[PATCH] x86-64: Only define used SSE/AVX/AVX512 run-time resolvers
H.J. Lu
hjl.tools@gmail.com
Mon Jun 27 20:19:38 GMT 2022
On Mon, Jun 27, 2022 at 1:09 PM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
>
> On Mon, Jun 27, 2022 at 1:00 PM H.J. Lu <hjl.tools@gmail.com> wrote:
> >
> > On Mon, Jun 27, 2022 at 12:51 PM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
> > >
> > > On Mon, Jun 27, 2022 at 12:44 PM H.J. Lu <hjl.tools@gmail.com> wrote:
> > > >
> > > > On Mon, Jun 27, 2022 at 12:23 PM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
> > > > >
> > > > > On Mon, Jun 27, 2022 at 12:07 PM H.J. Lu <hjl.tools@gmail.com> wrote:
> > > > > >
> > > > > > When glibc is built with x86-64 ISA level v3, SSE run-time resolvers
> > > > > > aren't used. For x86-64 ISA level v4 build, both SSE and AVX resolvers
> > > > > > are unused.
> > > > > >
> > > > > > 1. Move X86_ISA_CPU_FEATURE_USABLE_P and X86_ISA_CPU_FEATURES_ARCH_P to
> > > > > > isa-level.h.
> > > > > > 2. Check the minimum x86-64 ISA level to exclude the unused run-time
> > > > > > resolvers.
> > > > > > ---
> > > > > > sysdeps/x86/isa-ifunc-macros.h | 27 ----------------
> > > > > > sysdeps/x86/isa-level.h | 26 +++++++++++++++
> > > > > > sysdeps/x86_64/dl-machine.h | 12 ++++---
> > > > > > sysdeps/x86_64/dl-trampoline.S | 59 ++++++++++++++++++----------------
> > > > > > 4 files changed, 66 insertions(+), 58 deletions(-)
> > > > > >
> > > > > > diff --git a/sysdeps/x86/isa-ifunc-macros.h b/sysdeps/x86/isa-ifunc-macros.h
> > > > > > index d69905689b..f967a1bec6 100644
> > > > > > --- a/sysdeps/x86/isa-ifunc-macros.h
> > > > > > +++ b/sysdeps/x86/isa-ifunc-macros.h
> > > > > > @@ -56,31 +56,4 @@
> > > > > > # define X86_IFUNC_IMPL_ADD_V1(...)
> > > > > > #endif
> > > > > >
> > > > > > -/* Both X86_ISA_CPU_FEATURE_USABLE_P and X86_ISA_CPU_FEATURES_ARCH_P
> > > > > > - macros are wrappers for the the respective
> > > > > > - CPU_FEATURE{S}_{USABLE|ARCH}_P runtime checks. They differ in two
> > > > > > - ways.
> > > > > > -
> > > > > > - 1. The USABLE_P version is evaluated to true when the feature
> > > > > > - is enabled.
> > > > > > -
> > > > > > - 2. The ARCH_P version has a third argument `not`. The `not`
> > > > > > - argument can either be '!' or empty. If the feature is
> > > > > > - enabled above an ISA level, the third argument should be empty
> > > > > > - and the expression is evaluated to true when the feature is
> > > > > > - enabled. If the feature is disabled above an ISA level, the
> > > > > > - third argument should be `!` and the expression is evaluated
> > > > > > - to true when the feature is disabled.
