| Message ID | 20150408100239.GO19273@tucnak.redhat.com |
|---|---|
| State | New |
| Headers | show |
On 04/08/2015 06:02 AM, Jakub Jelinek wrote: > (cp_build_qualified_type_real): Use check_base_type. Build a > variant and copy over even TYPE_CONTEXT and > TYPE_ALIGN/TYPE_USER_ALIGN if any of those are different. This seems wrong. If there is an array with the same name, attributes and element type, it should have the same alignment; if it doesn't, that probably means that one of the types hasn't been laid out yet. We don't want to have two variants of the same array that are distinguished only by whether they've been laid out, especially since later probably both will be laid out and the two types will be the same. Jason
On Wed, Apr 08, 2015 at 10:47:15AM -0400, Jason Merrill wrote: > On 04/08/2015 06:02 AM, Jakub Jelinek wrote: > > (cp_build_qualified_type_real): Use check_base_type. Build a > > variant and copy over even TYPE_CONTEXT and > > TYPE_ALIGN/TYPE_USER_ALIGN if any of those are different. > > This seems wrong. If there is an array with the same name, attributes and > element type, it should have the same alignment; if it doesn't, that > probably means that one of the types hasn't been laid out yet. We don't > want to have two variants of the same array that are distinguished only by > whether they've been laid out, especially since later probably both will be > laid out and the two types will be the same. But that is how handle_aligned_attribute works, since forever (checked it back to 3.2). In <= 3.4.x, it used to create it using build_type_copy, since 4.0.0 using build_variant_type_copy, but both those routines behave the same - build a type variant which is linked in the TYPE_NEXT_VARIANT chain, and differs from the other type in there possibly just by TYPE_ALIGN/TYPE_USER_ALIGN. Perhaps it should check TYPE_ALIGN only if at least one of the two types has TYPE_USER_ALIGN set? As for why TYPE_ATTRIBUTES are NULL, the reason for that is that these are attributes on a typedef, so the attributes go into DECL_ATTRIBUTES of the TYPE_DECL instead. Anyway, the P1 regression is just about the first hunk, so if you have issues just with the second hunk and not the first hunk (from either of the patches), I can just comment out the tests for alignof (const T), and open a separate PR for that for later. Jakub
> On 04/08/2015 06:02 AM, Jakub Jelinek wrote: > > (cp_build_qualified_type_real): Use check_base_type. Build a > > variant and copy over even TYPE_CONTEXT and > > TYPE_ALIGN/TYPE_USER_ALIGN if any of those are different. > > This seems wrong. If there is an array with the same name, > attributes and element type, it should have the same alignment; if One of problems is that cp_build_qualified_type rebuilds the array from scratch and never copies the attribute list around (as oposed to build_qualified_type that just memcpy the type node) Honza > it doesn't, that probably means that one of the types hasn't been > laid out yet. We don't want to have two variants of the same array > that are distinguished only by whether they've been laid out, > especially since later probably both will be laid out and the two > types will be the same. > > Jason
On Wed, Apr 08, 2015 at 06:22:10PM +0200, Jan Hubicka wrote: > > On 04/08/2015 06:02 AM, Jakub Jelinek wrote: > > > (cp_build_qualified_type_real): Use check_base_type. Build a > > > variant and copy over even TYPE_CONTEXT and > > > TYPE_ALIGN/TYPE_USER_ALIGN if any of those are different. > > > > This seems wrong. If there is an array with the same name, > > attributes and element type, it should have the same alignment; if > > One of problems is that cp_build_qualified_type rebuilds the array from > scratch and never copies the attribute list around (as oposed to > build_qualified_type that just memcpy the type node) As I said earlier, TYPE_ATTRIBUTES is NULL here anyway, because the attributes hang in DECL_ATTRIBUTES of TYPE_DECL. And, except for config/sol2.c (which looks wrong), nothing ever calls lookup_attribute for "aligned" anyway, the user aligned stuff is encoded in TYPE_USER_ALIGN and/or DECL_USER_ALIGN and TYPE_ALIGN/DECL_ALIGN. Jakub
