system.h asm fixes

system.h asm fixes

[Justin Carlson]

Looks to me like we're missing some proper asm clobber markers:

diff -u -r1.13 system.h
--- system.h	4 Feb 2002 17:35:51 -0000	1.13
+++ system.h	17 Jun 2002 18:18:56 -0000
@@ -41,7 +41,8 @@
 		"__sti"
 		: /* no outputs */
 		: /* no inputs */
-		: "memory");
+		: "$1","memory"
+		);
 }
 
 /*
@@ -73,7 +74,7 @@
 		"__cli"
 		: /* no outputs */
 		: /* no inputs */
-		: "memory");
+		: "$1","memory");
 }
 
 __asm__ (
@@ -110,7 +111,7 @@
 	"__save_and_cli\t%0"						\
 	: "=r" (x)							\
 	: /* no inputs */						\
-	: "memory")
+	: "$1","memory")
 
 __asm__(".macro\t__restore_flags flags\n\t"
 	".set\tpush\n\t"
@@ -136,7 +137,7 @@
 		"__restore_flags\t%0"					\
 		: "=r" (__tmp1)						\
 		: "0" (flags)						\
-		: "memory");						\
+		: "$1","memory");					\
 } while(0)
 
 #ifdef CONFIG_SMP

Re: system.h asm fixes

[Daniel Jacobowitz]

On Mon, Jun 17, 2002 at 11:20:42AM -0700, Justin Carlson wrote:
> Looks to me like we're missing some proper asm clobber markers:

Not really, I think those actually caused a problem at some point but
that might be my imagination.  They certainly aren't necessary.  The
compiler can not use $1 for anything, because it doesn't know what the
assembler might be doing with it.

-- 
Daniel Jacobowitz                           Carnegie Mellon University
MontaVista Software                         Debian GNU/Linux Developer

Re: system.h asm fixes

[Thiemo Seufer]

Daniel Jacobowitz wrote:
> On Mon, Jun 17, 2002 at 11:20:42AM -0700, Justin Carlson wrote:
> > Looks to me like we're missing some proper asm clobber markers:
> 
> Not really, I think those actually caused a problem at some point but
> that might be my imagination.  They certainly aren't necessary.  The
> compiler can not use $1 for anything, because it doesn't know what the
> assembler might be doing with it.

But other assembler routines which run in the meantime may do.


Thiemo

Re: system.h asm fixes

[Ralf Baechle]

In-Reply-To: <1024338042.1463.21.camel@localhost.localdomain>; from justin@cs.cmu.edu on Mon, Jun 17, 2002 at 11:20:42AM -0700

You may consider configuring a hostname for your machine :)

On Mon, Jun 17, 2002 at 11:20:42AM -0700, Justin Carlson wrote:

> Looks to me like we're missing some proper asm clobber markers:

No, as per convention $1 is never used by the compiler per convention,
so clobbering not necessary.  I recently removed all "$1" clobbers to
make the code a bit easier to read.

  Ralf

Re: system.h asm fixes

[Thiemo Seufer]

Ralf Baechle wrote:
[snip]
> > Looks to me like we're missing some proper asm clobber markers:
> 
> No, as per convention $1 is never used by the compiler per convention,
> so clobbering not necessary.  I recently removed all "$1" clobbers to
> make the code a bit easier to read.

How can this work? A grep shows many instances of $1 usage,
I don't think all of this code is interrupt safe.


Thiemo

Re: system.h asm fixes

[Justin Carlson]

On Mon, 2002-06-17 at 15:36, Thiemo Seufer wrote:
> Ralf Baechle wrote:
> [snip]
> > > Looks to me like we're missing some proper asm clobber markers:
> > 
> > No, as per convention $1 is never used by the compiler per convention,
> > so clobbering not necessary.  I recently removed all "$1" clobbers to
> > make the code a bit easier to read.
> 
> How can this work? A grep shows many instances of $1 usage,
> I don't think all of this code is interrupt safe.
> 

The "$x" clobber markers exist so that the *compiler* won't expect
values in those registers to be preserved across the asm call.  It has
nothing to do with safety across interrupts.  In other words, it's there
to prevent something like this:

	foo = 2;   // Compiler sticks foo a register, say, $2	
	asm (
		"  lw $2, baz\n"
		"  sw $2, (%0)\n"
		::"r" (sna));   // Assembly code uses $2 as a temp reg
	(*bar) = foo;      // Compiler erroneously does (*bar) = baz;

In terms of interrupt safety, there is no issue with $1; any interrupt
would save the register before mucking with it, and restore it before
returning.  The only registers for which this is not true are k0 and k1.

As others have rightly pointed out, the compiler isn't allowed to muck
with $1 anyways, as it is defined to be a temporary register for the
assembler.  So the clobber notation isn't necessary.

Make sense?

-Justin

Re: system.h asm fixes

[Ralf Baechle]

On Tue, Jun 18, 2002 at 12:36:50AM +0200, Thiemo Seufer wrote:

> Ralf Baechle wrote:
> [snip]
> > > Looks to me like we're missing some proper asm clobber markers:
> > 
> > No, as per convention $1 is never used by the compiler per convention,
> > so clobbering not necessary.  I recently removed all "$1" clobbers to
> > make the code a bit easier to read.
> 
> How can this work? A grep shows many instances of $1 usage,

Uses by assembler code only, not gcc.  Therefoe we don't need to protect
C code against use of $at.

> I don't think all of this code is interrupt safe.

How this should relate to interrupts is beyond me.  Exception handlers do
their own register saving thing.

  Ralf

Re: system.h asm fixes

[Maciej W. Rozycki]

On Tue, 18 Jun 2002, Ralf Baechle wrote:

> > How can this work? A grep shows many instances of $1 usage,
> 
> Uses by assembler code only, not gcc.  Therefoe we don't need to protect
> C code against use of $at.

 Moreover, gcc assumes ".set at" is always in effect and includes macros
in its output that are known to clobber $1 and don't work with ".set
noat". 

-- 
+  Maciej W. Rozycki, Technical University of Gdansk, Poland   +
+--------------------------------------------------------------+
+        e-mail: macro@ds2.pg.gda.pl, PGP key available        +