[Patch] / 0 should send SIGFPE not SIGTRAP...

[Patch] / 0 should send SIGFPE not SIGTRAP...

[David Daney]

I am getting a SIGTRAP whenever an integer divide by 0 happens.  It 
should be sending SIGFPE.

It looks like kernel/traps.c is a little messed up.

The attached patch fixes it for me.

The decoding of the break instruction was selecting the wrong bits.  It 
looks like the trap instruction decoding was messed up also.  The patch 
fixes trap also, but I could not figure out how to get gcc to generate 
the trap form of division, so that part is untested.

David Daney.

Re: [Patch] / 0 should send SIGFPE not SIGTRAP...

[David Daney]

[-- Attachment #1: Type: text/plain, Size: 571 bytes --]

It might help if I attached the patch.  Here it is...

David Daney wrote:

> I am getting a SIGTRAP whenever an integer divide by 0 happens.  It 
> should be sending SIGFPE.
>
> It looks like kernel/traps.c is a little messed up.
>
> The attached patch fixes it for me.
>
> The decoding of the break instruction was selecting the wrong bits.  
> It looks like the trap instruction decoding was messed up also.  The 
> patch fixes trap also, but I could not figure out how to get gcc to 
> generate the trap form of division, so that part is untested.
>
> David Daney.
>


[-- Attachment #2: traps.diff --]
[-- Type: text/plain, Size: 699 bytes --]

--- ../linux-avtrex/linux/arch/mips/kernel/traps.c	2004-02-26 11:14:09.000000000 -0800
+++ arch/mips/kernel/traps.c	2004-06-11 10:13:59.000000000 -0700
@@ -598,7 +598,7 @@
 	 * code starts left to bit 16 instead to bit 6 in the opcode.
 	 * Gas is bug-compatible ...
 	 */
-	bcode = ((opcode >> 16) & ((1 << 20) - 1));
+	bcode = ((opcode >> 6) & ((1 << 20) - 1));
 
 	/*
 	 * (A short test says that IRIX 5.3 sends SIGTRAP for all break
@@ -633,7 +633,7 @@
 
 	/* Immediate versions don't provide a code.  */
 	if (!(opcode & OPCODE))
-		tcode = ((opcode >> 6) & ((1 << 20) - 1));
+		tcode = ((opcode >> 6) & ((1 << 10) - 1));
 
 	/*
 	 * (A short test says that IRIX 5.3 sends SIGTRAP for all trap

Re: [Patch] / 0 should send SIGFPE not SIGTRAP...

[David Daney]

I guess I didn't fully read the comment above this section of code.

    /*
     * There is the ancient bug in the MIPS assemblers that the break
     * code starts left to bit 16 instead to bit 6 in the opcode.
     * Gas is bug-compatible ...
     */

I am using gcc-3.3.1/binutils-2.15.  With this toolchain, I get break 
instructions that comply with the MIPS documentation for break instructions:


00000000 <do_div>:
   0:   3c1c0000        lui     gp,0x0
   4:   279c0000        addiu   gp,gp,0
   8:   0399e021        addu    gp,gp,t9
   c:   0085001a        div     zero,a0,a1
  10:   14a00002        bnez    a1,1c <do_div+0x1c>
  14:   00000000        nop
  18:   000001cd        break   0x7
  1c:   00001012        mflo    v0
  20:   03e00008        jr      ra
  24:   00000000        nop
 

What to do?

David Daney

David Daney wrote:

> It might help if I attached the patch.  Here it is...
>
> David Daney wrote:
>
>> I am getting a SIGTRAP whenever an integer divide by 0 happens.  It 
>> should be sending SIGFPE.
>>
>> It looks like kernel/traps.c is a little messed up.
>>
>> The attached patch fixes it for me.
>>
>> The decoding of the break instruction was selecting the wrong bits.  
>> It looks like the trap instruction decoding was messed up also.  The 
>> patch fixes trap also, but I could not figure out how to get gcc to 
>> generate the trap form of division, so that part is untested.
>>
>> David Daney.
>>
>
>------------------------------------------------------------------------
>
>--- ../linux-avtrex/linux/arch/mips/kernel/traps.c	2004-02-26 11:14:09.000000000 -0800
>+++ arch/mips/kernel/traps.c	2004-06-11 10:13:59.000000000 -0700
>@@ -598,7 +598,7 @@
> 	 * code starts left to bit 16 instead to bit 6 in the opcode.
> 	 * Gas is bug-compatible ...
> 	 */
>-	bcode = ((opcode >> 16) & ((1 << 20) - 1));
>+	bcode = ((opcode >> 6) & ((1 << 20) - 1));
> 
> 	/*
> 	 * (A short test says that IRIX 5.3 sends SIGTRAP for all break
>@@ -633,7 +633,7 @@
> 
> 	/* Immediate versions don't provide a code.  */
> 	if (!(opcode & OPCODE))
>-		tcode = ((opcode >> 6) & ((1 << 20) - 1));
>+		tcode = ((opcode >> 6) & ((1 << 10) - 1));
> 
> 	/*
> 	 * (A short test says that IRIX 5.3 sends SIGTRAP for all trap
>  
>

Re: [Patch] / 0 should send SIGFPE not SIGTRAP...

[Maciej W. Rozycki]

On Fri, 11 Jun 2004, David Daney wrote:

> I guess I didn't fully read the comment above this section of code.
> 
>     /*
>      * There is the ancient bug in the MIPS assemblers that the break
>      * code starts left to bit 16 instead to bit 6 in the opcode.
>      * Gas is bug-compatible ...
>      */
> 
> I am using gcc-3.3.1/binutils-2.15.  With this toolchain, I get break 
> instructions that comply with the MIPS documentation for break instructions:

 Well, I did some research and I'm afraid it dates back to a commit from
2000-12-01, when a different interpretation of "break" was introduced for
the MIPS32/64 ISA.  So the interpretation of the "break" instruction
depends on the "-march" setting and moreover, only for instructions
requested explicitly.  For ones emitted implicitly as a result of division
and multiplication macros, the interpretation is always the "traditional"
one (the "c" vs the "B" code).

