arne
/
cryptopp
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Fix GCC optimizing away feature tests

pull/157/merge
Jeffrey Walton 2016-05-30 03:08:53 -04:00
parent be572877f6
commit 89ec42b1f2
1 changed files with 99 additions and 55 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

142
cpu.cpp
View File

@ -48,19 +48,19 @@ bool CpuId(word32 input, word32 output[4])
#else #else
#ifndef CRYPTOPP_MS_STYLE_INLINE_ASSEMBLY #ifndef CRYPTOPP_MS_STYLE_INLINE_ASSEMBLY
extern "C" { extern "C"
static jmp_buf s_jmpNoCPUID;
static void SigIllHandlerCPUID(int)
{ {
static jmp_buf s_jmpNoCPUID;
static void SigIllHandlerCPUID(int)
{
longjmp(s_jmpNoCPUID, 1); longjmp(s_jmpNoCPUID, 1);
} }
static jmp_buf s_jmpNoSSE2; static jmp_buf s_jmpNoSSE2;
static void SigIllHandlerSSE2(int) static void SigIllHandlerSSE2(int)
{ {
longjmp(s_jmpNoSSE2, 1); longjmp(s_jmpNoSSE2, 1);
} }
} }
#endif #endif
@ -97,10 +97,14 @@ bool CpuId(word32 input, word32 output[4])
// http://github.com/weidai11/cryptopp/issues/24 and http://stackoverflow.com/q/7721854 // http://github.com/weidai11/cryptopp/issues/24 and http://stackoverflow.com/q/7721854
volatile bool result = true; volatile bool result = true;
SigHandler oldHandler = signal(SIGILL, SigIllHandlerCPUID); volatile SigHandler oldHandler = signal(SIGILL, SigIllHandlerCPUID);
if (oldHandler == SIG_ERR) if (oldHandler == SIG_ERR)
result = false; result = false;
volatile sigset_t oldMask;
if (sigprocmask(0, NULL, (sigset_t*)&oldMask))
result = false;
if (setjmp(s_jmpNoCPUID)) if (setjmp(s_jmpNoCPUID))
result = false; result = false;
else else
@ -119,6 +123,7 @@ bool CpuId(word32 input, word32 output[4])
); );
} }
sigprocmask(SIG_SETMASK, (sigset_t*)&oldMask, NULL);
signal(SIGILL, oldHandler); signal(SIGILL, oldHandler);
return result; return result;
#endif #endif
@ -151,7 +156,7 @@ static bool TrySSE2()
// http://github.com/weidai11/cryptopp/issues/24 and http://stackoverflow.com/q/7721854 // http://github.com/weidai11/cryptopp/issues/24 and http://stackoverflow.com/q/7721854
volatile bool result = true; volatile bool result = true;
SigHandler oldHandler = signal(SIGILL, SigIllHandlerSSE2); volatile SigHandler oldHandler = signal(SIGILL, SigIllHandlerSSE2);
if (oldHandler == SIG_ERR) if (oldHandler == SIG_ERR)
return false; return false;
@ -167,6 +172,7 @@ static bool TrySSE2()
#endif #endif
} }
sigprocmask(SIG_SETMASK, (sigset_t*)&oldMask, NULL);
signal(SIGILL, oldHandler); signal(SIGILL, oldHandler);
return result; return result;
#endif #endif
@ -298,16 +304,11 @@ void DetectX86Features()
*((volatile bool*)&g_x86DetectionDone) = true; *((volatile bool*)&g_x86DetectionDone) = true;
} }
// http://community.arm.com/groups/android-community/blog/2014/10/10/runtime-detection-of-cpu-features-on-an-armv8-a-cpu
// http://stackoverflow.com/questions/26701262/how-to-check-the-existence-of-neon-on-arm
#elif (CRYPTOPP_BOOL_ARM32 || CRYPTOPP_BOOL_ARM64) #elif (CRYPTOPP_BOOL_ARM32 || CRYPTOPP_BOOL_ARM64)
bool g_ArmDetectionDone = false; // The ARM equivalent of CPUID probing is reading a MSR. The code requires Exception Level 1 (EL1) and above, but user space runs at EL0.
bool g_hasNEON = false, g_hasCRC32 = false, g_hasAES = false, g_hasSHA1 = false, g_hasSHA2 = false; // Attempting to run the code results in a SIGILL and termination.
