[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] [PATCH 4/4] x86: use POPCNT for hweight<N>() when available


  • To: Jan Beulich <JBeulich@xxxxxxxx>, xen-devel <xen-devel@xxxxxxxxxxxxxxxxxxxx>
  • From: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>
  • Date: Fri, 31 May 2019 13:43:36 -0700
  • Authentication-results: esa1.hc3370-68.iphmx.com; dkim=none (message not signed) header.i=none; spf=None smtp.pra=andrew.cooper3@xxxxxxxxxx; spf=Pass smtp.mailfrom=Andrew.Cooper3@xxxxxxxxxx; spf=None smtp.helo=postmaster@xxxxxxxxxxxxxxxxxxxxxxxxxx
  • Autocrypt: addr=andrew.cooper3@xxxxxxxxxx; prefer-encrypt=mutual; keydata= mQINBFLhNn8BEADVhE+Hb8i0GV6mihnnr/uiQQdPF8kUoFzCOPXkf7jQ5sLYeJa0cQi6Penp VtiFYznTairnVsN5J+ujSTIb+OlMSJUWV4opS7WVNnxHbFTPYZVQ3erv7NKc2iVizCRZ2Kxn srM1oPXWRic8BIAdYOKOloF2300SL/bIpeD+x7h3w9B/qez7nOin5NzkxgFoaUeIal12pXSR Q354FKFoy6Vh96gc4VRqte3jw8mPuJQpfws+Pb+swvSf/i1q1+1I4jsRQQh2m6OTADHIqg2E ofTYAEh7R5HfPx0EXoEDMdRjOeKn8+vvkAwhviWXTHlG3R1QkbE5M/oywnZ83udJmi+lxjJ5 YhQ5IzomvJ16H0Bq+TLyVLO/VRksp1VR9HxCzItLNCS8PdpYYz5TC204ViycobYU65WMpzWe LFAGn8jSS25XIpqv0Y9k87dLbctKKA14Ifw2kq5OIVu2FuX+3i446JOa2vpCI9GcjCzi3oHV e00bzYiHMIl0FICrNJU0Kjho8pdo0m2uxkn6SYEpogAy9pnatUlO+erL4LqFUO7GXSdBRbw5 gNt25XTLdSFuZtMxkY3tq8MFss5QnjhehCVPEpE6y9ZjI4XB8ad1G4oBHVGK5LMsvg22PfMJ ISWFSHoF/B5+lHkCKWkFxZ0gZn33ju5n6/FOdEx4B8cMJt+cWwARAQABtClBbmRyZXcgQ29v cGVyIDxhbmRyZXcuY29vcGVyM0BjaXRyaXguY29tPokCOgQTAQgAJAIbAwULCQgHAwUVCgkI CwUWAgMBAAIeAQIXgAUCWKD95wIZAQAKCRBlw/kGpdefoHbdD/9AIoR3k6fKl+RFiFpyAhvO 59ttDFI7nIAnlYngev2XUR3acFElJATHSDO0ju+hqWqAb8kVijXLops0gOfqt3VPZq9cuHlh IMDquatGLzAadfFx2eQYIYT+FYuMoPZy/aTUazmJIDVxP7L383grjIkn+7tAv+qeDfE+txL4 SAm1UHNvmdfgL2/lcmL3xRh7sub3nJilM93RWX1Pe5LBSDXO45uzCGEdst6uSlzYR/MEr+5Z JQQ32JV64zwvf/aKaagSQSQMYNX9JFgfZ3TKWC1KJQbX5ssoX/5hNLqxMcZV3TN7kU8I3kjK mPec9+1nECOjjJSO/h4P0sBZyIUGfguwzhEeGf4sMCuSEM4xjCnwiBwftR17sr0spYcOpqET ZGcAmyYcNjy6CYadNCnfR40vhhWuCfNCBzWnUW0lFoo12wb0YnzoOLjvfD6OL3JjIUJNOmJy RCsJ5IA/Iz33RhSVRmROu+TztwuThClw63g7+hoyewv7BemKyuU6FTVhjjW+XUWmS/FzknSi dAG+insr0746cTPpSkGl3KAXeWDGJzve7/SBBfyznWCMGaf8E2P1oOdIZRxHgWj0zNr1+ooF /PzgLPiCI4OMUttTlEKChgbUTQ+5o0P080JojqfXwbPAyumbaYcQNiH1/xYbJdOFSiBv9rpt TQTBLzDKXok86LkCDQRS4TZ/ARAAkgqudHsp+hd82UVkvgnlqZjzz2vyrYfz7bkPtXaGb9H4 Rfo7mQsEQavEBdWWjbga6eMnDqtu+FC+qeTGYebToxEyp2lKDSoAsvt8w82tIlP/EbmRbDVn 7bhjBlfRcFjVYw8uVDPptT0TV47vpoCVkTwcyb6OltJrvg/QzV9f07DJswuda1JH3/qvYu0p vjPnYvCq4NsqY2XSdAJ02HrdYPFtNyPEntu1n1KK+gJrstjtw7KsZ4ygXYrsm/oCBiVW/OgU g/XIlGErkrxe4vQvJyVwg6YH653YTX5hLLUEL1NS4TCo47RP+wi6y+TnuAL36UtK/uFyEuPy wwrDVcC4cIFhYSfsO0BumEI65yu7a8aHbGfq2lW251UcoU48Z27ZUUZd2Dr6O/n8poQHbaTd 6bJJSjzGGHZVbRP9UQ3lkmkmc0+XCHmj5WhwNNYjgbbmML7y0fsJT5RgvefAIFfHBg7fTY/i kBEimoUsTEQz+N4hbKwo1hULfVxDJStE4sbPhjbsPCrlXf6W9CxSyQ0qmZ2bXsLQYRj2xqd1 bpA+1o1j2N4/au1R/uSiUFjewJdT/LX1EklKDcQwpk06Af/N7VZtSfEJeRV04unbsKVXWZAk uAJyDDKN99ziC0Wz5kcPyVD1HNf8bgaqGDzrv3TfYjwqayRFcMf7xJaL9xXedMcAEQEAAYkC HwQYAQgACQUCUuE2fwIbDAAKCRBlw/kGpdefoG4XEACD1Qf/er8EA7g23HMxYWd3FXHThrVQ HgiGdk5Yh632vjOm9L4sd/GCEACVQKjsu98e8o3ysitFlznEns5EAAXEbITrgKWXDDUWGYxd pnjj2u+GkVdsOAGk0kxczX6s+VRBhpbBI2PWnOsRJgU2n10PZ3mZD4Xu9kU2IXYmuW+e5KCA vTArRUdCrAtIa1k01sPipPPw6dfxx2e5asy21YOytzxuWFfJTGnVxZZSCyLUO83sh6OZhJkk b9rxL9wPmpN/t2IPaEKoAc0FTQZS36wAMOXkBh24PQ9gaLJvfPKpNzGD8XWR5HHF0NLIJhgg 4ZlEXQ2fVp3XrtocHqhu4UZR4koCijgB8sB7Tb0GCpwK+C4UePdFLfhKyRdSXuvY3AHJd4CP 4JzW0Bzq/WXY3XMOzUTYApGQpnUpdOmuQSfpV9MQO+/jo7r6yPbxT7CwRS5dcQPzUiuHLK9i nvjREdh84qycnx0/6dDroYhp0DFv4udxuAvt1h4wGwTPRQZerSm4xaYegEFusyhbZrI0U9tJ B8WrhBLXDiYlyJT6zOV2yZFuW47VrLsjYnHwn27hmxTC/7tvG3euCklmkn9Sl9IAKFu29RSo d5bD8kMSCYsTqtTfT6W4A3qHGvIDta3ptLYpIAOD2sY3GYq2nf3Bbzx81wZK14JdDDHUX2Rs 6+ahAA==
  • Cc: George Dunlap <George.Dunlap@xxxxxxxxxxxxx>, Wei Liu <wl@xxxxxxx>, Roger Pau Monne <roger.pau@xxxxxxxxxx>
  • Delivery-date: Fri, 31 May 2019 22:11:18 +0000
  • Ironport-sdr: XJfp5lzJQ6+3ChXFR0AWCCHZmnhX8vdRJyJNK34OwO1SA6/GsUIGmIOFyu7WRplduXH/V0o7pz Oz2pBHq6PFcHaSPZfiXMzrGp4IVyGlnxf97eq5Lql3bomq4DzAMr8oBMpYgCKaAQbWoMnBI2Jm wn7BEllORayRk/kqmMyeMXxpnDxheMGTKY5Z87N5PBc4voqYqGjSLZVLgRfsW4l/yNf1X8Ymo+ wIrjlUjaTdhxnM66bXRztvj7uPq27akgF3+u8KP6FFjBnESJIV3V73SDB464Y3c1/6G4z/6UgC jpo=
  • List-id: Xen developer discussion <xen-devel.lists.xenproject.org>
  • Openpgp: preference=signencrypt

