Re: Gas is broken for x86-64

Re: Gas is broken for x86-64

[Jan Beulich]

>>> Alan Modra <amodra@bigpond.net.au> 08.11.04 09:38:07 >>>
>On Mon, Nov 08, 2004 at 09:20:22AM +0100, Jan Beulich wrote:
>> I would think your patch should have added both suffix-less movsb
and
>> movzb, or neither.
>
>No.  movsb is a string move instruction.  movzb is an abbreviation of
>movzbl or movzbw depending on data size.

Certainly. But movsb is also a suffix-less variant of movsb[wlq], and
since movz and movs are fully in sync with one another, allowing
abbreviation of one should also allow the same for the other.
One of the points here is that the abbreviation (or suffix) guessing
depends on only the mnemonic, not the operands, thus making things like
movsb (movs[bwlq] vs. movsb[wlq]), movsd (movs[bwdq] vs. movs[sd]), or
cmpsd (cmps[bwdq] vs. cmps[sd]) conflict. But there are more such issues
(like AT&T/Intel syntax insns being visible in the opposite mode). I
think some signficiant amount of cleanup work is necessary here, and
I'll try to get to this hopefully not too far in the future.

Jan

Re: Gas is broken for x86-64

[Alan Modra]

On Mon, Nov 08, 2004 at 09:56:13AM +0100, Jan Beulich wrote:
> >>> Alan Modra <amodra@bigpond.net.au> 08.11.04 09:38:07 >>>
> >On Mon, Nov 08, 2004 at 09:20:22AM +0100, Jan Beulich wrote:
> >> I would think your patch should have added both suffix-less movsb
> and
> >> movzb, or neither.
> >
> >No.  movsb is a string move instruction.  movzb is an abbreviation of
> >movzbl or movzbw depending on data size.
> 
> Certainly. But movsb is also a suffix-less variant of movsb[wlq], and

Huh?  Of course it isn't.  We can't allow that precisely because movsb
is a string move.

> since movz and movs are fully in sync with one another, allowing
> abbreviation of one should also allow the same for the other.

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre

Re: Gas is broken for x86-64

[Jan Beulich]

>>> "H. J. Lu" <hjl@lucon.org> 06.11.04 00:17:59 >>>
>On Fri, Nov 05, 2004 at 03:04:29PM -0800, H. J. Lu wrote:
>> On Fri, Nov 05, 2004 at 03:00:27PM -0800, H. J. Lu wrote:
>> > On Fri, Nov 05, 2004 at 02:47:58PM -0800, H. J. Lu wrote:
>> > > Hi Jan,
>> > > 
>> > > Your i386 opcode change breaks x86-64. I got
>> > > 
>> > > [hjl@gnu-64 gas]$ cat /tmp/x.s
>> > > movzb (%rax),%edx
>> > > [hjl@gnu-64 gas]$ ./as-new -o x.o /tmp/x.s
>> > > /tmp/x.s: Assembler messages:
>> > > /tmp/x.s:1: Error: no such instruction: `movzb (%rax),%edx'
>> > > 
>> > > 
>> > 
>> > Your patch
>> > 
>> >
http://sources.redhat.com/ml/binutils/2004-11/msg00033/binutils-mainline-x86-intel-syntax.patch

>> > 
>> > removed movzb. It is wrong.

This change was for consistency reasons, as indicated also by an added
comment. If inconsistencies are intentional, then these should be very
clearly marked in my opinion. In this case, since there's no support
for

movsb (%eax), %eax

I can't see why you'd want

movzb (%eax), %eax

>> BTW, do you know what other instructions you have removed? We can't
>> remove any of them unless we have a very good reason.

None.

>I checked in a patch to restore "movzb":
>
>http://sources.redhat.com/ml/binutils-cvs/2004-11/msg00045.html 

I would think your patch should have added both suffix-less movsb and
movzb, or neither.

Jan

Re: Gas is broken for x86-64

[Alan Modra]

On Mon, Nov 08, 2004 at 09:20:22AM +0100, Jan Beulich wrote:
> I would think your patch should have added both suffix-less movsb and
> movzb, or neither.

No.  movsb is a string move instruction.  movzb is an abbreviation of
movzbl or movzbw depending on data size.

