[PATCH 0/2] x86: Add thresholds for "rep movsb/stosb" to tunables

classic Classic list List threaded Threaded
14 messages Options
Reply | Threaded
Open this post in threaded view
|

[PATCH 0/2] x86: Add thresholds for "rep movsb/stosb" to tunables

Sourceware - libc-alpha mailing list
1. Add __tunable_update_val to update tunable min/max values and move x86
processor cache info to cpu_features.
2. Add x86_rep_movsb_threshold and x86_rep_stosb_threshold to tunables
to update thresholds for "rep movsb" and "rep stosb" at run-time.

H.J. Lu (2):
  Update tunable min/max values
  x86: Add thresholds for "rep movsb/stosb" to tunables

 elf/dl-tunables.c                             |  51 +-
 elf/dl-tunables.h                             |  15 +
 manual/tunables.texi                          |  14 +
 sysdeps/i386/cacheinfo.c                      |   3 -
 sysdeps/x86/Makefile                          |   2 +-
 sysdeps/x86/cacheinfo.c                       | 870 ++--------------
 sysdeps/x86/cpu-features.c                    |  19 +-
 sysdeps/x86/cpu-features.h                    |  30 +
 sysdeps/x86/dl-cacheinfo.c                    | 926 ++++++++++++++++++
 sysdeps/x86/dl-tunables.list                  |   6 +
 sysdeps/x86/init-arch.h                       |   3 +
 .../multiarch/memmove-vec-unaligned-erms.S    |  16 +-
 .../multiarch/memset-vec-unaligned-erms.S     |  12 +-
 13 files changed, 1107 insertions(+), 860 deletions(-)
 delete mode 100644 sysdeps/i386/cacheinfo.c
 create mode 100644 sysdeps/x86/dl-cacheinfo.c

--
2.26.2

Reply | Threaded
Open this post in threaded view
|

[PATCH 1/2] Update tunable min/max values

Sourceware - libc-alpha mailing list
Add __tunable_update_val to update tunable min/max values and move x86
processor cache info to cpu_features.
---
 elf/dl-tunables.c          |  51 ++-
 elf/dl-tunables.h          |  15 +
 sysdeps/i386/cacheinfo.c   |   3 -
 sysdeps/x86/Makefile       |   2 +-
 sysdeps/x86/cacheinfo.c    | 852 +++--------------------------------
 sysdeps/x86/cpu-features.c |  19 +-
 sysdeps/x86/cpu-features.h |  26 ++
 sysdeps/x86/dl-cacheinfo.c | 888 +++++++++++++++++++++++++++++++++++++
 sysdeps/x86/init-arch.h    |   3 +
 9 files changed, 1024 insertions(+), 835 deletions(-)
 delete mode 100644 sysdeps/i386/cacheinfo.c
 create mode 100644 sysdeps/x86/dl-cacheinfo.c

diff --git a/elf/dl-tunables.c b/elf/dl-tunables.c
index 26e6e26612..7c9f1ca31f 100644
--- a/elf/dl-tunables.c
+++ b/elf/dl-tunables.c
@@ -100,31 +100,39 @@ get_next_env (char **envp, char **name, size_t *namelen, char **val,
     }      \
 })
 
+#define TUNABLE_UPDATE_VAL(__cur, __val, __min, __max, __type)      \
+({      \
+  (__cur)->type.min = (__min);      \
+  (__cur)->type.max = (__max);      \
+  (__cur)->val.numval = (__val);      \
+  (__cur)->initialized = true;      \
+})
+
 static void
-do_tunable_update_val (tunable_t *cur, const void *valp)
+do_tunable_update_val (tunable_t *cur, const void *valp,
+       const void *minp, const void *maxp)
 {
-  uint64_t val;
+  uint64_t val, min, max;
 
   if (cur->type.type_code != TUNABLE_TYPE_STRING)
-    val = *((int64_t *) valp);
+    {
+      val = *((int64_t *) valp);
+      if (minp)
+ min = *((int64_t *) minp);
+      if (maxp)
+ max = *((int64_t *) maxp);
+    }
 
   switch (cur->type.type_code)
     {
     case TUNABLE_TYPE_INT_32:
- {
-  TUNABLE_SET_VAL_IF_VALID_RANGE (cur, val, int64_t);
-  break;
- }
     case TUNABLE_TYPE_UINT_64:
- {
-  TUNABLE_SET_VAL_IF_VALID_RANGE (cur, val, uint64_t);
-  break;
- }
     case TUNABLE_TYPE_SIZE_T:
- {
-  TUNABLE_SET_VAL_IF_VALID_RANGE (cur, val, uint64_t);
-  break;
- }
+      if (minp && maxp)
+ TUNABLE_UPDATE_VAL (cur, val, min, max, int64_t);
+      else
+ TUNABLE_SET_VAL_IF_VALID_RANGE (cur, val, int64_t);
+      break;
     case TUNABLE_TYPE_STRING:
  {
   cur->val.strval = valp;
@@ -153,7 +161,7 @@ tunable_initialize (tunable_t *cur, const char *strval)
       cur->initialized = true;
       valp = strval;
     }
-  do_tunable_update_val (cur, valp);
+  do_tunable_update_val (cur, valp, NULL, NULL);
 }
 
 void
@@ -161,8 +169,17 @@ __tunable_set_val (tunable_id_t id, void *valp)
 {
   tunable_t *cur = &tunable_list[id];
 
-  do_tunable_update_val (cur, valp);
+  do_tunable_update_val (cur, valp, NULL, NULL);
+}
+
+void
+__tunable_update_val (tunable_id_t id, void *valp, void *minp, void *maxp)
+{
+  tunable_t *cur = &tunable_list[id];
+
+  do_tunable_update_val (cur, valp, minp, maxp);
 }
+rtld_hidden_def (__tunable_update_val)
 
 #if TUNABLES_FRONTEND == TUNABLES_FRONTEND_valstring
 /* Parse the tunable string TUNESTR and adjust it to drop any tunables that may
diff --git a/elf/dl-tunables.h b/elf/dl-tunables.h
index f05eb50c2f..f6bf7379af 100644
--- a/elf/dl-tunables.h
+++ b/elf/dl-tunables.h
@@ -71,8 +71,10 @@ typedef struct _tunable tunable_t;
 extern void __tunables_init (char **);
 extern void __tunable_get_val (tunable_id_t, void *, tunable_callback_t);
 extern void __tunable_set_val (tunable_id_t, void *);
+extern void __tunable_update_val (tunable_id_t, void *, void *, void *);
 rtld_hidden_proto (__tunables_init)
 rtld_hidden_proto (__tunable_get_val)
+rtld_hidden_proto (__tunable_update_val)
 
 /* Define TUNABLE_GET and TUNABLE_SET in short form if TOP_NAMESPACE and
    TUNABLE_NAMESPACE are defined.  This is useful shorthand to get and set
@@ -82,11 +84,16 @@ rtld_hidden_proto (__tunable_get_val)
   TUNABLE_GET_FULL (TOP_NAMESPACE, TUNABLE_NAMESPACE, __id, __type, __cb)
 # define TUNABLE_SET(__id, __type, __val) \
   TUNABLE_SET_FULL (TOP_NAMESPACE, TUNABLE_NAMESPACE, __id, __type, __val)
+# define TUNABLE_UPDATE(__id, __type, __val, __min, __max) \
+  TUNABLE_UPDATE_FULL (TOP_NAMESPACE, TUNABLE_NAMESPACE, __id, __type, \
+       __val, __min, __max)
 #else
 # define TUNABLE_GET(__top, __ns, __id, __type, __cb) \
   TUNABLE_GET_FULL (__top, __ns, __id, __type, __cb)
 # define TUNABLE_SET(__top, __ns, __id, __type, __val) \
   TUNABLE_SET_FULL (__top, __ns, __id, __type, __val)
+# define TUNABLE_UPDATE(__top, __ns, __id, __type, __val, __min, __max) \
+  TUNABLE_UPDATE_FULL (__top, __ns, __id, __type, __val, __min, __max)
 #endif
 
 /* Get and return a tunable value.  If the tunable was set externally and __CB
@@ -106,6 +113,14 @@ rtld_hidden_proto (__tunable_get_val)
  & (__type) {__val});      \
 })
 
+/* Update a tunable value.  */
+# define TUNABLE_UPDATE_FULL(__top, __ns, __id, __type, __val, __min, __max) \
+({      \
+  __tunable_update_val (TUNABLE_ENUM_NAME (__top, __ns, __id),      \
+ & (__type) {__val},  & (__type) {__min},      \
+ & (__type) {__max});      \
+})
+
 /* Namespace sanity for callback functions.  Use this macro to keep the
    namespace of the modules clean.  */
 # define TUNABLE_CALLBACK(__name) _dl_tunable_ ## __name
diff --git a/sysdeps/i386/cacheinfo.c b/sysdeps/i386/cacheinfo.c
deleted file mode 100644
index f15fe0779a..0000000000
--- a/sysdeps/i386/cacheinfo.c
+++ /dev/null
@@ -1,3 +0,0 @@
-#define DISABLE_PREFETCHW
-
-#include <sysdeps/x86/cacheinfo.c>
diff --git a/sysdeps/x86/Makefile b/sysdeps/x86/Makefile
index beab426f67..0872e0e655 100644
--- a/sysdeps/x86/Makefile
+++ b/sysdeps/x86/Makefile
@@ -3,7 +3,7 @@ gen-as-const-headers += cpu-features-offsets.sym
 endif
 
 ifeq ($(subdir),elf)
-sysdep-dl-routines += dl-get-cpu-features
+sysdep-dl-routines += dl-get-cpu-features dl-cacheinfo
 
 tests += tst-get-cpu-features tst-get-cpu-features-static
 tests-static += tst-get-cpu-features-static
diff --git a/sysdeps/x86/cacheinfo.c b/sysdeps/x86/cacheinfo.c
index 311502dee3..8c4c7f9972 100644
--- a/sysdeps/x86/cacheinfo.c
+++ b/sysdeps/x86/cacheinfo.c
@@ -18,498 +18,9 @@
 
 #if IS_IN (libc)
 
-#include <assert.h>
-#include <stdbool.h>
-#include <stdlib.h>
 #include <unistd.h>
-#include <cpuid.h>
 #include <init-arch.h>
 
-static const struct intel_02_cache_info
-{
-  unsigned char idx;
-  unsigned char assoc;
-  unsigned char linesize;
-  unsigned char rel_name;
-  unsigned int size;
-} intel_02_known [] =
-  {
-#define M(sc) ((sc) - _SC_LEVEL1_ICACHE_SIZE)
-    { 0x06,  4, 32, M(_SC_LEVEL1_ICACHE_SIZE),    8192 },
-    { 0x08,  4, 32, M(_SC_LEVEL1_ICACHE_SIZE),   16384 },
-    { 0x09,  4, 32, M(_SC_LEVEL1_ICACHE_SIZE),   32768 },
-    { 0x0a,  2, 32, M(_SC_LEVEL1_DCACHE_SIZE),    8192 },
-    { 0x0c,  4, 32, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
-    { 0x0d,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
-    { 0x0e,  6, 64, M(_SC_LEVEL1_DCACHE_SIZE),   24576 },
-    { 0x21,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
-    { 0x22,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),   524288 },
-    { 0x23,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  1048576 },
-    { 0x25,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
-    { 0x29,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
-    { 0x2c,  8, 64, M(_SC_LEVEL1_DCACHE_SIZE),   32768 },
-    { 0x30,  8, 64, M(_SC_LEVEL1_ICACHE_SIZE),   32768 },
-    { 0x39,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
-    { 0x3a,  6, 64, M(_SC_LEVEL2_CACHE_SIZE),   196608 },
-    { 0x3b,  2, 64, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
-    { 0x3c,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
-    { 0x3d,  6, 64, M(_SC_LEVEL2_CACHE_SIZE),   393216 },
-    { 0x3e,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
-    { 0x3f,  2, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
-    { 0x41,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
-    { 0x42,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
-    { 0x43,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
-    { 0x44,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
-    { 0x45,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),  2097152 },
-    { 0x46,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
-    { 0x47,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
-    { 0x48, 12, 64, M(_SC_LEVEL2_CACHE_SIZE),  3145728 },
-    { 0x49, 16, 64, M(_SC_LEVEL2_CACHE_SIZE),  4194304 },
-    { 0x4a, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  6291456 },
-    { 0x4b, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
-    { 0x4c, 12, 64, M(_SC_LEVEL3_CACHE_SIZE), 12582912 },
-    { 0x4d, 16, 64, M(_SC_LEVEL3_CACHE_SIZE), 16777216 },
-    { 0x4e, 24, 64, M(_SC_LEVEL2_CACHE_SIZE),  6291456 },
-    { 0x60,  8, 64, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
-    { 0x66,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),    8192 },
-    { 0x67,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
-    { 0x68,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),   32768 },
-    { 0x78,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
-    { 0x79,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
-    { 0x7a,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
-    { 0x7b,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
-    { 0x7c,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
-    { 0x7d,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  2097152 },
-    { 0x7f,  2, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
-    { 0x80,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
-    { 0x82,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
-    { 0x83,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
-    { 0x84,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
-    { 0x85,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),  2097152 },
-    { 0x86,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
-    { 0x87,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
-    { 0xd0,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),   524288 },
-    { 0xd1,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),  1048576 },
-    { 0xd2,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
-    { 0xd6,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  1048576 },
-    { 0xd7,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
-    { 0xd8,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
-    { 0xdc, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
-    { 0xdd, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
-    { 0xde, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
-    { 0xe2, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
-    { 0xe3, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
-    { 0xe4, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
-    { 0xea, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 12582912 },
-    { 0xeb, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 18874368 },
-    { 0xec, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 25165824 },
-  };
-
-#define nintel_02_known (sizeof (intel_02_known) / sizeof (intel_02_known [0]))
-
-static int
-intel_02_known_compare (const void *p1, const void *p2)
-{
-  const struct intel_02_cache_info *i1;
-  const struct intel_02_cache_info *i2;
-
-  i1 = (const struct intel_02_cache_info *) p1;
-  i2 = (const struct intel_02_cache_info *) p2;
-
-  if (i1->idx == i2->idx)
-    return 0;
-
-  return i1->idx < i2->idx ? -1 : 1;
-}
-
-
-static long int
-__attribute__ ((noinline))
-intel_check_word (int name, unsigned int value, bool *has_level_2,
-  bool *no_level_2_or_3,
-  const struct cpu_features *cpu_features)
-{
-  if ((value & 0x80000000) != 0)
-    /* The register value is reserved.  */
-    return 0;
-
-  /* Fold the name.  The _SC_ constants are always in the order SIZE,
-     ASSOC, LINESIZE.  */
-  int folded_rel_name = (M(name) / 3) * 3;
-
-  while (value != 0)
-    {
-      unsigned int byte = value & 0xff;
-
-      if (byte == 0x40)
- {
-  *no_level_2_or_3 = true;
-
-  if (folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))
-    /* No need to look further.  */
-    break;
- }
-      else if (byte == 0xff)
- {
-  /* CPUID leaf 0x4 contains all the information.  We need to
-     iterate over it.  */
-  unsigned int eax;
-  unsigned int ebx;
-  unsigned int ecx;
-  unsigned int edx;
-
-  unsigned int round = 0;
-  while (1)
-    {
-      __cpuid_count (4, round, eax, ebx, ecx, edx);
-
-      enum { null = 0, data = 1, inst = 2, uni = 3 } type = eax & 0x1f;
-      if (type == null)
- /* That was the end.  */
- break;
-
-      unsigned int level = (eax >> 5) & 0x7;
-
-      if ((level == 1 && type == data
-   && folded_rel_name == M(_SC_LEVEL1_DCACHE_SIZE))
-  || (level == 1 && type == inst
-      && folded_rel_name == M(_SC_LEVEL1_ICACHE_SIZE))
-  || (level == 2 && folded_rel_name == M(_SC_LEVEL2_CACHE_SIZE))
-  || (level == 3 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))
-  || (level == 4 && folded_rel_name == M(_SC_LEVEL4_CACHE_SIZE)))
- {
-  unsigned int offset = M(name) - folded_rel_name;
-
-  if (offset == 0)
-    /* Cache size.  */
-    return (((ebx >> 22) + 1)
-    * (((ebx >> 12) & 0x3ff) + 1)
-    * ((ebx & 0xfff) + 1)
-    * (ecx + 1));
-  if (offset == 1)
-    return (ebx >> 22) + 1;
-
-  assert (offset == 2);
-  return (ebx & 0xfff) + 1;
- }
-
-      ++round;
-    }
-  /* There is no other cache information anywhere else.  */
-  break;
- }
-      else
- {
-  if (byte == 0x49 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))
-    {
-      /* Intel reused this value.  For family 15, model 6 it
- specifies the 3rd level cache.  Otherwise the 2nd
- level cache.  */
-      unsigned int family = cpu_features->basic.family;
-      unsigned int model = cpu_features->basic.model;
-
-      if (family == 15 && model == 6)
- {
-  /* The level 3 cache is encoded for this model like
-     the level 2 cache is for other models.  Pretend
-     the caller asked for the level 2 cache.  */
-  name = (_SC_LEVEL2_CACHE_SIZE
-  + (name - _SC_LEVEL3_CACHE_SIZE));
-  folded_rel_name = M(_SC_LEVEL2_CACHE_SIZE);
- }
-    }
-
-  struct intel_02_cache_info *found;
-  struct intel_02_cache_info search;
-
-  search.idx = byte;
-  found = bsearch (&search, intel_02_known, nintel_02_known,
-   sizeof (intel_02_known[0]), intel_02_known_compare);
-  if (found != NULL)
-    {
-      if (found->rel_name == folded_rel_name)
- {
-  unsigned int offset = M(name) - folded_rel_name;
-
-  if (offset == 0)
-    /* Cache size.  */
-    return found->size;
-  if (offset == 1)
-    return found->assoc;
-
-  assert (offset == 2);
-  return found->linesize;
- }
-
-      if (found->rel_name == M(_SC_LEVEL2_CACHE_SIZE))
- *has_level_2 = true;
-    }
- }
-
-      /* Next byte for the next round.  */
-      value >>= 8;
-    }
-
-  /* Nothing found.  */
-  return 0;
-}
-
-
-static long int __attribute__ ((noinline))
-handle_intel (int name, const struct cpu_features *cpu_features)
-{
-  unsigned int maxidx = cpu_features->basic.max_cpuid;
-
-  /* Return -1 for older CPUs.  */
-  if (maxidx < 2)
-    return -1;
-
-  /* OK, we can use the CPUID instruction to get all info about the
-     caches.  */
-  unsigned int cnt = 0;
-  unsigned int max = 1;
-  long int result = 0;
-  bool no_level_2_or_3 = false;
-  bool has_level_2 = false;
-
-  while (cnt++ < max)
-    {
-      unsigned int eax;
-      unsigned int ebx;
-      unsigned int ecx;
-      unsigned int edx;
-      __cpuid (2, eax, ebx, ecx, edx);
-
-      /* The low byte of EAX in the first round contain the number of
- rounds we have to make.  At least one, the one we are already
- doing.  */
-      if (cnt == 1)
- {
-  max = eax & 0xff;
-  eax &= 0xffffff00;
- }
-
-      /* Process the individual registers' value.  */
-      result = intel_check_word (name, eax, &has_level_2,
- &no_level_2_or_3, cpu_features);
-      if (result != 0)
- return result;
-
-      result = intel_check_word (name, ebx, &has_level_2,
- &no_level_2_or_3, cpu_features);
-      if (result != 0)
- return result;
-
-      result = intel_check_word (name, ecx, &has_level_2,
- &no_level_2_or_3, cpu_features);
-      if (result != 0)
- return result;
-
-      result = intel_check_word (name, edx, &has_level_2,
- &no_level_2_or_3, cpu_features);
-      if (result != 0)
- return result;
-    }
-
-  if (name >= _SC_LEVEL2_CACHE_SIZE && name <= _SC_LEVEL3_CACHE_LINESIZE
-      && no_level_2_or_3)
-    return -1;
-
-  return 0;
-}
-
-
-static long int __attribute__ ((noinline))
-handle_amd (int name)
-{
-  unsigned int eax;
-  unsigned int ebx;
-  unsigned int ecx;
-  unsigned int edx;
-  __cpuid (0x80000000, eax, ebx, ecx, edx);
-
-  /* No level 4 cache (yet).  */
-  if (name > _SC_LEVEL3_CACHE_LINESIZE)
-    return 0;
-
-  unsigned int fn = 0x80000005 + (name >= _SC_LEVEL2_CACHE_SIZE);
-  if (eax < fn)
-    return 0;
-
-  __cpuid (fn, eax, ebx, ecx, edx);
-
-  if (name < _SC_LEVEL1_DCACHE_SIZE)
-    {
-      name += _SC_LEVEL1_DCACHE_SIZE - _SC_LEVEL1_ICACHE_SIZE;
-      ecx = edx;
-    }
-
-  switch (name)
-    {
-    case _SC_LEVEL1_DCACHE_SIZE:
-      return (ecx >> 14) & 0x3fc00;
-
-    case _SC_LEVEL1_DCACHE_ASSOC:
-      ecx >>= 16;
-      if ((ecx & 0xff) == 0xff)
- /* Fully associative.  */
- return (ecx << 2) & 0x3fc00;
-      return ecx & 0xff;
-
-    case _SC_LEVEL1_DCACHE_LINESIZE:
-      return ecx & 0xff;
-
-    case _SC_LEVEL2_CACHE_SIZE:
-      return (ecx & 0xf000) == 0 ? 0 : (ecx >> 6) & 0x3fffc00;
-
-    case _SC_LEVEL2_CACHE_ASSOC:
-      switch ((ecx >> 12) & 0xf)
- {
- case 0:
- case 1:
- case 2:
- case 4:
-  return (ecx >> 12) & 0xf;
- case 6:
-  return 8;
- case 8:
-  return 16;
- case 10:
-  return 32;
- case 11:
-  return 48;
- case 12:
-  return 64;
- case 13:
-  return 96;
- case 14:
-  return 128;
- case 15:
-  return ((ecx >> 6) & 0x3fffc00) / (ecx & 0xff);
- default:
-  return 0;
- }
-      /* NOTREACHED */
-
-    case _SC_LEVEL2_CACHE_LINESIZE:
-      return (ecx & 0xf000) == 0 ? 0 : ecx & 0xff;
-
-    case _SC_LEVEL3_CACHE_SIZE:
-      return (edx & 0xf000) == 0 ? 0 : (edx & 0x3ffc0000) << 1;
-
-    case _SC_LEVEL3_CACHE_ASSOC:
-      switch ((edx >> 12) & 0xf)
- {
- case 0:
- case 1:
- case 2:
- case 4:
-  return (edx >> 12) & 0xf;
- case 6:
-  return 8;
- case 8:
-  return 16;
- case 10:
-  return 32;
- case 11:
-  return 48;
- case 12:
-  return 64;
- case 13:
-  return 96;
- case 14:
-  return 128;
- case 15:
-  return ((edx & 0x3ffc0000) << 1) / (edx & 0xff);
- default:
-  return 0;
- }
-      /* NOTREACHED */
-
-    case _SC_LEVEL3_CACHE_LINESIZE:
-      return (edx & 0xf000) == 0 ? 0 : edx & 0xff;
-
-    default:
-      assert (! "cannot happen");
-    }
-  return -1;
-}
-
-
-static long int __attribute__ ((noinline))
-handle_zhaoxin (int name)
-{
-  unsigned int eax;
-  unsigned int ebx;
-  unsigned int ecx;
-  unsigned int edx;
-
-  int folded_rel_name = (M(name) / 3) * 3;
-
-  unsigned int round = 0;
-  while (1)
-    {
-      __cpuid_count (4, round, eax, ebx, ecx, edx);
-
-      enum { null = 0, data = 1, inst = 2, uni = 3 } type = eax & 0x1f;
-      if (type == null)
-        break;
-
-      unsigned int level = (eax >> 5) & 0x7;
-
-      if ((level == 1 && type == data
-        && folded_rel_name == M(_SC_LEVEL1_DCACHE_SIZE))
-        || (level == 1 && type == inst
-            && folded_rel_name == M(_SC_LEVEL1_ICACHE_SIZE))
-        || (level == 2 && folded_rel_name == M(_SC_LEVEL2_CACHE_SIZE))
-        || (level == 3 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE)))
-        {
-          unsigned int offset = M(name) - folded_rel_name;
-
-          if (offset == 0)
-            /* Cache size.  */
-            return (((ebx >> 22) + 1)
-                * (((ebx >> 12) & 0x3ff) + 1)
-                * ((ebx & 0xfff) + 1)
-                * (ecx + 1));
-          if (offset == 1)
-            return (ebx >> 22) + 1;
-
-          assert (offset == 2);
-          return (ebx & 0xfff) + 1;
-        }
-
-      ++round;
-    }
-
-  /* Nothing found.  */
-  return 0;
-}
-
-
-/* Get the value of the system variable NAME.  */
-long int
-attribute_hidden
-__cache_sysconf (int name)
-{
-  const struct cpu_features *cpu_features = __get_cpu_features ();
-
-  if (cpu_features->basic.kind == arch_kind_intel)
-    return handle_intel (name, cpu_features);
-
-  if (cpu_features->basic.kind == arch_kind_amd)
-    return handle_amd (name);
-
-  if (cpu_features->basic.kind == arch_kind_zhaoxin)
-    return handle_zhaoxin (name);
-
-  // XXX Fill in more vendors.
-
-  /* CPU not known, we have no information.  */
-  return 0;
-}
-
-
 /* Data cache size for use in memory and string routines, typically
    L1 size, rounded to multiple of 256 bytes.  */
 long int __x86_data_cache_size_half attribute_hidden = 32 * 1024 / 2;
@@ -530,348 +41,85 @@ long int __x86_raw_shared_cache_size attribute_hidden = 1024 * 1024;
 /* Threshold to use non temporal store.  */
 long int __x86_shared_non_temporal_threshold attribute_hidden;
 
-#ifndef DISABLE_PREFETCHW
+#ifndef __x86_64__
 /* PREFETCHW support flag for use in memory and string routines.  */
 int __x86_prefetchw attribute_hidden;
 #endif
 
-
-static void
-get_common_cache_info (long int *shared_ptr, unsigned int *threads_ptr,
-                long int core)
+/* Get the value of the system variable NAME.  */
+long int
+attribute_hidden
+__cache_sysconf (int name)
 {
-  unsigned int eax;
-  unsigned int ebx;
-  unsigned int ecx;
-  unsigned int edx;
-
-  /* Number of logical processors sharing L2 cache.  */
-  int threads_l2;
-
-  /* Number of logical processors sharing L3 cache.  */
-  int threads_l3;
-
   const struct cpu_features *cpu_features = __get_cpu_features ();
-  int max_cpuid = cpu_features->basic.max_cpuid;
-  unsigned int family = cpu_features->basic.family;
-  unsigned int model = cpu_features->basic.model;
-  long int shared = *shared_ptr;
-  unsigned int threads = *threads_ptr;
-  bool inclusive_cache = true;
-  bool support_count_mask = true;
-
-  /* Try L3 first.  */
-  unsigned int level = 3;
-
-  if (cpu_features->basic.kind == arch_kind_zhaoxin && family == 6)
-    support_count_mask = false;
-
-  if (shared <= 0)
-    {
-      /* Try L2 otherwise.  */
-      level  = 2;
-      shared = core;
-      threads_l2 = 0;
-      threads_l3 = -1;
-    }
-  else
-    {
-      threads_l2 = 0;
-      threads_l3 = 0;
-    }
-
-  /* A value of 0 for the HTT bit indicates there is only a single
-     logical processor.  */
-  if (HAS_CPU_FEATURE (HTT))
+  switch (name)
     {
-      /* Figure out the number of logical threads that share the
-         highest cache level.  */
-      if (max_cpuid >= 4)
-        {
-          int i = 0;
-
-          /* Query until cache level 2 and 3 are enumerated.  */
-          int check = 0x1 | (threads_l3 == 0) << 1;
-          do
-            {
-              __cpuid_count (4, i++, eax, ebx, ecx, edx);
+    case _SC_LEVEL1_ICACHE_SIZE:
+      return cpu_features->level1_icache_size;
 
-              /* There seems to be a bug in at least some Pentium Ds
-                 which sometimes fail to iterate all cache parameters.
-                 Do not loop indefinitely here, stop in this case and
-                 assume there is no such information.  */
-              if (cpu_features->basic.kind == arch_kind_intel
-                  && (eax & 0x1f) == 0 )
-                goto intel_bug_no_cache_info;
+    case _SC_LEVEL1_DCACHE_SIZE:
+      return cpu_features->level1_dcache_size;
 
-              switch ((eax >> 5) & 0x7)
-                {
-                  default:
-                    break;
-                  case 2:
-                    if ((check & 0x1))
-                      {
-                        /* Get maximum number of logical processors
-                           sharing L2 cache.  */
-                        threads_l2 = (eax >> 14) & 0x3ff;
-                        check &= ~0x1;
-                      }
-                    break;
-                  case 3:
-                    if ((check & (0x1 << 1)))
-                      {
-                        /* Get maximum number of logical processors
-                           sharing L3 cache.  */
-                        threads_l3 = (eax >> 14) & 0x3ff;
+    case _SC_LEVEL1_DCACHE_ASSOC:
+      return cpu_features->level1_dcache_assoc;
 
-                        /* Check if L2 and L3 caches are inclusive.  */
-                        inclusive_cache = (edx & 0x2) != 0;
-                        check &= ~(0x1 << 1);
-                      }
-                    break;
-                }
-            }
-          while (check);
+    case _SC_LEVEL1_DCACHE_LINESIZE:
+      return cpu_features->level1_dcache_linesize;
 
-          /* If max_cpuid >= 11, THREADS_L2/THREADS_L3 are the maximum
-             numbers of addressable IDs for logical processors sharing
-             the cache, instead of the maximum number of threads
-             sharing the cache.  */
-          if (max_cpuid >= 11 && support_count_mask)
-            {
-              /* Find the number of logical processors shipped in
-                 one core and apply count mask.  */
-              i = 0;
+    case _SC_LEVEL2_CACHE_SIZE:
+      return cpu_features->level2_cache_size;
 
-              /* Count SMT only if there is L3 cache.  Always count
-                 core if there is no L3 cache.  */
-              int count = ((threads_l2 > 0 && level == 3)
-                           | ((threads_l3 > 0
-                               || (threads_l2 > 0 && level == 2)) << 1));
+    case _SC_LEVEL2_CACHE_ASSOC:
+      return cpu_features->level2_cache_assoc;
 
-              while (count)
-                {
-                  __cpuid_count (11, i++, eax, ebx, ecx, edx);
+    case _SC_LEVEL2_CACHE_LINESIZE:
+      return cpu_features->level2_cache_linesize;
 
-                  int shipped = ebx & 0xff;
-                  int type = ecx & 0xff00;
-                  if (shipped == 0 || type == 0)
-                    break;
-                  else if (type == 0x100)
-                    {
-                      /* Count SMT.  */
-                      if ((count & 0x1))
-                        {
-                          int count_mask;
+    case _SC_LEVEL3_CACHE_SIZE:
+      return cpu_features->level3_cache_size;
 
-                          /* Compute count mask.  */
-                          asm ("bsr %1, %0"
-                               : "=r" (count_mask) : "g" (threads_l2));
-                          count_mask = ~(-1 << (count_mask + 1));
-                          threads_l2 = (shipped - 1) & count_mask;
-                          count &= ~0x1;
-                        }
-                    }
-                  else if (type == 0x200)
-                    {
-                      /* Count core.  */
-                      if ((count & (0x1 << 1)))
-                        {
-                          int count_mask;
-                          int threads_core
-                            = (level == 2 ? threads_l2 : threads_l3);
+    case _SC_LEVEL3_CACHE_ASSOC:
+      return cpu_features->level3_cache_assoc;
 
-                          /* Compute count mask.  */
-                          asm ("bsr %1, %0"
-                               : "=r" (count_mask) : "g" (threads_core));
-                          count_mask = ~(-1 << (count_mask + 1));
-                          threads_core = (shipped - 1) & count_mask;
-                          if (level == 2)
-                            threads_l2 = threads_core;
-                          else
-                            threads_l3 = threads_core;
-                          count &= ~(0x1 << 1);
-                        }
-                    }
-                }
-            }
-          if (threads_l2 > 0)
-            threads_l2 += 1;
-          if (threads_l3 > 0)
-            threads_l3 += 1;
-          if (level == 2)
-            {
-              if (threads_l2)
-                {
-                  threads = threads_l2;
-                  if (cpu_features->basic.kind == arch_kind_intel
-                      && threads > 2
-                      && family == 6)
-                    switch (model)
-                      {
-                        case 0x37:
-                        case 0x4a:
-                        case 0x4d:
-                        case 0x5a:
-                        case 0x5d:
-                          /* Silvermont has L2 cache shared by 2 cores.  */
-                          threads = 2;
-                          break;
-                        default:
-                          break;
-                      }
-                }
-            }
-          else if (threads_l3)
-            threads = threads_l3;
-        }
-      else
-        {
-intel_bug_no_cache_info:
-          /* Assume that all logical threads share the highest cache
-             level.  */
-          threads
-            = ((cpu_features->cpuid[COMMON_CPUID_INDEX_1].ebx
-                >> 16) & 0xff);
-        }
+    case _SC_LEVEL3_CACHE_LINESIZE:
+      return cpu_features->level3_cache_linesize;
 
-        /* Cap usage of highest cache level to the number of supported
-           threads.  */
-        if (shared > 0 && threads > 0)
-          shared /= threads;
-    }
+    case _SC_LEVEL4_CACHE_SIZE:
+      return cpu_features->level4_cache_size;
 
-  /* Account for non-inclusive L2 and L3 caches.  */
-  if (!inclusive_cache)
-    {
-      if (threads_l2 > 0)
-        core /= threads_l2;
-      shared += core;
+    default:
+      break;
     }
-
-  *shared_ptr = shared;
-  *threads_ptr = threads;
+  return -1;
 }
 
-
 static void
 __attribute__((constructor))
 init_cacheinfo (void)
 {
-  /* Find out what brand of processor.  */
-  unsigned int ebx;
-  unsigned int ecx;
-  unsigned int edx;
-  int max_cpuid_ex;
-  long int data = -1;
-  long int shared = -1;
-  long int core;
-  unsigned int threads = 0;
   const struct cpu_features *cpu_features = __get_cpu_features ();
+  long int data = cpu_features->data_cache_size;
+  __x86_raw_data_cache_size_half = data / 2;
+  __x86_raw_data_cache_size = data;
+  /* Round data cache size to multiple of 256 bytes.  */
+  data = data & ~255L;
+  __x86_data_cache_size_half = data / 2;
+  __x86_data_cache_size = data;
+
+  long int shared = cpu_features->shared_cache_size;
+  __x86_raw_shared_cache_size_half = shared / 2;
+  __x86_raw_shared_cache_size = shared;
+  /* Round shared cache size to multiple of 256 bytes.  */
+  shared = shared & ~255L;
+  __x86_shared_cache_size_half = shared / 2;
+  __x86_shared_cache_size = shared;
 
-  if (cpu_features->basic.kind == arch_kind_intel)
-    {
-      data = handle_intel (_SC_LEVEL1_DCACHE_SIZE, cpu_features);
-      core = handle_intel (_SC_LEVEL2_CACHE_SIZE, cpu_features);
-      shared = handle_intel (_SC_LEVEL3_CACHE_SIZE, cpu_features);
-
-      get_common_cache_info (&shared, &threads, core);
-    }
-  else if (cpu_features->basic.kind == arch_kind_zhaoxin)
-    {
-      data = handle_zhaoxin (_SC_LEVEL1_DCACHE_SIZE);
-      core = handle_zhaoxin (_SC_LEVEL2_CACHE_SIZE);
-      shared = handle_zhaoxin (_SC_LEVEL3_CACHE_SIZE);
-
-      get_common_cache_info (&shared, &threads, core);
-    }
-  else if (cpu_features->basic.kind == arch_kind_amd)
-    {
-      data   = handle_amd (_SC_LEVEL1_DCACHE_SIZE);
-      long int core = handle_amd (_SC_LEVEL2_CACHE_SIZE);
-      shared = handle_amd (_SC_LEVEL3_CACHE_SIZE);
-
-      /* Get maximum extended function. */
-      __cpuid (0x80000000, max_cpuid_ex, ebx, ecx, edx);
-
-      if (shared <= 0)
- /* No shared L3 cache.  All we have is the L2 cache.  */
- shared = core;
-      else
- {
-  /* Figure out the number of logical threads that share L3.  */
-  if (max_cpuid_ex >= 0x80000008)
-    {
-      /* Get width of APIC ID.  */
-      __cpuid (0x80000008, max_cpuid_ex, ebx, ecx, edx);
-      threads = 1 << ((ecx >> 12) & 0x0f);
-    }
-
-  if (threads == 0)
-    {
-      /* If APIC ID width is not available, use logical
- processor count.  */
-      __cpuid (0x00000001, max_cpuid_ex, ebx, ecx, edx);
-
-      if ((edx & (1 << 28)) != 0)
- threads = (ebx >> 16) & 0xff;
-    }
-
-  /* Cap usage of highest cache level to the number of
-     supported threads.  */
-  if (threads > 0)
-    shared /= threads;
-
-  /* Account for exclusive L2 and L3 caches.  */
-  shared += core;
- }
+  __x86_shared_non_temporal_threshold
+    = cpu_features->non_temporal_threshold;
 
-#ifndef DISABLE_PREFETCHW
-      if (max_cpuid_ex >= 0x80000001)
- {
-  unsigned int eax;
-  __cpuid (0x80000001, eax, ebx, ecx, edx);
-  /*  PREFETCHW     || 3DNow!  */
-  if ((ecx & 0x100) || (edx & 0x80000000))
-    __x86_prefetchw = -1;
- }
+#ifndef __x86_64__
+  __x86_prefetchw = cpu_features->prefetchw;
 #endif
-    }
-
-  if (cpu_features->data_cache_size != 0)
-    data = cpu_features->data_cache_size;
-
-  if (data > 0)
-    {
-      __x86_raw_data_cache_size_half = data / 2;
-      __x86_raw_data_cache_size = data;
-      /* Round data cache size to multiple of 256 bytes.  */
-      data = data & ~255L;
-      __x86_data_cache_size_half = data / 2;
-      __x86_data_cache_size = data;
-    }
-
-  if (cpu_features->shared_cache_size != 0)
-    shared = cpu_features->shared_cache_size;
-
-  if (shared > 0)
-    {
-      __x86_raw_shared_cache_size_half = shared / 2;
-      __x86_raw_shared_cache_size = shared;
-      /* Round shared cache size to multiple of 256 bytes.  */
-      shared = shared & ~255L;
-      __x86_shared_cache_size_half = shared / 2;
-      __x86_shared_cache_size = shared;
-    }
-
-  /* The large memcpy micro benchmark in glibc shows that 6 times of
-     shared cache size is the approximate value above which non-temporal
-     store becomes faster on a 8-core processor.  This is the 3/4 of the
-     total shared cache size.  */
-  __x86_shared_non_temporal_threshold
-    = (cpu_features->non_temporal_threshold != 0
-       ? cpu_features->non_temporal_threshold
-       : __x86_shared_cache_size * threads * 3 / 4);
 }
 
 #endif
diff --git a/sysdeps/x86/cpu-features.c b/sysdeps/x86/cpu-features.c
index c351bdd54a..e718204c18 100644
--- a/sysdeps/x86/cpu-features.c
+++ b/sysdeps/x86/cpu-features.c
@@ -19,6 +19,7 @@
 #include <cpuid.h>
 #include <cpu-features.h>
 #include <dl-hwcap.h>
+#include <init-arch.h>
 #include <libc-pointer-arith.h>
 
 #if HAVE_TUNABLES
@@ -602,20 +603,14 @@ no_cpuid:
   cpu_features->basic.model = model;
   cpu_features->basic.stepping = stepping;
 
+  __init_cacheinfo ();
+
 #if HAVE_TUNABLES
   TUNABLE_GET (hwcaps, tunable_val_t *, TUNABLE_CALLBACK (set_hwcaps));
-  cpu_features->non_temporal_threshold
-    = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
-  cpu_features->data_cache_size
-    = TUNABLE_GET (x86_data_cache_size, long int, NULL);
-  cpu_features->shared_cache_size
-    = TUNABLE_GET (x86_shared_cache_size, long int, NULL);
-#endif
-
-  /* Reuse dl_platform, dl_hwcap and dl_hwcap_mask for x86.  */
-#if !HAVE_TUNABLES && defined SHARED
-  /* The glibc.cpu.hwcap_mask tunable is initialized already, so no need to do
-     this.  */
+#elif defined SHARED
+  /* Reuse dl_platform, dl_hwcap and dl_hwcap_mask for x86.  The
+     glibc.cpu.hwcap_mask tunable is initialized already, so no
+     need to do this.  */
   GLRO(dl_hwcap_mask) = HWCAP_IMPORTANT;
 #endif
 
diff --git a/sysdeps/x86/cpu-features.h b/sysdeps/x86/cpu-features.h
index d66dc206f7..3aaed33cbc 100644
--- a/sysdeps/x86/cpu-features.h
+++ b/sysdeps/x86/cpu-features.h
@@ -102,6 +102,32 @@ struct cpu_features
   unsigned long int shared_cache_size;
   /* Threshold to use non temporal store.  */
   unsigned long int non_temporal_threshold;
+  /* _SC_LEVEL1_ICACHE_SIZE.  */
+  unsigned long int level1_icache_size;
+  /* _SC_LEVEL1_DCACHE_SIZE.  */
+  unsigned long int level1_dcache_size;
+  /* _SC_LEVEL1_DCACHE_ASSOC.  */
+  unsigned long int level1_dcache_assoc;
+  /* _SC_LEVEL1_DCACHE_LINESIZE.  */
+  unsigned long int level1_dcache_linesize;
+  /* _SC_LEVEL2_CACHE_ASSOC.  */
+  unsigned long int level2_cache_size;
+  /* _SC_LEVEL2_DCACHE_ASSOC.  */
+  unsigned long int level2_cache_assoc;
+  /* _SC_LEVEL2_CACHE_LINESIZE.  */
+  unsigned long int level2_cache_linesize;
+  /* /_SC_LEVEL3_CACHE_SIZE.  */
+  unsigned long int level3_cache_size;
+  /* _SC_LEVEL3_CACHE_ASSOC.  */
+  unsigned long int level3_cache_assoc;
+  /* _SC_LEVEL3_CACHE_LINESIZE.  */
+  unsigned long int level3_cache_linesize;
+  /* /_SC_LEVEL4_CACHE_SIZE.  */
+  unsigned long int level4_cache_size;
+#ifndef __x86_64__
+  /* PREFETCHW support flag for use in memory and string routines.  */
+  unsigned long int prefetchw;
+#endif
 };
 
