Message ID | 20240822230635.954557-6-andrew.cooper3@citrix.com (mailing list archive) |
---|---|
State | Superseded |
Headers | show |
Series | xen/bitops: hweight() cleanup/improvements | expand |
On 23.08.2024 01:06, Andrew Cooper wrote: --- a/xen/include/xen/bitops.h +++ b/xen/include/xen/bitops.h @@ -35,6 +35,12 @@ extern void __bitop_bad_size(void); unsigned int __pure generic_ffsl(unsigned long x); unsigned int __pure generic_flsl(unsigned long x); > +/* > + * Hamming Weight, also called Population Count. Returns the number of set > + * bits in @x. > + */ > +unsigned int __pure generic_hweightl(unsigned long x); Aren't this and ... > @@ -284,6 +290,18 @@ static always_inline __pure unsigned int fls64(uint64_t x) > (_v & (_v - 1)) != 0; \ > }) > > +static always_inline __pure unsigned int hweightl(unsigned long x) ... this even __attribute_const__? > +{ > + if ( __builtin_constant_p(x) ) > + return __builtin_popcountl(x); How certain are you that compilers (even old ones) will reliably fold constant expressions here, and never emit a libgcc call instead? The conditions look to be more tight than just __builtin_constant_p(); a pretty absurd example: unsigned ltest(void) { return __builtin_constant_p("") ? __builtin_popcountl((unsigned long)"") : ~0; } > --- /dev/null > +++ b/xen/lib/generic-hweightl.c > @@ -0,0 +1,46 @@ > +/* SPDX-License-Identifier: GPL-2.0-only */ > + > +#include <xen/bitops.h> > +#include <xen/init.h> > +#include <xen/self-tests.h> > + > +/* Mask value @b broadcast to every byte in a long */ > +#if BITS_PER_LONG == 32 > +# define MASK(b) ((b) * 0x01010101UL) > +#elif BITS_PER_LONG == 64 > +# define MASK(b) ((b) * 0x0101010101010101UL) > +#else > +# error Extend me please > +#endif > + > +unsigned int generic_hweightl(unsigned long x) > +{ > + x -= (x >> 1) & MASK(0x55); > + x = (x & MASK(0x33)) + ((x >> 2) & MASK(0x33)); > + x = (x + (x >> 4)) & MASK(0x0f); > + > + if ( IS_ENABLED(CONFIG_HAS_FAST_MULTIPLY) ) > + return (x * MASK(0x01)) >> (BITS_PER_LONG - 8); I realize it's nitpicking, yet especially this use isn't really "mask"- like. Could I talk you into naming the macro e.g. BCST()? > + x += x >> 8; > + x += x >> 16; > +#if BITS_PER_LONG > 32 > + x += x >> 32; > +#endif > + > + return x & 0xff; > +} Perhaps #undef MASK here, or else ... > +#ifdef CONFIG_SELF_TESTS > +static void __init __constructor test_generic_hweightl(void) > +{ > + RUNTIME_CHECK(generic_hweightl, 0, 0); > + RUNTIME_CHECK(generic_hweightl, 1, 1); > + RUNTIME_CHECK(generic_hweightl, 3, 2); > + RUNTIME_CHECK(generic_hweightl, 7, 3); > + RUNTIME_CHECK(generic_hweightl, 0xff, 8); > + > + RUNTIME_CHECK(generic_hweightl, 1 | (1UL << (BITS_PER_LONG - 1)), 2); > + RUNTIME_CHECK(generic_hweightl, -1UL, BITS_PER_LONG); > +} ... actually use it some here, to have a few more cases? Jan
On 26/08/2024 12:40 pm, Jan Beulich wrote: > On 23.08.2024 01:06, Andrew Cooper wrote: > --- a/xen/include/xen/bitops.h > +++ b/xen/include/xen/bitops.h > @@ -35,6 +35,12 @@ extern void __bitop_bad_size(void); > unsigned int __pure generic_ffsl(unsigned long x); > unsigned int __pure generic_flsl(unsigned long x); > >> +/* >> + * Hamming Weight, also called Population Count. Returns the number of set >> + * bits in @x. >> + */ >> +unsigned int __pure generic_hweightl(unsigned long x); > Aren't this and ... > >> @@ -284,6 +290,18 @@ static