[Bug tree-optimization/42889] New: "-fcompare-debug failure (length)" with "-O1 -fgcse"

[Bug tree-optimization/42889] New: "-fcompare-debug failure (length)" with "-O1 -fgcse"

[zsojka at seznam dot cz]

Command line:
gcc -O1 -fgcse -fcompare-debug -c testcase.c

Tested revisions:
r156293 - crash
r155833 - crash
r155363 - OK
r154886 - OK
r153685 - OK

Output:
$ /mnt/svn/gcc-trunk/binary-155833-lto/bin/gcc -O1 -fgcse -fcompare-debug -c
testcase.c
gcc: testcase.c: -fcompare-debug failure (length)

I noticed this because gentoo's weekly gcc-4.5.0_alpha builds started failing
with 20100107, last working 20091231, at both x86 and x86_64.
I reproduced bootstrap failure when configure with --disable-checking, fails at
builtins.o (there is actually different code being emited).


-- 
           Summary: "-fcompare-debug failure (length)" with "-O1 -fgcse"
           Product: gcc
           Version: 4.5.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: zsojka at seznam dot cz
  GCC host triplet: x86_64-pc-linux-gnu
GCC target triplet: x86_64-pc-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug tree-optimization/42889] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[zsojka at seznam dot cz]

------- Comment #1 from zsojka at seznam dot cz  2010-01-28 00:26 -------
Created an attachment (id=19733)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=19733&action=view)
reduced testcase

Reduced from gcc/builtins.c

Command line:
gcc -O1 -fgcse -fcompare-debug -c pr42889.c


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug tree-optimization/42889] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[steven at gcc dot gnu dot org]

-- 

steven at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         AssignedTo|unassigned at gcc dot gnu   |steven at gcc dot gnu dot
                   |dot org                     |org
             Status|UNCONFIRMED                 |ASSIGNED
     Ever Confirmed|0                           |1
   Last reconfirmed|0000-00-00 00:00:00         |2010-01-28 06:56:42
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug tree-optimization/42889] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[steven at gcc dot gnu dot org]

------- Comment #2 from steven at gcc dot gnu dot org  2010-01-28 09:07 -------
Caused by an RTL PRE transformation.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug tree-optimization/42889] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[steven at gcc dot gnu dot org]

------- Comment #3 from steven at gcc dot gnu dot org  2010-01-28 09:40 -------
combine fails to combine the following insns.

Before combine (.init_regs dump):

(insn 20 19 21 4 PR42889.c:7 (parallel [
            (set (reg:SI 70)
                (and:SI (reg/v:SI 64 [ c ])
                    (const_int 4 [0x4])))
            (clobber (reg:CC 17 flags))
        ]) 375 {*andsi_1} (expr_list:REG_DEAD (reg/v:SI 64 [ c ])
        (expr_list:REG_UNUSED (reg:CC 17 flags)
            (nil))))

(insn 21 20 48 4 PR42889.c:7 (set (reg:CCZ 17 flags)
        (compare:CCZ (reg:SI 70)
            (const_int 0 [0x0]))) 2 {*cmpsi_ccno_1} (expr_list:REG_UNUSED
(reg:CCZ 17 flags)
        (nil)))


After combine (.combine dump):
(note 20 19 21 4 NOTE_INSN_DELETED)

(insn 21 20 48 4 PR42889.c:7 (parallel [
            (set (reg:CCZ 17 flags)
                (compare:CCZ (and:SI (reg/v:SI 64 [ c ])
                        (const_int 4 [0x4]))
                    (const_int 0 [0x0])))
            (set (reg:SI 70)
                (and:SI (reg/v:SI 64 [ c ])
                    (const_int 4 [0x4])))
        ]) 384 {*andsi_2} (expr_list:REG_DEAD (reg/v:SI 64 [ c ])
        (expr_list:REG_UNUSED (reg:CCZ 17 flags)
            (nil))))




When compiled with -g that combine doesn't happen because the compare insn has
been changed somewhere before combine.