 /* Used from outside of glibc to get access to the CPU features
diff --git a/sysdeps/x86/dl-cacheinfo.c b/sysdeps/x86/dl-cacheinfo.c
new file mode 100644
index 0000000000..8e2a6f552c
--- /dev/null
+++ b/sysdeps/x86/dl-cacheinfo.c
@@ -0,0 +1,888 @@
+/* x86 cache info.
+   Copyright (C) 2020 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
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include <assert.h>
+#include <stdbool.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <cpuid.h>
+#include <init-arch.h>
+#if HAVE_TUNABLES
+# define TUNABLE_NAMESPACE cpu
+# include <elf/dl-tunables.h>
+#endif
+
+static const struct intel_02_cache_info
+{
+  unsigned char idx;
+  unsigned char assoc;
+  unsigned char linesize;
+  unsigned char rel_name;
+  unsigned int size;
+} intel_02_known [] =
+  {
+#define M(sc) ((sc) - _SC_LEVEL1_ICACHE_SIZE)
+    { 0x06,  4, 32, M(_SC_LEVEL1_ICACHE_SIZE),    8192 },
+    { 0x08,  4, 32, M(_SC_LEVEL1_ICACHE_SIZE),   16384 },
+    { 0x09,  4, 32, M(_SC_LEVEL1_ICACHE_SIZE),   32768 },
+    { 0x0a,  2, 32, M(_SC_LEVEL1_DCACHE_SIZE),    8192 },
+    { 0x0c,  4, 32, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
+    { 0x0d,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
+    { 0x0e,  6, 64, M(_SC_LEVEL1_DCACHE_SIZE),   24576 },
+    { 0x21,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
+    { 0x22,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),   524288 },
+    { 0x23,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  1048576 },
+    { 0x25,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
+    { 0x29,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
+    { 0x2c,  8, 64, M(_SC_LEVEL1_DCACHE_SIZE),   32768 },
+    { 0x30,  8, 64, M(_SC_LEVEL1_ICACHE_SIZE),   32768 },
+    { 0x39,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
+    { 0x3a,  6, 64, M(_SC_LEVEL2_CACHE_SIZE),   196608 },
+    { 0x3b,  2, 64, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
+    { 0x3c,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
+    { 0x3d,  6, 64, M(_SC_LEVEL2_CACHE_SIZE),   393216 },
+    { 0x3e,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
+    { 0x3f,  2, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
+    { 0x41,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
+    { 0x42,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
+    { 0x43,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
+    { 0x44,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
+    { 0x45,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),  2097152 },
+    { 0x46,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
+    { 0x47,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
+    { 0x48, 12, 64, M(_SC_LEVEL2_CACHE_SIZE),  3145728 },
+    { 0x49, 16, 64, M(_SC_LEVEL2_CACHE_SIZE),  4194304 },
+    { 0x4a, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  6291456 },
+    { 0x4b, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
+    { 0x4c, 12, 64, M(_SC_LEVEL3_CACHE_SIZE), 12582912 },
+    { 0x4d, 16, 64, M(_SC_LEVEL3_CACHE_SIZE), 16777216 },
+    { 0x4e, 24, 64, M(_SC_LEVEL2_CACHE_SIZE),  6291456 },
+    { 0x60,  8, 64, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
+    { 0x66,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),    8192 },
+    { 0x67,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
+    { 0x68,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),   32768 },
+    { 0x78,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
+    { 0x79,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
+    { 0x7a,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
+    { 0x7b,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
+    { 0x7c,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
+    { 0x7d,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  2097152 },
+    { 0x7f,  2, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
+    { 0x80,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
+    { 0x82,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
+    { 0x83,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
+    { 0x84,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
+    { 0x85,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),  2097152 },
+    { 0x86,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
+    { 0x87,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
+    { 0xd0,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),   524288 },
+    { 0xd1,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),  1048576 },
+    { 0xd2,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
+    { 0xd6,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  1048576 },
+    { 0xd7,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
+    { 0xd8,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
+    { 0xdc, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
+    { 0xdd, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
+    { 0xde, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
+    { 0xe2, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
+    { 0xe3, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
+    { 0xe4, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
+    { 0xea, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 12582912 },
+    { 0xeb, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 18874368 },
+    { 0xec, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 25165824 },
+  };
+
+#define nintel_02_known (sizeof (intel_02_known) / sizeof (intel_02_known [0]))
+
+static int
+intel_02_known_compare (const void *p1, const void *p2)
+{
+  const struct intel_02_cache_info *i1;
+  const struct intel_02_cache_info *i2;
+
+  i1 = (const struct intel_02_cache_info *) p1;
+  i2 = (const struct intel_02_cache_info *) p2;
+
+  if (i1->idx == i2->idx)
+    return 0;
+
+  return i1->idx < i2->idx ? -1 : 1;
+}
+
+
+static long int
+__attribute__ ((noinline))
+intel_check_word (int name, unsigned int value, bool *has_level_2,
+  bool *no_level_2_or_3,
+  const struct cpu_features *cpu_features)
+{
+  if ((value & 0x80000000) != 0)
+    /* The register value is reserved.  */
+    return 0;
+
+  /* Fold the name.  The _SC_ constants are always in the order SIZE,
+     ASSOC, LINESIZE.  */
+  int folded_rel_name = (M(name) / 3) * 3;
+
+  while (value != 0)
+    {
+      unsigned int byte = value & 0xff;
+
+      if (byte == 0x40)
+ {
+  *no_level_2_or_3 = true;
+
+  if (folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))
+    /* No need to look further.  */
+    break;
+ }
+      else if (byte == 0xff)
+ {
+  /* CPUID leaf 0x4 contains all the information.  We need to
+     iterate over it.  */
+  unsigned int eax;
+  unsigned int ebx;
+  unsigned int ecx;
+  unsigned int edx;
+
+  unsigned int round = 0;
+  while (1)
+    {
+      __cpuid_count (4, round, eax, ebx, ecx, edx);
+
+      enum { null = 0, data = 1, inst = 2, uni = 3 } type = eax & 0x1f;
+      if (type == null)
+ /* That was the end.  */
+ break;
+
+      unsigned int level = (eax >> 5) & 0x7;
+
+      if ((level == 1 && type == data
+   && folded_rel_name == M(_SC_LEVEL1_DCACHE_SIZE))
+  || (level == 1 && type == inst
+      && folded_rel_name == M(_SC_LEVEL1_ICACHE_SIZE))
+  || (level == 2 && folded_rel_name == M(_SC_LEVEL2_CACHE_SIZE))
+  || (level == 3 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))
+  || (level == 4 && folded_rel_name == M(_SC_LEVEL4_CACHE_SIZE)))
+ {
+  unsigned int offset = M(name) - folded_rel_name;
+
+  if (offset == 0)
+    /* Cache size.  */
+    return (((ebx >> 22) + 1)
+    * (((ebx >> 12) & 0x3ff) + 1)
+    * ((ebx & 0xfff) + 1)
+    * (ecx + 1));
+  if (offset == 1)
+    return (ebx >> 22) + 1;
+
+  assert (offset == 2);
+  return (ebx & 0xfff) + 1;
+ }
+
+      ++round;
+    }
+  /* There is no other cache information anywhere else.  */
+  break;
+ }
+      else
+ {
+  if (byte == 0x49 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))
+    {
+      /* Intel reused this value.  For family 15, model 6 it
+ specifies the 3rd level cache.  Otherwise the 2nd
+ level cache.  */
+      unsigned int family = cpu_features->basic.family;
+      unsigned int model = cpu_features->basic.model;
+
+      if (family == 15 && model == 6)
+ {
+  /* The level 3 cache is encoded for this model like
+     the level 2 cache is for other models.  Pretend
+     the caller asked for the level 2 cache.  */
+  name = (_SC_LEVEL2_CACHE_SIZE
+  + (name - _SC_LEVEL3_CACHE_SIZE));
+  folded_rel_name = M(_SC_LEVEL2_CACHE_SIZE);
+ }
+    }
+
+  struct intel_02_cache_info *found;
+  struct intel_02_cache_info search;
+
+  search.idx = byte;
+  found = bsearch (&search, intel_02_known, nintel_02_known,
+   sizeof (intel_02_known[0]), intel_02_known_compare);
+  if (found != NULL)
+    {
+      if (found->rel_name == folded_rel_name)
+ {
+  unsigned int offset = M(name) - folded_rel_name;
+
+  if (offset == 0)
+    /* Cache size.  */
+    return found->size;
+  if (offset == 1)
+    return found->assoc;
+
+  assert (offset == 2);
+  return found->linesize;
+ }
+
+      if (found->rel_name == M(_SC_LEVEL2_CACHE_SIZE))
+ *has_level_2 = true;
+    }
+ }
+
+      /* Next byte for the next round.  */
+      value >>= 8;
+    }
+
+  /* Nothing found.  */
+  return 0;
+}
+
+
+static long int __attribute__ ((noinline))
+handle_intel (int name, const struct cpu_features *cpu_features)
+{
+  unsigned int maxidx = cpu_features->basic.max_cpuid;
+
+  /* Return -1 for older CPUs.  */
+  if (maxidx < 2)
+    return -1;
+
+  /* OK, we can use the CPUID instruction to get all info about the
+     caches.  */
+  unsigned int cnt = 0;
+  unsigned int max = 1;
+  long int result = 0;
+  bool no_level_2_or_3 = false;
+  bool has_level_2 = false;
+
+  while (cnt++ < max)
+    {
+      unsigned int eax;
+      unsigned int ebx;
+      unsigned int ecx;
+      unsigned int edx;
+      __cpuid (2, eax, ebx, ecx, edx);
+
+      /* The low byte of EAX in the first round contain the number of
+ rounds we have to make.  At least one, the one we are already
+ doing.  */
+      if (cnt == 1)
+ {
+  max = eax & 0xff;
+  eax &= 0xffffff00;
+ }
+
+      /* Process the individual registers' value.  */
+      result = intel_check_word (name, eax, &has_level_2,
+ &no_level_2_or_3, cpu_features);
+      if (result != 0)
+ return result;
+
+      result = intel_check_word (name, ebx, &has_level_2,
+ &no_level_2_or_3, cpu_features);
+      if (result != 0)
+ return result;
+
+      result = intel_check_word (name, ecx, &has_level_2,
+ &no_level_2_or_3, cpu_features);
+      if (result != 0)
+ return result;
+
+      result = intel_check_word (name, edx, &has_level_2,
+ &no_level_2_or_3, cpu_features);
+      if (result != 0)
+ return result;
+    }
+
+  if (name >= _SC_LEVEL2_CACHE_SIZE && name <= _SC_LEVEL3_CACHE_LINESIZE
+      && no_level_2_or_3)
+    return -1;
+
+  return 0;
+}
+
+
+static long int __attribute__ ((noinline))
+handle_amd (int name)
+{
+  unsigned int eax;
+  unsigned int ebx;
+  unsigned int ecx;
+  unsigned int edx;
+  __cpuid (0x80000000, eax, ebx, ecx, edx);
+
+  /* No level 4 cache (yet).  */
+  if (name > _SC_LEVEL3_CACHE_LINESIZE)
+    return 0;
+
+  unsigned int fn = 0x80000005 + (name >= _SC_LEVEL2_CACHE_SIZE);
+  if (eax < fn)
+    return 0;
+
+  __cpuid (fn, eax, ebx, ecx, edx);
+
+  if (name < _SC_LEVEL1_DCACHE_SIZE)
+    {
+      name += _SC_LEVEL1_DCACHE_SIZE - _SC_LEVEL1_ICACHE_SIZE;
+      ecx = edx;
+    }
+
+  switch (name)
+    {
+    case _SC_LEVEL1_DCACHE_SIZE:
+      return (ecx >> 14) & 0x3fc00;
+
+    case _SC_LEVEL1_DCACHE_ASSOC:
+      ecx >>= 16;
+      if ((ecx & 0xff) == 0xff)
+ /* Fully associative.  */
+ return (ecx << 2) & 0x3fc00;
+      return ecx & 0xff;
+
+    case _SC_LEVEL1_DCACHE_LINESIZE:
+      return ecx & 0xff;
+
+    case _SC_LEVEL2_CACHE_SIZE:
+      return (ecx & 0xf000) == 0 ? 0 : (ecx >> 6) & 0x3fffc00;
+
+    case _SC_LEVEL2_CACHE_ASSOC:
+      switch ((ecx >> 12) & 0xf)
+ {
+ case 0:
+ case 1:
+ case 2:
+ case 4:
+  return (ecx >> 12) & 0xf;
+ case 6:
+  return 8;
+ case 8:
+  return 16;
+ case 10:
+  return 32;
+ case 11:
+  return 48;
+ case 12:
+  return 64;
+ case 13:
+  return 96;
+ case 14:
+  return 128;
+ case 15:
+  return ((ecx >> 6) & 0x3fffc00) / (ecx & 0xff);
+ default:
+  return 0;
+ }
+      /* NOTREACHED */
+
+    case _SC_LEVEL2_CACHE_LINESIZE:
+      return (ecx & 0xf000) == 0 ? 0 : ecx & 0xff;
+
+    case _SC_LEVEL3_CACHE_SIZE:
+      return (edx & 0xf000) == 0 ? 0 : (edx & 0x3ffc0000) << 1;
+
+    case _SC_LEVEL3_CACHE_ASSOC:
+      switch ((edx >> 12) & 0xf)
+ {
+ case 0:
+ case 1:
+ case 2:
+ case 4:
+  return (edx >> 12) & 0xf;
+ case 6:
+  return 8;
+ case 8:
+  return 16;
+ case 10:
+  return 32;
+ case 11:
+  return 48;
+ case 12:
+  return 64;
+ case 13:
+  return 96;
+ case 14:
+  return 128;
+ case 15:
+  return ((edx & 0x3ffc0000) << 1) / (edx & 0xff);
+ default:
+  return 0;
+ }
+      /* NOTREACHED */
+
+    case _SC_LEVEL3_CACHE_LINESIZE:
+      return (edx & 0xf000) == 0 ? 0 : edx & 0xff;
+
+    default:
+      assert (! "cannot happen");
+    }
+  return -1;
+}
+
+
+static long int __attribute__ ((noinline))
+handle_zhaoxin (int name)
+{
+  unsigned int eax;
+  unsigned int ebx;
+  unsigned int ecx;
+  unsigned int edx;
+
+  int folded_rel_name = (M(name) / 3) * 3;
+
+  unsigned int round = 0;
+  while (1)
+    {
+      __cpuid_count (4, round, eax, ebx, ecx, edx);
+
+      enum { null = 0, data = 1, inst = 2, uni = 3 } type = eax & 0x1f;
+      if (type == null)
+        break;
+
+      unsigned int level = (eax >> 5) & 0x7;
+
+      if ((level == 1 && type == data
+        && folded_rel_name == M(_SC_LEVEL1_DCACHE_SIZE))
+        || (level == 1 && type == inst
+            && folded_rel_name == M(_SC_LEVEL1_ICACHE_SIZE))
+        || (level == 2 && folded_rel_name == M(_SC_LEVEL2_CACHE_SIZE))
+        || (level == 3 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE)))
+        {
+          unsigned int offset = M(name) - folded_rel_name;
+
+          if (offset == 0)
+            /* Cache size.  */
+            return (((ebx >> 22) + 1)
+                * (((ebx >> 12) & 0x3ff) + 1)
+                * ((ebx & 0xfff) + 1)
+                * (ecx + 1));
+          if (offset == 1)
+            return (ebx >> 22) + 1;
+
+          assert (offset == 2);
+          return (ebx & 0xfff) + 1;
+        }
+
+      ++round;
+    }
+
+  /* Nothing found.  */
+  return 0;
+}
+
+
+static void
+get_common_cache_info (long int *shared_ptr, unsigned int *threads_ptr,
+                long int core)
+{
+  unsigned int eax;
+  unsigned int ebx;
+  unsigned int ecx;
+  unsigned int edx;
+
+  /* Number of logical processors sharing L2 cache.  */
+  int threads_l2;
+
+  /* Number of logical processors sharing L3 cache.  */
+  int threads_l3;
+
+  const struct cpu_features *cpu_features = __get_cpu_features ();
+  int max_cpuid = cpu_features->basic.max_cpuid;
+  unsigned int family = cpu_features->basic.family;
+  unsigned int model = cpu_features->basic.model;
+  long int shared = *shared_ptr;
+  unsigned int threads = *threads_ptr;
+  bool inclusive_cache = true;
+  bool support_count_mask = true;
+
+  /* Try L3 first.  */
+  unsigned int level = 3;
+
+  if (cpu_features->basic.kind == arch_kind_zhaoxin && family == 6)
+    support_count_mask = false;
+
+  if (shared <= 0)
+    {
+      /* Try L2 otherwise.  */
+      level  = 2;
+      shared = core;
+      threads_l2 = 0;
+      threads_l3 = -1;
+    }
+  else
+    {
+      threads_l2 = 0;
+      threads_l3 = 0;
+    }
+
+  /* A value of 0 for the HTT bit indicates there is only a single
+     logical processor.  */
+  if (HAS_CPU_FEATURE (HTT))
+    {
+      /* Figure out the number of logical threads that share the
+         highest cache level.  */
+      if (max_cpuid >= 4)
+        {
+          int i = 0;
+
+          /* Query until cache level 2 and 3 are enumerated.  */
+          int check = 0x1 | (threads_l3 == 0) << 1;
+          do
+            {
+              __cpuid_count (4, i++, eax, ebx, ecx, edx);
+
+              /* There seems to be a bug in at least some Pentium Ds
+                 which sometimes fail to iterate all cache parameters.
+                 Do not loop indefinitely here, stop in this case and
+                 assume there is no such information.  */
+              if (cpu_features->basic.kind == arch_kind_intel
+                  && (eax & 0x1f) == 0 )
+                goto intel_bug_no_cache_info;
+
+              switch ((eax >> 5) & 0x7)
+                {
+                  default:
+                    break;
+                  case 2:
+                    if ((check & 0x1))
+                      {
+                        /* Get maximum number of logical processors
+                           sharing L2 cache.  */
+                        threads_l2 = (eax >> 14) & 0x3ff;
+                        check &= ~0x1;
+                      }
+                    break;
+                  case 3:
+                    if ((check & (0x1 << 1)))
+                      {
+                        /* Get maximum number of logical processors
+                           sharing L3 cache.  */
+                        threads_l3 = (eax >> 14) & 0x3ff;
+
+                        /* Check if L2 and L3 caches are inclusive.  */
+                        inclusive_cache = (edx & 0x2) != 0;
+                        check &= ~(0x1 << 1);
+                      }
+                    break;
+                }
+            }
+          while (check);
+
+          /* If max_cpuid >= 11, THREADS_L2/THREADS_L3 are the maximum
+             numbers of addressable IDs for logical processors sharing
+             the cache, instead of the maximum number of threads
+             sharing the cache.  */
+          if (max_cpuid >= 11 && support_count_mask)
+            {
+              /* Find the number of logical processors shipped in
+                 one core and apply count mask.  */
+              i = 0;
+
+              /* Count SMT only if there is L3 cache.  Always count
+                 core if there is no L3 cache.  */
+              int count = ((threads_l2 > 0 && level == 3)
+                           | ((threads_l3 > 0
+                               || (threads_l2 > 0 && level == 2)) << 1));
+
+              while (count)
+                {
+                  __cpuid_count (11, i++, eax, ebx, ecx, edx);
+
+                  int shipped = ebx & 0xff;
+                  int type = ecx & 0xff00;
+                  if (shipped == 0 || type == 0)
+                    break;
+                  else if (type == 0x100)
+                    {
+                      /* Count SMT.  */
+                      if ((count & 0x1))
+                        {
+                          int count_mask;
+
+                          /* Compute count mask.  */
+                          asm ("bsr %1, %0"
+                               : "=r" (count_mask) : "g" (threads_l2));
+                          count_mask = ~(-1 << (count_mask + 1));
+                          threads_l2 = (shipped - 1) & count_mask;
+                          count &= ~0x1;
+                        }
+                    }
+                  else if (type == 0x200)
+                    {
+                      /* Count core.  */
+                      if ((count & (0x1 << 1)))
+                        {
+                          int count_mask;
+                          int threads_core
+                            = (level == 2 ? threads_l2 : threads_l3);
+
+                          /* Compute count mask.  */
+                          asm ("bsr %1, %0"
+                               : "=r" (count_mask) : "g" (threads_core));
+                          count_mask = ~(-1 << (count_mask + 1));
+                          threads_core = (shipped - 1) & count_mask;
+                          if (level == 2)
+                            threads_l2 = threads_core;
+                          else
+                            threads_l3 = threads_core;
+                          count &= ~(0x1 << 1);
+                        }
+                    }
+                }
+            }
+          if (threads_l2 > 0)
+            threads_l2 += 1;
+          if (threads_l3 > 0)
+            threads_l3 += 1;
+          if (level == 2)
+            {
+              if (threads_l2)
+                {
+                  threads = threads_l2;
+                  if (cpu_features->basic.kind == arch_kind_intel
+                      && threads > 2
+                      && family == 6)
+                    switch (model)
+                      {
+                        case 0x37:
+                        case 0x4a:
+                        case 0x4d:
+                        case 0x5a:
+                        case 0x5d:
+                          /* Silvermont has L2 cache shared by 2 cores.  */
+                          threads = 2;
+                          break;
+                        default:
+                          break;
+                      }
+                }
+            }
+          else if (threads_l3)
+            threads = threads_l3;
+        }
+      else
+        {
+intel_bug_no_cache_info:
+          /* Assume that all logical threads share the highest cache
+             level.  */
+          threads
+            = ((cpu_features->cpuid[COMMON_CPUID_INDEX_1].ebx
+                >> 16) & 0xff);
+        }
+
+        /* Cap usage of highest cache level to the number of supported
+           threads.  */
+        if (shared > 0 && threads > 0)
+          shared /= threads;
+    }
+
+  /* Account for non-inclusive L2 and L3 caches.  */
+  if (!inclusive_cache)
+    {
+      if (threads_l2 > 0)
+        core /= threads_l2;
+      shared += core;
+    }
+
+  *shared_ptr = shared;
+  *threads_ptr = threads;
+}
+
+void
+__init_cacheinfo (void)
+{
+  /* Find out what brand of processor.  */
+  unsigned int ebx;
+  unsigned int ecx;
+  unsigned int edx;
+  int max_cpuid_ex;
+  long int data = -1;
+  long int shared = -1;
+  long int core;
+  unsigned int threads = 0;
+  unsigned long int level1_icache_size = -1;
+  unsigned long int level1_dcache_size = -1;
+  unsigned long int level1_dcache_assoc = -1;
+  unsigned long int level1_dcache_linesize = -1;
+  unsigned long int level2_cache_size = -1;
+  unsigned long int level2_cache_assoc = -1;
+  unsigned long int level2_cache_linesize = -1;
+  unsigned long int level3_cache_size = -1;
+  unsigned long int level3_cache_assoc = -1;
+  unsigned long int level3_cache_linesize = -1;
+  unsigned long int level4_cache_size = -1;
+  struct cpu_features *cpu_features = __get_cpu_features ();
+
+  if (cpu_features->basic.kind == arch_kind_intel)
+    {
+      data = handle_intel (_SC_LEVEL1_DCACHE_SIZE, cpu_features);
+      core = handle_intel (_SC_LEVEL2_CACHE_SIZE, cpu_features);
+      shared = handle_intel (_SC_LEVEL3_CACHE_SIZE, cpu_features);
+
+      level1_icache_size
+ = handle_intel (_SC_LEVEL1_ICACHE_SIZE, cpu_features);
+      level1_dcache_size = data;
+      level1_dcache_assoc
+ = handle_intel (_SC_LEVEL1_DCACHE_ASSOC, cpu_features);
+      level1_dcache_linesize
+ = handle_intel (_SC_LEVEL1_DCACHE_LINESIZE, cpu_features);
+      level2_cache_size = core;
+      level2_cache_assoc
+ = handle_intel (_SC_LEVEL2_CACHE_ASSOC, cpu_features);
+      level2_cache_linesize
+ = handle_intel (_SC_LEVEL2_CACHE_LINESIZE, cpu_features);
+      level3_cache_size = shared;
+      level3_cache_assoc
+ = handle_intel (_SC_LEVEL3_CACHE_ASSOC, cpu_features);
+      level3_cache_linesize
+ = handle_intel (_SC_LEVEL3_CACHE_LINESIZE, cpu_features);
+      level4_cache_size
+ = handle_intel (_SC_LEVEL4_CACHE_SIZE, cpu_features);
+
+      get_common_cache_info (&shared, &threads, core);
+    }
+  else if (cpu_features->basic.kind == arch_kind_zhaoxin)
+    {
+      data = handle_zhaoxin (_SC_LEVEL1_DCACHE_SIZE);
+      core = handle_zhaoxin (_SC_LEVEL2_CACHE_SIZE);
+      shared = handle_zhaoxin (_SC_LEVEL3_CACHE_SIZE);
+
+      level1_icache_size = handle_zhaoxin (_SC_LEVEL1_ICACHE_SIZE);
+      level1_dcache_size = data;
+      level1_dcache_assoc = handle_zhaoxin (_SC_LEVEL1_DCACHE_ASSOC);
+      level1_dcache_linesize = handle_zhaoxin (_SC_LEVEL1_DCACHE_LINESIZE);
+      level2_cache_size = core;
+      level2_cache_assoc = handle_zhaoxin (_SC_LEVEL2_CACHE_ASSOC);
+      level2_cache_linesize = handle_zhaoxin (_SC_LEVEL2_CACHE_LINESIZE);
+      level3_cache_size = shared;
+      level3_cache_assoc = handle_zhaoxin (_SC_LEVEL3_CACHE_ASSOC);
+      level3_cache_linesize = handle_zhaoxin (_SC_LEVEL3_CACHE_LINESIZE);
+
+      get_common_cache_info (&shared, &threads, core);
+    }
+  else if (cpu_features->basic.kind == arch_kind_amd)
+    {
+      data  = handle_amd (_SC_LEVEL1_DCACHE_SIZE);
+      core = handle_amd (_SC_LEVEL2_CACHE_SIZE);
+      shared = handle_amd (_SC_LEVEL3_CACHE_SIZE);
+
+      level1_icache_size = handle_amd (_SC_LEVEL1_ICACHE_SIZE);
+      level1_dcache_size = data;
+      level1_dcache_assoc = handle_amd (_SC_LEVEL1_DCACHE_ASSOC);
+      level1_dcache_linesize = handle_amd (_SC_LEVEL1_DCACHE_LINESIZE);
+      level2_cache_size = core;
+      level2_cache_assoc = handle_amd (_SC_LEVEL2_CACHE_ASSOC);
+      level2_cache_linesize = handle_amd (_SC_LEVEL2_CACHE_LINESIZE);
+      level3_cache_size = shared;
+      level3_cache_assoc = handle_amd (_SC_LEVEL3_CACHE_ASSOC);
+      level3_cache_linesize = handle_amd (_SC_LEVEL3_CACHE_LINESIZE);
+
+      /* Get maximum extended function. */
+      __cpuid (0x80000000, max_cpuid_ex, ebx, ecx, edx);
+
+      if (shared <= 0)
+ /* No shared L3 cache.  All we have is the L2 cache.  */
+ shared = core;
+      else
+ {
+  /* Figure out the number of logical threads that share L3.  */
+  if (max_cpuid_ex >= 0x80000008)
+    {
+      /* Get width of APIC ID.  */
+      __cpuid (0x80000008, max_cpuid_ex, ebx, ecx, edx);
+      threads = 1 << ((ecx >> 12) & 0x0f);
+    }
+
+  if (threads == 0)
+    {
+      /* If APIC ID width is not available, use logical
+ processor count.  */
+      __cpuid (0x00000001, max_cpuid_ex, ebx, ecx, edx);
+
+      if ((edx & (1 << 28)) != 0)
+ threads = (ebx >> 16) & 0xff;
+    }
+
+  /* Cap usage of highest cache level to the number of
+     supported threads.  */
+  if (threads > 0)
+    shared /= threads;
+
+  /* Account for exclusive L2 and L3 caches.  */
+  shared += core;
+ }
+
+#ifndef __x86_64__
+      if (max_cpuid_ex >= 0x80000001)
+ {
+  unsigned int eax;
+  __cpuid (0x80000001, eax, ebx, ecx, edx);
+  /*  PREFETCHW     || 3DNow!  */
+  if ((ecx & 0x100) || (edx & 0x80000000))
+    cpu_features->prefetchw = -1;
+ }
+#endif
+    }
+
+  cpu_features->level1_icache_size = level1_icache_size;
+  cpu_features->level1_dcache_size = level1_dcache_size;
+  cpu_features->level1_dcache_assoc = level1_dcache_assoc;
+  cpu_features->level1_dcache_linesize = level1_dcache_linesize;
+  cpu_features->level2_cache_size = level2_cache_size;
+  cpu_features->level2_cache_assoc = level2_cache_assoc;
+  cpu_features->level2_cache_linesize = level2_cache_linesize;
+  cpu_features->level3_cache_size = level3_cache_size;
+  cpu_features->level3_cache_assoc = level3_cache_assoc;
+  cpu_features->level3_cache_linesize = level3_cache_linesize;
+  cpu_features->level4_cache_size = level4_cache_size;
+
+  /* The large memcpy micro benchmark in glibc shows that 6 times of
+     shared cache size is the approximate value above which non-temporal
+     store becomes faster on a 8-core processor.  This is the 3/4 of the
+     total shared cache size.  */
+  unsigned long int non_temporal_threshold = (shared * threads * 3 / 4);
+
+#if HAVE_TUNABLES
+  long int tunable_size;
+  tunable_size = TUNABLE_GET (x86_data_cache_size, long int, NULL);
+  if (tunable_size != 0)
+    data = tunable_size;
+  tunable_size = TUNABLE_GET (x86_shared_cache_size, long int, NULL);
+  if (tunable_size != 0)
+    shared = tunable_size;
+  tunable_size = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
+  if (tunable_size != 0)
+    non_temporal_threshold = tunable_size;
+#endif
+
+  cpu_features->data_cache_size = data;
+  cpu_features->shared_cache_size = shared;
+  cpu_features->non_temporal_threshold = non_temporal_threshold;
+
+#if HAVE_TUNABLES
+  TUNABLE_UPDATE (x86_data_cache_size, long int,
+  data, 0, (long int) -1);
+  TUNABLE_UPDATE (x86_shared_cache_size, long int,
+  shared, 0, (long int) -1);
+  TUNABLE_UPDATE (x86_non_temporal_threshold, long int,
+  non_temporal_threshold, 0, (long int) -1);
+#endif
+}
diff --git a/sysdeps/x86/init-arch.h b/sysdeps/x86/init-arch.h
index d6f59cf962..272ed10902 100644
--- a/sysdeps/x86/init-arch.h
+++ b/sysdeps/x86/init-arch.h
@@ -23,6 +23,9 @@
 #include <ifunc-init.h>
 #include <isa.h>
 
+extern void __init_cacheinfo (void)
+  __attribute__ ((visibility ("hidden")));
+
 #ifndef __x86_64__
 /* Due to the reordering and the other nifty extensions in i686, it is
    not really good to use heavily i586 optimized code on an i686.  It's
--
2.26.2

Reply | Threaded
Open this post in threaded view
|

[PATCH 2/2] x86: Add thresholds for "rep movsb/stosb" to tunables

Sourceware - libc-alpha mailing list
In reply to this post by Sourceware - libc-alpha mailing list
Add x86_rep_movsb_threshold and x86_rep_stosb_threshold to tunables
to update thresholds for "rep movsb" and "rep stosb" at run-time.

Note that the user specified threshold for "rep movsb" smaller than the
minimum threshold will be ignored.
---
 manual/tunables.texi                          | 14 +++++++
 sysdeps/x86/cacheinfo.c                       | 20 ++++++++++
 sysdeps/x86/cpu-features.h                    |  4 ++
 sysdeps/x86/dl-cacheinfo.c                    | 38 +++++++++++++++++++
 sysdeps/x86/dl-tunables.list                  |  6 +++
 .../multiarch/memmove-vec-unaligned-erms.S    | 16 +-------
 .../multiarch/memset-vec-unaligned-erms.S     | 12 +-----
 7 files changed, 84 insertions(+), 26 deletions(-)

diff --git a/manual/tunables.texi b/manual/tunables.texi
index ec18b10834..61edd62425 100644
--- a/manual/tunables.texi
+++ b/manual/tunables.texi
@@ -396,6 +396,20 @@ to set threshold in bytes for non temporal store.
 This tunable is specific to i386 and x86-64.
 @end deftp
 
+@deftp Tunable glibc.cpu.x86_rep_movsb_threshold
+The @code{glibc.cpu.x86_rep_movsb_threshold} tunable allows the user
+to set threshold in bytes to start using "rep movsb".
+
+This tunable is specific to i386 and x86-64.
+@end deftp
+
+@deftp Tunable glibc.cpu.x86_rep_stosb_threshold
+The @code{glibc.cpu.x86_rep_stosb_threshold} tunable allows the user
+to set threshold in bytes to start using "rep stosb".
+
+This tunable is specific to i386 and x86-64.
+@end deftp
+
 @deftp Tunable glibc.cpu.x86_ibt
 The @code{glibc.cpu.x86_ibt} tunable allows the user to control how
 indirect branch tracking (IBT) should be enabled.  Accepted values are
diff --git a/sysdeps/x86/cacheinfo.c b/sysdeps/x86/cacheinfo.c
index 8c4c7f9972..bb536d96ef 100644
--- a/sysdeps/x86/cacheinfo.c
+++ b/sysdeps/x86/cacheinfo.c
@@ -41,6 +41,23 @@ long int __x86_raw_shared_cache_size attribute_hidden = 1024 * 1024;
 /* Threshold to use non temporal store.  */
 long int __x86_shared_non_temporal_threshold attribute_hidden;
 
+/* Threshold to use Enhanced REP MOVSB.  Since there is overhead to set
+   up REP MOVSB operation, REP MOVSB isn't faster on short data.  The
+   memcpy micro benchmark in glibc shows that 2KB is the approximate
+   value above which REP MOVSB becomes faster than SSE2 optimization
+   on processors with Enhanced REP MOVSB.  Since larger register size
+   can move more data with a single load and store, the threshold is
+   higher with larger register size.  */
+long int __x86_rep_movsb_threshold attribute_hidden = 2048;
+
+/* Threshold to use Enhanced REP STOSB.  Since there is overhead to set
+   up REP STOSB operation, REP STOSB isn't faster on short data.  The
+   memset micro benchmark in glibc shows that 2KB is the approximate
+   value above which REP STOSB becomes faster on processors with
+   Enhanced REP STOSB.  Since the stored value is fixed, larger register
+   size has minimal impact on threshold.  */
+long int __x86_rep_stosb_threshold attribute_hidden = 2048;
+
 #ifndef __x86_64__
 /* PREFETCHW support flag for use in memory and string routines.  */
 int __x86_prefetchw attribute_hidden;
@@ -117,6 +134,9 @@ init_cacheinfo (void)
   __x86_shared_non_temporal_threshold
     = cpu_features->non_temporal_threshold;
 
+  __x86_rep_movsb_threshold = cpu_features->rep_movsb_threshold;
+  __x86_rep_stosb_threshold = cpu_features->rep_stosb_threshold;
+
 #ifndef __x86_64__
   __x86_prefetchw = cpu_features->prefetchw;
 #endif
diff --git a/sysdeps/x86/cpu-features.h b/sysdeps/x86/cpu-features.h
index 3aaed33cbc..002e12e11f 100644
--- a/sysdeps/x86/cpu-features.h
+++ b/sysdeps/x86/cpu-features.h
@@ -128,6 +128,10 @@ struct cpu_features
   /* PREFETCHW support flag for use in memory and string routines.  */
   unsigned long int prefetchw;
 #endif
+  /* Threshold to use "rep movsb".  */
+  unsigned long int rep_movsb_threshold;
+  /* Threshold to use "rep stosb".  */
+  unsigned long int rep_stosb_threshold;
 };
 
 /* Used from outside of glibc to get access to the CPU features
diff --git a/sysdeps/x86/dl-cacheinfo.c b/sysdeps/x86/dl-cacheinfo.c
index 8e2a6f552c..aff9bd1067 100644
--- a/sysdeps/x86/dl-cacheinfo.c
+++ b/sysdeps/x86/dl-cacheinfo.c
@@ -860,6 +860,31 @@ __init_cacheinfo (void)
      total shared cache size.  */
   unsigned long int non_temporal_threshold = (shared * threads * 3 / 4);
 
+  /* NB: The REP MOVSB threshold must be greater than VEC_SIZE * 8.  */
+  unsigned long int minimum_rep_movsb_threshold;
+  /* NB: The default REP MOVSB threshold is 2048 * (VEC_SIZE / 16).  See
+     comments for __x86_rep_movsb_threshold in cacheinfo.c.  */
+  unsigned long int rep_movsb_threshold;
+  if (CPU_FEATURES_ARCH_P (cpu_features, AVX512F_Usable)
+      && !CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_AVX512))
+    {
+      rep_movsb_threshold = 2048 * (64 / 16);
+      minimum_rep_movsb_threshold = 64 * 8;
+    }
+  else if (CPU_FEATURES_ARCH_P (cpu_features,
+ AVX_Fast_Unaligned_Load))
+    {
+      rep_movsb_threshold = 2048 * (32 / 16);
+      minimum_rep_movsb_threshold = 32 * 8;
+    }
+  else
+    {
+      rep_movsb_threshold = 2048 * (16 / 16);
+      minimum_rep_movsb_threshold = 16 * 8;
+    }
+  /* NB: See comments for __x86_rep_stosb_threshold in cacheinfo.c.  */
+  unsigned long int rep_stosb_threshold = 2048;
+
 #if HAVE_TUNABLES
   long int tunable_size;
   tunable_size = TUNABLE_GET (x86_data_cache_size, long int, NULL);
@@ -871,11 +896,19 @@ __init_cacheinfo (void)
   tunable_size = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
   if (tunable_size != 0)
     non_temporal_threshold = tunable_size;
+  tunable_size = TUNABLE_GET (x86_rep_movsb_threshold, long int, NULL);
+  if (tunable_size > minimum_rep_movsb_threshold)
+    rep_movsb_threshold = tunable_size;
+  tunable_size = TUNABLE_GET (x86_rep_stosb_threshold, long int, NULL);
+  if (tunable_size != 0)
+    rep_stosb_threshold = tunable_size;
 #endif
 
   cpu_features->data_cache_size = data;
   cpu_features->shared_cache_size = shared;
   cpu_features->non_temporal_threshold = non_temporal_threshold;
+  cpu_features->rep_movsb_threshold = rep_movsb_threshold;
+  cpu_features->rep_stosb_threshold = rep_stosb_threshold;
 
 #if HAVE_TUNABLES
   TUNABLE_UPDATE (x86_data_cache_size, long int,
@@ -884,5 +917,10 @@ __init_cacheinfo (void)
   shared, 0, (long int) -1);
   TUNABLE_UPDATE (x86_non_temporal_threshold, long int,
   non_temporal_threshold, 0, (long int) -1);
+  TUNABLE_UPDATE (x86_rep_movsb_threshold, long int,
+  rep_movsb_threshold, minimum_rep_movsb_threshold,
+  (long int) -1);
+  TUNABLE_UPDATE (x86_rep_stosb_threshold, long int,
+  rep_stosb_threshold, 0, (long int) -1);
 #endif
 }
diff --git a/sysdeps/x86/dl-tunables.list b/sysdeps/x86/dl-tunables.list
index 251b926ce4..43bf6c2389 100644
--- a/sysdeps/x86/dl-tunables.list
+++ b/sysdeps/x86/dl-tunables.list
@@ -30,6 +30,12 @@ glibc {
     x86_non_temporal_threshold {
       type: SIZE_T
     }
+    x86_rep_movsb_threshold {
+      type: SIZE_T
+    }
+    x86_rep_stosb_threshold {
+      type: SIZE_T
+    }
     x86_data_cache_size {
       type: SIZE_T
     }
diff --git a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
index 74953245aa..bd5dc1a3f3 100644
--- a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
+++ b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
@@ -56,17 +56,6 @@
 # endif
 #endif
 
-/* Threshold to use Enhanced REP MOVSB.  Since there is overhead to set
-   up REP MOVSB operation, REP MOVSB isn't faster on short data.  The
-   memcpy micro benchmark in glibc shows that 2KB is the approximate
-   value above which REP MOVSB becomes faster than SSE2 optimization
-   on processors with Enhanced REP MOVSB.  Since larger register size
-   can move more data with a single load and store, the threshold is
-   higher with larger register size.  */
-#ifndef REP_MOVSB_THRESHOLD
-# define REP_MOVSB_THRESHOLD (2048 * (VEC_SIZE / 16))
-#endif
-
 #ifndef PREFETCH
 # define PREFETCH(addr) prefetcht0 addr
 #endif
@@ -253,9 +242,6 @@ L(movsb):
  leaq (%rsi,%rdx), %r9
  cmpq %r9, %rdi
  /* Avoid slow backward REP MOVSB.  */
-# if REP_MOVSB_THRESHOLD <= (VEC_SIZE * 8)
-#  error Unsupported REP_MOVSB_THRESHOLD and VEC_SIZE!
-# endif
  jb L(more_8x_vec_backward)
 1:
  mov %RDX_LP, %RCX_LP
@@ -331,7 +317,7 @@ L(between_2_3):
 
 #if defined USE_MULTIARCH && IS_IN (libc)
 L(movsb_more_2x_vec):
- cmpq $REP_MOVSB_THRESHOLD, %rdx
+ cmp __x86_rep_movsb_threshold(%rip), %RDX_LP
  ja L(movsb)
 #endif
 L(more_2x_vec):
diff --git a/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S b/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
index af2299709c..2bfc95de05 100644
--- a/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
+++ b/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
@@ -58,16 +58,6 @@
 # endif
 #endif
 
-/* Threshold to use Enhanced REP STOSB.  Since there is overhead to set
-   up REP STOSB operation, REP STOSB isn't faster on short data.  The
-   memset micro benchmark in glibc shows that 2KB is the approximate
-   value above which REP STOSB becomes faster on processors with
-   Enhanced REP STOSB.  Since the stored value is fixed, larger register
-   size has minimal impact on threshold.  */
-#ifndef REP_STOSB_THRESHOLD
-# define REP_STOSB_THRESHOLD 2048
-#endif
-
 #ifndef SECTION
 # error SECTION is not defined!
 #endif
@@ -181,7 +171,7 @@ ENTRY (MEMSET_SYMBOL (__memset, unaligned_erms))
  ret
 
 L(stosb_more_2x_vec):
- cmpq $REP_STOSB_THRESHOLD, %rdx
+ cmp __x86_rep_stosb_threshold(%rip), %RDX_LP
  ja L(stosb)
 #endif
 L(more_2x_vec):
--
2.26.2

Reply | Threaded
Open this post in threaded view
|

Re: [PATCH 1/2] Update tunable min/max values

Sourceware - libc-alpha mailing list
In reply to this post by Sourceware - libc-alpha mailing list
On 7/3/20 1:52 PM, H.J. Lu wrote:
> Add __tunable_update_val to update tunable min/max values and move x86
> processor cache info to cpu_features.

