2.6.19-rc5-mm1 fails to compile with gcc 4.2

2.6.19-rc5-mm1 fails to compile with gcc 4.2

[Bernhard Rosenkraenzer]

drivers/kvm/kvm_main.c: In function 'kvm_dev_ioctl_run':
drivers/kvm/kvm_main.c:153: error: 'asm' operand has impossible constraints
drivers/kvm/kvm_main.c:158: error: 'asm' operand has impossible constraints

Re: 2.6.19-rc5-mm1 fails to compile with gcc 4.2

[Avi Kivity]

Bernhard Rosenkraenzer wrote:
> drivers/kvm/kvm_main.c: In function 'kvm_dev_ioctl_run':
> drivers/kvm/kvm_main.c:153: error: 'asm' operand has impossible constraints
> drivers/kvm/kvm_main.c:158: error: 'asm' operand has impossible constraints
>   

Smells like a gcc regression.  Can you send .config?

Or better yet, preprocessed source and full gcc command line (as seen on 
'make V=1').

-- 
error compiling committee.c: too many arguments to function

Re: 2.6.19-rc5-mm1 fails to compile with gcc 4.2

[Bernhard Rosenkraenzer]

On Sunday, 12. November 2006 09:22, Avi Kivity wrote:
> Bernhard Rosenkraenzer wrote:
> > drivers/kvm/kvm_main.c: In function 'kvm_dev_ioctl_run':
> > drivers/kvm/kvm_main.c:153: error: 'asm' operand has impossible
> > constraints drivers/kvm/kvm_main.c:158: error: 'asm' operand has
> > impossible constraints
>
> Smells like a gcc regression.  Can you send .config?
>
> Or better yet, preprocessed source and full gcc command line (as seen on
> 'make V=1').

It does look like a gcc bug -- -O0 makes it go away.
Details at http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29808

Re: 2.6.19-rc5-mm1 fails to compile with gcc 4.2

[Avi Kivity]

Bernhard Rosenkraenzer wrote:
> On Sunday, 12. November 2006 09:22, Avi Kivity wrote:
>   
>> Bernhard Rosenkraenzer wrote:
>>     
>>> drivers/kvm/kvm_main.c: In function 'kvm_dev_ioctl_run':
>>> drivers/kvm/kvm_main.c:153: error: 'asm' operand has impossible
>>> constraints drivers/kvm/kvm_main.c:158: error: 'asm' operand has
>>> impossible constraints
>>>       
>> Smells like a gcc regression.  Can you send .config?
>>
>> Or better yet, preprocessed source and full gcc command line (as seen on
>> 'make V=1').
>>     
>
> It does look like a gcc bug -- -O0 makes it go away.
> Details at http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29808
>   

That's a different bug, gcc generates code that the assembler can't 
handle.  Might be an assembler bug.

Can you compile it with -S and post the generated assembly?

-- 
error compiling committee.c: too many arguments to function

Re: 2.6.19-rc5-mm1 fails to compile with gcc 4.2

[Segher Boessenkool]

>>> drivers/kvm/kvm_main.c: In function 'kvm_dev_ioctl_run':
>>> drivers/kvm/kvm_main.c:153: error: 'asm' operand has impossible
>>> constraints drivers/kvm/kvm_main.c:158: error: 'asm' operand has
>>> impossible constraints
>>
>> Smells like a gcc regression.  Can you send .config?
>>
>> Or better yet, preprocessed source and full gcc command line (as  
>> seen on
>> 'make V=1').

Just the function containing those lines (please mark which
lines they are) would do probably.

> It does look like a gcc bug -- -O0 makes it go away.
> Details at http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29808

PR29808 is not a GCC bug, but invalid code.


