Re: ‘void free(void*)’ called on unallocated object ‘yyssa’ [-Wfree-nonheap-object]

[Akim Demaille]

[Paul, at the bottom of this message we have a warning from GCC11
that I guess you already had to keep quiet.]


Hi Christoph,

> Le 18 janv. 2021 à 22:01, Christoph Grüninger <foss@grueninger.de> a écrit :
> 
> Dear Bison community,
> while compiling CMake using GCC 11 (not yet released), I came across several 
> -Wfree-nonheap-object warnings. They came from a file generated by GNU Bison 
> 3.4.2. [1]
> 
> I was hoping the error might be fixed in a newer version, so
> 1. I downloaded your latest release 3.7.4
> 2. Re-generated the file cmCommandArgumentParser.y with
>     bison --yacc --name-prefix=cmCommandArgument_yy 
> --defines=cmCommandArgumentParserTokens.h -ocmCommandArgumentParser.cxx 
> cmCommandArgumentParser.y

It makes little sense to use --yacc while using Bison features.  Why
would you do that?  Bison generates warnings for Bison features used
in Yacc mode.


> 3. I removed the lines 1717 and 1745 as the label yyerrlab1 must be defined.

Could you please clarify what you are doing here, and why?


> 4. I recompiled CMake and get the following error:
>    Scanning dependencies of target CMakeLib
>    [ 54%] Building CXX object 
> Source/CMakeFiles/CMakeLib.dir/LexerParser/cmCommandArgumentParser.cxx.o
>    cmCommandArgumentParser.cxx: In function ‘int 
> cmCommandArgument_yyparse(yyscan_t)’:
>    cmCommandArgumentParser.cxx:1838:18: warning: ‘void free(void*)’ called on 
> unallocated object ‘yyssa’ [-Wfree-nonheap-object]
>    cmCommandArgumentParser.cxx:1203:16: note: declared here
> 
> I attached both the .y and he generated .cxx file.
> 
> Is this a known issue?

Nope.


> Is the GCC 11 warning wrong (false positive)? I assume not, as yyssa is an 
> array:
>    yy_state_t yyssa[YYINITDEPTH];
> 
> Can you change Bison that it will generate code that behaves correctly?

You can't seriously expect code used in the wild for many years to be
calling free on a stack-allocated value...

GCC 11 is wrong here.  The full story (taken from your file) is:

> // Create.
>     yy_state_t yyssa[YYINITDEPTH];
>     yy_state_t *yyss = yyssa;
> [...]
> // Grow.
>       yystacksize *= 2;
>      {
>         yy_state_t *yyss1 = yyss;
>         union yyalloc *yyptr =
>           YY_CAST (union yyalloc *,
>                    YYSTACK_ALLOC (YY_CAST (YYSIZE_T, YYSTACK_BYTES 
> (yystacksize))));
>         if (! yyptr)
>           goto yyexhaustedlab;
>         YYSTACK_RELOCATE (yyss_alloc, yyss);
>         if (yyss1 != yyssa)
>           YYSTACK_FREE (yyss1);
>       }
> [...]
> // Clean up
>   if (yyss != yyssa)
>     YYSTACK_FREE (yyss);

Or, in words, we use a stack-allocated stack until it's too small and
then we use a heap-allocated stack, and in that case, and that case
only, we free it at the end.  Said another way: we never ever call
free(yyssa).

So GCC's warning is a false positive.

This is probably the kind of annoying warnings that Paul knows how
to defeat.  I would use pragmas, but maybe there is something more
clever to do?  Thanks in advance Paul!

Cheers!