On 31/05/2019 02:54, Jan Beulich wrote:
> This is faster than using the software implementation, and the insn is
> available on all half-way recent hardware. Therefore convert
> generic_hweight<N>() to out-of-line functions (without affecting Arm)
> and use alternatives patching to replace the function calls.
>
> Suggested-by: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>
> Signed-off-by: Jan Beulich <jbeulich@xxxxxxxx>

So, I trust you weren't expecting to just ack this and let it go in?

The principle of the patch (use popcnt when available) is an improvement
which I'm entirely in agreement with, but everything else is a problem.

The long and the short of it is that I'm not going to accept any version
of this which isn't the Linux version.

>From a microarchitectural standpoint, the tradeoff between fractional
register scheduling flexibility (which in practice is largely bound
anyway by real function calls in surrounding code) and increased icache
pressure/coldness (from the redundant function copies) falls largely in
favour of the Linux way of doing it, a cold icache line is
disproportionally more expensive than requiring the compiler to order
its registers differently (especially as all non-obsolete processors
these days have zero-cost register renaming internally, for the purpose
of superscalar execution).

If however you can demonstrate that a number of CPU microarchitects are
wrong about what is faster in practice, then I will happily accept a
patch which matches Linux, once you've improved the performance of Linux.

> @@ -245,6 +246,9 @@ boot/mkelf32: boot/mkelf32.c
>  efi/mkreloc: efi/mkreloc.c
>       $(HOSTCC) $(HOSTCFLAGS) -g -o $@ $<
>  
> +nocov-y += hweight.o

Irrespective of the exact specifics of how the patch ends up, I don't
think the nocov restriction is a direction we want to take.

Coverage may not be a thing used in production, but when it is used for
development, it needs to not have random holes missing in the results data.

> +hweight.o: CFLAGS += $(foreach reg,cx dx si 8 9 10 11,-ffixed-r$(reg))
> +

Does this work with Clang?


> --- /dev/null
> +++ b/xen/arch/x86/hweight.c
> @@ -0,0 +1,28 @@
> +#define generic_hweight64 _hweight64
> +#define generic_hweight32 _hweight32
> +#define generic_hweight16 _hweight16
> +#define generic_hweight8  _hweight8
> +
> +#include <xen/compiler.h>
> +
> +#undef inline
> +#define inline always_inline
> +
> +#include <xen/bitops.h>
> +
> +#undef generic_hweight8
> +#undef generic_hweight16
> +#undef generic_hweight32
> +#undef generic_hweight64
> +
> +#define HWEIGHT(n)                                             \
> +typeof(_hweight##n) generic_hweight##n;                        \
> +unsigned int generic_hweight##n(typeof((uint##n##_t)0 + 0U) x) \

A question to the rest of xen-devel.  Is there anyone else who can
actually work out what this construct is doing?

I'd like to get a feel for how many people can even follow some of our C.

> --- a/xen/include/asm-x86/bitops.h
> +++ b/xen/include/asm-x86/bitops.h
> @@ -469,15 +469,35 @@ static inline int fls(unsigned int x)
>      return r + 1;
>  }
>  
> +/* POPCNT encodings with %{r,e}di input and %{r,e}ax output: */
> +#define POPCNT_64 ".byte 0xF3, 0x48, 0x0F, 0xB8, 0xC7"
> +#define POPCNT_32 ".byte 0xF3, 0x0F, 0xB8, 0xC7"

So (the dangers of false micro-optimsiation aside), POPCNT_32 will
probably be better using a redundant %ds prefix.

The reason is that the result needs padding to 5 bytes, as the original
instruction is `call disp32`, meaning that a single byte nop needs
inserting.  The version with a single nop takes two decode ports as
opposed to one, and single byte nops are forced to take an execution
delay for backwards compatibility with DoS.

OTOH, I also bet that noone could observe a difference without using
perf counters and fetch/decode uarch events.

~Andrew

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel

 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.