> On Wed, Apr 08, 2015 at 06:22:10PM +0200, Jan Hubicka wrote: > > > On 04/08/2015 06:02 AM, Jakub Jelinek wrote: > > > > (cp_build_qualified_type_real): Use check_base_type. Build a > > > > variant and copy over even TYPE_CONTEXT and > > > > TYPE_ALIGN/TYPE_USER_ALIGN if any of those are different. > > > > > > This seems wrong. If there is an array with the same name, > > > attributes and element type, it should have the same alignment; if > > > > One of problems is that cp_build_qualified_type rebuilds the array from > > scratch and never copies the attribute list around (as oposed to > > build_qualified_type that just memcpy the type node) > > As I said earlier, TYPE_ATTRIBUTES is NULL here anyway, because the > attributes hang in DECL_ATTRIBUTES of TYPE_DECL. And, except for > config/sol2.c (which looks wrong), nothing ever calls lookup_attribute for > "aligned" anyway, the user aligned stuff is encoded in TYPE_USER_ALIGN > and/or DECL_USER_ALIGN and TYPE_ALIGN/DECL_ALIGN. This is interesting too. I did know that alignment is "lowered" into TYPE_USER_ALIGN/TYPE_ALIGN values, but there is a lot of other code that looks for type attributes by searching TYPE_ATTRIBUTES, not DECL_ATTRIBUTES of TYPE_DECL (such as nonnul_arg_p in tree-vrp) or alloc_object_size. Does it mean that those attributes are ignored for C++ produced types? Honza > > Jakub
On Wed, Apr 08, 2015 at 06:32:29PM +0200, Jan Hubicka wrote: > > On Wed, Apr 08, 2015 at 06:22:10PM +0200, Jan Hubicka wrote: > > > > On 04/08/2015 06:02 AM, Jakub Jelinek wrote: > > > > > (cp_build_qualified_type_real): Use check_base_type. Build a > > > > > variant and copy over even TYPE_CONTEXT and > > > > > TYPE_ALIGN/TYPE_USER_ALIGN if any of those are different. > > > > > > > > This seems wrong. If there is an array with the same name, > > > > attributes and element type, it should have the same alignment; if > > > > > > One of problems is that cp_build_qualified_type rebuilds the array from > > > scratch and never copies the attribute list around (as oposed to > > > build_qualified_type that just memcpy the type node) > > > > As I said earlier, TYPE_ATTRIBUTES is NULL here anyway, because the > > attributes hang in DECL_ATTRIBUTES of TYPE_DECL. And, except for > > config/sol2.c (which looks wrong), nothing ever calls lookup_attribute for > > "aligned" anyway, the user aligned stuff is encoded in TYPE_USER_ALIGN > > and/or DECL_USER_ALIGN and TYPE_ALIGN/DECL_ALIGN. > > This is interesting too. I did know that alignment is "lowered" into > TYPE_USER_ALIGN/TYPE_ALIGN values, but there is a lot of other code > that looks for type attributes by searching TYPE_ATTRIBUTES, not DECL_ATTRIBUTES > of TYPE_DECL (such as nonnul_arg_p in tree-vrp) or alloc_object_size. > Does it mean that those attributes are ignored for C++ produced types? I don't think this has anything to do with C++. c-common.c has an attribute table, for each attribute it has 3 flags, whether a decl is required, type is required and/or fn type is required, and that determines to what the attributes go. These flags have the following combinations (decl/type/fntype # of attributes): TFF 51 (i.e. decl required) FTT 12 (i.e. function type required) FTF 9 (i.e. some type required) FFF 7 (applies to both types and decls) Which means most of the attributes require a decl and thus go into DECL_ATTRIBUTES, then some require function types and go to the TYPE_ATTRIBUTES of the function, then others go solely to TYPE_ATTRIBUTES. The last row are attributes that don't really care where they apply to, and that is { "packed", 0, 