[Bug inline-asm/37562] New: [4.2] -funroll-loops destroys inline asm code von powerpc

[Bug inline-asm/37562] New: [4.2] -funroll-loops destroys inline asm code von powerpc

[rbuergel at web dot de]

typedef unsigned int UInt32;
typedef unsigned char UInt8;

struct Data
{
  UInt8 data[16];

  const UInt8* getData() const
  {
    return data + 4;
  }
};

struct Value {
  UInt32 value;
  static void update (Value& signal, const Data& source, UInt32 startBit);
};

void Value::update(Value& signal, const Data& source, UInt32 startByte)
{
  UInt32& value = signal.value;
  const UInt8* data = source.getData();
  asm( "lwbrx %0,%1,%2;\n":"=r"(value): "r"(data), "r"(startByte), "m" (*(data
+ startByte)) );
}

generated for gcc-4.2.4 on powerpc:
00000000 <_ZN5Value6updateERS_RK4Dataj>:
   0:   38 04 00 04     addi    r0,r4,4
   4:   7c 80 2c 2c     lwbrx   r4,0,r5
   8:   90 83 00 00     stw     r4,0(r3)
   c:   4e 80 00 20     blr

the correct version without -funroll-loops reads:
00000000 <_ZN5Value6updateERS_RK4Dataj>:
   0:   38 84 00 04     addi    r4,r4,4
   4:   7c 84 2c 2c     lwbrx   r4,r4,r5
   8:   90 83 00 00     stw     r4,0(r3)
   c:   4e 80 00 20     blr


lwbrx sums the contents of r4 and r5 and accesses this address, swaps it's
value and stores it again in r4. The incorrect version accesses an invalid
address, which leads to a segmentation fault.

this is corrected at least with gcc-4.3.1, there i couldn't reproduce the
behaviour.


-- 
           Summary: [4.2] -funroll-loops destroys inline asm code von
                    powerpc
           Product: gcc
           Version: 4.2.4
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: inline-asm
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: rbuergel at web dot de


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37562

[Bug inline-asm/37562] [4.2] -funroll-loops destroys inline asm code von powerpc

[rbuergel at web dot de]

------- Comment #1 from rbuergel at web dot de  2008-09-17 15:44 -------
oops, i forgot to mention my command line to compile it:
powerpc-linux-g++ -O3 -funroll-loops -c -o GetBytes2.o GetBytes2.cpp

using -O3 is important for reproducing this error


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37562

Re: [Bug inline-asm/37562] New: [4.2] -funroll-loops destroys inline asm code von powerpc

[Andrew Thomas Pinski]

Sent from my iPhone

On Sep 17, 2008, at 8:42 AM, "rbuergel at web dot de" <gcc-bugzilla@gcc.gnu.org 
 > wrote:

> typedef unsigned int UInt32;
> typedef unsigned char UInt8;
>
> struct Data
> {
>  UInt8 data[16];
>
>  const UInt8* getData() const
>  {
>    return data + 4;
>  }
> };
>
> struct Value {
>  UInt32 value;
>  static void update (Value& signal, const Data& source, UInt32  
> startBit);
> };
>
> void Value::update(Value& signal, const Data& source, UInt32  
> startByte)
> {
>  UInt32& value = signal.value;
>  const UInt8* data = source.getData();
>  asm( "lwbrx %0,%1,%2;\n":"=r"(value): "r"(data), "r"(startByte),  
> "m" (*(data
> + startByte)) );
> }

The second constraint should be using "b" instead of "r" as b says  
don't use r0.


>
>
> generated for gcc-4.2.4 on powerpc:
> 00000000 <_ZN5Value6updateERS_RK4Dataj>:
>   0:   38 04 00 04     addi    r0,r4,4
>   4:   7c 80 2c 2c     lwbrx   r4,0,r5
>   8:   90 83 00 00     stw     r4,0(r3)
>   c:   4e 80 00 20     blr
>
> the correct version without -funroll-loops reads:
> 00000000 <_ZN5Value6updateERS_RK4Dataj>:
>   0:   38 84 00 04     addi    r4,r4,4
>   4:   7c 84 2c 2c     lwbrx   r4,r4,r5
>   8:   90 83 00 00     stw     r4,0(r3)
>   c:   4e 80 00 20     blr
>
>
> lwbrx sums the contents of r4 and r5 and accesses this address,  
> swaps it's
> value and stores it again in r4. The incorrect version accesses an  
> invalid
> address, which leads to a segmentation fault.
>
> this is corrected at least with gcc-4.3.1, there i couldn't  
> reproduce the
> behaviour.
>
>
> -- 
>           Summary: [4.2] -funroll-loops destroys inline asm code von
>                    powerpc
>           Product: gcc
>           Version: 4.2.4
>            Status: UNCONFIRMED
>          Severity: normal
>          Priority: P3
>         Component: inline-asm
>        AssignedTo: unassigned at gcc dot gnu dot org
>        ReportedBy: rbuergel at web dot de
>
>
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37562
>

