[Bug c/36359] New: [Regression] compile error in linux-kernel 2.6.26-rc4 with -O2

[Bug c/36359] New: [Regression] compile error in linux-kernel 2.6.26-rc4 with -O2

[mt-ml at gmx dot de]

Hi
I've tested gcc-4.4 with the current developer kernel 2.6.26-rc4 and got the
following compile error:

...
Setup is 12236 bytes (padded to 12288 bytes).
System is 2521 kB
CRC 735d1682
Kernel: arch/x86/boot/bzImage is ready  (#62)
  Building modules, stage 2.
  MODPOST 222 modules
ERROR: "____ilog2_NaN" [drivers/usb/core/usbcore.ko] undefined!
make[1]: *** [__modpost] Error 1
make: *** [modules] Error 2

This can be avoided when switching on the OPTIMIZE_SIZE kconfig option which in
fact replaces the standard -O2 option with -Os, so I think there must be
something wrong in what -O2 does. 
This happens only with gcc-4.4 not with gcc-4.3 and before (I've also tested
gcc-4.1). The latest gcc-4.4 version I tested was 20080526.

I compiled gcc without any options to ./configure.

I'll try to attach my config file so you might be able to reproduce the error.

This has been reported to lkml. The discussion can be read here:
http://lkml.org/lkml/2008/5/25/169.

Thanks
Mirco


-- 
           Summary: [Regression] compile error in linux-kernel 2.6.26-rc4
                    with -O2
           Product: gcc
           Version: 4.4.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: mt-ml at gmx dot de
 GCC build triplet: none
  GCC host triplet: ubuntu linux 8.10 (dev)
GCC target triplet: x86_64


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug c/36359] [Regression] compile error in linux-kernel 2.6.26-rc4 with -O2

[mt-ml at gmx dot de]

------- Comment #1 from mt-ml at gmx dot de  2008-05-28 22:14 -------
Created an attachment (id=15695)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=15695&action=view)
kernel config-file to reproduce error


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug c/36359] [Regression] compile error in linux-kernel 2.6.26-rc4 with -O2

[rguenth at gcc dot gnu dot org]

------- Comment #2 from rguenth at gcc dot gnu dot org  2008-05-28 22:20 -------
This sounds like a kernel bug.  At least we need preprocessed source of the
source file in drivers/usb/core/usbcore.ko that generates the reference
to ____ilog2_NaN.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |WAITING


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug c/36359] [Regression] compile error in linux-kernel 2.6.26-rc4 with -O2

[pinskia at gcc dot gnu dot org]

------- Comment #3 from pinskia at gcc dot gnu dot org  2008-05-28 22:52 -------
http://lkml.org/lkml/2008/5/25/233

It says something about __builtin_constant_p being broken but I doubt it
really.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug c/36359] [Regression] compile error in linux-kernel 2.6.26-rc4 with -O2

[mt-ml at gmx dot de]

------- Comment #4 from mt-ml at gmx dot de  2008-05-28 23:08 -------
(In reply to comment #2)
> This sounds like a kernel bug.  At least we need preprocessed source of the
> source file in drivers/usb/core/usbcore.ko that generates the reference
> to ____ilog2_NaN.
> 
Can you advise me please how to create them? I've tried adding -save-temps to
the EXTRA_CFLAGS in driver/usb/core/Makefile but that didn't seem to have
changed anything. There are only sources and .o/.ko files in that directory.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug c/36359] [Regression] compile error in linux-kernel 2.6.26-rc4 with -O2

[mt-ml at gmx dot de]

------- Comment #5 from mt-ml at gmx dot de  2008-05-31 13:58 -------
Created an attachment (id=15707)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=15707&action=view)
preprocessed sources of urb.c


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug c/36359] [Regression] compile error in linux-kernel 2.6.26-rc4 with -O2

[mt-ml at gmx dot de]