Before combine:
(insn 20 19 50 4 PR42889.c:7 (parallel [
            (set (reg:SI 70)
                (and:SI (reg/v:SI 64 [ c ])
                    (const_int 4 [0x4])))
            (clobber (reg:CC 17 flags))
        ]) 375 {*andsi_1} (expr_list:REG_DEAD (reg/v:SI 64 [ c ])
        (expr_list:REG_UNUSED (reg:CC 17 flags)
            (nil))))

(insn 50 20 51 4 PR42889.c:7 (set (reg:CC 17 flags)
        (compare:CC (reg:SI 70)
            (const_int 1 [0x1]))) 6 {*cmpsi_1} (expr_list:REG_DEAD (reg:SI 70)
        (nil)))




The difference starts in the cprop3 pass:

+ DCE: Deleting insn 21
+ deleting insn with uid = 21.

         (expr_list:REG_UNUSED (reg:CC 17 flags)
             (nil))))

-(insn# # # 4 PR42889.c:7 (set (reg:CCZ 17 flags)
-        (compare:CCZ (reg:SI 70)
-            (const_int 0 [0x0])))# {*cmpsi_ccno_1} (expr_list:REG_DEAD (reg:SI
70)
-        (nil)))
-
 (insn# # # 4 PR42889.c:7 (set (reg:CC 17 flags)
         (compare:CC (reg:SI 70)
             (const_int 1 [0x1])))# {*cmpsi_1} (nil))


Interestingly: with -g0 "df_analyze called" happens, but there is no DF output
and it almost looks as if DF simply doesn't run (and therefore also DCE does
not run).


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug rtl-optimization/42889] [4.5 Regression] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[rguenth at gcc dot gnu dot org]

-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|tree-optimization           |rtl-optimization
           Priority|P3                          |P1
            Summary|"-fcompare-debug failure    |[4.5 Regression] "-fcompare-
                   |(length)" with "-O1 -fgcse" |debug failure (length)" with
                   |                            |"-O1 -fgcse"
   Target Milestone|---                         |4.5.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug rtl-optimization/42889] [4.5 Regression] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[steven at gcc dot gnu dot org]

------- Comment #4 from steven at gcc dot gnu dot org  2010-01-28 10:14 -------
The fast DCE is not called because df_lr_finalize is never called with -g0 --
nothing has changed so there is nothing to finalize.

Unassigning. Aleksandre's problem.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug rtl-optimization/42889] [4.5 Regression] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[steven at gcc dot gnu dot org]

------- Comment #5 from steven at gcc dot gnu dot org  2010-01-28 10:17 -------
FWIW, DF rescanning runs with -g1 and not with -g0 because copy propagation
updates a debug_insn:

LOCAL COPY-PROP: Replacing reg 63 in insn 28 with reg 72

(insn 27 26 28 4 PR42889.c:7 (set (reg/v:SI 63 [ x ])
        (reg:SI 72)) 47 {*movsi_1} (expr_list:REG_DEAD (reg:SI 72)
        (nil)))

(debug_insn 28 27 29 4 PR42889.c:7 (var_location:SI x (reg:SI 72)) -1 (nil))


This clobbers the DF solution and triggers a re-run of all the DF problems. And
that, in turn, causes the extra DCE run.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug rtl-optimization/42889] [4.5 Regression] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[rguenth at gcc dot gnu dot org]

------- Comment #6 from rguenth at gcc dot gnu dot org  2010-01-28 11:05 -------
(In reply to comment #5)
> FWIW, DF rescanning runs with -g1 and not with -g0 because copy propagation
> updates a debug_insn:
> 
> LOCAL COPY-PROP: Replacing reg 63 in insn 28 with reg 72
> 
> (insn 27 26 28 4 PR42889.c:7 (set (reg/v:SI 63 [ x ])
>         (reg:SI 72)) 47 {*movsi_1} (expr_list:REG_DEAD (reg:SI 72)
>         (nil)))
> 
> (debug_insn 28 27 29 4 PR42889.c:7 (var_location:SI x (reg:SI 72)) -1 (nil))
> 
> 
> This clobbers the DF solution and triggers a re-run of all the DF problems. And
> that, in turn, causes the extra DCE run.

Ugh.  It shouldn't trigger a DCE run at least.  I guess it has to trigger
a re-run of the DF problems though.