> 00000000 <do_div>:
>    0:   3c1c0000        lui     gp,0x0
>    4:   279c0000        addiu   gp,gp,0
>    8:   0399e021        addu    gp,gp,t9
>    c:   0085001a        div     zero,a0,a1
>   10:   14a00002        bnez    a1,1c <do_div+0x1c>
>   14:   00000000        nop
>   18:   000001cd        break   0x7
>   1c:   00001012        mflo    v0
>   20:   03e00008        jr      ra
>   24:   00000000        nop
>  
> 
> What to do?

1. I think Linux can intercept both the "traditional" and MIPS32/64 values
-- break codes are not used that extensively this would risk running out
of them.  The set of known ones can be seen in <asm/break.h>.  However,
this won't aid userland trying to interpret code (there may be none such, 
though).

2. Gas should definitely use the codes consistently.  And it's a pity the
ABI got broken -- I think another mnemonic should have been chosen for the
correct implementation of "break", available to any ISA.

3. GCC should probably use traps for anything above MIPS I, anyway.  
Perhaps with an option, like for gas, to select the alternative.

4. Perhaps something else.

  Maciej

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

Re: [Patch] / 0 should send SIGFPE not SIGTRAP...

[cgd]

At Fri, 11 Jun 2004 19:14:11 +0000 (UTC), "Maciej W. Rozycki" wrote:
> 2. Gas should definitely use the codes consistently.  And it's a pity the
> ABI got broken -- I think another mnemonic should have been chosen for the
> correct implementation of "break", available to any ISA.

in retrospect, the 'B' variation probably wasn't the greatest idea.

If it were removed (leaving 'c' and 'c','q' variations), I don't know
that any real harm would occur.

It may be very confusing to people who expect that the break code will
translate into the instruction in an obvious way, and obviously it
would mess up use of 20-bit codes, but i don't know how prevalent that
is.

Unfortunately, at this point, Linux should probably accept the
divide-by-zero code in both locations.


(Really, from day one, assemblers probably should have accepted a
20-bit code.  I just checked my copy of the Kane r2000/r3000 book, and
it was 20-bit all the way back then.  If i had to guess, i'd guess
that gas was copying a non-gnu assembler's behaviour.  In any case,
water under the bridge.)



cgd

Re: [Patch] / 0 should send SIGFPE not SIGTRAP...

[Maciej W. Rozycki]

On Fri, 11 Jun 2004 cgd@broadcom.com wrote:

> > 2. Gas should definitely use the codes consistently.  And it's a pity the
> > ABI got broken -- I think another mnemonic should have been chosen for the
> > correct implementation of "break", available to any ISA.
> 
> in retrospect, the 'B' variation probably wasn't the greatest idea.

 I guess it may be useful for something to have 20-bit codes available.  
Though except these few special cases, breaks tend to be inserted at the
run time, so it's the interested software that decides how to interpret
them, not gas.

> It may be very confusing to people who expect that the break code will
> translate into the instruction in an obvious way, and obviously it
> would mess up use of 20-bit codes, but i don't know how prevalent that
> is.

 I was surprised at first, too.

> Unfortunately, at this point, Linux should probably accept the
> divide-by-zero code in both locations.

 I think that's not a big trouble for Linux -- the path is rare and not
critical for performance.

> (Really, from day one, assemblers probably should have accepted a
> 20-bit code.  I just checked my copy of the Kane r2000/r3000 book, and
> it was 20-bit all the way back then.  If i had to guess, i'd guess
> that gas was copying a non-gnu assembler's behaviour.  In any case,
> water under the bridge.)

 Definitely they should have.  It's bug-compatibility with the original
MIPS assembler, I'm told.

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

Re: [Patch] / 0 should send SIGFPE not SIGTRAP...

[David Daney]

[-- Attachment #1: Type: text/plain, Size: 364 bytes --]

Maciej W. Rozycki wrote:

>On Fri, 11 Jun 2004 cgd@broadcom.com wrote:
>  
>
>>Unfortunately, at this point, Linux should probably accept the
>>divide-by-zero code in both locations.
>>    
>>
>
> I think that's not a big trouble for Linux -- the path is rare and not
>critical for performance.
>
>  
>
How about the attached (lightly tested) patch?

David Daney.

[-- Attachment #2: traps.diff --]
[-- Type: text/plain, Size: 2111 bytes --]

*** ../linux-avtrex/linux/arch/mips/kernel/traps.c	2004-02-26 11:14:09.000000000 -0800
--- arch/mips/kernel/traps.c	2004-06-11 13:43:00.000000000 -0700
***************
*** 597,615 ****
  	 * There is the ancient bug in the MIPS assemblers that the break
  	 * code starts left to bit 16 instead to bit 6 in the opcode.
  	 * Gas is bug-compatible ...
! 	 */
! 	bcode = ((opcode >> 16) & ((1 << 20) - 1));
! 
! 	/*
  	 * (A short test says that IRIX 5.3 sends SIGTRAP for all break
  	 * insns, even for break codes that indicate arithmetic failures.
  	 * Weird ...)
  	 * But should we continue the brokenness???  --macro
  	 */
  	switch (bcode) {
! 	case 6:
! 	case 7:
! 		if (bcode == 7)
  			info.si_code = FPE_INTDIV;
  		else
  			info.si_code = FPE_INTOVF;
--- 597,621 ----
  	 * There is the ancient bug in the MIPS assemblers that the break
  	 * code starts left to bit 16 instead to bit 6 in the opcode.
  	 * Gas is bug-compatible ...
! 	 *
  	 * (A short test says that IRIX 5.3 sends SIGTRAP for all break
  	 * insns, even for break codes that indicate arithmetic failures.
  	 * Weird ...)
  	 * But should we continue the brokenness???  --macro
+          *
+          * It seems some assemblers (binutils-2.15 for example) assemble
+          * break correctly.  So we check for the break code in either
+          * position.
+          *
  	 */
+ 
+ 	bcode = ((opcode >> 6) & ((1 << 20) - 1));
  	switch (bcode) {
! 	case 0x0006:
! 	case 0x0007:
!         case 0x1800: /* 6 << 10 */
!         case 0x1c00: /* 7 << 10 */
! 		if (bcode == 0x7 || bcode == 0x1c00)
  			info.si_code = FPE_INTDIV;
  		else
  			info.si_code = FPE_INTOVF;
***************
*** 633,639 ****
  