------- Comment #6 from mt-ml at gmx dot de  2008-05-31 13:59 -------
I was able to find the function in which the error occurs. It's usb_submit_urb
in drivers/usb/core/urb.c I think.
The command:
$ gcc -c -save-temps drivers/usb/core/urb.c -I include
executed in the linux-kernel main directory gives me the attached urb.i. Is
that what you need? Executing this gives me the following errors, though:
In file included from
/home/mirco/source/kernel/linux.git/include/asm/bitops.h:421,
                 from
/home/mirco/source/kernel/linux.git/include/linux/bitops.h:17,
                 from
/home/mirco/source/kernel/linux.git/include/linux/bitmap.h:7,
                 from
/home/mirco/source/kernel/linux.git/include/linux/cpumask.h:90,
                 from
/home/mirco/source/kernel/linux.git/include/asm/processor.h:25,
                 from
/home/mirco/source/kernel/linux.git/include/linux/prefetch.h:14,
                 from
/home/mirco/source/kernel/linux.git/include/linux/list.h:6,
                 from
/home/mirco/source/kernel/linux.git/include/linux/module.h:9,
                 from urb.c:1:
/home/mirco/source/kernel/linux.git/include/asm-generic/bitops/fls64.h:33:2:
error: #error BITS_PER_LONG not 32 or 64
In file included from
/home/mirco/source/kernel/linux.git/include/linux/prefetch.h:14,
                 from
/home/mirco/source/kernel/linux.git/include/linux/list.h:6,
                 from
/home/mirco/source/kernel/linux.git/include/linux/module.h:9,
                 from urb.c:1:
/home/mirco/source/kernel/linux.git/include/asm/processor.h:152:1: warning:
"cache_line_size" redefined
In file included from /home/mirco/source/kernel/linux.git/include/asm/pda.h:7,
                 from
/home/mirco/source/kernel/linux.git/include/asm/current_64.h:7,
                 from
/home/mirco/source/kernel/linux.git/include/asm/current.h:4,
                 from
/home/mirco/source/kernel/linux.git/include/asm/processor.h:15,
                 from
/home/mirco/source/kernel/linux.git/include/linux/prefetch.h:14,
                 from
/home/mirco/source/kernel/linux.git/include/linux/list.h:6,
                 from
/home/mirco/source/kernel/linux.git/include/linux/module.h:9,
                 from urb.c:1:
/home/mirco/source/kernel/linux.git/include/linux/cache.h:64:1: warning: this
is the location of the previous definition
In file included from
/home/mirco/source/kernel/linux.git/include/linux/module.h:21,
                 from urb.c:1:
/home/mirco/source/kernel/linux.git/include/asm/module.h:70:2: error: #error
unknown processor family


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug c/36359] [Regression] compile error in linux-kernel 2.6.26-rc4 with -O2

[rguenth at gcc dot gnu dot org]

------- Comment #7 from rguenth at gcc dot gnu dot org  2008-05-31 15:45 -------
Looking at http://www.readcode.org/code/linux-2.6.20/include/linux/log2.h it
seems that either the argument to ilog _is_ zero or the compiler thinks so.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |NEW
     Ever Confirmed|0                           |1
   Last reconfirmed|0000-00-00 00:00:00         |2008-05-31 15:45:30
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug c/36359] [Regression] compile error in linux-kernel 2.6.26-rc4 with -O2

[mt-ml at gmx dot de]

------- Comment #8 from mt-ml at gmx dot de  2008-06-08 00:17 -------
I've been able to identify the code where ilog2 is called. If I do not choose
usbcore as a module I get the following error:
drivers/built-in.o: In function `usb_submit_urb':
(.text+0xc247f): undefined reference to `____ilog2_NaN'
make: *** [.tmp_vmlinux1] Fehler 1

Observing that function usb_submit_urb() in drivers/usb/core/urb.c shows the
interesting lines should be:
        case USB_ENDPOINT_XFER_INT:
                /* too small? */
                if (urb->interval <= 0)
                        return -EINVAL;
                /* too big? */
                switch (dev->speed) {
                case USB_SPEED_HIGH:    /* units are microframes */
                        /* NOTE usb handles 2^15 */
                        if (urb->interval > (1024 * 8))
                                urb->interval = 1024 * 8;
                        max = 1024 * 8;
                        break;
                case USB_SPEED_FULL:    /* units are frames/msec */
                case USB_SPEED_LOW:
                        if (xfertype == USB_ENDPOINT_XFER_INT) {
                                if (urb->interval > 255)
                                        return -EINVAL;
                                /* NOTE ohci only handles up to 32 */
                                max = 128;
                        } else {
                                if (urb->interval > 1024)
                                        urb->interval = 1024;
                                /* NOTE usb and ohci handle up to 2^15 */
                                max = 1024;
                        }
                        break;
                default:
                        return -EINVAL;
                }
                /* Round down to a power of 2, no more than max */
                urb->interval = min(max, 1 << ilog2(urb->interval));
        }

As you can see, the argument to ilog2 is urb->interval which can't be zero at
that point IMHO.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

Re: [Bug c/36359] [Regression] compile error in linux-kernel 2.6.26-rc4 with -O2

[Andrew Pinski]

On Sat, Jun 7, 2008 at 5:17 PM, mt-ml at gmx dot de
<gcc-bugzilla@gcc.gnu.org> wrote:
> As you can see, the argument to ilog2 is urb->interval which can't be zero at
> that point IMHO.

