[Bug rtl-optimization/67920] New: wrong code with -O3

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67920

            Bug ID: 67920
           Summary: wrong code with -O3
           Product: gcc
           Version: 6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jamrial at gmail dot com
  Target Milestone: ---

Some change to trunk made in the last 18 days is making gcc miscompile the
files libavcodec/jpeg2000dec.c libavcodec/j2kenc.c and libavcodec/avuidec.c
from ffmpeg git head when -O3 is used. This doesn't happen with -O2.
I tested with x86_64-pc-linux-gnu only.

How to reproduce:
git clone git://git.videolan.org/ffmpeg.git && cd ffmpeg && ./configure
make fate-vsynth1-jpeg2000 fate-vsynth1-avui

The tests will fail, the former showing both the encoder and decoder are
producing garbage, and the latter only the decoder. Adding -fno-strict-aliasing
-fwrapv does not make a difference.

The command Makefile uses to compile all three files looks like this:
gcc -I. -I/home/jamrial/ffmpeg/ -D_ISOC99_SOURCE -D_FILE_OFFSET_BITS=64
-D_LARGEFILE_SOURCE -D_POSIX_C_SOURCE=200112 -D_XOPEN_SOURCE=600 -DZLIB_CONST
-DHAVE_AV_CONFIG_H -std=c99 -fomit-frame-pointer -pthread -g
-Wdeclaration-after-statement -Wall -Wdisabled-optimization -Wpointer-arith
-Wredundant-decls -Wwrite-strings -Wtype-limits -Wundef -Wmissing-prototypes
-Wno-pointer-to-int-cast -Wstrict-prototypes -Wempty-body -Wno-parentheses
-Wno-switch -Wno-format-zero-length -Wno-pointer-sign -O3 -fno-math-errno
-fno-signed-zeros -fno-tree-vectorize -Werror=format-security
-Werror=implicit-function-declaration -Werror=missing-prototypes
-Werror=return-type -Werror=vla -Wformat -fdiagnostics-color=auto
-Wno-maybe-uninitialized -c -o OUTPUT.o INPUT.c

You can get that kind of verbose output by running make with V=1.
Keep in mind you'll probably run into pr67794 while bisecting gcc trunk to find
the commit that introduced this regression.

Apologizes for not having a simpler test case.