  	/* Immediate versions don't provide a code.  */
  	if (!(opcode & OPCODE))
! 		tcode = ((opcode >> 6) & ((1 << 20) - 1));
  
  	/*
  	 * (A short test says that IRIX 5.3 sends SIGTRAP for all trap
--- 639,645 ----
  
  	/* Immediate versions don't provide a code.  */
  	if (!(opcode & OPCODE))
! 		tcode = ((opcode >> 6) & ((1 << 10) - 1));
  
  	/*
  	 * (A short test says that IRIX 5.3 sends SIGTRAP for all trap

Re: [Patch] (revised patch) / 0 should send SIGFPE not SIGTRAP

[David Daney]

[-- Attachment #1: Type: text/plain, Size: 260 bytes --]

David Daney wrote:

> How about the attached (lightly tested) patch?
>
I will quit sending patches after this one.  It is equivalent to the 
previous version, except it uses the symbolic names of the break codes 
instead of the numeric values.

David Daney.



[-- Attachment #2: traps.diff --]
[-- Type: text/plain, Size: 2382 bytes --]

*** ../linux-avtrex/linux/arch/mips/kernel/traps.c	2004-02-26 11:14:09.000000000 -0800
--- arch/mips/kernel/traps.c	2004-06-11 14:04:52.000000000 -0700
***************
*** 38,43 ****
--- 38,44 ----
  #include <asm/traps.h>
  #include <asm/uaccess.h>
  #include <asm/mmu_context.h>
+ #include <asm/break.h>
  
  extern asmlinkage void handle_mod(void);
  extern asmlinkage void handle_tlbl(void);
***************
*** 597,615 ****
  	 * There is the ancient bug in the MIPS assemblers that the break
  	 * code starts left to bit 16 instead to bit 6 in the opcode.
  	 * Gas is bug-compatible ...
! 	 */
! 	bcode = ((opcode >> 16) & ((1 << 20) - 1));
! 
! 	/*
  	 * (A short test says that IRIX 5.3 sends SIGTRAP for all break
  	 * insns, even for break codes that indicate arithmetic failures.
  	 * Weird ...)
  	 * But should we continue the brokenness???  --macro
  	 */
  	switch (bcode) {
! 	case 6:
! 	case 7:
! 		if (bcode == 7)
  			info.si_code = FPE_INTDIV;
  		else
  			info.si_code = FPE_INTOVF;
--- 598,622 ----
  	 * There is the ancient bug in the MIPS assemblers that the break
  	 * code starts left to bit 16 instead to bit 6 in the opcode.
  	 * Gas is bug-compatible ...
! 	 *
  	 * (A short test says that IRIX 5.3 sends SIGTRAP for all break
  	 * insns, even for break codes that indicate arithmetic failures.
  	 * Weird ...)
  	 * But should we continue the brokenness???  --macro
+          *
+          * It seems some assemblers (binutils-2.15 for example) assemble
+          * break correctly.  So we check for the break code in either
+          * position.
+          *
  	 */
+ 
+ 	bcode = ((opcode >> 6) & ((1 << 20) - 1));
  	switch (bcode) {
! 	case BRK_OVERFLOW:
! 	case BRK_DIVZERO:
!         case BRK_OVERFLOW << 10:
!         case BRK_DIVZERO << 10:
! 		if (bcode == BRK_DIVZERO || bcode == (BRK_DIVZERO << 10))
  			info.si_code = FPE_INTDIV;
  		else
  			info.si_code = FPE_INTOVF;
***************
*** 633,639 ****
  
  	/* Immediate versions don't provide a code.  */
  	if (!(opcode & OPCODE))
! 		tcode = ((opcode >> 6) & ((1 << 20) - 1));
  
  	/*
  	 * (A short test says that IRIX 5.3 sends SIGTRAP for all trap
--- 640,646 ----
  
  	/* Immediate versions don't provide a code.  */
  	if (!(opcode & OPCODE))
! 		tcode = ((opcode >> 6) & ((1 << 10) - 1));
  
  	/*
  	 * (A short test says that IRIX 5.3 sends SIGTRAP for all trap

Re: [Patch] (revised patch) / 0 should send SIGFPE not SIGTRAP

[Geert Uytterhoeven]

On Fri, 11 Jun 2004, David Daney wrote:
> David Daney wrote:
> > How about the attached (lightly tested) patch?
> >
> I will quit sending patches after this one.  It is equivalent to the
> previous version, except it uses the symbolic names of the break codes
> instead of the numeric values.

Please send one more, where you use `diff -up' :-)

Gr{oetje,eeting}s,

						Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
							    -- Linus Torvalds

Re: [Patch] (revised patch) / 0 should send SIGFPE not SIGTRAP

[Maciej W. Rozycki]

On Sun, 13 Jun 2004, Geert Uytterhoeven wrote:

> Please send one more, where you use `diff -up' :-)

 No need to -- I've reimplemented it a bit differently meanwhile.  Any 
objections to the following changes?