always_inline __pure unsigned int fls64(uint64_t x) >> (_v & (_v - 1)) != 0; \ >> }) >> >> +static always_inline __pure unsigned int hweightl(unsigned long x) > ... this even __attribute_const__? Hmm. This is following fls(), but on further consideration, they should be const too. I'll do a prep patch fixing that, although I'm going to rename it to __attr_const for brevity. Much as I'd prefer __const, I expect that is going too far, making it too close to regular const. > >> +{ >> + if ( __builtin_constant_p(x) ) >> + return __builtin_popcountl(x); > How certain are you that compilers (even old ones) will reliably fold > constant expressions here, and never emit a libgcc call instead? The > conditions look to be more tight than just __builtin_constant_p(); a > pretty absurd example: > > unsigned ltest(void) { > return __builtin_constant_p("") ? __builtin_popcountl((unsigned long)"") : ~0; > } How do you express that in terms of a call to hweightl()? Again, this is following the layout started with fls() in order to avoid each arch opencoding different versions of constant folding. https://godbolt.org/z/r544c49oY When it's forced through the hweightl() interface, even GCC 4.1 decides that it's non-constant and falls back to generic_hweightl(). I did spend a *lot* of time with the fls() series checking that all compilers we supported did what we wanted in this case, so I don't expect it to be a problem. But, if a library call is emitted, it will be very obvious (link failure), and we can re-evaluate. > >> --- /dev/null >> +++ b/xen/lib/generic-hweightl.c >> @@ -0,0 +1,46 @@ >> +/* SPDX-License-Identifier: GPL-2.0-only */ >> + >> +#include <xen/bitops.h> >> +#include <xen/init.h> >> +#include <xen/self-tests.h> >> + >> +/* Mask value @b broadcast to every byte in a long */ >> +#if BITS_PER_LONG == 32 >> +# define MASK(b) ((b) * 0x01010101UL) >> +#elif BITS_PER_LONG == 64 >> +# define MASK(b) ((b) * 0x0101010101010101UL) >> +#else >> +# error Extend me please >> +#endif >> + >> +unsigned int generic_hweightl(unsigned long x) >> +{ >> + x -= (x >> 1) & MASK(0x55); >> + x = (x & MASK(0x33)) + ((x >> 2) & MASK(0x33)); >> + x = (x + (x >> 4)) & MASK(0x0f); >> + >> + if ( IS_ENABLED(CONFIG_HAS_FAST_MULTIPLY) ) >> + return (x * MASK(0x01)) >> (BITS_PER_LONG - 8); > I realize it's nitpicking, yet especially this use isn't really "mask"- > like. Could I talk you into naming the macro e.g. BCST()? Ok - I wasn't overly happy with the name MASK(), and BCST() is better. > >> + x += x >> 8; >> + x += x >> 16; >> +#if BITS_PER_LONG > 32 >> + x += x >> 32; >> +#endif >> + >> + return x & 0xff; >> +} > Perhaps #undef MASK here, or else ... > >> +#ifdef CONFIG_SELF_TESTS >> +static void __init __constructor test_generic_hweightl(void) >> +{ >> + RUNTIME_CHECK(generic_hweightl, 0, 0); >> + RUNTIME_CHECK(generic_hweightl, 1, 1); >> + RUNTIME_CHECK(generic_hweightl, 3, 2); >> + RUNTIME_CHECK(generic_hweightl, 7, 3); >> + RUNTIME_CHECK(generic_hweightl, 0xff, 8); >> + >> + RUNTIME_CHECK(generic_hweightl, 1 | (1UL << (BITS_PER_LONG - 1)), 2); >> + RUNTIME_CHECK(generic_hweightl, -1UL, BITS_PER_LONG); >> +} > ... actually use it some here, to have a few more cases? Hmm, why not. ~Andrew
