Fix -fprofile-use miscompilation of tramp3d and read-rtl.c

Fix -fprofile-use miscompilation of tramp3d and read-rtl.c

[Jan Hubicka]

Hi,
this patch fixes tuplification bug that makes us to compile working
tramp3d with -fprofile-use again. I also fixed one uninitialized var
warking showing during profiledbootstrap.  We now build -fprofile-use
cc1 that ICEs shortly in compiling libgcc on memory corruption.
Otherwise it seems to work  and produce same output as non-profiled.

It is about 8% faster compiling GCC modules than unprofiled compiler.
Binary shrinks from 940Kb to 900Kb as a result of optimize_*_for_size
changes that is IMO quite nice.

Interestingly the largest speedups we used to have on empty files seems
gone (at least compiling 100 empty files takes about same time).  I
think it is result of optimizing initialization of register move tables
and friends so we are now fully memory bound starting up.

Honza

	* value-prof.c (gimple_ic): Fix tuplification bug.
	* sched-deps.c (sched_insns_conditions_mutex_p): Silence unitialized var
	warning.
Index: value-prof.c
===================================================================
*** value-prof.c	(revision 140145)
--- value-prof.c	(working copy)
*************** gimple_ic (gimple stmt, gimple call, str
*** 1117,1124 ****
    bb1end = stmt3;
  
    stmt1 = gimple_copy (stmt);
!   gimple_call_set_fn (stmt,
! 		      build_addr (direct_call->decl, current_function_decl));
    gsi_insert_before (&gsi, stmt1, GSI_SAME_STMT);
    bb2end = stmt1;
    bb3end = stmt;
--- 1117,1123 ----
    bb1end = stmt3;
  
    stmt1 = gimple_copy (stmt);
!   gimple_call_set_fndecl (stmt1, direct_call->decl);
    gsi_insert_before (&gsi, stmt1, GSI_SAME_STMT);
    bb2end = stmt1;
    bb3end = stmt;
Index: sched-deps.c
===================================================================
*** sched-deps.c	(revision 140145)
--- sched-deps.c	(working copy)
*************** bool
*** 527,533 ****
  sched_insns_conditions_mutex_p (const_rtx insn1, const_rtx insn2)
  {
    rtx cond1, cond2;
!   bool rev1, rev2;
  
    /* df doesn't handle conditional lifetimes entirely correctly;
       calls mess up the conditional lifetimes.  */
--- 527,533 ----
  sched_insns_conditions_mutex_p (const_rtx insn1, const_rtx insn2)
  {
    rtx cond1, cond2;
!   bool rev1 = false, rev2 = false;
  
    /* df doesn't handle conditional lifetimes entirely correctly;
       calls mess up the conditional lifetimes.  */

Re: Fix -fprofile-use miscompilation of tramp3d and read-rtl.c

[Richard Guenther]

On Wed, Sep 10, 2008 at 11:48 AM, Jan Hubicka <jh@suse.cz> wrote:
> Hi,
> this patch fixes tuplification bug that makes us to compile working
> tramp3d with -fprofile-use again. I also fixed one uninitialized var
> warking showing during profiledbootstrap.  We now build -fprofile-use
> cc1 that ICEs shortly in compiling libgcc on memory corruption.
> Otherwise it seems to work  and produce same output as non-profiled.
>
> It is about 8% faster compiling GCC modules than unprofiled compiler.
> Binary shrinks from 940Kb to 900Kb as a result of optimize_*_for_size
> changes that is IMO quite nice.
>
> Interestingly the largest speedups we used to have on empty files seems
> gone (at least compiling 100 empty files takes about same time).  I
> think it is result of optimizing initialization of register move tables
> and friends so we are now fully memory bound starting up.

Ok.

Thanks,
Richard.

> Honza
>
>        * value-prof.c (gimple_ic): Fix tuplification bug.
>        * sched-deps.c (sched_insns_conditions_mutex_p): Silence unitialized var
>        warning.
> Index: value-prof.c
> ===================================================================
> *** value-prof.c        (revision 140145)
> --- value-prof.c        (working copy)
> *************** gimple_ic (gimple stmt, gimple call, str
> *** 1117,1124 ****
>    bb1end = stmt3;
>
>    stmt1 = gimple_copy (stmt);
> !   gimple_call_set_fn (stmt,
> !                     build_addr (direct_call->decl, current_function_decl));
>    gsi_insert_before (&gsi, stmt1, GSI_SAME_STMT);
>    bb2end = stmt1;
>    bb3end = stmt;
> --- 1117,1123 ----
>    bb1end = stmt3;
>
>    stmt1 = gimple_copy (stmt);
> !   gimple_call_set_fndecl (stmt1, direct_call->decl);
>    gsi_insert_before (&gsi, stmt1, GSI_SAME_STMT);
>    bb2end = stmt1;
>    bb3end = stmt;
> Index: sched-deps.c
> ===================================================================
> *** sched-deps.c        (revision 140145)
> --- sched-deps.c        (working copy)
> *************** bool
> *** 527,533 ****
>  sched_insns_conditions_mutex_p (const_rtx insn1, const_rtx insn2)
>  {
>    rtx cond1, cond2;
> !   bool rev1, rev2;
>
>    /* df doesn't handle conditional lifetimes entirely correctly;
>       calls mess up the conditional lifetimes.  */
> --- 527,533 ----
>  sched_insns_conditions_mutex_p (const_rtx insn1, const_rtx insn2)
>  {
>    rtx cond1, cond2;
> !   bool rev1 = false, rev2 = false;
>
>    /* df doesn't handle conditional lifetimes entirely correctly;
>       calls mess up the conditional lifetimes.  */
>