[PATCH, x86, 63534] Fix '-p' profile for 32 bit PIC mode

[Evgeny Stupachenko]

The following patch align stack for mcount and there should be no
problems with unwind as ix86_frame_pointer_required is true when
crtl->profile is true and flag_fentry is false (we call mcount after
function prolog).
When flag_fentry is true it is set to false in 32bit PIC mode:
  if (!TARGET_64BIT_P (opts->x_ix86_isa_flags) && opts->x_flag_pic)
    {
      if (opts->x_flag_fentry > 0)
        sorry ("-mfentry isn%'t supported for 32-bit in combination "
      "with -fpic");
      opts->x_flag_fentry = 0;
    }



2014-10-24  Evgeny Stupachenko  <evstupac@gmail.com>

        PR target/63534
        * config/i386/i386.c (x86_function_profiler): Add GOT register init
        for mcount call.

diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index 6235c4f..2dff29c 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -39124,13 +39124,22 @@ x86_function_profiler (FILE *file, int
labelno ATTRIBUTE_UNUSED)
       else
        x86_print_call_or_nop (file, mcount_name);
     }
+  /* At this stage we can't detrmine where GOT register is, as RA can allocate
+     it to any hard register.  Therefore we need to set it once again.  */
   else if (flag_pic)
     {
+      pic_labels_used |= 1 << BX_REG;
+      fprintf (file,"\tsub\t$16, %%esp\n");
+      fprintf (file,"\tmovl\t%%ebx, (%%esp)\n");
+      fprintf (file,"\tcall\t__x86.get_pc_thunk.bx\n");
+      fprintf (file,"\taddl\t$_GLOBAL_OFFSET_TABLE_, %%ebx\n");
 #ifndef NO_PROFILE_COUNTERS
       fprintf (file, "\tleal\t%sP%d@GOTOFF(%%ebx),%%"
PROFILE_COUNT_REGISTER "\n",
               LPREFIX, labelno);
 #endif
       fprintf (file, "1:\tcall\t*%s@GOT(%%ebx)\n", mcount_name);
+      fprintf (file,"\tmovl\t(%%esp), %%ebx\n");
+      fprintf (file,"\tadd\t$16, %%esp\n");
     }
   else
     {

On Fri, Oct 17, 2014 at 6:38 PM, Jakub Jelinek <jakub@redhat.com> wrote:
> On Fri, Oct 17, 2014 at 06:30:42PM +0400, Evgeny Stupachenko wrote:
>> Hi,
>>
>> The patch fixes profile in 32bits PIC mode (only -p option affected).
>>
>> x86 bootstrap, make check passed
>>
>> spec2000 o2 -p train data on Corei7:
>> CINT -5%
>> CFP  +1,5
>> compared to a compiler before "enabling ebx".
>>
>> There is a potential performance improve after the patch applied
>> suggested by Jakub:
>> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63534#c8
>> There is opened bug on this: PR63527. However the fix of the bug is
>> more complicated.
>>
>> Is it ok?
>
> Unfortunately I don't think it is ok.
> 1) you don't set the appropriate bit in pic_labels_used (for ebx)
> 2) more importantly, it causes the stack to be misaligned (i.e. violating
>    ABI) for the _mcount call, and, break unwind info.
>
>> 2014-10-16  Evgeny Stupachenko  <evstupac@gmail.com>
>>
>>         PR target/63534
>>         * config/i386/i386.c (x86_function_profiler): Add GOT register init
>>         for mcount call.
>>
>> diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
>> index a3ca2ed..5117572 100644
>> --- a/gcc/config/i386/i386.c
>> +++ b/gcc/config/i386/i386.c
>> @@ -39119,11 +39126,15 @@ x86_function_profiler (FILE *file, int
>> labelno ATTRIBUTE_UNUSED)
>>      }
>>    else if (flag_pic)
>>      {
>> +      fprintf (file,"\tpush\t%%ebx\n");
>> +      fprintf (file,"\tcall\t__x86.get_pc_thunk.bx\n");
>> +      fprintf (file,"\taddl\t$_GLOBAL_OFFSET_TABLE_, %%ebx\n");
>>  #ifndef NO_PROFILE_COUNTERS
>>        fprintf (file, "\tleal\t%sP%d@GOTOFF(%%ebx),%%"
>> PROFILE_COUNT_REGISTER "\n",
>>                LPREFIX, labelno);
>>  #endif
>>        fprintf (file, "1:\tcall\t*%s@GOT(%%ebx)\n", mcount_name);
>> +      fprintf (file,"\tpop\t%%ebx\n");
>>      }
>>    else
>>      {
>
>         Jakub