//
word32 g_cacheLineSize = CRYPTOPP_L1_CACHE_LINE_SIZE;
// The ARM equivalent of CPUID is reading a MSR. For example, fetch crypto capabilities with:
// #if defined(__arm64__) || defined(__aarch64__) // #if defined(__arm64__) || defined(__aarch64__)
// word64 caps = 0; // Read ID_AA64ISAR0_EL1 // word64 caps = 0; // Read ID_AA64ISAR0_EL1
// __asm __volatile("mrs %0, " "id_aa64isar0_el1" : "=r" (caps)); // __asm __volatile("mrs %0, " "id_aa64isar0_el1" : "=r" (caps));
@ -315,12 +316,18 @@ word32 g_cacheLineSize = CRYPTOPP_L1_CACHE_LINE_SIZE;
// word32 caps = 0; // Read ID_ISAR5_EL1 // word32 caps = 0; // Read ID_ISAR5_EL1
// __asm __volatile("mrs %0, " "id_isar5_el1" : "=r" (caps)); // __asm __volatile("mrs %0, " "id_isar5_el1" : "=r" (caps));
// #endif // #endif
// The code requires Exception Level 1 (EL1) and above, but user space runs at EL0. //
// Attempting to run the code results in a SIGILL and termination. // The following does not work well either. Its appears to be missing constants, and it does not detect Aarch32 execution environments on Aarch64
// http://community.arm.com/groups/android-community/blog/2014/10/10/runtime-detection-of-cpu-features-on-an-armv8-a-cpu
//
bool g_ArmDetectionDone = false;
bool g_hasNEON = false, g_hasCRC32 = false, g_hasAES = false, g_hasSHA1 = false, g_hasSHA2 = false;
word32 g_cacheLineSize = CRYPTOPP_L1_CACHE_LINE_SIZE;
#ifndef CRYPTOPP_MS_STYLE_INLINE_ASSEMBLY #ifndef CRYPTOPP_MS_STYLE_INLINE_ASSEMBLY
extern "C" { extern "C"
{
static jmp_buf s_jmpNoNEON; static jmp_buf s_jmpNoNEON;
static void SigIllHandlerNEON(int) static void SigIllHandlerNEON(int)
{ {
@ -359,15 +366,15 @@ static bool TryNEON()
# if defined(CRYPTOPP_MS_STYLE_INLINE_ASSEMBLY) # if defined(CRYPTOPP_MS_STYLE_INLINE_ASSEMBLY)
__try __try
{ {
static const uint32_t v1[4] = {1,1,1,1}; uint32_t v1[4] = {1,1,1,1};
uint32x4_t x1 = vld1q_u32(v1); uint32x4_t x1 = vld1q_u32(v1);
static const uint64_t v2[2] = {1,1}; uint64_t v2[2] = {1,1};
uint64x2_t x2 = vld1q_u64(v2); uint64x2_t x2 = vld1q_u64(v2);
uint32x4_t x3 = vdupq_n_u32(0); uint32x4_t x3 = {0,0,0,0};
x3 = vsetq_lane_u32(vgetq_lane_u32(x1,0),x3,0); x3 = vsetq_lane_u32(vgetq_lane_u32(x1,0),x3,0);
x3 = vsetq_lane_u32(vgetq_lane_u32(x1,3),x3,3); x3 = vsetq_lane_u32(vgetq_lane_u32(x1,3),x3,3);
uint64x2_t x4 = vdupq_n_u64(0); uint64x2_t x4 = {0,0};
x4 = vsetq_lane_u64(vgetq_lane_u64(x2,0),x4,0); x4 = vsetq_lane_u64(vgetq_lane_u64(x2,0),x4,0);
x4 = vsetq_lane_u64(vgetq_lane_u64(x2,1),x4,1); x4 = vsetq_lane_u64(vgetq_lane_u64(x2,1),x4,1);
} }
@ -381,27 +388,35 @@ static bool TryNEON()
// http://github.com/weidai11/cryptopp/issues/24 and http://stackoverflow.com/q/7721854 // http://github.com/weidai11/cryptopp/issues/24 and http://stackoverflow.com/q/7721854
volatile bool result = true; volatile bool result = true;
SigHandler oldHandler = signal(SIGILL, SigIllHandlerNEON); volatile SigHandler oldHandler = signal(SIGILL, SigIllHandlerNEON);
if (oldHandler == SIG_ERR) if (oldHandler == SIG_ERR)
result = false; result = false;
volatile sigset_t oldMask;
if (sigprocmask(0, NULL, (sigset_t*)&oldMask))
result = false;
if (setjmp(s_jmpNoNEON)) if (setjmp(s_jmpNoNEON))
