analyzer: detect zero-assignment in phis (PR 93544)
authorDavid Malcolm <dmalcolm@redhat.com>
Mon, 3 Feb 2020 16:23:09 +0000 (11:23 -0500)
committerDavid Malcolm <dmalcolm@redhat.com>
Mon, 3 Feb 2020 19:29:08 +0000 (14:29 -0500)
commit8525d1f5f57b11fe04a97674cc2fc2b7727621d0
tree54e3cdd1ac7812b005b343a24924232fa26b3fe0
parent73f386581bddc4d630b93eeb0cddd32943bf24e7
analyzer: detect zero-assignment in phis (PR 93544)

PR analyzer/93544 reports an ICE when attempting to report a double-free
within diagnostic_manager::prune_for_sm_diagnostic, in which the
variable of interest has become an INTEGER_CST.  Additionally, it picks
a nonsensical path through the function in which the pointer being
double-freed is known to be NULL, which we shouldn't complain about.

The dump shows that it picks the INTEGER_CST when updating var at a phi
node:
    considering event 4, with var: ‘iftmp.0_2’, state: ‘start’
    updating from ‘iftmp.0_2’ to ‘0B’ based on phi node
      phi: iftmp.0_2 = PHI <iftmp.0_6(3), 0B(2)>
    considering event 3, with var: ‘0B’, state: ‘start’
and that it has picked the shortest path through the exploded graph,
and on this path the pointer has been assigned NULL.

The root cause is that the state machine's on_stmt isn't called for phi
nodes (and wouldn't make much sense, as we wouldn't know which arg to
choose).  malloc state machine::on_stmt "sees" a GIMPLE_ASSIGN to NULL
and handles it by transitioning the lhs to the "null" state, but never
"sees" GIMPLE_PHI nodes.

This patch fixes the ICE by wiring up phi-handling with state machines,
so that state machines have an on_phi vfunc.  It updates the only current
user of "is_zero_assignment" (the malloc sm) to implement equivalent
logic for phi nodes.  Doing so ensures that the pointer is in a separate
sm-state for the NULL vs non-NULL cases, and so gets separate exploded
nodes, and hence the path-finding logic chooses the correct path, and
the correct non-NULL phi argument.

The patch also adds some bulletproofing to prune_for_sm_diagnostic to
avoid crashing in the event of a bad path.

gcc/analyzer/ChangeLog:
PR analyzer/93544
* diagnostic-manager.cc
(diagnostic_manager::prune_for_sm_diagnostic): Bulletproof
against bad choices due to bad paths.
* engine.cc (impl_region_model_context::on_phi): New.
* exploded-graph.h (impl_region_model_context::on_phi): New decl.
* region-model.cc (region_model::on_longjmp): Likewise.
(region_model::handle_phi): Add phi param.  Call the ctxt's on_phi
vfunc.
(region_model::update_for_phis): Pass phi to handle_phi.
* region-model.h (region_model::handle_phi): Add phi param.
(region_model_context::on_phi): New vfunc.
(test_region_model_context::on_phi): New.
* sm-malloc.cc (malloc_state_machine::on_phi): New.
(malloc_state_machine::on_zero_assignment): New.
* sm.h (state_machine::on_phi): New vfunc.

gcc/testsuite/ChangeLog:
PR analyzer/93544
* gcc.dg/analyzer/torture/pr93544.c: New test.
gcc/analyzer/ChangeLog
gcc/analyzer/diagnostic-manager.cc
gcc/analyzer/engine.cc
gcc/analyzer/exploded-graph.h
gcc/analyzer/region-model.cc
gcc/analyzer/region-model.h
gcc/analyzer/sm-malloc.cc
gcc/analyzer/sm.h
gcc/testsuite/ChangeLog
gcc/testsuite/gcc.dg/analyzer/torture/pr93544.c [new file with mode: 0644]