Nees a little more cleanup. I'd like to avoid more macro APIs if possible.  

Post v2 please.

> ---
>  elf/dl-tunables.c          |  51 ++-
>  elf/dl-tunables.h          |  15 +
>  sysdeps/i386/cacheinfo.c   |   3 -
>  sysdeps/x86/Makefile       |   2 +-
>  sysdeps/x86/cacheinfo.c    | 852 +++--------------------------------
>  sysdeps/x86/cpu-features.c |  19 +-
>  sysdeps/x86/cpu-features.h |  26 ++
>  sysdeps/x86/dl-cacheinfo.c | 888 +++++++++++++++++++++++++++++++++++++
>  sysdeps/x86/init-arch.h    |   3 +
>  9 files changed, 1024 insertions(+), 835 deletions(-)
>  delete mode 100644 sysdeps/i386/cacheinfo.c
>  create mode 100644 sysdeps/x86/dl-cacheinfo.c
>
> diff --git a/elf/dl-tunables.c b/elf/dl-tunables.c
> index 26e6e26612..7c9f1ca31f 100644
> --- a/elf/dl-tunables.c
> +++ b/elf/dl-tunables.c
> @@ -100,31 +100,39 @@ get_next_env (char **envp, char **name, size_t *namelen, char **val,
>      }      \
>  })
>  
> +#define TUNABLE_UPDATE_VAL(__cur, __val, __min, __max, __type)      \
> +({      \
> +  (__cur)->type.min = (__min);      \
> +  (__cur)->type.max = (__max);      \
> +  (__cur)->val.numval = (__val);      \
> +  (__cur)->initialized = true;      \
> +})

Can we update TUNABLE_SET_VAL_IF_VALID_RANGE to take the extra parameters?

If the value is not within [__min, __max] it should not update numval.

> +
>  static void
> -do_tunable_update_val (tunable_t *cur, const void *valp)
> +do_tunable_update_val (tunable_t *cur, const void *valp,
> +       const void *minp, const void *maxp)
>  {
> -  uint64_t val;
> +  uint64_t val, min, max;
>  
>    if (cur->type.type_code != TUNABLE_TYPE_STRING)
> -    val = *((int64_t *) valp);
> +    {
> +      val = *((int64_t *) valp);
> +      if (minp)
> + min = *((int64_t *) minp);
> +      if (maxp)
> + max = *((int64_t *) maxp);
> +    }

OK.

>  
>    switch (cur->type.type_code)
>      {
>      case TUNABLE_TYPE_INT_32:
> - {
> -  TUNABLE_SET_VAL_IF_VALID_RANGE (cur, val, int64_t);
> -  break;
> - }
>      case TUNABLE_TYPE_UINT_64:
> - {
> -  TUNABLE_SET_VAL_IF_VALID_RANGE (cur, val, uint64_t);
> -  break;
> - }
>      case TUNABLE_TYPE_SIZE_T:
> - {
> -  TUNABLE_SET_VAL_IF_VALID_RANGE (cur, val, uint64_t);
> -  break;
> - }
> +      if (minp && maxp)
> + TUNABLE_UPDATE_VAL (cur, val, min, max, int64_t);
> +      else
> + TUNABLE_SET_VAL_IF_VALID_RANGE (cur, val, int64_t);
> +      break;

Merging the two would result in a one call here and only one macro API
for set/update. I don't see the requirement for a different set and update.

>      case TUNABLE_TYPE_STRING:
>   {
>    cur->val.strval = valp;
> @@ -153,7 +161,7 @@ tunable_initialize (tunable_t *cur, const char *strval)
>        cur->initialized = true;
>        valp = strval;
>      }
> -  do_tunable_update_val (cur, valp);
> +  do_tunable_update_val (cur, valp, NULL, NULL);
>  }
>  
>  void
> @@ -161,8 +169,17 @@ __tunable_set_val (tunable_id_t id, void *valp)
>  {
>    tunable_t *cur = &tunable_list[id];
>  
> -  do_tunable_update_val (cur, valp);
> +  do_tunable_update_val (cur, valp, NULL, NULL);
> +}
> +
> +void
> +__tunable_update_val (tunable_id_t id, void *valp, void *minp, void *maxp)
> +{
> +  tunable_t *cur = &tunable_list[id];
> +
> +  do_tunable_update_val (cur, valp, minp, maxp);
>  }
> +rtld_hidden_def (__tunable_update_val)
>  
>  #if TUNABLES_FRONTEND == TUNABLES_FRONTEND_valstring
>  /* Parse the tunable string TUNESTR and adjust it to drop any tunables that may
> diff --git a/elf/dl-tunables.h b/elf/dl-tunables.h
> index f05eb50c2f..f6bf7379af 100644
> --- a/elf/dl-tunables.h
> +++ b/elf/dl-tunables.h
> @@ -71,8 +71,10 @@ typedef struct _tunable tunable_t;
>  extern void __tunables_init (char **);
>  extern void __tunable_get_val (tunable_id_t, void *, tunable_callback_t);
>  extern void __tunable_set_val (tunable_id_t, void *);
> +extern void __tunable_update_val (tunable_id_t, void *, void *, void *);
>  rtld_hidden_proto (__tunables_init)
>  rtld_hidden_proto (__tunable_get_val)
> +rtld_hidden_proto (__tunable_update_val)
>  
>  /* Define TUNABLE_GET and TUNABLE_SET in short form if TOP_NAMESPACE and
>     TUNABLE_NAMESPACE are defined.  This is useful shorthand to get and set
> @@ -82,11 +84,16 @@ rtld_hidden_proto (__tunable_get_val)
>    TUNABLE_GET_FULL (TOP_NAMESPACE, TUNABLE_NAMESPACE, __id, __type, __cb)
>  # define TUNABLE_SET(__id, __type, __val) \
>    TUNABLE_SET_FULL (TOP_NAMESPACE, TUNABLE_NAMESPACE, __id, __type, __val)
> +# define TUNABLE_UPDATE(__id, __type, __val, __min, __max) \
> +  TUNABLE_UPDATE_FULL (TOP_NAMESPACE, TUNABLE_NAMESPACE, __id, __type, \
> +       __val, __min, __max)

Please use TUNABLE_SET?

I would like avoid needing new macro APIs.

There doesn't need to be symmetry between TUNABLE_GET and TUNABLE_SET.

The TUNABLE_GET is for getting the value of the tunable during subsystem
startup, and the tunable framework handles: setting the default value,
processing user settings, enforcing minimums, enforcing maximums etc.

The TUNABLE_SET side doesn't have to be symmetric it may need to set more
than just the tunable value e.g. min/max.

>  #else
>  # define TUNABLE_GET(__top, __ns, __id, __type, __cb) \
>    TUNABLE_GET_FULL (__top, __ns, __id, __type, __cb)
>  # define TUNABLE_SET(__top, __ns, __id, __type, __val) \
>    TUNABLE_SET_FULL (__top, __ns, __id, __type, __val)
> +# define TUNABLE_UPDATE(__top, __ns, __id, __type, __val, __min, __max) \
> +  TUNABLE_UPDATE_FULL (__top, __ns, __id, __type, __val, __min, __max)
>  #endif
>  
>  /* Get and return a tunable value.  If the tunable was set externally and __CB
> @@ -106,6 +113,14 @@ rtld_hidden_proto (__tunable_get_val)
>   & (__type) {__val});      \
>  })
>  
> +/* Update a tunable value.  */
> +# define TUNABLE_UPDATE_FULL(__top, __ns, __id, __type, __val, __min, __max) \
> +({      \
> +  __tunable_update_val (TUNABLE_ENUM_NAME (__top, __ns, __id),      \
> + & (__type) {__val},  & (__type) {__min},      \
> + & (__type) {__max});      \
> +})
> +
>  /* Namespace sanity for callback functions.  Use this macro to keep the
>     namespace of the modules clean.  */
>  # define TUNABLE_CALLBACK(__name) _dl_tunable_ ## __name
> diff --git a/sysdeps/i386/cacheinfo.c b/sysdeps/i386/cacheinfo.c
> deleted file mode 100644
> index f15fe0779a..0000000000
> --- a/sysdeps/i386/cacheinfo.c
> +++ /dev/null
> @@ -1,3 +0,0 @@
> -#define DISABLE_PREFETCHW
> -
> -#include <sysdeps/x86/cacheinfo.c>
> diff --git a/sysdeps/x86/Makefile b/sysdeps/x86/Makefile
> index beab426f67..0872e0e655 100644
> --- a/sysdeps/x86/Makefile
> +++ b/sysdeps/x86/Makefile
> @@ -3,7 +3,7 @@ gen-as-const-headers += cpu-features-offsets.sym
>  endif
>  
>  ifeq ($(subdir),elf)
> -sysdep-dl-routines += dl-get-cpu-features
> +sysdep-dl-routines += dl-get-cpu-features dl-cacheinfo
>  
>  tests += tst-get-cpu-features tst-get-cpu-features-static
>  tests-static += tst-get-cpu-features-static
> diff --git a/sysdeps/x86/cacheinfo.c b/sysdeps/x86/cacheinfo.c
> index 311502dee3..8c4c7f9972 100644
> --- a/sysdeps/x86/cacheinfo.c
> +++ b/sysdeps/x86/cacheinfo.c
> @@ -18,498 +18,9 @@
>  
>  #if IS_IN (libc)
>  
> -#include <assert.h>
> -#include <stdbool.h>
> -#include <stdlib.h>
>  #include <unistd.h>
> -#include <cpuid.h>
>  #include <init-arch.h>
>  
> -static const struct intel_02_cache_info
> -{
> -  unsigned char idx;
> -  unsigned char assoc;
> -  unsigned char linesize;
> -  unsigned char rel_name;
> -  unsigned int size;
> -} intel_02_known [] =
> -  {
> -#define M(sc) ((sc) - _SC_LEVEL1_ICACHE_SIZE)
> -    { 0x06,  4, 32, M(_SC_LEVEL1_ICACHE_SIZE),    8192 },
> -    { 0x08,  4, 32, M(_SC_LEVEL1_ICACHE_SIZE),   16384 },
> -    { 0x09,  4, 32, M(_SC_LEVEL1_ICACHE_SIZE),   32768 },
> -    { 0x0a,  2, 32, M(_SC_LEVEL1_DCACHE_SIZE),    8192 },
> -    { 0x0c,  4, 32, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
> -    { 0x0d,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
> -    { 0x0e,  6, 64, M(_SC_LEVEL1_DCACHE_SIZE),   24576 },
> -    { 0x21,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
> -    { 0x22,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),   524288 },
> -    { 0x23,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  1048576 },
> -    { 0x25,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
> -    { 0x29,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
> -    { 0x2c,  8, 64, M(_SC_LEVEL1_DCACHE_SIZE),   32768 },
> -    { 0x30,  8, 64, M(_SC_LEVEL1_ICACHE_SIZE),   32768 },
> -    { 0x39,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
> -    { 0x3a,  6, 64, M(_SC_LEVEL2_CACHE_SIZE),   196608 },
> -    { 0x3b,  2, 64, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
> -    { 0x3c,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
> -    { 0x3d,  6, 64, M(_SC_LEVEL2_CACHE_SIZE),   393216 },
> -    { 0x3e,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
> -    { 0x3f,  2, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
> -    { 0x41,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
> -    { 0x42,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
> -    { 0x43,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
> -    { 0x44,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
> -    { 0x45,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),  2097152 },
> -    { 0x46,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
> -    { 0x47,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
> -    { 0x48, 12, 64, M(_SC_LEVEL2_CACHE_SIZE),  3145728 },
> -    { 0x49, 16, 64, M(_SC_LEVEL2_CACHE_SIZE),  4194304 },
> -    { 0x4a, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  6291456 },
> -    { 0x4b, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
> -    { 0x4c, 12, 64, M(_SC_LEVEL3_CACHE_SIZE), 12582912 },
> -    { 0x4d, 16, 64, M(_SC_LEVEL3_CACHE_SIZE), 16777216 },
> -    { 0x4e, 24, 64, M(_SC_LEVEL2_CACHE_SIZE),  6291456 },
> -    { 0x60,  8, 64, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
> -    { 0x66,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),    8192 },
> -    { 0x67,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
> -    { 0x68,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),   32768 },
> -    { 0x78,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
> -    { 0x79,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
> -    { 0x7a,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
> -    { 0x7b,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
> -    { 0x7c,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
> -    { 0x7d,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  2097152 },
> -    { 0x7f,  2, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
> -    { 0x80,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
> -    { 0x82,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
> -    { 0x83,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
> -    { 0x84,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
> -    { 0x85,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),  2097152 },
> -    { 0x86,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
> -    { 0x87,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
> -    { 0xd0,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),   524288 },
> -    { 0xd1,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),  1048576 },
> -    { 0xd2,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
> -    { 0xd6,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  1048576 },
> -    { 0xd7,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
> -    { 0xd8,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
> -    { 0xdc, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
> -    { 0xdd, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
> -    { 0xde, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
> -    { 0xe2, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
> -    { 0xe3, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
> -    { 0xe4, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
> -    { 0xea, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 12582912 },
> -    { 0xeb, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 18874368 },
> -    { 0xec, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 25165824 },
> -  };
> -
> -#define nintel_02_known (sizeof (intel_02_known) / sizeof (intel_02_known [0]))
> -
> -static int
> -intel_02_known_compare (const void *p1, const void *p2)
> -{
> -  const struct intel_02_cache_info *i1;
> -  const struct intel_02_cache_info *i2;
> -
> -  i1 = (const struct intel_02_cache_info *) p1;
> -  i2 = (const struct intel_02_cache_info *) p2;
> -
> -  if (i1->idx == i2->idx)
> -    return 0;
> -
> -  return i1->idx < i2->idx ? -1 : 1;
> -}
> -
> -
> -static long int
> -__attribute__ ((noinline))
> -intel_check_word (int name, unsigned int value, bool *has_level_2,
> -  bool *no_level_2_or_3,
> -  const struct cpu_features *cpu_features)
> -{
> -  if ((value & 0x80000000) != 0)
> -    /* The register value is reserved.  */
> -    return 0;
> -
> -  /* Fold the name.  The _SC_ constants are always in the order SIZE,
> -     ASSOC, LINESIZE.  */
> -  int folded_rel_name = (M(name) / 3) * 3;
> -
> -  while (value != 0)
> -    {
> -      unsigned int byte = value & 0xff;
> -
> -      if (byte == 0x40)
> - {
> -  *no_level_2_or_3 = true;
> -
> -  if (folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))
> -    /* No need to look further.  */
> -    break;
> - }
> -      else if (byte == 0xff)
> - {
> -  /* CPUID leaf 0x4 contains all the information.  We need to
> -     iterate over it.  */
> -  unsigned int eax;
> -  unsigned int ebx;
> -  unsigned int ecx;
> -  unsigned int edx;
> -
> -  unsigned int round = 0;
> -  while (1)
> -    {
> -      __cpuid_count (4, round, eax, ebx, ecx, edx);
> -
> -      enum { null = 0, data = 1, inst = 2, uni = 3 } type = eax & 0x1f;
> -      if (type == null)
> - /* That was the end.  */
> - break;
> -
> -      unsigned int level = (eax >> 5) & 0x7;
> -
> -      if ((level == 1 && type == data
> -   && folded_rel_name == M(_SC_LEVEL1_DCACHE_SIZE))
> -  || (level == 1 && type == inst
> -      && folded_rel_name == M(_SC_LEVEL1_ICACHE_SIZE))
> -  || (level == 2 && folded_rel_name == M(_SC_LEVEL2_CACHE_SIZE))
> -  || (level == 3 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))
> -  || (level == 4 && folded_rel_name == M(_SC_LEVEL4_CACHE_SIZE)))
> - {
> -  unsigned int offset = M(name) - folded_rel_name;
> -
> -  if (offset == 0)
> -    /* Cache size.  */
> -    return (((ebx >> 22) + 1)
> -    * (((ebx >> 12) & 0x3ff) + 1)
> -    * ((ebx & 0xfff) + 1)
> -    * (ecx + 1));
> -  if (offset == 1)
> -    return (ebx >> 22) + 1;
> -
> -  assert (offset == 2);
> -  return (ebx & 0xfff) + 1;
> - }
> -
> -      ++round;
> -    }
> -  /* There is no other cache information anywhere else.  */
> -  break;
> - }
> -      else
> - {
> -  if (byte == 0x49 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))
> -    {
> -      /* Intel reused this value.  For family 15, model 6 it
> - specifies the 3rd level cache.  Otherwise the 2nd
> - level cache.  */
> -      unsigned int family = cpu_features->basic.family;
> -      unsigned int model = cpu_features->basic.model;
> -
> -      if (family == 15 && model == 6)
> - {
> -  /* The level 3 cache is encoded for this model like
> -     the level 2 cache is for other models.  Pretend
> -     the caller asked for the level 2 cache.  */
> -  name = (_SC_LEVEL2_CACHE_SIZE
> -  + (name - _SC_LEVEL3_CACHE_SIZE));
> -  folded_rel_name = M(_SC_LEVEL2_CACHE_SIZE);
> - }
> -    }
> -
> -  struct intel_02_cache_info *found;
> -  struct intel_02_cache_info search;
> -
> -  search.idx = byte;
> -  found = bsearch (&search, intel_02_known, nintel_02_known,
> -   sizeof (intel_02_known[0]), intel_02_known_compare);
> -  if (found != NULL)
> -    {
> -      if (found->rel_name == folded_rel_name)
> - {
> -  unsigned int offset = M(name) - folded_rel_name;
> -
> -  if (offset == 0)
> -    /* Cache size.  */
> -    return found->size;
> -  if (offset == 1)
> -    return found->assoc;
> -
> -  assert (offset == 2);
> -  return found->linesize;
> - }
> -
> -      if (found->rel_name == M(_SC_LEVEL2_CACHE_SIZE))
> - *has_level_2 = true;
> -    }
> - }
> -
> -      /* Next byte for the next round.  */
> -      value >>= 8;
> -    }
> -
> -  /* Nothing found.  */
> -  return 0;
> -}
> -
> -
> -static long int __attribute__ ((noinline))
> -handle_intel (int name, const struct cpu_features *cpu_features)
> -{
> -  unsigned int maxidx = cpu_features->basic.max_cpuid;
> -
> -  /* Return -1 for older CPUs.  */
> -  if (maxidx < 2)
> -    return -1;
> -
> -  /* OK, we can use the CPUID instruction to get all info about the
> -     caches.  */
> -  unsigned int cnt = 0;
> -  unsigned int max = 1;
> -  long int result = 0;
> -  bool no_level_2_or_3 = false;
> -  bool has_level_2 = false;
> -
> -  while (cnt++ < max)
> -    {
> -      unsigned int eax;
> -      unsigned int ebx;
> -      unsigned int ecx;
> -      unsigned int edx;
> -      __cpuid (2, eax, ebx, ecx, edx);
> -
> -      /* The low byte of EAX in the first round contain the number of
> - rounds we have to make.  At least one, the one we are already
> - doing.  */
> -      if (cnt == 1)
> - {
> -  max = eax & 0xff;
> -  eax &= 0xffffff00;
> - }
> -
> -      /* Process the individual registers' value.  */
> -      result = intel_check_word (name, eax, &has_level_2,
> - &no_level_2_or_3, cpu_features);
> -      if (result != 0)
> - return result;
> -
> -      result = intel_check_word (name, ebx, &has_level_2,
> - &no_level_2_or_3, cpu_features);
> -      if (result != 0)
> - return result;
> -
> -      result = intel_check_word (name, ecx, &has_level_2,
> - &no_level_2_or_3, cpu_features);
> -      if (result != 0)
> - return result;
> -
> -      result = intel_check_word (name, edx, &has_level_2,
> - &no_level_2_or_3, cpu_features);
> -      if (result != 0)
> - return result;
> -    }
> -
> -  if (name >= _SC_LEVEL2_CACHE_SIZE && name <= _SC_LEVEL3_CACHE_LINESIZE
> -      && no_level_2_or_3)
> -    return -1;
> -
> -  return 0;
> -}
> -
> -
> -static long int __attribute__ ((noinline))
> -handle_amd (int name)
> -{
> -  unsigned int eax;
> -  unsigned int ebx;
> -  unsigned int ecx;
> -  unsigned int edx;
> -  __cpuid (0x80000000, eax, ebx, ecx, edx);
> -
> -  /* No level 4 cache (yet).  */
> -  if (name > _SC_LEVEL3_CACHE_LINESIZE)
> -    return 0;
> -
> -  unsigned int fn = 0x80000005 + (name >= _SC_LEVEL2_CACHE_SIZE);
> -  if (eax < fn)
> -    return 0;
> -
> -  __cpuid (fn, eax, ebx, ecx, edx);
> -
> -  if (name < _SC_LEVEL1_DCACHE_SIZE)
> -    {
> -      name += _SC_LEVEL1_DCACHE_SIZE - _SC_LEVEL1_ICACHE_SIZE;
> -      ecx = edx;
> -    }
> -
> -  switch (name)
> -    {
> -    case _SC_LEVEL1_DCACHE_SIZE:
> -      return (ecx >> 14) & 0x3fc00;
> -
> -    case _SC_LEVEL1_DCACHE_ASSOC:
> -      ecx >>= 16;
> -      if ((ecx & 0xff) == 0xff)
> - /* Fully associative.  */
> - return (ecx << 2) & 0x3fc00;
> -      return ecx & 0xff;
> -
> -    case _SC_LEVEL1_DCACHE_LINESIZE:
> -      return ecx & 0xff;
> -
> -    case _SC_LEVEL2_CACHE_SIZE:
> -      return (ecx & 0xf000) == 0 ? 0 : (ecx >> 6) & 0x3fffc00;
> -
> -    case _SC_LEVEL2_CACHE_ASSOC:
> -      switch ((ecx >> 12) & 0xf)
> - {
> - case 0:
> - case 1:
> - case 2:
> - case 4:
> -  return (ecx >> 12) & 0xf;
> - case 6:
> -  return 8;
> - case 8:
> -  return 16;
> - case 10:
> -  return 32;
> - case 11:
> -  return 48;
> - case 12:
> -  return 64;
> - case 13:
> -  return 96;
> - case 14:
> -  return 128;
> - case 15:
> -  return ((ecx >> 6) & 0x3fffc00) / (ecx & 0xff);
> - default:
> -  return 0;
> - }
> -      /* NOTREACHED */
> -
> -    case _SC_LEVEL2_CACHE_LINESIZE:
> -      return (ecx & 0xf000) == 0 ? 0 : ecx & 0xff;
> -
> -    case _SC_LEVEL3_CACHE_SIZE:
> -      return (edx & 0xf000) == 0 ? 0 : (edx & 0x3ffc0000) << 1;
> -
> -    case _SC_LEVEL3_CACHE_ASSOC:
> -      switch ((edx >> 12) & 0xf)
> - {
> - case 0:
> - case 1:
> - case 2:
> - case 4:
> -  return (edx >> 12) & 0xf;
> - case 6:
> -  return 8;
> - case 8:
> -  return 16;
> - case 10:
> -  return 32;
> - case 11:
> -  return 48;
> - case 12:
> -  return 64;
> - case 13:
> -  return 96;
> - case 14:
> -  return 128;
> - case 15:
> -  return ((edx & 0x3ffc0000) << 1) / (edx & 0xff);
> - default:
> -  return 0;
> - }
> -      /* NOTREACHED */
> -
> -    case _SC_LEVEL3_CACHE_LINESIZE:
> -      return (edx & 0xf000) == 0 ? 0 : edx & 0xff;
> -
> -    default:
> -      assert (! "cannot happen");
> -    }
> -  return -1;
> -}
> -
> -
> -static long int __attribute__ ((noinline))
> -handle_zhaoxin (int name)
> -{
> -  unsigned int eax;
> -  unsigned int ebx;
> -  unsigned int ecx;
> -  unsigned int edx;
> -
> -  int folded_rel_name = (M(name) / 3) * 3;
> -
> -  unsigned int round = 0;
> -  while (1)
> -    {
> -      __cpuid_count (4, round, eax, ebx, ecx, edx);
> -
> -      enum { null = 0, data = 1, inst = 2, uni = 3 } type = eax & 0x1f;
> -      if (type == null)
> -        break;
> -
> -      unsigned int level = (eax >> 5) & 0x7;
> -
> -      if ((level == 1 && type == data
> -        && folded_rel_name == M(_SC_LEVEL1_DCACHE_SIZE))
> -        || (level == 1 && type == inst
> -            && folded_rel_name == M(_SC_LEVEL1_ICACHE_SIZE))
> -        || (level == 2 && folded_rel_name == M(_SC_LEVEL2_CACHE_SIZE))
> -        || (level == 3 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE)))
> -        {
> -          unsigned int offset = M(name) - folded_rel_name;
> -
> -          if (offset == 0)
> -            /* Cache size.  */
> -            return (((ebx >> 22) + 1)
> -                * (((ebx >> 12) & 0x3ff) + 1)
> -                * ((ebx & 0xfff) + 1)
> -                * (ecx + 1));
> -          if (offset == 1)
> -            return (ebx >> 22) + 1;
> -
> -          assert (offset == 2);
> -          return (ebx & 0xfff) + 1;
> -        }
> -
> -      ++round;
> -    }
> -
> -  /* Nothing found.  */
> -  return 0;
> -}
> -
> -
> -/* Get the value of the system variable NAME.  */
> -long int
> -attribute_hidden
> -__cache_sysconf (int name)
> -{
> -  const struct cpu_features *cpu_features = __get_cpu_features ();
> -
> -  if (cpu_features->basic.kind == arch_kind_intel)
> -    return handle_intel (name, cpu_features);
> -
> -  if (cpu_features->basic.kind == arch_kind_amd)
> -    return handle_amd (name);
> -
> -  if (cpu_features->basic.kind == arch_kind_zhaoxin)
> -    return handle_zhaoxin (name);
> -
> -  // XXX Fill in more vendors.
> -
> -  /* CPU not known, we have no information.  */
> -  return 0;
> -}
> -
> -

OK.

>  /* Data cache size for use in memory and string routines, typically
>     L1 size, rounded to multiple of 256 bytes.  */
>  long int __x86_data_cache_size_half attribute_hidden = 32 * 1024 / 2;
> @@ -530,348 +41,85 @@ long int __x86_raw_shared_cache_size attribute_hidden = 1024 * 1024;
>  /* Threshold to use non temporal store.  */
>  long int __x86_shared_non_temporal_threshold attribute_hidden;
>  
> -#ifndef DISABLE_PREFETCHW
> +#ifndef __x86_64__
>  /* PREFETCHW support flag for use in memory and string routines.  */
>  int __x86_prefetchw attribute_hidden;
>  #endif
>  
> -
> -static void
> -get_common_cache_info (long int *shared_ptr, unsigned int *threads_ptr,
> -                long int core)
> +/* Get the value of the system variable NAME.  */
> +long int
> +attribute_hidden
> +__cache_sysconf (int name)
>  {
> -  unsigned int eax;
> -  unsigned int ebx;
> -  unsigned int ecx;
> -  unsigned int edx;
> -
> -  /* Number of logical processors sharing L2 cache.  */
> -  int threads_l2;
> -
> -  /* Number of logical processors sharing L3 cache.  */
> -  int threads_l3;
> -
>    const struct cpu_features *cpu_features = __get_cpu_features ();
> -  int max_cpuid = cpu_features->basic.max_cpuid;
> -  unsigned int family = cpu_features->basic.family;
> -  unsigned int model = cpu_features->basic.model;
> -  long int shared = *shared_ptr;
> -  unsigned int threads = *threads_ptr;
> -  bool inclusive_cache = true;
> -  bool support_count_mask = true;
> -
> -  /* Try L3 first.  */
> -  unsigned int level = 3;
> -
> -  if (cpu_features->basic.kind == arch_kind_zhaoxin && family == 6)
> -    support_count_mask = false;
> -
> -  if (shared <= 0)
> -    {
> -      /* Try L2 otherwise.  */
> -      level  = 2;
> -      shared = core;
> -      threads_l2 = 0;
> -      threads_l3 = -1;
> -    }
> -  else
> -    {
> -      threads_l2 = 0;
> -      threads_l3 = 0;
> -    }
> -
> -  /* A value of 0 for the HTT bit indicates there is only a single
> -     logical processor.  */
> -  if (HAS_CPU_FEATURE (HTT))
> +  switch (name)
>      {
> -      /* Figure out the number of logical threads that share the
> -         highest cache level.  */
> -      if (max_cpuid >= 4)
> -        {
> -          int i = 0;
> -
> -          /* Query until cache level 2 and 3 are enumerated.  */
> -          int check = 0x1 | (threads_l3 == 0) << 1;
> -          do
> -            {
> -              __cpuid_count (4, i++, eax, ebx, ecx, edx);
> +    case _SC_LEVEL1_ICACHE_SIZE:
> +      return cpu_features->level1_icache_size;
>  
> -              /* There seems to be a bug in at least some Pentium Ds
> -                 which sometimes fail to iterate all cache parameters.
> -                 Do not loop indefinitely here, stop in this case and
> -                 assume there is no such information.  */
> -              if (cpu_features->basic.kind == arch_kind_intel
> -                  && (eax & 0x1f) == 0 )
> -                goto intel_bug_no_cache_info;
> +    case _SC_LEVEL1_DCACHE_SIZE:
> +      return cpu_features->level1_dcache_size;
>  
> -              switch ((eax >> 5) & 0x7)
> -                {
> -                  default:
> -                    break;
> -                  case 2:
> -                    if ((check & 0x1))
> -                      {
> -                        /* Get maximum number of logical processors
> -                           sharing L2 cache.  */
> -                        threads_l2 = (eax >> 14) & 0x3ff;
> -                        check &= ~0x1;
> -                      }
> -                    break;
> -                  case 3:
> -                    if ((check & (0x1 << 1)))
> -                      {
> -                        /* Get maximum number of logical processors
> -                           sharing L3 cache.  */
> -                        threads_l3 = (eax >> 14) & 0x3ff;
> +    case _SC_LEVEL1_DCACHE_ASSOC:
> +      return cpu_features->level1_dcache_assoc;
>  
> -                        /* Check if L2 and L3 caches are inclusive.  */
> -                        inclusive_cache = (edx & 0x2) != 0;
> -                        check &= ~(0x1 << 1);
> -                      }
> -                    break;
> -                }
> -            }
> -          while (check);
> +    case _SC_LEVEL1_DCACHE_LINESIZE:
> +      return cpu_features->level1_dcache_linesize;
>  
> -          /* If max_cpuid >= 11, THREADS_L2/THREADS_L3 are the maximum
> -             numbers of addressable IDs for logical processors sharing
> -             the cache, instead of the maximum number of threads
> -             sharing the cache.  */
> -          if (max_cpuid >= 11 && support_count_mask)
> -            {
> -              /* Find the number of logical processors shipped in
> -                 one core and apply count mask.  */
> -              i = 0;
> +    case _SC_LEVEL2_CACHE_SIZE:
> +      return cpu_features->level2_cache_size;
>  
> -              /* Count SMT only if there is L3 cache.  Always count
> -                 core if there is no L3 cache.  */
> -              int count = ((threads_l2 > 0 && level == 3)
> -                           | ((threads_l3 > 0
> -                               || (threads_l2 > 0 && level == 2)) << 1));
> +    case _SC_LEVEL2_CACHE_ASSOC:
> +      return cpu_features->level2_cache_assoc;
>  
> -              while (count)
> -                {
> -                  __cpuid_count (11, i++, eax, ebx, ecx, edx);
> +    case _SC_LEVEL2_CACHE_LINESIZE:
> +      return cpu_features->level2_cache_linesize;
>  
> -                  int shipped = ebx & 0xff;
> -                  int type = ecx & 0xff00;
> -                  if (shipped == 0 || type == 0)
> -                    break;
> -                  else if (type == 0x100)
> -                    {
> -                      /* Count SMT.  */
> -                      if ((count & 0x1))
> -                        {
> -                          int count_mask;
> +    case _SC_LEVEL3_CACHE_SIZE:
> +      return cpu_features->level3_cache_size;
>  
> -                          /* Compute count mask.  */
> -                          asm ("bsr %1, %0"
> -                               : "=r" (count_mask) : "g" (threads_l2));
> -                          count_mask = ~(-1 << (count_mask + 1));
> -                          threads_l2 = (shipped - 1) & count_mask;
> -                          count &= ~0x1;
> -                        }
> -                    }
> -                  else if (type == 0x200)
> -                    {
> -                      /* Count core.  */
> -                      if ((count & (0x1 << 1)))
> -                        {
> -                          int count_mask;
> -                          int threads_core
> -                            = (level == 2 ? threads_l2 : threads_l3);
> +    case _SC_LEVEL3_CACHE_ASSOC:
> +      return cpu_features->level3_cache_assoc;
>  
> -                          /* Compute count mask.  */
> -                          asm ("bsr %1, %0"
> -                               : "=r" (count_mask) : "g" (threads_core));
> -                          count_mask = ~(-1 << (count_mask + 1));
> -                          threads_core = (shipped - 1) & count_mask;
> -                          if (level == 2)
> -                            threads_l2 = threads_core;
> -                          else
> -                            threads_l3 = threads_core;
> -                          count &= ~(0x1 << 1);
> -                        }
> -                    }
> -                }
> -            }
> -          if (threads_l2 > 0)
> -            threads_l2 += 1;
> -          if (threads_l3 > 0)
> -            threads_l3 += 1;
> -          if (level == 2)
> -            {
> -              if (threads_l2)
> -                {
> -                  threads = threads_l2;
> -                  if (cpu_features->basic.kind == arch_kind_intel
> -                      && threads > 2
> -                      && family == 6)
> -                    switch (model)
> -                      {
> -                        case 0x37:
> -                        case 0x4a:
> -                        case 0x4d:
> -                        case 0x5a:
> -                        case 0x5d:
> -                          /* Silvermont has L2 cache shared by 2 cores.  */
> -                          threads = 2;
> -                          break;
> -                        default:
> -                          break;
> -                      }
> -                }
> -            }
> -          else if (threads_l3)
> -            threads = threads_l3;
> -        }
> -      else
> -        {
> -intel_bug_no_cache_info:
> -          /* Assume that all logical threads share the highest cache
> -             level.  */
> -          threads
> -            = ((cpu_features->cpuid[COMMON_CPUID_INDEX_1].ebx
> -                >> 16) & 0xff);
> -        }
> +    case _SC_LEVEL3_CACHE_LINESIZE:
> +      return cpu_features->level3_cache_linesize;
>  
> -        /* Cap usage of highest cache level to the number of supported
> -           threads.  */
> -        if (shared > 0 && threads > 0)
> -          shared /= threads;
> -    }
> +    case _SC_LEVEL4_CACHE_SIZE:
> +      return cpu_features->level4_cache_size;
>  
> -  /* Account for non-inclusive L2 and L3 caches.  */
> -  if (!inclusive_cache)
> -    {
> -      if (threads_l2 > 0)
> -        core /= threads_l2;
> -      shared += core;
> +    default:
> +      break;
>      }
> -
> -  *shared_ptr = shared;
> -  *threads_ptr = threads;
> +  return -1;
>  }
>  
> -
>  static void
>  __attribute__((constructor))

OK. I would eventually like to see this constructor go away.

>  init_cacheinfo (void)
>  {
> -  /* Find out what brand of processor.  */
> -  unsigned int ebx;
> -  unsigned int ecx;
> -  unsigned int edx;
> -  int max_cpuid_ex;
> -  long int data = -1;
> -  long int shared = -1;
> -  long int core;
> -  unsigned int threads = 0;
>    const struct cpu_features *cpu_features = __get_cpu_features ();
> +  long int data = cpu_features->data_cache_size;
> +  __x86_raw_data_cache_size_half = data / 2;
> +  __x86_raw_data_cache_size = data;
> +  /* Round data cache size to multiple of 256 bytes.  */
> +  data = data & ~255L;
> +  __x86_data_cache_size_half = data / 2;
> +  __x86_data_cache_size = data;
> +
> +  long int shared = cpu_features->shared_cache_size;
> +  __x86_raw_shared_cache_size_half = shared / 2;
> +  __x86_raw_shared_cache_size = shared;
> +  /* Round shared cache size to multiple of 256 bytes.  */
> +  shared = shared & ~255L;
> +  __x86_shared_cache_size_half = shared / 2;
> +  __x86_shared_cache_size = shared;
>  
> -  if (cpu_features->basic.kind == arch_kind_intel)
> -    {
> -      data = handle_intel (_SC_LEVEL1_DCACHE_SIZE, cpu_features);
> -      core = handle_intel (_SC_LEVEL2_CACHE_SIZE, cpu_features);
> -      shared = handle_intel (_SC_LEVEL3_CACHE_SIZE, cpu_features);
> -
> -      get_common_cache_info (&shared, &threads, core);
> -    }
> -  else if (cpu_features->basic.kind == arch_kind_zhaoxin)
> -    {
> -      data = handle_zhaoxin (_SC_LEVEL1_DCACHE_SIZE);
> -      core = handle_zhaoxin (_SC_LEVEL2_CACHE_SIZE);
> -      shared = handle_zhaoxin (_SC_LEVEL3_CACHE_SIZE);
> -
> -      get_common_cache_info (&shared, &threads, core);
> -    }
> -  else if (cpu_features->basic.kind == arch_kind_amd)
> -    {
> -      data   = handle_amd (_SC_LEVEL1_DCACHE_SIZE);
> -      long int core = handle_amd (_SC_LEVEL2_CACHE_SIZE);
> -      shared = handle_amd (_SC_LEVEL3_CACHE_SIZE);
> -
> -      /* Get maximum extended function. */
> -      __cpuid (0x80000000, max_cpuid_ex, ebx, ecx, edx);
> -
> -      if (shared <= 0)
> - /* No shared L3 cache.  All we have is the L2 cache.  */
> - shared = core;
> -      else
> - {
> -  /* Figure out the number of logical threads that share L3.  */
> -  if (max_cpuid_ex >= 0x80000008)
> -    {
> -      /* Get width of APIC ID.  */
> -      __cpuid (0x80000008, max_cpuid_ex, ebx, ecx, edx);
> -      threads = 1 << ((ecx >> 12) & 0x0f);
> -    }
> -
> -  if (threads == 0)
> -    {
> -      /* If APIC ID width is not available, use logical
> - processor count.  */
> -      __cpuid (0x00000001, max_cpuid_ex, ebx, ecx, edx);
> -
> -      if ((edx & (1 << 28)) != 0)
> - threads = (ebx >> 16) & 0xff;
> -    }
> -
> -  /* Cap usage of highest cache level to the number of
> -     supported threads.  */
> -  if (threads > 0)
> -    shared /= threads;
> -
> -  /* Account for exclusive L2 and L3 caches.  */
> -  shared += core;
> - }
> +  __x86_shared_non_temporal_threshold
> +    = cpu_features->non_temporal_threshold;
>  
> -#ifndef DISABLE_PREFETCHW
> -      if (max_cpuid_ex >= 0x80000001)
> - {
> -  unsigned int eax;
> -  __cpuid (0x80000001, eax, ebx, ecx, edx);
> -  /*  PREFETCHW     || 3DNow!  */
> -  if ((ecx & 0x100) || (edx & 0x80000000))
> -    __x86_prefetchw = -1;
> - }
> +#ifndef __x86_64__
> +  __x86_prefetchw = cpu_features->prefetchw;
>  #endif
> -    }
> -
> -  if (cpu_features->data_cache_size != 0)
> -    data = cpu_features->data_cache_size;
> -
> -  if (data > 0)
> -    {
> -      __x86_raw_data_cache_size_half = data / 2;
> -      __x86_raw_data_cache_size = data;
> -      /* Round data cache size to multiple of 256 bytes.  */
> -      data = data & ~255L;
> -      __x86_data_cache_size_half = data / 2;
> -      __x86_data_cache_size = data;
> -    }
> -
> -  if (cpu_features->shared_cache_size != 0)
> -    shared = cpu_features->shared_cache_size;
> -
> -  if (shared > 0)
> -    {
> -      __x86_raw_shared_cache_size_half = shared / 2;
> -      __x86_raw_shared_cache_size = shared;
> -      /* Round shared cache size to multiple of 256 bytes.  */
> -      shared = shared & ~255L;
> -      __x86_shared_cache_size_half = shared / 2;
> -      __x86_shared_cache_size = shared;
> -    }
> -
> -  /* The large memcpy micro benchmark in glibc shows that 6 times of
> -     shared cache size is the approximate value above which non-temporal
> -     store becomes faster on a 8-core processor.  This is the 3/4 of the
> -     total shared cache size.  */
> -  __x86_shared_non_temporal_threshold
> -    = (cpu_features->non_temporal_threshold != 0
> -       ? cpu_features->non_temporal_threshold
> -       : __x86_shared_cache_size * threads * 3 / 4);
>  }
>  
>  #endif
> diff --git a/sysdeps/x86/cpu-features.c b/sysdeps/x86/cpu-features.c
> index c351bdd54a..e718204c18 100644
> --- a/sysdeps/x86/cpu-features.c
> +++ b/sysdeps/x86/cpu-features.c
> @@ -19,6 +19,7 @@
>  #include <cpuid.h>
>  #include <cpu-features.h>
>  #include <dl-hwcap.h>
> +#include <init-arch.h>
>  #include <libc-pointer-arith.h>
>  
>  #if HAVE_TUNABLES
> @@ -602,20 +603,14 @@ no_cpuid:
>    cpu_features->basic.model = model;
>    cpu_features->basic.stepping = stepping;
>  
> +  __init_cacheinfo ();

OK.

