Re: [PATCH v2] libstdc++: Add support for POWER9 DARN instruction to std::random_device

[Bill Schmidt <wschmidt@linux.ibm.com>]

For posterity:  This was discussed briefly on IRC, and Segher approved with some 
simplifications and a request to implement a fail/retry check.

Thanks,
Bill

On 11/3/21 10:02 AM, Jonathan Wakely wrote:
> On Wed, 3 Nov 2021 at 15:01, Jonathan Wakely wrote: Any feedback from POWER maintainers about this? I'd like to push it soon if there's nothing wrong with it. With the updated patch attached again this time ... ‍ ‍ ‍ ‍ ‍ ‍ ‍ ‍ ‍ ‍ ‍ ‍ ‍ ‍ ‍ ‍ ‍ ‍ ‍ ‍ ZjQcmQRYFpfptBannerStart
> This Message Is From an External Sender
> This message came from outside your organization.
> ZjQcmQRYFpfptBannerEnd
> On Wed, 3 Nov 2021 at 15:01, Jonathan Wakely wrote:
>
>     Any feedback from POWER maintainers about this? I'd like to push it soon if there's nothing wrong with it.
>
>
> With the updated patch attached again this time ...
>
>
>
>     On Wed, 20 Oct 2021 at 14:00, Jonathan Wakely via Libstdc++ <libstdc++@gcc.gnu.org <mailto:libstdc%2B%2B@gcc.gnu.org>> wrote:
>
>         On 20/10/21 10:12 +0100, Jonathan Wakely wrote:
>         >On 19/10/21 17:47 +0100, Jonathan Wakely wrote:
>         >>The ISA-3.0 instruction set includes DARN ("deliver a random number")
>         >>which can be used similar to the existing support for RDRAND and RDSEED.
>         >>
>         >>libstdc++-v3/ChangeLog:
>         >>
>         >>      * src/c++11/random.cc (USE_DARN): Define.
>         >>      (__ppc_darn): New function to use POWER9 DARN instruction.
>         >>      (Which): Add 'darn' enumerator.
>         >>      (which_source): Check for __ppc_darn.
>         >>      (random_device::_M_init): Support "darn" and "hw" tokens.
>         >>      (random_device::_M_getentropy): Add darn to switch.
>         >>      * testsuite/26_numerics/random/random_device/cons/token.cc:
>         >>      Check "darn" token.
>         >>      * testsuite/26_numerics/random/random_device/entropy.cc:
>         >>      Likewise.
>         >>
>         >>Tested powerpc64le-linux (power8 and power9) and x86_64-linux.
>         >>
>         >>The new "darn" (power-specific) and "hw" (x86 and power)
>         >>strings should be documented, but I'll do that if this gets committed.
>         >>
>         >>Most of this patch is just "more of the same", similar to the existing
>         >>code for RDRAND and RDSEED on x86, but the parts of the patch I'd like
>         >>more eyes on are:
>         >>
>         >>
>         >>+#elif defined __powerpc__ && defined __BUILTIN_CPU_SUPPORTS__
>         >>+# define USE_DARN 1
>         >>#endif
>         >
>         >This means DARN can only be used when __builtin_cpu_supports is
>         >available, which means glibc 2.23 ... is that acceptable? It means
>         >RHEL 7 wouldn't be able to use DARN, but RHEL 8 would.
>         >
>         >There certainly are POWER9 machines running RHEL 7 and similar
>         >vintages (the GCC compile farm has one) so if there's another way to
>         >check for ISA 3.0 then I could use that.
>         >
>         >If __POWER9_VECTOR__ is defined when building libstdc++, presumably
>         >that means the whole library can only be run on POWER9 hardware. So
>         >would that mean we don't need to check __builtin_cpu_supports("darn")
>         >when __POWER9_VECTOR__ is defined? Or is it possible to build with
>         >-mcpu=power8 -mpower9-vector and run it on h/w without the DARN
>         >instruction?
>         >
>         >Also, I forgot to add a configure check that the assembler supports
>         >darn, which is another prerequisite for using it here.
>         >
>         >>@@ -135,6 +137,15 @@ namespace std _GLIBCXX_VISIBILITY(default)
>         >>#endif
>         >>#endif
>         >>
>         >>+#ifdef USE_DARN
>         >>+    unsigned int
>         >>+    __attribute__((target("power9")))
>         >
>         >Oops, that should be "cpu=power9".
>         >
>         >With that change it works on a POWER9 machine (9009-42A) with glibc
>         >2.34 and binutils 2.35.
>         >
>
>         Here's the updated patch with a configure check for assembler support,
>         and the target attribute fixed.
>
>         This still requires Glibc 2.23 for __builtin_cpu_supports, which I'm
>         assuming is acceptable.
>
>
>
>