Segher

Re: 2.6.19-rc5-mm1 fails to compile with gcc 4.2

[Bernhard Rosenkraenzer]

On Sunday, 12. November 2006 10:24, Avi Kivity wrote:
> >> Or better yet, preprocessed source and full gcc command line (as seen on
> >> 'make V=1').

gcc -m32 -Wp,-MD,drivers/kvm/.kvm_main.o.d  -nostdinc -isystem /usr/lib/gcc/i586-ark-linux/4.2.0/include -D__KERNEL__ -Iinclude  -include 
include/linux/autoconf.h -Wall -Wundef -Wstrict-prototypes -Wno-trigraphs -fno-strict-aliasing -fno-common -O2 -pipe -msoft-float -mpreferred-stack-boundary=2  -march=i686 -mtune=pentium3 -maccumulate-outgoing-args -DCONFIG_AS_CFI=1 -DCONFIG_AS_CFI_SIGNAL_FRAME=1 -Iinclude/asm-i386/mach-default -fomit-frame-pointer  -fno-stack-protector -Wdeclaration-after-statement -Wno-pointer-sign   -DMODULE -D"KBUILD_STR(s)=#s" -D"KBUILD_BASENAME=KBUILD_STR(kvm_main)"  -D"KBUILD_MODNAME=KBUILD_STR(kvm)" -c -o 
drivers/kvm/.tmp_kvm_main.o drivers/kvm/kvm_main.c
drivers/kvm/kvm_main.c: In function 'kvm_dev_ioctl_run':
drivers/kvm/kvm_main.c:153: error: 'asm' operand has impossible constraints
drivers/kvm/kvm_main.c:158: error: 'asm' operand has impossible constraints

> > It does look like a gcc bug -- -O0 makes it go away.
> > Details at http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29808
>
> That's a different bug, gcc generates code that the assembler can't
> handle.  Might be an assembler bug.

It's the same thing, the code is taken from kvm_main.c:

static void load_fs(u16 sel)
{
        asm ("mov %0, %%fs" : : "g"(sel));	<--- line 153
}

static void load_gs(u16 sel)
{
        asm ("mov %0, %%gs" : : "g"(sel));	<--- line 158
}


> Can you compile it with -S and post the generated assembly?

It can't generate assembly with asm() constructs it perceives as invalid -- -S 
produces the same error.

Re: 2.6.19-rc5-mm1 fails to compile with gcc 4.2

[Avi Kivity]

Bernhard Rosenkraenzer wrote:
> On Sunday, 12. November 2006 10:24, Avi Kivity wrote:
>   
>>>> Or better yet, preprocessed source and full gcc command line (as seen on
>>>> 'make V=1').
>>>>         
>
> gcc -m32 -Wp,-MD,drivers/kvm/.kvm_main.o.d  -nostdinc -isystem /usr/lib/gcc/i586-ark-linux/4.2.0/include -D__KERNEL__ -Iinclude  -include 
> include/linux/autoconf.h -Wall -Wundef -Wstrict-prototypes -Wno-trigraphs -fno-strict-aliasing -fno-common -O2 -pipe -msoft-float -mpreferred-stack-boundary=2  -march=i686 -mtune=pentium3 -maccumulate-outgoing-args -DCONFIG_AS_CFI=1 -DCONFIG_AS_CFI_SIGNAL_FRAME=1 -Iinclude/asm-i386/mach-default -fomit-frame-pointer  -fno-stack-protector -Wdeclaration-after-statement -Wno-pointer-sign   -DMODULE -D"KBUILD_STR(s)=#s" -D"KBUILD_BASENAME=KBUILD_STR(kvm_main)"  -D"KBUILD_MODNAME=KBUILD_STR(kvm)" -c -o 
> drivers/kvm/.tmp_kvm_main.o drivers/kvm/kvm_main.c
> drivers/kvm/kvm_main.c: In function 'kvm_dev_ioctl_run':
> drivers/kvm/kvm_main.c:153: error: 'asm' operand has impossible constraints
> drivers/kvm/kvm_main.c:158: error: 'asm' operand has impossible constraints
>
>   
>>> It does look like a gcc bug -- -O0 makes it go away.
>>> Details at http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29808
>>>       
>> That's a different bug, gcc generates code that the assembler can't
>> handle.  Might be an assembler bug.
>>     
>
> It's the same thing, the code is taken from kvm_main.c:
>
> static void load_fs(u16 sel)
> {
>         asm ("mov %0, %%fs" : : "g"(sel));	<--- line 153
> }
>
> static void load_gs(u16 sel)
> {
>         asm ("mov %0, %%gs" : : "g"(sel));	<--- line 158
> }
>
>   

'sel' is a variable, so gcc can't provide it as an immediate operand.  
Specifying it as "rm" instead of "g" would have been better, but can't 
have any real influence.