> +
>  #if HAVE_TUNABLES
>    TUNABLE_GET (hwcaps, tunable_val_t *, TUNABLE_CALLBACK (set_hwcaps));
> -  cpu_features->non_temporal_threshold
> -    = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
> -  cpu_features->data_cache_size
> -    = TUNABLE_GET (x86_data_cache_size, long int, NULL);
> -  cpu_features->shared_cache_size
> -    = TUNABLE_GET (x86_shared_cache_size, long int, NULL);
> -#endif
> -
> -  /* Reuse dl_platform, dl_hwcap and dl_hwcap_mask for x86.  */
> -#if !HAVE_TUNABLES && defined SHARED
> -  /* The glibc.cpu.hwcap_mask tunable is initialized already, so no need to do
> -     this.  */
> +#elif defined SHARED
> +  /* Reuse dl_platform, dl_hwcap and dl_hwcap_mask for x86.  The
> +     glibc.cpu.hwcap_mask tunable is initialized already, so no
> +     need to do this.  */
>    GLRO(dl_hwcap_mask) = HWCAP_IMPORTANT;
>  #endif
>  
> diff --git a/sysdeps/x86/cpu-features.h b/sysdeps/x86/cpu-features.h
> index d66dc206f7..3aaed33cbc 100644
> --- a/sysdeps/x86/cpu-features.h
> +++ b/sysdeps/x86/cpu-features.h
> @@ -102,6 +102,32 @@ struct cpu_features
>    unsigned long int shared_cache_size;
>    /* Threshold to use non temporal store.  */
>    unsigned long int non_temporal_threshold;
> +  /* _SC_LEVEL1_ICACHE_SIZE.  */
> +  unsigned long int level1_icache_size;
> +  /* _SC_LEVEL1_DCACHE_SIZE.  */
> +  unsigned long int level1_dcache_size;
> +  /* _SC_LEVEL1_DCACHE_ASSOC.  */
> +  unsigned long int level1_dcache_assoc;
> +  /* _SC_LEVEL1_DCACHE_LINESIZE.  */
> +  unsigned long int level1_dcache_linesize;
> +  /* _SC_LEVEL2_CACHE_ASSOC.  */
> +  unsigned long int level2_cache_size;
> +  /* _SC_LEVEL2_DCACHE_ASSOC.  */
> +  unsigned long int level2_cache_assoc;
> +  /* _SC_LEVEL2_CACHE_LINESIZE.  */
> +  unsigned long int level2_cache_linesize;
> +  /* /_SC_LEVEL3_CACHE_SIZE.  */
> +  unsigned long int level3_cache_size;
> +  /* _SC_LEVEL3_CACHE_ASSOC.  */
> +  unsigned long int level3_cache_assoc;
> +  /* _SC_LEVEL3_CACHE_LINESIZE.  */
> +  unsigned long int level3_cache_linesize;
> +  /* /_SC_LEVEL4_CACHE_SIZE.  */
> +  unsigned long int level4_cache_size;
> +#ifndef __x86_64__
> +  /* PREFETCHW support flag for use in memory and string routines.  */
> +  unsigned long int prefetchw;
> +#endif
>  };
>  
>  /* Used from outside of glibc to get access to the CPU features
> diff --git a/sysdeps/x86/dl-cacheinfo.c b/sysdeps/x86/dl-cacheinfo.c
> new file mode 100644
> index 0000000000..8e2a6f552c
> --- /dev/null
> +++ b/sysdeps/x86/dl-cacheinfo.c
> @@ -0,0 +1,888 @@
> +/* x86 cache info.
> +   Copyright (C) 2020 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
> +   modify it under the terms of the GNU Lesser General Public
> +   License as published by the Free Software Foundation; either
> +   version 2.1 of the License, or (at your option) any later version.
> +
> +   The GNU C Library is distributed in the hope that it will be useful,
> +   but WITHOUT ANY WARRANTY; without even the implied warranty of
> +   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> +   Lesser General Public License for more details.
> +
> +   You should have received a copy of the GNU Lesser General Public
> +   License along with the GNU C Library; if not, see
> +   <https://www.gnu.org/licenses/>.  */
> +
> +#include <assert.h>
> +#include <stdbool.h>
> +#include <stdlib.h>
> +#include <unistd.h>
> +#include <cpuid.h>
> +#include <init-arch.h>
> +#if HAVE_TUNABLES
> +# define TUNABLE_NAMESPACE cpu
> +# include <elf/dl-tunables.h>
> +#endif
> +
> +static const struct intel_02_cache_info
> +{
> +  unsigned char idx;
> +  unsigned char assoc;
> +  unsigned char linesize;
> +  unsigned char rel_name;
> +  unsigned int size;
> +} intel_02_known [] =
> +  {
> +#define M(sc) ((sc) - _SC_LEVEL1_ICACHE_SIZE)
> +    { 0x06,  4, 32, M(_SC_LEVEL1_ICACHE_SIZE),    8192 },
> +    { 0x08,  4, 32, M(_SC_LEVEL1_ICACHE_SIZE),   16384 },
> +    { 0x09,  4, 32, M(_SC_LEVEL1_ICACHE_SIZE),   32768 },
> +    { 0x0a,  2, 32, M(_SC_LEVEL1_DCACHE_SIZE),    8192 },
> +    { 0x0c,  4, 32, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
> +    { 0x0d,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
> +    { 0x0e,  6, 64, M(_SC_LEVEL1_DCACHE_SIZE),   24576 },
> +    { 0x21,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
> +    { 0x22,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),   524288 },
> +    { 0x23,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  1048576 },
> +    { 0x25,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
> +    { 0x29,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
> +    { 0x2c,  8, 64, M(_SC_LEVEL1_DCACHE_SIZE),   32768 },
> +    { 0x30,  8, 64, M(_SC_LEVEL1_ICACHE_SIZE),   32768 },
> +    { 0x39,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
> +    { 0x3a,  6, 64, M(_SC_LEVEL2_CACHE_SIZE),   196608 },
> +    { 0x3b,  2, 64, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
> +    { 0x3c,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
> +    { 0x3d,  6, 64, M(_SC_LEVEL2_CACHE_SIZE),   393216 },
> +    { 0x3e,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
> +    { 0x3f,  2, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
> +    { 0x41,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
> +    { 0x42,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
> +    { 0x43,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
> +    { 0x44,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
> +    { 0x45,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),  2097152 },
> +    { 0x46,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
> +    { 0x47,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
> +    { 0x48, 12, 64, M(_SC_LEVEL2_CACHE_SIZE),  3145728 },
> +    { 0x49, 16, 64, M(_SC_LEVEL2_CACHE_SIZE),  4194304 },
> +    { 0x4a, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  6291456 },
> +    { 0x4b, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
> +    { 0x4c, 12, 64, M(_SC_LEVEL3_CACHE_SIZE), 12582912 },
> +    { 0x4d, 16, 64, M(_SC_LEVEL3_CACHE_SIZE), 16777216 },
> +    { 0x4e, 24, 64, M(_SC_LEVEL2_CACHE_SIZE),  6291456 },
> +    { 0x60,  8, 64, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
> +    { 0x66,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),    8192 },
> +    { 0x67,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
> +    { 0x68,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),   32768 },
> +    { 0x78,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
> +    { 0x79,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
> +    { 0x7a,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
> +    { 0x7b,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
> +    { 0x7c,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
> +    { 0x7d,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  2097152 },
> +    { 0x7f,  2, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
> +    { 0x80,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
> +    { 0x82,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
> +    { 0x83,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
> +    { 0x84,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
> +    { 0x85,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),  2097152 },
> +    { 0x86,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
> +    { 0x87,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
> +    { 0xd0,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),   524288 },
> +    { 0xd1,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),  1048576 },
> +    { 0xd2,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
> +    { 0xd6,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  1048576 },
> +    { 0xd7,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
> +    { 0xd8,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
> +    { 0xdc, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
> +    { 0xdd, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
> +    { 0xde, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
> +    { 0xe2, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
> +    { 0xe3, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
> +    { 0xe4, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
> +    { 0xea, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 12582912 },
> +    { 0xeb, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 18874368 },
> +    { 0xec, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 25165824 },
> +  };
> +
> +#define nintel_02_known (sizeof (intel_02_known) / sizeof (intel_02_known [0]))
> +
> +static int
> +intel_02_known_compare (const void *p1, const void *p2)
> +{
> +  const struct intel_02_cache_info *i1;
> +  const struct intel_02_cache_info *i2;
> +
> +  i1 = (const struct intel_02_cache_info *) p1;
> +  i2 = (const struct intel_02_cache_info *) p2;
> +
> +  if (i1->idx == i2->idx)
> +    return 0;
> +
> +  return i1->idx < i2->idx ? -1 : 1;
> +}
> +
> +
> +static long int
> +__attribute__ ((noinline))
> +intel_check_word (int name, unsigned int value, bool *has_level_2,
> +  bool *no_level_2_or_3,
> +  const struct cpu_features *cpu_features)
> +{
> +  if ((value & 0x80000000) != 0)
> +    /* The register value is reserved.  */
> +    return 0;
> +
> +  /* Fold the name.  The _SC_ constants are always in the order SIZE,
> +     ASSOC, LINESIZE.  */
> +  int folded_rel_name = (M(name) / 3) * 3;
> +
> +  while (value != 0)
> +    {
> +      unsigned int byte = value & 0xff;
> +
> +      if (byte == 0x40)
> + {
> +  *no_level_2_or_3 = true;
> +
> +  if (folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))
> +    /* No need to look further.  */
> +    break;
> + }
> +      else if (byte == 0xff)
> + {
> +  /* CPUID leaf 0x4 contains all the information.  We need to
> +     iterate over it.  */
> +  unsigned int eax;
> +  unsigned int ebx;
> +  unsigned int ecx;
> +  unsigned int edx;
> +
> +  unsigned int round = 0;
> +  while (1)
> +    {
> +      __cpuid_count (4, round, eax, ebx, ecx, edx);
> +
> +      enum { null = 0, data = 1, inst = 2, uni = 3 } type = eax & 0x1f;
> +      if (type == null)
> + /* That was the end.  */
> + break;
> +
> +      unsigned int level = (eax >> 5) & 0x7;
> +
> +      if ((level == 1 && type == data
> +   && folded_rel_name == M(_SC_LEVEL1_DCACHE_SIZE))
> +  || (level == 1 && type == inst
> +      && folded_rel_name == M(_SC_LEVEL1_ICACHE_SIZE))
> +  || (level == 2 && folded_rel_name == M(_SC_LEVEL2_CACHE_SIZE))
> +  || (level == 3 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))
> +  || (level == 4 && folded_rel_name == M(_SC_LEVEL4_CACHE_SIZE)))
> + {
> +  unsigned int offset = M(name) - folded_rel_name;
> +
> +  if (offset == 0)
> +    /* Cache size.  */
> +    return (((ebx >> 22) + 1)
> +    * (((ebx >> 12) & 0x3ff) + 1)
> +    * ((ebx & 0xfff) + 1)
> +    * (ecx + 1));
> +  if (offset == 1)
> +    return (ebx >> 22) + 1;
> +
> +  assert (offset == 2);
> +  return (ebx & 0xfff) + 1;
> + }
> +
> +      ++round;
> +    }
> +  /* There is no other cache information anywhere else.  */
> +  break;
> + }
> +      else
> + {
> +  if (byte == 0x49 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))
> +    {
> +      /* Intel reused this value.  For family 15, model 6 it
> + specifies the 3rd level cache.  Otherwise the 2nd
> + level cache.  */
> +      unsigned int family = cpu_features->basic.family;
> +      unsigned int model = cpu_features->basic.model;
> +
> +      if (family == 15 && model == 6)
> + {
> +  /* The level 3 cache is encoded for this model like
> +     the level 2 cache is for other models.  Pretend
> +     the caller asked for the level 2 cache.  */
> +  name = (_SC_LEVEL2_CACHE_SIZE
> +  + (name - _SC_LEVEL3_CACHE_SIZE));
> +  folded_rel_name = M(_SC_LEVEL2_CACHE_SIZE);
> + }
> +    }
> +
> +  struct intel_02_cache_info *found;
> +  struct intel_02_cache_info search;
> +
> +  search.idx = byte;
> +  found = bsearch (&search, intel_02_known, nintel_02_known,
> +   sizeof (intel_02_known[0]), intel_02_known_compare);
> +  if (found != NULL)
> +    {
> +      if (found->rel_name == folded_rel_name)
> + {
> +  unsigned int offset = M(name) - folded_rel_name;
> +
> +  if (offset == 0)
> +    /* Cache size.  */
> +    return found->size;
> +  if (offset == 1)
> +    return found->assoc;
> +
> +  assert (offset == 2);
> +  return found->linesize;
> + }
> +
> +      if (found->rel_name == M(_SC_LEVEL2_CACHE_SIZE))
> + *has_level_2 = true;
> +    }
> + }
> +
> +      /* Next byte for the next round.  */
> +      value >>= 8;
> +    }
> +
> +  /* Nothing found.  */
> +  return 0;
> +}
> +
> +
> +static long int __attribute__ ((noinline))
> +handle_intel (int name, const struct cpu_features *cpu_features)
> +{
> +  unsigned int maxidx = cpu_features->basic.max_cpuid;
> +
> +  /* Return -1 for older CPUs.  */
> +  if (maxidx < 2)
> +    return -1;
> +
> +  /* OK, we can use the CPUID instruction to get all info about the
> +     caches.  */
> +  unsigned int cnt = 0;
> +  unsigned int max = 1;
> +  long int result = 0;
> +  bool no_level_2_or_3 = false;
> +  bool has_level_2 = false;
> +
> +  while (cnt++ < max)
> +    {
> +      unsigned int eax;
> +      unsigned int ebx;
> +      unsigned int ecx;
> +      unsigned int edx;
> +      __cpuid (2, eax, ebx, ecx, edx);
> +
> +      /* The low byte of EAX in the first round contain the number of
> + rounds we have to make.  At least one, the one we are already
> + doing.  */
> +      if (cnt == 1)
> + {
> +  max = eax & 0xff;
> +  eax &= 0xffffff00;
> + }
> +
> +      /* Process the individual registers' value.  */
> +      result = intel_check_word (name, eax, &has_level_2,
> + &no_level_2_or_3, cpu_features);
> +      if (result != 0)
> + return result;
> +
> +      result = intel_check_word (name, ebx, &has_level_2,
> + &no_level_2_or_3, cpu_features);
> +      if (result != 0)
> + return result;
> +
> +      result = intel_check_word (name, ecx, &has_level_2,
> + &no_level_2_or_3, cpu_features);
> +      if (result != 0)
> + return result;
> +
> +      result = intel_check_word (name, edx, &has_level_2,
> + &no_level_2_or_3, cpu_features);
> +      if (result != 0)
> + return result;
> +    }
> +
> +  if (name >= _SC_LEVEL2_CACHE_SIZE && name <= _SC_LEVEL3_CACHE_LINESIZE
> +      && no_level_2_or_3)
> +    return -1;
> +
> +  return 0;
> +}
> +
> +
> +static long int __attribute__ ((noinline))
> +handle_amd (int name)
> +{
> +  unsigned int eax;
> +  unsigned int ebx;
> +  unsigned int ecx;
> +  unsigned int edx;
> +  __cpuid (0x80000000, eax, ebx, ecx, edx);
> +
> +  /* No level 4 cache (yet).  */
> +  if (name > _SC_LEVEL3_CACHE_LINESIZE)
> +    return 0;
> +
> +  unsigned int fn = 0x80000005 + (name >= _SC_LEVEL2_CACHE_SIZE);
> +  if (eax < fn)
> +    return 0;
> +
> +  __cpuid (fn, eax, ebx, ecx, edx);
> +
> +  if (name < _SC_LEVEL1_DCACHE_SIZE)
> +    {
> +      name += _SC_LEVEL1_DCACHE_SIZE - _SC_LEVEL1_ICACHE_SIZE;
> +      ecx = edx;
> +    }
> +
> +  switch (name)
> +    {
> +    case _SC_LEVEL1_DCACHE_SIZE:
> +      return (ecx >> 14) & 0x3fc00;
> +
> +    case _SC_LEVEL1_DCACHE_ASSOC:
> +      ecx >>= 16;
> +      if ((ecx & 0xff) == 0xff)
> + /* Fully associative.  */
> + return (ecx << 2) & 0x3fc00;
> +      return ecx & 0xff;
> +
> +    case _SC_LEVEL1_DCACHE_LINESIZE:
> +      return ecx & 0xff;
> +
> +    case _SC_LEVEL2_CACHE_SIZE:
> +      return (ecx & 0xf000) == 0 ? 0 : (ecx >> 6) & 0x3fffc00;
> +
> +    case _SC_LEVEL2_CACHE_ASSOC:
> +      switch ((ecx >> 12) & 0xf)
> + {
> + case 0:
> + case 1:
> + case 2:
> + case 4:
> +  return (ecx >> 12) & 0xf;
> + case 6:
> +  return 8;
> + case 8:
> +  return 16;
> + case 10:
> +  return 32;
> + case 11:
> +  return 48;
> + case 12:
> +  return 64;
> + case 13:
> +  return 96;
> + case 14:
> +  return 128;
> + case 15:
> +  return ((ecx >> 6) & 0x3fffc00) / (ecx & 0xff);
> + default:
> +  return 0;
> + }
> +      /* NOTREACHED */
> +
> +    case _SC_LEVEL2_CACHE_LINESIZE:
> +      return (ecx & 0xf000) == 0 ? 0 : ecx & 0xff;
> +
> +    case _SC_LEVEL3_CACHE_SIZE:
> +      return (edx & 0xf000) == 0 ? 0 : (edx & 0x3ffc0000) << 1;
> +
> +    case _SC_LEVEL3_CACHE_ASSOC:
> +      switch ((edx >> 12) & 0xf)
> + {
> + case 0:
> + case 1:
> + case 2:
> + case 4:
> +  return (edx >> 12) & 0xf;
> + case 6:
> +  return 8;
> + case 8:
> +  return 16;
> + case 10:
> +  return 32;
> + case 11:
> +  return 48;
> + case 12:
> +  return 64;
> + case 13:
> +  return 96;
> + case 14:
> +  return 128;
> + case 15:
> +  return ((edx & 0x3ffc0000) << 1) / (edx & 0xff);
> + default:
> +  return 0;
> + }
> +      /* NOTREACHED */
> +
> +    case _SC_LEVEL3_CACHE_LINESIZE:
> +      return (edx & 0xf000) == 0 ? 0 : edx & 0xff;
> +
> +    default:
> +      assert (! "cannot happen");
> +    }
> +  return -1;
> +}
> +
> +
> +static long int __attribute__ ((noinline))
> +handle_zhaoxin (int name)
> +{
> +  unsigned int eax;
> +  unsigned int ebx;
> +  unsigned int ecx;
> +  unsigned int edx;
> +
> +  int folded_rel_name = (M(name) / 3) * 3;
> +
> +  unsigned int round = 0;
> +  while (1)
> +    {
> +      __cpuid_count (4, round, eax, ebx, ecx, edx);
> +
> +      enum { null = 0, data = 1, inst = 2, uni = 3 } type = eax & 0x1f;
> +      if (type == null)
> +        break;
> +
> +      unsigned int level = (eax >> 5) & 0x7;
> +
> +      if ((level == 1 && type == data
> +        && folded_rel_name == M(_SC_LEVEL1_DCACHE_SIZE))
> +        || (level == 1 && type == inst
> +            && folded_rel_name == M(_SC_LEVEL1_ICACHE_SIZE))
> +        || (level == 2 && folded_rel_name == M(_SC_LEVEL2_CACHE_SIZE))
> +        || (level == 3 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE)))
> +        {
> +          unsigned int offset = M(name) - folded_rel_name;
> +
> +          if (offset == 0)
> +            /* Cache size.  */
> +            return (((ebx >> 22) + 1)
> +                * (((ebx >> 12) & 0x3ff) + 1)
> +                * ((ebx & 0xfff) + 1)
> +                * (ecx + 1));
> +          if (offset == 1)
> +            return (ebx >> 22) + 1;
> +
> +          assert (offset == 2);
> +          return (ebx & 0xfff) + 1;
> +        }
> +
> +      ++round;
> +    }
> +
> +  /* Nothing found.  */
> +  return 0;
> +}
> +
> +
> +static void
> +get_common_cache_info (long int *shared_ptr, unsigned int *threads_ptr,
> +                long int core)
> +{
> +  unsigned int eax;
> +  unsigned int ebx;
> +  unsigned int ecx;
> +  unsigned int edx;
> +
> +  /* Number of logical processors sharing L2 cache.  */
> +  int threads_l2;
> +
> +  /* Number of logical processors sharing L3 cache.  */
> +  int threads_l3;
> +
> +  const struct cpu_features *cpu_features = __get_cpu_features ();
> +  int max_cpuid = cpu_features->basic.max_cpuid;
> +  unsigned int family = cpu_features->basic.family;
> +  unsigned int model = cpu_features->basic.model;
> +  long int shared = *shared_ptr;
> +  unsigned int threads = *threads_ptr;
> +  bool inclusive_cache = true;
> +  bool support_count_mask = true;
> +
> +  /* Try L3 first.  */
> +  unsigned int level = 3;
> +
> +  if (cpu_features->basic.kind == arch_kind_zhaoxin && family == 6)
> +    support_count_mask = false;
> +
> +  if (shared <= 0)
> +    {
> +      /* Try L2 otherwise.  */
> +      level  = 2;
> +      shared = core;
> +      threads_l2 = 0;
> +      threads_l3 = -1;
> +    }
> +  else
> +    {
> +      threads_l2 = 0;
> +      threads_l3 = 0;
> +    }
> +
> +  /* A value of 0 for the HTT bit indicates there is only a single
> +     logical processor.  */
> +  if (HAS_CPU_FEATURE (HTT))
> +    {
> +      /* Figure out the number of logical threads that share the
> +         highest cache level.  */
> +      if (max_cpuid >= 4)
> +        {
> +          int i = 0;
> +
> +          /* Query until cache level 2 and 3 are enumerated.  */
> +          int check = 0x1 | (threads_l3 == 0) << 1;
> +          do
> +            {
> +              __cpuid_count (4, i++, eax, ebx, ecx, edx);
> +
> +              /* There seems to be a bug in at least some Pentium Ds
> +                 which sometimes fail to iterate all cache parameters.
> +                 Do not loop indefinitely here, stop in this case and
> +                 assume there is no such information.  */
> +              if (cpu_features->basic.kind == arch_kind_intel
> +                  && (eax & 0x1f) == 0 )
> +                goto intel_bug_no_cache_info;
> +
> +              switch ((eax >> 5) & 0x7)
> +                {
> +                  default:
> +                    break;
> +                  case 2:
> +                    if ((check & 0x1))
> +                      {
> +                        /* Get maximum number of logical processors
> +                           sharing L2 cache.  */
> +                        threads_l2 = (eax >> 14) & 0x3ff;
> +                        check &= ~0x1;
> +                      }
> +                    break;
> +                  case 3:
> +                    if ((check & (0x1 << 1)))
> +                      {
> +                        /* Get maximum number of logical processors
> +                           sharing L3 cache.  */
> +                        threads_l3 = (eax >> 14) & 0x3ff;
> +
> +                        /* Check if L2 and L3 caches are inclusive.  */
> +                        inclusive_cache = (edx & 0x2) != 0;
> +                        check &= ~(0x1 << 1);
> +                      }
> +                    break;
> +                }
> +            }
> +          while (check);
> +
> +          /* If max_cpuid >= 11, THREADS_L2/THREADS_L3 are the maximum
> +             numbers of addressable IDs for logical processors sharing
> +             the cache, instead of the maximum number of threads
> +             sharing the cache.  */
> +          if (max_cpuid >= 11 && support_count_mask)
> +            {
> +              /* Find the number of logical processors shipped in
> +                 one core and apply count mask.  */
> +              i = 0;
> +
> +              /* Count SMT only if there is L3 cache.  Always count
> +                 core if there is no L3 cache.  */
> +              int count = ((threads_l2 > 0 && level == 3)
> +                           | ((threads_l3 > 0
> +                               || (threads_l2 > 0 && level == 2)) << 1));
> +
> +              while (count)
> +                {
> +                  __cpuid_count (11, i++, eax, ebx, ecx, edx);
> +
> +                  int shipped = ebx & 0xff;
> +                  int type = ecx & 0xff00;
> +                  if (shipped == 0 || type == 0)
> +                    break;
> +                  else if (type == 0x100)
> +                    {
> +                      /* Count SMT.  */
> +                      if ((count & 0x1))
> +                        {
> +                          int count_mask;
> +
> +                          /* Compute count mask.  */
> +                          asm ("bsr %1, %0"
> +                               : "=r" (count_mask) : "g" (threads_l2));
> +                          count_mask = ~(-1 << (count_mask + 1));
> +                          threads_l2 = (shipped - 1) & count_mask;
> +                          count &= ~0x1;
> +                        }
> +                    }
> +                  else if (type == 0x200)
> +                    {
> +                      /* Count core.  */
> +                      if ((count & (0x1 << 1)))
> +                        {
> +                          int count_mask;
> +                          int threads_core
> +                            = (level == 2 ? threads_l2 : threads_l3);
> +
> +                          /* Compute count mask.  */
> +                          asm ("bsr %1, %0"
> +                               : "=r" (count_mask) : "g" (threads_core));
> +                          count_mask = ~(-1 << (count_mask + 1));
> +                          threads_core = (shipped - 1) & count_mask;
> +                          if (level == 2)
> +                            threads_l2 = threads_core;
> +                          else
> +                            threads_l3 = threads_core;
> +                          count &= ~(0x1 << 1);
> +                        }
> +                    }
> +                }
> +            }
> +          if (threads_l2 > 0)
> +            threads_l2 += 1;
> +          if (threads_l3 > 0)
> +            threads_l3 += 1;
> +          if (level == 2)
> +            {
> +              if (threads_l2)
> +                {
> +                  threads = threads_l2;
> +                  if (cpu_features->basic.kind == arch_kind_intel
> +                      && threads > 2
> +                      && family == 6)
> +                    switch (model)
> +                      {
> +                        case 0x37:
> +                        case 0x4a:
> +                        case 0x4d:
> +                        case 0x5a:
> +                        case 0x5d:
> +                          /* Silvermont has L2 cache shared by 2 cores.  */
> +                          threads = 2;
> +                          break;
> +                        default:
> +                          break;
> +                      }
> +                }
> +            }
> +          else if (threads_l3)
> +            threads = threads_l3;
> +        }
> +      else
> +        {
> +intel_bug_no_cache_info:
> +          /* Assume that all logical threads share the highest cache
> +             level.  */
> +          threads
> +            = ((cpu_features->cpuid[COMMON_CPUID_INDEX_1].ebx
> +                >> 16) & 0xff);
> +        }
> +
> +        /* Cap usage of highest cache level to the number of supported
> +           threads.  */
> +        if (shared > 0 && threads > 0)
> +          shared /= threads;
> +    }
> +
> +  /* Account for non-inclusive L2 and L3 caches.  */
> +  if (!inclusive_cache)
> +    {
> +      if (threads_l2 > 0)
> +        core /= threads_l2;
> +      shared += core;
> +    }
> +
> +  *shared_ptr = shared;
> +  *threads_ptr = threads;
> +}
> +
> +void
> +__init_cacheinfo (void)
> +{
> +  /* Find out what brand of processor.  */
> +  unsigned int ebx;
> +  unsigned int ecx;
> +  unsigned int edx;
> +  int max_cpuid_ex;
> +  long int data = -1;
> +  long int shared = -1;
> +  long int core;
> +  unsigned int threads = 0;
> +  unsigned long int level1_icache_size = -1;
> +  unsigned long int level1_dcache_size = -1;
> +  unsigned long int level1_dcache_assoc = -1;
> +  unsigned long int level1_dcache_linesize = -1;
> +  unsigned long int level2_cache_size = -1;
> +  unsigned long int level2_cache_assoc = -1;
> +  unsigned long int level2_cache_linesize = -1;
> +  unsigned long int level3_cache_size = -1;
> +  unsigned long int level3_cache_assoc = -1;
> +  unsigned long int level3_cache_linesize = -1;
> +  unsigned long int level4_cache_size = -1;
> +  struct cpu_features *cpu_features = __get_cpu_features ();
> +
> +  if (cpu_features->basic.kind == arch_kind_intel)
> +    {
> +      data = handle_intel (_SC_LEVEL1_DCACHE_SIZE, cpu_features);
> +      core = handle_intel (_SC_LEVEL2_CACHE_SIZE, cpu_features);
> +      shared = handle_intel (_SC_LEVEL3_CACHE_SIZE, cpu_features);
> +
> +      level1_icache_size
> + = handle_intel (_SC_LEVEL1_ICACHE_SIZE, cpu_features);
> +      level1_dcache_size = data;
> +      level1_dcache_assoc
> + = handle_intel (_SC_LEVEL1_DCACHE_ASSOC, cpu_features);
> +      level1_dcache_linesize
> + = handle_intel (_SC_LEVEL1_DCACHE_LINESIZE, cpu_features);
> +      level2_cache_size = core;
> +      level2_cache_assoc
> + = handle_intel (_SC_LEVEL2_CACHE_ASSOC, cpu_features);
> +      level2_cache_linesize
> + = handle_intel (_SC_LEVEL2_CACHE_LINESIZE, cpu_features);
> +      level3_cache_size = shared;
> +      level3_cache_assoc
> + = handle_intel (_SC_LEVEL3_CACHE_ASSOC, cpu_features);
> +      level3_cache_linesize
> + = handle_intel (_SC_LEVEL3_CACHE_LINESIZE, cpu_features);
> +      level4_cache_size
> + = handle_intel (_SC_LEVEL4_CACHE_SIZE, cpu_features);
> +
> +      get_common_cache_info (&shared, &threads, core);
> +    }
> +  else if (cpu_features->basic.kind == arch_kind_zhaoxin)
> +    {
> +      data = handle_zhaoxin (_SC_LEVEL1_DCACHE_SIZE);
> +      core = handle_zhaoxin (_SC_LEVEL2_CACHE_SIZE);
> +      shared = handle_zhaoxin (_SC_LEVEL3_CACHE_SIZE);
> +
> +      level1_icache_size = handle_zhaoxin (_SC_LEVEL1_ICACHE_SIZE);
> +      level1_dcache_size = data;
> +      level1_dcache_assoc = handle_zhaoxin (_SC_LEVEL1_DCACHE_ASSOC);
> +      level1_dcache_linesize = handle_zhaoxin (_SC_LEVEL1_DCACHE_LINESIZE);
> +      level2_cache_size = core;
> +      level2_cache_assoc = handle_zhaoxin (_SC_LEVEL2_CACHE_ASSOC);
> +      level2_cache_linesize = handle_zhaoxin (_SC_LEVEL2_CACHE_LINESIZE);
> +      level3_cache_size = shared;
> +      level3_cache_assoc = handle_zhaoxin (_SC_LEVEL3_CACHE_ASSOC);
> +      level3_cache_linesize = handle_zhaoxin (_SC_LEVEL3_CACHE_LINESIZE);
> +
> +      get_common_cache_info (&shared, &threads, core);
> +    }
> +  else if (cpu_features->basic.kind == arch_kind_amd)
> +    {
> +      data  = handle_amd (_SC_LEVEL1_DCACHE_SIZE);
> +      core = handle_amd (_SC_LEVEL2_CACHE_SIZE);
> +      shared = handle_amd (_SC_LEVEL3_CACHE_SIZE);
> +
> +      level1_icache_size = handle_amd (_SC_LEVEL1_ICACHE_SIZE);
> +      level1_dcache_size = data;
> +      level1_dcache_assoc = handle_amd (_SC_LEVEL1_DCACHE_ASSOC);
> +      level1_dcache_linesize = handle_amd (_SC_LEVEL1_DCACHE_LINESIZE);
> +      level2_cache_size = core;
> +      level2_cache_assoc = handle_amd (_SC_LEVEL2_CACHE_ASSOC);
> +      level2_cache_linesize = handle_amd (_SC_LEVEL2_CACHE_LINESIZE);
> +      level3_cache_size = shared;
> +      level3_cache_assoc = handle_amd (_SC_LEVEL3_CACHE_ASSOC);
> +      level3_cache_linesize = handle_amd (_SC_LEVEL3_CACHE_LINESIZE);
> +
> +      /* Get maximum extended function. */
> +      __cpuid (0x80000000, max_cpuid_ex, ebx, ecx, edx);
> +
> +      if (shared <= 0)
> + /* No shared L3 cache.  All we have is the L2 cache.  */
> + shared = core;
> +      else
> + {
> +  /* Figure out the number of logical threads that share L3.  */
> +  if (max_cpuid_ex >= 0x80000008)
> +    {
> +      /* Get width of APIC ID.  */
> +      __cpuid (0x80000008, max_cpuid_ex, ebx, ecx, edx);
> +      threads = 1 << ((ecx >> 12) & 0x0f);
> +    }
> +
> +  if (threads == 0)
> +    {
> +      /* If APIC ID width is not available, use logical
> + processor count.  */
> +      __cpuid (0x00000001, max_cpuid_ex, ebx, ecx, edx);
> +
> +      if ((edx & (1 << 28)) != 0)
> + threads = (ebx >> 16) & 0xff;
> +    }
> +
> +  /* Cap usage of highest cache level to the number of
> +     supported threads.  */
> +  if (threads > 0)
> +    shared /= threads;
> +
> +  /* Account for exclusive L2 and L3 caches.  */
> +  shared += core;
> + }
> +
> +#ifndef __x86_64__
> +      if (max_cpuid_ex >= 0x80000001)
> + {
> +  unsigned int eax;
> +  __cpuid (0x80000001, eax, ebx, ecx, edx);
> +  /*  PREFETCHW     || 3DNow!  */
> +  if ((ecx & 0x100) || (edx & 0x80000000))
> +    cpu_features->prefetchw = -1;
> + }
> +#endif
> +    }
> +
> +  cpu_features->level1_icache_size = level1_icache_size;
> +  cpu_features->level1_dcache_size = level1_dcache_size;
> +  cpu_features->level1_dcache_assoc = level1_dcache_assoc;
> +  cpu_features->level1_dcache_linesize = level1_dcache_linesize;
> +  cpu_features->level2_cache_size = level2_cache_size;
> +  cpu_features->level2_cache_assoc = level2_cache_assoc;
> +  cpu_features->level2_cache_linesize = level2_cache_linesize;
> +  cpu_features->level3_cache_size = level3_cache_size;
> +  cpu_features->level3_cache_assoc = level3_cache_assoc;
> +  cpu_features->level3_cache_linesize = level3_cache_linesize;
> +  cpu_features->level4_cache_size = level4_cache_size;
> +
> +  /* The large memcpy micro benchmark in glibc shows that 6 times of
> +     shared cache size is the approximate value above which non-temporal
> +     store becomes faster on a 8-core processor.  This is the 3/4 of the
> +     total shared cache size.  */
> +  unsigned long int non_temporal_threshold = (shared * threads * 3 / 4);
> +
> +#if HAVE_TUNABLES
> +  long int tunable_size;
> +  tunable_size = TUNABLE_GET (x86_data_cache_size, long int, NULL);
> +  if (tunable_size != 0)
> +    data = tunable_size;
> +  tunable_size = TUNABLE_GET (x86_shared_cache_size, long int, NULL);
> +  if (tunable_size != 0)
> +    shared = tunable_size;
> +  tunable_size = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
> +  if (tunable_size != 0)
> +    non_temporal_threshold = tunable_size;

This is wrong, you have excluded 0 from the list of allowed options?
Why not raise the dl-tunables.list minimum to 1?
Then the above code just becomes what we used to have.

e.g.
  cpu_features->non_temporal_threshold
    = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);

The tunables framework should make it so you don't have to write this
boilerplate.

There should be only *one* variable that you have to handle specially
for the dynamic minimum with a TUNABLE_GET + TUNABLE_SET.

> +#endif
> +
> +  cpu_features->data_cache_size = data;
> +  cpu_features->shared_cache_size = shared;
> +  cpu_features->non_temporal_threshold = non_temporal_threshold;

Not needed, you can just fold the cpu_features->foo update like we used
to have it. The ranges are not being updated to any dynamic values.

> +
> +#if HAVE_TUNABLES
> +  TUNABLE_UPDATE (x86_data_cache_size, long int,
> +  data, 0, (long int) -1);
> +  TUNABLE_UPDATE (x86_shared_cache_size, long int,
> +  shared, 0, (long int) -1);
> +  TUNABLE_UPDATE (x86_non_temporal_threshold, long int,
> +  non_temporal_threshold, 0, (long int) -1);

Delete all of these. They don't need updating?

> +#endif
> +}
> diff --git a/sysdeps/x86/init-arch.h b/sysdeps/x86/init-arch.h
> index d6f59cf962..272ed10902 100644
> --- a/sysdeps/x86/init-arch.h
> +++ b/sysdeps/x86/init-arch.h
> @@ -23,6 +23,9 @@
>  #include <ifunc-init.h>
>  #include <isa.h>
>  
> +extern void __init_cacheinfo (void)
> +  __attribute__ ((visibility ("hidden")));
> +
>  #ifndef __x86_64__
>  /* Due to the reordering and the other nifty extensions in i686, it is
>     not really good to use heavily i586 optimized code on an i686.  It's
>


--
Cheers,
Carlos.

Reply | Threaded
Open this post in threaded view
|

Re: [PATCH 2/2] x86: Add thresholds for "rep movsb/stosb" to tunables

Sourceware - libc-alpha mailing list
In reply to this post by Sourceware - libc-alpha mailing list
On 7/3/20 1:52 PM, H.J. Lu wrote:
> Add x86_rep_movsb_threshold and x86_rep_stosb_threshold to tunables
> to update thresholds for "rep movsb" and "rep stosb" at run-time.
>
> Note that the user specified threshold for "rep movsb" smaller than the
> minimum threshold will be ignored.

Post v2 please. Almost there.

> ---
>  manual/tunables.texi                          | 14 +++++++
>  sysdeps/x86/cacheinfo.c                       | 20 ++++++++++
>  sysdeps/x86/cpu-features.h                    |  4 ++
>  sysdeps/x86/dl-cacheinfo.c                    | 38 +++++++++++++++++++
>  sysdeps/x86/dl-tunables.list                  |  6 +++
>  .../multiarch/memmove-vec-unaligned-erms.S    | 16 +-------
>  .../multiarch/memset-vec-unaligned-erms.S     | 12 +-----
>  7 files changed, 84 insertions(+), 26 deletions(-)
>
> diff --git a/manual/tunables.texi b/manual/tunables.texi
> index ec18b10834..61edd62425 100644
> --- a/manual/tunables.texi
> +++ b/manual/tunables.texi
> @@ -396,6 +396,20 @@ to set threshold in bytes for non temporal store.
>  This tunable is specific to i386 and x86-64.
>  @end deftp
>  
> +@deftp Tunable glibc.cpu.x86_rep_movsb_threshold
> +The @code{glibc.cpu.x86_rep_movsb_threshold} tunable allows the user
> +to set threshold in bytes to start using "rep movsb".
> +
> +This tunable is specific to i386 and x86-64.
> +@end deftp
> +
> +@deftp Tunable glibc.cpu.x86_rep_stosb_threshold
> +The @code{glibc.cpu.x86_rep_stosb_threshold} tunable allows the user
> +to set threshold in bytes to start using "rep stosb".
> +
> +This tunable is specific to i386 and x86-64.
> +@end deftp
> +
>  @deftp Tunable glibc.cpu.x86_ibt
>  The @code{glibc.cpu.x86_ibt} tunable allows the user to control how
>  indirect branch tracking (IBT) should be enabled.  Accepted values are
> diff --git a/sysdeps/x86/cacheinfo.c b/sysdeps/x86/cacheinfo.c
> index 8c4c7f9972..bb536d96ef 100644
> --- a/sysdeps/x86/cacheinfo.c
> +++ b/sysdeps/x86/cacheinfo.c
> @@ -41,6 +41,23 @@ long int __x86_raw_shared_cache_size attribute_hidden = 1024 * 1024;
>  /* Threshold to use non temporal store.  */
>  long int __x86_shared_non_temporal_threshold attribute_hidden;
>  
> +/* Threshold to use Enhanced REP MOVSB.  Since there is overhead to set
> +   up REP MOVSB operation, REP MOVSB isn't faster on short data.  The
> +   memcpy micro benchmark in glibc shows that 2KB is the approximate
> +   value above which REP MOVSB becomes faster than SSE2 optimization
> +   on processors with Enhanced REP MOVSB.  Since larger register size
> +   can move more data with a single load and store, the threshold is
> +   higher with larger register size.  */
> +long int __x86_rep_movsb_threshold attribute_hidden = 2048;
> +
> +/* Threshold to use Enhanced REP STOSB.  Since there is overhead to set
> +   up REP STOSB operation, REP STOSB isn't faster on short data.  The
> +   memset micro benchmark in glibc shows that 2KB is the approximate
> +   value above which REP STOSB becomes faster on processors with
> +   Enhanced REP STOSB.  Since the stored value is fixed, larger register
> +   size has minimal impact on threshold.  */
> +long int __x86_rep_stosb_threshold attribute_hidden = 2048;
> +
>  #ifndef __x86_64__
>  /* PREFETCHW support flag for use in memory and string routines.  */
>  int __x86_prefetchw attribute_hidden;
> @@ -117,6 +134,9 @@ init_cacheinfo (void)
>    __x86_shared_non_temporal_threshold
>      = cpu_features->non_temporal_threshold;
>  
> +  __x86_rep_movsb_threshold = cpu_features->rep_movsb_threshold;
> +  __x86_rep_stosb_threshold = cpu_features->rep_stosb_threshold;
> +

OK. Update global from cpu_features with values.

I would really like to see some kind of "assert (cpu_features->initialized);"
that way we know we didn't break the startup sequence unintentionally.

>  #ifndef __x86_64__
>    __x86_prefetchw = cpu_features->prefetchw;
>  #endif
> diff --git a/sysdeps/x86/cpu-features.h b/sysdeps/x86/cpu-features.h
> index 3aaed33cbc..002e12e11f 100644
> --- a/sysdeps/x86/cpu-features.h
> +++ b/sysdeps/x86/cpu-features.h
> @@ -128,6 +128,10 @@ struct cpu_features
>    /* PREFETCHW support flag for use in memory and string routines.  */
>    unsigned long int prefetchw;
>  #endif
> +  /* Threshold to use "rep movsb".  */
> +  unsigned long int rep_movsb_threshold;
> +  /* Threshold to use "rep stosb".  */
> +  unsigned long int rep_stosb_threshold;

OK.

>  };
>  
>  /* Used from outside of glibc to get access to the CPU features
> diff --git a/sysdeps/x86/dl-cacheinfo.c b/sysdeps/x86/dl-cacheinfo.c
> index 8e2a6f552c..aff9bd1067 100644
> --- a/sysdeps/x86/dl-cacheinfo.c
> +++ b/sysdeps/x86/dl-cacheinfo.c
> @@ -860,6 +860,31 @@ __init_cacheinfo (void)
>       total shared cache size.  */
>    unsigned long int non_temporal_threshold = (shared * threads * 3 / 4);
>  
> +  /* NB: The REP MOVSB threshold must be greater than VEC_SIZE * 8.  */
> +  unsigned long int minimum_rep_movsb_threshold;
> +  /* NB: The default REP MOVSB threshold is 2048 * (VEC_SIZE / 16).  See
> +     comments for __x86_rep_movsb_threshold in cacheinfo.c.  */
> +  unsigned long int rep_movsb_threshold;
> +  if (CPU_FEATURES_ARCH_P (cpu_features, AVX512F_Usable)
> +      && !CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_AVX512))
> +    {
> +      rep_movsb_threshold = 2048 * (64 / 16);
> +      minimum_rep_movsb_threshold = 64 * 8;
> +    }
> +  else if (CPU_FEATURES_ARCH_P (cpu_features,
> + AVX_Fast_Unaligned_Load))
> +    {
> +      rep_movsb_threshold = 2048 * (32 / 16);
> +      minimum_rep_movsb_threshold = 32 * 8;
> +    }
> +  else
> +    {
> +      rep_movsb_threshold = 2048 * (16 / 16);
> +      minimum_rep_movsb_threshold = 16 * 8;
> +    }
> +  /* NB: See comments for __x86_rep_stosb_threshold in cacheinfo.c.  */
> +  unsigned long int rep_stosb_threshold = 2048;
> +
>  #if HAVE_TUNABLES
>    long int tunable_size;
>    tunable_size = TUNABLE_GET (x86_data_cache_size, long int, NULL);
> @@ -871,11 +896,19 @@ __init_cacheinfo (void)
>    tunable_size = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
>    if (tunable_size != 0)
>      non_temporal_threshold = tunable_size;

> +  tunable_size = TUNABLE_GET (x86_rep_movsb_threshold, long int, NULL);
> +  if (tunable_size > minimum_rep_movsb_threshold)
> +    rep_movsb_threshold = tunable_size;

OK. Good, we only set rep_movsb_threshold if it's greater than min.