> > > > > > - */
> > > > > > -
> > > > > > -#define X86_ISA_CPU_FEATURE_USABLE_P(ptr, name) \
> > > > > > - (((name##_X86_ISA_LEVEL) <= MINIMUM_X86_ISA_LEVEL) \
> > > > > > - || CPU_FEATURE_USABLE_P (ptr, name))
> > > > > > -
> > > > > > -
> > > > > > -#define X86_ISA_CPU_FEATURES_ARCH_P(ptr, name, not) \
> > > > > > - (((name##_X86_ISA_LEVEL) <= MINIMUM_X86_ISA_LEVEL) \
> > > > > > - || not CPU_FEATURES_ARCH_P (ptr, name))
> > > > > > -
> > > > > > -
> > > > > > #endif
> > > > > > diff --git a/sysdeps/x86/isa-level.h b/sysdeps/x86/isa-level.h
> > > > > > index 075e7c6ee1..f293aea906 100644
> > > > > > --- a/sysdeps/x86/isa-level.h
> > > > > > +++ b/sysdeps/x86/isa-level.h
> > > > > > @@ -68,10 +68,12 @@
> > > > > > compile-time constant.. */
> > > > > >
> > > > > > /* ISA level >= 4 guaranteed includes. */
> > > > > > +#define AVX512F_X86_ISA_LEVEL 4
> > > > > > #define AVX512VL_X86_ISA_LEVEL 4
> > > > > > #define AVX512BW_X86_ISA_LEVEL 4
> > > > > >
> > > > > > /* ISA level >= 3 guaranteed includes. */
> > > > > > +#define AVX_X86_ISA_LEVEL 3
> > > > > > #define AVX2_X86_ISA_LEVEL 3
> > > > > > #define BMI2_X86_ISA_LEVEL 3
> > > > > >
> > > > > > @@ -87,6 +89,30 @@
> > > > > > when ISA level < 3. */
> > > > > > #define Prefer_No_VZEROUPPER_X86_ISA_LEVEL 3
> > > > > >
> > > > > > +/* Both X86_ISA_CPU_FEATURE_USABLE_P and X86_ISA_CPU_FEATURES_ARCH_P
> > > > > > + macros are wrappers for the respective CPU_FEATURE{S}_{USABLE|ARCH}_P
> > > > > > + runtime checks. They differ in two ways.
> > > > > > +
> > > > > > + 1. The USABLE_P version is evaluated to true when the feature
> > > > > > + is enabled.
> > > > > > +
> > > > > > + 2. The ARCH_P version has a third argument `not`. The `not`
> > > > > > + argument can either be `!` or empty. If the feature is
> > > > > > + enabled above an ISA level, the third argument should be empty
> > > > > > + and the expression is evaluated to true when the feature is
> > > > > > + enabled. If the feature is disabled above an ISA level, the
> > > > > > + third argument should be `!` and the expression is evaluated
> > > > > > + to true when the feature is disabled.
> > > > > > + */
> > > > > > +
> > > > > > +#define X86_ISA_CPU_FEATURE_USABLE_P(ptr, name) \
> > > > > > + (((name##_X86_ISA_LEVEL) <= MINIMUM_X86_ISA_LEVEL) \
> > > > > > + || CPU_FEATURE_USABLE_P (ptr, name))
> > > > > > +
> > > > > > +#define X86_ISA_CPU_FEATURES_ARCH_P(ptr, name, not) \
> > > > > > + (((name##_X86_ISA_LEVEL) <= MINIMUM_X86_ISA_LEVEL) \
> > > > > > + || not CPU_FEATURES_ARCH_P (ptr, name))
> > > > > > +
> > > > > > #define ISA_SHOULD_BUILD(isa_build_level) \
> > > > > > (MINIMUM_X86_ISA_LEVEL <= (isa_build_level) && IS_IN (libc)) \
> > > > > > || defined ISA_DEFAULT_IMPL
> > > > > > diff --git a/sysdeps/x86_64/dl-machine.h b/sysdeps/x86_64/dl-machine.h
> > > > > > index 34766325ae..005d089501 100644
> > > > > > --- a/sysdeps/x86_64/dl-machine.h
> > > > > > +++ b/sysdeps/x86_64/dl-machine.h
> > > > > > @@ -28,6 +28,7 @@
> > > > > > #include <dl-tlsdesc.h>
> > > > > > #include <dl-static-tls.h>
> > > > > > #include <dl-machine-rel.h>
> > > > > > +#include <isa-level.h>
> > > > > >
> > > > > > /* Return nonzero iff ELF header is compatible with the running host. */
