[Bug libc/6816] New: Support for POWER7 VSX registers in *context routines

[Bug libc/6816] New: Support for POWER7 VSX registers in *context routines

[cseo at linux dot vnet dot ibm dot com]

The POWER7 processor will introduce a new register set called VSX (Vector Scalar
Extension), which consists of 64 128-bit wide registers. The new register layout
is as follows:

             VSR doubleword 0               VSR doubleword 1
----------------------------------------------------------------
VSR[0]  |             FPR[0]            |                       |   
----------------------------------------------------------------
VSR[1]  |             FPR[1]            |                       |   
----------------------------------------------------------------
        |              ...              |                       |
        |              ...              |                       |    
----------------------------------------------------------------
VSR[30] |             FPR[30]           |                       |
----------------------------------------------------------------
VSR[31] |             FPR[31]           |                       |   
----------------------------------------------------------------
VSR[32] |             FPR[32]         VR[0]                     |
----------------------------------------------------------------
VSR[33] |             FPR[33]         VR[1]                     |   
----------------------------------------------------------------
        |                              ...                      |
        |                              ...                      |
----------------------------------------------------------------
VSR[62] |             FPR[62]         VR[30]                    |
----------------------------------------------------------------
VSR[63] |             FPR[63]         VR[31]                    |
----------------------------------------------------------------

The first doubleword of registers VSR[0] to VSR[31] overlap with the current
existing floating point registers FPR[0] to FPR[31], while the registers VSR[32]
to VSR[63] overlap with the current existing vector registers (Altivec/VMX)
VR[0] to VR[31].

The VSX ISA introduces a new set of scalar and SIMD instructions that operate in
all VSR registers. Legacy BFP and DFP instructions operate only on
VSR[0]-VSR[31] doubleword 0, and legacy VMX instructions operate only on
VSR[31]-VSR[63].

Initial support for VSX registers and instructions in binutils was added
recently, providing the instructions required for glibc to save the extended
signal frames, as per discussion in:

http://sourceware.org/ml/binutils/2008-08/msg00016.html

and gcc support is being discussed in
http://gcc.gnu.org/ml/gcc-patches/2008-08/msg00162.html .

In order to maintain backward compatibility with older applications, we  propose
extending the ucontext, in a way that the doubleword 1 of VSR[0]-VSR[31] is
appended to the end of the ucontext, after VR[31] (or VSR[63]). In other words,
we propose extending the vmx_reserve array within struct sigcontext, in a way
that the doubleword 1 elements from VSRs 0-31 follow the VRSAVE quadword.  Note
that this means doubleword 0 and doubleword 1 of VSR[0]-VSR[31] are not stored
in adjacent memory locations. So, in our proposed approach, we have to use
slightly more complicated code than simple loads and stores to save and restore
the register state from the ucontext.  However, it is still possible to
accomplish saving and restoring the entire VSX register state using only 64
stores/loads, as it will be shown shortly in a suggested implementation for this.

-- 
           Summary: Support for POWER7 VSX registers in *context routines
           Product: glibc
           Version: unspecified
            Status: NEW
          Severity: enhancement
          Priority: P2
         Component: libc
        AssignedTo: drepper at redhat dot com
        ReportedBy: cseo at linux dot vnet dot ibm dot com
                CC: glibc-bugs at sources dot redhat dot com,rsa at us dot
                    ibm dot com,sjmunroe at us dot ibm dot com


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[cseo at linux dot vnet dot ibm dot com]

------- Additional Comments From cseo at linux dot vnet dot ibm dot com  2008-08-04 18:16 -------
Created an attachment (id=2887)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=2887&action=view)
VSX implementation for *context routines (powerpc32)


-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[cseo at linux dot vnet dot ibm dot com]

------- Additional Comments From cseo at linux dot vnet dot ibm dot com  2008-08-04 18:15 -------
Created an attachment (id=2886)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=2886&action=view)
configure check for .machine "power7"


-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[cseo at linux dot vnet dot ibm dot com]

------- Additional Comments From cseo at linux dot vnet dot ibm dot com  2008-08-04 18:17 -------
Created an attachment (id=2888)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=2888&action=view)
VSX implementation for *context routines (powerpc64)


-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[cseo at linux dot vnet dot ibm dot com]

------- Additional Comments From cseo at linux dot vnet dot ibm dot com  2008-08-04 18:17 -------
Created an attachment (id=2889)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=2889&action=view)
VSX extension to ucontext


