[Bug target/70098] PowerPC64: eigen hits ICE following invalid register assignment

[segher at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70098

--- Comment #8 from Segher Boessenkool <segher at gcc dot gnu.org> ---
Author: segher
Date: Mon Mar 14 14:18:44 2016
New Revision: 234181

URL: https://gcc.gnu.org/viewcvs?rev=234181&root=gcc&view=rev
Log:
rs6000: Handle "d" output in the bd*z patterns (PR70098)

In the rs6000 port, FLOAT_REGS can contain DImode values when compiling
for 64-bit targets.  Some instructions (like "fcfid" in the testcase,
convert from integer to DP float) only work on floating point registers.
So, we do want to allow DImode in these regs.

Now, in unusual cases IRA will assign FLOAT_REGS to some allocno where
some insns cannot handle FLOAT_REGS there, so they will need a reload.
Maybe IRA can be made smarter, but it isn't doing anything wrong here,
so we should be able to handle it.

The place it goes wrong is in the output of the *ctrdi_internal[1256]
pattern: the "bdz" and "bdnz" instructions.  GCC refuses to do output
reloads on JUMP_INSNs, probably because it is hard to do, needs different
strategies than "normal" reloads do, and it cannot even be done at all
for general patterns.  So JUMP_INSNs need to be able to handle every
possible output for the register class used.

These patterns already handle writing to "c" (the base insn case), and
to "r", "m", and "c" or "l"; all those via splitters.  We just need to
handle "d" as well.  That is what this patch does.  [A predicate in one
of the splitters needs to be touched up so that the correct splitter
is used for the FLOAT_REGS case.]

But, that leaves another problem.  One of the insns that are split to
is a move from a GPR to an FPR.  That work fine on targets with direct
move (which does exactly that), i.e. power8 and up.  But older targets
need memory to do the move, and this splitter runs after reload so
it cannot allocate memory; and allocating memory beforehand for every
bdnz insn is pretty horrible as well.

This patch implements the easy part.  With it, power8 works, where it
didn't before.


        PR target/70098
        * config/rs6000/rs6000.md (*ctr<mode>_internal1, *ctr<mode>_internal2,
        *ctr<mode>_internal5, *ctr<mode>_internal6): Also allow "d" as output.
        (define_split for the GPR case): Use int_reg_operand instead of
        gpc_reg_operand for the output.

gcc/testsuite/
        PR target/70098
        * lib/target-supports.exp (check_effective_target_powerpc64_no_dm):
        New function.
        * g++.dg/pr70098.C: New testcase.

Added:
    trunk/gcc/testsuite/g++.dg/pr70098.C
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/rs6000/rs6000.md
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/testsuite/lib/target-supports.exp