[Bug rtl-optimization/60501] New: LRA emits add patterns which might clobber cc

[krebbel at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60501

            Bug ID: 60501
           Summary: LRA emits add patterns which might clobber cc
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Severity: critical
          Priority: P3
         Component: rtl-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: krebbel at gcc dot gnu.org

Created attachment 32333
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=32333&action=edit
Experimental fix

This actually is a follow-on to #57604.

While debugging 435.gromacs I've noticed that LRA still emits add patterns
which might clobber cc without actually checking whether this is valid.

no lra:
  4982a6:       eb e6 00 01 00 df       sllk    %r14,%r6,1
  4982ac:       a7 78 00 02             lhi     %r7,2
  4982b0:       40 7e a0 00             sth     %r7,0(%r14,%r10)
  4982b4:       40 7e a0 02             sth     %r7,2(%r14,%r10)
  4982b8:       40 7e a0 04             sth     %r7,4(%r14,%r10)
  4982bc:       b9 f2 30 06             locrnle %r0,%r6

lra:
  49823e:       eb 9e 00 01 00 df       sllk    %r9,%r14,1 <-- cc set
  498244:       18 69                   lr      %r6,%r9
  498246:       1a 6a                   ar      %r6,%r10   <-- cc clobber
  498248:       e5 44 60 00 00 02       mvhhi   0(%r6),2
  49824e:       a7 78 00 02             lhi     %r7,2
  498252:       a7 38 00 04             lhi     %r3,4
  498256:       1a 7a                   ar      %r7,%r10   <-- cc clobber
  498258:       1a 3a                   ar      %r3,%r10   <-- cc clobber
  49825a:       1a 79                   ar      %r7,%r9    <-- cc clobber
  49825c:       1a 39                   ar      %r3,%r9    <-- cc clobber
  49825e:       e5 44 70 00 00 02       mvhhi   0(%r7),2
  498264:       b9 f2 30 0e             locrnle %r0,%r14   <-- cc use

Apart from the no lra code looking much better the LRA variant is dies with a
miscompare.

A potential solution has already been discussed in:
http://gcc.gnu.org/ml/gcc-patches/2013-06/msg01174.html

Richard proposed to have standard pattern names for pointer arithmetic. I've
tried to implement this for S/390. With the attached (untested) patch I get the
following:

     97e:       eb 9e 00 01 00 df       sllk    %r9,%r14,1
     984:       41 69 a0 00             la      %r6,0(%r9,%r10)
     988:       e5 44 60 00 00 02       mvhhi   0(%r6),2
     98e:       19 0e                   cr      %r0,%r14
     990:       a7 78 00 02             lhi     %r7,2
     994:       41 37 a0 00             la      %r3,0(%r7,%r10)
     998:       41 63 90 00             la      %r6,0(%r3,%r9)
     99c:       e5 44 60 00 00 02       mvhhi   0(%r6),2
     9a2:       b9 f2 30 0e             locrnle %r0,%r14