|
Packit |
6c4009 |
/* x86_64 cache info.
|
|
Packit |
6c4009 |
Copyright (C) 2003-2018 Free Software Foundation, Inc.
|
|
Packit |
6c4009 |
This file is part of the GNU C Library.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
The GNU C Library is free software; you can redistribute it and/or
|
|
Packit |
6c4009 |
modify it under the terms of the GNU Lesser General Public
|
|
Packit |
6c4009 |
License as published by the Free Software Foundation; either
|
|
Packit |
6c4009 |
version 2.1 of the License, or (at your option) any later version.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
The GNU C Library is distributed in the hope that it will be useful,
|
|
Packit |
6c4009 |
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
Packit |
6c4009 |
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
Packit |
6c4009 |
Lesser General Public License for more details.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
You should have received a copy of the GNU Lesser General Public
|
|
Packit |
6c4009 |
License along with the GNU C Library; if not, see
|
|
Packit |
6c4009 |
<http://www.gnu.org/licenses/>. */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#if IS_IN (libc)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#include <assert.h>
|
|
Packit |
6c4009 |
#include <stdbool.h>
|
|
Packit |
6c4009 |
#include <stdlib.h>
|
|
Packit |
6c4009 |
#include <unistd.h>
|
|
Packit |
6c4009 |
#include <cpuid.h>
|
|
Packit |
6c4009 |
#include <init-arch.h>
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
static const struct intel_02_cache_info
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
unsigned char idx;
|
|
Packit |
6c4009 |
unsigned char assoc;
|
|
Packit |
6c4009 |
unsigned char linesize;
|
|
Packit |
6c4009 |
unsigned char rel_name;
|
|
Packit |
6c4009 |
unsigned int size;
|
|
Packit |
6c4009 |
} intel_02_known [] =
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
#define M(sc) ((sc) - _SC_LEVEL1_ICACHE_SIZE)
|
|
Packit |
6c4009 |
{ 0x06, 4, 32, M(_SC_LEVEL1_ICACHE_SIZE), 8192 },
|
|
Packit |
6c4009 |
{ 0x08, 4, 32, M(_SC_LEVEL1_ICACHE_SIZE), 16384 },
|
|
Packit |
6c4009 |
{ 0x09, 4, 32, M(_SC_LEVEL1_ICACHE_SIZE), 32768 },
|
|
Packit |
6c4009 |
{ 0x0a, 2, 32, M(_SC_LEVEL1_DCACHE_SIZE), 8192 },
|
|
Packit |
6c4009 |
{ 0x0c, 4, 32, M(_SC_LEVEL1_DCACHE_SIZE), 16384 },
|
|
Packit |
6c4009 |
{ 0x0d, 4, 64, M(_SC_LEVEL1_DCACHE_SIZE), 16384 },
|
|
Packit |
6c4009 |
{ 0x0e, 6, 64, M(_SC_LEVEL1_DCACHE_SIZE), 24576 },
|
|
Packit |
6c4009 |
{ 0x21, 8, 64, M(_SC_LEVEL2_CACHE_SIZE), 262144 },
|
|
Packit |
6c4009 |
{ 0x22, 4, 64, M(_SC_LEVEL3_CACHE_SIZE), 524288 },
|
|
Packit |
6c4009 |
{ 0x23, 8, 64, M(_SC_LEVEL3_CACHE_SIZE), 1048576 },
|
|
Packit |
6c4009 |
{ 0x25, 8, 64, M(_SC_LEVEL3_CACHE_SIZE), 2097152 },
|
|
Packit |
6c4009 |
{ 0x29, 8, 64, M(_SC_LEVEL3_CACHE_SIZE), 4194304 },
|
|
Packit |
6c4009 |
{ 0x2c, 8, 64, M(_SC_LEVEL1_DCACHE_SIZE), 32768 },
|
|
Packit |
6c4009 |
{ 0x30, 8, 64, M(_SC_LEVEL1_ICACHE_SIZE), 32768 },
|
|
Packit |
6c4009 |
{ 0x39, 4, 64, M(_SC_LEVEL2_CACHE_SIZE), 131072 },
|
|
Packit |
6c4009 |
{ 0x3a, 6, 64, M(_SC_LEVEL2_CACHE_SIZE), 196608 },
|
|
Packit |
6c4009 |
{ 0x3b, 2, 64, M(_SC_LEVEL2_CACHE_SIZE), 131072 },
|
|
Packit |
6c4009 |
{ 0x3c, 4, 64, M(_SC_LEVEL2_CACHE_SIZE), 262144 },
|
|
Packit |
6c4009 |
{ 0x3d, 6, 64, M(_SC_LEVEL2_CACHE_SIZE), 393216 },
|
|
Packit |
6c4009 |
{ 0x3e, 4, 64, M(_SC_LEVEL2_CACHE_SIZE), 524288 },
|
|
Packit |
6c4009 |
{ 0x3f, 2, 64, M(_SC_LEVEL2_CACHE_SIZE), 262144 },
|
|
Packit |
6c4009 |
{ 0x41, 4, 32, M(_SC_LEVEL2_CACHE_SIZE), 131072 },
|
|
Packit |
6c4009 |
{ 0x42, 4, 32, M(_SC_LEVEL2_CACHE_SIZE), 262144 },
|
|
Packit |
6c4009 |
{ 0x43, 4, 32, M(_SC_LEVEL2_CACHE_SIZE), 524288 },
|
|
Packit |
6c4009 |
{ 0x44, 4, 32, M(_SC_LEVEL2_CACHE_SIZE), 1048576 },
|
|
Packit |
6c4009 |
{ 0x45, 4, 32, M(_SC_LEVEL2_CACHE_SIZE), 2097152 },
|
|
Packit |
6c4009 |
{ 0x46, 4, 64, M(_SC_LEVEL3_CACHE_SIZE), 4194304 },
|
|
Packit |
6c4009 |
{ 0x47, 8, 64, M(_SC_LEVEL3_CACHE_SIZE), 8388608 },
|
|
Packit |
6c4009 |
{ 0x48, 12, 64, M(_SC_LEVEL2_CACHE_SIZE), 3145728 },
|
|
Packit |
6c4009 |
{ 0x49, 16, 64, M(_SC_LEVEL2_CACHE_SIZE), 4194304 },
|
|
Packit |
6c4009 |
{ 0x4a, 12, 64, M(_SC_LEVEL3_CACHE_SIZE), 6291456 },
|
|
Packit |
6c4009 |
{ 0x4b, 16, 64, M(_SC_LEVEL3_CACHE_SIZE), 8388608 },
|
|
Packit |
6c4009 |
{ 0x4c, 12, 64, M(_SC_LEVEL3_CACHE_SIZE), 12582912 },
|
|
Packit |
6c4009 |
{ 0x4d, 16, 64, M(_SC_LEVEL3_CACHE_SIZE), 16777216 },
