Re: [PATCH] Fix PR preprocessor/15167

[Zack Weinberg <zack at codesourcery dot com>]

Eric Botcazou <ebotcazou@libertysurf.fr> writes:

> Hi,
>
> This is a regression on the mainline and 3.4 branch related to #pragma once.  
> The compiler segfaults when a #pragma-onced header is included from two 
> different directories (e.g. parent and child directory).
>
> The segfault occurs because f->buffer is NULL when memcmp is invoked in 
> should_stack_file here
>
>       if ((import || f->once_only)
> 	  && f->err_no == 0
> 	  && f->st.st_mtime == file->st.st_mtime
> 	  && f->st.st_size == file->st.st_size
> 	  && read_file (pfile, f)
> 	  /* Size might have changed in read_file().  */
> 	  && f->st.st_size == file->st.st_size
> 	  && !memcmp (f->buffer, file->buffer, f->st.st_size))
> 	break;
>
> although read_file was invoked on f.  The problem originates in 
> _cpp_pop_file_buffer, which doesn't reset f->buffer_valid.

I think f->buffer_valid is supposed to be reset by _cpp_stack_file.
How did it get this far with buffer_valid true?

> Bootstrapped/regtested on amd64-mandrake-linux-gnu.  No testcase
> because it is made up of 5 files.  OK for mainline and 3.4 branch?

You're allowed to create a test case made up of five files (and/or
directories) if you need to.  Several tests in gcc.dg/cpp are already
like that.  And it would help me review this if you showed an example,
so that I could see how this was happening.

zw