"
no time to investigate right now, so I'm putting this on hold.
Eric, can you see if the opt34.adb FAIL is "harmless"?
Richard.
> Richard.
>
> 2018-06-21 Richard Biener <rguent...@suse.de>
>
> PR middle-end/86223
> * gimplify.c (gimplify_init_constructor): For an all-zero
> constructor emit a block-clear.
>
> * gcc.dg/pr86223-1.c: New testcase.
>
> Index: gcc/gimplify.c
> ===================================================================
> --- gcc/gimplify.c (revision 261839)
> +++ gcc/gimplify.c (working copy)
> @@ -4805,6 +4805,10 @@ gimplify_init_constructor (tree *expr_p,
> requires trickery to avoid quadratic compile-time behavior in
> large cases or excessive memory use in small cases. */
> cleared = !CONSTRUCTOR_NO_CLEARING (ctor);
> + else if (num_nonzero_elements == 0)
> + /* If all elements are zero it is most efficient to block-clear
> + things. */
> + cleared = true;
> else if (num_ctor_elements - num_nonzero_elements
> > CLEAR_RATIO (optimize_function_for_speed_p (cfun))
> && num_nonzero_elements < num_ctor_elements / 4)
> Index: gcc/testsuite/gcc.dg/pr86223-1.c
> ===================================================================
> --- gcc/testsuite/gcc.dg/pr86223-1.c (revision 0)
> +++ gcc/testsuite/gcc.dg/pr86223-1.c (working copy)
> @@ -0,0 +1,17 @@
> +/* { dg-do compile } */
> +/* { dg-options "-fdump-tree-gimple" } */
> +
> +void f (int *);
> +void g ()
> +{
> + int a[3] = { 0, 0, 0 };
> + f (a);
> +}
> +void h ()
> +{
> + int a[3] = { 0 };
> + f (a);
> +}
> +
> +/* We should use block-clearing for the initializer of a in both cases. */
> +/* { dg-final { scan-tree-dump-times "a = {};" 2 "gimple" } } */
>
--
Richard Biener <rguent...@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB
21284 (AG Nuernberg)