|
|
Packit |
6c4009 |
{ 0x4e, 24, 64, M(_SC_LEVEL2_CACHE_SIZE), 6291456 },
|
|
Packit |
6c4009 |
{ 0x60, 8, 64, M(_SC_LEVEL1_DCACHE_SIZE), 16384 },
|
|
Packit |
6c4009 |
{ 0x66, 4, 64, M(_SC_LEVEL1_DCACHE_SIZE), 8192 },
|
|
Packit |
6c4009 |
{ 0x67, 4, 64, M(_SC_LEVEL1_DCACHE_SIZE), 16384 },
|
|
Packit |
6c4009 |
{ 0x68, 4, 64, M(_SC_LEVEL1_DCACHE_SIZE), 32768 },
|
|
Packit |
6c4009 |
{ 0x78, 8, 64, M(_SC_LEVEL2_CACHE_SIZE), 1048576 },
|
|
Packit |
6c4009 |
{ 0x79, 8, 64, M(_SC_LEVEL2_CACHE_SIZE), 131072 },
|
|
Packit |
6c4009 |
{ 0x7a, 8, 64, M(_SC_LEVEL2_CACHE_SIZE), 262144 },
|
|
Packit |
6c4009 |
{ 0x7b, 8, 64, M(_SC_LEVEL2_CACHE_SIZE), 524288 },
|
|
Packit |
6c4009 |
{ 0x7c, 8, 64, M(_SC_LEVEL2_CACHE_SIZE), 1048576 },
|
|
Packit |
6c4009 |
{ 0x7d, 8, 64, M(_SC_LEVEL2_CACHE_SIZE), 2097152 },
|
|
Packit |
6c4009 |
{ 0x7f, 2, 64, M(_SC_LEVEL2_CACHE_SIZE), 524288 },
|
|
Packit |
6c4009 |
{ 0x80, 8, 64, M(_SC_LEVEL2_CACHE_SIZE), 524288 },
|
|
Packit |
6c4009 |
{ 0x82, 8, 32, M(_SC_LEVEL2_CACHE_SIZE), 262144 },
|
|
Packit |
6c4009 |
{ 0x83, 8, 32, M(_SC_LEVEL2_CACHE_SIZE), 524288 },
|
|
Packit |
6c4009 |
{ 0x84, 8, 32, M(_SC_LEVEL2_CACHE_SIZE), 1048576 },
|
|
Packit |
6c4009 |
{ 0x85, 8, 32, M(_SC_LEVEL2_CACHE_SIZE), 2097152 },
|
|
Packit |
6c4009 |
{ 0x86, 4, 64, M(_SC_LEVEL2_CACHE_SIZE), 524288 },
|
|
Packit |
6c4009 |
{ 0x87, 8, 64, M(_SC_LEVEL2_CACHE_SIZE), 1048576 },
|
|
Packit |
6c4009 |
{ 0xd0, 4, 64, M(_SC_LEVEL3_CACHE_SIZE), 524288 },
|
|
Packit |
6c4009 |
{ 0xd1, 4, 64, M(_SC_LEVEL3_CACHE_SIZE), 1048576 },
|
|
Packit |
6c4009 |
{ 0xd2, 4, 64, M(_SC_LEVEL3_CACHE_SIZE), 2097152 },
|
|
Packit |
6c4009 |
{ 0xd6, 8, 64, M(_SC_LEVEL3_CACHE_SIZE), 1048576 },
|
|
Packit |
6c4009 |
{ 0xd7, 8, 64, M(_SC_LEVEL3_CACHE_SIZE), 2097152 },
|
|
Packit |
6c4009 |
{ 0xd8, 8, 64, M(_SC_LEVEL3_CACHE_SIZE), 4194304 },
|
|
Packit |
6c4009 |
{ 0xdc, 12, 64, M(_SC_LEVEL3_CACHE_SIZE), 2097152 },
|
|
Packit |
6c4009 |
{ 0xdd, 12, 64, M(_SC_LEVEL3_CACHE_SIZE), 4194304 },
|
|
Packit |
6c4009 |
{ 0xde, 12, 64, M(_SC_LEVEL3_CACHE_SIZE), 8388608 },
|
|
Packit |
6c4009 |
{ 0xe2, 16, 64, M(_SC_LEVEL3_CACHE_SIZE), 2097152 },
|
|
Packit |
6c4009 |
{ 0xe3, 16, 64, M(_SC_LEVEL3_CACHE_SIZE), 4194304 },
|
|
Packit |
6c4009 |
{ 0xe4, 16, 64, M(_SC_LEVEL3_CACHE_SIZE), 8388608 },
|
|
Packit |
6c4009 |
{ 0xea, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 12582912 },
|
|
Packit |
6c4009 |
{ 0xeb, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 18874368 },
|
|
Packit |
6c4009 |
{ 0xec, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 25165824 },
|
|
Packit |
6c4009 |
};
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#define nintel_02_known (sizeof (intel_02_known) / sizeof (intel_02_known [0]))
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
static int
|
|
Packit |
6c4009 |
intel_02_known_compare (const void *p1, const void *p2)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
const struct intel_02_cache_info *i1;
|
|
Packit |
6c4009 |
const struct intel_02_cache_info *i2;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
i1 = (const struct intel_02_cache_info *) p1;
|
|
Packit |
6c4009 |
i2 = (const struct intel_02_cache_info *) p2;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
if (i1->idx == i2->idx)
|
|
Packit |
6c4009 |
return 0;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
return i1->idx < i2->idx ? -1 : 1;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
static long int
|
|
Packit |
6c4009 |
__attribute__ ((noinline))
|
|
Packit |
6c4009 |
intel_check_word (int name, unsigned int value, bool *has_level_2,
|
|
Packit |
6c4009 |
bool *no_level_2_or_3,
|
|
Packit |
6c4009 |
const struct cpu_features *cpu_features)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
if ((value & 0x80000000) != 0)
|
|
Packit |
6c4009 |
/* The register value is reserved. */
|
|
Packit |
6c4009 |
return 0;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Fold the name. The _SC_ constants are always in the order SIZE,
|
|
Packit |
6c4009 |
ASSOC, LINESIZE. */
|
|
Packit |
6c4009 |
int folded_rel_name = (M(name) / 3) * 3;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
while (value != 0)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
unsigned int byte = value & 0xff;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
if (byte == 0x40)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
*no_level_2_or_3 = true;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