-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[cseo at linux dot vnet dot ibm dot com]

------- Additional Comments From cseo at linux dot vnet dot ibm dot com  2008-08-04 18:18 -------
Created an attachment (id=2890)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=2890&action=view)
Added versioned symbols for GLIBC-_2.9 (getcontext,setcontext, swapcontext)


-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[rsa at us dot ibm dot com]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
   GCC host triplet|                            |powerpc64-linux
 GCC target triplet|                            |powerpc64-linux
   Target Milestone|---                         |2.9


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[sjmunroe at us dot ibm dot com]

------- Additional Comments From sjmunroe at us dot ibm dot com  2008-08-05 17:53 -------
Might as well fix this comment typo while we are at it:

diff -urN libc28-cvs-20080709/sysdeps/unix/sysv/linux/powerpc/sys/ucontext.h
libc28/sysdeps/unix/sysv/linux/powerpc/sys/ucontext.h
--- libc28-cvs-20080709/sysdeps/unix/sysv/linux/powerpc/sys/ucontext.h
2005-03-15 16:14:24.000000000 -0600
+++ libc28/sysdeps/unix/sysv/linux/powerpc/sys/ucontext.h	2008-08-05
10:13:09.000000000 -0500
@@ -65,8 +72,8 @@
 /* For 64-bit kernels with Altivec support, a machine context is exactly 
  * a sigcontext.  For older kernel (without Altivec) the sigcontext matches 
  * the mcontext upto but not including the v_regs field.  For kernels that 
- * don't AT_HWCAP or return AT_HWCAP without PPC_FEATURE_HAS_ALTIVEC the 
- * v_regs field may not exit and should not be referenced.  The v_regd field
+ * don't have AT_HWCAP or return AT_HWCAP without PPC_FEATURE_HAS_ALTIVEC the 
+ * v_regs field may not exist and should not be referenced.  The v_regd field
  * can be refernced safely only after verifying that PPC_FEATURE_HAS_ALTIVEC
  * is set in AT_HWCAP.  */


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|2.9                         |---


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[rsa at us dot ibm dot com]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |2.9


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[drepper at redhat dot com]

------- Additional Comments From drepper at redhat dot com  2008-08-25 22:53 -------
The structure size changes.  This cannot be handled using symbol versioning
alone.  We have to resort to the same technique used by stat.  I.e.:

- assign a non-default version to [gs]etcontext so that you cannot link against
  it anymore (see atexit for how to do this)

- create new __x[sg]etcontent interfaces which take the usual interfaces plus a
  version number

- add new [sg]etcontext interfaces into libc_nonshared.a which call the
  __x[sg]etcontent interfaces with the appropriate version argument

This still isn't solving the problem completely.  Is there some padding space in
the old structure?  If yes, use it for a version number.  If there is no room,
add such a field to the new structure somewhere.

The code then must be changed to check for this value so that references are not
passed incorrectly.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |WAITING


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[cseo at linux dot vnet dot ibm dot com]

------- Additional Comments From cseo at linux dot vnet dot ibm dot com  2008-09-12 21:58 -------
Created an attachment (id=2947)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=2947&action=view)
VSX implementation for *context routines (powerpc32)


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
Attachment #2887 is|0                           |1
           obsolete|                            |


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[cseo at linux dot vnet dot ibm dot com]

------- Additional Comments From cseo at linux dot vnet dot ibm dot com  2008-09-12 21:57 -------
Created an attachment (id=2946)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=2946&action=view)
Makefile changes required for the new routines


-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[cseo at linux dot vnet dot ibm dot com]

------- Additional Comments From cseo at linux dot vnet dot ibm dot com  2008-09-12 21:59 -------
Created an attachment (id=2948)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=2948&action=view)
VSX implementation for *context routines (powerpc64)


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
Attachment #2888 is|0                           |1
           obsolete|                            |


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[cseo at linux dot vnet dot ibm dot com]

------- Additional Comments From cseo at linux dot vnet dot ibm dot com  2008-09-12 22:00 -------
Created an attachment (id=2949)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=2949&action=view)
New interfaces for *context functions


-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[cseo at linux dot vnet dot ibm dot com]

------- Additional Comments From cseo at linux dot vnet dot ibm dot com  2008-09-12 22:01 -------
Created an attachment (id=2951)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=2951&action=view)
VSX extension to ucontext


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
Attachment #2889 is|0                           |1
           obsolete|                            |


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[cseo at linux dot vnet dot ibm dot com]

