Re: [ppc] Include hwcap as ifunc argument

Re: [ppc] Include hwcap as ifunc argument

[Richard Henderson]

[-- Attachment #1: Type: text/plain, Size: 780 bytes --]

On 05/21/2012 09:58 AM, Richard Henderson wrote:
> Aside from the x86 port, there are only two targets that have ifunc
> support in glibc mainline: ppc and arm.  It would be awesomely handy
> if we switch these targets to mirror the sparc lead in passing along
> the dl_hwcap contents as an argument to the resolver, *and* to do it
> in the same release as adding getauxval.  There are some clever things
> we can do at static link time if we know that getauxval corresponds to
> the availability of the ifunc argument.
>
> Dave, dunno if you wanna change the signature of the arg from int to
> unsigned long for 64-bit?  It just seemed prudent to do that here...
>
> I'll present a similar patch for arm shortly...

Like so.  Tested on armv7.  And a ping for the ppc patch?


r~

[-- Attachment #2: d-hwcap-6 --]
[-- Type: text/plain, Size: 404 bytes --]

diff --git a/sysdeps/arm/dl-irel.h b/sysdeps/arm/dl-irel.h
index 5141031..5b1964e 100644
--- a/sysdeps/arm/dl-irel.h
+++ b/sysdeps/arm/dl-irel.h
@@ -29,7 +29,7 @@ static inline Elf32_Addr
 __attribute ((always_inline))
 elf_ifunc_invoke (Elf32_Addr addr)
 {
-  return ((Elf32_Addr (*) (void)) (addr)) ();
+  return ((Elf32_Addr (*) (unsigned long int)) (addr)) (GLRO(dl_hwcap));
 }
 
 static inline void

Re: [ppc] Include hwcap as ifunc argument

[Joseph S. Myers]

On Thu, 24 May 2012, Richard Henderson wrote:

> On 05/21/2012 09:58 AM, Richard Henderson wrote:
> > Aside from the x86 port, there are only two targets that have ifunc
> > support in glibc mainline: ppc and arm.  It would be awesomely handy
> > if we switch these targets to mirror the sparc lead in passing along
> > the dl_hwcap contents as an argument to the resolver, *and* to do it
> > in the same release as adding getauxval.  There are some clever things
> > we can do at static link time if we know that getauxval corresponds to
> > the availability of the ifunc argument.
> > 
> > Dave, dunno if you wanna change the signature of the arg from int to
> > unsigned long for 64-bit?  It just seemed prudent to do that here...
> > 
> > I'll present a similar patch for arm shortly...
> 
> Like so.  Tested on armv7.  And a ping for the ppc patch?

This ARM patch is OK with the usual copyright dates update.

-- 
Joseph S. Myers
joseph@codesourcery.com

Re: [ppc] Include hwcap as ifunc argument

[David Miller]

From: Richard Henderson <rth@twiddle.net>
Date: Thu, 24 May 2012 07:40:03 -0700

> On 05/21/2012 09:58 AM, Richard Henderson wrote:
>> Aside from the x86 port, there are only two targets that have ifunc
>> support in glibc mainline: ppc and arm.  It would be awesomely handy
>> if we switch these targets to mirror the sparc lead in passing along
>> the dl_hwcap contents as an argument to the resolver, *and* to do it
>> in the same release as adding getauxval.  There are some clever things
>> we can do at static link time if we know that getauxval corresponds to
>> the availability of the ifunc argument.
>>
>> Dave, dunno if you wanna change the signature of the arg from int to
>> unsigned long for 64-bit?  It just seemed prudent to do that here...
>>
>> I'll present a similar patch for arm shortly...
> 
> Like so.  Tested on armv7.  And a ping for the ppc patch?

You also have to adjust the IFUNC invocations in dl-machine.h, and
potentially elsewhere in the port.

Include hwcap as ifunc argument

[Richard Henderson]

[-- Attachment #1: Type: text/plain, Size: 64 bytes --]

Committed as follows, after re-testing on arm, ppc[32/64].


r~

[-- Attachment #2: z1 --]
[-- Type: text/plain, Size: 1686 bytes --]

        * sysdeps/arm/dl-irel.h (elf_ifunc_invoke): Pass dl_hwcap.
        * sysdeps/arm/dl-machine.h (elf_machine_rel): Use elf_ifunc_invoke.
        (elf_machine_rela): Likewise.



diff --git a/sysdeps/arm/dl-irel.h b/sysdeps/arm/dl-irel.h
index 5141031..5b1964e 100644
--- a/sysdeps/arm/dl-irel.h
+++ b/sysdeps/arm/dl-irel.h
@@ -29,7 +29,7 @@ static inline Elf32_Addr
 __attribute ((always_inline))
 elf_ifunc_invoke (Elf32_Addr addr)
 {
-  return ((Elf32_Addr (*) (void)) (addr)) ();
+  return ((Elf32_Addr (*) (unsigned long int)) (addr)) (GLRO(dl_hwcap));
 }
 
 static inline void
diff --git a/sysdeps/arm/dl-machine.h b/sysdeps/arm/dl-machine.h
index 58cf418..8d905e8 100644
--- a/sysdeps/arm/dl-machine.h
+++ b/sysdeps/arm/dl-machine.h
@@ -25,6 +25,7 @@
 #include <sys/param.h>
 #include <tls.h>
 #include <dl-tlsdesc.h>
+#include <dl-irel.h>
 