if (folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))
|
|
Packit |
6c4009 |
/* No need to look further. */
|
|
Packit |
6c4009 |
break;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
else if (byte == 0xff)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
/* CPUID leaf 0x4 contains all the information. We need to
|
|
Packit |
6c4009 |
iterate over it. */
|
|
Packit |
6c4009 |
unsigned int eax;
|
|
Packit |
6c4009 |
unsigned int ebx;
|
|
Packit |
6c4009 |
unsigned int ecx;
|
|
Packit |
6c4009 |
unsigned int edx;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
unsigned int round = 0;
|
|
Packit |
6c4009 |
while (1)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
__cpuid_count (4, round, eax, ebx, ecx, edx);
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
enum { null = 0, data = 1, inst = 2, uni = 3 } type = eax & 0x1f;
|
|
Packit |
6c4009 |
if (type == null)
|
|
Packit |
6c4009 |
/* That was the end. */
|
|
Packit |
6c4009 |
break;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
unsigned int level = (eax >> 5) & 0x7;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
if ((level == 1 && type == data
|
|
Packit |
6c4009 |
&& folded_rel_name == M(_SC_LEVEL1_DCACHE_SIZE))
|
|
Packit |
6c4009 |
|| (level == 1 && type == inst
|
|
Packit |
6c4009 |
&& folded_rel_name == M(_SC_LEVEL1_ICACHE_SIZE))
|
|
Packit |
6c4009 |
|| (level == 2 && folded_rel_name == M(_SC_LEVEL2_CACHE_SIZE))
|
|
Packit |
6c4009 |
|| (level == 3 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))
|
|
Packit |
6c4009 |
|| (level == 4 && folded_rel_name == M(_SC_LEVEL4_CACHE_SIZE)))
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
unsigned int offset = M(name) - folded_rel_name;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
if (offset == 0)
|
|
Packit |
6c4009 |
/* Cache size. */
|
|
Packit |
6c4009 |
return (((ebx >> 22) + 1)
|
|
Packit |
6c4009 |
* (((ebx >> 12) & 0x3ff) + 1)
|
|
Packit |
6c4009 |
* ((ebx & 0xfff) + 1)
|
|
Packit |
6c4009 |
* (ecx + 1));
|
|
Packit |
6c4009 |
if (offset == 1)
|
|
Packit |
6c4009 |
return (ebx >> 22) + 1;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
assert (offset == 2);
|
|
Packit |
6c4009 |
return (ebx & 0xfff) + 1;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
++round;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
/* There is no other cache information anywhere else. */
|
|
Packit |
6c4009 |
break;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
else
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
if (byte == 0x49 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
/* Intel reused this value. For family 15, model 6 it
|
|
Packit |
6c4009 |
specifies the 3rd level cache. Otherwise the 2nd
|
|
Packit |
6c4009 |
level cache. */
|
|
Packit |
6c4009 |
unsigned int family = cpu_features->family;
|
|
Packit |
6c4009 |
unsigned int model = cpu_features->model;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
if (family == 15 && model == 6)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
/* The level 3 cache is encoded for this model like
|
|
Packit |
6c4009 |
the level 2 cache is for other models. Pretend
|
|
Packit |
6c4009 |
the caller asked for the level 2 cache. */
|
|
Packit |
6c4009 |
name = (_SC_LEVEL2_CACHE_SIZE
|
|
Packit |
6c4009 |
+ (name - _SC_LEVEL3_CACHE_SIZE));
|
|
Packit |
6c4009 |
folded_rel_name = M(_SC_LEVEL2_CACHE_SIZE);
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
struct intel_02_cache_info *found;
|
|
Packit |
6c4009 |
struct intel_02_cache_info search;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
search.idx = byte;
|
|
Packit |
6c4009 |
found = bsearch (&search, intel_02_known, nintel_02_known,
|
|
Packit |
6c4009 |
sizeof (intel_02_known[0]), intel_02_known_compare);
|
|
Packit |
6c4009 |
if (found != NULL)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
if (found->rel_name == folded_rel_name)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
unsigned int offset = M(name) - folded_rel_name;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
if (offset == 0)
|
|
Packit |
6c4009 |
/* Cache size. */
|
|
Packit |
6c4009 |
return found->size;
|
|
Packit |
6c4009 |
if (offset == 1)
|
|
Packit |
6c4009 |
return found->assoc;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
assert (offset == 2);
|
|
Packit |
6c4009 |
return found->linesize;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
if (found->rel_name == M(_SC_LEVEL2_CACHE_SIZE))
|
|
Packit |
6c4009 |