------- Additional Comments From cseo at linux dot vnet dot ibm dot com  2008-09-12 22:00 -------
Created an attachment (id=2950)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=2950&action=view)
Updated testcases


-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[cseo at linux dot vnet dot ibm dot com]

------- Additional Comments From cseo at linux dot vnet dot ibm dot com  2008-09-12 22:02 -------
Created an attachment (id=2952)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=2952&action=view)
Added new __x*context symbols to GLIBC_2.9


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
Attachment #2890 is|0                           |1
           obsolete|                            |


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[cseo at linux dot vnet dot ibm dot com]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |ASSIGNED


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[cseo at linux dot vnet dot ibm dot com]

------- Additional Comments From cseo at linux dot vnet dot ibm dot com  2008-09-12 22:14 -------
I've just attached the patches containing the implementation changes you
suggested. One consequence of that is that we had to allow the internal version
of ucontext.h to be overridden (please see power7-ucontext-20080912.patch).

Waiting for review.

-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[drepper at redhat dot com]

------- Additional Comments From drepper at redhat dot com  2008-10-15 22:22 -------
I've asked this before but I don't remember getting a reply.

On x86/x86-64 the additional registers are not preserved in function calls. 
This means the context functions don't have to handle the registers.  The kernel
still provides the register content to signal handlers but that's a different story.

Why has the decision been made that the registers have to be saved (assuming
they are, otherwise the whole extension business is mood)?

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |WAITING


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[rsa at us dot ibm dot com]

------- Additional Comments From rsa at us dot ibm dot com  2008-10-16 14:00 -------
We want to allow user space applications continued API access to the context via
the get/set/swap context interface to the new VSX aware swapcontext syscall.

The kernel will support the new context size none-the-less and, as you've
pointed out, a user may copy the context on a signal and then attempt to use the
setcontext API on that context.  We'd like to support that.  You are correct,
the getcontext common code should have the VSX register saves removed.  If the
VSX aware swapcontext syscall is available the get/set/swap/context code will
simply set up some scaffolding and invoke the syscall.

I have a patch ready to go with the getcontext VSX save in place and I'll simply
remove that portion of the patch and respin.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |ASSIGNED
   Last reconfirmed|0000-00-00 00:00:00         |2008-10-16 14:00:48
               date|                            |


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[drepper at redhat dot com]

------- Additional Comments From drepper at redhat dot com  2008-10-16 14:56 -------
(In reply to comment #17)
> We want to allow user space applications continued API access to the context via
> the get/set/swap context interface to the new VSX aware swapcontext syscall.

The context shouldn't contain the new registers, that's the point.  The *context
functions are function calls.  Hence registers which are not preserved don't
have to be included in the context data structures.  This eliminates the
versioning problem and it is exactly the way this is handled on x86.



> The kernel will support the new context size none-the-less and, as you've
> pointed out, a user may copy the context on a signal and then attempt to use the
> setcontext API on that context.

That's fine, they can do this.  But only the registers which are callee-saved
have any guarantee.  There is no reason IMO to add the new registers.

This will all _significantly_ slow down code as well.  All those big registers
have to be stored and reloaded.

-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[sjmunroe at us dot ibm dot com]

------- Additional Comments From sjmunroe at us dot ibm dot com  2008-10-16 16:53 -------
>> We want to allow user space applications continued API access to the context via
>> the get/set/swap context interface to the new VSX aware swapcontext syscall.

>The context shouldn't contain the new registers, that's the point.  The *context
>functions are function calls.  Hence registers which are not preserved don't
>have to be included in the context data structures.  This eliminates the
>versioning problem and it is exactly the way this is handled on x86.

That is true for getcontext but may not be true for set/swapcontext depending on
the source of the context (ie. signal frames).

So it seem like the setcontext and the kernels swapcontext syscall should handle
the larger context size. 

There is a separate question if the size_of(ucontext_t) includes the VSX state
or not. If as you suggest (this time) we can avoid the versioning for the user
state functions as long as the kernel swapcontext syscall recognizes that for
MSR:VSX =1 the context is a the bigger size. This would be the case for a signal
ucontext in an application using VSX.

It follows that for getcontext we only pass the current size and the kernel does
not set MSR:VSX even if VSX is use (because the new VSX state is volatile!).

-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[drepper at redhat dot com]

------- Additional Comments From drepper at redhat dot com  2008-10-16 17:05 -------
(In reply to comment #19)
> That is true for getcontext but may not be true for set/swapcontext depending on
> the source of the context (ie. signal frames).
> 
> So it seem like the setcontext and the kernels swapcontext syscall should handle
> the larger context size. 

