[Bug lto/49302] New: [4.7 Regression] FAIL: gcc.dg/lto/20110201-1 c_lto_20110201-1_0.o-c_lto_20110201-1_0.o

[Bug lto/49302] New: [4.7 Regression] FAIL: gcc.dg/lto/20110201-1 c_lto_20110201-1_0.o-c_lto_20110201-1_0.o

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49302

           Summary: [4.7 Regression] FAIL: gcc.dg/lto/20110201-1
                    c_lto_20110201-1_0.o-c_lto_20110201-1_0.o
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: lto
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: hjl.tools@gmail.com


On Linux/x86. revision 174704 gave

FAIL: gcc.dg/lto/20110201-1 c_lto_20110201-1_0.o-c_lto_20110201-1_0.o link, 
-O0 -flto 

Revision 174700 is OK.

[Bug lto/49302] [4.7 Regression] FAIL: gcc.dg/lto/20110201-1 c_lto_20110201-1_0.o-c_lto_20110201-1_0.o

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49302

H.J. Lu <hjl.tools at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |wschmidt at gcc dot gnu.org

--- Comment #1 from H.J. Lu <hjl.tools at gmail dot com> 2011-06-07 04:13:48 UTC ---
It is caused by revision 174701:

http://gcc.gnu.org/ml/gcc-cvs/2011-06/msg00189.html

I got

Executing on host: /export/regression/rrs/174701/bld/gcc/xgcc
-B/export/regression/rrs/174701/bld/gcc/ c_lto_20110201-1_0.o   -O0 -flto  -O2
-ffast-math -fuse-linker-plugin      -m32 -o gcc-dg-lto-20110201-1-01   
(timeout = 300)
spawn -ignore SIGHUP /export/regression/rrs/174701/bld/gcc/xgcc
-B/export/regression/rrs/174701/bld/gcc/ c_lto_20110201-1_0.o -O0 -flto -O2
-ffast-math -fuse-linker-plugin -m32 -o gcc-dg-lto-20110201-1-01^M
/tmp/cc5ZLSPu.ltrans0.ltrans.o: In function `foo':^M
cc5ZLSPu.ltrans0.o:(.text+0x45): undefined reference to `pow'^M
cc5ZLSPu.ltrans0.o:(.text+0x5d): undefined reference to `pow'^M
collect2: error: ld returned 1 exit status^M

[Bug lto/49302] [4.7 Regression] FAIL: gcc.dg/lto/20110201-1 c_lto_20110201-1_0.o-c_lto_20110201-1_0.o

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49302

--- Comment #2 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2011-06-07 11:51:56 UTC ---
I'll have a look.  With -ffast-math, cabs is converted to
sqrt(pow(rpart,2.0),pow(ipart,2.0)).  Not sure yet why the pow calls aren't
converted back to x*x form.  We may have to restore just that transformation to
the expander for builtin pow.  Will dig deeper today.

[Bug lto/49302] [4.7 Regression] FAIL: gcc.dg/lto/20110201-1 c_lto_20110201-1_0.o-c_lto_20110201-1_0.o

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49302

--- Comment #3 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2011-06-07 12:08:17 UTC ---
(In reply to comment #2)
> I'll have a look.  With -ffast-math, cabs is converted to
> sqrt(pow(rpart,2.0),pow(ipart,2.0)).  Not sure yet why the pow calls aren't
> converted back to x*x form.  We may have to restore just that transformation to
> the expander for builtin pow.  Will dig deeper today.

Too early to be typing.  Of course I meant sqrt(pow(rpart,2.0)+pow(ipart,2.0)).

[Bug lto/49302] [4.7 Regression] FAIL: gcc.dg/lto/20110201-1 c_lto_20110201-1_0.o-c_lto_20110201-1_0.o

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49302

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2011.06.07 12:34:08
   Target Milestone|---                         |4.7.0
     Ever Confirmed|0                           |1

--- Comment #4 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-06-07 12:34:08 UTC ---
Well, it fails at the -O0 (non-LTO) link.  Hmm.  The testcase is from PR47566.

[Bug lto/49302] [4.7 Regression] FAIL: gcc.dg/lto/20110201-1 c_lto_20110201-1_0.o-c_lto_20110201-1_0.o

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49302

William J. Schmidt <wschmidt at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |bergner at gcc dot gnu.org

--- Comment #5 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2011-06-07 13:16:12 UTC ---
(In reply to comment #4)
> Well, it fails at the -O0 (non-LTO) link.  Hmm.  The testcase is from PR47566.

OK.  So the LTO writer assumed -O2 -ffast-math, causing the pow calls to be
generated, but at -O0 the gimple expand code for pow was skipped.

I _think_ that

[Bug lto/49302] [4.7 Regression] FAIL: gcc.dg/lto/20110201-1 c_lto_20110201-1_0.o-c_lto_20110201-1_0.o

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49302

--- Comment #6 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2011-06-07 13:23:06 UTC ---
(In reply to comment #5)
Please ignore that.  It leaked in when I was trying to CC somebody.