>   
>> Can you compile it with -S and post the generated assembly?
>>     
>
> It can't generate assembly with asm() constructs it perceives as invalid -- -S 
> produces the same error.
>   

Well, for the code you posted in in the gcc bug, it probaby generated 
something like

    mov $0, %fs

which is indeed invalid assembly.  But the kvm miscompile is something 
else (running out of registers or something like that).


-- 
error compiling committee.c: too many arguments to function

Re: 2.6.19-rc5-mm1 fails to compile with gcc 4.2

[Bernhard Rosenkraenzer]

[-- Attachment #1: Type: text/plain, Size: 446 bytes --]

On Sunday, 12. November 2006 09:22, Avi Kivity wrote:
> Bernhard Rosenkraenzer wrote:
> > drivers/kvm/kvm_main.c: In function 'kvm_dev_ioctl_run':
> > drivers/kvm/kvm_main.c:153: error: 'asm' operand has impossible
> > constraints drivers/kvm/kvm_main.c:158: error: 'asm' operand has
> > impossible constraints

The attached patch makes it compile (but I'm not 100% sure it's the right 
thing to do, I'm not very experienced with gcc-style asm).

[-- Attachment #2: kvm_main-compilefix.patch --]
[-- Type: text/x-diff, Size: 422 bytes --]

--- linux-2.6.18/drivers/kvm/kvm_main.c.ark	2006-11-12 14:40:09.000000000 +0100
+++ linux-2.6.18/drivers/kvm/kvm_main.c	2006-11-12 14:38:51.000000000 +0100
@@ -150,12 +150,12 @@
 
 static void load_fs(u16 sel)
 {
-	asm ("mov %0, %%fs" : : "g"(sel));
+	asm ("mov %0, %%fs" : : "m"(sel));
 }
 
 static void load_gs(u16 sel)
 {
-	asm ("mov %0, %%gs" : : "g"(sel));
+	asm ("mov %0, %%gs" : : "m"(sel));
 }
 
 #ifndef load_ldt

Re: 2.6.19-rc5-mm1 fails to compile with gcc 4.2

[Avi Kivity]

Bernhard Rosenkraenzer wrote:
> The attached patch makes it compile (but I'm not 100% sure it's the right 
> thing to do, I'm not very experienced with gcc-style asm).
>   
> ------------------------------------------------------------------------
>
> --- linux-2.6.18/drivers/kvm/kvm_main.c.ark	2006-11-12 14:40:09.000000000 +0100
> +++ linux-2.6.18/drivers/kvm/kvm_main.c	2006-11-12 14:38:51.000000000 +0100
> @@ -150,12 +150,12 @@
>  
>  static void load_fs(u16 sel)
>  {
> -	asm ("mov %0, %%fs" : : "g"(sel));
> +	asm ("mov %0, %%fs" : : "m"(sel));
>  }
>  
>  static void load_gs(u16 sel)
>  {
> -	asm ("mov %0, %%gs" : : "g"(sel));
> +	asm ("mov %0, %%gs" : : "m"(sel));
>  }
>  
>  #ifndef load_ldt
>   

Does "rm" instead of "m" work as well?


-- 
error compiling committee.c: too many arguments to function

Re: 2.6.19-rc5-mm1 fails to compile with gcc 4.2

[Bernhard Rosenkraenzer]

[-- Attachment #1: Type: text/plain, Size: 744 bytes --]

On Sunday, 12. November 2006 14:43, Avi Kivity wrote:
> 'sel' is a variable, so gcc can't provide it as an immediate operand.
> Specifying it as "rm" instead of "g" would have been better, but can't
> have any real influence.

Specifying it as "rm" instead of "g" does fix it -- patch attached.