*has_level_2 = true;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Next byte for the next round. */
|
|
Packit |
6c4009 |
value >>= 8;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Nothing found. */
|
|
Packit |
6c4009 |
return 0;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
static long int __attribute__ ((noinline))
|
|
Packit |
6c4009 |
handle_intel (int name, const struct cpu_features *cpu_features)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
unsigned int maxidx = cpu_features->max_cpuid;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Return -1 for older CPUs. */
|
|
Packit |
6c4009 |
if (maxidx < 2)
|
|
Packit |
6c4009 |
return -1;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* OK, we can use the CPUID instruction to get all info about the
|
|
Packit |
6c4009 |
caches. */
|
|
Packit |
6c4009 |
unsigned int cnt = 0;
|
|
Packit |
6c4009 |
unsigned int max = 1;
|
|
Packit |
6c4009 |
long int result = 0;
|
|
Packit |
6c4009 |
bool no_level_2_or_3 = false;
|
|
Packit |
6c4009 |
bool has_level_2 = false;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
while (cnt++ < max)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
unsigned int eax;
|
|
Packit |
6c4009 |
unsigned int ebx;
|
|
Packit |
6c4009 |
unsigned int ecx;
|
|
Packit |
6c4009 |
unsigned int edx;
|
|
Packit |
6c4009 |
__cpuid (2, eax, ebx, ecx, edx);
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* The low byte of EAX in the first round contain the number of
|
|
Packit |
6c4009 |
rounds we have to make. At least one, the one we are already
|
|
Packit |
6c4009 |
doing. */
|
|
Packit |
6c4009 |
if (cnt == 1)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
max = eax & 0xff;
|
|
Packit |
6c4009 |
eax &= 0xffffff00;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Process the individual registers' value. */
|
|
Packit |
6c4009 |
result = intel_check_word (name, eax, &has_level_2,
|
|
Packit |
6c4009 |
&no_level_2_or_3, cpu_features);
|
|
Packit |
6c4009 |
if (result != 0)
|
|
Packit |
6c4009 |
return result;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
result = intel_check_word (name, ebx, &has_level_2,
|
|
Packit |
6c4009 |
&no_level_2_or_3, cpu_features);
|
|
Packit |
6c4009 |
if (result != 0)
|
|
Packit |
6c4009 |
return result;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
result = intel_check_word (name, ecx, &has_level_2,
|
|
Packit |
6c4009 |
&no_level_2_or_3, cpu_features);
|
|
Packit |
6c4009 |
if (result != 0)
|
|
Packit |
6c4009 |
return result;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
result = intel_check_word (name, edx, &has_level_2,
|
|
Packit |
6c4009 |
&no_level_2_or_3, cpu_features);
|
|
Packit |
6c4009 |
if (result != 0)
|
|
Packit |
6c4009 |
return result;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
if (name >= _SC_LEVEL2_CACHE_SIZE && name <= _SC_LEVEL3_CACHE_LINESIZE
|
|
Packit |
6c4009 |
&& no_level_2_or_3)
|
|
Packit |
6c4009 |
return -1;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
return 0;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
static long int __attribute__ ((noinline))
|
|
Packit |
6c4009 |
handle_amd (int name)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
unsigned int eax;
|
|
Packit |
6c4009 |
unsigned int ebx;
|
|
Packit |
6c4009 |
unsigned int ecx;
|
|
Packit |
6c4009 |
unsigned int edx;
|
|
Packit |
6c4009 |
__cpuid (0x80000000, eax, ebx, ecx, edx);
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* No level 4 cache (yet). */
|
|
Packit |
6c4009 |
if (name > _SC_LEVEL3_CACHE_LINESIZE)
|
|
Packit |
6c4009 |
return 0;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
unsigned int fn = 0x80000005 + (name >= _SC_LEVEL2_CACHE_SIZE);
|
|
Packit |
6c4009 |
if (eax < fn)
|
|
Packit |
6c4009 |
return 0;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
__cpuid (fn, eax, ebx, ecx, edx);
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
if (name < _SC_LEVEL1_DCACHE_SIZE)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
name += _SC_LEVEL1_DCACHE_SIZE - _SC_LEVEL1_ICACHE_SIZE;
|
|
Packit |
6c4009 |
ecx = edx;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
switch (name)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
case _SC_LEVEL1_DCACHE_SIZE:
|
|
Packit |
6c4009 |
return (ecx >> 14) & 0x3fc00;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
case _SC_LEVEL1_DCACHE_ASSOC:
|
|
Packit |
6c4009 |
ecx >>= 16;
|
|
Packit |
6c4009 |
if ((ecx & 0xff) == 0xff)
|
|
Packit |
6c4009 |