I guess similar things could happen on the tree-level with scheduling
of cleanup_cfg after something changed.  Double-ugh.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rguenth at gcc dot gnu dot
                   |                            |org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug rtl-optimization/42889] [4.5 Regression] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[steven at gcc dot gnu dot org]

-- 

steven at gcc dot gnu dot org changed:

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


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug rtl-optimization/42889] [4.5 Regression] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[steven at gcc dot gnu dot org]

------- Comment #7 from steven at gcc dot gnu dot org  2010-01-28 12:07 -------
Well let's face it: All this happening because of DEBUG_INSN/DEBUG_STMT is deep
in "hate to say I told you so" teritory...


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug rtl-optimization/42889] [4.5 Regression] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[rguenther at suse dot de]

------- Comment #8 from rguenther at suse dot de  2010-01-28 12:09 -------
Subject: Re:  [4.5 Regression] "-fcompare-debug
 failure (length)" with "-O1 -fgcse"

On Thu, 28 Jan 2010, steven at gcc dot gnu dot org wrote:

> ------- Comment #7 from steven at gcc dot gnu dot org  2010-01-28 12:07 -------
> Well let's face it: All this happening because of DEBUG_INSN/DEBUG_STMT is deep
> in "hate to say I told you so" teritory...

True - this happens when all concerns get overturned by
"let's make progress with debuginfo" ...


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug rtl-optimization/42889] [4.5 Regression] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[jakub at gcc dot gnu dot org]

------- Comment #9 from jakub at gcc dot gnu dot org  2010-01-28 13:39 -------
Created an attachment (id=19739)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=19739&action=view)
gcc45-pr42889.patch

Patch that cures this by not running DCE if only DEBUG_INSNs changed.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug rtl-optimization/42889] [4.5 Regression] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[steven at gcc dot gnu dot org]

------- Comment #10 from steven at gcc dot gnu dot org  2010-01-28 13:53 -------
The patch of comment #9 is a hack that doesn't solve anything. There are loads
of other passes that may inadvertedly trigger a fast DCE if only a DEBUG_INSN
is modified and needs to be rescanned.

IMHO the only "proper" fix would be to teach DF not to run DCE (and any other
DF-problem other than scan) if only DEBUG_INSNs are modified.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug rtl-optimization/42889] [4.5 Regression] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[dominiq at lps dot ens dot fr]

------- Comment #11 from dominiq at lps dot ens dot fr  2010-01-28 14:59 -------
Created an attachment (id=19741)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=19741&action=view)
Code giving a suspicious error with the patch.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug rtl-optimization/42889] [4.5 Regression] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[jakub at gcc dot gnu dot org]

------- Comment #12 from jakub at gcc dot gnu dot org  2010-01-28 15:11 -------
So not call df_set_bb_dirty in df_insn_rescan if DEBUG_INSN_P (insn), or not
mark dirty only certain problems?
I believe DEBUG_INSNs aren't meant to extend lifetime of pseudos, so perhaps
lr/live problems don't need to be marked dirty, what about others?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug rtl-optimization/42889] [4.5 Regression] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[dominiq at lps dot ens dot fr]

------- Comment #13 from dominiq at lps dot ens dot fr  2010-01-28 15:37 -------
Wrong pr!-(sorry for the noise).


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug rtl-optimization/42889] [4.5 Regression] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[steven at gcc dot gnu dot org]

------- Comment #14 from steven at gcc dot gnu dot org  2010-01-28 16:28 -------
Re. comment #12: I think both: Do not call df_set_bb_dirty in df_insn_rescan
rescans a DEBUG_INSN, *and* mark dirty only certain problems. The last part is
probably the most tricky bit: which problems to mark. I think the RD and CHAINS
problems, at least, still need updating.

Updating a DF problem doesn't have to be a problem if the update doesn't result
in any code changes. The fact that we run DCE in an LR update is a Bad Thing
that was an unfortunate but necessary hack to fix some PR or another. But in
general the DF problems shouldn't do any code modifications.

I guess it should be documented somewhere that all DF problems except LR should
not even think about modifying any insns...


