[Bug rtl-optimization/54110] New: lower-subreg related code quality for long long function return

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

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54110

             Bug #: 54110
           Summary: lower-subreg related code quality for long long
                    function return
    Classification: Unclassified
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: amodra@gmail.com


powerpc-linux-gcc -m32 -O2
long long ll (long long *x) { return *x; }

Rev 189440 output
 lwz r4,4(r3)
 lwz r3 0(r3)
 blr

Rev 189441 output
 mr r9,r3
 lwz r3,0(r9)
 lwz r4,4(r9)
 blr

Not a big deal, but it gets a little worse for
long long llo (long long *x) { return x[4095]; }

Current mainline
 lwz r11,32764(r3)
 lwz r10,32760(r3)
 mr r4,r11
 mr r3,r10
 blr

Current mainline less r189441 patch
 lwz r4,32764(r3)
 lwz r3,32760(r3)
 blr

Noticed when developing testcases for pr53914.  I happened to be working with a
copy of mainline a day before r189441 and saw ideal code being generated with
my pr53914 fix.

Some notes
- Obviously we don't want to revert r189441 as without that the lower-subreg
pass is effectively disabled on powerpc.
- Without lower-subreg, combine merges (set (reg:DI 124) (mem:DI (...));
(set (reg:DI 121) (reg:DI 124)); (set (reg:DI 3) (reg:DI 121)
into (set (reg:DI 3) (mem:DI (...));  Combine can't do that when lowered to
SImode.
- For llo.c, lower-subreg leaves the mem as DImode due to offsettable_memref
quirks.  For smaller offsets the mem is split and you get addi,lwz,lwz,blr.
- If we have a DImode mem it persists until the split after reload.  It is
there that rs6000_split_multireg_move gets into play, and is reponsible for
reordering the loads.
- I'm not sure under what conditions loads that might trap can be reordered.