/* Fully associative. */
|
|
Packit |
6c4009 |
return (ecx << 2) & 0x3fc00;
|
|
Packit |
6c4009 |
return ecx & 0xff;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
case _SC_LEVEL1_DCACHE_LINESIZE:
|
|
Packit |
6c4009 |
return ecx & 0xff;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
case _SC_LEVEL2_CACHE_SIZE:
|
|
Packit |
6c4009 |
return (ecx & 0xf000) == 0 ? 0 : (ecx >> 6) & 0x3fffc00;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
case _SC_LEVEL2_CACHE_ASSOC:
|
|
Packit |
6c4009 |
switch ((ecx >> 12) & 0xf)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
case 0:
|
|
Packit |
6c4009 |
case 1:
|
|
Packit |
6c4009 |
case 2:
|
|
Packit |
6c4009 |
case 4:
|
|
Packit |
6c4009 |
return (ecx >> 12) & 0xf;
|
|
Packit |
6c4009 |
case 6:
|
|
Packit |
6c4009 |
return 8;
|
|
Packit |
6c4009 |
case 8:
|
|
Packit |
6c4009 |
return 16;
|
|
Packit |
6c4009 |
case 10:
|
|
Packit |
6c4009 |
return 32;
|
|
Packit |
6c4009 |
case 11:
|
|
Packit |
6c4009 |
return 48;
|
|
Packit |
6c4009 |
case 12:
|
|
Packit |
6c4009 |
return 64;
|
|
Packit |
6c4009 |
case 13:
|
|
Packit |
6c4009 |
return 96;
|
|
Packit |
6c4009 |
case 14:
|
|
Packit |
6c4009 |
return 128;
|
|
Packit |
6c4009 |
case 15:
|
|
Packit |
6c4009 |
return ((ecx >> 6) & 0x3fffc00) / (ecx & 0xff);
|
|
Packit |
6c4009 |
default:
|
|
Packit |
6c4009 |
return 0;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
/* NOTREACHED */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
case _SC_LEVEL2_CACHE_LINESIZE:
|
|
Packit |
6c4009 |
return (ecx & 0xf000) == 0 ? 0 : ecx & 0xff;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
case _SC_LEVEL3_CACHE_SIZE:
|
|
Packit |
6c4009 |
return (edx & 0xf000) == 0 ? 0 : (edx & 0x3ffc0000) << 1;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
case _SC_LEVEL3_CACHE_ASSOC:
|
|
Packit |
6c4009 |
switch ((edx >> 12) & 0xf)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
case 0:
|
|
Packit |
6c4009 |
case 1:
|
|
Packit |
6c4009 |
case 2:
|
|
Packit |
6c4009 |
case 4:
|
|
Packit |
6c4009 |
return (edx >> 12) & 0xf;
|
|
Packit |
6c4009 |
case 6:
|
|
Packit |
6c4009 |
return 8;
|
|
Packit |
6c4009 |
case 8:
|
|
Packit |
6c4009 |
return 16;
|
|
Packit |
6c4009 |
case 10:
|
|
Packit |
6c4009 |
return 32;
|
|
Packit |
6c4009 |
case 11:
|
|
Packit |
6c4009 |
return 48;
|
|
Packit |
6c4009 |
case 12:
|
|
Packit |
6c4009 |
return 64;
|
|
Packit |
6c4009 |
case 13:
|
|
Packit |
6c4009 |
return 96;
|
|
Packit |
6c4009 |
case 14:
|
|
Packit |
6c4009 |
return 128;
|
|
Packit |
6c4009 |
case 15:
|
|
Packit |
6c4009 |
return ((edx & 0x3ffc0000) << 1) / (edx & 0xff);
|
|
Packit |
6c4009 |
default:
|
|
Packit |
6c4009 |
return 0;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
/* NOTREACHED */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
case _SC_LEVEL3_CACHE_LINESIZE:
|
|
Packit |
6c4009 |
return (edx & 0xf000) == 0 ? 0 : edx & 0xff;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
default:
|
|
Packit |
6c4009 |
assert (! "cannot happen");
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
return -1;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Get the value of the system variable NAME. */
|
|
Packit |
6c4009 |
long int
|
|
Packit |
6c4009 |
attribute_hidden
|
|
Packit |
6c4009 |
__cache_sysconf (int name)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
const struct cpu_features *cpu_features = __get_cpu_features ();
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
if (cpu_features->kind == arch_kind_intel)
|
|
Packit |
6c4009 |
return handle_intel (name, cpu_features);
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
if (cpu_features->kind == arch_kind_amd)
|
|
Packit |
6c4009 |
return handle_amd (name);
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
// XXX Fill in more vendors.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* CPU not known, we have no information. */
|
|
Packit |
6c4009 |
return 0;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Data cache size for use in memory and string routines, typically
|
|
Packit |
6c4009 |
L1 size, rounded to multiple of 256 bytes. */
|
|
Packit |
6c4009 |
long int __x86_data_cache_size_half attribute_hidden = 32 * 1024 / 2;
|
|
Packit |
6c4009 |
long int __x86_data_cache_size attribute_hidden = 32 * 1024;
|
|
Packit |
6c4009 |
/* Similar to __x86_data_cache_size_half, but not rounded. */
|
|
Packit |
6c4009 |
long int __x86_raw_data_cache_size_half attribute_hidden = 32 * 1024 / 2;
|
|
Packit |
6c4009 |