> +  tunable_size = TUNABLE_GET (x86_rep_stosb_threshold, long int, NULL);
> +  if (tunable_size != 0)
> +    rep_stosb_threshold = tunable_size;

This should be min=1, default=2048 in dl-tunables.list, and would remove
this code since the range is not dynamic.

The point of the tunables framework is to remove such boiler plate for
range a default processing and clearing parameters for security settings.

>  #endif
>  
>    cpu_features->data_cache_size = data;
>    cpu_features->shared_cache_size = shared;
>    cpu_features->non_temporal_threshold = non_temporal_threshold;
> +  cpu_features->rep_movsb_threshold = rep_movsb_threshold;
> +  cpu_features->rep_stosb_threshold = rep_stosb_threshold;
>  
>  #if HAVE_TUNABLES
>    TUNABLE_UPDATE (x86_data_cache_size, long int,
> @@ -884,5 +917,10 @@ __init_cacheinfo (void)
>    shared, 0, (long int) -1);
>    TUNABLE_UPDATE (x86_non_temporal_threshold, long int,
>    non_temporal_threshold, 0, (long int) -1);
> +  TUNABLE_UPDATE (x86_rep_movsb_threshold, long int,
> +  rep_movsb_threshold, minimum_rep_movsb_threshold,
> +  (long int) -1);

OK. Store the new value and the computed minimum.

> +  TUNABLE_UPDATE (x86_rep_stosb_threshold, long int,
> +  rep_stosb_threshold, 0, (long int) -1);

This one can be deleted.

>  #endif
>  }
> diff --git a/sysdeps/x86/dl-tunables.list b/sysdeps/x86/dl-tunables.list
> index 251b926ce4..43bf6c2389 100644
> --- a/sysdeps/x86/dl-tunables.list
> +++ b/sysdeps/x86/dl-tunables.list
> @@ -30,6 +30,12 @@ glibc {
>      x86_non_temporal_threshold {
>        type: SIZE_T
>      }
> +    x86_rep_movsb_threshold {
> +      type: SIZE_T
> +    }
> +    x86_rep_stosb_threshold {
> +      type: SIZE_T

min: 1
default: 2048

> +    }
>      x86_data_cache_size {
>        type: SIZE_T
>      }
> diff --git a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
> index 74953245aa..bd5dc1a3f3 100644
> --- a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
> +++ b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
> @@ -56,17 +56,6 @@
>  # endif
>  #endif
>  
> -/* Threshold to use Enhanced REP MOVSB.  Since there is overhead to set
> -   up REP MOVSB operation, REP MOVSB isn't faster on short data.  The
> -   memcpy micro benchmark in glibc shows that 2KB is the approximate
> -   value above which REP MOVSB becomes faster than SSE2 optimization
> -   on processors with Enhanced REP MOVSB.  Since larger register size
> -   can move more data with a single load and store, the threshold is
> -   higher with larger register size.  */
> -#ifndef REP_MOVSB_THRESHOLD
> -# define REP_MOVSB_THRESHOLD (2048 * (VEC_SIZE / 16))> -#endif

OK.

> -
>  #ifndef PREFETCH
>  # define PREFETCH(addr) prefetcht0 addr
>  #endif
> @@ -253,9 +242,6 @@ L(movsb):
>   leaq (%rsi,%rdx), %r9
>   cmpq %r9, %rdi
>   /* Avoid slow backward REP MOVSB.  */
> -# if REP_MOVSB_THRESHOLD <= (VEC_SIZE * 8)
> -#  error Unsupported REP_MOVSB_THRESHOLD and VEC_SIZE!
> -# endif

OK.

>   jb L(more_8x_vec_backward)
>  1:
>   mov %RDX_LP, %RCX_LP
> @@ -331,7 +317,7 @@ L(between_2_3):
>  
>  #if defined USE_MULTIARCH && IS_IN (libc)
>  L(movsb_more_2x_vec):
> - cmpq $REP_MOVSB_THRESHOLD, %rdx
> + cmp __x86_rep_movsb_threshold(%rip), %RDX_LP

OK.

>   ja L(movsb)
>  #endif
>  L(more_2x_vec):
> diff --git a/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S b/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
> index af2299709c..2bfc95de05 100644
> --- a/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
> +++ b/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
> @@ -58,16 +58,6 @@
>  # endif
>  #endif
>  
> -/* Threshold to use Enhanced REP STOSB.  Since there is overhead to set
> -   up REP STOSB operation, REP STOSB isn't faster on short data.  The
> -   memset micro benchmark in glibc shows that 2KB is the approximate
> -   value above which REP STOSB becomes faster on processors with
> -   Enhanced REP STOSB.  Since the stored value is fixed, larger register
> -   size has minimal impact on threshold.  */
> -#ifndef REP_STOSB_THRESHOLD
> -# define REP_STOSB_THRESHOLD 2048
> -#endif
> -
>  #ifndef SECTION
>  # error SECTION is not defined!
>  #endif
> @@ -181,7 +171,7 @@ ENTRY (MEMSET_SYMBOL (__memset, unaligned_erms))
>   ret
>  
>  L(stosb_more_2x_vec):
> - cmpq $REP_STOSB_THRESHOLD, %rdx
> + cmp __x86_rep_stosb_threshold(%rip), %RDX_LP

OK.

>   ja L(stosb)
>  #endif
>  L(more_2x_vec):
>


--
Cheers,
Carlos.

Reply | Threaded
Open this post in threaded view
|

V2 [PATCH 1/2] Update tunable min/max values

Sourceware - libc-alpha mailing list
In reply to this post by Sourceware - libc-alpha mailing list
On Fri, Jul 03, 2020 at 03:49:16PM -0400, Carlos O'Donell wrote:

> On 7/3/20 1:52 PM, H.J. Lu wrote:
> > Add __tunable_update_val to update tunable min/max values and move x86
> > processor cache info to cpu_features.
>
> Nees a little more cleanup. I'd like to avoid more macro APIs if possible.  
>
> Post v2 please.
>
> > ---
> >  elf/dl-tunables.c          |  51 ++-
> >  elf/dl-tunables.h          |  15 +
> >  sysdeps/i386/cacheinfo.c   |   3 -
> >  sysdeps/x86/Makefile       |   2 +-
> >  sysdeps/x86/cacheinfo.c    | 852 +++--------------------------------
> >  sysdeps/x86/cpu-features.c |  19 +-
> >  sysdeps/x86/cpu-features.h |  26 ++
> >  sysdeps/x86/dl-cacheinfo.c | 888 +++++++++++++++++++++++++++++++++++++
> >  sysdeps/x86/init-arch.h    |   3 +
> >  9 files changed, 1024 insertions(+), 835 deletions(-)
> >  delete mode 100644 sysdeps/i386/cacheinfo.c
> >  create mode 100644 sysdeps/x86/dl-cacheinfo.c
> >
> > diff --git a/elf/dl-tunables.c b/elf/dl-tunables.c
> > index 26e6e26612..7c9f1ca31f 100644
> > --- a/elf/dl-tunables.c
> > +++ b/elf/dl-tunables.c
> > @@ -100,31 +100,39 @@ get_next_env (char **envp, char **name, size_t *namelen, char **val,
> >      }      \
> >  })
> >  
> > +#define TUNABLE_UPDATE_VAL(__cur, __val, __min, __max, __type)      \
> > +({      \
> > +  (__cur)->type.min = (__min);      \
> > +  (__cur)->type.max = (__max);      \
> > +  (__cur)->val.numval = (__val);      \
> > +  (__cur)->initialized = true;      \
> > +})
>
> Can we update TUNABLE_SET_VAL_IF_VALID_RANGE to take the extra parameters?
>
> If the value is not within [__min, __max] it should not update numval.

I removed TUNABLE_UPDATE_VAL.

>
> > +
> >  static void
> > -do_tunable_update_val (tunable_t *cur, const void *valp)
> > +do_tunable_update_val (tunable_t *cur, const void *valp,
> > +       const void *minp, const void *maxp)
> >  {
> > -  uint64_t val;
> > +  uint64_t val, min, max;
> >  
> >    if (cur->type.type_code != TUNABLE_TYPE_STRING)
> > -    val = *((int64_t *) valp);
> > +    {
> > +      val = *((int64_t *) valp);
> > +      if (minp)
> > + min = *((int64_t *) minp);
> > +      if (maxp)
> > + max = *((int64_t *) maxp);
> > +    }
>
> OK.
>
> >  
> >    switch (cur->type.type_code)
> >      {
> >      case TUNABLE_TYPE_INT_32:
> > - {
> > -  TUNABLE_SET_VAL_IF_VALID_RANGE (cur, val, int64_t);
> > -  break;
> > - }
> >      case TUNABLE_TYPE_UINT_64:
> > - {
> > -  TUNABLE_SET_VAL_IF_VALID_RANGE (cur, val, uint64_t);
> > -  break;
> > - }
> >      case TUNABLE_TYPE_SIZE_T:
> > - {
> > -  TUNABLE_SET_VAL_IF_VALID_RANGE (cur, val, uint64_t);
> > -  break;
> > - }
> > +      if (minp && maxp)
> > + TUNABLE_UPDATE_VAL (cur, val, min, max, int64_t);
> > +      else
> > + TUNABLE_SET_VAL_IF_VALID_RANGE (cur, val, int64_t);
> > +      break;
>
> Merging the two would result in a one call here and only one macro API
> for set/update. I don't see the requirement for a different set and update.

I left this unchanged.

>
> >      case TUNABLE_TYPE_STRING:
> >   {
> >    cur->val.strval = valp;
> > @@ -153,7 +161,7 @@ tunable_initialize (tunable_t *cur, const char *strval)
> >        cur->initialized = true;
> >        valp = strval;
> >      }
> > -  do_tunable_update_val (cur, valp);
> > +  do_tunable_update_val (cur, valp, NULL, NULL);
> >  }
> >  
> >  void
> > @@ -161,8 +169,17 @@ __tunable_set_val (tunable_id_t id, void *valp)
> >  {
> >    tunable_t *cur = &tunable_list[id];
> >  
> > -  do_tunable_update_val (cur, valp);
> > +  do_tunable_update_val (cur, valp, NULL, NULL);
> > +}
> > +
> > +void
> > +__tunable_update_val (tunable_id_t id, void *valp, void *minp, void *maxp)
> > +{
> > +  tunable_t *cur = &tunable_list[id];
> > +
> > +  do_tunable_update_val (cur, valp, minp, maxp);
> >  }
> > +rtld_hidden_def (__tunable_update_val)
> >  
> >  #if TUNABLES_FRONTEND == TUNABLES_FRONTEND_valstring
> >  /* Parse the tunable string TUNESTR and adjust it to drop any tunables that may
> > diff --git a/elf/dl-tunables.h b/elf/dl-tunables.h
> > index f05eb50c2f..f6bf7379af 100644
> > --- a/elf/dl-tunables.h
> > +++ b/elf/dl-tunables.h
> > @@ -71,8 +71,10 @@ typedef struct _tunable tunable_t;
> >  extern void __tunables_init (char **);
> >  extern void __tunable_get_val (tunable_id_t, void *, tunable_callback_t);
> >  extern void __tunable_set_val (tunable_id_t, void *);
> > +extern void __tunable_update_val (tunable_id_t, void *, void *, void *);
> >  rtld_hidden_proto (__tunables_init)
> >  rtld_hidden_proto (__tunable_get_val)
> > +rtld_hidden_proto (__tunable_update_val)
> >  
> >  /* Define TUNABLE_GET and TUNABLE_SET in short form if TOP_NAMESPACE and
> >     TUNABLE_NAMESPACE are defined.  This is useful shorthand to get and set
> > @@ -82,11 +84,16 @@ rtld_hidden_proto (__tunable_get_val)
> >    TUNABLE_GET_FULL (TOP_NAMESPACE, TUNABLE_NAMESPACE, __id, __type, __cb)
> >  # define TUNABLE_SET(__id, __type, __val) \
> >    TUNABLE_SET_FULL (TOP_NAMESPACE, TUNABLE_NAMESPACE, __id, __type, __val)
> > +# define TUNABLE_UPDATE(__id, __type, __val, __min, __max) \
> > +  TUNABLE_UPDATE_FULL (TOP_NAMESPACE, TUNABLE_NAMESPACE, __id, __type, \
> > +       __val, __min, __max)
>
> Please use TUNABLE_SET?
>
> I would like avoid needing new macro APIs.
>
> There doesn't need to be symmetry between TUNABLE_GET and TUNABLE_SET.
>
> The TUNABLE_GET is for getting the value of the tunable during subsystem
> startup, and the tunable framework handles: setting the default value,
> processing user settings, enforcing minimums, enforcing maximums etc.
>
> The TUNABLE_SET side doesn't have to be symmetric it may need to set more
> than just the tunable value e.g. min/max.

TUNABLE_SET is needed to only set value.  I added TUNABLE_SET_ALL to
set value as well as min/max.

>
> > -}
> > -
> > -
>
> OK.
>
> >  /* Data cache size for use in memory and string routines, typically
> OK.
>
> > +
> >  #if HAVE_TUNABLES
> >    TUNABLE_GET (hwcaps, tunable_val_t *, TUNABLE_CALLBACK (set_hwcaps));
> > -  cpu_features->non_temporal_threshold
> > +  long int tunable_size;
> > +  tunable_size = TUNABLE_GET (x86_data_cache_size, long int, NULL);
> > +  if (tunable_size != 0)
> > +    data = tunable_size;
> > +  tunable_size = TUNABLE_GET (x86_shared_cache_size, long int, NULL);
> > +  if (tunable_size != 0)
> > +    shared = tunable_size;
> > +  tunable_size = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
> > +  if (tunable_size != 0)
> > +    non_temporal_threshold = tunable_size;
>
> This is wrong, you have excluded 0 from the list of allowed options?

0 is invalid and means it is unset in tunables.

> Why not raise the dl-tunables.list minimum to 1?
> Then the above code just becomes what we used to have.

What we used to have are

  cpu_features->non_temporal_threshold
    = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
...
  __x86_shared_non_temporal_threshold
    = (cpu_features->non_temporal_threshold != 0
       ? cpu_features->non_temporal_threshold
       : __x86_shared_cache_size * threads * 3 / 4);

I just changed the order.

>
> e.g.
>   cpu_features->non_temporal_threshold
>     = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
>
> The tunables framework should make it so you don't have to write this
> boilerplate.
>
> There should be only *one* variable that you have to handle specially
> for the dynamic minimum with a TUNABLE_GET + TUNABLE_SET.
>
> > +#endif
> > +
> > +  cpu_features->data_cache_size = data;
> > +  cpu_features->shared_cache_size = shared;
> > +  cpu_features->non_temporal_threshold = non_temporal_threshold;
>
> Not needed, you can just fold the cpu_features->foo update like we used
> to have it. The ranges are not being updated to any dynamic values.

The purpose of these changes is to make these values always valid
so that they can be used directly.  Otherwise, they have to be checked
against invalid 0 value before using.

> > +
> > +#if HAVE_TUNABLES
> > +  TUNABLE_UPDATE (x86_data_cache_size, long int,
> > +  data, 0, (long int) -1);
> > +  TUNABLE_UPDATE (x86_shared_cache_size, long int,
> > +  shared, 0, (long int) -1);
> > +  TUNABLE_UPDATE (x86_non_temporal_threshold, long int,
> > +  non_temporal_threshold, 0, (long int) -1);
>
> Delete all of these. They don't need updating?

Done.

> > +#endif
> > +}
> > diff --git a/sysdeps/x86/init-arch.h b/sysdeps/x86/init-arch.h
> > index d6f59cf962..272ed10902 100644
> > --- a/sysdeps/x86/init-arch.h
> > +++ b/sysdeps/x86/init-arch.h
> > @@ -23,6 +23,9 @@
> >  #include <ifunc-init.h>
> >  #include <isa.h>
> >  
> > +extern void __init_cacheinfo (void)
> > +  __attribute__ ((visibility ("hidden")));
> > +
> >  #ifndef __x86_64__
> >  /* Due to the reordering and the other nifty extensions in i686, it is
> >     not really good to use heavily i586 optimized code on an i686.  It's
> >


H.J.
---
From b66f08560b08ad1d718b3d19a8b0bee66c08ba49 Mon Sep 17 00:00:00 2001
From: "H.J. Lu" <[hidden email]>
Date: Mon, 1 Jun 2020 14:11:32 -0700
Subject: [PATCH 1/2] Set tunable value as well as min/max values

Update __tunable_set_val to set tunable value as well as min/max values.
TUNABLE_SET_ALL macro is added, but not used.  Move x86 processor cache
info to cpu_features.
---
 elf/dl-tunables.c          |  20 +-
 elf/dl-tunables.h          |  18 +-
 sysdeps/i386/cacheinfo.c   |   3 -
 sysdeps/x86/Makefile       |   2 +-
 sysdeps/x86/cacheinfo.c    | 852 +++--------------------------------
 sysdeps/x86/cpu-features.c |  19 +-
 sysdeps/x86/cpu-features.h |  26 ++
 sysdeps/x86/dl-cacheinfo.c | 879 +++++++++++++++++++++++++++++++++++++
 sysdeps/x86/init-arch.h    |   3 +
 9 files changed, 996 insertions(+), 826 deletions(-)
 delete mode 100644 sysdeps/i386/cacheinfo.c
 create mode 100644 sysdeps/x86/dl-cacheinfo.c

diff --git a/elf/dl-tunables.c b/elf/dl-tunables.c
index 26e6e26612..762b366705 100644
--- a/elf/dl-tunables.c
+++ b/elf/dl-tunables.c
@@ -101,12 +101,19 @@ get_next_env (char **envp, char **name, size_t *namelen, char **val,
 })
 
 static void
-do_tunable_update_val (tunable_t *cur, const void *valp)
+do_tunable_update_val (tunable_t *cur, const void *valp,
+       const void *minp, const void *maxp)
 {
-  uint64_t val;
+  uint64_t val, min, max;
 
   if (cur->type.type_code != TUNABLE_TYPE_STRING)
-    val = *((int64_t *) valp);
+    {
+      val = *((int64_t *) valp);
+      if (minp)
+ cur->type.min = min = *((int64_t *) minp);
+      if (maxp)
+ cur->type.min = max = *((int64_t *) maxp);
+    }
 
   switch (cur->type.type_code)
     {
@@ -153,16 +160,17 @@ tunable_initialize (tunable_t *cur, const char *strval)
       cur->initialized = true;
       valp = strval;
     }
-  do_tunable_update_val (cur, valp);
+  do_tunable_update_val (cur, valp, NULL, NULL);
 }
 
 void
-__tunable_set_val (tunable_id_t id, void *valp)
+__tunable_set_val (tunable_id_t id, void *valp, void *minp, void *maxp)
 {
   tunable_t *cur = &tunable_list[id];
 
-  do_tunable_update_val (cur, valp);
+  do_tunable_update_val (cur, valp, minp, maxp);
 }
+rtld_hidden_def (__tunable_set_val)
 
 #if TUNABLES_FRONTEND == TUNABLES_FRONTEND_valstring
 /* Parse the tunable string TUNESTR and adjust it to drop any tunables that may
diff --git a/elf/dl-tunables.h b/elf/dl-tunables.h
index f05eb50c2f..1bad31240d 100644
--- a/elf/dl-tunables.h
+++ b/elf/dl-tunables.h
@@ -70,9 +70,10 @@ typedef struct _tunable tunable_t;
 
 extern void __tunables_init (char **);
 extern void __tunable_get_val (tunable_id_t, void *, tunable_callback_t);
-extern void __tunable_set_val (tunable_id_t, void *);
+extern void __tunable_set_val (tunable_id_t, void *, void *, void *);
 rtld_hidden_proto (__tunables_init)
 rtld_hidden_proto (__tunable_get_val)
+rtld_hidden_proto (__tunable_set_val)
 
 /* Define TUNABLE_GET and TUNABLE_SET in short form if TOP_NAMESPACE and
    TUNABLE_NAMESPACE are defined.  This is useful shorthand to get and set
@@ -82,11 +83,16 @@ rtld_hidden_proto (__tunable_get_val)
   TUNABLE_GET_FULL (TOP_NAMESPACE, TUNABLE_NAMESPACE, __id, __type, __cb)
 # define TUNABLE_SET(__id, __type, __val) \
   TUNABLE_SET_FULL (TOP_NAMESPACE, TUNABLE_NAMESPACE, __id, __type, __val)
+# define TUNABLE_SET_ALL(__id, __type, __val, __min, __max) \
+  TUNABLE_SET_ALL_FULL (TOP_NAMESPACE, TUNABLE_NAMESPACE, __id, __type, \
+ __val, __min, __max)
 #else
 # define TUNABLE_GET(__top, __ns, __id, __type, __cb) \
   TUNABLE_GET_FULL (__top, __ns, __id, __type, __cb)
 # define TUNABLE_SET(__top, __ns, __id, __type, __val) \
   TUNABLE_SET_FULL (__top, __ns, __id, __type, __val)
+# define TUNABLE_SET_ALL(__top, __ns, __id, __type, __val, __min, __max) \
+  TUNABLE_SET_ALL_FULL (__top, __ns, __id, __type, __val, __min, __max)
 #endif
 
 /* Get and return a tunable value.  If the tunable was set externally and __CB
@@ -103,7 +109,15 @@ rtld_hidden_proto (__tunable_get_val)
 # define TUNABLE_SET_FULL(__top, __ns, __id, __type, __val) \
 ({      \
   __tunable_set_val (TUNABLE_ENUM_NAME (__top, __ns, __id),      \
- & (__type) {__val});      \
+ & (__type) {__val}, NULL, NULL);      \
+})
+
+/* Set a tunable value together with min/max values.  */
+# define TUNABLE_SET_ALL_FULL(__top, __ns, __id, __type, __val, __min, __max) \
+({      \
+  __tunable_set_val (TUNABLE_ENUM_NAME (__top, __ns, __id),      \
+     & (__type) {__val},  & (__type) {__min},      \
+     & (__type) {__max});      \
 })
 
 /* Namespace sanity for callback functions.  Use this macro to keep the
diff --git a/sysdeps/i386/cacheinfo.c b/sysdeps/i386/cacheinfo.c
deleted file mode 100644
index f15fe0779a..0000000000
--- a/sysdeps/i386/cacheinfo.c
+++ /dev/null
@@ -1,3 +0,0 @@
-#define DISABLE_PREFETCHW
-
-#include <sysdeps/x86/cacheinfo.c>
diff --git a/sysdeps/x86/Makefile b/sysdeps/x86/Makefile
index beab426f67..0872e0e655 100644
--- a/sysdeps/x86/Makefile
+++ b/sysdeps/x86/Makefile
@@ -3,7 +3,7 @@ gen-as-const-headers += cpu-features-offsets.sym
 endif
 
 ifeq ($(subdir),elf)
-sysdep-dl-routines += dl-get-cpu-features
+sysdep-dl-routines += dl-get-cpu-features dl-cacheinfo
 
 tests += tst-get-cpu-features tst-get-cpu-features-static
 tests-static += tst-get-cpu-features-static
diff --git a/sysdeps/x86/cacheinfo.c b/sysdeps/x86/cacheinfo.c
index 311502dee3..8c4c7f9972 100644
--- a/sysdeps/x86/cacheinfo.c
+++ b/sysdeps/x86/cacheinfo.c
@@ -18,498 +18,9 @@
 
 #if IS_IN (libc)
 
-#include <assert.h>
-#include <stdbool.h>
-#include <stdlib.h>
 #include <unistd.h>
-#include <cpuid.h>
 #include <init-arch.h>
 
-static const struct intel_02_cache_info
-{
-  unsigned char idx;
-  unsigned char assoc;
-  unsigned char linesize;
-  unsigned char rel_name;
-  unsigned int size;
-} intel_02_known [] =
-  {
-#define M(sc) ((sc) - _SC_LEVEL1_ICACHE_SIZE)
-    { 0x06,  4, 32, M(_SC_LEVEL1_ICACHE_SIZE),    8192 },
-    { 0x08,  4, 32, M(_SC_LEVEL1_ICACHE_SIZE),   16384 },
-    { 0x09,  4, 32, M(_SC_LEVEL1_ICACHE_SIZE),   32768 },
-    { 0x0a,  2, 32, M(_SC_LEVEL1_DCACHE_SIZE),    8192 },
-    { 0x0c,  4, 32, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
-    { 0x0d,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
-    { 0x0e,  6, 64, M(_SC_LEVEL1_DCACHE_SIZE),   24576 },
-    { 0x21,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
-    { 0x22,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),   524288 },
-    { 0x23,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  1048576 },
-    { 0x25,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
-    { 0x29,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
-    { 0x2c,  8, 64, M(_SC_LEVEL1_DCACHE_SIZE),   32768 },
-    { 0x30,  8, 64, M(_SC_LEVEL1_ICACHE_SIZE),   32768 },
-    { 0x39,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
-    { 0x3a,  6, 64, M(_SC_LEVEL2_CACHE_SIZE),   196608 },
-    { 0x3b,  2, 64, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
-    { 0x3c,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
-    { 0x3d,  6, 64, M(_SC_LEVEL2_CACHE_SIZE),   393216 },
-    { 0x3e,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
-    { 0x3f,  2, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
-    { 0x41,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
-    { 0x42,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
-    { 0x43,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
-    { 0x44,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
-    { 0x45,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),  2097152 },
-    { 0x46,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
-    { 0x47,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
-    { 0x48, 12, 64, M(_SC_LEVEL2_CACHE_SIZE),  3145728 },
-    { 0x49, 16, 64, M(_SC_LEVEL2_CACHE_SIZE),  4194304 },
-    { 0x4a, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  6291456 },
-    { 0x4b, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
-    { 0x4c, 12, 64, M(_SC_LEVEL3_CACHE_SIZE), 12582912 },
-    { 0x4d, 16, 64, M(_SC_LEVEL3_CACHE_SIZE), 16777216 },
-    { 0x4e, 24, 64, M(_SC_LEVEL2_CACHE_SIZE),  6291456 },
-    { 0x60,  8, 64, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
-    { 0x66,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),    8192 },
-    { 0x67,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
-    { 0x68,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),   32768 },
-    { 0x78,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
-    { 0x79,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
-    { 0x7a,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
-    { 0x7b,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
-    { 0x7c,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
-    { 0x7d,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  2097152 },
-    { 0x7f,  2, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
-    { 0x80,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
-    { 0x82,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
-    { 0x83,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
-    { 0x84,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
-    { 0x85,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),  2097152 },
-    { 0x86,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
-    { 0x87,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
-    { 0xd0,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),   524288 },
-    { 0xd1,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),  1048576 },
-    { 0xd2,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
-    { 0xd6,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  1048576 },
-    { 0xd7,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
-    { 0xd8,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
-    { 0xdc, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
-    { 0xdd, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
-    { 0xde, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
-    { 0xe2, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
-    { 0xe3, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
-    { 0xe4, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
-    { 0xea, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 12582912 },
-    { 0xeb, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 18874368 },
-    { 0xec, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 25165824 },
-  };
-
-#define nintel_02_known (sizeof (intel_02_known) / sizeof (intel_02_known [0]))
-
-static int
-intel_02_known_compare (const void *p1, const void *p2)
-{
-  const struct intel_02_cache_info *i1;
-  const struct intel_02_cache_info *i2;
-
-  i1 = (const struct intel_02_cache_info *) p1;
-  i2 = (const struct intel_02_cache_info *) p2;
-
-  if (i1->idx == i2->idx)
-    return 0;
-
-  return i1->idx < i2->idx ? -1 : 1;
-}
-
-
-static long int
-__attribute__ ((noinline))
-intel_check_word (int name, unsigned int value, bool *has_level_2,
-  bool *no_level_2_or_3,
-  const struct cpu_features *cpu_features)
-{
-  if ((value & 0x80000000) != 0)
-    /* The register value is reserved.  */
-    return 0;
-
-  /* Fold the name.  The _SC_ constants are always in the order SIZE,
-     ASSOC, LINESIZE.  */
-  int folded_rel_name = (M(name) / 3) * 3;
-
-  while (value != 0)
-    {
-      unsigned int byte = value & 0xff;
-
-      if (byte == 0x40)
- {
-  *no_level_2_or_3 = true;
-
-  if (folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))
-    /* No need to look further.  */
-    break;
- }
-      else if (byte == 0xff)
- {
-  /* CPUID leaf 0x4 contains all the information.  We need to
-     iterate over it.  */
-  unsigned int eax;
-  unsigned int ebx;
-  unsigned int ecx;
-  unsigned int edx;
-
-  unsigned int round = 0;
-  while (1)
-    {
-      __cpuid_count (4, round, eax, ebx, ecx, edx);
-
-      enum { null = 0, data = 1, inst = 2, uni = 3 } type = eax & 0x1f;
-      if (type == null)
- /* That was the end.  */
- break;
-
-      unsigned int level = (eax >> 5) & 0x7;
-
-      if ((level == 1 && type == data
-   && folded_rel_name == M(_SC_LEVEL1_DCACHE_SIZE))
-  || (level == 1 && type == inst
-      && folded_rel_name == M(_SC_LEVEL1_ICACHE_SIZE))
-  || (level == 2 && folded_rel_name == M(_SC_LEVEL2_CACHE_SIZE))
-  || (level == 3 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))
-  || (level == 4 && folded_rel_name == M(_SC_LEVEL4_CACHE_SIZE)))
- {
-  unsigned int offset = M(name) - folded_rel_name;
-
-  if (offset == 0)
-    /* Cache size.  */
-    return (((ebx >> 22) + 1)
-    * (((ebx >> 12) & 0x3ff) + 1)
-    * ((ebx & 0xfff) + 1)
-    * (ecx + 1));
-  if (offset == 1)
-    return (ebx >> 22) + 1;
-
-  assert (offset == 2);
-  return (ebx & 0xfff) + 1;
- }
-
-      ++round;
-    }
-  /* There is no other cache information anywhere else.  */
-  break;
- }
-      else
- {
-  if (byte == 0x49 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))
-    {
-      /* Intel reused this value.  For family 15, model 6 it
- specifies the 3rd level cache.  Otherwise the 2nd
- level cache.  */
-      unsigned int family = cpu_features->basic.family;
-      unsigned int model = cpu_features->basic.model;
-
-      if (family == 15 && model == 6)
- {
-  /* The level 3 cache is encoded for this model like
-     the level 2 cache is for other models.  Pretend
-     the caller asked for the level 2 cache.  */
-  name = (_SC_LEVEL2_CACHE_SIZE
-  + (name - _SC_LEVEL3_CACHE_SIZE));
-  folded_rel_name = M(_SC_LEVEL2_CACHE_SIZE);
- }
-    }
-
-  struct intel_02_cache_info *found;
-  struct intel_02_cache_info search;
-
-  search.idx = byte;
-  found = bsearch (&search, intel_02_known, nintel_02_known,
-   sizeof (intel_02_known[0]), intel_02_known_compare);
-  if (found != NULL)
-    {
-      if (found->rel_name == folded_rel_name)
- {
-  unsigned int offset = M(name) - folded_rel_name;
-
-  if (offset == 0)
-    /* Cache size.  */
-    return found->size;
-  if (offset == 1)
-    return found->assoc;
-
-  assert (offset == 2);
-  return found->linesize;
- }
-
-      if (found->rel_name == M(_SC_LEVEL2_CACHE_SIZE))
- *has_level_2 = true;
-    }
- }
-
-      /* Next byte for the next round.  */
-      value >>= 8;
-    }
-
-  /* Nothing found.  */
-  return 0;
-}
-
-
-static long int __attribute__ ((noinline))
-handle_intel (int name, const struct cpu_features *cpu_features)
-{
-  unsigned int maxidx = cpu_features->basic.max_cpuid;
-
-  /* Return -1 for older CPUs.  */
-  if (maxidx < 2)
-    return -1;
-
-  /* OK, we can use the CPUID instruction to get all info about the
-     caches.  */
-  unsigned int cnt = 0;
-  unsigned int max = 1;
-  long int result = 0;
-  bool no_level_2_or_3 = false;
-  bool has_level_2 = false;
-
-  while (cnt++ < max)
-    {
-      unsigned int eax;
-      unsigned int ebx;
-      unsigned int ecx;
-      unsigned int edx;
-      __cpuid (2, eax, ebx, ecx, edx);
-
-      /* The low byte of EAX in the first round contain the number of
- rounds we have to make.  At least one, the one we are already
- doing.  */
-      if (cnt == 1)
- {
-  max = eax & 0xff;
-  eax &= 0xffffff00;
- }
-
-      /* Process the individual registers' value.  */
-      result = intel_check_word (name, eax, &has_level_2,
- &no_level_2_or_3, cpu_features);
-      if (result != 0)
- return result;
-
-      result = intel_check_word (name, ebx, &has_level_2,
- &no_level_2_or_3, cpu_features);
-      if (result != 0)
- return result;
-
-      result = intel_check_word (name, ecx, &has_level_2,
- &no_level_2_or_3, cpu_features);
-      if (result != 0)
- return result;
-
-      result = intel_check_word (name, edx, &has_level_2,
- &no_level_2_or_3, cpu_features);
-      if (result != 0)
- return result;
-    }
-
-  if (name >= _SC_LEVEL2_CACHE_SIZE && name <= _SC_LEVEL3_CACHE_LINESIZE
-      && no_level_2_or_3)
-    return -1;
-
-  return 0;
-}
-
-
-static long int __attribute__ ((noinline))
-handle_amd (int name)
-{
-  unsigned int eax;
-  unsigned int ebx;
-  unsigned int ecx;
-  unsigned int edx;
-  __cpuid (0x80000000, eax, ebx, ecx, edx);
-
-  /* No level 4 cache (yet).  */
-  if (name > _SC_LEVEL3_CACHE_LINESIZE)
-    return 0;
-
-  unsigned int fn = 0x80000005 + (name >= _SC_LEVEL2_CACHE_SIZE);
-  if (eax < fn)
-    return 0;
-
-  __cpuid (fn, eax, ebx, ecx, edx);
-
-  if (name < _SC_LEVEL1_DCACHE_SIZE)
-    {
-      name += _SC_LEVEL1_DCACHE_SIZE - _SC_LEVEL1_ICACHE_SIZE;
-      ecx = edx;
-    }
-
-  switch (name)
-    {
-    case _SC_LEVEL1_DCACHE_SIZE:
-      return (ecx >> 14) & 0x3fc00;
-
-    case _SC_LEVEL1_DCACHE_ASSOC:
-      ecx >>= 16;
-      if ((ecx & 0xff) == 0xff)
- /* Fully associative.  */
- return (ecx << 2) & 0x3fc00;
-      return ecx & 0xff;
-
-    case _SC_LEVEL1_DCACHE_LINESIZE:
-      return ecx & 0xff;
-
-    case _SC_LEVEL2_CACHE_SIZE:
-      return (ecx & 0xf000) == 0 ? 0 : (ecx >> 6) & 0x3fffc00;
-
-    case _SC_LEVEL2_CACHE_ASSOC:
-      switch ((ecx >> 12) & 0xf)
- {
- case 0:
- case 1:
- case 2:
- case 4:
-  return (ecx >> 12) & 0xf;
- case 6:
-  return 8;
- case 8:
-  return 16;
- case 10:
-  return 32;
- case 11:
-  return 48;
- case 12:
-  return 64;
- case 13:
-  return 96;
- case 14:
-  return 128;
- case 15:
-  return ((ecx >> 6) & 0x3fffc00) / (ecx & 0xff);
- default:
-  return 0;
- }
-      /* NOTREACHED */
-
-    case _SC_LEVEL2_CACHE_LINESIZE:
-      return (ecx & 0xf000) == 0 ? 0 : ecx & 0xff;
-
-    case _SC_LEVEL3_CACHE_SIZE:
-      return (edx & 0xf000) == 0 ? 0 : (edx & 0x3ffc0000) << 1;
-
-    case _SC_LEVEL3_CACHE_ASSOC:
-      switch ((edx >> 12) & 0xf)
- {
- case 0:
- case 1:
- case 2:
- case 4:
-  return (edx >> 12) & 0xf;
- case 6:
-  return 8;
- case 8:
-  return 16;
- case 10:
-  return 32;
- case 11:
-  return 48;
- case 12:
-  return 64;
- case 13:
-  return 96;
- case 14:
-  return 128;
- case 15:
-  return ((edx & 0x3ffc0000) << 1) / (edx & 0xff);
- default:
-  return 0;
- }
-      /* NOTREACHED */
-
-    case _SC_LEVEL3_CACHE_LINESIZE:
-      return (edx & 0xf000) == 0 ? 0 : edx & 0xff;
-
-    default:
-      assert (! "cannot happen");
-    }
-  return -1;
-}
-
-
-static long int __attribute__ ((noinline))
-handle_zhaoxin (int name)
-{
-  unsigned int eax;
-  unsigned int ebx;
-  unsigned int ecx;
-  unsigned int edx;
-
-  int folded_rel_name = (M(name) / 3) * 3;
-
-  unsigned int round = 0;
-  while (1)
-    {
-      __cpuid_count (4, round, eax, ebx, ecx, edx);
-
-      enum { null = 0, data = 1, inst = 2, uni = 3 } type = eax & 0x1f;
-      if (type == null)
-        break;
-
-      unsigned int level = (eax >> 5) & 0x7;
-
-      if ((level == 1 && type == data
-        && folded_rel_name == M(_SC_LEVEL1_DCACHE_SIZE))
-        || (level == 1 && type == inst
-            && folded_rel_name == M(_SC_LEVEL1_ICACHE_SIZE))
-        || (level == 2 && folded_rel_name == M(_SC_LEVEL2_CACHE_SIZE))
-        || (level == 3 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE)))
-        {
-          unsigned int offset = M(name) - folded_rel_name;
-
-          if (offset == 0)
-            /* Cache size.  */
-            return (((ebx >> 22) + 1)
-                * (((ebx >> 12) & 0x3ff) + 1)
-                * ((ebx & 0xfff) + 1)
-                * (ecx + 1));
-          if (offset == 1)
-            return (ebx >> 22) + 1;
-
-          assert (offset == 2);
-          return (ebx & 0xfff) + 1;
-        }
-
-      ++round;
-    }
-
-  /* Nothing found.  */
-  return 0;
-}
-
-
-/* Get the value of the system variable NAME.  */
-long int
-attribute_hidden
-__cache_sysconf (int name)
-{
-  const struct cpu_features *cpu_features = __get_cpu_features ();
-
-  if (cpu_features->basic.kind == arch_kind_intel)
-    return handle_intel (name, cpu_features);
-
-  if (cpu_features->basic.kind == arch_kind_amd)
-    return handle_amd (name);
-
-  if (cpu_features->basic.kind == arch_kind_zhaoxin)
-    return handle_zhaoxin (name);
-
-  // XXX Fill in more vendors.
-
-  /* CPU not known, we have no information.  */
-  return 0;
-}
-
-
 /* Data cache size for use in memory and string routines, typically
    L1 size, rounded to multiple of 256 bytes.  */
 long int __x86_data_cache_size_half attribute_hidden = 32 * 1024 / 2;
@@ -530,348 +41,85 @@ long int __x86_raw_shared_cache_size attribute_hidden = 1024 * 1024;
 /* Threshold to use non temporal store.  */
 long int __x86_shared_non_temporal_threshold attribute_hidden;
 
-#ifndef DISABLE_PREFETCHW
+#ifndef __x86_64__
 /* PREFETCHW support flag for use in memory and string routines.  */
 int __x86_prefetchw attribute_hidden;
 #endif
 
-
-static void
-get_common_cache_info (long int *shared_ptr, unsigned int *threads_ptr,
-                long int core)
+/* Get the value of the system variable NAME.  */
+long int
+attribute_hidden
+__cache_sysconf (int name)
 {
-  unsigned int eax;
-  unsigned int ebx;
-  unsigned int ecx;
-  unsigned int edx;
-
-  /* Number of logical processors sharing L2 cache.  */
-  int threads_l2;
-
-  /* Number of logical processors sharing L3 cache.  */
-  int threads_l3;
-
   const struct cpu_features *cpu_features = __get_cpu_features ();
-  int max_cpuid = cpu_features->basic.max_cpuid;
-  unsigned int family = cpu_features->basic.family;
-  unsigned int model = cpu_features->basic.model;
-  long int shared = *shared_ptr;
-  unsigned int threads = *threads_ptr;
-  bool inclusive_cache = true;
-  bool support_count_mask = true;
-
-  /* Try L3 first.  */
-  unsigned int level = 3;
-
-  if (cpu_features->basic.kind == arch_kind_zhaoxin && family == 6)
-    support_count_mask = false;
-
-  if (shared <= 0)
-    {
-      /* Try L2 otherwise.  */
-      level  = 2;
-      shared = core;
-      threads_l2 = 0;
-      threads_l3 = -1;
-    }
-  else
-    {
-      threads_l2 = 0;
-      threads_l3 = 0;
-    }
-
-  /* A value of 0 for the HTT bit indicates there is only a single
-     logical processor.  */
-  if (HAS_CPU_FEATURE (HTT))
+  switch (name)
     {
-      /* Figure out the number of logical threads that share the
-         highest cache level.  */
-      if (max_cpuid >= 4)
-        {
-          int i = 0;
-
-          /* Query until cache level 2 and 3 are enumerated.  */
-          int check = 0x1 | (threads_l3 == 0) << 1;
-          do
-            {
-              __cpuid_count (4, i++, eax, ebx, ecx, edx);
+    case _SC_LEVEL1_ICACHE_SIZE:
+      return cpu_features->level1_icache_size;
 
-              /* There seems to be a bug in at least some Pentium Ds
-                 which sometimes fail to iterate all cache parameters.
-                 Do not loop indefinitely here, stop in this case and
-                 assume there is no such information.  */
-              if (cpu_features->basic.kind == arch_kind_intel
-                  && (eax & 0x1f) == 0 )
-                goto intel_bug_no_cache_info;
+    case _SC_LEVEL1_DCACHE_SIZE:
+      return cpu_features->level1_dcache_size;
 