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

patch-mips-2.4.26-20040531-mips-bp-tr-0
diff -up --recursive --new-file linux-mips-2.4.26-20040531.macro/arch/mips/kernel/traps.c linux-mips-2.4.26-20040531/arch/mips/kernel/traps.c
--- linux-mips-2.4.26-20040531.macro/arch/mips/kernel/traps.c	2004-03-13 03:56:44.000000000 +0000
+++ linux-mips-2.4.26-20040531/arch/mips/kernel/traps.c	2004-06-13 20:24:32.000000000 +0000
@@ -9,7 +9,7 @@
  * Copyright (C) 1999 Silicon Graphics, Inc.
  * Kevin D. Kissell, kevink@mips.com and Carsten Langgaard, carstenl@mips.com
  * Copyright (C) 2000, 01 MIPS Technologies, Inc.
- * Copyright (C) 2002, 2003  Maciej W. Rozycki
+ * Copyright (C) 2002, 2003, 2004  Maciej W. Rozycki
  */
 #include <linux/config.h>
 #include <linux/init.h>
@@ -22,6 +22,7 @@
 
 #include <asm/bootinfo.h>
 #include <asm/branch.h>
+#include <asm/break.h>
 #include <asm/cpu.h>
 #include <asm/fpu.h>
 #include <asm/cachectl.h>
@@ -596,9 +597,12 @@ asmlinkage void do_bp(struct pt_regs *re
 	/*
 	 * There is the ancient bug in the MIPS assemblers that the break
 	 * code starts left to bit 16 instead to bit 6 in the opcode.
-	 * Gas is bug-compatible ...
+	 * Gas is bug-compatible, but not always, grrr...
+	 * We handle both cases with a simple heuristics.  --macro
 	 */
-	bcode = ((opcode >> 16) & ((1 << 20) - 1));
+	bcode = ((opcode >> 6) & ((1 << 20) - 1));
+	if (bcode < (1 << 10))
+		bcode <<= 10;
 
 	/*
 	 * (A short test says that IRIX 5.3 sends SIGTRAP for all break
@@ -607,9 +611,9 @@ asmlinkage void do_bp(struct pt_regs *re
 	 * But should we continue the brokenness???  --macro
 	 */
 	switch (bcode) {
-	case 6:
-	case 7:
-		if (bcode == 7)
+	case BRK_OVERFLOW << 10:
+	case BRK_DIVZERO << 10:
+		if (bcode == (BRK_DIVZERO << 10))
 			info.si_code = FPE_INTDIV;
 		else
 			info.si_code = FPE_INTOVF;
@@ -633,7 +637,7 @@ asmlinkage void do_tr(struct pt_regs *re
 
 	/* Immediate versions don't provide a code.  */
 	if (!(opcode & OPCODE))
-		tcode = ((opcode >> 6) & ((1 << 20) - 1));
+		tcode = ((opcode >> 6) & ((1 << 10) - 1));
 
 	/*
 	 * (A short test says that IRIX 5.3 sends SIGTRAP for all trap
@@ -642,9 +646,9 @@ asmlinkage void do_tr(struct pt_regs *re
 	 * But should we continue the brokenness???  --macro
 	 */
 	switch (tcode) {
-	case 6:
-	case 7:
-		if (tcode == 7)
+	case BRK_OVERFLOW:
+	case BRK_DIVZERO:
+		if (tcode == BRK_DIVZERO)
 			info.si_code = FPE_INTDIV;
 		else
 			info.si_code = FPE_INTOVF;
diff -up --recursive --new-file linux-mips-2.4.26-20040531.macro/arch/mips64/kernel/traps.c linux-mips-2.4.26-20040531/arch/mips64/kernel/traps.c
--- linux-mips-2.4.26-20040531.macro/arch/mips64/kernel/traps.c	2004-03-13 03:56:45.000000000 +0000
+++ linux-mips-2.4.26-20040531/arch/mips64/kernel/traps.c	2004-06-13 20:26:01.000000000 +0000
@@ -9,7 +9,7 @@
  * Copyright (C) 1999 Silicon Graphics, Inc.
  * Kevin D. Kissell, kevink@mips.com and Carsten Langgaard, carstenl@mips.com
  * Copyright (C) 2000, 01 MIPS Technologies, Inc.
- * Copyright (C) 2002, 2003  Maciej W. Rozycki
+ * Copyright (C) 2002, 2003, 2004  Maciej W. Rozycki
  */
 #include <linux/config.h>
 #include <linux/init.h>
@@ -22,6 +22,7 @@
 
 #include <asm/bootinfo.h>
 #include <asm/branch.h>
+#include <asm/break.h>
 #include <asm/cpu.h>
 #include <asm/fpu.h>
 #include <asm/module.h>
@@ -606,9 +607,12 @@ asmlinkage void do_bp(struct pt_regs *re
 	/*
 	 * There is the ancient bug in the MIPS assemblers that the break
 	 * code starts left to bit 16 instead to bit 6 in the opcode.
-	 * Gas is bug-compatible ...
+	 * Gas is bug-compatible, but not always, grrr...
+	 * We handle both cases with a simple heuristics.  --macro
 	 */
-	bcode = ((opcode >> 16) & ((1 << 20) - 1));
+	bcode = ((opcode >> 6) & ((1 << 20) - 1));
+	if (bcode < (1 << 10))
+		bcode <<= 10;
 
