1 
2 // crc_simd.cpp - written and placed in the public domain by
3 //                Jeffrey Walton, Uri Blumenthal and Marcel Raad.
4 //
5 //    This source file uses intrinsics to gain access to ARMv7a and
6 //    ARMv8a NEON instructions. A separate source file is needed
7 //    because additional CXXFLAGS are required to enable the
8 //    appropriate instructions sets in some build configurations.
9 //    For Linux and Unix additional flags are not required.
10 
11 #include "pch.h"
12 #include "config.h"
13 #include "stdcpp.h"
14 
15 #if (CRYPTOPP_ARM_NEON_HEADER)
16 # include <arm_neon.h>
17 #endif
18 
19 #if (CRYPTOPP_ARM_ACLE_HEADER)
20 # include <stdint.h>
21 # include <arm_acle.h>
22 #endif
23 
24 #ifdef CRYPTOPP_GNU_STYLE_INLINE_ASSEMBLY
25 # include <signal.h>
26 # include <setjmp.h>
27 #endif
28 
29 #ifndef EXCEPTION_EXECUTE_HANDLER
30 # define EXCEPTION_EXECUTE_HANDLER 1
31 #endif
32 
33 // Squash MS LNK4221 and libtool warnings
34 extern const char NEON_SIMD_FNAME[] = __FILE__;
35 
NAMESPACE_BEGIN(CryptoPP)36 NAMESPACE_BEGIN(CryptoPP)
37 
38 #ifdef CRYPTOPP_GNU_STYLE_INLINE_ASSEMBLY
39 extern "C" {
40     typedef void (*SigHandler)(int);
41 
42     static jmp_buf s_jmpSIGILL;
43     static void SigIllHandler(int)
44     {
45         longjmp(s_jmpSIGILL, 1);
46     }
47 }
48 #endif  // Not CRYPTOPP_MS_STYLE_INLINE_ASSEMBLY
49 
CPU_ProbeARMv7()50 bool CPU_ProbeARMv7()
51 {
52 #if defined(__aarch32__) || defined(__aarch64__) || defined(_M_ARM64)
53     return false;
54 #elif defined(CRYPTOPP_NO_CPU_FEATURE_PROBES)
55     return false;
56 #elif CRYPTOPP_ARM_NEON_AVAILABLE
57 # if defined(CRYPTOPP_MS_STYLE_INLINE_ASSEMBLY)
58     volatile bool result = true;
59     __try
60     {
61         // Modern MS hardware is ARMv7
62         result = true;
63     }
64     __except (EXCEPTION_EXECUTE_HANDLER)
65     {
66         return false;
67     }
68     return result;
69 # else
70     // longjmp and clobber warnings. Volatile is required.
71     // http://github.com/weidai11/cryptopp/issues/24 and http://stackoverflow.com/q/7721854
72     volatile bool result = true;
73 
74     volatile SigHandler oldHandler = signal(SIGILL, SigIllHandler);
75     if (oldHandler == SIG_ERR)
76         return false;
77 
78     volatile sigset_t oldMask;
79     if (sigprocmask(0, NULLPTR, (sigset_t*)&oldMask))
80     {
81         signal(SIGILL, oldHandler);
82         return false;
83     }
84 
85     if (setjmp(s_jmpSIGILL))
86         result = false;
87     else
88     {
89 
90 #if 0
91         // ARMv7 added movt and movw
92         int a;
93         asm volatile("movw %0,%1 \n"
94                      "movt %0,%1 \n"
95                      : "=r"(a) : "i"(0x1234));
96 
97 00000010 <_Z5test2v>:  // ARM
98   10:   e3010234        movw    r0, #4660       ; 0x1234
99   14:   e3410234        movt    r0, #4660       ; 0x1234
100   18:   e12fff1e        bx      lr
101 
102 0000001c <_Z5test3v>:  // Thumb
103   1c:   f241 2034       movw    r0, #4660       ; 0x1234
104   20:   f2c1 2034       movt    r0, #4660       ; 0x1234
105   24:   e12fff1e        bx      lr
106 #endif
107 
108         unsigned int a;
109         asm volatile (
110 #if defined(__thumb__)
111             ".inst.n 0xf241, 0x2034  \n\t"   // movw r0, 0x1234
112             ".inst.n 0xf2c1, 0x2034  \n\t"   // movt r0, 0x1234
113             "mov %0, r0              \n\t"   // mov [a], r0
114 #else
115             ".inst 0xe3010234  \n\t"   // movw r0, 0x1234
116             ".inst 0xe3410234  \n\t"   // movt r0, 0x1234
117             "mov %0, r0        \n\t"   // mov [a], r0
118 #endif
119             : "=r" (a) : : "r0");
120 
121         result = (a == 0x12341234);
122     }
123 
124     sigprocmask(SIG_SETMASK, (sigset_t*)&oldMask, NULLPTR);
125     signal(SIGILL, oldHandler);
126     return result;
127 # endif
128 #else
129     return false;
130 #endif  // CRYPTOPP_ARM_NEON_AVAILABLE
131 }
132 
CPU_ProbeNEON()133 bool CPU_ProbeNEON()
134 {
135 #if defined(__aarch32__) || defined(__aarch64__) || defined(_M_ARM64)
136     return true;
137 #elif defined(CRYPTOPP_NO_CPU_FEATURE_PROBES)
138     return false;
139 #elif CRYPTOPP_ARM_NEON_AVAILABLE
140 # if defined(CRYPTOPP_MS_STYLE_INLINE_ASSEMBLY)
141     volatile bool result = true;
142     __try
143     {
144         uint32x4_t x = vdupq_n_u32(1);
145         uint32x4_t y = vshlq_n_u32(x, 4);
146 
147         word32 z[4]; vst1q_u32(z, y);
148         return (z[0] & z[1] & z[2] & z[3]) == 16;
149     }
150     __except (EXCEPTION_EXECUTE_HANDLER)
151     {
152         return false;
153     }
154     return result;
155 # else
156     // longjmp and clobber warnings. Volatile is required.
157     // http://github.com/weidai11/cryptopp/issues/24 and http://stackoverflow.com/q/7721854
158     volatile bool result = true;
159 
160     volatile SigHandler oldHandler = signal(SIGILL, SigIllHandler);
161     if (oldHandler == SIG_ERR)
162         return false;
163 
164     volatile sigset_t oldMask;
165     if (sigprocmask(0, NULLPTR, (sigset_t*)&oldMask))
166     {
167         signal(SIGILL, oldHandler);
168         return false;
169     }
170 
171     if (setjmp(s_jmpSIGILL))
172         result = false;
173     else
174     {
175         // This is risky... When we hand encode the instructions
176         // for vmov.u32 and vshl.u32 we get a SIGILL. Apparently
177         // we need more than just the instructions. Using
178         // intrinsics introduces the risk because the whole
179         // file gets built with ISA options, and the higher ISA
180         // may escape the try block with the SIGILL guard.
181         uint32x4_t x = vdupq_n_u32(1);
182         uint32x4_t y = vshlq_n_u32(x, 4);
183 
184         word32 z[4]; vst1q_u32(z, y);
185         result = (z[0] & z[1] & z[2] & z[3]) == 16;
186     }
187 
188     sigprocmask(SIG_SETMASK, (sigset_t*)&oldMask, NULLPTR);
189     signal(SIGILL, oldHandler);
190     return result;
191 # endif
192 #else
193     return false;
194 #endif  // CRYPTOPP_ARM_NEON_AVAILABLE
195 }
196 
197 NAMESPACE_END
198