[Bug middle-end/88032] [9 Regression] ICE in operand_subword_force, at emit-rtl.c:1793

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88032

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |jakub at gcc dot gnu.org

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 45009
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=45009&action=edit
gcc9-pr88032.patch

Untested fix.  The problem as I see is that while operand_subword will for mode
== VOIDmode just pick the mode from the operand's mode, so it is important to
pass it only if the operand has VOIDmode, operand_subword_force actually with
VOIDmode argument acts just as operand_subword that asserts the result is
non-NULL.  That is something we don't want in this case, e.g. because op is a
hard register that we need to force into a pseudo to be able to create a subreg
out of it.