This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] PR lto/61048 Define missed builtins on demand


On Wed, Oct 15, 2014 at 10:18 AM, Ilya Palachev <i.palachev@samsung.com> wrote:
> On 15.10.2014 10:59, Jakub Jelinek wrote:
>>
>> On Tue, Oct 14, 2014 at 08:39:40PM +0400, Ilya Palachev wrote:
>>>
>>> Attached patch fixes PR lto/61048 -
>>> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61048
>>
>>
>> Given that the patch just replaces an ICE with a necessary link failure,
>> I'd say
>> it is done at the wrong place, instead during the LTO option handling you
>> should error out if there are incompatibilities in -fsanitize options
>> (any object compiled with flag_sanitize & SANITIZE_USER_ADDRESS, but
>> link done without that, ditto for SANITIZE_KERNEL_ADDRESS,
>> SANITIZE_THREAD.
>> And finally if flag_sanitize & (SANITIZE_UNDEFINED | SANITIZE_NONDEFALT)
>> is non-zero but during linking it is zero (it doesn't really matter which
>> exact undefined sanitization options are used at what time).
>
>
> As I understand you suggest that the compiler should:
> - Write compiler options with which the object file was compiled to the ELF
> format in a special section.
> - Read these sections during the linking command
> - Compare collected sets of options with options that were set at linking
> command.
> ?
>
> Does compiler have any opportunity to save the set of options with which the
> object file was compiled inside some special ELF section? For LTO they the
> special section ".gnu.lto_.opts" is used. But not all compiler options are
> written to it:
>
> $ gcc -fsanitize=address -c test.cpp -o test.o -flto
> $ objdump --full-contents --section=.gnu.lto_.opts test.o
>
> test.o:     file format elf64-x86-64
>
> Contents of section .gnu.lto_.opts:
>  0000 272d6665 78636570 74696f6e 73272027  '-fexceptions' '
>  0010 2d666d61 74682d65 72726e6f 2720272d  -fmath-errno' '-
>  0020 66736967 6e65642d 7a65726f 73272027  fsigned-zeros' '
>  0030 2d667472 61707069 6e672d6d 61746827  -ftrapping-math'
>  0040 20272d66 6e6f2d74 72617076 2720272d   '-fno-trapv' '-
>  0050 666e6f2d 73747269 63742d6f 76657266  fno-strict-overf
>  0060 6c6f7727 20272d6d 74756e65 3d67656e  low' '-mtune=gen
>  0070 65726963 2720272d 6d617263 683d7838  eric' '-march=x8
>  0080 362d3634 2720272d 666c746f 2700      6-64' '-flto'.
>
> As you can see there is no "-fsanitize" option here. Only options that are
> handled in function lto_write_options in lto-opts.c are written there.
>
> Does gcc has an opportunity to write something like ELF section ".gnu.opts"
> to save all compiler options with which the object file was compiled?
> Or we need to introduce such section? Or anything else?

You need to handle them in lto-opts.c and output them to the existing option
section.  Not sure why they don't appear there already - ah, because of

      /* Also drop all options that are handled by the driver as well,
         which includes things like -o and -v or -fhelp for example.
         We do not need those.  Also drop all diagnostic options.  */
      if (cl_options[option->opt_index].flags & (CL_DRIVER|CL_WARNING))
        continue;

and -fsanitize= being marked as Driver.  Looks like we have to drop
things explicitely here given that -o is also Common Driver.  Or simply
have

      if (cl_options[option->opt_index].flags & (CL_DRIVER|CL_WARNING))
           switch (option->opt_index)
             {
             case OPT_fsanitize_:
                break;
             default:
                continue;
             }

instead.

Richard.

>>
>> BTW, in your patches please watch formatting, you didn't use space before
>> (.
>
>
> Sorry for that. I'll check it in further patches.
>
>>
>>         Jakub
>>
>


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]