[Bug target/84741] [7/8 Regression] ICE in ix86_expand_prologue, at config/i386/i386.c:13893

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84741

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |dmalcolm at gcc dot gnu.org,
                   |                            |jakub at gcc dot gnu.org

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I don't think this is a regression, the __RTL stuff has been introduced in GCC
7.  And, there are infinite number of ways to get ICEs with __RTL or __GIMPLE
by just adding invalid IL, __RTL and __GIMPLE are just meant for testing the
compiler.

The reason this ICEs is that if we call run_rtl_passes we bypass the
initialization done at expand time, in particular prepare_function_start
which among other things does:
 if (flag_stack_usage_info)
    {
      cfun->su = ggc_cleared_alloc<stack_usage> ();
      cfun->su->static_stack_size = -1;
    }
and so cfun->su is uninitialized.
Note, if I replace in the testcase __RTL (startwith ("vregs")) with just __RTL,
it ICEs exactly in prepare_function_start on the assertion that 
gcc_assert (!get_last_insn ());
Note prepare_function_start initializes a bunch of other global vars:
  cse_not_expected = ! optimize;

  /* Caller save not needed yet.  */
  caller_save_needed = 0;

  /* We haven't done register allocation yet.  */
  reg_renumber = 0;

  /* Indicate that we have not instantiated virtual registers yet.  */
  virtuals_instantiated = 0;

  /* Indicate that we want CONCATs now.  */
  generating_concat_p = 1;

  /* Indicate we have no need of a frame pointer yet.  */
  frame_pointer_needed = 0;

If we don't start with expand pass, I guess we want !generating_concat_p, but
the others really depend on what exact pass it starts with.  I'm afraid the RTL
passes has lots of other state that I don't really see how run_rtl_passes would
make it work.  E.g. reload_completed, epilogue_completed, regstack_completed,
...  All those start initially clear (or as set above), and at some point are
changed after some pass.