 	/*
 	 * (A short test says that IRIX 5.3 sends SIGTRAP for all break
@@ -617,9 +621,9 @@ asmlinkage void do_bp(struct pt_regs *re
 	 * But should we continue the brokenness???  --macro
 	 */
 	switch (bcode) {
-	case 6:
-	case 7:
-		if (bcode == 7)
+	case BRK_OVERFLOW << 10:
+	case BRK_DIVZERO << 10:
+		if (bcode == (BRK_DIVZERO << 10))
 			info.si_code = FPE_INTDIV;
 		else
 			info.si_code = FPE_INTOVF;
@@ -643,7 +647,7 @@ asmlinkage void do_tr(struct pt_regs *re
 
 	/* Immediate versions don't provide a code.  */
 	if (!(opcode & OPCODE))
-		tcode = ((opcode >> 6) & ((1 << 20) - 1));
+		tcode = ((opcode >> 6) & ((1 << 10) - 1));
 
 	/*
 	 * (A short test says that IRIX 5.3 sends SIGTRAP for all trap
@@ -652,9 +656,9 @@ asmlinkage void do_tr(struct pt_regs *re
 	 * But should we continue the brokenness???  --macro
 	 */
 	switch (tcode) {
-	case 6:
-	case 7:
-		if (tcode == 7)
+	case BRK_OVERFLOW:
+	case BRK_DIVZERO:
+		if (tcode == BRK_DIVZERO)
 			info.si_code = FPE_INTDIV;
 		else
 			info.si_code = FPE_INTOVF;

Re: [Patch] / 0 should send SIGFPE not SIGTRAP...

[Maciej W. Rozycki]

On Fri, 11 Jun 2004 cgd@broadcom.com wrote:

> in retrospect, the 'B' variation probably wasn't the greatest idea.
> 
> If it were removed (leaving 'c' and 'c','q' variations), I don't know
> that any real harm would occur.
> 
> It may be very confusing to people who expect that the break code will
> translate into the instruction in an obvious way, and obviously it
> would mess up use of 20-bit codes, but i don't know how prevalent that
> is.
> 
> Unfortunately, at this point, Linux should probably accept the
> divide-by-zero code in both locations.
> 
> 
> (Really, from day one, assemblers probably should have accepted a
> 20-bit code.  I just checked my copy of the Kane r2000/r3000 book, and
> it was 20-bit all the way back then.  If i had to guess, i'd guess
> that gas was copying a non-gnu assembler's behaviour.  In any case,
> water under the bridge.)

 As it's at least annoying to have different break codes for divisions 
expanded by gcc explicitly and ones created implicitly by gas, here's the 
most reasonable (IMO) approach to fix that.  I think it should have been 
implemented this way originally (if at all).

gas/testsuite/:
2004-06-22  Maciej W. Rozycki  <macro@ds2.pg.gda.pl>

	* gas/mips/break20.s: Test the "break20" alias.
	* gas/mips/break20.d: Results for the test.
	* gas/mips/mips32.s: Replace "break" with "break20".
	* gas/mips/set-arch.s: Likewise.
	* gas/mips/mips32.d: Adjust for the new output.
	* gas/mips/set-arch.d: Likewise.

opcodes/:
2004-06-22  Maciej W. Rozycki  <macro@ds2.pg.gda.pl>

	* mips-opc.c (mips_builtin_opcodes): Replace the MIPS32 ISA 
	specific "break" encoding with a "break20" alias accepted for any 
	ISA.

 I decided to give a precedence to "break x,y" over "break20 z" to avoid 
perhaps a bit surprising output in `objdump'.

 The question is: does anyone know of possible troubles with this change?  
Chances are someone depends on the current semantics for
-march=mips32/mips64, but that's fragile anyway as building with a
different setting results in different code.

 Or should we get rid of the 20-bit "break" completely?  The two-argument
version provides the same functionality, although the 10-bit codes to be
used do not map to the 20-bit equivalent "optically" very well.  
Especially if decimal notation is used.

 Note the same problem appears to be the case with "sdbbp".  It could be
handled similarly, but given the limited use of "sdbbp" for
non-MIPS32/MIPS64 ISAs it may not be worth the hassle.  I see no problem
writing a similar fix if desired, though.

  Maciej

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

binutils-2.15.91-20040615-mips-break20.patch
diff -up --recursive --new-file binutils-2.15.91-20040615.macro/gas/testsuite/gas/mips/break20.d binutils-2.15.91-20040615/gas/testsuite/gas/mips/break20.d
--- binutils-2.15.91-20040615.macro/gas/testsuite/gas/mips/break20.d	2003-05-08 03:25:31.000000000 +0000
+++ binutils-2.15.91-20040615/gas/testsuite/gas/mips/break20.d	2004-06-17 21:03:30.000000000 +0000
@@ -10,9 +10,13 @@ Disassembly of section .text:
 0+0008 <[^>]*> break	0x14
 0+000c <[^>]*> break	0x14,0x28
 0+0010 <[^>]*> break	0x3ff,0x3ff
-0+0014 <[^>]*> sdbbp
-0+0018 <[^>]*> sdbbp
-0+001c <[^>]*> sdbbp	0x14
-0+0020 <[^>]*> sdbbp	0x14,0x28
-0+0024 <[^>]*> sdbbp	0x3ff,0x3ff
+0+0014 <[^>]*> break
+0+0018 <[^>]*> break	0x0,0x14
+0+001c <[^>]*> break	0x14,0x28
+0+0020 <[^>]*> break	0x3ff,0x3ff
+0+0024 <[^>]*> sdbbp
+0+0028 <[^>]*> sdbbp
+0+002c <[^>]*> sdbbp	0x14
+0+0030 <[^>]*> sdbbp	0x14,0x28
+0+0034 <[^>]*> sdbbp	0x3ff,0x3ff
 	...
diff -up --recursive --new-file binutils-2.15.91-20040615.macro/gas/testsuite/gas/mips/break20.s binutils-2.15.91-20040615/gas/testsuite/gas/mips/break20.s
--- binutils-2.15.91-20040615.macro/gas/testsuite/gas/mips/break20.s	1999-08-10 14:24:45.000000000 +0000
+++ binutils-2.15.91-20040615/gas/testsuite/gas/mips/break20.s	2004-06-17 19:57:51.000000000 +0000
@@ -6,6 +6,11 @@ foo:	
 	break	20,40
 	break	1023,1023
 	
