[Bug ipa/105160] [12 regression] ipa modref marks functions with asm volatile as const or pure

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105160

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |hubicka at gcc dot gnu.org

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
The odd thing is that during IPA propagation (modref_propagate_in_scc), the
summary of sctlr_read is

(gdb) p *cur_summary
$2 = {loads = 0x7ffff665af00, stores = 0x7ffff665af10, 
  kills = {<vec<modref_access_node, va_heap, vl_ptr>> = {
      m_vec = 0x0}, <No data fields>}, 
  arg_flags = {<vec<unsigned short, va_heap, vl_ptr>> = {
      m_vec = 0x0}, <No data fields>}, retslot_flags = 0, 
  static_chain_flags = 0, writes_errno = 0, side_effects = 0, 
  nondeterministic = 0, calls_interposable = 0, load_accesses = 0, 
  global_memory_read = 0, global_memory_written = 0, try_dse = 1}

which is because we access the summary of cur->inlined_to (sctlr_rmw) here
but that hasn't been processed.  But the non-inlined node doesn't have a
summary.

So somehow we fail to merge the info from the inlined clones?

But even using ->clone_of instead of ->inlined_to doesn't make a difference
(no summary for that node either).

Something is wrong.  Honza?