Fix expand_builtin_atomic_fetch_op for pre-op (PR80902)
authorSegher Boessenkool <segher@kernel.crashing.org>
Fri, 23 Jun 2017 17:37:27 +0000 (19:37 +0200)
committerSegher Boessenkool <segher@gcc.gnu.org>
Fri, 23 Jun 2017 17:37:27 +0000 (19:37 +0200)
commit08c273bb9ca995bc8c8f58a3cbdb2be6a47765bd
tree89c7f9311baba6c0a9a0a452d30f482c63850b2a
parenteded3fe59237591c17259f1b6df9174b25643673
Fix expand_builtin_atomic_fetch_op for pre-op (PR80902)

__atomic_add_fetch adds a value to some memory, and returns the result.
If there is no direct support for this, expand_builtin_atomic_fetch_op
is asked to implement this as __atomic_fetch_add (which returns the
original value of the mem), followed by the addition.  Now, the
__atomic_add_fetch could have been a tail call, but we shouldn't
perform the __atomic_fetch_add as a tail call: following code would
not be executed, and in fact thrown away because there is a barrier
after tail calls.

This fixes it.

PR middle-end/80902
* builtins.c (expand_builtin_atomic_fetch_op): If emitting code after
a call, force the call to not be a tail call.

From-SVN: r249603
gcc/ChangeLog
gcc/builtins.c