+	break20	0
+	break20	20
+	break20	20520
+	break20	1048575
+
 	sdbbp
 	sdbbp	0
 	sdbbp	20
diff -up --recursive --new-file binutils-2.15.91-20040615.macro/gas/testsuite/gas/mips/mips32.d binutils-2.15.91-20040615/gas/testsuite/gas/mips/mips32.d
--- binutils-2.15.91-20040615.macro/gas/testsuite/gas/mips/mips32.d	2003-05-08 03:25:35.000000000 +0000
+++ binutils-2.15.91-20040615/gas/testsuite/gas/mips/mips32.d	2004-06-17 21:15:04.000000000 +0000
@@ -48,7 +48,7 @@ Disassembly of section .text:
 0+0098 <[^>]*> 4359e260 	wait	0x56789
 0+009c <[^>]*> 0000000d 	break
 0+00a0 <[^>]*> 0000000d 	break
-0+00a4 <[^>]*> 0048d14d 	break	0x12345
+0+00a4 <[^>]*> 0048d14d 	break	0x48,0x345
 0+00a8 <[^>]*> 7000003f 	sdbbp
 0+00ac <[^>]*> 7000003f 	sdbbp
 0+00b0 <[^>]*> 7159e27f 	sdbbp	0x56789
diff -up --recursive --new-file binutils-2.15.91-20040615.macro/gas/testsuite/gas/mips/mips32.s binutils-2.15.91-20040615/gas/testsuite/gas/mips/mips32.s
--- binutils-2.15.91-20040615.macro/gas/testsuite/gas/mips/mips32.s	2002-09-05 03:25:40.000000000 +0000
+++ binutils-2.15.91-20040615/gas/testsuite/gas/mips/mips32.s	2004-06-17 21:13:22.000000000 +0000
@@ -62,7 +62,7 @@ text_label:
       # different.
       break
       break   0                       # disassembles without code
-      break   0x12345
+      break20 0x12345
       sdbbp
       sdbbp   0                       # disassembles without code
       sdbbp   0x56789
diff -up --recursive --new-file binutils-2.15.91-20040615.macro/gas/testsuite/gas/mips/set-arch.d binutils-2.15.91-20040615/gas/testsuite/gas/mips/set-arch.d
--- binutils-2.15.91-20040615.macro/gas/testsuite/gas/mips/set-arch.d	2003-10-01 03:25:36.000000000 +0000
+++ binutils-2.15.91-20040615/gas/testsuite/gas/mips/set-arch.d	2004-06-17 21:17:46.000000000 +0000
@@ -160,7 +160,7 @@ Disassembly of section \.text:
 00000260 <[^>]*> 4359e260 	wait	0x56789
 00000264 <[^>]*> 0000000d 	break
 00000268 <[^>]*> 0000000d 	break
-0000026c <[^>]*> 0048d14d 	break	0x12345
+0000026c <[^>]*> 0048d14d 	break	0x48,0x345
 00000270 <[^>]*> 7000003f 	sdbbp
 00000274 <[^>]*> 7000003f 	sdbbp
 00000278 <[^>]*> 7159e27f 	sdbbp	0x56789
diff -up --recursive --new-file binutils-2.15.91-20040615.macro/gas/testsuite/gas/mips/set-arch.s binutils-2.15.91-20040615/gas/testsuite/gas/mips/set-arch.s
--- binutils-2.15.91-20040615.macro/gas/testsuite/gas/mips/set-arch.s	2003-06-29 19:41:33.000000000 +0000
+++ binutils-2.15.91-20040615/gas/testsuite/gas/mips/set-arch.s	2004-06-17 21:18:01.000000000 +0000
@@ -204,7 +204,7 @@ text_label:	
 	# different.
 	break
 	break   0                       # disassembles without code
-	break   0x12345
+	break20 0x12345
 	sdbbp
 	sdbbp   0                       # disassembles without code
 	sdbbp   0x56789
diff -up --recursive --new-file binutils-2.15.91-20040615.macro/opcodes/mips-opc.c binutils-2.15.91-20040615/opcodes/mips-opc.c
--- binutils-2.15.91-20040615.macro/opcodes/mips-opc.c	2003-11-19 04:25:23.000000000 +0000
+++ binutils-2.15.91-20040615/opcodes/mips-opc.c	2004-06-17 19:46:23.000000000 +0000
@@ -274,9 +274,9 @@ const struct mips_opcode mips_builtin_op
 {"bnel",    "s,t,p",	0x54000000, 0xfc000000,	CBL|RD_s|RD_t, 		I2|T3	},
 {"bnel",    "s,I,p",	0,    (int) M_BNEL_I,	INSN_MACRO,		I2|T3	},
 {"break",   "",		0x0000000d, 0xffffffff,	TRAP,			I1	},
-{"break",   "B",        0x0000000d, 0xfc00003f, TRAP,           	I32     },
 {"break",   "c",	0x0000000d, 0xfc00ffff,	TRAP,			I1	},
 {"break",   "c,q",	0x0000000d, 0xfc00003f,	TRAP,			I1	},
+{"break20", "B",        0x0000000d, 0xfc00003f, TRAP,           	I1      },
 {"c.f.d",   "S,T",	0x46200030, 0xffe007ff,	RD_S|RD_T|WR_CC|FP_D,	I1	},
 {"c.f.d",   "M,S,T",    0x46200030, 0xffe000ff, RD_S|RD_T|WR_CC|FP_D,   I4|I32	},
 {"c.f.s",   "S,T",      0x46000030, 0xffe007ff, RD_S|RD_T|WR_CC|FP_S,   I1      },

Re: [Patch] / 0 should send SIGFPE not SIGTRAP...

