Array subscript is above array bounds when compiling NLSF2A.c

[rawr51919]

I happened to come across this bug with spearmint, available at spearmint.pw. The description of this bug is, when you go to compile opus, the compiler gives out 4 warnings. Here's an excerpt of the warning as it happens:

CC code/opus-1.1/silk/NLSF2A.c
code/opus-1.1/silk/NLSF2A.c: In function ‘silk_NLSF2A’:
code/opus-1.1/silk/NLSF2A.c:121:17: warning: array subscript is above array bounds [-Warray-bounds]
Ptmp = P[ k+1 ] + P[ k ];
^
code/opus-1.1/silk/NLSF2A.c:122:17: warning: array subscript is above array bounds [-Warray-bounds]
Qtmp = Q[ k+1 ] - Q[ k ];
^
code/opus-1.1/silk/NLSF2A.c:121:17: warning: array subscript is above array bounds [-Warray-bounds]
Ptmp = P[ k+1 ] + P[ k ];
^
code/opus-1.1/silk/NLSF2A.c:122:17: warning: array subscript is above array bounds [-Warray-bounds]
Qtmp = Q[ k+1 ] - Q[ k ];
^

This also happens when compiling any project which uses opus, meaning it's probably also a bug with opus.

[rillian]

I think these are spurious. On line 89 we have

silk_assert( d==10||d==16 );

then on line 113, the loop bound is set to

dd = silk_RSHIFT( d, 1 );

so dd is either 5 or 8. Meanwhile the array itself is size SILK_MAX_ORDER_LPC / 2 + 1 which is 9 (by SigProc_FIX.h:38) so the k+1 array access is safe.

Any suggestions how we can silence the warning? Which compiler version shows the issue?

[rawr51919]

Latest Cygwin, rillian. I'm currently using the x64 version.

[rillian]

what does gcc --version say?

[rawr51919]

It reports gcc 5.3.0.

[rawr51919]

zturtleman fixed this error in Spearmint. Here's the changelog for the error fix: zturtleman/spearmint@a3da9be

[rillian]

Thanks for the link. I'd rather not just turn the warning off, since it can find real bugs elsewhere.

Oddly, I can't reproduce with gcc 5.3 or 6.1 on linux, nor with cross-mingw32-gcc 6.1 on linux with the opus autotools build, and not with gcc 6.0 on linux with the spearmint build (although I do see array-bounds warnings for several other spearmint submodules). I also don't see any differences between the silk code in the opus master branch verses the spearmint fork which would change the array bound calculation.

[rawr51919]

He reverted the changes explained here because it broke compilation on Mac OS X 10.6.8.