On 27.08.2024 12:39, Andrew Cooper wrote: > On 26/08/2024 12:40 pm, Jan Beulich wrote: >> On 23.08.2024 01:06, Andrew Cooper wrote: >> --- a/xen/include/xen/bitops.h >> +++ b/xen/include/xen/bitops.h >> @@ -35,6 +35,12 @@ extern void __bitop_bad_size(void); >> unsigned int __pure generic_ffsl(unsigned long x); >> unsigned int __pure generic_flsl(unsigned long x); >> >>> +/* >>> + * Hamming Weight, also called Population Count. Returns the number of set >>> + * bits in @x. >>> + */ >>> +unsigned int __pure generic_hweightl(unsigned long x); >> Aren't this and ... >> >>> @@ -284,6 +290,18 @@ static always_inline __pure unsigned int fls64(uint64_t x) >>> (_v & (_v - 1)) != 0; \ >>> }) >>> >>> +static always_inline __pure unsigned int hweightl(unsigned long x) >> ... this even __attribute_const__? > > Hmm. This is following fls(), but on further consideration, they should > be const too. > > I'll do a prep patch fixing that, although I'm going to rename it to > __attr_const for brevity. > > Much as I'd prefer __const, I expect that is going too far, making it > too close to regular const. I was actually going to suggest using that name, if we want to shorten __attribute_const__. Yes, gcc treats __const (and __const__) as keywords, but do we care (much)? All it requires is that we don't start using __const as a (real) keyword. Of course __const is a good example of why really we shouldn't use double-underscore prefixed names anywhere. Any of them can be assigned a meaning by the compiler, and here that's clearly the case. Therefore, taking your planned rename, maybe better make it attr_const then? And eventually switch stuff like __packed, __pure, and __weak to attr_* as well? Or even introduce something like #define attr(attr...) __attribute__((attr)) and use attr(const) here? >>> +{ >>> + if ( __builtin_constant_p(x) ) >>> + return __builtin_popcountl(x); >> How certain are you that compilers (even old ones) will reliably fold >> constant expressions here, and never emit a libgcc call instead? The >> conditions look to be more tight than just __builtin_constant_p(); a >> pretty absurd example: >> >> unsigned ltest(void) { >> return __builtin_constant_p("") ? __builtin_popcountl((unsigned long)"") : ~0; >> } > > How do you express that in terms of a call to hweightl()? hweightl((unsigned long)""); Yet as said - it's absurd. It merely serves to make the point that what __builtin_constant_p() returns true for doesn't necessarily constant- fold in expressions. > Again, this is following the layout started with fls() in order to avoid > each arch opencoding different versions of constant folding. > > https://godbolt.org/z/r544c49oY > > When it's forced through the hweightl() interface, even GCC 4.1 decides > that it's non-constant and falls back to generic_hweightl(). > > > I did spend a *lot* of time with the fls() series checking that all > compilers we supported did what we wanted in this case, so I don't > expect it to be a problem. Right, and I guess I was pointlessly more concerned about popcount than I was for ffs() / fls(). The criteria upon which gcc decides whether to constant-fold the uses is exactly the same. > But, if a library call is emitted, it will > be very obvious (link failure), and we can re-evaluate. Indeed, we certainly would notice, albeit the diagnostic may be cryptic to people. Bottom line - keep it as is. Jan