> > > > > > static inline int __attribute__ ((unused))
> > > > > > @@ -86,6 +87,8 @@ elf_machine_runtime_setup (struct link_map *l, struct r_scope_elem *scope[],
> > > > > > /* Identify this shared object. */
> > > > > > *(ElfW(Addr) *) (got + 1) = (ElfW(Addr)) l;
> > > > > >
> > > > > > + const struct cpu_features* cpu_features = __get_cpu_features ();
> > > > > > +
> > > > > > /* The got[2] entry contains the address of a function which gets
> > > > > > called to get the address of a so far unresolved function and
> > > > > > jump to it. The profiling extension of the dynamic linker allows
> > > > > > @@ -94,9 +97,9 @@ elf_machine_runtime_setup (struct link_map *l, struct r_scope_elem *scope[],
> > > > > > end in this function. */
> > > > > > if (__glibc_unlikely (profile))
> > > > > > {
> > > > > > - if (CPU_FEATURE_USABLE (AVX512F))
> > > > > > + if (X86_ISA_CPU_FEATURE_USABLE_P (cpu_features, AVX512F))
> > > > > > *(ElfW(Addr) *) (got + 2) = (ElfW(Addr)) &_dl_runtime_profile_avx512;
> > > > > > - else if (CPU_FEATURE_USABLE (AVX))
> > > > > > + else if (X86_ISA_CPU_FEATURE_USABLE_P (cpu_features, AVX))
> > > > > > *(ElfW(Addr) *) (got + 2) = (ElfW(Addr)) &_dl_runtime_profile_avx;
> > > > > > else
> > > > > > *(ElfW(Addr) *) (got + 2) = (ElfW(Addr)) &_dl_runtime_profile_sse;
> > > > > > @@ -112,9 +115,10 @@ elf_machine_runtime_setup (struct link_map *l, struct r_scope_elem *scope[],
> > > > > > /* This function will get called to fix up the GOT entry
> > > > > > indicated by the offset on the stack, and then jump to
> > > > > > the resolved address. */
> > > > > > - if (GLRO(dl_x86_cpu_features).xsave_state_size != 0)
> > > > > > + if (MINIMUM_X86_ISA_LEVEL >= AVX_X86_ISA_LEVEL
> > > > > > + || GLRO(dl_x86_cpu_features).xsave_state_size != 0)
> > > > > > *(ElfW(Addr) *) (got + 2)
> > > > > > - = (CPU_FEATURE_USABLE (XSAVEC)
> > > > > > + = (CPU_FEATURE_USABLE_P (cpu_features, XSAVEC)
> > > > > > ? (ElfW(Addr)) &_dl_runtime_resolve_xsavec
> > > > > > : (ElfW(Addr)) &_dl_runtime_resolve_xsave);
> > > > > > else
> > > > > > diff --git a/sysdeps/x86_64/dl-trampoline.S b/sysdeps/x86_64/dl-trampoline.S
> > > > > > index 831a654713..f669805ac5 100644
> > > > > > --- a/sysdeps/x86_64/dl-trampoline.S
> > > > > > +++ b/sysdeps/x86_64/dl-trampoline.S
> > > > > > @@ -20,6 +20,7 @@
> > > > > > #include <sysdep.h>
> > > > > > #include <cpu-features-offsets.h>
> > > > > > #include <link-defines.h>
> > > > > > +#include <isa-level.h>
> > > > > >
> > > > > > #ifndef DL_STACK_ALIGNMENT
> > > > > > /* Due to GCC bug:
> > > > > > @@ -62,35 +63,39 @@
> > > > >
> > > > > Can you also add a guard for isa level v4 on the avx512 version?
> > > >
> > > > It is needed only when we need to save and restore new registers
> > > > in _dl_runtime_profile, which requires a new _dl_runtime_profile.
> > > > We can add the check then.
> > >
> > > ?
> > > Why does the prohibit the check?
> >
> > I don't think we need another _dl_runtime_XXX any time soon.
> > Adding a check may indicate otherwise.
>
> Guess my thought here is everywhere else we are adding for all
> ISA levels. Just seems like an inconsistency that we will potentially
> miss sometime in the future.
For this one, a check may lead to the question when the check will
be false.
