SPEC's CPU2000 253.perlbmk fails with -O3 using armhf 2014.05 toolchain

[frank smith]

Built SPEC's CPU2000 suite using ARMHF toolcahin 2014.05.

The benchmark 253.perlbmk failed to correctly execute (mis-compared) when built with -O3. Changing to -O2 is successfully ran. However with the 2014.04 release of this toolchain it built and ran with the -O3 flag which makes this a regression

Command I used below with the only difference being -O3 changing to -O2

COMP: /opt/gcc-linaro-arm-linux-gnueabihf-4.9-2014.05_linux/bin/arm-linux-gnueabihf-gcc -c -o options.o -static -DSPEC_CPU2000_LINUX_I386 -DSPEC_CPU2000_NEED_BOOL -O2 -march=armv7-a -mfpu=vfpv3-d16 -fno-use-linker-plugin -flto -marm -mfloat-abi=hard -ffast-math -ftree-vectorize -g -static

LINK: /opt/gcc-linaro-arm-linux-gnueabihf-4.9-2014.05_linux/bin/arm-linux-gnueabihf-gcc -O2 -march=armv7-a -mfpu=vfpv3-d16 -fno-use-linker-plugin -flto -marm -mfloat-abi=hard -ffast-math -ftree-vectorize -g -static -static -lm -o options

-----------
Note that with the AArch64 2014.04 toolchain 253.perlbmk failed to run correctly with -O3 but ran correctly with -O2. This still occurs in AArch64 2014.05 toolchain but since this problem existed in 2014.04 this, for me at least, was not a regression.

I can add the AArch64 flags if requested.