0, false, false, false, { "unused", 0, 0, false, false, false, { "transparent_union", 0, 0, false, false, false, { "aligned", 0, 1, false, false, false, { "deprecated", 0, 1, false, false, false, { "visibility", 1, 1, false, false, false, { "warn_unused", 0, 0, false, false, false, where the first 6 really don't care about what is stored in {TYPE,DECL}_ATTRIBUTES because the attributes are encoded differently in generic, and the last one sounds like a mistake (perhaps one that can't be undone anymore) where it doesn't require a type, but only stores it on types and warns otherwise. Jakub
On Wed, Apr 08, 2015 at 12:02:39PM +0200, Jakub Jelinek wrote: > The included (first) patch has been successfully bootstrapped/regtested on > x86_64-linux and i686-linux, the attached patch not, but I can > bootstrap/regtest it if you prefer it. The attached version unfortunately regresses pr60226.c typedef int __attribute__ ((aligned (1 << 28))) int28; int28 foo[4] = {}; /* { dg-error "alignment of array elements is greater than element size" } */ error is not reported then. But the first version with the layout_type call in there works. > @@ -1057,21 +1064,23 @@ cp_build_qualified_type_real (tree type, > should be equivalent to those in check_qualified_type. */ > for (t = TYPE_MAIN_VARIANT (type); t; t = TYPE_NEXT_VARIANT (t)) > if (TREE_TYPE (t) == element_type > - && TYPE_NAME (t) == TYPE_NAME (type) > - && TYPE_CONTEXT (t) == TYPE_CONTEXT (type) > - && attribute_list_equal (TYPE_ATTRIBUTES (t), > - TYPE_ATTRIBUTES (type))) > + && check_base_type (t, type)) So, would instead of this change adding && TYPE_USER_ALIGN (t) == TYPE_USER_ALIGN (type) && (!TYPE_USER_ALIGN (t) || TYPE_ALIGN (t) == TYPE_ALIGN (type)) to the conditional work for you? Or supposedly no change at all, as the attributes would be either in TYPE_ATTRIBUTES, or on TYPE_DECL, but then it would be covered by the TYPE_NAME comparison. > break; > > if (!t) > { > t = build_cplus_array_type (element_type, TYPE_DOMAIN (type)); > > - /* Keep the typedef name. */ > - if (TYPE_NAME (t) != TYPE_NAME (type)) > + /* Keep the typedef name, context and alignment. */ > + if (TYPE_NAME (t) != TYPE_NAME (type) > + || TYPE_CONTEXT (t) != TYPE_CONTEXT (type) > + || TYPE_ALIGN (t) != TYPE_ALIGN (type)) Then supposedly similar change to the above one (the TYPE_USER_ALIGN stuff). Jakub
On 04/08/2015 11:07 AM, Jakub Jelinek wrote: > But that is how handle_aligned_attribute works, since forever (checked it > back to 3.2). In <= 3.4.x, it used to create it using build_type_copy, > since 4.0.0 using build_variant_type_copy, but both those routines behave > the same - build a type variant which is linked in the TYPE_NEXT_VARIANT > chain, and differs from the other type in there possibly just by > TYPE_ALIGN/TYPE_USER_ALIGN. Perhaps it should check TYPE_ALIGN only if > at least one of the two types has TYPE_USER_ALIGN set? > As for why TYPE_ATTRIBUTES are NULL, the reason for that is that > these are attributes on a typedef, so the attributes go into DECL_ATTRIBUTES > of the TYPE_DECL instead. I see. So yes, I think we want to look at TYPE_USER_ALIGN as well. > Anyway, the P1 regression is just about the first hunk, so if you have > issues just with the second hunk and not the first hunk (from either of the > patches), I can just comment out the tests for alignof (const T), and open > a separate PR for that for later. The second patch seems to cause a regression on c-c++-common/pr60226.c because we avoid calling layout_type, so we lose the over-alignment error. So the first hunk of the first patch is OK. Incidentally, for template testcases it seems we also need to fix layout_type and fixup_type_sizes to respect already-set TYPE_USER_ALIGN: template <class T> struct A { T t; }; typedef A<int> T[4] alignas (2 * alignof (int)); A<int> a[4]; typedef A<char> T2[4] alignas (2 * alignof (int)); #define SA(X) static_assert((X),#X) SA(alignof (T) == 2 * alignof(int)); SA(alignof (T2) == 2 * alignof(int)); Jason
