[Bug optimization/9240] wierd scheduling on v850 unless -fno-sched-spec specified

[Bug optimization/9240] wierd scheduling on v850 unless -fno-sched-spec specified

[pinskia@physics.uc.edu]

PLEASE REPLY TO gcc-bugzilla@gcc.gnu.org ONLY, *NOT* gcc-bugs@gcc.gnu.org.

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=9240


pinskia@physics.uc.edu changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |NEW
     Ever Confirmed|                            |1
   Last reconfirmed|0000-00-00 00:00:00         |2003-06-02 04:29:26
               date|                            |


------- Additional Comments From pinskia@physics.uc.edu  2003-06-02 04:29 -------
Acording to the submitter the problem is still there and looks like -fno-sched-spec fixes 
the problem.



------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug optimization/9240] weird scheduling on v850 unless -fno-sched-spec specified

[dhazeghi at yahoo dot com]

PLEASE REPLY TO gcc-bugzilla@gcc.gnu.org ONLY, *NOT* gcc-bugs@gcc.gnu.org.

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=9240


dhazeghi at yahoo dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|3.4                         |---

[Bug optimization/9240] weird scheduling on v850 unless -fno-sched-spec specified

[vmakarov at redhat dot com]

------- Additional Comments From vmakarov at redhat dot com  2004-01-15 17:22 -------
Some analysis:

1. Scheduler uses own algorithm of probablity calculation (3 basic
   blocks of the test have probability 100/50/100%).

2. It should use common infrastructure bb frequency and edge
   probablility.

3. Even if it used the common infrastructure, bb frequnecies would be
   still wrong (10000, 9001, 10000 or 100/90/100%).  So builtin-expect
   is ignored in the infrastructure too.  I saw broken builtin-expect
   many times during my 6 years work on gcc.

4. Becuase the probability of the 2nd bb is 50%, it includes 2
   insns from it to schedule into the 1st bb.

So fixing the problem could be:

  1. Taking builtin_expect into account in scheduler algorithm of
     probablity calculation.

  2. Start using the common infrastructure bb frequency and edge
     probablility and fixing there builtin_expect.

  Neither the 1st solution nor the 2nd one is easy. And they include
the risk of new bugs.  So I think it is not reasonable to fix it for
3.4.  It is a just generation of non-optimal code (scheduler works by
heuristics.  there will be always some unoptimal schedules.  we are
just oriented to guarantee that scheduler improves code in overall on
credible benchmarks like Spec).  The problem was pressent in 3.3.

  I'll try to fix the problem in 3-4 months.

Vlad


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |vmakarov at redhat dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=9240

[Bug rtl-optimization/9240] weird scheduling on v850 unless -fno-sched-spec specified

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2005-09-26 15:21 -------
I think the patch at
http://gcc.gnu.org/ml/gcc-patches/2005-09/msg00370.html
will fix this.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=9240

[Bug rtl-optimization/9240] weird scheduling on v850 unless -fno-sched-spec specified

[steinmtz at us dot ibm dot com]

------- Additional Comments From steinmtz at us dot ibm dot com  2005-09-26 22:07 -------
Unless the common infrastructure bb frequency and edge probabilities have been 
updated to reflect builtin_expect, http://gcc.gnu.org/ml/gcc-patches/2005-
09/msg00370.html probably won't fix this problem.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=9240

[Bug rtl-optimization/9240] weird scheduling on v850 unless -fno-sched-spec specified

[law at redhat dot com]

------- Comment #8 from law at redhat dot com  2010-02-24 21:20 -------
This was fixed at some point; I don't know if it was the patch referenced in
comment #6, or some other patch or some combination thereof.

You can see that probabilities are being propagated down through to the
scheduler by looking at the .sched dump:

(insn 9 7 10 2 j.c:8 (set (cc0)
        (compare (reg:SI 39 [ prephitmp___4 ])
            (reg:SI 42))) 8 {*cmpsi} (expr_list:REG_DEAD (reg:SI 42)
        (expr_list:REG_EQUAL (compare (reg:SI 39 [ prephitmp___4 ])
                (const_int -125 [0xffffffffffffff83]))
            (nil))))

(jump_insn 10 9 11 2 j.c:8 (set (pc)
        (if_then_else (lt (cc0)
                (const_int 0 [0x0]))
            (label_ref 15)
            (pc))) 41 {*branch_normal} (expr_list:REG_BR_PROB (const_int 9996
[0x270c])
        (nil))
 -> 15)
;; End of basic block 2 -> ( 3 4)
;; lr  out       3 [sp] 29 [r29] 32 [.fp] 33 [.ap] 39
;; live  out     3 [sp] 29 [r29] 32 [.fp] 33 [.ap] 39


;; Succ edge  3 [0.0%]  (fallthru)
;; Succ edge  4 [100.0%]

Note how the probability of the edge 2->3 is now zero, reflecting the unlikely
attribute while the edge 2->4 is 100%.  With the correct probabilities the
scheduler naturally does the right thing.


-- 

law at redhat dot com changed:

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


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=9240