-              switch ((eax >> 5) & 0x7)
-                {
-                  default:
-                    break;
-                  case 2:
-                    if ((check & 0x1))
-                      {
-                        /* Get maximum number of logical processors
-                           sharing L2 cache.  */
-                        threads_l2 = (eax >> 14) & 0x3ff;
-                        check &= ~0x1;
-                      }
-                    break;
-                  case 3:
-                    if ((check & (0x1 << 1)))
-                      {
-                        /* Get maximum number of logical processors
-                           sharing L3 cache.  */
-                        threads_l3 = (eax >> 14) & 0x3ff;
+    case _SC_LEVEL1_DCACHE_ASSOC:
+      return cpu_features->level1_dcache_assoc;
 
-                        /* Check if L2 and L3 caches are inclusive.  */
-                        inclusive_cache = (edx & 0x2) != 0;
-                        check &= ~(0x1 << 1);
-                      }
-                    break;
-                }
-            }
-          while (check);
+    case _SC_LEVEL1_DCACHE_LINESIZE:
+      return cpu_features->level1_dcache_linesize;
 
-          /* If max_cpuid >= 11, THREADS_L2/THREADS_L3 are the maximum
-             numbers of addressable IDs for logical processors sharing
-             the cache, instead of the maximum number of threads
-             sharing the cache.  */
-          if (max_cpuid >= 11 && support_count_mask)
-            {
-              /* Find the number of logical processors shipped in
-                 one core and apply count mask.  */
-              i = 0;
+    case _SC_LEVEL2_CACHE_SIZE:
+      return cpu_features->level2_cache_size;
 
-              /* Count SMT only if there is L3 cache.  Always count
-                 core if there is no L3 cache.  */
-              int count = ((threads_l2 > 0 && level == 3)
-                           | ((threads_l3 > 0
-                               || (threads_l2 > 0 && level == 2)) << 1));
+    case _SC_LEVEL2_CACHE_ASSOC:
+      return cpu_features->level2_cache_assoc;
 
-              while (count)
-                {
-                  __cpuid_count (11, i++, eax, ebx, ecx, edx);
+    case _SC_LEVEL2_CACHE_LINESIZE:
+      return cpu_features->level2_cache_linesize;
 
-                  int shipped = ebx & 0xff;
-                  int type = ecx & 0xff00;
-                  if (shipped == 0 || type == 0)
-                    break;
-                  else if (type == 0x100)
-                    {
-                      /* Count SMT.  */
-                      if ((count & 0x1))
-                        {
-                          int count_mask;
+    case _SC_LEVEL3_CACHE_SIZE:
+      return cpu_features->level3_cache_size;
 
-                          /* Compute count mask.  */
-                          asm ("bsr %1, %0"
-                               : "=r" (count_mask) : "g" (threads_l2));
-                          count_mask = ~(-1 << (count_mask + 1));
-                          threads_l2 = (shipped - 1) & count_mask;
-                          count &= ~0x1;
-                        }
-                    }
-                  else if (type == 0x200)
-                    {
-                      /* Count core.  */
-                      if ((count & (0x1 << 1)))
-                        {
-                          int count_mask;
-                          int threads_core
-                            = (level == 2 ? threads_l2 : threads_l3);
+    case _SC_LEVEL3_CACHE_ASSOC:
+      return cpu_features->level3_cache_assoc;
 
-                          /* Compute count mask.  */
-                          asm ("bsr %1, %0"
-                               : "=r" (count_mask) : "g" (threads_core));
-                          count_mask = ~(-1 << (count_mask + 1));
-                          threads_core = (shipped - 1) & count_mask;
-                          if (level == 2)
-                            threads_l2 = threads_core;
-                          else
-                            threads_l3 = threads_core;
-                          count &= ~(0x1 << 1);
-                        }
-                    }
-                }
-            }
-          if (threads_l2 > 0)
-            threads_l2 += 1;
-          if (threads_l3 > 0)
-            threads_l3 += 1;
-          if (level == 2)
-            {
-              if (threads_l2)
-                {
-                  threads = threads_l2;
-                  if (cpu_features->basic.kind == arch_kind_intel
-                      && threads > 2
-                      && family == 6)
-                    switch (model)
-                      {
-                        case 0x37:
-                        case 0x4a:
-                        case 0x4d:
-                        case 0x5a:
-                        case 0x5d:
-                          /* Silvermont has L2 cache shared by 2 cores.  */
-                          threads = 2;
-                          break;
-                        default:
-                          break;
-                      }
-                }
-            }
-          else if (threads_l3)
-            threads = threads_l3;
-        }
-      else
-        {
-intel_bug_no_cache_info:
-          /* Assume that all logical threads share the highest cache
-             level.  */
-          threads
-            = ((cpu_features->cpuid[COMMON_CPUID_INDEX_1].ebx
-                >> 16) & 0xff);
-        }
+    case _SC_LEVEL3_CACHE_LINESIZE:
+      return cpu_features->level3_cache_linesize;
 
-        /* Cap usage of highest cache level to the number of supported
-           threads.  */
-        if (shared > 0 && threads > 0)
-          shared /= threads;
-    }
+    case _SC_LEVEL4_CACHE_SIZE:
+      return cpu_features->level4_cache_size;
 
-  /* Account for non-inclusive L2 and L3 caches.  */
-  if (!inclusive_cache)
-    {
-      if (threads_l2 > 0)
-        core /= threads_l2;
-      shared += core;
+    default:
+      break;
     }
-
-  *shared_ptr = shared;
-  *threads_ptr = threads;
+  return -1;
 }
 
-
 static void
 __attribute__((constructor))
 init_cacheinfo (void)
 {
-  /* Find out what brand of processor.  */
-  unsigned int ebx;
-  unsigned int ecx;
-  unsigned int edx;
-  int max_cpuid_ex;
-  long int data = -1;
-  long int shared = -1;
-  long int core;
-  unsigned int threads = 0;
   const struct cpu_features *cpu_features = __get_cpu_features ();
+  long int data = cpu_features->data_cache_size;
+  __x86_raw_data_cache_size_half = data / 2;
+  __x86_raw_data_cache_size = data;
+  /* Round data cache size to multiple of 256 bytes.  */
+  data = data & ~255L;
+  __x86_data_cache_size_half = data / 2;
+  __x86_data_cache_size = data;
+
+  long int shared = cpu_features->shared_cache_size;
+  __x86_raw_shared_cache_size_half = shared / 2;
+  __x86_raw_shared_cache_size = shared;
+  /* Round shared cache size to multiple of 256 bytes.  */
+  shared = shared & ~255L;
+  __x86_shared_cache_size_half = shared / 2;
+  __x86_shared_cache_size = shared;
 
-  if (cpu_features->basic.kind == arch_kind_intel)
-    {
-      data = handle_intel (_SC_LEVEL1_DCACHE_SIZE, cpu_features);
-      core = handle_intel (_SC_LEVEL2_CACHE_SIZE, cpu_features);
-      shared = handle_intel (_SC_LEVEL3_CACHE_SIZE, cpu_features);
-
-      get_common_cache_info (&shared, &threads, core);
-    }
-  else if (cpu_features->basic.kind == arch_kind_zhaoxin)
-    {
-      data = handle_zhaoxin (_SC_LEVEL1_DCACHE_SIZE);
-      core = handle_zhaoxin (_SC_LEVEL2_CACHE_SIZE);
-      shared = handle_zhaoxin (_SC_LEVEL3_CACHE_SIZE);
-
-      get_common_cache_info (&shared, &threads, core);
-    }
-  else if (cpu_features->basic.kind == arch_kind_amd)
-    {
-      data   = handle_amd (_SC_LEVEL1_DCACHE_SIZE);
-      long int core = handle_amd (_SC_LEVEL2_CACHE_SIZE);
-      shared = handle_amd (_SC_LEVEL3_CACHE_SIZE);
-
-      /* Get maximum extended function. */
-      __cpuid (0x80000000, max_cpuid_ex, ebx, ecx, edx);
-
-      if (shared <= 0)
- /* No shared L3 cache.  All we have is the L2 cache.  */
- shared = core;
-      else
- {
-  /* Figure out the number of logical threads that share L3.  */
-  if (max_cpuid_ex >= 0x80000008)
-    {
-      /* Get width of APIC ID.  */
-      __cpuid (0x80000008, max_cpuid_ex, ebx, ecx, edx);
-      threads = 1 << ((ecx >> 12) & 0x0f);
-    }
-
-  if (threads == 0)
-    {
-      /* If APIC ID width is not available, use logical
- processor count.  */
-      __cpuid (0x00000001, max_cpuid_ex, ebx, ecx, edx);
-
-      if ((edx & (1 << 28)) != 0)
- threads = (ebx >> 16) & 0xff;
-    }
-
-  /* Cap usage of highest cache level to the number of
-     supported threads.  */
-  if (threads > 0)
-    shared /= threads;
-
-  /* Account for exclusive L2 and L3 caches.  */
-  shared += core;
- }
+  __x86_shared_non_temporal_threshold
+    = cpu_features->non_temporal_threshold;
 
-#ifndef DISABLE_PREFETCHW
-      if (max_cpuid_ex >= 0x80000001)
- {
-  unsigned int eax;
-  __cpuid (0x80000001, eax, ebx, ecx, edx);
-  /*  PREFETCHW     || 3DNow!  */
-  if ((ecx & 0x100) || (edx & 0x80000000))
-    __x86_prefetchw = -1;
- }
+#ifndef __x86_64__
+  __x86_prefetchw = cpu_features->prefetchw;
 #endif
-    }
-
-  if (cpu_features->data_cache_size != 0)
-    data = cpu_features->data_cache_size;
-
-  if (data > 0)
-    {
-      __x86_raw_data_cache_size_half = data / 2;
-      __x86_raw_data_cache_size = data;
-      /* Round data cache size to multiple of 256 bytes.  */
-      data = data & ~255L;
-      __x86_data_cache_size_half = data / 2;
-      __x86_data_cache_size = data;
-    }
-
-  if (cpu_features->shared_cache_size != 0)
-    shared = cpu_features->shared_cache_size;
-
-  if (shared > 0)
-    {
-      __x86_raw_shared_cache_size_half = shared / 2;
-      __x86_raw_shared_cache_size = shared;
-      /* Round shared cache size to multiple of 256 bytes.  */
-      shared = shared & ~255L;
-      __x86_shared_cache_size_half = shared / 2;
-      __x86_shared_cache_size = shared;
-    }
-
-  /* The large memcpy micro benchmark in glibc shows that 6 times of
-     shared cache size is the approximate value above which non-temporal
-     store becomes faster on a 8-core processor.  This is the 3/4 of the
-     total shared cache size.  */
-  __x86_shared_non_temporal_threshold
-    = (cpu_features->non_temporal_threshold != 0
-       ? cpu_features->non_temporal_threshold
-       : __x86_shared_cache_size * threads * 3 / 4);
 }
 
 #endif
diff --git a/sysdeps/x86/cpu-features.c b/sysdeps/x86/cpu-features.c
index c351bdd54a..e718204c18 100644
--- a/sysdeps/x86/cpu-features.c
+++ b/sysdeps/x86/cpu-features.c
@@ -19,6 +19,7 @@
 #include <cpuid.h>
 #include <cpu-features.h>
 #include <dl-hwcap.h>
+#include <init-arch.h>
 #include <libc-pointer-arith.h>
 
 #if HAVE_TUNABLES
@@ -602,20 +603,14 @@ no_cpuid:
   cpu_features->basic.model = model;
   cpu_features->basic.stepping = stepping;
 
+  __init_cacheinfo ();
+
 #if HAVE_TUNABLES
   TUNABLE_GET (hwcaps, tunable_val_t *, TUNABLE_CALLBACK (set_hwcaps));
-  cpu_features->non_temporal_threshold
-    = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
-  cpu_features->data_cache_size
-    = TUNABLE_GET (x86_data_cache_size, long int, NULL);
-  cpu_features->shared_cache_size
-    = TUNABLE_GET (x86_shared_cache_size, long int, NULL);
-#endif
-
-  /* Reuse dl_platform, dl_hwcap and dl_hwcap_mask for x86.  */
-#if !HAVE_TUNABLES && defined SHARED
-  /* The glibc.cpu.hwcap_mask tunable is initialized already, so no need to do
-     this.  */
+#elif defined SHARED
+  /* Reuse dl_platform, dl_hwcap and dl_hwcap_mask for x86.  The
+     glibc.cpu.hwcap_mask tunable is initialized already, so no
+     need to do this.  */
   GLRO(dl_hwcap_mask) = HWCAP_IMPORTANT;
 #endif
 
diff --git a/sysdeps/x86/cpu-features.h b/sysdeps/x86/cpu-features.h
index d66dc206f7..3aaed33cbc 100644
--- a/sysdeps/x86/cpu-features.h
+++ b/sysdeps/x86/cpu-features.h
@@ -102,6 +102,32 @@ struct cpu_features
   unsigned long int shared_cache_size;
   /* Threshold to use non temporal store.  */
   unsigned long int non_temporal_threshold;
+  /* _SC_LEVEL1_ICACHE_SIZE.  */
+  unsigned long int level1_icache_size;
+  /* _SC_LEVEL1_DCACHE_SIZE.  */
+  unsigned long int level1_dcache_size;
+  /* _SC_LEVEL1_DCACHE_ASSOC.  */
+  unsigned long int level1_dcache_assoc;
+  /* _SC_LEVEL1_DCACHE_LINESIZE.  */
+  unsigned long int level1_dcache_linesize;
+  /* _SC_LEVEL2_CACHE_ASSOC.  */
+  unsigned long int level2_cache_size;
+  /* _SC_LEVEL2_DCACHE_ASSOC.  */
+  unsigned long int level2_cache_assoc;
+  /* _SC_LEVEL2_CACHE_LINESIZE.  */
+  unsigned long int level2_cache_linesize;
+  /* /_SC_LEVEL3_CACHE_SIZE.  */
+  unsigned long int level3_cache_size;
+  /* _SC_LEVEL3_CACHE_ASSOC.  */
+  unsigned long int level3_cache_assoc;
+  /* _SC_LEVEL3_CACHE_LINESIZE.  */
+  unsigned long int level3_cache_linesize;
+  /* /_SC_LEVEL4_CACHE_SIZE.  */
+  unsigned long int level4_cache_size;
+#ifndef __x86_64__
+  /* PREFETCHW support flag for use in memory and string routines.  */
+  unsigned long int prefetchw;
+#endif
 };
 
 /* Used from outside of glibc to get access to the CPU features
diff --git a/sysdeps/x86/dl-cacheinfo.c b/sysdeps/x86/dl-cacheinfo.c
new file mode 100644
index 0000000000..2a1b1b9fc2
--- /dev/null
+++ b/sysdeps/x86/dl-cacheinfo.c
@@ -0,0 +1,879 @@
+/* x86 cache info.
+   Copyright (C) 2020 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
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include <assert.h>
+#include <stdbool.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <cpuid.h>
+#include <init-arch.h>
+#if HAVE_TUNABLES
+# define TUNABLE_NAMESPACE cpu
+# include <elf/dl-tunables.h>
+#endif
+
+static const struct intel_02_cache_info
+{
+  unsigned char idx;
+  unsigned char assoc;
+  unsigned char linesize;
+  unsigned char rel_name;
+  unsigned int size;
+} intel_02_known [] =
+  {
+#define M(sc) ((sc) - _SC_LEVEL1_ICACHE_SIZE)
+    { 0x06,  4, 32, M(_SC_LEVEL1_ICACHE_SIZE),    8192 },
+    { 0x08,  4, 32, M(_SC_LEVEL1_ICACHE_SIZE),   16384 },
+    { 0x09,  4, 32, M(_SC_LEVEL1_ICACHE_SIZE),   32768 },
+    { 0x0a,  2, 32, M(_SC_LEVEL1_DCACHE_SIZE),    8192 },
+    { 0x0c,  4, 32, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
+    { 0x0d,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
+    { 0x0e,  6, 64, M(_SC_LEVEL1_DCACHE_SIZE),   24576 },
+    { 0x21,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
+    { 0x22,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),   524288 },
+    { 0x23,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  1048576 },
+    { 0x25,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
+    { 0x29,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
+    { 0x2c,  8, 64, M(_SC_LEVEL1_DCACHE_SIZE),   32768 },
+    { 0x30,  8, 64, M(_SC_LEVEL1_ICACHE_SIZE),   32768 },
+    { 0x39,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
+    { 0x3a,  6, 64, M(_SC_LEVEL2_CACHE_SIZE),   196608 },
+    { 0x3b,  2, 64, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
+    { 0x3c,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
+    { 0x3d,  6, 64, M(_SC_LEVEL2_CACHE_SIZE),   393216 },
+    { 0x3e,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
+    { 0x3f,  2, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
+    { 0x41,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
+    { 0x42,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
+    { 0x43,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
+    { 0x44,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
+    { 0x45,  4, 32, M(_SC_LEVEL2_CACHE_SIZE),  2097152 },
+    { 0x46,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
+    { 0x47,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
+    { 0x48, 12, 64, M(_SC_LEVEL2_CACHE_SIZE),  3145728 },
+    { 0x49, 16, 64, M(_SC_LEVEL2_CACHE_SIZE),  4194304 },
+    { 0x4a, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  6291456 },
+    { 0x4b, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
+    { 0x4c, 12, 64, M(_SC_LEVEL3_CACHE_SIZE), 12582912 },
+    { 0x4d, 16, 64, M(_SC_LEVEL3_CACHE_SIZE), 16777216 },
+    { 0x4e, 24, 64, M(_SC_LEVEL2_CACHE_SIZE),  6291456 },
+    { 0x60,  8, 64, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
+    { 0x66,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),    8192 },
+    { 0x67,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),   16384 },
+    { 0x68,  4, 64, M(_SC_LEVEL1_DCACHE_SIZE),   32768 },
+    { 0x78,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
+    { 0x79,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   131072 },
+    { 0x7a,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
+    { 0x7b,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
+    { 0x7c,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
+    { 0x7d,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  2097152 },
+    { 0x7f,  2, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
+    { 0x80,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
+    { 0x82,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),   262144 },
+    { 0x83,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
+    { 0x84,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
+    { 0x85,  8, 32, M(_SC_LEVEL2_CACHE_SIZE),  2097152 },
+    { 0x86,  4, 64, M(_SC_LEVEL2_CACHE_SIZE),   524288 },
+    { 0x87,  8, 64, M(_SC_LEVEL2_CACHE_SIZE),  1048576 },
+    { 0xd0,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),   524288 },
+    { 0xd1,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),  1048576 },
+    { 0xd2,  4, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
+    { 0xd6,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  1048576 },
+    { 0xd7,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
+    { 0xd8,  8, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
+    { 0xdc, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
+    { 0xdd, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
+    { 0xde, 12, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
+    { 0xe2, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  2097152 },
+    { 0xe3, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  4194304 },
+    { 0xe4, 16, 64, M(_SC_LEVEL3_CACHE_SIZE),  8388608 },
+    { 0xea, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 12582912 },
+    { 0xeb, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 18874368 },
+    { 0xec, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 25165824 },
+  };
+
+#define nintel_02_known (sizeof (intel_02_known) / sizeof (intel_02_known [0]))
+
+static int
+intel_02_known_compare (const void *p1, const void *p2)
+{
+  const struct intel_02_cache_info *i1;
+  const struct intel_02_cache_info *i2;
+
+  i1 = (const struct intel_02_cache_info *) p1;
+  i2 = (const struct intel_02_cache_info *) p2;
+
+  if (i1->idx == i2->idx)
+    return 0;
+
+  return i1->idx < i2->idx ? -1 : 1;
+}
+
+
+static long int
+__attribute__ ((noinline))
+intel_check_word (int name, unsigned int value, bool *has_level_2,
+  bool *no_level_2_or_3,
+  const struct cpu_features *cpu_features)
+{
+  if ((value & 0x80000000) != 0)
+    /* The register value is reserved.  */
+    return 0;
+
+  /* Fold the name.  The _SC_ constants are always in the order SIZE,
+     ASSOC, LINESIZE.  */
+  int folded_rel_name = (M(name) / 3) * 3;
+
+  while (value != 0)
+    {
+      unsigned int byte = value & 0xff;
+
+      if (byte == 0x40)
+ {
+  *no_level_2_or_3 = true;
+
+  if (folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))
+    /* No need to look further.  */
+    break;
+ }
+      else if (byte == 0xff)
+ {
+  /* CPUID leaf 0x4 contains all the information.  We need to
+     iterate over it.  */
+  unsigned int eax;
+  unsigned int ebx;
+  unsigned int ecx;
+  unsigned int edx;
+
+  unsigned int round = 0;
+  while (1)
+    {
+      __cpuid_count (4, round, eax, ebx, ecx, edx);
+
+      enum { null = 0, data = 1, inst = 2, uni = 3 } type = eax & 0x1f;
+      if (type == null)
+ /* That was the end.  */
+ break;
+
+      unsigned int level = (eax >> 5) & 0x7;
+
+      if ((level == 1 && type == data
+   && folded_rel_name == M(_SC_LEVEL1_DCACHE_SIZE))
+  || (level == 1 && type == inst
+      && folded_rel_name == M(_SC_LEVEL1_ICACHE_SIZE))
+  || (level == 2 && folded_rel_name == M(_SC_LEVEL2_CACHE_SIZE))
+  || (level == 3 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))
+  || (level == 4 && folded_rel_name == M(_SC_LEVEL4_CACHE_SIZE)))
+ {
+  unsigned int offset = M(name) - folded_rel_name;
+
+  if (offset == 0)
+    /* Cache size.  */
+    return (((ebx >> 22) + 1)
+    * (((ebx >> 12) & 0x3ff) + 1)
+    * ((ebx & 0xfff) + 1)
+    * (ecx + 1));
+  if (offset == 1)
+    return (ebx >> 22) + 1;
+
+  assert (offset == 2);
+  return (ebx & 0xfff) + 1;
+ }
+
+      ++round;
+    }
+  /* There is no other cache information anywhere else.  */
+  break;
+ }
+      else
+ {
+  if (byte == 0x49 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))
+    {
+      /* Intel reused this value.  For family 15, model 6 it
+ specifies the 3rd level cache.  Otherwise the 2nd
+ level cache.  */
+      unsigned int family = cpu_features->basic.family;
+      unsigned int model = cpu_features->basic.model;
+
+      if (family == 15 && model == 6)
+ {
+  /* The level 3 cache is encoded for this model like
+     the level 2 cache is for other models.  Pretend
+     the caller asked for the level 2 cache.  */
+  name = (_SC_LEVEL2_CACHE_SIZE
+  + (name - _SC_LEVEL3_CACHE_SIZE));
+  folded_rel_name = M(_SC_LEVEL2_CACHE_SIZE);
+ }
+    }
+
+  struct intel_02_cache_info *found;
+  struct intel_02_cache_info search;
+
+  search.idx = byte;
+  found = bsearch (&search, intel_02_known, nintel_02_known,
+   sizeof (intel_02_known[0]), intel_02_known_compare);
+  if (found != NULL)
+    {
+      if (found->rel_name == folded_rel_name)
+ {
+  unsigned int offset = M(name) - folded_rel_name;
+
+  if (offset == 0)
+    /* Cache size.  */
+    return found->size;
+  if (offset == 1)
+    return found->assoc;
+
+  assert (offset == 2);
+  return found->linesize;
+ }
+
+      if (found->rel_name == M(_SC_LEVEL2_CACHE_SIZE))
+ *has_level_2 = true;
+    }
+ }
+
+      /* Next byte for the next round.  */
+      value >>= 8;
+    }
+
+  /* Nothing found.  */
+  return 0;
+}
+
+
+static long int __attribute__ ((noinline))
+handle_intel (int name, const struct cpu_features *cpu_features)
+{
+  unsigned int maxidx = cpu_features->basic.max_cpuid;
+
+  /* Return -1 for older CPUs.  */
+  if (maxidx < 2)
+    return -1;
+
+  /* OK, we can use the CPUID instruction to get all info about the
+     caches.  */
+  unsigned int cnt = 0;
+  unsigned int max = 1;
+  long int result = 0;
+  bool no_level_2_or_3 = false;
+  bool has_level_2 = false;
+
+  while (cnt++ < max)
+    {
+      unsigned int eax;
+      unsigned int ebx;
+      unsigned int ecx;
+      unsigned int edx;
+      __cpuid (2, eax, ebx, ecx, edx);
+
+      /* The low byte of EAX in the first round contain the number of
+ rounds we have to make.  At least one, the one we are already
+ doing.  */
+      if (cnt == 1)
+ {
+  max = eax & 0xff;
+  eax &= 0xffffff00;
+ }
+
+      /* Process the individual registers' value.  */
+      result = intel_check_word (name, eax, &has_level_2,
+ &no_level_2_or_3, cpu_features);
+      if (result != 0)
+ return result;
+
+      result = intel_check_word (name, ebx, &has_level_2,
+ &no_level_2_or_3, cpu_features);
+      if (result != 0)
+ return result;
+
+      result = intel_check_word (name, ecx, &has_level_2,
+ &no_level_2_or_3, cpu_features);
+      if (result != 0)
+ return result;
+
+      result = intel_check_word (name, edx, &has_level_2,
+ &no_level_2_or_3, cpu_features);
+      if (result != 0)
+ return result;
+    }
+
+  if (name >= _SC_LEVEL2_CACHE_SIZE && name <= _SC_LEVEL3_CACHE_LINESIZE
+      && no_level_2_or_3)
+    return -1;
+
+  return 0;
+}
+
+
+static long int __attribute__ ((noinline))
+handle_amd (int name)
+{
+  unsigned int eax;
+  unsigned int ebx;
+  unsigned int ecx;
+  unsigned int edx;
+  __cpuid (0x80000000, eax, ebx, ecx, edx);
+
+  /* No level 4 cache (yet).  */
+  if (name > _SC_LEVEL3_CACHE_LINESIZE)
+    return 0;
+
+  unsigned int fn = 0x80000005 + (name >= _SC_LEVEL2_CACHE_SIZE);
+  if (eax < fn)
+    return 0;
+
+  __cpuid (fn, eax, ebx, ecx, edx);
+
+  if (name < _SC_LEVEL1_DCACHE_SIZE)
+    {
+      name += _SC_LEVEL1_DCACHE_SIZE - _SC_LEVEL1_ICACHE_SIZE;
+      ecx = edx;
+    }
+
+  switch (name)
+    {
+    case _SC_LEVEL1_DCACHE_SIZE:
+      return (ecx >> 14) & 0x3fc00;
+
+    case _SC_LEVEL1_DCACHE_ASSOC:
+      ecx >>= 16;
+      if ((ecx & 0xff) == 0xff)
+ /* Fully associative.  */
+ return (ecx << 2) & 0x3fc00;
+      return ecx & 0xff;
+
+    case _SC_LEVEL1_DCACHE_LINESIZE:
+      return ecx & 0xff;
+
+    case _SC_LEVEL2_CACHE_SIZE:
+      return (ecx & 0xf000) == 0 ? 0 : (ecx >> 6) & 0x3fffc00;
+
+    case _SC_LEVEL2_CACHE_ASSOC:
+      switch ((ecx >> 12) & 0xf)
+ {
+ case 0:
+ case 1:
+ case 2:
+ case 4:
+  return (ecx >> 12) & 0xf;
+ case 6:
+  return 8;
+ case 8:
+  return 16;
+ case 10:
+  return 32;
+ case 11:
+  return 48;
+ case 12:
+  return 64;
+ case 13:
+  return 96;
+ case 14:
+  return 128;
+ case 15:
+  return ((ecx >> 6) & 0x3fffc00) / (ecx & 0xff);
+ default:
+  return 0;
+ }
+      /* NOTREACHED */
+
+    case _SC_LEVEL2_CACHE_LINESIZE:
+      return (ecx & 0xf000) == 0 ? 0 : ecx & 0xff;
+
+    case _SC_LEVEL3_CACHE_SIZE:
+      return (edx & 0xf000) == 0 ? 0 : (edx & 0x3ffc0000) << 1;
+
+    case _SC_LEVEL3_CACHE_ASSOC:
+      switch ((edx >> 12) & 0xf)
+ {
+ case 0:
+ case 1:
+ case 2:
+ case 4:
+  return (edx >> 12) & 0xf;
+ case 6:
+  return 8;
+ case 8:
+  return 16;
+ case 10:
+  return 32;
+ case 11:
+  return 48;
+ case 12:
+  return 64;
+ case 13:
+  return 96;
+ case 14:
+  return 128;
+ case 15:
+  return ((edx & 0x3ffc0000) << 1) / (edx & 0xff);
+ default:
+  return 0;
+ }
+      /* NOTREACHED */
+
+    case _SC_LEVEL3_CACHE_LINESIZE:
+      return (edx & 0xf000) == 0 ? 0 : edx & 0xff;
+
+    default:
+      assert (! "cannot happen");
+    }
+  return -1;
+}
+
+
+static long int __attribute__ ((noinline))
+handle_zhaoxin (int name)
+{
+  unsigned int eax;
+  unsigned int ebx;
+  unsigned int ecx;
+  unsigned int edx;
+
+  int folded_rel_name = (M(name) / 3) * 3;
+
+  unsigned int round = 0;
+  while (1)
+    {
+      __cpuid_count (4, round, eax, ebx, ecx, edx);
+
+      enum { null = 0, data = 1, inst = 2, uni = 3 } type = eax & 0x1f;
+      if (type == null)
+        break;
+
+      unsigned int level = (eax >> 5) & 0x7;
+
+      if ((level == 1 && type == data
+        && folded_rel_name == M(_SC_LEVEL1_DCACHE_SIZE))
+        || (level == 1 && type == inst
+            && folded_rel_name == M(_SC_LEVEL1_ICACHE_SIZE))
+        || (level == 2 && folded_rel_name == M(_SC_LEVEL2_CACHE_SIZE))
+        || (level == 3 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE)))
+        {
+          unsigned int offset = M(name) - folded_rel_name;
+
+          if (offset == 0)
+            /* Cache size.  */
+            return (((ebx >> 22) + 1)
+                * (((ebx >> 12) & 0x3ff) + 1)
+                * ((ebx & 0xfff) + 1)
+                * (ecx + 1));
+          if (offset == 1)
+            return (ebx >> 22) + 1;
+
+          assert (offset == 2);
+          return (ebx & 0xfff) + 1;
+        }
+
+      ++round;
+    }
+
+  /* Nothing found.  */
+  return 0;
+}
+
+
+static void
+get_common_cache_info (long int *shared_ptr, unsigned int *threads_ptr,
+                long int core)
+{
+  unsigned int eax;
+  unsigned int ebx;
+  unsigned int ecx;
+  unsigned int edx;
+
+  /* Number of logical processors sharing L2 cache.  */
+  int threads_l2;
+
+  /* Number of logical processors sharing L3 cache.  */
+  int threads_l3;
+
+  const struct cpu_features *cpu_features = __get_cpu_features ();
+  int max_cpuid = cpu_features->basic.max_cpuid;
+  unsigned int family = cpu_features->basic.family;
+  unsigned int model = cpu_features->basic.model;
+  long int shared = *shared_ptr;
+  unsigned int threads = *threads_ptr;
+  bool inclusive_cache = true;
+  bool support_count_mask = true;
+
+  /* Try L3 first.  */
+  unsigned int level = 3;
+
+  if (cpu_features->basic.kind == arch_kind_zhaoxin && family == 6)
+    support_count_mask = false;
+
+  if (shared <= 0)
+    {
+      /* Try L2 otherwise.  */
+      level  = 2;
+      shared = core;
+      threads_l2 = 0;
+      threads_l3 = -1;
+    }
+  else
+    {
+      threads_l2 = 0;
+      threads_l3 = 0;
+    }
+
+  /* A value of 0 for the HTT bit indicates there is only a single
+     logical processor.  */
+  if (HAS_CPU_FEATURE (HTT))
+    {
+      /* Figure out the number of logical threads that share the
+         highest cache level.  */
+      if (max_cpuid >= 4)
+        {
+          int i = 0;
+
+          /* Query until cache level 2 and 3 are enumerated.  */
+          int check = 0x1 | (threads_l3 == 0) << 1;
+          do
+            {
+              __cpuid_count (4, i++, eax, ebx, ecx, edx);
+
+              /* There seems to be a bug in at least some Pentium Ds
+                 which sometimes fail to iterate all cache parameters.
+                 Do not loop indefinitely here, stop in this case and
+                 assume there is no such information.  */
+              if (cpu_features->basic.kind == arch_kind_intel
+                  && (eax & 0x1f) == 0 )
+                goto intel_bug_no_cache_info;
+
+              switch ((eax >> 5) & 0x7)
+                {
+                  default:
+                    break;
+                  case 2:
+                    if ((check & 0x1))
+                      {
+                        /* Get maximum number of logical processors
+                           sharing L2 cache.  */
+                        threads_l2 = (eax >> 14) & 0x3ff;
+                        check &= ~0x1;
+                      }
+                    break;
+                  case 3:
+                    if ((check & (0x1 << 1)))
+                      {
+                        /* Get maximum number of logical processors
+                           sharing L3 cache.  */
+                        threads_l3 = (eax >> 14) & 0x3ff;
+
+                        /* Check if L2 and L3 caches are inclusive.  */
+                        inclusive_cache = (edx & 0x2) != 0;
+                        check &= ~(0x1 << 1);
+                      }
+                    break;
+                }
+            }
+          while (check);
+
+          /* If max_cpuid >= 11, THREADS_L2/THREADS_L3 are the maximum
+             numbers of addressable IDs for logical processors sharing
+             the cache, instead of the maximum number of threads
+             sharing the cache.  */
+          if (max_cpuid >= 11 && support_count_mask)
+            {
+              /* Find the number of logical processors shipped in
+                 one core and apply count mask.  */
+              i = 0;
+
+              /* Count SMT only if there is L3 cache.  Always count
+                 core if there is no L3 cache.  */
+              int count = ((threads_l2 > 0 && level == 3)
+                           | ((threads_l3 > 0
+                               || (threads_l2 > 0 && level == 2)) << 1));
+
+              while (count)
+                {
+                  __cpuid_count (11, i++, eax, ebx, ecx, edx);
+
+                  int shipped = ebx & 0xff;
+                  int type = ecx & 0xff00;
+                  if (shipped == 0 || type == 0)
+                    break;
+                  else if (type == 0x100)
+                    {
+                      /* Count SMT.  */
+                      if ((count & 0x1))
+                        {
+                          int count_mask;
+
+                          /* Compute count mask.  */
+                          asm ("bsr %1, %0"
+                               : "=r" (count_mask) : "g" (threads_l2));
+                          count_mask = ~(-1 << (count_mask + 1));
+                          threads_l2 = (shipped - 1) & count_mask;
+                          count &= ~0x1;
+                        }
+                    }
+                  else if (type == 0x200)
+                    {
+                      /* Count core.  */
+                      if ((count & (0x1 << 1)))
+                        {
+                          int count_mask;
+                          int threads_core
+                            = (level == 2 ? threads_l2 : threads_l3);
+
+                          /* Compute count mask.  */
+                          asm ("bsr %1, %0"
+                               : "=r" (count_mask) : "g" (threads_core));
+                          count_mask = ~(-1 << (count_mask + 1));
+                          threads_core = (shipped - 1) & count_mask;
+                          if (level == 2)
+                            threads_l2 = threads_core;
+                          else
+                            threads_l3 = threads_core;
+                          count &= ~(0x1 << 1);
+                        }
+                    }
+                }
+            }
+          if (threads_l2 > 0)
+            threads_l2 += 1;
+          if (threads_l3 > 0)
+            threads_l3 += 1;
+          if (level == 2)
+            {
+              if (threads_l2)
+                {
+                  threads = threads_l2;
+                  if (cpu_features->basic.kind == arch_kind_intel
+                      && threads > 2
+                      && family == 6)
+                    switch (model)
+                      {
+                        case 0x37:
+                        case 0x4a:
+                        case 0x4d:
+                        case 0x5a:
+                        case 0x5d:
+                          /* Silvermont has L2 cache shared by 2 cores.  */
+                          threads = 2;
+                          break;
+                        default:
+                          break;
+                      }
+                }
+            }
+          else if (threads_l3)
+            threads = threads_l3;
+        }
+      else
+        {
+intel_bug_no_cache_info:
+          /* Assume that all logical threads share the highest cache
+             level.  */
+          threads
+            = ((cpu_features->cpuid[COMMON_CPUID_INDEX_1].ebx
+                >> 16) & 0xff);
+        }
+
+        /* Cap usage of highest cache level to the number of supported
+           threads.  */
+        if (shared > 0 && threads > 0)
+          shared /= threads;
+    }
+
+  /* Account for non-inclusive L2 and L3 caches.  */
+  if (!inclusive_cache)
+    {
+      if (threads_l2 > 0)
+        core /= threads_l2;
+      shared += core;
+    }
+
+  *shared_ptr = shared;
+  *threads_ptr = threads;
+}
+
+void
+__init_cacheinfo (void)
+{
+  /* Find out what brand of processor.  */
+  unsigned int ebx;
+  unsigned int ecx;
+  unsigned int edx;
+  int max_cpuid_ex;
+  long int data = -1;
+  long int shared = -1;
+  long int core;
+  unsigned int threads = 0;
+  unsigned long int level1_icache_size = -1;
+  unsigned long int level1_dcache_size = -1;
+  unsigned long int level1_dcache_assoc = -1;
+  unsigned long int level1_dcache_linesize = -1;
+  unsigned long int level2_cache_size = -1;
+  unsigned long int level2_cache_assoc = -1;
+  unsigned long int level2_cache_linesize = -1;
+  unsigned long int level3_cache_size = -1;
+  unsigned long int level3_cache_assoc = -1;
+  unsigned long int level3_cache_linesize = -1;
+  unsigned long int level4_cache_size = -1;
+  struct cpu_features *cpu_features = __get_cpu_features ();
+
+  if (cpu_features->basic.kind == arch_kind_intel)
+    {
+      data = handle_intel (_SC_LEVEL1_DCACHE_SIZE, cpu_features);
+      core = handle_intel (_SC_LEVEL2_CACHE_SIZE, cpu_features);
+      shared = handle_intel (_SC_LEVEL3_CACHE_SIZE, cpu_features);
+
+      level1_icache_size
+ = handle_intel (_SC_LEVEL1_ICACHE_SIZE, cpu_features);
+      level1_dcache_size = data;
+      level1_dcache_assoc
+ = handle_intel (_SC_LEVEL1_DCACHE_ASSOC, cpu_features);
+      level1_dcache_linesize
+ = handle_intel (_SC_LEVEL1_DCACHE_LINESIZE, cpu_features);
+      level2_cache_size = core;
+      level2_cache_assoc
+ = handle_intel (_SC_LEVEL2_CACHE_ASSOC, cpu_features);
+      level2_cache_linesize
+ = handle_intel (_SC_LEVEL2_CACHE_LINESIZE, cpu_features);
+      level3_cache_size = shared;
+      level3_cache_assoc
+ = handle_intel (_SC_LEVEL3_CACHE_ASSOC, cpu_features);
+      level3_cache_linesize
+ = handle_intel (_SC_LEVEL3_CACHE_LINESIZE, cpu_features);
+      level4_cache_size
+ = handle_intel (_SC_LEVEL4_CACHE_SIZE, cpu_features);
+
+      get_common_cache_info (&shared, &threads, core);
+    }
+  else if (cpu_features->basic.kind == arch_kind_zhaoxin)
+    {
+      data = handle_zhaoxin (_SC_LEVEL1_DCACHE_SIZE);
+      core = handle_zhaoxin (_SC_LEVEL2_CACHE_SIZE);
+      shared = handle_zhaoxin (_SC_LEVEL3_CACHE_SIZE);
+
+      level1_icache_size = handle_zhaoxin (_SC_LEVEL1_ICACHE_SIZE);
+      level1_dcache_size = data;
+      level1_dcache_assoc = handle_zhaoxin (_SC_LEVEL1_DCACHE_ASSOC);
+      level1_dcache_linesize = handle_zhaoxin (_SC_LEVEL1_DCACHE_LINESIZE);
+      level2_cache_size = core;
+      level2_cache_assoc = handle_zhaoxin (_SC_LEVEL2_CACHE_ASSOC);
+      level2_cache_linesize = handle_zhaoxin (_SC_LEVEL2_CACHE_LINESIZE);
+      level3_cache_size = shared;
+      level3_cache_assoc = handle_zhaoxin (_SC_LEVEL3_CACHE_ASSOC);
+      level3_cache_linesize = handle_zhaoxin (_SC_LEVEL3_CACHE_LINESIZE);
+
+      get_common_cache_info (&shared, &threads, core);
+    }
+  else if (cpu_features->basic.kind == arch_kind_amd)
+    {
+      data  = handle_amd (_SC_LEVEL1_DCACHE_SIZE);
+      core = handle_amd (_SC_LEVEL2_CACHE_SIZE);
+      shared = handle_amd (_SC_LEVEL3_CACHE_SIZE);
+
+      level1_icache_size = handle_amd (_SC_LEVEL1_ICACHE_SIZE);
+      level1_dcache_size = data;
+      level1_dcache_assoc = handle_amd (_SC_LEVEL1_DCACHE_ASSOC);
+      level1_dcache_linesize = handle_amd (_SC_LEVEL1_DCACHE_LINESIZE);
+      level2_cache_size = core;
+      level2_cache_assoc = handle_amd (_SC_LEVEL2_CACHE_ASSOC);
+      level2_cache_linesize = handle_amd (_SC_LEVEL2_CACHE_LINESIZE);
+      level3_cache_size = shared;
+      level3_cache_assoc = handle_amd (_SC_LEVEL3_CACHE_ASSOC);
+      level3_cache_linesize = handle_amd (_SC_LEVEL3_CACHE_LINESIZE);
+
+      /* Get maximum extended function. */
+      __cpuid (0x80000000, max_cpuid_ex, ebx, ecx, edx);
+
+      if (shared <= 0)
+ /* No shared L3 cache.  All we have is the L2 cache.  */
+ shared = core;
+      else
+ {
+  /* Figure out the number of logical threads that share L3.  */
+  if (max_cpuid_ex >= 0x80000008)
+    {
+      /* Get width of APIC ID.  */
+      __cpuid (0x80000008, max_cpuid_ex, ebx, ecx, edx);
+      threads = 1 << ((ecx >> 12) & 0x0f);
+    }
+
+  if (threads == 0)
+    {
+      /* If APIC ID width is not available, use logical
+ processor count.  */
+      __cpuid (0x00000001, max_cpuid_ex, ebx, ecx, edx);
+
+      if ((edx & (1 << 28)) != 0)
+ threads = (ebx >> 16) & 0xff;
+    }
+
+  /* Cap usage of highest cache level to the number of
+     supported threads.  */
+  if (threads > 0)
+    shared /= threads;
+
+  /* Account for exclusive L2 and L3 caches.  */
+  shared += core;
+ }
+
+#ifndef __x86_64__
+      if (max_cpuid_ex >= 0x80000001)
+ {
+  unsigned int eax;
+  __cpuid (0x80000001, eax, ebx, ecx, edx);
+  /*  PREFETCHW     || 3DNow!  */
+  if ((ecx & 0x100) || (edx & 0x80000000))
+    cpu_features->prefetchw = -1;
+ }
+#endif
+    }
+
+  cpu_features->level1_icache_size = level1_icache_size;
+  cpu_features->level1_dcache_size = level1_dcache_size;
+  cpu_features->level1_dcache_assoc = level1_dcache_assoc;
+  cpu_features->level1_dcache_linesize = level1_dcache_linesize;
+  cpu_features->level2_cache_size = level2_cache_size;
+  cpu_features->level2_cache_assoc = level2_cache_assoc;
+  cpu_features->level2_cache_linesize = level2_cache_linesize;
+  cpu_features->level3_cache_size = level3_cache_size;
+  cpu_features->level3_cache_assoc = level3_cache_assoc;
+  cpu_features->level3_cache_linesize = level3_cache_linesize;
+  cpu_features->level4_cache_size = level4_cache_size;
+
+  /* The large memcpy micro benchmark in glibc shows that 6 times of
+     shared cache size is the approximate value above which non-temporal
+     store becomes faster on a 8-core processor.  This is the 3/4 of the
+     total shared cache size.  */
+  unsigned long int non_temporal_threshold = (shared * threads * 3 / 4);
+
+#if HAVE_TUNABLES
+  long int tunable_size;
+  tunable_size = TUNABLE_GET (x86_data_cache_size, long int, NULL);
+  if (tunable_size != 0)
+    data = tunable_size;
+  tunable_size = TUNABLE_GET (x86_shared_cache_size, long int, NULL);
+  if (tunable_size != 0)
+    shared = tunable_size;
+  tunable_size = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
+  if (tunable_size != 0)
+    non_temporal_threshold = tunable_size;
+#endif
+
+  cpu_features->data_cache_size = data;
+  cpu_features->shared_cache_size = shared;
+  cpu_features->non_temporal_threshold = non_temporal_threshold;
+}
diff --git a/sysdeps/x86/init-arch.h b/sysdeps/x86/init-arch.h
index d6f59cf962..272ed10902 100644
--- a/sysdeps/x86/init-arch.h
+++ b/sysdeps/x86/init-arch.h
@@ -23,6 +23,9 @@
 #include <ifunc-init.h>
 #include <isa.h>
 