Yes and that requires a lot of jump threading and many other
optimizations to detect that really.  First PRE has to be working and
then VRP has to figure out it will be greater than 0.  So really this
code depends on optimization to work correctly and the kernel is
broken that way again (not suppressing really)

-- Pinski

[Bug c/36359] [Regression] compile error in linux-kernel 2.6.26-rc4 with -O2

[pinskia at gmail dot com]

------- Comment #9 from pinskia at gmail dot com  2008-06-08 00:33 -------
Subject: Re:  [Regression] compile error in linux-kernel 2.6.26-rc4 with -O2

On Sat, Jun 7, 2008 at 5:17 PM, mt-ml at gmx dot de
<gcc-bugzilla@gcc.gnu.org> wrote:
> As you can see, the argument to ilog2 is urb->interval which can't be zero at
> that point IMHO.

Yes and that requires a lot of jump threading and many other
optimizations to detect that really.  First PRE has to be working and
then VRP has to figure out it will be greater than 0.  So really this
code depends on optimization to work correctly and the kernel is
broken that way again (not suppressing really)

-- Pinski


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug c/36359] [Regression] compile error in linux-kernel 2.6.26-rc4 with -O2

[bunk at stusta dot de]

------- Comment #10 from bunk at stusta dot de  2008-06-10 10:11 -------
Created an attachment (id=15745)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=15745&action=view)
Mirco's usbcore.o


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug c/36359] [Regression] compile error in linux-kernel 2.6.26-rc4 with -O2

[bunk at stusta dot de]

------- Comment #11 from bunk at stusta dot de  2008-06-10 10:28 -------
(In reply to comment #7)
> Looking at http://www.readcode.org/code/linux-2.6.20/include/linux/log2.h it
> seems that either the argument to ilog _is_ zero or the compiler thinks so.

If you disassemble Mirco's usbcore.o you'll see why I said "I'd have said your
gcc has a bug regarding __builtin_constant_p()": The compiler does not think
the argument to ilog() was zero - it expanded _all_ the code for the case when
__builtin_constant_p(n) returns true.


-- 

bunk at stusta dot de changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |bunk at stusta dot de


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug c/36359] [Regression] compile error in linux-kernel 2.6.26-rc4 with -O2

[aldot at gcc dot gnu dot org]

------- Comment #12 from aldot at gcc dot gnu dot org  2008-06-10 11:07 -------
smallish testcase, fwiw.

$ gcc -Wall -Wundef -Wstrict-prototypes -Wno-trigraphs -fno-strict-aliasing
-fno-common -Werror-implicit-function-declaration -fno-stack-protector
-mregparm=3 -freg-struct-return -mpreferred-stack-boundary=2  -march=i686
-ffreestanding -pipe -fno-asynchronous-unwind-tables -mno-sse -mno-mmx
-mno-sse2 -mno-3dnow -O2 -c -o pr.o pr36359.c
$ nm pr.o 
         U argh
00000000 T foo
         U urb
$ cat pr36359.c 
extern int urb;
int argh(void);
void foo(void)
{
        if (urb <= 0)
                return;
        urb = __builtin_constant_p(urb) ? 0x0815 : argh();
}


-- 

aldot at gcc dot gnu dot org changed:

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


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug c/36359] [Regression] compile error in linux-kernel 2.6.26-rc4 with -O2

[aldot at gcc dot gnu dot org]

------- Comment #13 from aldot at gcc dot gnu dot org  2008-06-10 11:20 -------
should have been
int argh(void) __attribute__((error ("dce?")));

works without __builtin_constant_p and fails even with
-Os -ftree-dce -fweb -ftree-fre -ftree-pre

pr36359.c: In function 'foo':
pr36359.c:7: error: call to 'argh' declared with attribute error: dce?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug c/36359] [Regression] compile error in linux-kernel 2.6.26-rc4 with -O2

[pinskia at gcc dot gnu dot org]

------- Comment #14 from pinskia at gcc dot gnu dot org  2008-11-15 00:28 -------
(In reply to comment #12)
> smallish testcase, fwiw.
No that should have a call to argh.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug middle-end/36359] missed optimization in some cases with PRE VRP and other passes combined together

[pinskia at gcc dot gnu dot org]

------- Comment #15 from pinskia at gcc dot gnu dot org  2008-11-15 00:31 -------
As I mentioned this requires so many optimizations to get this done correctly
it is not a regression.