 #define CLEAR_CACHE(BEG,END)						\
   INTERNAL_SYSCALL_ARM (cacheflush, , 3, (BEG), (END), 0)
@@ -377,7 +378,7 @@ elf_machine_rel (struct link_map *map, const Elf32_Rel *reloc,
 	  && __builtin_expect (ELFW(ST_TYPE) (sym->st_info) == STT_GNU_IFUNC, 0)
 	  && __builtin_expect (sym->st_shndx != SHN_UNDEF, 1)
 	  && __builtin_expect (!skip_ifunc, 1))
-	value = ((Elf32_Addr (*) (void)) value) ();
+	value = elf_ifunc_invoke (value);
 
       switch (r_type)
 	{
@@ -551,7 +552,7 @@ elf_machine_rela (struct link_map *map, const Elf32_Rela *reloc,
 	  && __builtin_expect (ELFW(ST_TYPE) (sym->st_info) == STT_GNU_IFUNC, 0)
 	  && __builtin_expect (sym->st_shndx != SHN_UNDEF, 1)
 	  && __builtin_expect (!skip_ifunc, 1))
-	value = ((Elf32_Addr (*) (void)) value) ();
+	value = elf_ifunc_invoke (value);
 
       switch (r_type)
 	{

[-- Attachment #3: z2 --]
[-- Type: text/plain, Size: 3723 bytes --]

        * sysdeps/powerpc/powerpc32/dl-irel.h (elf_ifunc_invoke): Pass
        dl_hwcap to ifunc resolver.
        * sysdeps/powerpc/powerpc32/dl-machine.h (elf_machine_rela): Use
        elf_ifunc_invoke.
        * sysdeps/powerpc/powerpc64/dl-irel.h (elf_ifunc_invoke): Pass
        dl_hwcap to ifunc resolver.
        * sysdeps/powerpc/powerpc64/dl-machine.h (resolve_ifunc): Likewise.



diff --git a/sysdeps/powerpc/powerpc32/dl-irel.h b/sysdeps/powerpc/powerpc32/dl-irel.h
index c8e50ee..ebaf44a 100644
--- a/sysdeps/powerpc/powerpc32/dl-irel.h
+++ b/sysdeps/powerpc/powerpc32/dl-irel.h
@@ -1,6 +1,6 @@
 /* Machine-dependent ELF indirect relocation inline functions.
    PowerPC version.
-   Copyright (C) 2009 Free Software Foundation, Inc.
+   Copyright (C) 2009-2012 Free Software Foundation, Inc.
    This file is part of the GNU C Library.
 
    The GNU C Library is free software; you can redistribute it and/or
@@ -29,7 +29,7 @@ static inline Elf32_Addr
 __attribute ((always_inline))
 elf_ifunc_invoke (Elf32_Addr addr)
 {
-  return ((Elf32_Addr (*) (void)) (addr)) ();
+  return ((Elf32_Addr (*) (unsigned long int)) (addr)) (GLRO(dl_hwcap));
 }
 
 static inline void
diff --git a/sysdeps/powerpc/powerpc32/dl-machine.h b/sysdeps/powerpc/powerpc32/dl-machine.h
index 45868f5..3ae27a4 100644
--- a/sysdeps/powerpc/powerpc32/dl-machine.h
+++ b/sysdeps/powerpc/powerpc32/dl-machine.h
@@ -1,5 +1,5 @@
 /* Machine-dependent ELF dynamic relocation inline functions.  PowerPC version.
-   Copyright (C) 1995-2003, 2005, 2006, 2011 Free Software Foundation, Inc.
+   Copyright (C) 1995-2012 Free Software Foundation, Inc.
    This file is part of the GNU C Library.
 
    The GNU C Library is free software; you can redistribute it and/or
@@ -23,6 +23,7 @@
 
 #include <assert.h>
 #include <dl-tls.h>
+#include <dl-irel.h>
 
 /* Translate a processor specific dynamic tag to the index
    in l_info array.  */
@@ -308,7 +309,7 @@ elf_machine_rela (struct link_map *map, const Elf32_Rela *reloc,
       && __builtin_expect (ELFW(ST_TYPE) (sym->st_info) == STT_GNU_IFUNC, 0)
       && __builtin_expect (sym->st_shndx != SHN_UNDEF, 1)
       && __builtin_expect (!skip_ifunc, 1))
-    value = ((Elf32_Addr (*) (void)) value) ();
+    value = elf_ifunc_invoke (value);
 
   /* A small amount of code is duplicated here for speed.  In libc,
      more than 90% of the relocs are R_PPC_RELATIVE; in the X11 shared
diff --git a/sysdeps/powerpc/powerpc64/dl-irel.h b/sysdeps/powerpc/powerpc64/dl-irel.h
index 7270275..ced0f3f 100644
--- a/sysdeps/powerpc/powerpc64/dl-irel.h
+++ b/sysdeps/powerpc/powerpc64/dl-irel.h
@@ -1,6 +1,6 @@
 /* Machine-dependent ELF indirect relocation inline functions.
    PowerPC64 version.
-   Copyright (C) 2009, 2011 Free Software Foundation, Inc.
+   Copyright (C) 2009-2012 Free Software Foundation, Inc.
    This file is part of the GNU C Library.
 