+extern void __init_cacheinfo (void)
+  __attribute__ ((visibility ("hidden")));
+
 #ifndef __x86_64__
 /* Due to the reordering and the other nifty extensions in i686, it is
    not really good to use heavily i586 optimized code on an i686.  It's
--
2.26.2

Reply | Threaded
Open this post in threaded view
|

V2 [PATCH] x86: Add thresholds for "rep movsb/stosb" to tunables

Sourceware - libc-alpha mailing list
In reply to this post by Sourceware - libc-alpha mailing list
On Fri, Jul 03, 2020 at 03:49:21PM -0400, Carlos O'Donell wrote:

> On 7/3/20 1:52 PM, H.J. Lu wrote:
> > Add x86_rep_movsb_threshold and x86_rep_stosb_threshold to tunables
> > to update thresholds for "rep movsb" and "rep stosb" at run-time.
> >
> > Note that the user specified threshold for "rep movsb" smaller than the
> > minimum threshold will be ignored.
>
> Post v2 please. Almost there.
>
> > ---
> >  manual/tunables.texi                          | 14 +++++++
> >  sysdeps/x86/cacheinfo.c                       | 20 ++++++++++
> >  sysdeps/x86/cpu-features.h                    |  4 ++
> >  sysdeps/x86/dl-cacheinfo.c                    | 38 +++++++++++++++++++
> >  sysdeps/x86/dl-tunables.list                  |  6 +++
> >  .../multiarch/memmove-vec-unaligned-erms.S    | 16 +-------
> >  .../multiarch/memset-vec-unaligned-erms.S     | 12 +-----
> >  7 files changed, 84 insertions(+), 26 deletions(-)
> >
> > diff --git a/manual/tunables.texi b/manual/tunables.texi
> > index ec18b10834..61edd62425 100644
> > --- a/manual/tunables.texi
> > +++ b/manual/tunables.texi
> > @@ -396,6 +396,20 @@ to set threshold in bytes for non temporal store.
> >  This tunable is specific to i386 and x86-64.
> >  @end deftp
> >  
> > +@deftp Tunable glibc.cpu.x86_rep_movsb_threshold
> > +The @code{glibc.cpu.x86_rep_movsb_threshold} tunable allows the user
> > +to set threshold in bytes to start using "rep movsb".
> > +
> > +This tunable is specific to i386 and x86-64.
> > +@end deftp
> > +
> > +@deftp Tunable glibc.cpu.x86_rep_stosb_threshold
> > +The @code{glibc.cpu.x86_rep_stosb_threshold} tunable allows the user
> > +to set threshold in bytes to start using "rep stosb".
> > +
> > +This tunable is specific to i386 and x86-64.
> > +@end deftp
> > +
> >  @deftp Tunable glibc.cpu.x86_ibt
> >  The @code{glibc.cpu.x86_ibt} tunable allows the user to control how
> >  indirect branch tracking (IBT) should be enabled.  Accepted values are
> > diff --git a/sysdeps/x86/cacheinfo.c b/sysdeps/x86/cacheinfo.c
> > index 8c4c7f9972..bb536d96ef 100644
> > --- a/sysdeps/x86/cacheinfo.c
> > +++ b/sysdeps/x86/cacheinfo.c
> > @@ -41,6 +41,23 @@ long int __x86_raw_shared_cache_size attribute_hidden = 1024 * 1024;
> >  /* Threshold to use non temporal store.  */
> >  long int __x86_shared_non_temporal_threshold attribute_hidden;
> >  
> > +/* Threshold to use Enhanced REP MOVSB.  Since there is overhead to set
> > +   up REP MOVSB operation, REP MOVSB isn't faster on short data.  The
> > +   memcpy micro benchmark in glibc shows that 2KB is the approximate
> > +   value above which REP MOVSB becomes faster than SSE2 optimization
> > +   on processors with Enhanced REP MOVSB.  Since larger register size
> > +   can move more data with a single load and store, the threshold is
> > +   higher with larger register size.  */
> > +long int __x86_rep_movsb_threshold attribute_hidden = 2048;
> > +
> > +/* Threshold to use Enhanced REP STOSB.  Since there is overhead to set
> > +   up REP STOSB operation, REP STOSB isn't faster on short data.  The
> > +   memset micro benchmark in glibc shows that 2KB is the approximate
> > +   value above which REP STOSB becomes faster on processors with
> > +   Enhanced REP STOSB.  Since the stored value is fixed, larger register
> > +   size has minimal impact on threshold.  */
> > +long int __x86_rep_stosb_threshold attribute_hidden = 2048;
> > +
> >  #ifndef __x86_64__
> >  /* PREFETCHW support flag for use in memory and string routines.  */
> >  int __x86_prefetchw attribute_hidden;
> > @@ -117,6 +134,9 @@ init_cacheinfo (void)
> >    __x86_shared_non_temporal_threshold
> >      = cpu_features->non_temporal_threshold;
> >  
> > +  __x86_rep_movsb_threshold = cpu_features->rep_movsb_threshold;
> > +  __x86_rep_stosb_threshold = cpu_features->rep_stosb_threshold;
> > +
>
> OK. Update global from cpu_features with values.
>
> I would really like to see some kind of "assert (cpu_features->initialized);"
> that way we know we didn't break the startup sequence unintentionally.
>
> >  #ifndef __x86_64__
> >    __x86_prefetchw = cpu_features->prefetchw;
> >  #endif
> > diff --git a/sysdeps/x86/cpu-features.h b/sysdeps/x86/cpu-features.h
> > index 3aaed33cbc..002e12e11f 100644
> > --- a/sysdeps/x86/cpu-features.h
> > +++ b/sysdeps/x86/cpu-features.h
> > @@ -128,6 +128,10 @@ struct cpu_features
> >    /* PREFETCHW support flag for use in memory and string routines.  */
> >    unsigned long int prefetchw;
> >  #endif
> > +  /* Threshold to use "rep movsb".  */
> > +  unsigned long int rep_movsb_threshold;
> > +  /* Threshold to use "rep stosb".  */
> > +  unsigned long int rep_stosb_threshold;
>
> OK.
>
> >  };
> >  
> >  /* Used from outside of glibc to get access to the CPU features
> > diff --git a/sysdeps/x86/dl-cacheinfo.c b/sysdeps/x86/dl-cacheinfo.c
> > index 8e2a6f552c..aff9bd1067 100644
> > --- a/sysdeps/x86/dl-cacheinfo.c
> > +++ b/sysdeps/x86/dl-cacheinfo.c
> > @@ -860,6 +860,31 @@ __init_cacheinfo (void)
> >       total shared cache size.  */
> >    unsigned long int non_temporal_threshold = (shared * threads * 3 / 4);
> >  
> > +  /* NB: The REP MOVSB threshold must be greater than VEC_SIZE * 8.  */
> > +  unsigned long int minimum_rep_movsb_threshold;
> > +  /* NB: The default REP MOVSB threshold is 2048 * (VEC_SIZE / 16).  See
> > +     comments for __x86_rep_movsb_threshold in cacheinfo.c.  */
> > +  unsigned long int rep_movsb_threshold;
> > +  if (CPU_FEATURES_ARCH_P (cpu_features, AVX512F_Usable)
> > +      && !CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_AVX512))
> > +    {
> > +      rep_movsb_threshold = 2048 * (64 / 16);
> > +      minimum_rep_movsb_threshold = 64 * 8;
> > +    }
> > +  else if (CPU_FEATURES_ARCH_P (cpu_features,
> > + AVX_Fast_Unaligned_Load))
> > +    {
> > +      rep_movsb_threshold = 2048 * (32 / 16);
> > +      minimum_rep_movsb_threshold = 32 * 8;
> > +    }
> > +  else
> > +    {
> > +      rep_movsb_threshold = 2048 * (16 / 16);
> > +      minimum_rep_movsb_threshold = 16 * 8;
> > +    }
> > +  /* NB: See comments for __x86_rep_stosb_threshold in cacheinfo.c.  */
> > +  unsigned long int rep_stosb_threshold = 2048;
> > +
> >  #if HAVE_TUNABLES
> >    long int tunable_size;
> >    tunable_size = TUNABLE_GET (x86_data_cache_size, long int, NULL);
> > @@ -871,11 +896,19 @@ __init_cacheinfo (void)
> >    tunable_size = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
> >    if (tunable_size != 0)
> >      non_temporal_threshold = tunable_size;
>
> > +  tunable_size = TUNABLE_GET (x86_rep_movsb_threshold, long int, NULL);
> > +  if (tunable_size > minimum_rep_movsb_threshold)
> > +    rep_movsb_threshold = tunable_size;
>
> OK. Good, we only set rep_movsb_threshold if it's greater than min.
>
> > +  tunable_size = TUNABLE_GET (x86_rep_stosb_threshold, long int, NULL);
> > +  if (tunable_size != 0)
> > +    rep_stosb_threshold = tunable_size;
>
> This should be min=1, default=2048 in dl-tunables.list, and would remove
> this code since the range is not dynamic.
>
> The point of the tunables framework is to remove such boiler plate for
> range a default processing and clearing parameters for security settings.
>
> >  #endif
> >  
> >    cpu_features->data_cache_size = data;
> >    cpu_features->shared_cache_size = shared;
> >    cpu_features->non_temporal_threshold = non_temporal_threshold;
> > +  cpu_features->rep_movsb_threshold = rep_movsb_threshold;
> > +  cpu_features->rep_stosb_threshold = rep_stosb_threshold;
> >  
> >  #if HAVE_TUNABLES
> >    TUNABLE_UPDATE (x86_data_cache_size, long int,
> > @@ -884,5 +917,10 @@ __init_cacheinfo (void)
> >    shared, 0, (long int) -1);
> >    TUNABLE_UPDATE (x86_non_temporal_threshold, long int,
> >    non_temporal_threshold, 0, (long int) -1);
> > +  TUNABLE_UPDATE (x86_rep_movsb_threshold, long int,
> > +  rep_movsb_threshold, minimum_rep_movsb_threshold,
> > +  (long int) -1);
>
> OK. Store the new value and the computed minimum.
>
> > +  TUNABLE_UPDATE (x86_rep_stosb_threshold, long int,
> > +  rep_stosb_threshold, 0, (long int) -1);
>
> This one can be deleted.
>

We should go with this simple one for 2.32.


H.J.
---
Add x86_rep_movsb_threshold and x86_rep_stosb_threshold to tunables
to update thresholds for "rep movsb" and "rep stosb" at run-time.

Note that the user specified threshold for "rep movsb" smaller than
the minimum threshold will be ignored.
---
 manual/tunables.texi                          | 14 ++++++
 sysdeps/x86/cacheinfo.c                       | 46 +++++++++++++++++++
 sysdeps/x86/cpu-features.c                    |  4 ++
 sysdeps/x86/cpu-features.h                    |  4 ++
 sysdeps/x86/dl-tunables.list                  |  6 +++
 .../multiarch/memmove-vec-unaligned-erms.S    | 16 +------
 .../multiarch/memset-vec-unaligned-erms.S     | 12 +----
 7 files changed, 76 insertions(+), 26 deletions(-)

diff --git a/manual/tunables.texi b/manual/tunables.texi
index ec18b10834..61edd62425 100644
--- a/manual/tunables.texi
+++ b/manual/tunables.texi
@@ -396,6 +396,20 @@ to set threshold in bytes for non temporal store.
 This tunable is specific to i386 and x86-64.
 @end deftp
 
+@deftp Tunable glibc.cpu.x86_rep_movsb_threshold
+The @code{glibc.cpu.x86_rep_movsb_threshold} tunable allows the user
+to set threshold in bytes to start using "rep movsb".
+
+This tunable is specific to i386 and x86-64.
+@end deftp
+
+@deftp Tunable glibc.cpu.x86_rep_stosb_threshold
+The @code{glibc.cpu.x86_rep_stosb_threshold} tunable allows the user
+to set threshold in bytes to start using "rep stosb".
+
+This tunable is specific to i386 and x86-64.
+@end deftp
+
 @deftp Tunable glibc.cpu.x86_ibt
 The @code{glibc.cpu.x86_ibt} tunable allows the user to control how
 indirect branch tracking (IBT) should be enabled.  Accepted values are
diff --git a/sysdeps/x86/cacheinfo.c b/sysdeps/x86/cacheinfo.c
index 311502dee3..4322328a1b 100644
--- a/sysdeps/x86/cacheinfo.c
+++ b/sysdeps/x86/cacheinfo.c
@@ -530,6 +530,23 @@ long int __x86_raw_shared_cache_size attribute_hidden = 1024 * 1024;
 /* Threshold to use non temporal store.  */
 long int __x86_shared_non_temporal_threshold attribute_hidden;
 
+/* Threshold to use Enhanced REP MOVSB.  Since there is overhead to set
+   up REP MOVSB operation, REP MOVSB isn't faster on short data.  The
+   memcpy micro benchmark in glibc shows that 2KB is the approximate
+   value above which REP MOVSB becomes faster than SSE2 optimization
+   on processors with Enhanced REP MOVSB.  Since larger register size
+   can move more data with a single load and store, the threshold is
+   higher with larger register size.  */
+long int __x86_rep_movsb_threshold attribute_hidden = 2048;
+
+/* Threshold to use Enhanced REP STOSB.  Since there is overhead to set
+   up REP STOSB operation, REP STOSB isn't faster on short data.  The
+   memset micro benchmark in glibc shows that 2KB is the approximate
+   value above which REP STOSB becomes faster on processors with
+   Enhanced REP STOSB.  Since the stored value is fixed, larger register
+   size has minimal impact on threshold.  */
+long int __x86_rep_stosb_threshold attribute_hidden = 2048;
+
 #ifndef DISABLE_PREFETCHW
 /* PREFETCHW support flag for use in memory and string routines.  */
 int __x86_prefetchw attribute_hidden;
@@ -872,6 +889,35 @@ init_cacheinfo (void)
     = (cpu_features->non_temporal_threshold != 0
        ? cpu_features->non_temporal_threshold
        : __x86_shared_cache_size * threads * 3 / 4);
+
+  /* NB: The REP MOVSB threshold must be greater than VEC_SIZE * 8.  */
+  unsigned int minimum_rep_movsb_threshold;
+  /* NB: The default REP MOVSB threshold is 2048 * (VEC_SIZE / 16).  */
+  unsigned int rep_movsb_threshold;
+  if (CPU_FEATURES_ARCH_P (cpu_features, AVX512F_Usable)
+      && !CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_AVX512))
+    {
+      rep_movsb_threshold = 2048 * (64 / 16);
+      minimum_rep_movsb_threshold = 64 * 8;
+    }
+  else if (CPU_FEATURES_ARCH_P (cpu_features,
+ AVX_Fast_Unaligned_Load))
+    {
+      rep_movsb_threshold = 2048 * (32 / 16);
+      minimum_rep_movsb_threshold = 32 * 8;
+    }
+  else
+    {
+      rep_movsb_threshold = 2048 * (16 / 16);
+      minimum_rep_movsb_threshold = 16 * 8;
+    }
+  if (cpu_features->rep_movsb_threshold > minimum_rep_movsb_threshold)
+    __x86_rep_movsb_threshold = cpu_features->rep_movsb_threshold;
+  else
+    __x86_rep_movsb_threshold = rep_movsb_threshold;
+
+  if (cpu_features->rep_stosb_threshold)
+    __x86_rep_stosb_threshold = cpu_features->rep_stosb_threshold;
 }
 
 #endif
diff --git a/sysdeps/x86/cpu-features.c b/sysdeps/x86/cpu-features.c
index c351bdd54a..c7673a2eb9 100644
--- a/sysdeps/x86/cpu-features.c
+++ b/sysdeps/x86/cpu-features.c
@@ -606,6 +606,10 @@ no_cpuid:
   TUNABLE_GET (hwcaps, tunable_val_t *, TUNABLE_CALLBACK (set_hwcaps));
   cpu_features->non_temporal_threshold
     = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
+  cpu_features->rep_movsb_threshold
+    = TUNABLE_GET (x86_rep_movsb_threshold, long int, NULL);
+  cpu_features->rep_stosb_threshold
+    = TUNABLE_GET (x86_rep_stosb_threshold, long int, NULL);
   cpu_features->data_cache_size
     = TUNABLE_GET (x86_data_cache_size, long int, NULL);
   cpu_features->shared_cache_size
diff --git a/sysdeps/x86/cpu-features.h b/sysdeps/x86/cpu-features.h
index d66dc206f7..39d2b59d63 100644
--- a/sysdeps/x86/cpu-features.h
+++ b/sysdeps/x86/cpu-features.h
@@ -102,6 +102,10 @@ struct cpu_features
   unsigned long int shared_cache_size;
   /* Threshold to use non temporal store.  */
   unsigned long int non_temporal_threshold;
+  /* Threshold to use "rep movsb".  */
+  unsigned long int rep_movsb_threshold;
+  /* Threshold to use "rep stosb".  */
+  unsigned long int rep_stosb_threshold;
 };
 
 /* Used from outside of glibc to get access to the CPU features
diff --git a/sysdeps/x86/dl-tunables.list b/sysdeps/x86/dl-tunables.list
index 251b926ce4..43bf6c2389 100644
--- a/sysdeps/x86/dl-tunables.list
+++ b/sysdeps/x86/dl-tunables.list
@@ -30,6 +30,12 @@ glibc {
     x86_non_temporal_threshold {
       type: SIZE_T
     }
+    x86_rep_movsb_threshold {
+      type: SIZE_T
+    }
+    x86_rep_stosb_threshold {
+      type: SIZE_T
+    }
     x86_data_cache_size {
       type: SIZE_T
     }
diff --git a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
index 74953245aa..bd5dc1a3f3 100644
--- a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
+++ b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
@@ -56,17 +56,6 @@
 # endif
 #endif
 
-/* Threshold to use Enhanced REP MOVSB.  Since there is overhead to set
-   up REP MOVSB operation, REP MOVSB isn't faster on short data.  The
-   memcpy micro benchmark in glibc shows that 2KB is the approximate
-   value above which REP MOVSB becomes faster than SSE2 optimization
-   on processors with Enhanced REP MOVSB.  Since larger register size
-   can move more data with a single load and store, the threshold is
-   higher with larger register size.  */
-#ifndef REP_MOVSB_THRESHOLD
-# define REP_MOVSB_THRESHOLD (2048 * (VEC_SIZE / 16))
-#endif
-
 #ifndef PREFETCH
 # define PREFETCH(addr) prefetcht0 addr
 #endif
@@ -253,9 +242,6 @@ L(movsb):
  leaq (%rsi,%rdx), %r9
  cmpq %r9, %rdi
  /* Avoid slow backward REP MOVSB.  */
-# if REP_MOVSB_THRESHOLD <= (VEC_SIZE * 8)
-#  error Unsupported REP_MOVSB_THRESHOLD and VEC_SIZE!
-# endif
  jb L(more_8x_vec_backward)
 1:
  mov %RDX_LP, %RCX_LP
@@ -331,7 +317,7 @@ L(between_2_3):
 
 #if defined USE_MULTIARCH && IS_IN (libc)
 L(movsb_more_2x_vec):
- cmpq $REP_MOVSB_THRESHOLD, %rdx
+ cmp __x86_rep_movsb_threshold(%rip), %RDX_LP
  ja L(movsb)
 #endif
 L(more_2x_vec):
diff --git a/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S b/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
index af2299709c..2bfc95de05 100644
--- a/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
+++ b/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
@@ -58,16 +58,6 @@
 # endif
 #endif
 
-/* Threshold to use Enhanced REP STOSB.  Since there is overhead to set
-   up REP STOSB operation, REP STOSB isn't faster on short data.  The
-   memset micro benchmark in glibc shows that 2KB is the approximate
-   value above which REP STOSB becomes faster on processors with
-   Enhanced REP STOSB.  Since the stored value is fixed, larger register
-   size has minimal impact on threshold.  */
-#ifndef REP_STOSB_THRESHOLD
-# define REP_STOSB_THRESHOLD 2048
-#endif
-
 #ifndef SECTION
 # error SECTION is not defined!
 #endif
@@ -181,7 +171,7 @@ ENTRY (MEMSET_SYMBOL (__memset, unaligned_erms))
  ret
 
 L(stosb_more_2x_vec):
- cmpq $REP_STOSB_THRESHOLD, %rdx
+ cmp __x86_rep_stosb_threshold(%rip), %RDX_LP
  ja L(stosb)
 #endif
 L(more_2x_vec):
--
2.26.2

Reply | Threaded
Open this post in threaded view
|

Re: V2 [PATCH 1/2] Update tunable min/max values

Sourceware - libc-alpha mailing list
In reply to this post by Sourceware - libc-alpha mailing list
On 7/3/20 7:09 PM, H.J. Lu wrote:

>> Why not raise the dl-tunables.list minimum to 1?
>> Then the above code just becomes what we used to have.
>
> What we used to have are
>
>   cpu_features->non_temporal_threshold
>     = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
> ...
>   __x86_shared_non_temporal_threshold
>     = (cpu_features->non_temporal_threshold != 0
>        ? cpu_features->non_temporal_threshold
>        : __x86_shared_cache_size * threads * 3 / 4);
>
> I just changed the order.
This is a bug in sysdeps/x86/dl-tunables.list where we have failed
to set the minimum to 1. Please fix dl-tunables.list instead of
rewriting the code which is incorrect?

--
Cheers,
Carlos.

Reply | Threaded
Open this post in threaded view
|

Re: V2 [PATCH] x86: Add thresholds for "rep movsb/stosb" to tunables

Sourceware - libc-alpha mailing list
In reply to this post by Sourceware - libc-alpha mailing list
On 7/4/20 8:03 AM, H.J. Lu wrote:

> On Fri, Jul 03, 2020 at 03:49:21PM -0400, Carlos O'Donell wrote:
>> On 7/3/20 1:52 PM, H.J. Lu wrote:
>>> Add x86_rep_movsb_threshold and x86_rep_stosb_threshold to tunables
>>> to update thresholds for "rep movsb" and "rep stosb" at run-time.
>>>
>>> Note that the user specified threshold for "rep movsb" smaller than the
>>> minimum threshold will be ignored.
>>
>> Post v2 please. Almost there.
>>
>>> ---
>>>  manual/tunables.texi                          | 14 +++++++
>>>  sysdeps/x86/cacheinfo.c                       | 20 ++++++++++
>>>  sysdeps/x86/cpu-features.h                    |  4 ++
>>>  sysdeps/x86/dl-cacheinfo.c                    | 38 +++++++++++++++++++
>>>  sysdeps/x86/dl-tunables.list                  |  6 +++
>>>  .../multiarch/memmove-vec-unaligned-erms.S    | 16 +-------
>>>  .../multiarch/memset-vec-unaligned-erms.S     | 12 +-----
>>>  7 files changed, 84 insertions(+), 26 deletions(-)
>>>
>>> diff --git a/manual/tunables.texi b/manual/tunables.texi
>>> index ec18b10834..61edd62425 100644
>>> --- a/manual/tunables.texi
>>> +++ b/manual/tunables.texi
>>> @@ -396,6 +396,20 @@ to set threshold in bytes for non temporal store.
>>>  This tunable is specific to i386 and x86-64.
>>>  @end deftp
>>>  
>>> +@deftp Tunable glibc.cpu.x86_rep_movsb_threshold
>>> +The @code{glibc.cpu.x86_rep_movsb_threshold} tunable allows the user
>>> +to set threshold in bytes to start using "rep movsb".
>>> +
>>> +This tunable is specific to i386 and x86-64.
>>> +@end deftp
>>> +
>>> +@deftp Tunable glibc.cpu.x86_rep_stosb_threshold
>>> +The @code{glibc.cpu.x86_rep_stosb_threshold} tunable allows the user
>>> +to set threshold in bytes to start using "rep stosb".
>>> +
>>> +This tunable is specific to i386 and x86-64.
>>> +@end deftp
>>> +
>>>  @deftp Tunable glibc.cpu.x86_ibt
>>>  The @code{glibc.cpu.x86_ibt} tunable allows the user to control how
>>>  indirect branch tracking (IBT) should be enabled.  Accepted values are
>>> diff --git a/sysdeps/x86/cacheinfo.c b/sysdeps/x86/cacheinfo.c
>>> index 8c4c7f9972..bb536d96ef 100644
>>> --- a/sysdeps/x86/cacheinfo.c
>>> +++ b/sysdeps/x86/cacheinfo.c
>>> @@ -41,6 +41,23 @@ long int __x86_raw_shared_cache_size attribute_hidden = 1024 * 1024;
>>>  /* Threshold to use non temporal store.  */
>>>  long int __x86_shared_non_temporal_threshold attribute_hidden;
>>>  
>>> +/* Threshold to use Enhanced REP MOVSB.  Since there is overhead to set
>>> +   up REP MOVSB operation, REP MOVSB isn't faster on short data.  The
>>> +   memcpy micro benchmark in glibc shows that 2KB is the approximate
>>> +   value above which REP MOVSB becomes faster than SSE2 optimization
>>> +   on processors with Enhanced REP MOVSB.  Since larger register size
>>> +   can move more data with a single load and store, the threshold is
>>> +   higher with larger register size.  */
>>> +long int __x86_rep_movsb_threshold attribute_hidden = 2048;
>>> +
>>> +/* Threshold to use Enhanced REP STOSB.  Since there is overhead to set
>>> +   up REP STOSB operation, REP STOSB isn't faster on short data.  The
>>> +   memset micro benchmark in glibc shows that 2KB is the approximate
>>> +   value above which REP STOSB becomes faster on processors with
>>> +   Enhanced REP STOSB.  Since the stored value is fixed, larger register
>>> +   size has minimal impact on threshold.  */
>>> +long int __x86_rep_stosb_threshold attribute_hidden = 2048;
>>> +
>>>  #ifndef __x86_64__
>>>  /* PREFETCHW support flag for use in memory and string routines.  */
>>>  int __x86_prefetchw attribute_hidden;
>>> @@ -117,6 +134,9 @@ init_cacheinfo (void)
>>>    __x86_shared_non_temporal_threshold
>>>      = cpu_features->non_temporal_threshold;
>>>  
>>> +  __x86_rep_movsb_threshold = cpu_features->rep_movsb_threshold;
>>> +  __x86_rep_stosb_threshold = cpu_features->rep_stosb_threshold;
>>> +
>>
>> OK. Update global from cpu_features with values.
>>
>> I would really like to see some kind of "assert (cpu_features->initialized);"
>> that way we know we didn't break the startup sequence unintentionally.
>>
>>>  #ifndef __x86_64__
>>>    __x86_prefetchw = cpu_features->prefetchw;
>>>  #endif
>>> diff --git a/sysdeps/x86/cpu-features.h b/sysdeps/x86/cpu-features.h
>>> index 3aaed33cbc..002e12e11f 100644
>>> --- a/sysdeps/x86/cpu-features.h
>>> +++ b/sysdeps/x86/cpu-features.h
>>> @@ -128,6 +128,10 @@ struct cpu_features
>>>    /* PREFETCHW support flag for use in memory and string routines.  */
>>>    unsigned long int prefetchw;
>>>  #endif
>>> +  /* Threshold to use "rep movsb".  */
>>> +  unsigned long int rep_movsb_threshold;
>>> +  /* Threshold to use "rep stosb".  */
>>> +  unsigned long int rep_stosb_threshold;
>>
>> OK.
>>
>>>  };
>>>  
>>>  /* Used from outside of glibc to get access to the CPU features
>>> diff --git a/sysdeps/x86/dl-cacheinfo.c b/sysdeps/x86/dl-cacheinfo.c
>>> index 8e2a6f552c..aff9bd1067 100644
>>> --- a/sysdeps/x86/dl-cacheinfo.c
>>> +++ b/sysdeps/x86/dl-cacheinfo.c
>>> @@ -860,6 +860,31 @@ __init_cacheinfo (void)
>>>       total shared cache size.  */
>>>    unsigned long int non_temporal_threshold = (shared * threads * 3 / 4);
>>>  
>>> +  /* NB: The REP MOVSB threshold must be greater than VEC_SIZE * 8.  */
>>> +  unsigned long int minimum_rep_movsb_threshold;
>>> +  /* NB: The default REP MOVSB threshold is 2048 * (VEC_SIZE / 16).  See
>>> +     comments for __x86_rep_movsb_threshold in cacheinfo.c.  */
>>> +  unsigned long int rep_movsb_threshold;
>>> +  if (CPU_FEATURES_ARCH_P (cpu_features, AVX512F_Usable)
>>> +      && !CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_AVX512))
>>> +    {
>>> +      rep_movsb_threshold = 2048 * (64 / 16);
>>> +      minimum_rep_movsb_threshold = 64 * 8;
>>> +    }
>>> +  else if (CPU_FEATURES_ARCH_P (cpu_features,
>>> + AVX_Fast_Unaligned_Load))
>>> +    {
>>> +      rep_movsb_threshold = 2048 * (32 / 16);
>>> +      minimum_rep_movsb_threshold = 32 * 8;
>>> +    }
>>> +  else
>>> +    {
>>> +      rep_movsb_threshold = 2048 * (16 / 16);
>>> +      minimum_rep_movsb_threshold = 16 * 8;
>>> +    }
>>> +  /* NB: See comments for __x86_rep_stosb_threshold in cacheinfo.c.  */
>>> +  unsigned long int rep_stosb_threshold = 2048;
>>> +
>>>  #if HAVE_TUNABLES
>>>    long int tunable_size;
>>>    tunable_size = TUNABLE_GET (x86_data_cache_size, long int, NULL);
>>> @@ -871,11 +896,19 @@ __init_cacheinfo (void)
>>>    tunable_size = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
>>>    if (tunable_size != 0)
>>>      non_temporal_threshold = tunable_size;
>>
>>> +  tunable_size = TUNABLE_GET (x86_rep_movsb_threshold, long int, NULL);
>>> +  if (tunable_size > minimum_rep_movsb_threshold)
>>> +    rep_movsb_threshold = tunable_size;
>>
>> OK. Good, we only set rep_movsb_threshold if it's greater than min.
>>
>>> +  tunable_size = TUNABLE_GET (x86_rep_stosb_threshold, long int, NULL);
>>> +  if (tunable_size != 0)
>>> +    rep_stosb_threshold = tunable_size;
>>
>> This should be min=1, default=2048 in dl-tunables.list, and would remove
>> this code since the range is not dynamic.
>>
>> The point of the tunables framework is to remove such boiler plate for
>> range a default processing and clearing parameters for security settings.
>>
>>>  #endif
>>>  
>>>    cpu_features->data_cache_size = data;
>>>    cpu_features->shared_cache_size = shared;
>>>    cpu_features->non_temporal_threshold = non_temporal_threshold;
>>> +  cpu_features->rep_movsb_threshold = rep_movsb_threshold;
>>> +  cpu_features->rep_stosb_threshold = rep_stosb_threshold;
>>>  
>>>  #if HAVE_TUNABLES
>>>    TUNABLE_UPDATE (x86_data_cache_size, long int,
>>> @@ -884,5 +917,10 @@ __init_cacheinfo (void)
>>>    shared, 0, (long int) -1);
>>>    TUNABLE_UPDATE (x86_non_temporal_threshold, long int,
>>>    non_temporal_threshold, 0, (long int) -1);
>>> +  TUNABLE_UPDATE (x86_rep_movsb_threshold, long int,
>>> +  rep_movsb_threshold, minimum_rep_movsb_threshold,
>>> +  (long int) -1);
>>
>> OK. Store the new value and the computed minimum.
>>
>>> +  TUNABLE_UPDATE (x86_rep_stosb_threshold, long int,
>>> +  rep_stosb_threshold, 0, (long int) -1);
>>
>> This one can be deleted.
>>
>
> We should go with this simple one for 2.32.
 
I agree.

We can make this better in 2.33.

Please post V3 of this patch for final review.

See my notes below on how this is intended to be handled by tunables framework.
 

> H.J.
> ---
> Add x86_rep_movsb_threshold and x86_rep_stosb_threshold to tunables
> to update thresholds for "rep movsb" and "rep stosb" at run-time.
>
> Note that the user specified threshold for "rep movsb" smaller than
> the minimum threshold will be ignored.
> ---
>  manual/tunables.texi                          | 14 ++++++
>  sysdeps/x86/cacheinfo.c                       | 46 +++++++++++++++++++
>  sysdeps/x86/cpu-features.c                    |  4 ++
>  sysdeps/x86/cpu-features.h                    |  4 ++
>  sysdeps/x86/dl-tunables.list                  |  6 +++
>  .../multiarch/memmove-vec-unaligned-erms.S    | 16 +------
>  .../multiarch/memset-vec-unaligned-erms.S     | 12 +----
>  7 files changed, 76 insertions(+), 26 deletions(-)
>
> diff --git a/manual/tunables.texi b/manual/tunables.texi
> index ec18b10834..61edd62425 100644
> --- a/manual/tunables.texi
> +++ b/manual/tunables.texi
> @@ -396,6 +396,20 @@ to set threshold in bytes for non temporal store.
>  This tunable is specific to i386 and x86-64.
>  @end deftp
>  
> +@deftp Tunable glibc.cpu.x86_rep_movsb_threshold
> +The @code{glibc.cpu.x86_rep_movsb_threshold} tunable allows the user
> +to set threshold in bytes to start using "rep movsb".

Add: "The value must be greater than zero, and currently defaults to 2048 bytes."

> +
> +This tunable is specific to i386 and x86-64.
> +@end deftp
> +
> +@deftp Tunable glibc.cpu.x86_rep_stosb_threshold
> +The @code{glibc.cpu.x86_rep_stosb_threshold} tunable allows the user
> +to set threshold in bytes to start using "rep stosb".

Add: "The value must be greater than zero, and currently defaults to 2048 bytes."

> +
> +This tunable is specific to i386 and x86-64.
> +@end deftp

OK. Docs addition required and added here.

> +
>  @deftp Tunable glibc.cpu.x86_ibt
>  The @code{glibc.cpu.x86_ibt} tunable allows the user to control how
>  indirect branch tracking (IBT) should be enabled.  Accepted values are
> diff --git a/sysdeps/x86/cacheinfo.c b/sysdeps/x86/cacheinfo.c
> index 311502dee3..4322328a1b 100644
> --- a/sysdeps/x86/cacheinfo.c
> +++ b/sysdeps/x86/cacheinfo.c
> @@ -530,6 +530,23 @@ long int __x86_raw_shared_cache_size attribute_hidden = 1024 * 1024;
>  /* Threshold to use non temporal store.  */
>  long int __x86_shared_non_temporal_threshold attribute_hidden;
>  
> +/* Threshold to use Enhanced REP MOVSB.  Since there is overhead to set
> +   up REP MOVSB operation, REP MOVSB isn't faster on short data.  The
> +   memcpy micro benchmark in glibc shows that 2KB is the approximate
> +   value above which REP MOVSB becomes faster than SSE2 optimization
> +   on processors with Enhanced REP MOVSB.  Since larger register size
> +   can move more data with a single load and store, the threshold is
> +   higher with larger register size.  */

Comments should move to dl-tunables.list.

> +long int __x86_rep_movsb_threshold attribute_hidden = 2048;
> +
> +/* Threshold to use Enhanced REP STOSB.  Since there is overhead to set
> +   up REP STOSB operation, REP STOSB isn't faster on short data.  The
> +   memset micro benchmark in glibc shows that 2KB is the approximate
> +   value above which REP STOSB becomes faster on processors with
> +   Enhanced REP STOSB.  Since the stored value is fixed, larger register
> +   size has minimal impact on threshold.  */

Comments should move to dl-tunables.list.

> +long int __x86_rep_stosb_threshold attribute_hidden = 2048;

Are these globals used *before* tunables is initialized?

That should be the only case that causes us to require to set the default here.

Otherwise we should remove the default and express it in dl-tunables.list.

> +
>  #ifndef DISABLE_PREFETCHW
>  /* PREFETCHW support flag for use in memory and string routines.  */
>  int __x86_prefetchw attribute_hidden;
> @@ -872,6 +889,35 @@ init_cacheinfo (void)
>      = (cpu_features->non_temporal_threshold != 0
>         ? cpu_features->non_temporal_threshold
>         : __x86_shared_cache_size * threads * 3 / 4);
> +
> +  /* NB: The REP MOVSB threshold must be greater than VEC_SIZE * 8.  */
> +  unsigned int minimum_rep_movsb_threshold;
> +  /* NB: The default REP MOVSB threshold is 2048 * (VEC_SIZE / 16).  */
> +  unsigned int rep_movsb_threshold;
> +  if (CPU_FEATURES_ARCH_P (cpu_features, AVX512F_Usable)
> +      && !CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_AVX512))
> +    {
> +      rep_movsb_threshold = 2048 * (64 / 16);
> +      minimum_rep_movsb_threshold = 64 * 8;
> +    }
> +  else if (CPU_FEATURES_ARCH_P (cpu_features,
> + AVX_Fast_Unaligned_Load))
> +    {
> +      rep_movsb_threshold = 2048 * (32 / 16);
> +      minimum_rep_movsb_threshold = 32 * 8;
> +    }
> +  else
> +    {
> +      rep_movsb_threshold = 2048 * (16 / 16);
> +      minimum_rep_movsb_threshold = 16 * 8;
> +    }
> +  if (cpu_features->rep_movsb_threshold > minimum_rep_movsb_threshold)
> +    __x86_rep_movsb_threshold = cpu_features->rep_movsb_threshold;
> +  else
> +    __x86_rep_movsb_threshold = rep_movsb_threshold;

OK, use the cpu_features value otherwise the computed value threshold.

> +

OK.

> +  if (cpu_features->rep_stosb_threshold)
> +    __x86_rep_stosb_threshold = cpu_features->rep_stosb_threshold;

This code becomes:

__x86_rep_stosb_threshold = cpu_features->rep_stosb_threshold;

Because the tunables code ensured the thresold was > 0, otherwise
an invalid tunable would leave the default value of 2048.

>  }
>  
>  #endif
> diff --git a/sysdeps/x86/cpu-features.c b/sysdeps/x86/cpu-features.c
> index c351bdd54a..c7673a2eb9 100644
> --- a/sysdeps/x86/cpu-features.c
> +++ b/sysdeps/x86/cpu-features.c
> @@ -606,6 +606,10 @@ no_cpuid:
>    TUNABLE_GET (hwcaps, tunable_val_t *, TUNABLE_CALLBACK (set_hwcaps));
>    cpu_features->non_temporal_threshold
>      = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
> +  cpu_features->rep_movsb_threshold
> +    = TUNABLE_GET (x86_rep_movsb_threshold, long int, NULL);
> +  cpu_features->rep_stosb_threshold
> +    = TUNABLE_GET (x86_rep_stosb_threshold, long int, NULL);

OK.

>    cpu_features->data_cache_size
>      = TUNABLE_GET (x86_data_cache_size, long int, NULL);
>    cpu_features->shared_cache_size
> diff --git a/sysdeps/x86/cpu-features.h b/sysdeps/x86/cpu-features.h
> index d66dc206f7..39d2b59d63 100644
> --- a/sysdeps/x86/cpu-features.h
> +++ b/sysdeps/x86/cpu-features.h
> @@ -102,6 +102,10 @@ struct cpu_features
>    unsigned long int shared_cache_size;
>    /* Threshold to use non temporal store.  */
>    unsigned long int non_temporal_threshold;
> +  /* Threshold to use "rep movsb".  */
> +  unsigned long int rep_movsb_threshold;
> +  /* Threshold to use "rep stosb".  */
> +  unsigned long int rep_stosb_threshold;

OK.

>  };
>  
>  /* Used from outside of glibc to get access to the CPU features
> diff --git a/sysdeps/x86/dl-tunables.list b/sysdeps/x86/dl-tunables.list
> index 251b926ce4..43bf6c2389 100644
> --- a/sysdeps/x86/dl-tunables.list
> +++ b/sysdeps/x86/dl-tunables.list
> @@ -30,6 +30,12 @@ glibc {
>      x86_non_temporal_threshold {
>        type: SIZE_T
>      }
> +    x86_rep_movsb_threshold {
> +      type: SIZE_T

Add "minimum: 1"
Add "default: 2048"
Add comments about why it's 2048.

> +    }
> +    x86_rep_stosb_threshold {
> +      type: SIZE_T

Add "minimum: 1"
Add "default: 2048"
Add comments about why it's 2048.