>
> >
> > > >
> > > > > > #undef VMOVA
> > > > > > #undef VEC_SIZE
> > > > > >
> > > > > > -#define VEC_SIZE 32
> > > > > > -#define VMOVA vmovdqa
> > > > > > -#define VEC(i) ymm##i
> > > > > > -#define _dl_runtime_profile _dl_runtime_profile_avx
> > > > > > -#include "dl-trampoline.h"
> > > > > > -#undef _dl_runtime_profile
> > > > > > -#undef VEC
> > > > > > -#undef VMOVA
> > > > > > -#undef VEC_SIZE
> > > > > > +#if MINIMUM_X86_ISA_LEVEL <= AVX_X86_ISA_LEVEL
> > > > > Can this be <= __X86_ISA_LEVEL_V3
> > > >
> > > > No, since __X86_ISA_LEVEL_VN is defined either 0 or 1.
> > >
> > > Then <= 3 / <= 2?
> >
> > Since AVX and AVX512F are the features which _dl_runtime_XXX
> > supports, AVX_X86_ISA_LEVEL is more appropriate than a number.
> kk
> >
> > > >
> > > > > > +# define VEC_SIZE 32
> > > > > > +# define VMOVA vmovdqa
> > > > > > +# define VEC(i) ymm##i
> > > > > > +# define _dl_runtime_profile _dl_runtime_profile_avx
> > > > > > +# include "dl-trampoline.h"
> > > > > > +# undef _dl_runtime_profile
> > > > > > +# undef VEC
> > > > > > +# undef VMOVA
> > > > > > +# undef VEC_SIZE
> > > > > > +#endif
> > > > > >
> > > > > > +#if MINIMUM_X86_ISA_LEVEL < AVX_X86_ISA_LEVEL
> > > > >
> > > > > Can this be <= __X86_ISA_LEVEL_V2?
> > > > > > /* movaps/movups is 1-byte shorter. */
> > > > > > -#define VEC_SIZE 16
> > > > > > -#define VMOVA movaps
> > > > > > -#define VEC(i) xmm##i
> > > > > > -#define _dl_runtime_profile _dl_runtime_profile_sse
> > > > > > -#undef RESTORE_AVX
> > > > > > -#include "dl-trampoline.h"
> > > > > > -#undef _dl_runtime_profile
> > > > > > -#undef VEC
> > > > > > -#undef VMOVA
> > > > > > -#undef VEC_SIZE
> > > > > > -
> > > > > > -#define USE_FXSAVE
> > > > > > -#define STATE_SAVE_ALIGNMENT 16
> > > > > > -#define _dl_runtime_resolve _dl_runtime_resolve_fxsave
> > > > > > -#include "dl-trampoline.h"
> > > > > > -#undef _dl_runtime_resolve
> > > > > > -#undef USE_FXSAVE
> > > > > > -#undef STATE_SAVE_ALIGNMENT
> > > > > > +# define VEC_SIZE 16
> > > > > > +# define VMOVA movaps
> > > > > > +# define VEC(i) xmm##i
> > > > > > +# define _dl_runtime_profile _dl_runtime_profile_sse
> > > > > > +# undef RESTORE_AVX
> > > > > > +# include "dl-trampoline.h"
> > > > > > +# undef _dl_runtime_profile
> > > > > > +# undef VEC
> > > > > > +# undef VMOVA
> > > > > > +# undef VEC_SIZE
> > > > > > +
> > > > > > +# define USE_FXSAVE
> > > > > > +# define STATE_SAVE_ALIGNMENT 16
> > > > > > +# define _dl_runtime_resolve _dl_runtime_resolve_fxsave
> > > > > > +# include "dl-trampoline.h"
> > > > > > +# undef _dl_runtime_resolve
> > > > > > +# undef USE_FXSAVE
> > > > > > +# undef STATE_SAVE_ALIGNMENT
> > > > > > +#endif
> > > > > >
> > > > > > #define USE_XSAVE
> > > > > > #define STATE_SAVE_ALIGNMENT 64
> > > > > > --
> > > > > > 2.36.1
> > > > > >
> > > > >
> > > > > There are cases of unconditional `movaps` usage in dl-trampoline.h (L222)
> > > > >
> > > > > This appears to be on ISA level v3 / v4 path.
> > > > >
> > > > > For ISA level >= 3 builds can those become `vmovdqa`?
> > > >
> > > > It may be replaced with VMOVA in a separate patch.
> > >
> > > Kk.
> > > >
> > > > --
> > > > H.J.
> >
> >
> >
> > --
> > H.J.
--
H.J.
More information about the Libc-alpha
mailing list