    The GNU C Library is free software; you can redistribute it and/or
@@ -31,7 +31,7 @@ static inline Elf64_Addr
 __attribute ((always_inline))
 elf_ifunc_invoke (Elf64_Addr addr)
 {
-  return ((Elf64_Addr (*) (void)) (addr)) ();
+  return ((Elf64_Addr (*) (unsigned long int)) (addr)) (GLRO(dl_hwcap));
 }
 
 static inline void
diff --git a/sysdeps/powerpc/powerpc64/dl-machine.h b/sysdeps/powerpc/powerpc64/dl-machine.h
index a964a29..7c7d768 100644
--- a/sysdeps/powerpc/powerpc64/dl-machine.h
+++ b/sysdeps/powerpc/powerpc64/dl-machine.h
@@ -545,7 +545,7 @@ resolve_ifunc (Elf64_Addr value,
       value = (Elf64_Addr) &opd;
     }
 #endif
-  return ((Elf64_Addr (*) (void)) value) ();
+  return ((Elf64_Addr (*) (unsigned long int)) value) (GLRO(dl_hwcap));
 }
 
 /* Perform the relocation specified by RELOC and SYM (which is fully

Re: Include hwcap as ifunc argument

[Ryan S. Arnold]

On Fri, May 25, 2012 at 12:44 PM, Richard Henderson <rth@twiddle.net> wrote:
> Committed as follows, after re-testing on arm, ppc[32/64].

Hi Richard,

Peter Bergner reminded me today that on PowerPC the IFUNC resolver
will most likely/often use the value in AT_PLATFORM rather than
AT_HWCAP to determine which optimized function to use.  I suppose this
would diminish the necessity of passing the hwcap to the ifunc
resolver.

Ryan

Re: Include hwcap as ifunc argument

[Richard Henderson]

On 06/04/2012 02:21 PM, Ryan S. Arnold wrote:
> Peter Bergner reminded me today that on PowerPC the IFUNC resolver
> will most likely/often use the value in AT_PLATFORM rather than
> AT_HWCAP to determine which optimized function to use.  I suppose this
> would diminish the necessity of passing the hwcap to the ifunc
> resolver.

It depends on what's being tested, I suppose.

Within glibc itself you've got lots of platform-specific tuning, so 
you may well use AT_PLATFORM more.

Outside glibc, it might just be a test for Altivec or VDX or whatever.


r~

Re: Include hwcap as ifunc argument

[Andreas Krebbel]

On 06/05/2012 12:14 AM, Richard Henderson wrote:
> On 06/04/2012 02:21 PM, Ryan S. Arnold wrote:
>> Peter Bergner reminded me today that on PowerPC the IFUNC resolver
>> will most likely/often use the value in AT_PLATFORM rather than
>> AT_HWCAP to determine which optimized function to use.  I suppose this
>> would diminish the necessity of passing the hwcap to the ifunc
>> resolver.
> 
> It depends on what's being tested, I suppose.
> 
> Within glibc itself you've got lots of platform-specific tuning, so 
> you may well use AT_PLATFORM more.
> 
> Outside glibc, it might just be a test for Altivec or VDX or whatever.

From a S/390 perspective AT_PLATFORM would also be more helpful. We do not add every new
machine feature to the hwcaps vector so it is not possible to deduce the cpu level just
from the hwcaps vector.

I'm working on the S/390 ifunc support and hope to be able to come up with a patch soon.

Bye,

-Andreas-

Re: Include hwcap as ifunc argument

[Richard Henderson]

On 2012-06-08 07:00, Andreas Krebbel wrote:
>From a S/390 perspective AT_PLATFORM would also be more helpful. We do not add every new
> machine feature to the hwcaps vector so it is not possible to deduce the cpu level just
> from the hwcaps vector.

No, but the one important thing in s390's AT_HWCAP is HWCAP_S390_STFLE,
which tells you (without setting up sigill handlers) that you can go ahead
and find out *everything* you want to know via stfle.  More or less exactly
how the x86 port works atm with its cpuid bits.

But again, it depends on what you're trying to test.  Are you looking at
choosing a version based on scheduling, or choosing a version based on
features?

I'd would hope that any multiarch'd version of e.g. feraiseexcept would test
for the ieee-exception-simulation-facility (included in bit 41 of stfle results)
rather than using a strcmp of AT_PLATFORM vs "z109-xx" (or whatever).

Certainly for libatomic I would be most interested in testing for the 
presence of the interlocked-access and compare-and-swap-and-store-2 facilities.
 
> I'm working on the S/390 ifunc support and hope to be able to come up with a patch soon.

Excellent.


r~