[Bug middle-end/60418] New: [4.9 Regression] 435.gromacs in SPEC CPU 2006 is miscompiled

["hjl.tools at gmail dot com" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60418

            Bug ID: 60418
           Summary: [4.9 Regression] 435.gromacs in SPEC CPU 2006 is
                    miscompiled
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: hjl.tools at gmail dot com

Since r208165, 435.gromacs in SPEC CPU 2006 is miscompiled on x32 with

-mx32 -O3 -funroll-loops -ffast-math -fwhole-program -flto=jobserver
-fuse-linker-plugin

The failure is

  Running 435.gromacs ref peak lto default

*** Miscompare of gromacs.out; for details see
   
/export/project/git/gcc-regression/spec/2006/spec/benchspec/CPU2006/435.grom
acs/run/run_peak_ref_lto.0000/gromacs.out.mis

cat
/export/project/git/gcc-regression/spec/2006/spec/benchspec/CPU2006/435.gromacs/run/run_peak_ref_lto.0000/gromacs.out.mis
0002:  3.07684e+02
       3.03594e+02

The result of 3.03594e+02 is outside of tolerance for 3.07684e+02.
I added a static counter in tree_forwarder_block_p for the number of
times returned true above and return false if it is between getenv("from")
and getenv("to").  I noticed that as more basic blocks get removed,
the deviation is getting bigger. One removed basic block has

;; basic block 3, loop depth 0, count 0, freq 225, maybe hot 
;;  prev block 2, next block 4, flags: (NEW, REACHABLE)
;;  pred:       2
;; 
;;  succ:       4

and its successor has

;; basic block 4, loop depth 1, count 0, freq 2500, maybe hot 
;;  prev block 3, next block 5, flags: (NEW)
;;  pred:       3
;;              6
;; starting at line 722 
# gimple_phi <i_429, 0(3), [bondfree.c : 726:24] i_27(6)>

Is is safe to remove basic block 3?