[David Daney]

Maciej W. Rozycki wrote:

>On Fri, 11 Jun 2004 cgd@broadcom.com wrote:
>
>  
>
>>in retrospect, the 'B' variation probably wasn't the greatest idea.
>>
>>If it were removed (leaving 'c' and 'c','q' variations), I don't know
>>that any real harm would occur.
>>
>>It may be very confusing to people who expect that the break code will
>>translate into the instruction in an obvious way, and obviously it
>>would mess up use of 20-bit codes, but i don't know how prevalent that
>>is.
>>
>>Unfortunately, at this point, Linux should probably accept the
>>divide-by-zero code in both locations.
>>
>>
>>(Really, from day one, assemblers probably should have accepted a
>>20-bit code.  I just checked my copy of the Kane r2000/r3000 book, and
>>it was 20-bit all the way back then.  If i had to guess, i'd guess
>>that gas was copying a non-gnu assembler's behaviour.  In any case,
>>water under the bridge.)
>>    
>>
>
> As it's at least annoying to have different break codes for divisions 
>expanded by gcc explicitly and ones created implicitly by gas, here's the 
>most reasonable (IMO) approach to fix that.  I think it should have been 
>implemented this way originally (if at all).
>
>gas/testsuite/:
>2004-06-22  Maciej W. Rozycki  <macro@ds2.pg.gda.pl>
>
>	* gas/mips/break20.s: Test the "break20" alias.
>	* gas/mips/break20.d: Results for the test.
>	* gas/mips/mips32.s: Replace "break" with "break20".
>	* gas/mips/set-arch.s: Likewise.
>	* gas/mips/mips32.d: Adjust for the new output.
>	* gas/mips/set-arch.d: Likewise.
>
>opcodes/:
>2004-06-22  Maciej W. Rozycki  <macro@ds2.pg.gda.pl>
>
>	* mips-opc.c (mips_builtin_opcodes): Replace the MIPS32 ISA 
>	specific "break" encoding with a "break20" alias accepted for any 
>	ISA.
>
>  
>
.
.
.
Just out of curiosity, do you propose this patch in lieu of the patch to 
Linux's traps.c?

Or would you do both?

It seems like both would be best, as there are already "broken" binutils 
floating around out there.

Also nobody has objected to the kernel patch...

David Daney.

Re: [Patch] / 0 should send SIGFPE not SIGTRAP...

[Maciej W. Rozycki]

On Wed, 23 Jun 2004, David Daney wrote:

> Just out of curiosity, do you propose this patch in lieu of the patch to 
> Linux's traps.c?
> 
> Or would you do both?

 Both.

> It seems like both would be best, as there are already "broken" binutils 
> floating around out there.

 And broken binaries.

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

Re: [Patch] / 0 should send SIGFPE not SIGTRAP...

[Richard Sandiford]

"Maciej W. Rozycki" <macro@ds2.pg.gda.pl> writes:
>  Or should we get rid of the 20-bit "break" completely?  The two-argument
> version provides the same functionality, although the 10-bit codes to be
> used do not map to the 20-bit equivalent "optically" very well.  
> Especially if decimal notation is used.

I notice no-one's really responded to this question yet.  FWIW, on gut
instinct, I'd personally prefer to drop the 20-bit break than introduce
a new, non-standard name for it.

Just an opinion though.  I won't argue against anyone saying different. ;)

Richard

Re: [Patch] / 0 should send SIGFPE not SIGTRAP...

[Maciej W. Rozycki]

On Thu, 24 Jun 2004, Richard Sandiford wrote:

> >  Or should we get rid of the 20-bit "break" completely?  The two-argument
> > version provides the same functionality, although the 10-bit codes to be
> > used do not map to the 20-bit equivalent "optically" very well.  
> > Especially if decimal notation is used.
> 
> I notice no-one's really responded to this question yet.  FWIW, on gut
> instinct, I'd personally prefer to drop the 20-bit break than introduce
> a new, non-standard name for it.

 Well, this is essentially what the patch does.  Or do you mean: "drop it
and if anyone screams, consider an alternative?"  I'd find it acceptable,
actually, but it's not my opinion that really matters here.

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

Re: [Patch] / 0 should send SIGFPE not SIGTRAP...

[cgd]

At Thu, 24 Jun 2004 18:35:21 +0000 (UTC), "Maciej W. Rozycki" wrote:
>  Well, this is essentially what the patch does.  Or do you mean: "drop it
> and if anyone screams, consider an alternative?"  I'd find it acceptable,
> actually, but it's not my opinion that really matters here.

(it's fine w/ me.)

Re: [Patch] / 0 should send SIGFPE not SIGTRAP...

[Maciej W. Rozycki]

On Thu, 24 Jun 2004 cgd@broadcom.com wrote:

> >  Well, this is essentially what the patch does.  Or do you mean: "drop it
> > and if anyone screams, consider an alternative?"  I'd find it acceptable,
> > actually, but it's not my opinion that really matters here.
> 
> (it's fine w/ me.)

 Here's an updated patch I use currently in case this is the solution to
be agreed upon.

opcodes/:
2004-06-28  Maciej W. Rozycki  <macro@linux-mips.org>

	* mips-opc.c (mips_builtin_opcodes): Remove the MIPS32 
	ISA-specific "break" encoding.

gas/testsuite/:
2004-06-28  Maciej W. Rozycki  <macro@linux-mips.org>

	* gas/mips/mips32.s: Adjust for the unified "break" syntax.
	* gas/mips/set-arch.s: Likewise.
	* gas/mips/mips32.d: Adjust for the new output.
	* gas/mips/set-arch.d: Likewise.