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  GCC build triplet|none                        |
   GCC host triplet|ubuntu linux 8.10 (dev)     |
 GCC target triplet|x86_64                      |
            Summary|[Regression] compile error  |missed optimization in some
                   |in linux-kernel 2.6.26-rc4  |cases with PRE VRP and other
                   |with -O2                    |passes combined together


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug middle-end/36359] missed optimization in some cases with PRE VRP and other passes combined together

[pinskia at gcc dot gnu dot org]

------- Comment #16 from pinskia at gcc dot gnu dot org  2008-11-15 00:34 -------
The preprocessed source does not compile for me:
/home/mirco/source/kernel/linux.git/include/asm/current_64.h:11: error:
expected expression before 'struct'


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |WAITING


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug middle-end/36359] missed optimization in some cases with PRE VRP and other passes combined together

[edwintorok at gmail dot com]

------- Comment #17 from edwintorok at gmail dot com  2008-11-24 18:33 -------
Created an attachment (id=16761)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=16761&action=view)
drivers/scsi/sd.c

Similar problem with drivers/scsi/sd.c

I can compile the attached preprocessed on x86_64 linux:
$ gcc-4.3 -c x.i -o x.o -O2 && nm -u x.o|grep ilog2
$ ~/gcc_inst/bin/gcc -c x.i -o x.o -O2 && nm -u x.o|grep ilog2
U ____ilog2_NaN

IMHO the kernel is abusing gcc as a static analyzer, which gcc is not.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug middle-end/36359] missed optimization in some cases with PRE VRP and other passes combined together

[linuxl4 at sohu dot com]

------- Comment #18 from linuxl4 at sohu dot com  2008-12-17 20:08 -------
confirmed for linux kernel 2.6.28-rc8.


-- 

linuxl4 at sohu dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |linuxl4 at sohu dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug middle-end/36359] missed optimization in some cases with PRE VRP and other passes combined together

[gcc at spatium dot org]

------- Comment #19 from gcc at spatium dot org  2009-01-25 19:25 -------
This thing still exists in 2.6.28.2; what do you suggest should be changed in
the kernel to let it compile again?


-- 

gcc at spatium dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |gcc at spatium dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug middle-end/36359] missed optimization in some cases with PRE VRP and other passes combined together

[rguenth at gcc dot gnu dot org]

------- Comment #20 from rguenth at gcc dot gnu dot org  2009-01-25 21:10 -------
The testcase from #17 does not reproduce the issue for me with recent GCC 4.3.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug middle-end/36359] missed optimization in some cases with PRE VRP and other passes combined together

[bunk at stusta dot de]

------- Comment #21 from bunk at stusta dot de  2009-01-25 22:00 -------
(In reply to comment #20)
> The testcase from #17 does not reproduce the issue for me with recent GCC 4.3.

This bug is a regression in gcc 4.4, it was AFAIK never present in gcc 4.3.

Haven't tried more recent gcc versions, but I'm still seeing it with the
20090107-1 gcc-snapshot package (that is the SVN trunk).


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug middle-end/36359] missed optimization in some cases with PRE VRP and other passes combined together

[bunk at stusta dot de]

------- Comment #22 from bunk at stusta dot de  2009-01-25 22:05 -------
Check my comments #10 and #11 and the definition of ilog2() in
http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=blob;f=include/linux/log2.h;h=25b808631cd92c50d10cf6a31b2d9b9942b62ac9;hb=bce7f793daec3e65ec5c5705d2457b81fe7b5725
and you'll understand why I said 8 months ago "gcc has a bug regarding
__builtin_constant_p()".


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug middle-end/36359] missed optimization in some cases with PRE VRP and other passes combined together

[rguenth at gcc dot gnu dot org]

------- Comment #23 from rguenth at gcc dot gnu dot org  2009-01-25 22:17 -------
It doesn't reproduce for me with 4.4 either.  Maybe this is a dup of PR38789?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug middle-end/36359] missed optimization in some cases with PRE VRP and other passes combined together

[bunk at stusta dot de]

------- Comment #24 from bunk at stusta dot de  2009-01-26 00:49 -------
(In reply to comment #23)
> It doesn't reproduce for me with 4.4 either.  Maybe this is a dup of PR38789?


Seems so:
I've confirmed that the 4.4-20090109 snapshot is broken, and the 4.4-20090123
snapshot is fixed.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359

[Bug middle-end/36359] missed optimization in some cases with PRE VRP and other passes combined together

[rguenth at gcc dot gnu dot org]

------- Comment #25 from rguenth at gcc dot gnu dot org  2009-01-26 10:22 -------


*** This bug has been marked as a duplicate of 38789 ***


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |RESOLVED
         Resolution|                            |DUPLICATE


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36359