On 27/08/2024 12:41 pm, Jan Beulich wrote: > On 27.08.2024 12:39, Andrew Cooper wrote: >> On 26/08/2024 12:40 pm, Jan Beulich wrote: >>> On 23.08.2024 01:06, Andrew Cooper wrote: >>> --- a/xen/include/xen/bitops.h >>> +++ b/xen/include/xen/bitops.h >>> @@ -35,6 +35,12 @@ extern void __bitop_bad_size(void); >>> unsigned int __pure generic_ffsl(unsigned long x); >>> unsigned int __pure generic_flsl(unsigned long x); >>> >>>> +/* >>>> + * Hamming Weight, also called Population Count. Returns the number of set >>>> + * bits in @x. >>>> + */ >>>> +unsigned int __pure generic_hweightl(unsigned long x); >>> Aren't this and ... >>> >>>> @@ -284,6 +290,18 @@ static always_inline __pure unsigned int fls64(uint64_t x) >>>> (_v & (_v - 1)) != 0; \ >>>> }) >>>> >>>> +static always_inline __pure unsigned int hweightl(unsigned long x) >>> ... this even __attribute_const__? >> Hmm. This is following fls(), but on further consideration, they should >> be const too. >> >> I'll do a prep patch fixing that, although I'm going to rename it to >> __attr_const for brevity. >> >> Much as I'd prefer __const, I expect that is going too far, making it >> too close to regular const. > I was actually going to suggest using that name, if we want to shorten > __attribute_const__. Yes, gcc treats __const (and __const__) as > keywords, but do we care (much)? All it requires is that we don't start > using __const as a (real) keyword. Well also we'll get into more MISRA fun for overriding keywords. But yes - the fact that GCC treats __const to mean const is precisely why we shouldn't give it an unrelated meaning. > > Of course __const is a good example of why really we shouldn't use > double-underscore prefixed names anywhere. Any of them can be assigned > a meaning by the compiler, and here that's clearly the case. Therefore, > taking your planned rename, maybe better make it attr_const then? And > eventually switch stuff like __packed, __pure, and __weak to attr_* as > well? Or even introduce something like > > #define attr(attr...) __attribute__((attr)) > > and use attr(const) here? Hmm - that's an interesting approach, and for other attributes which we can use unconditionally. It will end up shorter than multiple separate __-prefixed names. As a tangent, I've got some work from playing with -fanalyzer which sprinkles some attr malloc/alloc_{size,align}()/free around. It does improve code generation (abeit marginally), but the function declaration size suffers. It won't work for attributes which are conditionally nothing (e.g. cf_check), or ones that contain multiple aspects (e.g. __constructer conataining cf_check). In practice this means we're always going to end up with a mix, so maybe attr_const is better for consistency. > >>>> +{ >>>> + if ( __builtin_constant_p(x) ) >>>> + return __builtin_popcountl(x); >>> How certain are you that compilers (even old ones) will reliably fold >>> constant expressions here, and never emit a libgcc call instead? The >>> conditions look to be more tight than just __builtin_constant_p(); a >>> pretty absurd example: >>> >>> unsigned ltest(void) { >>> return __builtin_constant_p("") ? __builtin_popcountl((unsigned long)"") : ~0; >>> } >> How do you express that in terms of a call to hweightl()? > hweightl((unsigned long)""); > > Yet as said - it's absurd. It merely serves to make the point that what > __builtin_constant_p() returns true for doesn't necessarily constant- > fold in expressions. Yes, but as shown in the