> +    }
>      x86_data_cache_size {
>        type: SIZE_T
>      }
> diff --git a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
> index 74953245aa..bd5dc1a3f3 100644
> --- a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
> +++ b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
> @@ -56,17 +56,6 @@
>  # endif
>  #endif
>  
> -/* Threshold to use Enhanced REP MOVSB.  Since there is overhead to set
> -   up REP MOVSB operation, REP MOVSB isn't faster on short data.  The
> -   memcpy micro benchmark in glibc shows that 2KB is the approximate
> -   value above which REP MOVSB becomes faster than SSE2 optimization
> -   on processors with Enhanced REP MOVSB.  Since larger register size
> -   can move more data with a single load and store, the threshold is
> -   higher with larger register size.  */
> -#ifndef REP_MOVSB_THRESHOLD
> -# define REP_MOVSB_THRESHOLD (2048 * (VEC_SIZE / 16))
> -#endif

OK.

> -
>  #ifndef PREFETCH
>  # define PREFETCH(addr) prefetcht0 addr
>  #endif
> @@ -253,9 +242,6 @@ L(movsb):
>   leaq (%rsi,%rdx), %r9
>   cmpq %r9, %rdi
>   /* Avoid slow backward REP MOVSB.  */
> -# if REP_MOVSB_THRESHOLD <= (VEC_SIZE * 8)
> -#  error Unsupported REP_MOVSB_THRESHOLD and VEC_SIZE!
> -# endif

OK.

>   jb L(more_8x_vec_backward)
>  1:
>   mov %RDX_LP, %RCX_LP
> @@ -331,7 +317,7 @@ L(between_2_3):
>  
>  #if defined USE_MULTIARCH && IS_IN (libc)
>  L(movsb_more_2x_vec):
> - cmpq $REP_MOVSB_THRESHOLD, %rdx
> + cmp __x86_rep_movsb_threshold(%rip), %RDX_LP

OK.

>   ja L(movsb)
>  #endif
>  L(more_2x_vec):
> diff --git a/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S b/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
> index af2299709c..2bfc95de05 100644
> --- a/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
> +++ b/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
> @@ -58,16 +58,6 @@
>  # endif
>  #endif
>  
> -/* Threshold to use Enhanced REP STOSB.  Since there is overhead to set
> -   up REP STOSB operation, REP STOSB isn't faster on short data.  The
> -   memset micro benchmark in glibc shows that 2KB is the approximate
> -   value above which REP STOSB becomes faster on processors with
> -   Enhanced REP STOSB.  Since the stored value is fixed, larger register
> -   size has minimal impact on threshold.  */
> -#ifndef REP_STOSB_THRESHOLD
> -# define REP_STOSB_THRESHOLD 2048
> -#endif

OK.

> -
>  #ifndef SECTION
>  # error SECTION is not defined!
>  #endif
> @@ -181,7 +171,7 @@ ENTRY (MEMSET_SYMBOL (__memset, unaligned_erms))
>   ret
>  
>  L(stosb_more_2x_vec):
> - cmpq $REP_STOSB_THRESHOLD, %rdx
> + cmp __x86_rep_stosb_threshold(%rip), %RDX_LP

OK.

>   ja L(stosb)
>  #endif
>  L(more_2x_vec):
>


--
Cheers,
Carlos.

Reply | Threaded
Open this post in threaded view
|

Re: V2 [PATCH 1/2] Update tunable min/max values

Sourceware - libc-alpha mailing list
In reply to this post by Sourceware - libc-alpha mailing list
On Mon, Jul 6, 2020 at 5:47 AM Carlos O'Donell <[hidden email]> wrote:

>
> On 7/3/20 7:09 PM, H.J. Lu wrote:
> >> Why not raise the dl-tunables.list minimum to 1?
> >> Then the above code just becomes what we used to have.
> >
> > What we used to have are
> >
> >   cpu_features->non_temporal_threshold
> >     = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
> > ...
> >   __x86_shared_non_temporal_threshold
> >     = (cpu_features->non_temporal_threshold != 0
> >        ? cpu_features->non_temporal_threshold
> >        : __x86_shared_cache_size * threads * 3 / 4);
> >
> > I just changed the order.
> This is a bug in sysdeps/x86/dl-tunables.list where we have failed
> to set the minimum to 1. Please fix dl-tunables.list instead of
> rewriting the code which is incorrect?

I will update this patch after 2.32 is branched since it isn't needed for
2.32.


--
H.J.
Reply | Threaded
Open this post in threaded view
|

Re: V2 [PATCH 1/2] Update tunable min/max values

Sourceware - libc-alpha mailing list
On 7/6/20 9:13 AM, H.J. Lu wrote:

> On Mon, Jul 6, 2020 at 5:47 AM Carlos O'Donell <[hidden email]> wrote:
>>
>> On 7/3/20 7:09 PM, H.J. Lu wrote:
>>>> Why not raise the dl-tunables.list minimum to 1?
>>>> Then the above code just becomes what we used to have.
>>>
>>> What we used to have are
>>>
>>>   cpu_features->non_temporal_threshold
>>>     = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
>>> ...
>>>   __x86_shared_non_temporal_threshold
>>>     = (cpu_features->non_temporal_threshold != 0
>>>        ? cpu_features->non_temporal_threshold
>>>        : __x86_shared_cache_size * threads * 3 / 4);
>>>
>>> I just changed the order.
>> This is a bug in sysdeps/x86/dl-tunables.list where we have failed
>> to set the minimum to 1. Please fix dl-tunables.list instead of
>> rewriting the code which is incorrect?
>
> I will update this patch after 2.32 is branched since it isn't needed for
> 2.32.

That's fine with me, and this patch is easily backported.

--
Cheers,
Carlos.

Reply | Threaded
Open this post in threaded view
|

Re: V2 [PATCH 1/2] Update tunable min/max values

Sourceware - libc-alpha mailing list
On Mon, Jul 6, 2020 at 6:15 AM Carlos O'Donell <[hidden email]> wrote:

>
> On 7/6/20 9:13 AM, H.J. Lu wrote:
> > On Mon, Jul 6, 2020 at 5:47 AM Carlos O'Donell <[hidden email]> wrote:
> >>
> >> On 7/3/20 7:09 PM, H.J. Lu wrote:
> >>>> Why not raise the dl-tunables.list minimum to 1?
> >>>> Then the above code just becomes what we used to have.
> >>>
> >>> What we used to have are
> >>>
> >>>   cpu_features->non_temporal_threshold
> >>>     = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
> >>> ...
> >>>   __x86_shared_non_temporal_threshold
> >>>     = (cpu_features->non_temporal_threshold != 0
> >>>        ? cpu_features->non_temporal_threshold
> >>>        : __x86_shared_cache_size * threads * 3 / 4);
> >>>
> >>> I just changed the order.
> >> This is a bug in sysdeps/x86/dl-tunables.list where we have failed
> >> to set the minimum to 1. Please fix dl-tunables.list instead of
> >> rewriting the code which is incorrect?
> >
> > I will update this patch after 2.32 is branched since it isn't needed for
> > 2.32.
>
> That's fine with me, and this patch is easily backported.
>

BTW, this patch exposed:

https://sourceware.org/bugzilla/show_bug.cgi?id=26203

which is fixed by

https://sourceware.org/pipermail/libc-alpha/2020-July/115788.html

--
H.J.
Reply | Threaded
Open this post in threaded view
|

V3 [PATCH] x86: Add thresholds for "rep movsb/stosb" to tunables

Sourceware - libc-alpha mailing list
In reply to this post by Sourceware - libc-alpha mailing list
On Mon, Jul 06, 2020 at 08:59:31AM -0400, Carlos O'Donell wrote:

> > We should go with this simple one for 2.32.
>  
> I agree.
>
> We can make this better in 2.33.
>
> Please post V3 of this patch for final review.
>
> See my notes below on how this is intended to be handled by tunables framework.
>  
> > H.J.
> > ---
> > Add x86_rep_movsb_threshold and x86_rep_stosb_threshold to tunables
> > to update thresholds for "rep movsb" and "rep stosb" at run-time.
> >
> > Note that the user specified threshold for "rep movsb" smaller than
> > the minimum threshold will be ignored.
> > ---
> >  manual/tunables.texi                          | 14 ++++++
> >  sysdeps/x86/cacheinfo.c                       | 46 +++++++++++++++++++
> >  sysdeps/x86/cpu-features.c                    |  4 ++
> >  sysdeps/x86/cpu-features.h                    |  4 ++
> >  sysdeps/x86/dl-tunables.list                  |  6 +++
> >  .../multiarch/memmove-vec-unaligned-erms.S    | 16 +------
> >  .../multiarch/memset-vec-unaligned-erms.S     | 12 +----
> >  7 files changed, 76 insertions(+), 26 deletions(-)
> >
> > diff --git a/manual/tunables.texi b/manual/tunables.texi
> > index ec18b10834..61edd62425 100644
> > --- a/manual/tunables.texi
> > +++ b/manual/tunables.texi
> > @@ -396,6 +396,20 @@ to set threshold in bytes for non temporal store.
> >  This tunable is specific to i386 and x86-64.
> >  @end deftp
> >  
> > +@deftp Tunable glibc.cpu.x86_rep_movsb_threshold
> > +The @code{glibc.cpu.x86_rep_movsb_threshold} tunable allows the user
> > +to set threshold in bytes to start using "rep movsb".
>
> Add: "The value must be greater than zero, and currently defaults to 2048 bytes."

Fixed.

>
> > +
> > +This tunable is specific to i386 and x86-64.
> > +@end deftp
> > +
> > +@deftp Tunable glibc.cpu.x86_rep_stosb_threshold
> > +The @code{glibc.cpu.x86_rep_stosb_threshold} tunable allows the user
> > +to set threshold in bytes to start using "rep stosb".
>
> Add: "The value must be greater than zero, and currently defaults to 2048 bytes."
>

Fixed.

> > +
> > +This tunable is specific to i386 and x86-64.
> > +@end deftp
>
> OK. Docs addition required and added here.
>
> > +
> >  @deftp Tunable glibc.cpu.x86_ibt
> >  The @code{glibc.cpu.x86_ibt} tunable allows the user to control how
> >  indirect branch tracking (IBT) should be enabled.  Accepted values are
> > diff --git a/sysdeps/x86/cacheinfo.c b/sysdeps/x86/cacheinfo.c
> > index 311502dee3..4322328a1b 100644
> > --- a/sysdeps/x86/cacheinfo.c
> > +++ b/sysdeps/x86/cacheinfo.c
> > @@ -530,6 +530,23 @@ long int __x86_raw_shared_cache_size attribute_hidden = 1024 * 1024;
> >  /* Threshold to use non temporal store.  */
> >  long int __x86_shared_non_temporal_threshold attribute_hidden;
> >  
> > +/* Threshold to use Enhanced REP MOVSB.  Since there is overhead to set
> > +   up REP MOVSB operation, REP MOVSB isn't faster on short data.  The
> > +   memcpy micro benchmark in glibc shows that 2KB is the approximate
> > +   value above which REP MOVSB becomes faster than SSE2 optimization
> > +   on processors with Enhanced REP MOVSB.  Since larger register size
> > +   can move more data with a single load and store, the threshold is
> > +   higher with larger register size.  */
>
> Comments should move to dl-tunables.list.

Fixed.

>
> > +long int __x86_rep_movsb_threshold attribute_hidden = 2048;
> > +
> > +/* Threshold to use Enhanced REP STOSB.  Since there is overhead to set
> > +   up REP STOSB operation, REP STOSB isn't faster on short data.  The
> > +   memset micro benchmark in glibc shows that 2KB is the approximate
> > +   value above which REP STOSB becomes faster on processors with
> > +   Enhanced REP STOSB.  Since the stored value is fixed, larger register
> > +   size has minimal impact on threshold.  */
>
> Comments should move to dl-tunables.list.

Fixed.

>
> > +long int __x86_rep_stosb_threshold attribute_hidden = 2048;
>
> Are these globals used *before* tunables is initialized?

No.

>
> That should be the only case that causes us to require to set the default here.

But since tunables is optional, they can be used without tunables.

>
> Otherwise we should remove the default and express it in dl-tunables.list.

The default is needed when tunables is disabled.

>
> > +
> >  #ifndef DISABLE_PREFETCHW
> >  /* PREFETCHW support flag for use in memory and string routines.  */
> >  int __x86_prefetchw attribute_hidden;
> > @@ -872,6 +889,35 @@ init_cacheinfo (void)
> >      = (cpu_features->non_temporal_threshold != 0
> >         ? cpu_features->non_temporal_threshold
> >         : __x86_shared_cache_size * threads * 3 / 4);
> > +
> > +  /* NB: The REP MOVSB threshold must be greater than VEC_SIZE * 8.  */
> > +  unsigned int minimum_rep_movsb_threshold;
> > +  /* NB: The default REP MOVSB threshold is 2048 * (VEC_SIZE / 16).  */
> > +  unsigned int rep_movsb_threshold;
> > +  if (CPU_FEATURES_ARCH_P (cpu_features, AVX512F_Usable)
> > +      && !CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_AVX512))
> > +    {
> > +      rep_movsb_threshold = 2048 * (64 / 16);
> > +      minimum_rep_movsb_threshold = 64 * 8;
> > +    }
> > +  else if (CPU_FEATURES_ARCH_P (cpu_features,
> > + AVX_Fast_Unaligned_Load))
> > +    {
> > +      rep_movsb_threshold = 2048 * (32 / 16);
> > +      minimum_rep_movsb_threshold = 32 * 8;
> > +    }
> > +  else
> > +    {
> > +      rep_movsb_threshold = 2048 * (16 / 16);
> > +      minimum_rep_movsb_threshold = 16 * 8;
> > +    }
> > +  if (cpu_features->rep_movsb_threshold > minimum_rep_movsb_threshold)
> > +    __x86_rep_movsb_threshold = cpu_features->rep_movsb_threshold;
> > +  else
> > +    __x86_rep_movsb_threshold = rep_movsb_threshold;
>
> OK, use the cpu_features value otherwise the computed value threshold.
>
> > +
>
> OK.
>
> > +  if (cpu_features->rep_stosb_threshold)
> > +    __x86_rep_stosb_threshold = cpu_features->rep_stosb_threshold;
>
> This code becomes:
>
> __x86_rep_stosb_threshold = cpu_features->rep_stosb_threshold;
>
> Because the tunables code ensured the thresold was > 0, otherwise
> an invalid tunable would leave the default value of 2048.

Done.

>
> >  }
> >  
> >  #endif
> > diff --git a/sysdeps/x86/cpu-features.c b/sysdeps/x86/cpu-features.c
> > index c351bdd54a..c7673a2eb9 100644
> > --- a/sysdeps/x86/cpu-features.c
> > +++ b/sysdeps/x86/cpu-features.c
> > @@ -606,6 +606,10 @@ no_cpuid:
> >    TUNABLE_GET (hwcaps, tunable_val_t *, TUNABLE_CALLBACK (set_hwcaps));
> >    cpu_features->non_temporal_threshold
> >      = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
> > +  cpu_features->rep_movsb_threshold
> > +    = TUNABLE_GET (x86_rep_movsb_threshold, long int, NULL);
> > +  cpu_features->rep_stosb_threshold
> > +    = TUNABLE_GET (x86_rep_stosb_threshold, long int, NULL);
>
> OK.
>
> >    cpu_features->data_cache_size
> >      = TUNABLE_GET (x86_data_cache_size, long int, NULL);
> >    cpu_features->shared_cache_size
> > diff --git a/sysdeps/x86/cpu-features.h b/sysdeps/x86/cpu-features.h
> > index d66dc206f7..39d2b59d63 100644
> > --- a/sysdeps/x86/cpu-features.h
> > +++ b/sysdeps/x86/cpu-features.h
> > @@ -102,6 +102,10 @@ struct cpu_features
> >    unsigned long int shared_cache_size;
> >    /* Threshold to use non temporal store.  */
> >    unsigned long int non_temporal_threshold;
> > +  /* Threshold to use "rep movsb".  */
> > +  unsigned long int rep_movsb_threshold;
> > +  /* Threshold to use "rep stosb".  */
> > +  unsigned long int rep_stosb_threshold;
>
> OK.
>
> >  };
> >  
> >  /* Used from outside of glibc to get access to the CPU features
> > diff --git a/sysdeps/x86/dl-tunables.list b/sysdeps/x86/dl-tunables.list
> > index 251b926ce4..43bf6c2389 100644
> > --- a/sysdeps/x86/dl-tunables.list
> > +++ b/sysdeps/x86/dl-tunables.list
> > @@ -30,6 +30,12 @@ glibc {
> >      x86_non_temporal_threshold {
> >        type: SIZE_T
> >      }
> > +    x86_rep_movsb_threshold {
> > +      type: SIZE_T
>
> Add "minimum: 1"
> Add "default: 2048"
> Add comments about why it's 2048.

Done.

>
> > +    }
> > +    x86_rep_stosb_threshold {
> > +      type: SIZE_T
>
> Add "minimum: 1"
> Add "default: 2048"
> Add comments about why it's 2048.

Done.

>
> > +    }
> >      x86_data_cache_size {
> >        type: SIZE_T
> >      }
> > diff --git a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
> > index 74953245aa..bd5dc1a3f3 100644
> > --- a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
> > +++ b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
> > @@ -56,17 +56,6 @@
> >  # endif
> >  #endif
> >  
> > -/* Threshold to use Enhanced REP MOVSB.  Since there is overhead to set
> > -   up REP MOVSB operation, REP MOVSB isn't faster on short data.  The
> > -   memcpy micro benchmark in glibc shows that 2KB is the approximate
> > -   value above which REP MOVSB becomes faster than SSE2 optimization
> > -   on processors with Enhanced REP MOVSB.  Since larger register size
> > -   can move more data with a single load and store, the threshold is
> > -   higher with larger register size.  */
> > -#ifndef REP_MOVSB_THRESHOLD
> > -# define REP_MOVSB_THRESHOLD (2048 * (VEC_SIZE / 16))
> > -#endif
>
> OK.
>
> > -
> >  #ifndef PREFETCH
> >  # define PREFETCH(addr) prefetcht0 addr
> >  #endif
> > @@ -253,9 +242,6 @@ L(movsb):
> >   leaq (%rsi,%rdx), %r9
> >   cmpq %r9, %rdi
> >   /* Avoid slow backward REP MOVSB.  */
> > -# if REP_MOVSB_THRESHOLD <= (VEC_SIZE * 8)
> > -#  error Unsupported REP_MOVSB_THRESHOLD and VEC_SIZE!
> > -# endif
>
> OK.
>
> >   jb L(more_8x_vec_backward)
> >  1:
> >   mov %RDX_LP, %RCX_LP
> > @@ -331,7 +317,7 @@ L(between_2_3):
> >  
> >  #if defined USE_MULTIARCH && IS_IN (libc)
> >  L(movsb_more_2x_vec):
> > - cmpq $REP_MOVSB_THRESHOLD, %rdx
> > + cmp __x86_rep_movsb_threshold(%rip), %RDX_LP
>
> OK.
>
> >   ja L(movsb)
> >  #endif
> >  L(more_2x_vec):
> > diff --git a/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S b/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
> > index af2299709c..2bfc95de05 100644
> > --- a/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
> > +++ b/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
> > @@ -58,16 +58,6 @@
> >  # endif
> >  #endif
> >  
> > -/* Threshold to use Enhanced REP STOSB.  Since there is overhead to set
> > -   up REP STOSB operation, REP STOSB isn't faster on short data.  The
> > -   memset micro benchmark in glibc shows that 2KB is the approximate
> > -   value above which REP STOSB becomes faster on processors with
> > -   Enhanced REP STOSB.  Since the stored value is fixed, larger register
> > -   size has minimal impact on threshold.  */
> > -#ifndef REP_STOSB_THRESHOLD
> > -# define REP_STOSB_THRESHOLD 2048
> > -#endif
>
> OK.
>
> > -
> >  #ifndef SECTION
> >  # error SECTION is not defined!
> >  #endif
> > @@ -181,7 +171,7 @@ ENTRY (MEMSET_SYMBOL (__memset, unaligned_erms))
> >   ret
> >  
> >  L(stosb_more_2x_vec):
> > - cmpq $REP_STOSB_THRESHOLD, %rdx
> > + cmp __x86_rep_stosb_threshold(%rip), %RDX_LP
>
> OK.
>
> >   ja L(stosb)
> >  #endif
> >  L(more_2x_vec):
> >
>
>

Here is the updated patch.  OK for master?

Thanks.

H.J.
---
Add x86_rep_movsb_threshold and x86_rep_stosb_threshold to tunables
to update thresholds for "rep movsb" and "rep stosb" at run-time.

Note that the user specified threshold for "rep movsb" smaller than
the minimum threshold will be ignored.
---
 manual/tunables.texi                          | 16 +++++++++
 sysdeps/x86/cacheinfo.c                       | 36 +++++++++++++++++++
 sysdeps/x86/cpu-features.c                    |  4 +++
 sysdeps/x86/cpu-features.h                    |  4 +++
 sysdeps/x86/dl-tunables.list                  | 24 +++++++++++++
 .../multiarch/memmove-vec-unaligned-erms.S    | 16 +--------
 .../multiarch/memset-vec-unaligned-erms.S     | 12 +------
 7 files changed, 86 insertions(+), 26 deletions(-)

diff --git a/manual/tunables.texi b/manual/tunables.texi
index ec18b10834..4e68c7ff91 100644
--- a/manual/tunables.texi
+++ b/manual/tunables.texi
@@ -396,6 +396,22 @@ to set threshold in bytes for non temporal store.
 This tunable is specific to i386 and x86-64.
 @end deftp
 
+@deftp Tunable glibc.cpu.x86_rep_movsb_threshold
+The @code{glibc.cpu.x86_rep_movsb_threshold} tunable allows the user to
+set threshold in bytes to start using "rep movsb".  The value must be
+greater than zero, and currently defaults to 2048 bytes.
+
+This tunable is specific to i386 and x86-64.
+@end deftp
+
+@deftp Tunable glibc.cpu.x86_rep_stosb_threshold
+The @code{glibc.cpu.x86_rep_stosb_threshold} tunable allows the user to
+set threshold in bytes to start using "rep stosb".  The value must be
+greater than zero, and currently defaults to 2048 bytes.
+
+This tunable is specific to i386 and x86-64.
+@end deftp
+
 @deftp Tunable glibc.cpu.x86_ibt
 The @code{glibc.cpu.x86_ibt} tunable allows the user to control how
 indirect branch tracking (IBT) should be enabled.  Accepted values are
diff --git a/sysdeps/x86/cacheinfo.c b/sysdeps/x86/cacheinfo.c
index 311502dee3..136809a6d8 100644
--- a/sysdeps/x86/cacheinfo.c
+++ b/sysdeps/x86/cacheinfo.c
@@ -530,6 +530,12 @@ long int __x86_raw_shared_cache_size attribute_hidden = 1024 * 1024;
 /* Threshold to use non temporal store.  */
 long int __x86_shared_non_temporal_threshold attribute_hidden;
 
+/* Threshold to use Enhanced REP MOVSB.  */
+long int __x86_rep_movsb_threshold attribute_hidden = 2048;
+
+/* Threshold to use Enhanced REP STOSB.  */
+long int __x86_rep_stosb_threshold attribute_hidden = 2048;
+
 #ifndef DISABLE_PREFETCHW
 /* PREFETCHW support flag for use in memory and string routines.  */
 int __x86_prefetchw attribute_hidden;
@@ -872,6 +878,36 @@ init_cacheinfo (void)
     = (cpu_features->non_temporal_threshold != 0
        ? cpu_features->non_temporal_threshold
        : __x86_shared_cache_size * threads * 3 / 4);
+
+  /* NB: The REP MOVSB threshold must be greater than VEC_SIZE * 8.  */
+  unsigned int minimum_rep_movsb_threshold;
+  /* NB: The default REP MOVSB threshold is 2048 * (VEC_SIZE / 16).  */
+  unsigned int rep_movsb_threshold;
+  if (CPU_FEATURES_ARCH_P (cpu_features, AVX512F_Usable)
+      && !CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_AVX512))
+    {
+      rep_movsb_threshold = 2048 * (64 / 16);
+      minimum_rep_movsb_threshold = 64 * 8;
+    }
+  else if (CPU_FEATURES_ARCH_P (cpu_features,
+ AVX_Fast_Unaligned_Load))
+    {
+      rep_movsb_threshold = 2048 * (32 / 16);
+      minimum_rep_movsb_threshold = 32 * 8;
+    }
+  else
+    {
+      rep_movsb_threshold = 2048 * (16 / 16);
+      minimum_rep_movsb_threshold = 16 * 8;
+    }
+  if (cpu_features->rep_movsb_threshold > minimum_rep_movsb_threshold)
+    __x86_rep_movsb_threshold = cpu_features->rep_movsb_threshold;
+  else
+    __x86_rep_movsb_threshold = rep_movsb_threshold;
+
+# if HAVE_TUNABLES
+  __x86_rep_stosb_threshold = cpu_features->rep_stosb_threshold;
+# endif
 }
 
 #endif
diff --git a/sysdeps/x86/cpu-features.c b/sysdeps/x86/cpu-features.c
index c351bdd54a..c7673a2eb9 100644
--- a/sysdeps/x86/cpu-features.c
+++ b/sysdeps/x86/cpu-features.c
@@ -606,6 +606,10 @@ no_cpuid:
   TUNABLE_GET (hwcaps, tunable_val_t *, TUNABLE_CALLBACK (set_hwcaps));
   cpu_features->non_temporal_threshold
     = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
+  cpu_features->rep_movsb_threshold
+    = TUNABLE_GET (x86_rep_movsb_threshold, long int, NULL);
+  cpu_features->rep_stosb_threshold
+    = TUNABLE_GET (x86_rep_stosb_threshold, long int, NULL);
   cpu_features->data_cache_size
     = TUNABLE_GET (x86_data_cache_size, long int, NULL);
   cpu_features->shared_cache_size
diff --git a/sysdeps/x86/cpu-features.h b/sysdeps/x86/cpu-features.h
index bc425462d6..0383131057 100644
--- a/sysdeps/x86/cpu-features.h
+++ b/sysdeps/x86/cpu-features.h
@@ -102,6 +102,10 @@ struct cpu_features
   unsigned long int shared_cache_size;
   /* Threshold to use non temporal store.  */
   unsigned long int non_temporal_threshold;
+  /* Threshold to use "rep movsb".  */
+  unsigned long int rep_movsb_threshold;
+  /* Threshold to use "rep stosb".  */
+  unsigned long int rep_stosb_threshold;
 };
 
 /* Used from outside of glibc to get access to the CPU features
diff --git a/sysdeps/x86/dl-tunables.list b/sysdeps/x86/dl-tunables.list
index 251b926ce4..1a4a93a070 100644
--- a/sysdeps/x86/dl-tunables.list
+++ b/sysdeps/x86/dl-tunables.list
@@ -30,6 +30,30 @@ glibc {
     x86_non_temporal_threshold {
       type: SIZE_T
     }
+    x86_rep_movsb_threshold {
+      type: SIZE_T
+      # Since there is overhead to set up REP MOVSB operation, REP MOVSB
+      # isn't faster on short data.  The memcpy micro benchmark in glibc
+      # shows that 2KB is the approximate value above which REP MOVSB
+      # becomes faster than SSE2 optimization on processors with Enhanced
+      # REP MOVSB.  Since larger register size can move more data with a
+      # single load and store, the threshold is higher with larger register
+      # size.  Note: Since the REP MOVSB threshold must be greater than 8
+      # times of vector size, the minium value must be updated at run-time.
+      minval: 1
+      default: 2048
+    }
+    x86_rep_stosb_threshold {
+      type: SIZE_T
+      # Since there is overhead to set up REP STOSB operation, REP STOSB
+      # isn't faster on short data.  The memset micro benchmark in glibc
+      # shows that 2KB is the approximate value above which REP STOSB
+      # becomes faster on processors with Enhanced REP STOSB.  Since the
+      # stored value is fixed, larger register size has minimal impact
+      # on threshold.
+      minval: 1
+      default: 2048
+    }
     x86_data_cache_size {
       type: SIZE_T
     }
diff --git a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
index 74953245aa..bd5dc1a3f3 100644
--- a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
+++ b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
@@ -56,17 +56,6 @@
 # endif
 #endif
 
-/* Threshold to use Enhanced REP MOVSB.  Since there is overhead to set
-   up REP MOVSB operation, REP MOVSB isn't faster on short data.  The
-   memcpy micro benchmark in glibc shows that 2KB is the approximate
-   value above which REP MOVSB becomes faster than SSE2 optimization
-   on processors with Enhanced REP MOVSB.  Since larger register size
-   can move more data with a single load and store, the threshold is
-   higher with larger register size.  */
-#ifndef REP_MOVSB_THRESHOLD
-# define REP_MOVSB_THRESHOLD (2048 * (VEC_SIZE / 16))
-#endif
-
 #ifndef PREFETCH
 # define PREFETCH(addr) prefetcht0 addr
 #endif
@@ -253,9 +242,6 @@ L(movsb):
  leaq (%rsi,%rdx), %r9
  cmpq %r9, %rdi
  /* Avoid slow backward REP MOVSB.  */
-# if REP_MOVSB_THRESHOLD <= (VEC_SIZE * 8)
-#  error Unsupported REP_MOVSB_THRESHOLD and VEC_SIZE!
-# endif
  jb L(more_8x_vec_backward)
 1:
  mov %RDX_LP, %RCX_LP
@@ -331,7 +317,7 @@ L(between_2_3):
 
 #if defined USE_MULTIARCH && IS_IN (libc)
 L(movsb_more_2x_vec):
- cmpq $REP_MOVSB_THRESHOLD, %rdx
+ cmp __x86_rep_movsb_threshold(%rip), %RDX_LP
  ja L(movsb)
 #endif
 L(more_2x_vec):
diff --git a/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S b/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
index af2299709c..2bfc95de05 100644
--- a/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
+++ b/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
@@ -58,16 +58,6 @@
 # endif
 #endif
 
-/* Threshold to use Enhanced REP STOSB.  Since there is overhead to set
-   up REP STOSB operation, REP STOSB isn't faster on short data.  The
-   memset micro benchmark in glibc shows that 2KB is the approximate
-   value above which REP STOSB becomes faster on processors with
-   Enhanced REP STOSB.  Since the stored value is fixed, larger register
-   size has minimal impact on threshold.  */
-#ifndef REP_STOSB_THRESHOLD
-# define REP_STOSB_THRESHOLD 2048
-#endif
-
 #ifndef SECTION
 # error SECTION is not defined!
 #endif
@@ -181,7 +171,7 @@ ENTRY (MEMSET_SYMBOL (__memset, unaligned_erms))
  ret
 
 L(stosb_more_2x_vec):
- cmpq $REP_STOSB_THRESHOLD, %rdx
+ cmp __x86_rep_stosb_threshold(%rip), %RDX_LP
  ja L(stosb)
 #endif
 L(more_2x_vec):
--
2.26.2

Reply | Threaded
Open this post in threaded view
|

Re: V3 [PATCH] x86: Add thresholds for "rep movsb/stosb" to tunables

Sourceware - libc-alpha mailing list
On 7/6/20 12:44 PM, H.J. Lu wrote:

>>> +long int __x86_rep_stosb_threshold attribute_hidden = 2048;
>>
>> Are these globals used *before* tunables is initialized?
>
> No.
>
>>
>> That should be the only case that causes us to require to set the default here.
>
> But since tunables is optional, they can be used without tunables.

Thanks, I forgot it's optional, so you need to set the globals.
 
>>
>> Otherwise we should remove the default and express it in dl-tunables.list.
>
> The default is needed when tunables is disabled.

OK.

> Here is the updated patch.  OK for master?

OK for master.

Reviewed-by: Carlos O'Donell <[hidden email]>
 

> Thanks.
>
> H.J.
> ---
> Add x86_rep_movsb_threshold and x86_rep_stosb_threshold to tunables
> to update thresholds for "rep movsb" and "rep stosb" at run-time.
>
> Note that the user specified threshold for "rep movsb" smaller than
> the minimum threshold will be ignored.
> ---
>  manual/tunables.texi                          | 16 +++++++++
>  sysdeps/x86/cacheinfo.c                       | 36 +++++++++++++++++++
>  sysdeps/x86/cpu-features.c                    |  4 +++
>  sysdeps/x86/cpu-features.h                    |  4 +++
>  sysdeps/x86/dl-tunables.list                  | 24 +++++++++++++
>  .../multiarch/memmove-vec-unaligned-erms.S    | 16 +--------
>  .../multiarch/memset-vec-unaligned-erms.S     | 12 +------
>  7 files changed, 86 insertions(+), 26 deletions(-)
>
> diff --git a/manual/tunables.texi b/manual/tunables.texi
> index ec18b10834..4e68c7ff91 100644
> --- a/manual/tunables.texi
> +++ b/manual/tunables.texi
> @@ -396,6 +396,22 @@ to set threshold in bytes for non temporal store.
>  This tunable is specific to i386 and x86-64.
>  @end deftp
>  
> +@deftp Tunable glibc.cpu.x86_rep_movsb_threshold
> +The @code{glibc.cpu.x86_rep_movsb_threshold} tunable allows the user to
> +set threshold in bytes to start using "rep movsb".  The value must be
> +greater than zero, and currently defaults to 2048 bytes.

OK.

> +
> +This tunable is specific to i386 and x86-64.
> +@end deftp
> +
> +@deftp Tunable glibc.cpu.x86_rep_stosb_threshold
> +The @code{glibc.cpu.x86_rep_stosb_threshold} tunable allows the user to
> +set threshold in bytes to start using "rep stosb".  The value must be
> +greater than zero, and currently defaults to 2048 bytes.

OK.

> +
> +This tunable is specific to i386 and x86-64.
> +@end deftp
> +
>  @deftp Tunable glibc.cpu.x86_ibt
>  The @code{glibc.cpu.x86_ibt} tunable allows the user to control how
>  indirect branch tracking (IBT) should be enabled.  Accepted values are
> diff --git a/sysdeps/x86/cacheinfo.c b/sysdeps/x86/cacheinfo.c
> index 311502dee3..136809a6d8 100644
> --- a/sysdeps/x86/cacheinfo.c
> +++ b/sysdeps/x86/cacheinfo.c
> @@ -530,6 +530,12 @@ long int __x86_raw_shared_cache_size attribute_hidden = 1024 * 1024;
>  /* Threshold to use non temporal store.  */
>  long int __x86_shared_non_temporal_threshold attribute_hidden;
>  
> +/* Threshold to use Enhanced REP MOVSB.  */
> +long int __x86_rep_movsb_threshold attribute_hidden = 2048;
> +
> +/* Threshold to use Enhanced REP STOSB.  */
> +long int __x86_rep_stosb_threshold attribute_hidden = 2048;
> +
>  #ifndef DISABLE_PREFETCHW
>  /* PREFETCHW support flag for use in memory and string routines.  */
>  int __x86_prefetchw attribute_hidden;
> @@ -872,6 +878,36 @@ init_cacheinfo (void)
>      = (cpu_features->non_temporal_threshold != 0
>         ? cpu_features->non_temporal_threshold
>         : __x86_shared_cache_size * threads * 3 / 4);
> +
> +  /* NB: The REP MOVSB threshold must be greater than VEC_SIZE * 8.  */
> +  unsigned int minimum_rep_movsb_threshold;
> +  /* NB: The default REP MOVSB threshold is 2048 * (VEC_SIZE / 16).  */
> +  unsigned int rep_movsb_threshold;
> +  if (CPU_FEATURES_ARCH_P (cpu_features, AVX512F_Usable)
> +      && !CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_AVX512))
> +    {
> +      rep_movsb_threshold = 2048 * (64 / 16);
> +      minimum_rep_movsb_threshold = 64 * 8;
> +    }
> +  else if (CPU_FEATURES_ARCH_P (cpu_features,
> + AVX_Fast_Unaligned_Load))
> +    {
> +      rep_movsb_threshold = 2048 * (32 / 16);
> +      minimum_rep_movsb_threshold = 32 * 8;
> +    }
> +  else
> +    {
> +      rep_movsb_threshold = 2048 * (16 / 16);
> +      minimum_rep_movsb_threshold = 16 * 8;
> +    }
> +  if (cpu_features->rep_movsb_threshold > minimum_rep_movsb_threshold)

OK. In theory if tunables are disabled this value is zero and so would
not be greater than the minimum and so we'd use the value from rep_movsb_threshold.

> +    __x86_rep_movsb_threshold = cpu_features->rep_movsb_threshold;
> +  else
> +    __x86_rep_movsb_threshold = rep_movsb_threshold;
> +
> +# if HAVE_TUNABLES
> +  __x86_rep_stosb_threshold = cpu_features->rep_stosb_threshold;
> +# endif
>  }
>  
>  #endif
> diff --git a/sysdeps/x86/cpu-features.c b/sysdeps/x86/cpu-features.c
> index c351bdd54a..c7673a2eb9 100644
> --- a/sysdeps/x86/cpu-features.c
> +++ b/sysdeps/x86/cpu-features.c
> @@ -606,6 +606,10 @@ no_cpuid:
>    TUNABLE_GET (hwcaps, tunable_val_t *, TUNABLE_CALLBACK (set_hwcaps));
>    cpu_features->non_temporal_threshold
>      = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL);
> +  cpu_features->rep_movsb_threshold
> +    = TUNABLE_GET (x86_rep_movsb_threshold, long int, NULL);
> +  cpu_features->rep_stosb_threshold
> +    = TUNABLE_GET (x86_rep_stosb_threshold, long int, NULL);

OK.

>    cpu_features->data_cache_size
>      = TUNABLE_GET (x86_data_cache_size, long int, NULL);
>    cpu_features->shared_cache_size
> diff --git a/sysdeps/x86/cpu-features.h b/sysdeps/x86/cpu-features.h
> index bc425462d6..0383131057 100644
> --- a/sysdeps/x86/cpu-features.h
> +++ b/sysdeps/x86/cpu-features.h
> @@ -102,6 +102,10 @@ struct cpu_features
>    unsigned long int shared_cache_size;
>    /* Threshold to use non temporal store.  */
>    unsigned long int non_temporal_threshold;
> +  /* Threshold to use "rep movsb".  */
> +  unsigned long int rep_movsb_threshold;
> +  /* Threshold to use "rep stosb".  */
> +  unsigned long int rep_stosb_threshold;

OK.

>  };
>  
>  /* Used from outside of glibc to get access to the CPU features
> diff --git a/sysdeps/x86/dl-tunables.list b/sysdeps/x86/dl-tunables.list
> index 251b926ce4..1a4a93a070 100644
> --- a/sysdeps/x86/dl-tunables.list
> +++ b/sysdeps/x86/dl-tunables.list
> @@ -30,6 +30,30 @@ glibc {
>      x86_non_temporal_threshold {
>        type: SIZE_T
>      }
> +    x86_rep_movsb_threshold {
> +      type: SIZE_T
> +      # Since there is overhead to set up REP MOVSB operation, REP MOVSB
> +      # isn't faster on short data.  The memcpy micro benchmark in glibc
> +      # shows that 2KB is the approximate value above which REP MOVSB
> +      # becomes faster than SSE2 optimization on processors with Enhanced
> +      # REP MOVSB.  Since larger register size can move more data with a
> +      # single load and store, the threshold is higher with larger register
> +      # size.  Note: Since the REP MOVSB threshold must be greater than 8
> +      # times of vector size, the minium value must be updated at run-time.
> +      minval: 1
> +      default: 2048
> +    }
> +    x86_rep_stosb_threshold {
> +      type: SIZE_T
> +      # Since there is overhead to set up REP STOSB operation, REP STOSB
> +      # isn't faster on short data.  The memset micro benchmark in glibc
> +      # shows that 2KB is the approximate value above which REP STOSB
> +      # becomes faster on processors with Enhanced REP STOSB.  Since the
> +      # stored value is fixed, larger register size has minimal impact
> +      # on threshold.
> +      minval: 1
> +      default: 2048
> +    }

OK.

>      x86_data_cache_size {
>        type: SIZE_T
>      }
> diff --git a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
> index 74953245aa..bd5dc1a3f3 100644
> --- a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
> +++ b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
> @@ -56,17 +56,6 @@
>  # endif
>  #endif
>  
> -/* Threshold to use Enhanced REP MOVSB.  Since there is overhead to set
> -   up REP MOVSB operation, REP MOVSB isn't faster on short data.  The
> -   memcpy micro benchmark in glibc shows that 2KB is the approximate
> -   value above which REP MOVSB becomes faster than SSE2 optimization
> -   on processors with Enhanced REP MOVSB.  Since larger register size
> -   can move more data with a single load and store, the threshold is
> -   higher with larger register size.  */
> -#ifndef REP_MOVSB_THRESHOLD
> -# define REP_MOVSB_THRESHOLD (2048 * (VEC_SIZE / 16))
> -#endif
> -
>  #ifndef PREFETCH
>  # define PREFETCH(addr) prefetcht0 addr
>  #endif
> @@ -253,9 +242,6 @@ L(movsb):
>   leaq (%rsi,%rdx), %r9
>   cmpq %r9, %rdi
>   /* Avoid slow backward REP MOVSB.  */
> -# if REP_MOVSB_THRESHOLD <= (VEC_SIZE * 8)
> -#  error Unsupported REP_MOVSB_THRESHOLD and VEC_SIZE!
> -# endif
>   jb L(more_8x_vec_backward)
>  1:
>   mov %RDX_LP, %RCX_LP
> @@ -331,7 +317,7 @@ L(between_2_3):
>  
>  #if defined USE_MULTIARCH && IS_IN (libc)
>  L(movsb_more_2x_vec):
> - cmpq $REP_MOVSB_THRESHOLD, %rdx
> + cmp __x86_rep_movsb_threshold(%rip), %RDX_LP
>   ja L(movsb)
>  #endif
>  L(more_2x_vec):
> diff --git a/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S b/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
> index af2299709c..2bfc95de05 100644
> --- a/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
> +++ b/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
> @@ -58,16 +58,6 @@
>  # endif
>  #endif
>  
> -/* Threshold to use Enhanced REP STOSB.  Since there is overhead to set
> -   up REP STOSB operation, REP STOSB isn't faster on short data.  The
> -   memset micro benchmark in glibc shows that 2KB is the approximate
> -   value above which REP STOSB becomes faster on processors with
> -   Enhanced REP STOSB.  Since the stored value is fixed, larger register
> -   size has minimal impact on threshold.  */
> -#ifndef REP_STOSB_THRESHOLD
> -# define REP_STOSB_THRESHOLD 2048
> -#endif
> -
>  #ifndef SECTION
>  # error SECTION is not defined!
>  #endif
> @@ -181,7 +171,7 @@ ENTRY (MEMSET_SYMBOL (__memset, unaligned_erms))
>   ret
>  
>  L(stosb_more_2x_vec):
> - cmpq $REP_STOSB_THRESHOLD, %rdx
> + cmp __x86_rep_stosb_threshold(%rip), %RDX_LP
>   ja L(stosb)
>  #endif
>  L(more_2x_vec):
>


--
Cheers,
Carlos.