Re: [PATCH, v2] Fortran: fix for absent array argument passed to optional dummy [PR101135]
Le 17/03/2024 à 23:10, Harald Anlauf a écrit : Hi Mikael, On 3/17/24 22:04, Mikael Morin wrote: diff --git a/gcc/fortran/trans-array.cc b/gcc/fortran/trans-array.cc index 3673fa40720..a7717a8107e 100644 --- a/gcc/fortran/trans-array.cc +++ b/gcc/fortran/trans-array.cc @@ -7526,6 +7526,17 @@ gfc_get_dataptr_offset (stmtblock_t *block, tree parm, tree desc, tree offset, /* Set the target data pointer. */ offset = gfc_build_addr_expr (gfc_array_dataptr_type (desc), tmp); + + /* Check for optional dummy argument being present. Arguments of BIND(C) + procedures are excepted here since they are handled differently. */ + if (expr->expr_type == EXPR_VARIABLE + && expr->symtree->n.sym->attr.dummy + && expr->symtree->n.sym->attr.optional + && !is_CFI_desc (NULL, expr)) I think the condition could additionally check the lack of subreferences. But it's maybe not worth the trouble, and the patch is conservatively correct as is, so OK. I have thought about the conditions here for some time and did not find better ones. They need to be broad enough to catch the case in gfortran.dg/missing_optional_dummy_6a.f90 that (according to the tree-dump) was not properly handled previously and would have triggered ubsan at some point in the future when someone tried to change that testcase from currently dg-do compile to dg-do run... No problem, as said it is conservatively correct. (After the patch it would pass, but I didn't dare to change the dg-do). Did it include cases not covered by the new testcase (which was quite complete already)? I have pushed the patch as-is, but feel free to post testcases not covered (or improperly covered) to narrow this down further... The case I had in mind would only be a missed optimization, and probably not that important, so let's move on.
Re: [PATCH, v2] Fortran: fix for absent array argument passed to optional dummy [PR101135]
Hi Mikael, On 3/17/24 22:04, Mikael Morin wrote: diff --git a/gcc/fortran/trans-array.cc b/gcc/fortran/trans-array.cc index 3673fa40720..a7717a8107e 100644 --- a/gcc/fortran/trans-array.cc +++ b/gcc/fortran/trans-array.cc @@ -7526,6 +7526,17 @@ gfc_get_dataptr_offset (stmtblock_t *block, tree parm, tree desc, tree offset, /* Set the target data pointer. */ offset = gfc_build_addr_expr (gfc_array_dataptr_type (desc), tmp); + + /* Check for optional dummy argument being present. Arguments of BIND(C) + procedures are excepted here since they are handled differently. */ + if (expr->expr_type == EXPR_VARIABLE + && expr->symtree->n.sym->attr.dummy + && expr->symtree->n.sym->attr.optional + && !is_CFI_desc (NULL, expr)) I think the condition could additionally check the lack of subreferences. But it's maybe not worth the trouble, and the patch is conservatively correct as is, so OK. I have thought about the conditions here for some time and did not find better ones. They need to be broad enough to catch the case in gfortran.dg/missing_optional_dummy_6a.f90 that (according to the tree-dump) was not properly handled previously and would have triggered ubsan at some point in the future when someone tried to change that testcase from currently dg-do compile to dg-do run... (After the patch it would pass, but I didn't dare to change the dg-do). I have pushed the patch as-is, but feel free to post testcases not covered (or improperly covered) to narrow this down further... Thanks for the patch. Thanks for the review! Harald + offset = build3_loc (input_location, COND_EXPR, TREE_TYPE (offset), + gfc_conv_expr_present (expr->symtree->n.sym), offset, + fold_convert (TREE_TYPE (offset), gfc_index_zero_node)); + gfc_conv_descriptor_data_set (block, parm, offset); }
Re: [PATCH, v2] Fortran: fix for absent array argument passed to optional dummy [PR101135]