godbolt link, this form changes GCC's mind about the __builtin_const-ness of the expression. > >> Again, this is following the layout started with fls() in order to avoid >> each arch opencoding different versions of constant folding. >> >> https://godbolt.org/z/r544c49oY >> >> When it's forced through the hweightl() interface, even GCC 4.1 decides >> that it's non-constant and falls back to generic_hweightl(). >> >> >> I did spend a *lot* of time with the fls() series checking that all >> compilers we supported did what we wanted in this case, so I don't >> expect it to be a problem. > Right, and I guess I was pointlessly more concerned about popcount than > I was for ffs() / fls(). The criteria upon which gcc decides whether to > constant-fold the uses is exactly the same. > >> But, if a library call is emitted, it will >> be very obvious (link failure), and we can re-evaluate. > Indeed, we certainly would notice, albeit the diagnostic may be cryptic > to people. > > Bottom line - keep it as is. Thanks. ~Andrew
diff --git a/xen/arch/ppc/include/asm/bitops.h b/xen/arch/ppc/include/asm/bitops.h index a62c4f99c3bb..64512e949530 100644 --- a/xen/arch/ppc/include/asm/bitops.h +++ b/xen/arch/ppc/include/asm/bitops.h @@ -124,6 +124,8 @@ static inline int test_and_set_bit(unsigned int nr, volatile void *addr) #define arch_fls(x) ((x) ? 32 - __builtin_clz(x) : 0) #define arch_flsl(x) ((x) ? BITS_PER_LONG - __builtin_clzl(x) : 0) +#define arch_hweightl(x) __builtin_popcountl(x) + /** * hweightN - returns the hamming weight of a N-bit word * @x: the word to weigh diff --git a/xen/common/bitops.c b/xen/common/bitops.c index 4545682aa8e0..d0c268b4994a 100644 --- a/xen/common/bitops.c +++ b/xen/common/bitops.c @@ -106,10 +106,24 @@ static void __init test_multiple_bits_set(void) CHECK(multiple_bits_set, 0xc000000000000000ULL, true); } +static void __init test_hweight(void) +{ + /* unsigned int hweightl(unsigned long) */ + CHECK(hweightl, 0, 0); + CHECK(hweightl, 1, 1); + CHECK(hweightl, 3, 2); + CHECK(hweightl, 7, 3); + CHECK(hweightl, 0xff, 8); + + CHECK(hweightl, 1 | (1UL << (BITS_PER_LONG - 1)), 2); + CHECK(hweightl, -1UL, BITS_PER_LONG); +} + static void __init __constructor test_bitops(void) { test_ffs(); test_fls(); test_multiple_bits_set(); + test_hweight(); } diff --git a/xen/include/xen/bitops.h b/xen/include/xen/bitops.h index 64d70a7a1cb5..3aac10b7f532 100644 --- a/xen/include/xen/bitops.h +++ b/xen/include/xen/bitops.h @@ -35,6 +35,12 @@ extern void __bitop_bad_size(void); unsigned int __pure generic_ffsl(unsigned long x); unsigned int __pure generic_flsl(unsigned long x); +/* + * Hamming Weight, also called Population Count. Returns the number of set + * bits in @x. + */ +unsigned int __pure generic_hweightl(unsigned long x); + /** * generic__test_and_set_bit - Set a bit and return its old value * @nr: Bit to set @@ -284,6 +290,18 @@ static always_inline __pure unsigned int fls64(uint64_t x) (_v & (_v - 1)) != 0; \ }) +static always_inline __pure unsigned int hweightl(unsigned long x) +{ + if ( __builtin_constant_p(x) ) + return __builtin_popcountl(x); + +#ifdef arch_hweightl + return arch_hweightl(x); +#else + return generic_hweightl(x); +#endif +} + /* --------------------- Please tidy below here --------------------- */ #ifndef find_next_bit diff --git a/xen/lib/Makefile b/xen/lib/Makefile index