| Message ID | 1320675134.18023.349.camel@triegel.csb |
|---|---|
| State | New |
| Headers | show |
On 11/07/2011 06:12 AM, Torvald Riegel wrote: > stmt = begin_transaction_stmt (input_location, NULL, flags); > tmp = RECUR (TRANSACTION_EXPR_BODY (t)); > + if (tmp) > + { > + /* No statements; handle this like an expression. */ In which case I'm pretty sure you ought to check for non-null TRANSACTION_EXPR_BODY first and not call begin_transaction_stmt. r~
On 11/07/2011 11:31 AM, Richard Henderson wrote: > On 11/07/2011 06:12 AM, Torvald Riegel wrote: >> stmt = begin_transaction_stmt (input_location, NULL, flags); >> tmp = RECUR (TRANSACTION_EXPR_BODY (t)); >> + if (tmp) >> + { >> + /* No statements; handle this like an expression. */ > > In which case I'm pretty sure you ought to check for non-null > TRANSACTION_EXPR_BODY first and not call begin_transaction_stmt. I believe TRANSACTION_EXPR_BODY is always non-null, but when tsubst_expr is called for a statement it adds the new statement via add_stmt and returns null. This approach is a bit funny, but ought to work fine, and there doesn't seem to be a simple test to distinguish between statements and expressions currently. Jason
diff --git a/gcc/cp/pt.c b/gcc/cp/pt.c index b1593fe..6e6dcdb 100644 --- a/gcc/cp/pt.c +++ b/gcc/cp/pt.c @@ -12960,6 +12960,16 @@ tsubst_expr (tree t, tree args, tsubst_flags_t complain, tree in_decl, stmt = begin_transaction_stmt (input_location, NULL, flags); tmp = RECUR (TRANSACTION_EXPR_BODY (t)); + if (tmp) + { + /* No statements; handle this like an expression. */ + pop_stmt_list (TRANSACTION_EXPR_BODY (stmt)); + stmt = build1 (TRANSACTION_EXPR, TREE_TYPE (tmp), tmp); + TRANSACTION_EXPR_OUTER (stmt) = TRANSACTION_EXPR_OUTER (t); + TRANSACTION_EXPR_RELAXED (stmt) = TRANSACTION_EXPR_RELAXED (t); + SET_EXPR_LOCATION (stmt, EXPR_LOCATION (t)); + return stmt; + } finish_transaction_stmt (stmt, NULL, flags); } break; diff --git a/gcc/testsuite/g++.dg/tm/template-1.C b/gcc/testsuite/g++.dg/tm/template-1.C new file mode 100644 index 0000000..b93828a --- /dev/null +++ b/gcc/testsuite/g++.dg/tm/template-1.C @@ -0,0 +1,35 @@ +// { dg-do compile } +// { dg-options "-fgnu-tm -O -fdump-tree-tmmark" } + +struct TrueFalse +{ + static bool v() { return true; } +}; + +int global; + +template<typename T> int foo() +{ + __transaction_atomic { global += 2; } + return __transaction_atomic (global + 1); +} + +template<typename T> int bar() __transaction_atomic +{ + return global + 3; +} + +template<typename T> void bar2() __transaction_atomic +{ + global += 4; +} + +int f1() +{ + bar2<TrueFalse>(); + return foo<TrueFalse>() + bar<TrueFalse>(); +} + +/* 4 transactions overall, two of the write to global: */ +/* { dg-final { scan-tree-dump-times "ITM_RU4\\s*\\(&global" 4 "tmmark" } } */ +/* { dg-final { scan-tree-dump-times "ITM_WU4\\s*\\(&global" 2 "tmmark" } } */
I stumbled upon this when working on noexcept. When txn expressions were processed in tsubst_expr(), the previous code incorrectly created empty transaction statements because no statements got added to the statement list used by (begin|finish)_transaction_stmt(). Also, "return __transaction_atomic (x+1);" incorrectly returned an error that this wouldn't return a value. With the patch, if RECUR(...) returns a value, then we assume that this is an expression and create a txn expression instead. No regressions for the TM tests, running a bootstrap currently. OK for branch? commit 437bd75328aa06561cb84f5e419431f979776115 Author: Torvald Riegel <triegel@redhat.com> Date: Mon Nov 7 14:51:50 2011 +0100 Fix instantiation of transaction expressions. * cp/pt.c (tsubst_expr) [TRANSACTION_EXPR]: If body is not a statement, create an expression instead. * testsuite/g++.dg/tm/template-1.C: New.