[Bug rtl-optimization/92176] New: LRA problem with reloads for subreg operands

[krebbel at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92176

            Bug ID: 92176
           Summary: LRA problem with reloads for subreg operands
           Product: gcc
           Version: 10.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: krebbel at gcc dot gnu.org
  Target Milestone: ---

Created attachment 47083
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=47083&action=edit
reduced testcase

The attached testcase core dumps when compiled with -O3 -march=z13 but executes
fine with -O3 -march=zEC12.

A QImode subreg on a pseudo gets assigned a floating point register and is
simplified to (reg:QI (...)). However, changing the mode of an FPR is
disallowed by the S/390 backend in s390_can_change_mode class. The problem
appears to be that simplify_operand_subreg in lra-constraints.c does not check
can_change_mode_class before assigning a register class. It always picks the
preferred register class for an expression as returned by the backend.

reload used to invoke can_change_mode_class when reloading a subreg - see
push_reload.

I tried to return just GENERAL_REGS in the preferred_reload_class whenever
being invoked with a subreg for which can_change_mode class would return false.
This helps in some cases. However, that function does not always get called
with the subreg expression, if there are other uses of that register not having
a subreg.