On 04/09/2015 05:38 AM, Jakub Jelinek wrote: > Or supposedly no change at all, as the attributes would be either in > TYPE_ATTRIBUTES, or on TYPE_DECL, but then it would be covered by > the TYPE_NAME comparison. Interesting point. So maybe all we need to do here is copy TYPE_ALIGN/TYPE_USER_ALIGN, not change any of the tests. Jason
--- gcc/cp/tree.c.jj 2015-04-01 15:29:33.000000000 +0200 +++ gcc/cp/tree.c 2015-04-08 09:09:45.326939354 +0200 @@ -880,12 +880,19 @@ build_cplus_array_type (tree elt_type, t { t = build_min_array_type (elt_type, index_type); set_array_type_canon (t, elt_type, index_type); + if (!dependent) + { + layout_type (t); + /* Make sure sizes are shared with the main variant. + layout_type can't be called after setting TYPE_NEXT_VARIANT, + as it will overwrite alignment etc. of all variants. */ + TYPE_SIZE (t) = TYPE_SIZE (m); + TYPE_SIZE_UNIT (t) = TYPE_SIZE_UNIT (m); + } TYPE_MAIN_VARIANT (t) = m; TYPE_NEXT_VARIANT (t) = TYPE_NEXT_VARIANT (m); TYPE_NEXT_VARIANT (m) = t; - if (!dependent) - layout_type (t); } } @@ -1057,21 +1064,23 @@ cp_build_qualified_type_real (tree type, should be equivalent to those in check_qualified_type. */ for (t = TYPE_MAIN_VARIANT (type); t; t = TYPE_NEXT_VARIANT (t)) if (TREE_TYPE (t) == element_type - && TYPE_NAME (t) == TYPE_NAME (type) - && TYPE_CONTEXT (t) == TYPE_CONTEXT (type) - && attribute_list_equal (TYPE_ATTRIBUTES (t), - TYPE_ATTRIBUTES (type))) + && check_base_type (t, type)) break; if (!t) { t = build_cplus_array_type (element_type, TYPE_DOMAIN (type)); - /* Keep the typedef name. */ - if (TYPE_NAME (t) != TYPE_NAME (type)) + /* Keep the typedef name, context and alignment. */ + if (TYPE_NAME (t) != TYPE_NAME (type) + || TYPE_CONTEXT (t) != TYPE_CONTEXT (type) + || TYPE_ALIGN (t) != TYPE_ALIGN (type)) { t = build_variant_type_copy (t); TYPE_NAME (t) = TYPE_NAME (type); + TYPE_CONTEXT (t) = TYPE_CONTEXT (type); + TYPE_ALIGN (t) = TYPE_ALIGN (type); + TYPE_USER_ALIGN (t) = TYPE_USER_ALIGN (type); } } --- gcc/testsuite/c-c++-common/attr-aligned-1.c.jj 2015-04-08 09:22:46.181427189 +0200 +++ gcc/testsuite/c-c++-common/attr-aligned-1.c 2015-04-08 09:26:41.315627195 +0200 @@ -0,0 +1,24 @@ +/* PR c++/65690 */ +/* { dg-do run } */ + +typedef double T[4][4] __attribute__((aligned (2 * __alignof__ (double)))); +void foo (const T); +struct S { T s; }; + +int +main () +{ + if (__alignof__ (struct S) != 2 * __alignof__ (double) + || __alignof__ (T) != 2 * __alignof__ (double) + || __alignof__ (const struct S) != 2 * __alignof__ (double) + || __alignof__ (const T) != 2 * __alignof__ (double)) + __builtin_abort (); + return 0; +} + +#if defined(__cplusplus) && __cplusplus >= 201103L +static_assert (alignof (S) == 2 * alignof (double), "alignment of S"); +static_assert (alignof (T) == 2 * alignof (double), "alignment of T"); +static_assert (alignof (const S) == 2 * alignof (double), "alignment of const S"); +static_assert (alignof (const T) == 2 * alignof (double), "alignment of const T"); +#endif
Hi! The following patches (included or attached) fix a regression on WebKit compilation. The first hunk basically reverts Honza's patch from December/January, because layout_type when the variants are already linked in doesn't layout just the current type, but also forcefully overwrites all the other variants, which is clearly highly undesirable. The attached patch has an alternate hunk for that, where it doesn't call layout_type at all and just copies over the needed fields from the main variant. The second hunk (the same in between both patches) fixes a problem that alignof (const T) is incorrect, but it has been that way already in 4.8 at least. I've been wondering if it would be possible that build_cplus_array_type would see incomplete element type on the main variant