/* Similar to __x86_data_cache_size, but not rounded. */
|
|
Packit |
6c4009 |
long int __x86_raw_data_cache_size attribute_hidden = 32 * 1024;
|
|
Packit |
6c4009 |
/* Shared cache size for use in memory and string routines, typically
|
|
Packit |
6c4009 |
L2 or L3 size, rounded to multiple of 256 bytes. */
|
|
Packit |
6c4009 |
long int __x86_shared_cache_size_half attribute_hidden = 1024 * 1024 / 2;
|
|
Packit |
6c4009 |
long int __x86_shared_cache_size attribute_hidden = 1024 * 1024;
|
|
Packit |
6c4009 |
/* Similar to __x86_shared_cache_size_half, but not rounded. */
|
|
Packit |
6c4009 |
long int __x86_raw_shared_cache_size_half attribute_hidden = 1024 * 1024 / 2;
|
|
Packit |
6c4009 |
/* Similar to __x86_shared_cache_size, but not rounded. */
|
|
Packit |
6c4009 |
long int __x86_raw_shared_cache_size attribute_hidden = 1024 * 1024;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Threshold to use non temporal store. */
|
|
Packit |
6c4009 |
long int __x86_shared_non_temporal_threshold attribute_hidden;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#ifndef DISABLE_PREFETCHW
|
|
Packit |
6c4009 |
/* PREFETCHW support flag for use in memory and string routines. */
|
|
Packit |
6c4009 |
int __x86_prefetchw attribute_hidden;
|
|
Packit |
6c4009 |
#endif
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
static void
|
|
Packit |
6c4009 |
__attribute__((constructor))
|
|
Packit |
6c4009 |
init_cacheinfo (void)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
/* Find out what brand of processor. */
|
|
Packit |
6c4009 |
unsigned int eax;
|
|
Packit |
6c4009 |
unsigned int ebx;
|
|
Packit |
6c4009 |
unsigned int ecx;
|
|
Packit |
6c4009 |
unsigned int edx;
|
|
Packit |
6c4009 |
int max_cpuid_ex;
|
|
Packit |
6c4009 |
long int data = -1;
|
|
Packit |
6c4009 |
long int shared = -1;
|
|
Packit |
6c4009 |
unsigned int level;
|
|
Packit |
6c4009 |
unsigned int threads = 0;
|
|
Packit |
6c4009 |
const struct cpu_features *cpu_features = __get_cpu_features ();
|
|
Packit |
6c4009 |
int max_cpuid = cpu_features->max_cpuid;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
if (cpu_features->kind == arch_kind_intel)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
data = handle_intel (_SC_LEVEL1_DCACHE_SIZE, cpu_features);
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
long int core = handle_intel (_SC_LEVEL2_CACHE_SIZE, cpu_features);
|
|
Packit |
6c4009 |
bool inclusive_cache = true;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Try L3 first. */
|
|
Packit |
6c4009 |
level = 3;
|
|
Packit |
6c4009 |
shared = handle_intel (_SC_LEVEL3_CACHE_SIZE, cpu_features);
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Number of logical processors sharing L2 cache. */
|
|
Packit |
6c4009 |
int threads_l2;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Number of logical processors sharing L3 cache. */
|
|
Packit |
6c4009 |
int threads_l3;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
if (shared <= 0)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
/* Try L2 otherwise. */
|
|
Packit |
6c4009 |
level = 2;
|
|
Packit |
6c4009 |
shared = core;
|
|
Packit |
6c4009 |
threads_l2 = 0;
|
|
Packit |
6c4009 |
threads_l3 = -1;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
else
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
threads_l2 = 0;
|
|
Packit |
6c4009 |
threads_l3 = 0;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* A value of 0 for the HTT bit indicates there is only a single
|
|
Packit |
6c4009 |
logical processor. */
|
|
Packit |
6c4009 |
if (HAS_CPU_FEATURE (HTT))
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
/* Figure out the number of logical threads that share the
|
|
Packit |
6c4009 |
highest cache level. */
|
|
Packit |
6c4009 |
if (max_cpuid >= 4)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
unsigned int family = cpu_features->family;
|
|
Packit |
6c4009 |
unsigned int model = cpu_features->model;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
int i = 0;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Query until cache level 2 and 3 are enumerated. */
|
|
Packit |
6c4009 |
int check = 0x1 | (threads_l3 == 0) << 1;
|
|
Packit |
6c4009 |
do
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
__cpuid_count (4, i++, eax, ebx, ecx, edx);
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* There seems to be a bug in at least some Pentium Ds
|
|
Packit |
6c4009 |
which sometimes fail to iterate all cache parameters.
