| Message ID | CAOvf_xwJ85txuJyEYw99A+FV0k6y7VS2a34zaG0xLOX=+5jXDg@mail.gmail.com |
|---|---|
| State | New |
| Headers | show |
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$_GLOBAL_OFFSET_TABLE_, %%ebx\n"); > #ifndef NO_PROFILE_COUNTERS > fprintf (file, "\tleal\t%sP%d@GOTOFF(%%ebx),%%" > PROFILE_COUNT_REGISTER "\n", > LPREFIX, labelno); > #endif > fprintf (file, "1:\tcall\t*%s@GOT(%%ebx)\n", mcount_name); > + fprintf (file,"\tpop\t%%ebx\n"); > } > else > { Jakub
diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c index a3ca2ed..5117572 100644 --- a/gcc/config/i386/i386.c +++ b/gcc/config/i386/i386.c @@ -39119,11 +39126,15 @@ x86_function_profiler (FILE *file, int labelno ATTRIBUTE_UNUSED) } else if (flag_pic) { + fprintf (file,"\tpush\t%%ebx\n"); + fprintf (file,"\tcall\t__x86.get_pc_thunk.bx\n"); + fprintf (file,"\taddl\t$_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"); }
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? ChangeLog 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. else {