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre

Gas is broken for x86-64

[H. J. Lu]

Hi Jan,

Your i386 opcode change breaks x86-64. I got

[hjl@gnu-64 gas]$ cat /tmp/x.s
movzb (%rax),%edx
[hjl@gnu-64 gas]$ ./as-new -o x.o /tmp/x.s
/tmp/x.s: Assembler messages:
/tmp/x.s:1: Error: no such instruction: `movzb (%rax),%edx'


H.J.

Re: Gas is broken for x86-64

[H. J. Lu]

On Fri, Nov 05, 2004 at 02:47:58PM -0800, H. J. Lu wrote:
> Hi Jan,
> 
> Your i386 opcode change breaks x86-64. I got
> 
> [hjl@gnu-64 gas]$ cat /tmp/x.s
> movzb (%rax),%edx
> [hjl@gnu-64 gas]$ ./as-new -o x.o /tmp/x.s
> /tmp/x.s: Assembler messages:
> /tmp/x.s:1: Error: no such instruction: `movzb (%rax),%edx'
> 
> 

Your patch

http://sources.redhat.com/ml/binutils/2004-11/msg00033/binutils-mainline-x86-intel-syntax.patch

removed movzb. It is wrong.

H.J.

Re: Gas is broken for x86-64

[H. J. Lu]

On Fri, Nov 05, 2004 at 03:00:27PM -0800, H. J. Lu wrote:
> On Fri, Nov 05, 2004 at 02:47:58PM -0800, H. J. Lu wrote:
> > Hi Jan,
> > 
> > Your i386 opcode change breaks x86-64. I got
> > 
> > [hjl@gnu-64 gas]$ cat /tmp/x.s
> > movzb (%rax),%edx
> > [hjl@gnu-64 gas]$ ./as-new -o x.o /tmp/x.s
> > /tmp/x.s: Assembler messages:
> > /tmp/x.s:1: Error: no such instruction: `movzb (%rax),%edx'
> > 
> > 
> 
> Your patch
> 
> http://sources.redhat.com/ml/binutils/2004-11/msg00033/binutils-mainline-x86-intel-syntax.patch
> 
> removed movzb. It is wrong.

BTW, do you know what other instructions you have removed? We can't
remove any of them unless we have a very good reason.


H.J.

Re: Gas is broken for x86-64

[H. J. Lu]

On Fri, Nov 05, 2004 at 03:04:29PM -0800, H. J. Lu wrote:
> On Fri, Nov 05, 2004 at 03:00:27PM -0800, H. J. Lu wrote:
> > On Fri, Nov 05, 2004 at 02:47:58PM -0800, H. J. Lu wrote:
> > > Hi Jan,
> > > 
> > > Your i386 opcode change breaks x86-64. I got
> > > 
> > > [hjl@gnu-64 gas]$ cat /tmp/x.s
> > > movzb (%rax),%edx
> > > [hjl@gnu-64 gas]$ ./as-new -o x.o /tmp/x.s
> > > /tmp/x.s: Assembler messages:
> > > /tmp/x.s:1: Error: no such instruction: `movzb (%rax),%edx'
> > > 
> > > 
> > 
> > Your patch
> > 
> > http://sources.redhat.com/ml/binutils/2004-11/msg00033/binutils-mainline-x86-intel-syntax.patch
> > 
> > removed movzb. It is wrong.
> 
> BTW, do you know what other instructions you have removed? We can't
> remove any of them unless we have a very good reason.
> 

I checked in a patch to restore "movzb":

http://sources.redhat.com/ml/binutils-cvs/2004-11/msg00045.html


H.J.