[Bug lto/49302] [4.7 Regression] FAIL: gcc.dg/lto/20110201-1 c_lto_20110201-1_0.o-c_lto_20110201-1_0.o

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49302

--- Comment #7 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-06-07 13:47:00 UTC ---
The original failure only occured with -O0 compile and -O2 -ffast-math link.
The linker-plugin use was already a workaround for the -O0 link by collect2,
so not sure why that isn't working anymore.  Thus, I guess this still needs
investigation.

I don't see the cabs -> sqrt (pow, pow) folding anymore.  Ah, because it
happens from PRE which folds and inserts cabs into the predecessor.  But
PRE runs after pass_sincos.

I think the cabs -> sqrt (pow, pow) folding should happen during
pass_sincos.

[Bug lto/49302] [4.7 Regression] FAIL: gcc.dg/lto/20110201-1 c_lto_20110201-1_0.o-c_lto_20110201-1_0.o

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49302

--- Comment #8 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2011-06-07 20:25:20 UTC ---
Because the compile uses -O0 and the link uses -O2 -ffast-math, we have a
mismatch of expectations.  With -O0 and no -ffast-math, builtin cabs isn't
folded into sqrt(pow+pow) during parsing.  With -O2 -ffast-math, we expect all
pow's to have been exposed before cse_sincos, but the folding doesn't have a
chance to occur until PRE.

It seems odd to use a -O0 compile when we're going to force a -O2 link, but
assuming that's WAD, then I agree with adding a cabs fold in cse_sincos.  I can
work on that.

I assume we don't want to mirror the other folds from
builtins.c:fold_builtin_cabs, as this is just a matter of ensuring things
compile correctly with -O0 -lto.

[Bug lto/49302] [4.7 Regression] FAIL: gcc.dg/lto/20110201-1 c_lto_20110201-1_0.o-c_lto_20110201-1_0.o

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49302

--- Comment #9 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2011-06-08 13:51:34 UTC ---
Hm, this isn't going to work for all cases.  When I reproduce the problem on
powerpc64-linux without forcing a hardware square root, I see:

spawn /home/wschmidt/gcc/build/gcc-mainline-pr49302/gcc/xgcc
-B/home/wschmidt/gcc/build/gcc-mainline-pr49302/gcc/ c_lto_20110201-1_0.o -O0
-flto -O2 -ffast-math -fuse-linker-plugin -m32 -o gcc-dg-lto-20110201-1-01

/tmp/cc1mUEPp.ltrans0.ltrans.o: In function `foo':

cc1mUEPp.ltrans0.o:(.text+0x5c): undefined reference to `pow'

cc1mUEPp.ltrans0.o:(.text+0x74): undefined reference to `sqrt'

collect2: error: ld returned 1 exit status

compiler exited with status 1
output is:
/tmp/cc1mUEPp.ltrans0.ltrans.o: In function `foo':

cc1mUEPp.ltrans0.o:(.text+0x5c): undefined reference to `pow'

cc1mUEPp.ltrans0.o:(.text+0x74): undefined reference to `sqrt'

collect2: error: ld returned 1 exit status


Expanding cabs to sqrt(r*r + i*i) will still leave an unresolved reference to
sqrt.

[Bug lto/49302] [4.7 Regression] FAIL: gcc.dg/lto/20110201-1 c_lto_20110201-1_0.o-c_lto_20110201-1_0.o

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49302

--- Comment #10 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2011-06-08 17:07:42 UTC ---
Proposed fix:  http://gcc.gnu.org/ml/gcc-patches/2011-06/msg00661.html

[Bug lto/49302] [4.7 Regression] FAIL: gcc.dg/lto/20110201-1 c_lto_20110201-1_0.o-c_lto_20110201-1_0.o

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49302

--- Comment #11 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2011-06-10 15:28:52 UTC ---
Author: wschmidt
Date: Fri Jun 10 15:28:46 2011
New Revision: 174912

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=174912
Log:
2011-06-10  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>

    PR lto/49302
    * tree-ssa-math-opts.c (powi_as_mults): Minor cleanup.
    (build_and_insert_call): Likewise.
    (build_and_insert_ref): New.
    (gimple_expand_builtin_pow): Minor cleanup.
    (gimple_expand_builtin_cabs): New.
    (execute_cse_sincos): Add case for BUILT_IN_CABS.


Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/tree-ssa-math-opts.c

[Bug lto/49302] [4.7 Regression] FAIL: gcc.dg/lto/20110201-1 c_lto_20110201-1_0.o-c_lto_20110201-1_0.o

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49302

--- Comment #12 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2011-06-10 15:33:53 UTC ---
Fixed.

[Bug lto/49302] [4.7 Regression] FAIL: gcc.dg/lto/20110201-1 c_lto_20110201-1_0.o-c_lto_20110201-1_0.o

[meissner at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49302

Michael Meissner <meissner at gcc dot gnu.org> changed:

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

[Bug lto/49302] [4.7 Regression] FAIL: gcc.dg/lto/20110201-1 c_lto_20110201-1_0.o-c_lto_20110201-1_0.o

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49302

William J. Schmidt <wschmidt at gcc dot gnu.org> changed:

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

--- Comment #13 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2011-06-15 17:45:05 UTC ---
Fixed.