> Well, for the code you posted in in the gcc bug, it probaby generated
> something like
>
>     mov $0, %fs
>
> which is indeed invalid assembly.  But the kvm miscompile is something
> else (running out of registers or something like that).

What am I overlooking? The code is the exact same (except I replaced "u16" 
with "unsigned short" to avoid the #include), and produces the exact same 
error message, and the fix is the same ("g" -> "rm").

[-- Attachment #2: kvm_main-compilefix.patch --]
[-- Type: text/x-diff, Size: 424 bytes --]

--- linux-2.6.18/drivers/kvm/kvm_main.c.ark	2006-11-12 14:40:09.000000000 +0100
+++ linux-2.6.18/drivers/kvm/kvm_main.c	2006-11-12 14:44:57.000000000 +0100
@@ -150,12 +150,12 @@
 
 static void load_fs(u16 sel)
 {
-	asm ("mov %0, %%fs" : : "g"(sel));
+	asm ("mov %0, %%fs" : : "rm"(sel));
 }
 
 static void load_gs(u16 sel)
 {
-	asm ("mov %0, %%gs" : : "g"(sel));
+	asm ("mov %0, %%gs" : : "rm"(sel));
 }
 
 #ifndef load_ldt

Re: 2.6.19-rc5-mm1 fails to compile with gcc 4.2

[Avi Kivity]

Bernhard Rosenkraenzer wrote:
> On Sunday, 12. November 2006 14:43, Avi Kivity wrote:
>   
>> 'sel' is a variable, so gcc can't provide it as an immediate operand.
>> Specifying it as "rm" instead of "g" would have been better, but can't
>> have any real influence.
>>     
>
> Specifying it as "rm" instead of "g" does fix it -- patch attached.
>
>   
>> Well, for the code you posted in in the gcc bug, it probaby generated
>> something like
>>
>>     mov $0, %fs
>>
>> which is indeed invalid assembly.  But the kvm miscompile is something
>> else (running out of registers or something like that).
>>     
>
> What am I overlooking? The code is the exact same (except I replaced "u16" 
> with "unsigned short" to avoid the #include), and produces the exact same 
> error message, and the fix is the same ("g" -> "rm").
>   

Well, since it works, I guess I'm overlooking something.  Maybe it's 
just a bad error message from gcc.

I'll apply this.  Thanks!

> ------------------------------------------------------------------------
>
> --- linux-2.6.18/drivers/kvm/kvm_main.c.ark	2006-11-12 14:40:09.000000000 +0100
> +++ linux-2.6.18/drivers/kvm/kvm_main.c	2006-11-12 14:44:57.000000000 +0100
> @@ -150,12 +150,12 @@
>  
>  static void load_fs(u16 sel)
>  {
> -	asm ("mov %0, %%fs" : : "g"(sel));
> +	asm ("mov %0, %%fs" : : "rm"(sel));
>  }
>  
>  static void load_gs(u16 sel)
>  {
> -	asm ("mov %0, %%gs" : : "g"(sel));
> +	asm ("mov %0, %%gs" : : "rm"(sel));
>  }
>  
>  #ifndef load_ldt
>   


-- 
error compiling committee.c: too many arguments to function

Re: 2.6.19-rc5-mm1 fails to compile with gcc 4.2

[Segher Boessenkool]

>>> It does look like a gcc bug -- -O0 makes it go away.
>>> Details at http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29808
>>
>> That's a different bug, gcc generates code that the assembler can't
>> handle.  Might be an assembler bug.
>
> It's the same thing, the code is taken from kvm_main.c:
>
> static void load_fs(u16 sel)
> {
>         asm ("mov %0, %%fs" : : "g"(sel));	<--- line 153
> }

Like I said in GCC PR29808, it's invalid -- use "r" instead.


Segher

Re: 2.6.19-rc5-mm1 fails to compile with gcc 4.2

[Segher Boessenkool]

> 'sel' is a variable, so gcc can't provide it as an immediate  
> operand.  Specifying it as "rm" instead of "g" would have been  
> better, but can't have any real influence.

It can become an immediate operand if the function
gets inlined into a caller.


Segher