  Maciej

binutils-2.15.91-20040625-mips-break20.patch
diff -up --recursive --new-file binutils-2.15.91-20040625.macro/gas/testsuite/gas/mips/mips32.d binutils-2.15.91-20040625/gas/testsuite/gas/mips/mips32.d
--- binutils-2.15.91-20040625.macro/gas/testsuite/gas/mips/mips32.d	2003-05-08 03:25:35.000000000 +0000
+++ binutils-2.15.91-20040625/gas/testsuite/gas/mips/mips32.d	2004-06-17 21:15:04.000000000 +0000
@@ -48,7 +48,7 @@ Disassembly of section .text:
 0+0098 <[^>]*> 4359e260 	wait	0x56789
 0+009c <[^>]*> 0000000d 	break
 0+00a0 <[^>]*> 0000000d 	break
-0+00a4 <[^>]*> 0048d14d 	break	0x12345
+0+00a4 <[^>]*> 0048d14d 	break	0x48,0x345
 0+00a8 <[^>]*> 7000003f 	sdbbp
 0+00ac <[^>]*> 7000003f 	sdbbp
 0+00b0 <[^>]*> 7159e27f 	sdbbp	0x56789
diff -up --recursive --new-file binutils-2.15.91-20040625.macro/gas/testsuite/gas/mips/mips32.s binutils-2.15.91-20040625/gas/testsuite/gas/mips/mips32.s
--- binutils-2.15.91-20040625.macro/gas/testsuite/gas/mips/mips32.s	2002-09-05 03:25:40.000000000 +0000
+++ binutils-2.15.91-20040625/gas/testsuite/gas/mips/mips32.s	2004-06-26 12:41:17.000000000 +0000
@@ -58,11 +58,13 @@ text_label:
       wait    0                       # disassembles without code
       wait    0x56789
 
-      # Instructions in previous ISAs or CPUs which are now slightly
-      # different.
+      # Instructions that used to have compatibility problems.
       break
       break   0                       # disassembles without code
-      break   0x12345
+      break   0x48,0x345
+
+      # Instructions in previous ISAs or CPUs which are now slightly
+      # different.
       sdbbp
       sdbbp   0                       # disassembles without code
       sdbbp   0x56789
diff -up --recursive --new-file binutils-2.15.91-20040625.macro/gas/testsuite/gas/mips/set-arch.d binutils-2.15.91-20040625/gas/testsuite/gas/mips/set-arch.d
--- binutils-2.15.91-20040625.macro/gas/testsuite/gas/mips/set-arch.d	2003-10-01 03:25:36.000000000 +0000
+++ binutils-2.15.91-20040625/gas/testsuite/gas/mips/set-arch.d	2004-06-17 21:17:46.000000000 +0000
@@ -160,7 +160,7 @@ Disassembly of section \.text:
 00000260 <[^>]*> 4359e260 	wait	0x56789
 00000264 <[^>]*> 0000000d 	break
 00000268 <[^>]*> 0000000d 	break
-0000026c <[^>]*> 0048d14d 	break	0x12345
+0000026c <[^>]*> 0048d14d 	break	0x48,0x345
 00000270 <[^>]*> 7000003f 	sdbbp
 00000274 <[^>]*> 7000003f 	sdbbp
 00000278 <[^>]*> 7159e27f 	sdbbp	0x56789
diff -up --recursive --new-file binutils-2.15.91-20040625.macro/gas/testsuite/gas/mips/set-arch.s binutils-2.15.91-20040625/gas/testsuite/gas/mips/set-arch.s
--- binutils-2.15.91-20040625.macro/gas/testsuite/gas/mips/set-arch.s	2003-06-29 19:41:33.000000000 +0000
+++ binutils-2.15.91-20040625/gas/testsuite/gas/mips/set-arch.s	2004-06-26 12:42:22.000000000 +0000
@@ -200,11 +200,13 @@ text_label:	
 	wait    0                       # disassembles without code
 	wait    0x56789
 
-	# Instructions in previous ISAs or CPUs which are now slightly
-	# different.
+	# Instructions that used to have compatibility problems.
 	break
 	break   0                       # disassembles without code
-	break   0x12345
+	break   0x48,0x345
+
+	# Instructions in previous ISAs or CPUs which are now slightly
+	# different.
 	sdbbp
 	sdbbp   0                       # disassembles without code
 	sdbbp   0x56789
diff -up --recursive --new-file binutils-2.15.91-20040625.macro/opcodes/mips-opc.c binutils-2.15.91-20040625/opcodes/mips-opc.c
--- binutils-2.15.91-20040625.macro/opcodes/mips-opc.c	2003-11-19 04:25:23.000000000 +0000
+++ binutils-2.15.91-20040625/opcodes/mips-opc.c	2004-06-26 12:42:46.000000000 +0000
@@ -274,7 +274,6 @@ const struct mips_opcode mips_builtin_op
 {"bnel",    "s,t,p",	0x54000000, 0xfc000000,	CBL|RD_s|RD_t, 		I2|T3	},
 {"bnel",    "s,I,p",	0,    (int) M_BNEL_I,	INSN_MACRO,		I2|T3	},
 {"break",   "",		0x0000000d, 0xffffffff,	TRAP,			I1	},
-{"break",   "B",        0x0000000d, 0xfc00003f, TRAP,           	I32     },
 {"break",   "c",	0x0000000d, 0xfc00ffff,	TRAP,			I1	},
 {"break",   "c,q",	0x0000000d, 0xfc00003f,	TRAP,			I1	},
 {"c.f.d",   "S,T",	0x46200030, 0xffe007ff,	RD_S|RD_T|WR_CC|FP_D,	I1	},