Le 15/11/2026 à 04:52, Jerry DeLisle a écrit :
See attached patch. For p1 = toto(p1), where toto deallocates and reallocates a component of its dummy, gfortran saves the components of p1 before the call and frees them after the assignment. The call has already freed them, so they are freed twice, and the components toto stored in p1 leak. The patch evaluates the call before the components are saved when an actual argument overlaps the lhs, as decided by gfc_check_dependency so that w%c = f(w) and w = f(w%c) are caught too, and its dummy is not INTENT(IN) or VALUE. -O2 code for other calls is unchanged.Such an assignment is not standard conforming (F2023 10.1.4 p2), but it should not corrupt the heap. The test is in gfortran.dg/asan because glibc usually reuses the freed block, so without -fsanitize=address it passes on trunk. The fix is rhs_may_redefine_lhs in trans-expr.cc and its use in gfc_trans_assignment_1; the rest is the new test. Regression tested on x86_64. OK for mainline? Regards, Jerry --- fortran: [PR52774] Double free when a function call redefines the lhs For lhs = f(...) with an lhs that has allocatable components, gfortran saved the lhs components before calling f and freed them after the assignment. When f redefined the lhs through an actual argument that overlaps it, f had already freed the saved components, so they were freed twice, and the ones f left in the lhs leaked. Evaluate such a call before the lhs components are saved. The assignment violates F2023 10.1.4 p2, which is not a constraint, so no diagnostic is added. Assisted-by: Claude Opus 5.5 PR fortran/52774 gcc/fortran/ChangeLog: * trans-expr.cc (rhs_may_redefine_lhs): New function. (gfc_trans_assignment_1): Evaluate an rhs that may redefine the lhs before the lhs components are saved for deallocation. gcc/testsuite/ChangeLog: * gfortran.dg/asan/pr52774.f90: New test.
OK.