|
|
Packit |
6c4009 |
Do not loop indefinitely here, stop in this case and
|
|
Packit |
6c4009 |
assume there is no such information. */
|
|
Packit |
6c4009 |
if ((eax & 0x1f) == 0)
|
|
Packit |
6c4009 |
goto intel_bug_no_cache_info;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
switch ((eax >> 5) & 0x7)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
default:
|
|
Packit |
6c4009 |
break;
|
|
Packit |
6c4009 |
case 2:
|
|
Packit |
6c4009 |
if ((check & 0x1))
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
/* Get maximum number of logical processors
|
|
Packit |
6c4009 |
sharing L2 cache. */
|
|
Packit |
6c4009 |
threads_l2 = (eax >> 14) & 0x3ff;
|
|
Packit |
6c4009 |
check &= ~0x1;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
break;
|
|
Packit |
6c4009 |
case 3:
|
|
Packit |
6c4009 |
if ((check & (0x1 << 1)))
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
/* Get maximum number of logical processors
|
|
Packit |
6c4009 |
sharing L3 cache. */
|
|
Packit |
6c4009 |
threads_l3 = (eax >> 14) & 0x3ff;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Check if L2 and L3 caches are inclusive. */
|
|
Packit |
6c4009 |
inclusive_cache = (edx & 0x2) != 0;
|
|
Packit |
6c4009 |
check &= ~(0x1 << 1);
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
break;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
while (check);
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* If max_cpuid >= 11, THREADS_L2/THREADS_L3 are the maximum
|
|
Packit |
6c4009 |
numbers of addressable IDs for logical processors sharing
|
|
Packit |
6c4009 |
the cache, instead of the maximum number of threads
|
|
Packit |
6c4009 |
sharing the cache. */
|
|
Packit |
6c4009 |
if (max_cpuid >= 11)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
/* Find the number of logical processors shipped in
|
|
Packit |
6c4009 |
one core and apply count mask. */
|
|
Packit |
6c4009 |
i = 0;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Count SMT only if there is L3 cache. Always count
|
|
Packit |
6c4009 |
core if there is no L3 cache. */
|
|
Packit |
6c4009 |
int count = ((threads_l2 > 0 && level == 3)
|
|
Packit |
6c4009 |
| ((threads_l3 > 0
|
|
Packit |
6c4009 |
|| (threads_l2 > 0 && level == 2)) << 1));
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
while (count)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
__cpuid_count (11, i++, eax, ebx, ecx, edx);
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
int shipped = ebx & 0xff;
|
|
Packit |
6c4009 |
int type = ecx & 0xff00;
|
|
Packit |
6c4009 |
if (shipped == 0 || type == 0)
|
|
Packit |
6c4009 |
break;
|
|
Packit |
6c4009 |
else if (type == 0x100)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
/* Count SMT. */
|
|
Packit |
6c4009 |
if ((count & 0x1))
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
int count_mask;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Compute count mask. */
|
|
Packit |
6c4009 |
asm ("bsr %1, %0"
|
|
Packit |
6c4009 |
: "=r" (count_mask) : "g" (threads_l2));
|
|
Packit |
6c4009 |
count_mask = ~(-1 << (count_mask + 1));
|
|
Packit |
6c4009 |
threads_l2 = (shipped - 1) & count_mask;
|
|
Packit |
6c4009 |
count &= ~0x1;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
else if (type == 0x200)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
/* Count core. */
|
|
Packit |
6c4009 |
if ((count & (0x1 << 1)))
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
int count_mask;
|
|
Packit |
6c4009 |
int threads_core
|
|
Packit |
6c4009 |
= (level == 2 ? threads_l2 : threads_l3);
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Compute count mask. */
|
|
Packit |
6c4009 |
asm ("bsr %1, %0"
|
|
Packit |
6c4009 |
: "=r" (count_mask) : "g" (threads_core));
|
|
Packit |
6c4009 |
count_mask = ~(-1 << (count_mask + 1));
|
|
Packit |
6c4009 |
threads_core = (shipped - 1) & count_mask;
|
|
Packit |
6c4009 |
if (level == 2)
|
|
Packit |
6c4009 |
threads_l2 = threads_core;
|
|
Packit |
6c4009 |
else
|
|
Packit |
6c4009 |
threads_l3 = threads_core;
|
|
Packit |
6c4009 |
count &= ~(0x1 << 1);
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
if (threads_l2 > 0)
|
|
Packit |
6c4009 |
threads_l2 += 1;
|
|
Packit |
6c4009 |
if (threads_l3 > 0)
|
|
Packit |
6c4009 |
threads_l3 += 1;
|
|
Packit |
6c4009 |
if (level == 2)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
if (threads_l2)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
threads = threads_l2;
|
|
Packit |
6c4009 |
if (threads > 2 && family == 6)
|
|
Packit |
6c4009 |
switch (model)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
case 0x37:
|
|
Packit |
6c4009 |
case 0x4a:
|
|
Packit |
6c4009 |
case 0x4d:
|
|
Packit |
6c4009 |
case 0x5a:
|
|
Packit |
6c4009 |
case 0x5d:
|
|
Packit |
6c4009 |
/* Silvermont has L2 cache shared by 2 cores. */
|
|
Packit |
6c4009 |
threads = 2;
|
|
Packit |
6c4009 |
break;
|
|
Packit |
6c4009 |
default:
|
|
Packit |
6c4009 |
break;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
else if (threads_l3)
|
|
Packit |
6c4009 |
threads = threads_l3;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
else
