Re: [PATCH] c++: Drop TREE_READONLY on vars (possibly) initialized by tls wrapper [PR109164]
On Sat, Mar 18, 2023 at 08:39:45AM -0400, Jason Merrill wrote: > On 3/17/23 13:51, Jakub Jelinek wrote: > > The following two testcases are miscompiled, because we keep TREE_READONLY > > on the vars even when they are (possibly) dynamically initialized by a TLS > > wrapper function. Normally cp_finish_decl drops TREE_READONLY from vars > > which need dynamic initialization, but for TLS we do this kind of > > initialization upon every access to those variables. > > Why not handle this case in cp_finish_decl, too? That is, add > DECL_THREAD_LOCAL_P to the TREE_STATIC check in The patch is mostly about DECL_EXTERNAL cases, the others are supposedly handled by the var_definition_p code there (or at least I assumed; testcases certainly test only DECL_EXTERNAL). I guess it could be done in cp_finish_decl, maybe better next to the /* A reference will be modified here, as it is initialized. */ if (! DECL_EXTERNAL (decl) && TREE_READONLY (decl) && TYPE_REF_P (type)) { was_readonly = 1; TREE_READONLY (decl) = 0; } spot, but we'd need to export the decl2.cc helpers for it, because not all DECL_THREAD_LOCAL_P vars need to be treated that way. if (VAR_P (decl) && CP_DECL_THREAD_LOCAL_P (decl) && var_needs_tls_wrapper (decl) && (!DECL_EXTERNAL (decl) || flag_extern_tls_init)) TREE_READONLY (decl) = 0; where var_needs_tls_wrapper would need to be exported from decl2.cc. Though, var_needs_tls_wrapper -> var_defined_without_dynamic_init needs DECL_NONTRIVIALLY_INITIALIZED_P/DECL_INITIALIZED_BY_CONSTANT_EXPRESSION_P, so perhaps that is accurate only closer to the end of cp_finish_decl? Jakub
Re: [PATCH] c++: Drop TREE_READONLY on vars (possibly) initialized by tls wrapper [PR109164]
On 3/17/23 13:51, Jakub Jelinek wrote: Hi! The following two testcases are miscompiled, because we keep TREE_READONLY on the vars even when they are (possibly) dynamically initialized by a TLS wrapper function. Normally cp_finish_decl drops TREE_READONLY from vars which need dynamic initialization, but for TLS we do this kind of initialization upon every access to those variables. Why not handle this case in cp_finish_decl, too? That is, add DECL_THREAD_LOCAL_P to the TREE_STATIC check in if (var_definition_p /* With -fmerge-all-constants, gimplify_init_constructor might add TREE_STATIC to aggregate variables. */ && (TREE_STATIC (decl) || (flag_merge_constants >= 2 && AGGREGATE_TYPE_P (type { /* If a TREE_READONLY variable needs initialization at runtime, it is no longer readonly and we need to avoid MEM_READONLY_P being set on RTL created for it. */ ? Keeping them TREE_READONLY means e.g. PRE can hoist loads from those before loops which contain the TLS wrapper calls, so we can access the TLS variables before they are initialized. Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? 2023-03-17 Jakub Jelinek PR c++/109164 * decl2.cc (get_tls_wrapper_fn): Clear TREE_READONLY on variables for which a TLS wrapper is added. * g++.dg/tls/thread_local13.C: New test. * g++.dg/tls/thread_local13-aux.cc: New file. * g++.dg/tls/thread_local14.C: New test. * g++.dg/tls/thread_local14-aux.cc: New file. --- gcc/cp/decl2.cc.jj 2023-03-07 21:20:31.800491531 +0100 +++ gcc/cp/decl2.cc 2023-03-17 12:20:11.960678291 +0100 @@ -3773,6 +3773,12 @@ get_tls_wrapper_fn (tree var) DECL_BEFRIENDING_CLASSES (fn) = var; set_global_binding (fn); + + /* The variable now needs dynamic initialization by the wrapper +function, we don't want to hoist accesses to it before the +wrapper. */ + if (TREE_READONLY (var)) + TREE_READONLY (var) = 0; } return fn; } --- gcc/testsuite/g++.dg/tls/thread_local13.C.jj2023-03-17 12:28:24.692427351 +0100 +++ gcc/testsuite/g++.dg/tls/thread_local13.C 2023-03-17 12:30:34.505519746 +0100 @@ -0,0 +1,21 @@ +// PR c++/109164 +// { dg-do run { target c++11 } } +// { dg-options "-O2" } +// { dg-add-options tls } +// { dg-require-effective-target tls_runtime } +// { dg-additional-sources "thread_local13-aux.cc" } + +struct S { virtual void foo (); int s; }; +extern thread_local S &t; +bool bar (); + +bool +baz () +{ + while (1) +{ + t.foo (); + if (!bar ()) +return false; +} +} --- gcc/testsuite/g++.dg/tls/thread_local13-aux.cc.jj 2023-03-17 12:28:28.721368058 +0100 +++ gcc/testsuite/g++.dg/tls/thread_local13-aux.cc 2023-03-17 12:37:53.952070861 +0100 @@ -0,0 +1,35 @@ +// PR c++/109164 + +struct S { virtual void foo (); int s; }; +extern bool baz (); + +void +S::foo () +{ + if (s != 42) +__builtin_abort (); +} + +S s; + +S & +qux () +{ + s.s = 