help with PR69133

[Jan Hubicka]

> On Mon, Jan 18, 2016 at 10:28 AM, Prathamesh Kulkarni
> <prathamesh.kulkarni@linaro.org> wrote:
> > On 17 January 2016 at 14:56, Prathamesh Kulkarni
> > <prathamesh.kulkarni@linaro.org> wrote:
> >> Hi,
> >> I was having a look at PR69133.
> >> It appears that with -flto-partition=none,
> >> cgraph_node::get_untransformed_body ()
> >> is called twice for node with asm_name _ZThn4_N11xercesc_3_11C5m_fn6ERKi.
> >> c++filt says it is: non-virtual thunk to xercesc_3_1::C::m_fn6(int const&)
> >>
> >> get_untransformed_body() calls
> >> lto_free_function_in_decl_state_for_node() which sets
> >> node->lto_file_data to NULL.
> >> Now 2nd time get_untransformed_body is called for the same node,
> >> node->lto_file_data is NULL and lto_get_decl_name_mapping() dereferences
> >> lto_file_data which results in segfault.
> >>
> >> I was wondering if gating on lto_file_data could be a reasonable solution ?
> >> if (!lto_file_data)
> >>   return false;
> >> I am not sure what value to return for this case (I chose false
> >> because of early exit).
> >> It prevents the ICE, and the patch passes bootstrap+test on
> >> x86_64-unknown-linux-gnu.
> >> With partitoning enabled, for the above test-case,
> >> get_untransformed_body () is called only once per node. However I
> >> don't understand why it gets called twice with -flto-partition=none.
> > Kugan pointed out to me the test-case doesn't ICE by passing -fno-ipa-icf.
> 
> It looks to me we are not supposed to call this twice on the same node
> but nothing guards against that (and we don't release the untransformed
> body either).  -flto-partition=none is special here becaue it combines WPA

You are suppose to read every function body just once from LTO stream. Doing
it multiple times is a waste.  You can call get_body and get_untransformed_body
as many times as you want as tHere is guard:
bool
cgraph_node::get_untransformed_body (void)
{
  lto_file_decl_data *file_data;
  const char *data, *name;
  size_t len;
  tree decl = this->decl;

  if (DECL_RESULT (decl))
    return false;

once you read the body DECL_RESULT is set.  Only way you can get past this
to ICE should be the case where you read the body and then trhow it away.

> and late stage.  IPA ICF is the only IPA pass needing untransformed bodies
> during WPA.

That is not exactly true.  We also read untransformed bodies when merging
profiles.

I will check the PR. ipa-icf is not supposed to get body of thunks.

Honza