Le 15/03/2024 à 20:32, Harald Anlauf a écrit : Dear all, as there has been some good progress in the handling of optional dummy arguments, I looked again at this PR and a patch for it that I withdrew as it turned out incomplete. It turned out that it now needs only a minor adjustment for optional dummy arguments of procedures with bind(c) attribute so that ubsan checking does not trigger. Along this way I extended the previous testcase to exercise to some extent combinations of bind(c) and non-bind(c) procedures and found one failure (since at least gcc-9) that is genuine: passing a missing optional from a bind(c) procedure to an assumed-rank dummy, see PR114355. The corresponding test is commented in the testcase. Regtested on x86_64-pc-linux-gnu. OK for mainline? Thanks, Harald (...) From b3079a82a220477704f8156207239e4af4103ea9 Mon Sep 17 00:00:00 2001 From: Harald Anlauf Date: Fri, 15 Mar 2024 20:14:07 +0100 Subject: [PATCH] Fortran: fix for absent array argument passed to optional dummy [PR101135] gcc/fortran/ChangeLog: PR fortran/101135 * trans-array.cc (gfc_get_dataptr_offset): Check for optional arguments being present before dereferencing data pointer. gcc/testsuite/ChangeLog: PR fortran/101135 * gfortran.dg/missing_optional_dummy_6a.f90: Adjust diagnostic pattern. * gfortran.dg/ubsan/missing_optional_dummy_8.f90: New test. --- gcc/fortran/trans-array.cc| 11 ++ .../gfortran.dg/missing_optional_dummy_6a.f90 | 2 +- .../ubsan/missing_optional_dummy_8.f90| 108 ++ 3 files changed, 120 insertions(+), 1 deletion(-) create mode 100644 gcc/testsuite/gfortran.dg/ubsan/missing_optional_dummy_8.f90 diff --git a/gcc/fortran/trans-array.cc b/gcc/fortran/trans-array.cc index 3673fa40720..a7717a8107e 100644 --- a/gcc/fortran/trans-array.cc +++ b/gcc/fortran/trans-array.cc @@ -7526,6 +7526,17 @@ gfc_get_dataptr_offset (stmtblock_t *block, tree parm, tree desc, tree offset, /* Set the target data pointer. */ offset = gfc_build_addr_expr (gfc_array_dataptr_type (desc), tmp); + + /* Check for optional dummy argument being present. Arguments of BIND(C) + procedures are excepted here since they are handled differently. */ + if (expr->expr_type == EXPR_VARIABLE + && expr->symtree->n.sym->attr.dummy + && expr->symtree->n.sym->attr.optional + && !is_CFI_desc (NULL, expr)) I think the condition could additionally check the lack of subreferences. But it's maybe not worth the trouble, and the patch is conservatively correct as is, so OK. Thanks for the patch. +offset = build3_loc (input_location, COND_EXPR, TREE_TYPE (offset), +gfc_conv_expr_present (expr->symtree->n.sym), offset, +fold_convert (TREE_TYPE (offset), gfc_index_zero_node)); + gfc_conv_descriptor_data_set (block, parm, offset); }
[PATCH, v2] Fortran: fix for absent array argument passed to optional dummy [PR101135]
Dear all, as there has been some good progress in the handling of optional dummy arguments, I looked again at this PR and a patch for it that I withdrew as it turned out incomplete. It turned out that it now needs only a minor adjustment for optional dummy arguments of procedures with bind(c) attribute so that ubsan checking does not trigger. Along this way I extended the previous testcase to exercise to some extent combinations of bind(c) and non-bind(c) procedures and found one failure (since at least gcc-9) that is genuine: passing a missing optional from a bind(c) procedure to an assumed-rank dummy, see PR114355. The corresponding test is commented in the testcase. Regtested on x86_64-pc-linux-gnu. OK for mainline? Thanks, Harald On 2/6/22 22:04, Harald Anlauf wrote: Hi Mikael, Am 04.02.22 um 11:45 schrieb Mikael Morin: Hello, Le 29/01/2022 à 22:41, Harald Anlauf via Fortran a écrit : The least invasive change - already pointed out by the reporter - is to