|
1 // Copyright (c) 2006-2008 The Chromium Authors. All rights reserved. |
|
2 // Use of this source code is governed by a BSD-style license that can be |
|
3 // found in the LICENSE file. |
|
4 |
|
5 // This module gets enough CPU information to optimize the |
|
6 // atomicops module on x86. |
|
7 |
|
8 #include <string.h> |
|
9 |
|
10 #include "base/atomicops.h" |
|
11 #include "base/basictypes.h" |
|
12 |
|
13 // This file only makes sense with atomicops_internals_x86_gcc.h -- it |
|
14 // depends on structs that are defined in that file. If atomicops.h |
|
15 // doesn't sub-include that file, then we aren't needed, and shouldn't |
|
16 // try to do anything. |
|
17 #ifdef BASE_ATOMICOPS_INTERNALS_X86_GCC_H_ |
|
18 |
|
19 // Inline cpuid instruction. In PIC compilations, %ebx contains the address |
|
20 // of the global offset table. To avoid breaking such executables, this code |
|
21 // must preserve that register's value across cpuid instructions. |
|
22 #if defined(__i386__) |
|
23 #define cpuid(a, b, c, d, inp) \ |
|
24 asm ("mov %%ebx, %%edi\n" \ |
|
25 "cpuid\n" \ |
|
26 "xchg %%edi, %%ebx\n" \ |
|
27 : "=a" (a), "=D" (b), "=c" (c), "=d" (d) : "a" (inp)) |
|
28 #elif defined (__x86_64__) |
|
29 #define cpuid(a, b, c, d, inp) \ |
|
30 asm ("mov %%rbx, %%rdi\n" \ |
|
31 "cpuid\n" \ |
|
32 "xchg %%rdi, %%rbx\n" \ |
|
33 : "=a" (a), "=D" (b), "=c" (c), "=d" (d) : "a" (inp)) |
|
34 #endif |
|
35 |
|
36 #if defined(cpuid) // initialize the struct only on x86 |
|
37 |
|
38 // Set the flags so that code will run correctly and conservatively, so even |
|
39 // if we haven't been initialized yet, we're probably single threaded, and our |
|
40 // default values should hopefully be pretty safe. |
|
41 struct AtomicOps_x86CPUFeatureStruct AtomicOps_Internalx86CPUFeatures = { |
|
42 false, // bug can't exist before process spawns multiple threads |
|
43 false, // no SSE2 |
|
44 }; |
|
45 |
|
46 // Initialize the AtomicOps_Internalx86CPUFeatures struct. |
|
47 static void AtomicOps_Internalx86CPUFeaturesInit() { |
|
48 uint32_t eax; |
|
49 uint32_t ebx; |
|
50 uint32_t ecx; |
|
51 uint32_t edx; |
|
52 |
|
53 // Get vendor string (issue CPUID with eax = 0) |
|
54 cpuid(eax, ebx, ecx, edx, 0); |
|
55 char vendor[13]; |
|
56 memcpy(vendor, &ebx, 4); |
|
57 memcpy(vendor + 4, &edx, 4); |
|
58 memcpy(vendor + 8, &ecx, 4); |
|
59 vendor[12] = 0; |
|
60 |
|
61 // get feature flags in ecx/edx, and family/model in eax |
|
62 cpuid(eax, ebx, ecx, edx, 1); |
|
63 |
|
64 int family = (eax >> 8) & 0xf; // family and model fields |
|
65 int model = (eax >> 4) & 0xf; |
|
66 if (family == 0xf) { // use extended family and model fields |
|
67 family += (eax >> 20) & 0xff; |
|
68 model += ((eax >> 16) & 0xf) << 4; |
|
69 } |
|
70 |
|
71 // Opteron Rev E has a bug in which on very rare occasions a locked |
|
72 // instruction doesn't act as a read-acquire barrier if followed by a |
|
73 // non-locked read-modify-write instruction. Rev F has this bug in |
|
74 // pre-release versions, but not in versions released to customers, |
|
75 // so we test only for Rev E, which is family 15, model 32..63 inclusive. |
|
76 if (strcmp(vendor, "AuthenticAMD") == 0 && // AMD |
|
77 family == 15 && |
|
78 32 <= model && model <= 63) { |
|
79 AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug = true; |
|
80 } else { |
|
81 AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug = false; |
|
82 } |
|
83 |
|
84 // edx bit 26 is SSE2 which we use to tell use whether we can use mfence |
|
85 AtomicOps_Internalx86CPUFeatures.has_sse2 = ((edx >> 26) & 1); |
|
86 } |
|
87 |
|
88 namespace { |
|
89 |
|
90 class AtomicOpsx86Initializer { |
|
91 public: |
|
92 AtomicOpsx86Initializer() { |
|
93 AtomicOps_Internalx86CPUFeaturesInit(); |
|
94 } |
|
95 }; |
|
96 |
|
97 // A global to get use initialized on startup via static initialization :/ |
|
98 AtomicOpsx86Initializer g_initer; |
|
99 |
|
100 } // namespace |
|
101 |
|
102 #endif // if x86 |
|
103 |
|
104 #endif // ifdef BASE_ATOMICOPS_INTERNALS_X86_GCC_H_ |