Re: Gas is broken for x86-64

[Daniel Jacobowitz]

On Fri, Nov 05, 2004 at 03:17:59PM -0800, H. J. Lu wrote:
> On Fri, Nov 05, 2004 at 03:04:29PM -0800, H. J. Lu wrote:
> > On Fri, Nov 05, 2004 at 03:00:27PM -0800, H. J. Lu wrote:
> > > On Fri, Nov 05, 2004 at 02:47:58PM -0800, H. J. Lu wrote:
> > > > Hi Jan,
> > > > 
> > > > Your i386 opcode change breaks x86-64. I got
> > > > 
> > > > [hjl@gnu-64 gas]$ cat /tmp/x.s
> > > > movzb (%rax),%edx
> > > > [hjl@gnu-64 gas]$ ./as-new -o x.o /tmp/x.s
> > > > /tmp/x.s: Assembler messages:
> > > > /tmp/x.s:1: Error: no such instruction: `movzb (%rax),%edx'
> > > > 
> > > > 
> > > 
> > > Your patch
> > > 
> > > http://sources.redhat.com/ml/binutils/2004-11/msg00033/binutils-mainline-x86-intel-syntax.patch
> > > 
> > > removed movzb. It is wrong.
> > 
> > BTW, do you know what other instructions you have removed? We can't
> > remove any of them unless we have a very good reason.
> > 
> 
> I checked in a patch to restore "movzb":
> 
> http://sources.redhat.com/ml/binutils-cvs/2004-11/msg00045.html

Then please add this to the testsuite.

-- 
Daniel Jacobowitz

Re: Gas is broken for x86-64

[H. J. Lu]

On Fri, Nov 05, 2004 at 07:41:36PM -0500, Daniel Jacobowitz wrote:
> On Fri, Nov 05, 2004 at 03:17:59PM -0800, H. J. Lu wrote:
> > On Fri, Nov 05, 2004 at 03:04:29PM -0800, H. J. Lu wrote:
> > > On Fri, Nov 05, 2004 at 03:00:27PM -0800, H. J. Lu wrote:
> > > > On Fri, Nov 05, 2004 at 02:47:58PM -0800, H. J. Lu wrote:
> > > > > Hi Jan,
> > > > > 
> > > > > Your i386 opcode change breaks x86-64. I got
> > > > > 
> > > > > [hjl@gnu-64 gas]$ cat /tmp/x.s
> > > > > movzb (%rax),%edx
> > > > > [hjl@gnu-64 gas]$ ./as-new -o x.o /tmp/x.s
> > > > > /tmp/x.s: Assembler messages:
> > > > > /tmp/x.s:1: Error: no such instruction: `movzb (%rax),%edx'
> > > > > 
> > > > > 
> > > > 
> > > > Your patch
> > > > 
> > > > http://sources.redhat.com/ml/binutils/2004-11/msg00033/binutils-mainline-x86-intel-syntax.patch
> > > > 
> > > > removed movzb. It is wrong.
> > > 
> > > BTW, do you know what other instructions you have removed? We can't
> > > remove any of them unless we have a very good reason.
> > > 
> > 
> > I checked in a patch to restore "movzb":
> > 
> > http://sources.redhat.com/ml/binutils-cvs/2004-11/msg00045.html
> 
> Then please add this to the testsuite.
> 

Good idea. I will check in this patch shortly.


H.J.
---
2004-11-05  H.J. Lu  <hongjiu.lu@intel.com>

	* gas/i386/general.s: Add movzb.
	* gas/i386/general.l: Updated.

--- gas/i386/general.l.movzb	2000-05-02 10:23:27.000000000 -0700
+++ gas/i386/general.l	2004-11-05 17:14:27.350938946 -0800
@@ -285,6 +285,10 @@
  212 021d 67668984 248C0000 		movl	%eax,140\(%esp\)
  212      00
  213                       	
- 214                       		# Force a good alignment.
- 215 0226 00000000 00000000 		.p2align	4,0
- 215      0000
+ 214                       	.code32
+ 215                       	# Make sure that we won't remove movzb by accident.
+ 216 0226 660FB6F8          		movzb	%al,%di
+ 217 022a 0FB6C8            		movzb	%al,%ecx
+ 218                       	
+ 219                       		# Force a good alignment.
+ 220 022d 000000            		.p2align	4,0
--- gas/i386/general.s.movzb	2000-04-03 09:46:40.000000000 -0700
+++ gas/i386/general.s	2004-11-05 17:14:32.479276317 -0800
@@ -211,5 +211,10 @@
 	leal	-1760(%ebp),%ebx
 	movl	%eax,140(%esp)
 
+.code32
+# Make sure that we won't remove movzb by accident.
+	movzb	%al,%di
+	movzb	%al,%ecx
+
 	# Force a good alignment.
 	.p2align	4,0