[Bug target/68973] [6 regression] Internal compiler error on power for gcc/testsuite/g++.dg/pr67211.C

["amodra at gmail dot com" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68973

Alan Modra <amodra at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |amodra at gmail dot com

--- Comment #14 from Alan Modra <amodra at gmail dot com> ---
This looks like a reload problem.  The pr67211.C testcase on powerpc64le gives
reload this insn

(insn 155 153 156 8 (set (reg/f:DI 178 [ p3$_M_first ])
        (mem/f:DI (pre_inc:DI (reg/f:DI 185 [ p3$_M_node ])) [3 MEM[base: _9,
offset: 0B]+0 S8 A64])) /src/gcc.git/gcc/testsuite/g++.dg/pr67211.C:28 550
{*movdi_internal64}
     (expr_list:REG_INC (reg/f:DI 185 [ p3$_M_node ])
        (nil)))

and reload chooses to put reg 185 in fr31, likely due to using that reg in
vector insns elsewhere.  Of course, (pre_inc:DI (reg/f:DI 63)) isn't a valid
memory address so it needs reloading.  reload does that by reloading the entire
address into a gpr rather than just reloading fr31.  Unfortunately neither the
rs6000 backend machinery nor the code in push_reload that normally handles
secondary memory sees that a secondary reload is needed.

There is some pre_inc code in find_reloads_address_1 that could reload just
fr31 but that is currently only active for regs that already have a mem. 
Making that code active for invalid hard regs cures this failure.

Another fix is to disable the auto_inc_dec pass (maybe we should do that anyway
for powerpc processors where the update forms of memory insns are slow).