[Bug rtl-optimization/18577] New: [3.3 regression] variable use moved before initialization

[Bug rtl-optimization/18577] New: [3.3 regression] variable use moved before initialization

[falk at debian dot org]

Test case:

static float tfcos12[3];
__attribute__((noinline)) double f(double x) { return x; }
int g;
int main(void) {
    int i, j;
    for (i = 0; i < 1; i++) 
	tfcos12[i] = 0.5;
    
    for (i = 0; i < 1; i++) {
	tfcos12[i] = 0.5 * f(i);
	for (j = 0; j < 12; j++)
	    g++;
    }
    return 0;
}

% gcc -funroll-all-loops -O2 test.c && ./a.out
zsh: floating point exception (core dumped)  ./a.out

The reason is that the constant 0.5 is first used and then loaded:

	[...]
	cpys $f10,$f10,$f2				# use $f10
	lds $f10,$LC0($1)		!gprellow	# before loading it
	[...]
	mult $f2,$f0,$f0				# and this FPEs

This does not happen without -funroll-all-loops. Also not at -O, or with
3.4.

-- 
           Summary: [3.3 regression] variable use moved before
                    initialization
           Product: gcc
           Version: 3.3.5
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: critical
          Priority: P2
         Component: rtl-optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: falk at debian dot org
                CC: gcc-bugs at gcc dot gnu dot org
 GCC build triplet: alphaev68-unknown-linux-gnu
  GCC host triplet: alphaev68-unknown-linux-gnu
GCC target triplet: alphaev68-unknown-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18577

[Bug rtl-optimization/18577] [3.3 regression] variable use moved before initialization

[pinskia at gcc dot gnu dot org]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |3.3.6


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18577

[Bug rtl-optimization/18577] [3.3 regression] variable use moved before initialization

[falk at debian dot org]

------- Additional Comments From falk at debian dot org  2004-11-21 02:33 -------
The bug seems to be caused by the loop unroller making a pseudo local
to be able to duplicate it while it is still needed outside of the loop.
Reverting Eric Botcazou's patch for rtl-optimization/11841, which is
also about this, fixes the problem. I don't really understand what's going
on, but is it possible we still have to look at the pseudo luids and not
only the note luids? Like this?

diff -u -p -r1.184.2.9 unroll.c
--- unroll.c    17 May 2004 21:05:48 -0000      1.184.2.9
+++ unroll.c    21 Nov 2004 02:30:41 -0000
@@ -794,6 +794,8 @@ unroll_loop (loop, insn_count, strength_
       for (r = FIRST_PSEUDO_REGISTER; r < max_reg_before_loop; ++r)
        if (REGNO_FIRST_UID (r) > 0 && REGNO_FIRST_UID (r) < max_uid_for_loop
            && REGNO_FIRST_LUID (r) >= copy_start_luid
+           && REGNO_LAST_UID (r) > 0 && REGNO_LAST_UID (r) < max_uid_for_loop
+           && REGNO_LAST_LUID (r) <= copy_end_luid
            && REGNO_LAST_NOTE_UID (r) > 0 && REGNO_LAST_NOTE_UID (r) <
max_uid_for_loop
            && REGNO_LAST_NOTE_LUID (r) <= copy_end_luid)
          {


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |ebotcazou at gcc dot gnu dot
                   |                            |org
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|                            |1


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18577

[Bug rtl-optimization/18577] [3.3 regression] variable use moved before initialization

[ebotcazou at gcc dot gnu dot org]

------- Additional Comments From ebotcazou at gcc dot gnu dot org  2004-11-21 07:14 -------
> The bug seems to be caused by the loop unroller making a pseudo local
> to be able to duplicate it while it is still needed outside of the loop.
> Reverting Eric Botcazou's patch for rtl-optimization/11841, which is
> also about this, fixes the problem.

Thanks for investigating.

> I don't really understand what's going on, but is it possible we still have to 
> look at the pseudo luids and not only the note luids? Like this?

IIRC note luids are supposed to be a superset of regular luids, so your check
should be redundant (see regclass.c:reg_scan_mark_refs).  Now something could
have invalidated this relationship.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|ebotcazou at gcc dot gnu dot|
                   |org                         |
         AssignedTo|unassigned at gcc dot gnu   |ebotcazou at gcc dot gnu dot
                   |dot org                     |org
             Status|NEW                         |ASSIGNED


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18577

[Bug rtl-optimization/18577] [3.3 regression] variable use moved before initialization

[ebotcazou at gcc dot gnu dot org]

------- Additional Comments From ebotcazou at gcc dot gnu dot org  2004-11-21 09:07 -------
> IIRC note luids are supposed to be a superset of regular luids, so your check
> should be redundant (see regclass.c:reg_scan_mark_refs).  Now something could
> have invalidated this relationship.

It's this chunk of code in load_mems:

	  set = gen_move_insn (reg, best);
	  set = loop_insn_hoist (loop, set);
	  if (REG_P (best))
	    {
	      for (p = prev_ebb_head; p != loop->start; p = NEXT_INSN (p))
		if (REGNO_LAST_UID (REGNO (best)) == INSN_UID (p))
		  {
		    REGNO_LAST_UID (REGNO (best)) = INSN_UID (set);
		    break;
		  }
	    }

which changes REGNO_LAST_UID without updating REGNO_LAST_NOTE_UID for pseudo 78.
 Note that there is no note mentioning pseudo 78 so REGNO_LAST_NOTE_UID points
to garbage after it is executed.

We could probably update REGNO_LAST_NOTE_UID too here, but there might be other
places with the same problem so I think your proposed fix is the safest one. 
I'm going to test it on x86.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18577

[Bug rtl-optimization/18577] [3.3 regression] variable use moved before initialization

[cvs-commit at gcc dot gnu dot org]

------- Additional Comments From cvs-commit at gcc dot gnu dot org  2004-11-27 16:59 -------
Subject: Bug 18577

CVSROOT:	/cvs/gcc
Module name:	gcc
Branch: 	gcc-3_3-branch
Changes by:	ebotcazou@gcc.gnu.org	2004-11-27 16:59:18

Modified files:
	gcc            : ChangeLog regclass.c regs.h unroll.c 
Added files:
	gcc/testsuite/gcc.dg: loop-6.c 

Log message:
	PR optimization/18577
	* unroll.c (unroll_loop): Test both REGNO_LAST_UID and
	REGNO_LAST_NOTE_UID to decide whether a pseudo is local
	to the loop.

Patches:
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/ChangeLog.diff?cvsroot=gcc&only_with_tag=gcc-3_3-branch&r1=1.16114.2.1027&r2=1.16114.2.1028
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/regclass.c.diff?cvsroot=gcc&only_with_tag=gcc-3_3-branch&r1=1.160.4.5&r2=1.160.4.6
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/regs.h.diff?cvsroot=gcc&only_with_tag=gcc-3_3-branch&r1=1.26.4.1&r2=1.26.4.2
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/unroll.c.diff?cvsroot=gcc&only_with_tag=gcc-3_3-branch&r1=1.184.2.9&r2=1.184.2.10
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/testsuite/gcc.dg/loop-6.c.diff?cvsroot=gcc&only_with_tag=gcc-3_3-branch&r1=NONE&r2=1.1.4.1



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18577

[Bug rtl-optimization/18577] [3.3 regression] variable use moved before initialization

[ebotcazou at gcc dot gnu dot org]

------- Additional Comments From ebotcazou at gcc dot gnu dot org  2004-11-27 17:05 -------
See http://gcc.gnu.org/ml/gcc-patches/2004-11/msg02385.html


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18577