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug rtl-optimization/42889] [4.5 Regression] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[jakub at gcc dot gnu dot org]

------- Comment #15 from jakub at gcc dot gnu dot org  2010-01-28 17:08 -------
If only LR can make code changes, perhaps we should just avoid marking LR dirty
and make all other problems dirty.  Add df_set_bb_dirty_nonlr and
df_mark_solutions_dirty_nonlr or something like that.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug rtl-optimization/42889] [4.5 Regression] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[jakub at gcc dot gnu dot org]

------- Comment #16 from jakub at gcc dot gnu dot org  2010-01-28 17:23 -------
Created an attachment (id=19743)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=19743&action=view)
gcc45-pr42889.patch

Patch that does that.
BTW, df_ref_create and df_ref_remove guard the df_set_bb_dirty calls with
!DEBUG_INSN_P tests, so perhaps we could just avoid calling df_set_bb_dirty
instead of calling df_set_bb_dirty_nonlr, or on the other side we could call
this new function even in df_ref_create/df_ref_remove.


-- 

jakub at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #19739|0                           |1
        is obsolete|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug rtl-optimization/42889] [4.5 Regression] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[steven at gcc dot gnu dot org]

------- Comment #17 from steven at gcc dot gnu dot org  2010-01-28 18:05 -------
How does this address the 30ish calls to df_set_bb_dirty from the rest of the
compiler? See cfgcleanup, cfgrtl, emit-rtl, ifcvt, and modulo-sched. Each of
them could result in a DCE run if someone adds DF_LR_RUN_DCE. At least in ifcvt
we already have this problem even now (i.e. directly from ifcvt.c). The others
could trigger a DCE run via one of the CFG manipulation functions in cfg* and
emit-rtl.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug rtl-optimization/42889] [4.5 Regression] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[jakub at gcc dot gnu dot org]

------- Comment #18 from jakub at gcc dot gnu dot org  2010-01-28 22:09 -------
Those need to be analyzed, perhaps some will need a similar change, many others
don't need any change, as they can't be called for DEBUG_INSNs.  E.g.
DEBUG_INSN isn't changing control flow, so any changes to it won't add/remove
basic blocks, etc.
But I fail to see why we'd need to solve all at once.

This patch has been successfully bootstrapped/regtested on x86_64-linux and
i686-linux BTW.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug rtl-optimization/42889] [4.5 Regression] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[steven at gcc dot gnu dot org]

------- Comment #19 from steven at gcc dot gnu dot org  2010-01-28 22:23 -------
I don't think there's a need to solve all at once, but it's only good practice
to look ahead for more bears on the road :-P


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug rtl-optimization/42889] [4.5 Regression] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[jakub at gcc dot gnu dot org]

------- Comment #20 from jakub at gcc dot gnu dot org  2010-01-29 12:15 -------
Subject: Bug 42889

Author: jakub
Date: Fri Jan 29 12:14:47 2010
New Revision: 156344

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=156344
Log:
        PR rtl-optimization/42889
        * df.h (df_set_bb_dirty_nonlr): New prototype.
        * df-core.c (df_set_bb_dirty_nonlr): New function.
        * df-scan.c (df_insn_rescan): Call it instead of
        df_set_bb_dirty for DEBUG_INSNs.

        * gcc.dg/pr42889.c: New test.

Added:
    trunk/gcc/testsuite/gcc.dg/pr42889.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/df-core.c
    trunk/gcc/df-scan.c
    trunk/gcc/df.h
    trunk/gcc/testsuite/ChangeLog


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug rtl-optimization/42889] [4.5 Regression] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[jakub at gcc dot gnu dot org]

------- Comment #21 from jakub at gcc dot gnu dot org  2010-01-29 17:26 -------
Fixed.


-- 

jakub at gcc dot gnu dot org changed:

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


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

[Bug rtl-optimization/42889] [4.5 Regression] "-fcompare-debug failure (length)" with "-O1 -fgcse"

[aoliva at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42889

--- Comment #22 from Alexandre Oliva <aoliva at gcc dot gnu.org> 2010-11-13 18:52:00 UTC ---
Author: aoliva
Date: Sat Nov 13 18:51:57 2010
New Revision: 166719

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=166719
Log:
PR debug/42889
* df-scan.c (df_insn_rescan): Don't mark BBs upon debug insns.
* df-core.c (df_set_bb_dirty_nonrl): Remove.
* df.h (df_set_bb_dirty_nonlr): Likewise.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/df-core.c
    trunk/gcc/df-scan.c
    trunk/gcc/df.h