and complete qualified element type, but have not succeeded with e.g. struct S { typedef S T[4][4] __attribute__((aligned (16))); static T t; static volatile T v; }; void foo (const S::T); volatile const S::T w; S::T S::t; volatile S::T S::v; The included (first) patch has been successfully bootstrapped/regtested on x86_64-linux and i686-linux, the attached patch not, but I can bootstrap/regtest it if you prefer it. 2015-04-08 Jakub Jelinek <jakub@redhat.com> Jan Hubicka <hubicka@ucw.cz> PR c++/65690 * tree.c (build_cplus_array_type): Layout type before variants are set, but copy over TYPE_SIZE and TYPE_SIZE_UNIT from the main variant. (cp_build_qualified_type_real): Use check_base_type. Build a variant and copy over even TYPE_CONTEXT and TYPE_ALIGN/TYPE_USER_ALIGN if any of those are different. * c-c++-common/attr-aligned-1.c: New test. Jakub 2015-04-08 Jakub Jelinek <jakub@redhat.com> Jan Hubicka <hubicka@ucw.cz> PR c++/65690 * tree.c (build_cplus_array_type): Don't layout the type, instead copy everything over from the main variant. (cp_build_qualified_type_real): Use check_base_type. Build a variant and copy over even TYPE_CONTEXT and TYPE_ALIGN/TYPE_USER_ALIGN if any of those are different. * c-c++-common/attr-aligned-1.c: New test. --- gcc/cp/tree.c.jj 2015-04-01 15:29:33.000000000 +0200 +++ gcc/cp/tree.c 2015-04-08 09:09:45.326939354 +0200 @@ -880,12 +880,19 @@ build_cplus_array_type (tree elt_type, t { t = build_min_array_type (elt_type, index_type); set_array_type_canon (t, elt_type, index_type); + if (!dependent) + { + TYPE_SIZE (t) = TYPE_SIZE (m); + TYPE_SIZE_UNIT (t) = TYPE_SIZE_UNIT (m); + TYPE_ALIGN (t) = TYPE_ALIGN (m); + TYPE_PRECISION (t) = TYPE_PRECISION (m); + TYPE_USER_ALIGN (t) = TYPE_USER_ALIGN (m); + SET_TYPE_MODE (t, TYPE_MODE (m)); + } TYPE_MAIN_VARIANT (t) = m; TYPE_NEXT_VARIANT (t) = TYPE_NEXT_VARIANT (m); TYPE_NEXT_VARIANT (m) = t; - if (!dependent) - layout_type (t); } } @@ -1057,21 +1064,23 @@ cp_build_qualified_type_real (tree type, should be equivalent to those in check_qualified_type. */ for (t = TYPE_MAIN_VARIANT (type); t; t = TYPE_NEXT_VARIANT (t)) if (TREE_TYPE (t) == element_type - && TYPE_NAME (t) == TYPE_NAME (type) - && TYPE_CONTEXT (t) == TYPE_CONTEXT (type) - && attribute_list_equal (TYPE_ATTRIBUTES (t), - TYPE_ATTRIBUTES (type))) + && check_base_type (t, type)) break; if (!t) { t = build_cplus_array_type (element_type, TYPE_DOMAIN (type)); - /* Keep the typedef name. */ - if (TYPE_NAME (t) != TYPE_NAME (type)) + /* Keep the typedef name, context and alignment. */ + if (TYPE_NAME (t) != TYPE_NAME (type) + || TYPE_CONTEXT (t) != TYPE_CONTEXT (type) + || TYPE_ALIGN (t) != TYPE_ALIGN (type)) { t = build_variant_type_copy (t); TYPE_NAME (t) = TYPE_NAME (type); + TYPE_CONTEXT (t) = TYPE_CONTEXT (type); + TYPE_ALIGN (t) = TYPE_ALIGN (type); + TYPE_USER_ALIGN (t) = TYPE_USER_ALIGN (type); } } --- gcc/testsuite/c-c++-common/attr-aligned-1.c.jj 2015-04-08 09:22:46.181427189 +0200 +++ gcc/testsuite/c-c++-common/attr-aligned-1.c 2015-04-08 09:26:41.315627195 +0200 @@ -0,0 +1,24 @@ +/* PR c++/65690 */ +/* { dg-do run } */ + +typedef double T[4][4] __attribute__((aligned (2 * __alignof__ (double)))); +void foo (const T); +struct S { T s; }; + +int +main () +{ + if (__alignof__ (struct S) != 2 * __alignof__ (double) + || __alignof__ (T) != 2 * __alignof__ (double) + || __alignof__ (const struct S) != 2 * __alignof__ (double) + || __alignof__ (const T) != 2 * __alignof__ (double)) + __builtin_abort (); + return 0; +} + +#if defined(__cplusplus) && __cplusplus >= 201103L +static_assert (alignof (S) == 2 * alignof (double), "alignment of S"); +static_assert (alignof (T) == 2 * alignof (double), "alignment of T"); +static_assert (alignof (const S) == 2 * alignof (double), "alignment of const S"); +static_assert (alignof (const T) == 2 * alignof (double), "alignment of const T"); +#endif