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
intel_bug_no_cache_info:
|
|
Packit |
6c4009 |
/* Assume that all logical threads share the highest cache
|
|
Packit |
6c4009 |
level. */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
threads
|
|
Packit |
6c4009 |
= ((cpu_features->cpuid[COMMON_CPUID_INDEX_1].ebx
|
|
Packit |
6c4009 |
>> 16) & 0xff);
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Cap usage of highest cache level to the number of supported
|
|
Packit |
6c4009 |
threads. */
|
|
Packit |
6c4009 |
if (shared > 0 && threads > 0)
|
|
Packit |
6c4009 |
shared /= threads;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Account for non-inclusive L2 and L3 caches. */
|
|
Packit |
6c4009 |
if (!inclusive_cache)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
if (threads_l2 > 0)
|
|
Packit |
6c4009 |
core /= threads_l2;
|
|
Packit |
6c4009 |
shared += core;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
else if (cpu_features->kind == arch_kind_amd)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
data = handle_amd (_SC_LEVEL1_DCACHE_SIZE);
|
|
Packit |
6c4009 |
long int core = handle_amd (_SC_LEVEL2_CACHE_SIZE);
|
|
Packit |
6c4009 |
shared = handle_amd (_SC_LEVEL3_CACHE_SIZE);
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Get maximum extended function. */
|
|
Packit |
6c4009 |
__cpuid (0x80000000, max_cpuid_ex, ebx, ecx, edx);
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
if (shared <= 0)
|
|
Packit |
6c4009 |
/* No shared L3 cache. All we have is the L2 cache. */
|
|
Packit |
6c4009 |
shared = core;
|
|
Packit |
6c4009 |
else
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
/* Figure out the number of logical threads that share L3. */
|
|
Packit |
6c4009 |
if (max_cpuid_ex >= 0x80000008)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
/* Get width of APIC ID. */
|
|
Packit |
6c4009 |
__cpuid (0x80000008, max_cpuid_ex, ebx, ecx, edx);
|
|
Packit |
6c4009 |
threads = 1 << ((ecx >> 12) & 0x0f);
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
if (threads == 0)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
/* If APIC ID width is not available, use logical
|
|
Packit |
6c4009 |
processor count. */
|
|
Packit |
6c4009 |
__cpuid (0x00000001, max_cpuid_ex, ebx, ecx, edx);
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
if ((edx & (1 << 28)) != 0)
|
|
Packit |
6c4009 |
threads = (ebx >> 16) & 0xff;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Cap usage of highest cache level to the number of
|
|
Packit |
6c4009 |
supported threads. */
|
|
Packit |
6c4009 |
if (threads > 0)
|
|
Packit |
6c4009 |
shared /= threads;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Account for exclusive L2 and L3 caches. */
|
|
Packit |
6c4009 |
shared += core;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#ifndef DISABLE_PREFETCHW
|
|
Packit |
6c4009 |
if (max_cpuid_ex >= 0x80000001)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
__cpuid (0x80000001, eax, ebx, ecx, edx);
|
|
Packit |
6c4009 |
/* PREFETCHW || 3DNow! */
|
|
Packit |
6c4009 |
if ((ecx & 0x100) || (edx & 0x80000000))
|
|
Packit |
6c4009 |
__x86_prefetchw = -1;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
#endif
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
if (cpu_features->data_cache_size != 0)
|
|
Packit |
6c4009 |
data = cpu_features->data_cache_size;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
if (data > 0)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
__x86_raw_data_cache_size_half = data / 2;
|
|
Packit |
6c4009 |
__x86_raw_data_cache_size = data;
|
|
Packit |
6c4009 |
/* Round data cache size to multiple of 256 bytes. */
|
|
Packit |
6c4009 |
data = data & ~255L;
|
|
Packit |
6c4009 |
__x86_data_cache_size_half = data / 2;
|
|
Packit |
6c4009 |
__x86_data_cache_size = data;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
if (cpu_features->shared_cache_size != 0)
|
|
Packit |
6c4009 |
shared = cpu_features->shared_cache_size;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
if (shared > 0)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
__x86_raw_shared_cache_size_half = shared / 2;
|
|
Packit |
6c4009 |
__x86_raw_shared_cache_size = shared;
|
|
Packit |
6c4009 |
/* Round shared cache size to multiple of 256 bytes. */
|
|
Packit |
6c4009 |
shared = shared & ~255L;
|
|
Packit |
6c4009 |
__x86_shared_cache_size_half = shared / 2;
|
|
Packit |
6c4009 |
__x86_shared_cache_size = shared;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* The large memcpy micro benchmark in glibc shows that 6 times of
|
|
Packit |
6c4009 |
shared cache size is the approximate value above which non-temporal
|
|
Packit |
6c4009 |
store becomes faster on a 8-core processor. This is the 3/4 of the
|
|
Packit |
6c4009 |
total shared cache size. */
|
|
Packit |
6c4009 |
__x86_shared_non_temporal_threshold
|
|
Packit |
6c4009 |
= (cpu_features->non_temporal_threshold != 0
|
|
Packit |
6c4009 |
? cpu_features->non_temporal_threshold
|
|
Packit |
6c4009 |
: __x86_shared_cache_size * threads * 3 / 4);
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#endif
|