check the presence of the argument before dereferencing the data pointer after the offset calculation. This requires adjusting the checking pattern for gfortran.dg/missing_optional_dummy_6a.f90. Regtesting reminded me that procedures with bind(c) attribute are doing their own stuff, which is why they need to be excluded here, otherwise testcase bind-c-contiguous-4.f90 would regress on the expected output. only after submitting the patch I figured that the patch is incomplete. When we have a call chain of procedures with and without bind(c), there are still cases left where the failure with the sanitizer is not fixed. Just add "bind(c)" to subroutine test_wrapper only in the original PR. I have added a corresponding comment in the PR. There is a potential alternative solution which I did not pursue, as I think it is more invasive, but also that I didn't succeed to implement: A non-present dummy array argument should not need to get its descriptor set up. Pursuing this is probably not the right thing to do during the current stage of development and could be implemented later. If somebody believes this is important, feel free to open a PR for this. I have an other (equally unimportant) concern that it may create an unnecessary conditional when passing a subobject of an optional argument. In that case we can assume that the optional is present. It’s not a correctness issue, so let’s not bother at this stage. Judging from the dump tree of the cases I looked at I did not see anything that would pose a problem to the optimizer. Regtested on x86_64-pc-linux-gnu. OK for mainline? OK. Given my latest observations I'd rather withdraw the current version of the patch and rethink. I also did not see an issue with bind(c) procedures calling alikes. It would help if one would not only know the properties of the actual argument, but also of the formal one, which is not available at that point in the code. I'll have another look and resubmit. Thanks. Thanks for the review! Harald From b3079a82a220477704f8156207239e4af4103ea9 Mon Sep 17 00:00:00 2001 From: Harald Anlauf Date: Fri, 15 Mar 2024 20:14:07 +0100 Subject: [PATCH] Fortran: fix for absent array argument passed to optional dummy [PR101135] gcc/fortran/ChangeLog: PR fortran/101135 * trans-array.cc (gfc_get_dataptr_offset): Check for optional arguments being present before dereferencing data pointer. gcc/testsuite/ChangeLog: PR fortran/101135 * gfortran.dg/missing_optional_dummy_6a.f90: Adjust diagnostic pattern. * gfortran.dg/ubsan/missing_optional_dummy_8.f90: New test. --- gcc/fortran/trans-array.cc| 11 ++ .../gfortran.dg/missing_optional_dummy_6a.f90 | 2 +- .../ubsan/missing_optional_dummy_8.f90| 108 ++ 3 files changed, 120 insertions(+), 1 deletion(-) create mode 100644 gcc/testsuite/gfortran.dg/ubsan/missing_optional_dummy_8.f90 diff --git a/gcc/fortran/trans-array.cc b/gcc/fortran/trans-array.cc index 3673fa40720..a7717a8107e 100644 --- a/gcc/fortran/trans-array.cc +++ b/gcc/fortran/trans-array.cc @@ -7526,6 +7526,17 @@ gfc_get_dataptr_offset (stmtblock_t *block, tree parm, tree desc, tree offset, /* Set the target data pointer. */ offset = gfc_build_addr_expr (gfc_array_dataptr_type (desc), tmp); + + /* Check for optional dummy argument being present. Arguments of BIND(C) + procedures are excepted here since they are handled differently. */ + if (expr->expr_type == EXPR_VARIABLE + && expr->symtree->n.sym->attr.dummy + && expr->symtree->n.sym->attr.optional + && !is_CFI_desc (NULL, expr)) +offset = build3_loc (input_location, COND_EXPR, TREE_TYPE (offset), + gfc_conv_expr_present (expr->symtree->n.sym), offset, + fold_convert (TREE_TYPE (offset), gfc_index_zero_node)); + gfc_conv_descriptor_data_set (block, parm, offset); } diff --git a/gcc/testsuite/gfortran.dg/missing_optional_dummy_6a.f90
