| Message ID | 201012151308.56233.ebotcazou@adacore.com |
|---|---|
| State | New |
| Headers | show |
On Wed, Dec 15, 2010 at 1:08 PM, Eric Botcazou <ebotcazou@adacore.com> wrote: > Hi, > > this is a regression present on mainline/4.5 branch. The compiler generates > wrong code at -O and above (for platforms that still use SJLJ exceptions). > > The problem is FRE computing that: > > D.2449_11 = P8b_3->d; > > loads 0 after: > > P8b_3->r{off: D.2453_9 * 4} = {}; > > although the 'd' and 'r' components don't overlap. > > > /* 2) Assignment from an empty CONSTRUCTOR. */ > else if (is_gimple_reg_type (vr->type) > && gimple_assign_single_p (def_stmt) > && gimple_assign_rhs_code (def_stmt) == CONSTRUCTOR > && CONSTRUCTOR_NELTS (gimple_assign_rhs1 (def_stmt)) == 0) > { > tree base2; > HOST_WIDE_INT offset2, size2, maxsize2; > base2 = get_ref_base_and_extent (gimple_assign_lhs (def_stmt), > &offset2, &size2, &maxsize2); > if (operand_equal_p (base, base2, 0) > && offset2 <= offset > && offset2 + size2 >= offset + maxsize) > { > tree val = fold_convert (vr->type, integer_zero_node); > unsigned int value_id = get_or_alloc_constant_value_id (val); > return vn_reference_insert_pieces (vuse, vr->set, vr->type, > VEC_copy (vn_reference_op_s, > heap, vr->operands), > val, value_id); > } > } > > (gdb) p debug_gimple_stmt(def_stmt) > # .MEM_21 = VDEF <.MEM_24> > P8b_3->r{off: D.2453_9 * 4} = {}; > > (gdb) p debug_generic_expr(base2) > *P8b_3 > $46 = void > (gdb) p offset2 > $47 = 0 > (gdb) p size2 > $48 = 192 > (gdb) p maxsize2 > $49 = -1 > > > Tested on x86_64-suse-linux, OK for the mainline and 4.5 branch when reopened? Oops. Thanks for fixing. Yes, the patch is ok, with the test adjusted to maxsize2 !=/== -1 (which is the canonical form we use elsewhere). Thanks, Richard. > > 2010-12-15 Eric Botcazou <ebotcazou@adacore.com> > > * tree-ssa-sccvn.c (vn_reference_lookup_3): Always punt if the call to > get_ref_base_and_extent returns -1 as the max size. > > > 2010-12-15 Eric Botcazou <ebotcazou@adacore.com> > > * gnat.dg/opt13.adb: New test. > * gnat.dg/opt13_pkg.ad[sb]: New helper. > > > -- > Eric Botcazou >
Index: tree-ssa-sccvn.c =================================================================== --- tree-ssa-sccvn.c (revision 167721) +++ tree-ssa-sccvn.c (working copy) @@ -1339,6 +1339,7 @@ vn_reference_lookup_3 (ao_ref *ref, tree size2 = TREE_INT_CST_LOW (gimple_call_arg (def_stmt, 2)) * 8; if ((unsigned HOST_WIDE_INT)size2 / 8 == TREE_INT_CST_LOW (gimple_call_arg (def_stmt, 2)) + && maxsize2 >= 0 && operand_equal_p (base, base2, 0) && offset2 <= offset && offset2 + size2 >= offset + maxsize) @@ -1362,7 +1363,8 @@ vn_reference_lookup_3 (ao_ref *ref, tree HOST_WIDE_INT offset2, size2, maxsize2; base2 = get_ref_base_and_extent (gimple_assign_lhs (def_stmt), &offset2, &size2, &maxsize2); - if (operand_equal_p (base, base2, 0) + if (maxsize2 >= 0 + && operand_equal_p (base, base2, 0) && offset2 <= offset && offset2 + size2 >= offset + maxsize) { @@ -1383,7 +1385,7 @@ vn_reference_lookup_3 (ao_ref *ref, tree || handled_component_p (gimple_assign_rhs1 (def_stmt)))) { tree base2; - HOST_WIDE_INT offset2, size2; + HOST_WIDE_INT offset2, size2, maxsize2; int i, j; VEC (vn_reference_op_s, heap) *rhs = NULL; vn_reference_op_t vro; @@ -1396,8 +1398,9 @@ vn_reference_lookup_3 (ao_ref *ref, tree base2 = ao_ref_base (&lhs_ref); offset2 = lhs_ref.offset; size2 = lhs_ref.size; - if ((base != base2 - && !operand_equal_p (base, base2, 0)) + maxsize2 = lhs_ref.max_size; + if (maxsize2 < 0 + || (base != base2 && !operand_equal_p (base, base2, 0)) || offset2 > offset || offset2 + size2 < offset + maxsize) return (void *)-1;
Hi, this is a regression present on mainline/4.5 branch. The compiler generates wrong code at -O and above (for platforms that still use SJLJ exceptions). The problem is FRE computing that: D.2449_11 = P8b_3->d; loads 0 after: P8b_3->r{off: D.2453_9 * 4} = {}; although the 'd' and 'r' components don't overlap. /* 2) Assignment from an empty CONSTRUCTOR. */ else if (is_gimple_reg_type (vr->type) && gimple_assign_single_p (def_stmt) && gimple_assign_rhs_code (def_stmt) == CONSTRUCTOR && CONSTRUCTOR_NELTS (gimple_assign_rhs1 (def_stmt)) == 0) { tree base2; HOST_WIDE_INT offset2, size2, maxsize2; base2 = get_ref_base_and_extent (gimple_assign_lhs (def_stmt), &offset2, &size2, &maxsize2); if (operand_equal_p (base, base2, 0) && offset2 <= offset && offset2 + size2 >= offset + maxsize) { tree val = fold_convert (vr->type, integer_zero_node); unsigned int value_id = get_or_alloc_constant_value_id (val); return vn_reference_insert_pieces (vuse, vr->set, vr->type, VEC_copy (vn_reference_op_s, heap, vr->operands), val, value_id); } } (gdb) p debug_gimple_stmt(def_stmt) # .MEM_21 = VDEF <.MEM_24> P8b_3->r{off: D.2453_9 * 4} = {}; (gdb) p debug_generic_expr(base2) *P8b_3 $46 = void (gdb) p offset2 $47 = 0 (gdb) p size2 $48 = 192 (gdb) p maxsize2 $49 = -1 Tested on x86_64-suse-linux, OK for the mainline and 4.5 branch when reopened? 2010-12-15 Eric Botcazou <ebotcazou@adacore.com> * tree-ssa-sccvn.c (vn_reference_lookup_3): Always punt if the call to get_ref_base_and_extent returns -1 as the max size. 2010-12-15 Eric Botcazou <ebotcazou@adacore.com> * gnat.dg/opt13.adb: New test. * gnat.dg/opt13_pkg.ad[sb]: New helper.