42; + return s; +} + +thread_local S &t = qux (); + +bool +bar () +{ + return false; +} + +int +main () +{ + if (baz ()) +__builtin_abort (); +} --- gcc/testsuite/g++.dg/tls/thread_local14.C.jj2023-03-17 12:35:48.951905245 +0100 +++ gcc/testsuite/g++.dg/tls/thread_local14.C 2023-03-17 12:49:03.456249628 +0100 @@ -0,0 +1,19 @@ +// PR c++/109164 +// { dg-do run { target c++11 } } +// { dg-options "-O2" } +// { dg-add-options tls } +// { dg-require-effective-target tls_runtime } +// { dg-additional-sources "thread_local14-aux.cc" } + +extern thread_local const int t; +bool bar (int); + +bool +baz () +{ + while (1) +{ + if (!bar (t)) +return false; +} +} --- gcc/testsuite/g++.dg/tls/thread_local14-aux.cc.jj 2023-03-17 12:36:58.724881322 +0100 +++ gcc/testsuite/g++.dg/tls/thread_local14-aux.cc 2023-03-17 12:48:53.914389421 +0100 @@ -0,0 +1,26 @@ +// PR c++/109164 + +extern bool baz (); + +int +qux () +{ + return 42; +} + +extern thread_local const int t = qux (); + +bool +bar (int x) +{ + if (x != 42) +__builtin_abort (); + return false; +} + +int +main () +{ + if (baz ()) +__builtin_abort (); +} Jakub
[PATCH] c++: Drop TREE_READONLY on vars (possibly) initialized by tls wrapper [PR109164]
Hi! The following two testcases are miscompiled, because we keep TREE_READONLY on the vars even when they are (possibly) dynamically initialized by a TLS wrapper function. Normally cp_finish_decl drops TREE_READONLY from vars which need dynamic initialization, but for TLS we do this kind of initialization upon every access to those variables. Keeping them TREE_READONLY means e.g. PRE can hoist loads from those before loops which contain the TLS wrapper calls, so we can access the TLS variables before they are initialized. Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? 2023-03-17 Jakub Jelinek PR c++/109164 * decl2.cc (get_tls_wrapper_fn): Clear TREE_READONLY on variables for which a TLS wrapper is added. * g++.dg/tls/thread_local13.C: New test. * g++.dg/tls/thread_local13-aux.cc: New file. * g++.dg/tls/thread_local14.C: New test. * g++.dg/tls/thread_local14-aux.cc: New file. --- gcc/cp/decl2.cc.jj 2023-03-07 21:20:31.800491531 +0100 +++ gcc/cp/decl2.cc 2023-03-17 12:20:11.960678291 +0100 @@ -3773,6 +3773,12 @@ get_tls_wrapper_fn (tree var) DECL_BEFRIENDING_CLASSES (fn) = var; set_global_binding (fn); + + /* The variable now needs dynamic initialization by the wrapper +function, we don't want to hoist accesses to it before the +wrapper. */ + if (TREE_READONLY (var)) + TREE_READONLY (var) = 0; } return fn; } --- gcc/testsuite/g++.dg/tls/thread_local13.C.jj2023-03-17 12:28:24.692427351 +0100 +++ gcc/testsuite/g++.dg/tls/thread_local13.C 2023-03-17 12:30:34.505519746 +0100 @@ -0,0 +1,21 @@ +// PR c++/109164 +// { dg-do run { target c++11 } } +// { dg-options "-O2" } +// { dg-add-options tls } +// { dg-require-effective-target tls_runtime } +// { dg-additional-sources "thread_local13-aux.cc" } + +struct S { virtual void foo (); int s; }; +extern thread_local S &t; +bool bar (); + +bool +baz () +{ + while (1) +{ + t.foo (); + if (!bar ()) +return false; +} +} --- gcc/testsuite/g++.dg/tls/thread_local13-aux.cc.jj 2023-03-17 12:28:28.721368058 +0100 +++ gcc/testsuite/g++.dg/tls/thread_local13-aux.cc 2023-03-17 12:37:53.952070861 +0100 @@ -0,0 +1,35 @@ +// PR c++/109164 + +struct S { virtual void foo (); int s; }; +extern bool baz (); + +void +S::foo () +{ + if (s != 42) +__builtin_abort (); +} + +S s; + +S & +qux () +{ + s.s = 42; + return s; +} + +thread_local S &t = qux (); + +bool +bar () +{ + return false; +} + +int +main () +{ + if (baz ()) +__builtin_abort (); +} --- gcc/testsuite/g++.dg/tls/thread_local14.C.jj2023-03-17 12:35:48.951905245 +0100 +++ gcc/testsuite/g++.dg/tls/thread_local14.C 2023-03-17 12:49:03.456249628 +0100 @@ -0,0 +1,19 @@ +// PR c++/109164 +// { dg-do run { target c++11 } } +// { dg-options "-O2" } +// { dg-add-options tls } +// { dg-require-effective-target tls_runtime } +// { dg-additional-sources "thread_local14-aux.cc" } + +extern thread_local const int t; +bool bar (int); + +bool +baz () +{ + while (1) +{ + if (!bar (t)) +return false; +} +} --- gcc/testsuite/g++.dg/tls/thread_local14-aux.cc.jj 2023-03-17 12:36:58.724881322 +0100 +++ gcc/testsuite/g++.dg/tls/thread_local14-aux.cc 2023-03-17 12:48:53.914389421 +0100 @@ -0,0 +1,26 @@ +// PR c++/109164 + +extern bool baz (); + +int +qux () +{ + return 42; +} + +extern thread_local const int t = qux (); + +bool +bar (int x) +{ + if (x != 42) +__builtin_abort (); + return false; +} + +int +main () +{ + if (baz ()) +__builtin_abort (); +} Jakub