[Bug inline-asm/37562] [4.2] -funroll-loops destroys inline asm code for powerpc

[pinskia at gmail dot com]

------- Comment #2 from pinskia at gmail dot com  2008-09-17 16:52 -------
Subject: Re:   New: [4.2] -funroll-loops destroys inline asm code von powerpc



Sent from my iPhone

On Sep 17, 2008, at 8:42 AM, "rbuergel at web dot de" <gcc-bugzilla@gcc.gnu.org 
 > wrote:

> typedef unsigned int UInt32;
> typedef unsigned char UInt8;
>
> struct Data
> {
>  UInt8 data[16];
>
>  const UInt8* getData() const
>  {
>    return data + 4;
>  }
> };
>
> struct Value {
>  UInt32 value;
>  static void update (Value& signal, const Data& source, UInt32  
> startBit);
> };
>
> void Value::update(Value& signal, const Data& source, UInt32  
> startByte)
> {
>  UInt32& value = signal.value;
>  const UInt8* data = source.getData();
>  asm( "lwbrx %0,%1,%2;\n":"=r"(value): "r"(data), "r"(startByte),  
> "m" (*(data
> + startByte)) );
> }

The second constraint should be using "b" instead of "r" as b says  
don't use r0.


>
>
> generated for gcc-4.2.4 on powerpc:
> 00000000 <_ZN5Value6updateERS_RK4Dataj>:
>   0:   38 04 00 04     addi    r0,r4,4
>   4:   7c 80 2c 2c     lwbrx   r4,0,r5
>   8:   90 83 00 00     stw     r4,0(r3)
>   c:   4e 80 00 20     blr
>
> the correct version without -funroll-loops reads:
> 00000000 <_ZN5Value6updateERS_RK4Dataj>:
>   0:   38 84 00 04     addi    r4,r4,4
>   4:   7c 84 2c 2c     lwbrx   r4,r4,r5
>   8:   90 83 00 00     stw     r4,0(r3)
>   c:   4e 80 00 20     blr
>
>
> lwbrx sums the contents of r4 and r5 and accesses this address,  
> swaps it's
> value and stores it again in r4. The incorrect version accesses an  
> invalid
> address, which leads to a segmentation fault.
>
> this is corrected at least with gcc-4.3.1, there i couldn't  
> reproduce the
> behaviour.
>
>
> -- 
>           Summary: [4.2] -funroll-loops destroys inline asm code von
>                    powerpc
>           Product: gcc
>           Version: 4.2.4
>            Status: UNCONFIRMED
>          Severity: normal
>          Priority: P3
>         Component: inline-asm
>        AssignedTo: unassigned at gcc dot gnu dot org
>        ReportedBy: rbuergel at web dot de
>
>
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37562
>


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37562

[Bug inline-asm/37562] [4.2] -funroll-loops destroys inline asm code for powerpc

[rbuergel at web dot de]

------- Comment #3 from rbuergel at web dot de  2008-09-17 18:46 -------
> The second constraint should be using "b" instead of "r" as b says
> don't use r0.
Is this documented anywhere? The gcc manual says "r" means "any general purpose
register" and "b" means "Address base register". Any Documentation (for example
http://developer.apple.com/documentation/DeveloperTools/Reference/Assembler/PPCInstructions/chapter_6_section_2.html
) i've found, lists r0-r31 as general purpose registers.

Anyway, ignoring this point, you seem to be right. I remember 4.3.1 using r6
instead of r0, probably that's why it works. Unfortunately, i can't check your
suggestion until monday, when i'm back at work.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37562

[Bug inline-asm/37562] [4.2] -funroll-loops destroys inline asm code for powerpc

[pinskia at gcc dot gnu dot org]

------- Comment #4 from pinskia at gcc dot gnu dot org  2008-09-17 18:50 -------
From:
http://gcc.gnu.org/onlinedocs/gcc-4.3.2/gcc/Machine-Constraints.html#Machine-Constraints

b
Address base register

Address base register is the standard saying rN or 0 (if r0 is used).  This is
just the way PPC asm works.


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |INVALID


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37562

[Bug inline-asm/37562] [4.2] -funroll-loops destroys inline asm code for powerpc

[rbuergel at web dot de]

------- Comment #5 from rbuergel at web dot de  2008-09-17 18:57 -------
Too bad for newbies to ppc asm. :)
But thank you for your explanations


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37562

[Bug inline-asm/37562] [4.2] -funroll-loops destroys inline asm code for powerpc

[pinskia at gcc dot gnu dot org]

------- Comment #6 from pinskia at gcc dot gnu dot org  2008-09-17 19:06 -------
You can also use __builtin_bswap32 in GCC 4.3 and above which gives you the
best code generation as it understands loading from memory and if the value is
in a register already.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37562