That has never been supported.  If it happened to work on ppc then this is
rather an accident.

Changing the data structure is hell.  It cannot be handled using versioning or
any other method complete without explicitly adding version information into the
data structure.  This would have to be in a completely backward compatible way.
 I.e., the version value would have to be in the range of the original data
structure and there must not be any ambiguity (i.e., you have to find a byte or
four in the original structure which has a limited range and use one of the
value which _never_ can happen in the real world).

And even that will not solve the problem of getcontext() where the data
structure can be uninitialized.  There really is no complete solution.

Hence, making such unnecessary changes just for the purpose of some programs
using undefined behavior is complete madness.

-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[sjmunroe at us dot ibm dot com]

------- Additional Comments From sjmunroe at us dot ibm dot com  2008-10-16 19:36 -------
>I.e., the version value would have to be in the range of the original data
>structure and there must not be any ambiguity (i.e., you have to find a byte or
>four in the original structure which has a limited range and use one of the
>value which _never_ can happen in the real world).

As I explained before the size of the ucontext is implied in an unambiguous way
by the feature enable bits (FP,VMX,VSX) of the MSR saved with the GPRs in the
ucontext. The kernal swapcontext swapcontext has set these since kernel 2.6.0,
before that getcontext zeroed the MSR. I believe that is unambiguous enough.

So if you want to not change sizeof(ucontext), we can do that. Just anyone crazy
enough to try copying a ucontext from a signal frame has to know to check for
MSR:VSX and copy a little bit more. And then use the swapcontext syscall with
the larger length. The kernel should handle this!

I assume we still need the wrappers to handle the previous version changes
(adding VMX) and any future changes?

-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[drepper at redhat dot com]

------- Additional Comments From drepper at redhat dot com  2009-04-07 18:10 -------
Don't change the size of the structure.  It's causing problems and it's
completely unnecessary.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |WAITING


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[rsa at us dot ibm dot com]

------- Additional Comments From rsa at us dot ibm dot com  2009-04-07 18:31 -------
Uli,

Yeah, we decided to not change the size.  Do you still think we should provide
the versioned structure mechanism for the existing interface to insulate against
future changes (since the code is pretty much ready to go)?

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |NEW
   Last reconfirmed|2008-10-16 14:00:48         |2009-04-07 18:31:25
               date|                            |


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[drepper at redhat dot com]

------- Additional Comments From drepper at redhat dot com  2009-04-07 18:38 -------
(In reply to comment #23)
> Do you still think we should provide
> the versioned structure mechanism for the existing interface to insulate against
> future changes (since the code is pretty much ready to go)?

No, don't break the ABI especially since no future change should do this either.
 The ABI is fixed.  No additional register is to be saved or at least depended on.

-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[pasky at suse dot cz]

------- Additional Comments From pasky at suse dot cz  2010-06-01 03:38 -------
Hmm, so what is actually the outcome of this bug? Is anything at all going to
happen yet, or can we close it now?

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |WAITING


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[rsa at us dot ibm dot com]

------- Additional Comments From rsa at us dot ibm dot com  2010-06-01 15:01 -------
All of the VSX registers are marked volatile, i.e. caller save so there's no
need to preserve or restore the register state in the context routines.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |SUSPENDED


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[rsa at us dot ibm dot com]

------- Additional Comments From rsa at us dot ibm dot com  2010-06-01 15:39 -------
We'll just mark this as invalid.

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


http://sourceware.org/bugzilla/show_bug.cgi?id=6816

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug libc/6816] Support for POWER7 VSX registers in *context routines

[jakub at redhat dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=6816

Jakub Jelinek <jakub at redhat dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at redhat dot com

--- Comment #28 from Jakub Jelinek <jakub at redhat dot com> ---
If the VSX state is indeed call clobbered in the ABI, supposedly glibc should
on successful getcontext and swapcontext syscalls just clear the MSR_VSX bit
from gregs[PT_MSR], because it knows it doesn't have room for the extra halves
of VSX registers and thus feeding it directly to setcontext or another
swapcontext would always fail?  Doing it in the kernel looks less desirable to
me, perhaps some code might be interested in knowing there is VSX state to be
saved and retry swapcontext syscall with a larger context or something (for
something doing the syscall by hand rather than using glibc APIs).

-- 
You are receiving this mail because:
You are on the CC list for the bug.