alpha combine/schedule conflict

alpha combine/schedule conflict

[Richard Henderson]

The attached code produces with -O2

  ;; Start of basic block 10, registers live: 15 [$15] 30 [$30] 63 [FP] 68 69 72
  (insn 553 145 187 (use (reg:DI 133)) -1 (nil)
      (expr_list:REG_DEAD (reg:DI 133)
          (nil)))

emitted from 

#2  0x1201a3048 in distribute_notes () at ../../egcs/gcc/combine.c:11257
#3  0x12017bd84 in try_combine () at ../../egcs/gcc/combine.c:2260
#4  0x120176300 in combine_instructions () at ../../egcs/gcc/combine.c:597

There are no other references to reg 133.  Schedule then aborts at

#1  0x1201f4498 in schedule_block () at ../../egcs/gcc/sched.c:3443

because it has a REG_DEAD note it doesn't know what to do with.

What is the correct fix -- teach distribute_notes not to emit the
use if there are no other references to the reg at all, or to teach
schedule_block not to barf?


r~

Re: alpha combine/schedule conflict

[Jeffrey A Law]

  In message < 19971027142124.62972@dot.cygnus.com >you write:
  > 
  > --lrZ03NoBR/3+SXJZ
  > Content-Type: text/plain; charset=us-ascii
  > 
  > The attached code produces with -O2
  > 
  >   ;; Start of basic block 10, registers live: 15 [$15] 30 [$30] 63 [FP] 68 
  > 69 72
  >   (insn 553 145 187 (use (reg:DI 133)) -1 (nil)
  >       (expr_list:REG_DEAD (reg:DI 133)
  >           (nil)))
  > 
  > emitted from 
  > 
  > #2  0x1201a3048 in distribute_notes () at ../../egcs/gcc/combine.c:11257
  > #3  0x12017bd84 in try_combine () at ../../egcs/gcc/combine.c:2260
  > #4  0x120176300 in combine_instructions () at ../../egcs/gcc/combine.c:597
  > 
  > There are no other references to reg 133.  Schedule then aborts at
  > 
  > #1  0x1201f4498 in schedule_block () at ../../egcs/gcc/sched.c:3443
  > 
  > because it has a REG_DEAD note it doesn't know what to do with.
  > 
  > What is the correct fix -- teach distribute_notes not to emit the
  > use if there are no other references to the reg at all, or to teach
  > schedule_block not to barf?
This may be related to a problem Jim tried to fix earlier.

Look for "optimization bug" in the Sept archives.  There should be messages
from Weiwen Liu, Jim Wilson & myself.  Jim had a patch, but apparently it
didn't work.

Jeff

Re: alpha combine/schedule conflict

[Richard Henderson]

On Mon, Oct 27, 1997 at 09:27:50PM -0800, Jim Wilson wrote:
> This is a known problem.  Combine is emitting a lot of USE insns that it
> should not be.  I tried coming up with a patch, but it is known to be
> incomplete/incorrect.

It looks to me that it solved the USE problem in the test case, but
exposed a new bug also related to REG_DEAD notes.  For the test case
I posted at the beginning of this thread I get

(insn 201 199 202 (set (reg/v:DI 71)
        (const_int -1)) 252 {movdi-1} (nil)
    (nil))

(insn 202 201 205 (set (reg/v:DI 71)
        (plus:DI (reg/v:DI 71)
            (const_int 536870912))) 7 {adddi3} (insn_list 201 (nil))
    (expr_list:REG_DEAD (reg/v:DI 71)
        (expr_list:REG_EQUAL (const_int 536870911)
            (nil))))

[...]

(insn 209 207 210 (set (reg:DI 137)
        (leu:DI (reg:DI 135)
            (reg/v:DI 71))) 131 {subdf3+4} (insn_list 202 (insn_list 207 (nil)))
    (expr_list:REG_DEAD (reg/v:DI 71)
        (nil)))

It appears to be simply due to incomplete checking for elim_i1.  The
attached patch (combined with yours because I'm lazy) solves the problem
for the test case.  A sanity three-stage is still going, but I'll let
you know.


r~

Re: alpha combine/schedule conflict

[Jeffrey A Law]

  In message <19971028111109.02890@dot.cygnus.com>you write:
  > 
  > Tue Oct 28 01:56:39 1997  Richard Henderson  <rth@cygnus.com>
  > 
  > 	* combine.c (try_combine): Don't elim_i1 when it is used in i2src.
  > 
  > Tue Sep 23 12:57:35 1997  Jim Wilson  <wilson@cygnus.com>
  > 
  >         * combine.c (try_combine): When setting elim_i2, check whether newi2pat
  >         sets i2dest.  When calling distribute_notes for i3dest_killed, pass
  >         elim_i2 and elim_i1.
Looks good to me.  Installed.

jeff

Re: alpha combine/schedule conflict

[Richard Henderson]

On Sun, Nov 02, 1997 at 11:29:04PM -0700, Jeffrey A Law wrote:
> Looks good to me.  Installed.

Actually, that patch isn't; it merely aborts under different
conditions.  I've been talking with Jim privately on this one;
we have another patch that seems to work but I was still trying
it on things.

I'll let you know how it turns out tomorrow.


r~

Re: alpha combine/schedule conflict

[Jeffrey A Law]

  In message < 19971102233429.46384@dot.cygnus.com >you write:
  > On Sun, Nov 02, 1997 at 11:29:04PM -0700, Jeffrey A Law wrote:
  > > Looks good to me.  Installed.
  > 
  > Actually, that patch isn't; it merely aborts under different
  > conditions.  I've been talking with Jim privately on this one;
  > we have another patch that seems to work but I was still trying
  > it on things.
  > 
  > I'll let you know how it turns out tomorrow.
Yow!  Well, we're waiting for the updated patch :-)

jeff