.set mips2 breaks 64bit kernel (Re: CVS Update@linux-mips.org: linux)

.set mips2 breaks 64bit kernel (Re: CVS Update@linux-mips.org: linux)

[Atsushi Nemoto]

>>>>> On Tue, 14 Jun 2005 18:35:06 +0100, macro@linux-mips.org said:
macro> Modified files:
macro> 	arch/mips/kernel: semaphore.c 
macro> 	include/asm-mips: atomic.h bitops.h system.h 

macro> Log message:
macro> 	Enable a suitable ISA for the assembler around ll/sc so that code
macro> 	builds even for processors that don't support the instructions.
macro> 	Plus minor formatting fixes.

Please do not set mips2 unconditionaly.  It breaks 64bit kernel.

---
Atsushi Nemoto

Re: .set mips2 breaks 64bit kernel (Re: CVS Update@linux-mips.org: linux)

[Maciej W. Rozycki]

On Wed, 22 Jun 2005, Atsushi Nemoto wrote:

> macro> Log message:
> macro> 	Enable a suitable ISA for the assembler around ll/sc so that code
> macro> 	builds even for processors that don't support the instructions.
> macro> 	Plus minor formatting fixes.
> 
> Please do not set mips2 unconditionaly.  It breaks 64bit kernel.

 Any specifics please?  It works for me -- I wouldn't have committed it 
otherwise.  In particular in the affected range there are no operations 
that would require a 64-bit ISA.

  Maciej

Re: .set mips2 breaks 64bit kernel (Re: CVS Update@linux-mips.org: linux)

[Kumba]

Maciej W. Rozycki wrote:
> 
>  Any specifics please?  It works for me -- I wouldn't have committed it 
> otherwise.  In particular in the affected range there are no operations 
> that would require a 64-bit ISA.
> 
>   Maciej

I've tested to see it break IP30, and Thiemo reported IP27 was also not booting 
(no text display after kernel entry).  Reversing this patch seems to allow both 
systems to boot.  Is it possible it breaks ELF64-only builds?  I haven't tested 
it on IP28 yet, so I can't confirm this (it's also ELF64).


--Kumba

-- 
Gentoo/MIPS Team Lead
Gentoo Foundation Board of Trustees

"Such is oft the course of deeds that move the wheels of the world: small hands 
do them because they must, while the eyes of the great are elsewhere."  --Elrond

Re: .set mips2 breaks 64bit kernel (Re: CVS Update@linux-mips.org: linux)

[Maciej W. Rozycki]

On Wed, 22 Jun 2005, Kumba wrote:

> I've tested to see it break IP30, and Thiemo reported IP27 was also not
> booting (no text display after kernel entry).  Reversing this patch seems to
> allow both systems to boot.  Is it possible it breaks ELF64-only builds?  I
> haven't tested it on IP28 yet, so I can't confirm this (it's also ELF64).

 Well, could I please ask people to be more specific when writing such 
reports?  Unfortunately "break" can mean anything and my crystal ball is 
currently away for a service.  At least you've mentioned problems at a 
bootstrap, so I infer it's not a build-time failure.

 That said, I'm only using ELF64 for my 64-bit builds, so that shouldn't 
be the cause.  But I think there is one possiblity of a problem -- 
obsolete versions of GCC may rely on gas expanding "ll" and "sc" as 
macros, i.e. substitute a name of a symbol rather than a valid 
machine-level address expression (e.g. "0($reg)" or "%lo(sym)($reg)") for 
memory constraints.  Well, by using "m" right now we sort of permit it to.  

 But shouldn't gas actually complain if the ISA or CPU selection forbids 
it doing the right address calculation for the ABI in use?  I think it 
should, so I consider it a bug in gas.  Of course, blaming gas doesn't 
magically make our code right, but the deficiency makes noticing such 
problems tougher.

 As we still strive to support old versions of GCC, I guess a reasonable 
solution is taking an approach similar to that I already have in 
include/asm-mips/bitops.h, i.e. defining __SET_MIPS appropriately and 
using it instead of hardcoded ".set mips2".  I'll prepare a fix shortly.

  Maciej

Re: .set mips2 breaks 64bit kernel

[Atsushi Nemoto]

>>>>> On Wed, 22 Jun 2005 14:30:41 +0100 (BST), "Maciej W. Rozycki" <macro@linux-mips.org> said:

macro> But I think there is one possiblity of a problem -- obsolete
macro> versions of GCC may rely on gas expanding "ll" and "sc" as
macro> macros, i.e. substitute a name of a symbol rather than a valid
macro> machine-level address expression (e.g. "0($reg)" or
macro> "%lo(sym)($reg)") for memory constraints.  Well, by using "m"
macro> right now we sort of permit it to.

Yes, this is my case.

A line in net/key/af_key.c:pfkey_create()

	atomic_inc(&pfkey_socks_nr);

was translated to:

		.set	mips2					
1:	ll	$3, pfkey_socks_nr		# atomic_add		
	addu	$3, 1					
	sc	$3, pfkey_socks_nr					
	beqz	$3, 1b					
	.set	mips0					

Then gas expands the 'll' to LUI and LL (no high 32bit).

I'm using gcc 3.4.4 and binutils 2.16.1.  Not so brand new but not so
obsolete (I hope :-))

---
Atsushi Nemoto

Re: .set mips2 breaks 64bit kernel

[Maciej W. Rozycki]

On Thu, 23 Jun 2005, Atsushi Nemoto wrote:

> Then gas expands the 'll' to LUI and LL (no high 32bit).

 I've fixed it by using ".set mips3" universally -- alternating between it 
and ".set mips2" would be too ugly and the only difference between these 
ISAs are 64-bit operations, so for code that does not use them explicitly 
these settings should be equivalent.

> I'm using gcc 3.4.4 and binutils 2.16.1.  Not so brand new but not so
> obsolete (I hope :-))

 Well, GCC 3.4.4 is a bit old-fashioned, but it's not the worst one and 
certainly we are going to support it for a while yet. ;-)  Binutils 2.16.1 
should be OK -- I don't think there's been any problems discovered with 
them yet that would urge an update.

  Maciej