a48541596470..b6558e108bd9 100644 --- a/xen/lib/Makefile +++ b/xen/lib/Makefile @@ -6,6 +6,7 @@ lib-y += ctype.o lib-y += find-next-bit.o lib-y += generic-ffsl.o lib-y += generic-flsl.o +lib-y += generic-hweightl.o lib-y += list-sort.o lib-y += memchr.o lib-y += memchr_inv.o diff --git a/xen/lib/generic-hweightl.c b/xen/lib/generic-hweightl.c new file mode 100644 index 000000000000..fa4bbec273ab --- /dev/null +++ b/xen/lib/generic-hweightl.c @@ -0,0 +1,46 @@ +/* SPDX-License-Identifier: GPL-2.0-only */ + +#include <xen/bitops.h> +#include <xen/init.h> +#include <xen/self-tests.h> + +/* Mask value @b broadcast to every byte in a long */ +#if BITS_PER_LONG == 32 +# define MASK(b) ((b) * 0x01010101UL) +#elif BITS_PER_LONG == 64 +# define MASK(b) ((b) * 0x0101010101010101UL) +#else +# error Extend me please +#endif + +unsigned int generic_hweightl(unsigned long x) +{ + x -= (x >> 1) & MASK(0x55); + x = (x & MASK(0x33)) + ((x >> 2) & MASK(0x33)); + x = (x + (x >> 4)) & MASK(0x0f); + + if ( IS_ENABLED(CONFIG_HAS_FAST_MULTIPLY) ) + return (x * MASK(0x01)) >> (BITS_PER_LONG - 8); + + x += x >> 8; + x += x >> 16; +#if BITS_PER_LONG > 32 + x += x >> 32; +#endif + + return x & 0xff; +} + +#ifdef CONFIG_SELF_TESTS +static void __init __constructor test_generic_hweightl(void) +{ + RUNTIME_CHECK(generic_hweightl, 0, 0); + RUNTIME_CHECK(generic_hweightl, 1, 1); + RUNTIME_CHECK(generic_hweightl, 3, 2); + RUNTIME_CHECK(generic_hweightl, 7, 3); + RUNTIME_CHECK(generic_hweightl, 0xff, 8); + + RUNTIME_CHECK(generic_hweightl, 1 | (1UL << (BITS_PER_LONG - 1)), 2); + RUNTIME_CHECK(generic_hweightl, -1UL, BITS_PER_LONG); +} +#endif /* CONFIG_SELF_TESTS */
There are 6 remaining callers in Xen: * The two hweight32() calls, _domain_struct_bits() and efi_find_gop_mode(), are __init only. * The two hweight_long() calls are both in bitmap_weight(). * The two hweight64() calls are hv_vpset_nr_banks() and x86_emulate(). Only bitmap_weight() and possibly hv_vpset_nr_banks() can be considered fast(ish) paths, and they're all of GPR-width form. Furthermore, the differences between a generic int and generic long form is only an ADD and SHIFT, and only in !CONFIG_HAS_FAST_MULTIPLY builds. Therefore, it is definitely not worth having both generic implemenations. Implement generic_hweightl() based on the current generic_hweight64(), adjusted to be compatible with ARM32, along with standard SELF_TESTS. Implement hweightl() with usual constant-folding and arch opt-in support. PPC is the only architecture that devates from generic, and it simply uses the builtin. No functional change. Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com> --- CC: Jan Beulich <JBeulich@suse.com> CC: Roger Pau Monné <roger.pau@citrix.com> CC: Wei Liu <wl@xen.org> CC: Stefano Stabellini <sstabellini@kernel.org> CC: Julien Grall <julien@xen.org> CC: Volodymyr Babchuk <Volodymyr_Babchuk@epam.com> CC: Bertrand Marquis <bertrand.marquis@arm.com> CC: Michal Orzel <michal.orzel@amd.com> CC: Oleksii Kurochko <oleksii.kurochko@gmail.com> CC: Shawn Anastasio <sanastasio@raptorengineering.com> --- xen/arch/ppc/include/asm/bitops.h | 2 ++ xen/common/bitops.c | 14 ++++++++++ xen/include/xen/bitops.h | 18 ++++++++++++ xen/lib/Makefile | 1 + xen/lib/generic-hweightl.c | 46 +++++++++++++++++++++++++++++++ 5 files changed, 81 insertions(+) create mode 100644 xen/lib/generic-hweightl.c