result = false; result = false;
else else
{ {
static const uint32_t v1[4] = {1,1,1,1}; uint32_t v1[4] = {1,1,1,1};
uint32x4_t x1 = vld1q_u32(v1); uint32x4_t x1 = vld1q_u32(v1);
static const uint64_t v2[2] = {1,1}; uint64_t v2[2] = {1,1};
uint64x2_t x2 = vld1q_u64(v2); uint64x2_t x2 = vld1q_u64(v2);
uint32x4_t x3 = vdupq_n_u32(0); uint32x4_t x3 = {0,0,0,0};
x3 = vsetq_lane_u32(vgetq_lane_u32(x1,0),x3,0); x3 = vsetq_lane_u32(vgetq_lane_u32(x1,0),x3,0);
x3 = vsetq_lane_u32(vgetq_lane_u32(x1,3),x3,3); x3 = vsetq_lane_u32(vgetq_lane_u32(x1,3),x3,3);
uint64x2_t x4 = vdupq_n_u64(0); uint64x2_t x4 = {0,0};
x4 = vsetq_lane_u64(vgetq_lane_u64(x2,0),x4,0); x4 = vsetq_lane_u64(vgetq_lane_u64(x2,0),x4,0);
x4 = vsetq_lane_u64(vgetq_lane_u64(x2,1),x4,1); x4 = vsetq_lane_u64(vgetq_lane_u64(x2,1),x4,1);
// Hack... GCC optimizes away the code and returns true
result = !!(vgetq_lane_u32(x3,0) | vgetq_lane_u64(x4,1));
} }
sigprocmask(SIG_SETMASK, (sigset_t*)&oldMask, NULL);
signal(SIGILL, oldHandler); signal(SIGILL, oldHandler);
return result; return result;
# endif # endif
@ -431,10 +446,14 @@ static bool TryCRC32()
// http://github.com/weidai11/cryptopp/issues/24 and http://stackoverflow.com/q/7721854 // http://github.com/weidai11/cryptopp/issues/24 and http://stackoverflow.com/q/7721854
volatile bool result = true; volatile bool result = true;
SigHandler oldHandler = signal(SIGILL, SigIllHandlerCRC32); volatile SigHandler oldHandler = signal(SIGILL, SigIllHandlerCRC32);
if (oldHandler == SIG_ERR) if (oldHandler == SIG_ERR)
result = false; result = false;
volatile sigset_t oldMask;
if (sigprocmask(0, NULL, (sigset_t*)&oldMask))
result = false;
if (setjmp(s_jmpNoCRC32)) if (setjmp(s_jmpNoCRC32))
result = false; result = false;
else else
@ -443,8 +462,12 @@ static bool TryCRC32()
w = __crc32cw(w,x); w = __crc32cw(w,x);
w = __crc32ch(w,y); w = __crc32ch(w,y);
w = __crc32cb(w,z); w = __crc32cb(w,z);
// Hack... GCC optimizes away the code and returns true
result = !!w;
} }
sigprocmask(SIG_SETMASK, (sigset_t*)&oldMask, NULL);
signal(SIGILL, oldHandler); signal(SIGILL, oldHandler);
return result; return result;
# endif # endif
@ -460,7 +483,7 @@ static bool TryAES()
__try __try
{ {
// AES encrypt and decrypt // AES encrypt and decrypt
static const uint8x16_t data = vdupq_n_u8(0), key = vdupq_n_u8(0); uint8x16_t data = vdupq_n_u8(0), key = vdupq_n_u8(0);
uint8x16_t r1 = vaeseq_u8(data, key); uint8x16_t r1 = vaeseq_u8(data, key);
uint8x16_t r2 = vaesdq_u8(data, key); uint8x16_t r2 = vaesdq_u8(data, key);
CRYPTOPP_UNUSED(r1), CRYPTOPP_UNUSED(r2); CRYPTOPP_UNUSED(r1), CRYPTOPP_UNUSED(r2);
@ -475,20 +498,27 @@ static bool TryAES()
// http://github.com/weidai11/cryptopp/issues/24 and http://stackoverflow.com/q/7721854 // http://github.com/weidai11/cryptopp/issues/24 and http://stackoverflow.com/q/7721854
volatile bool result = true; volatile bool result = true;
SigHandler oldHandler = signal(SIGILL, SigIllHandlerAES); volatile SigHandler oldHandler = signal(SIGILL, SigIllHandlerAES);
if (oldHandler == SIG_ERR) if (oldHandler == SIG_ERR)
result = false; result = false;
volatile sigset_t oldMask;
if (sigprocmask(0, NULL, (sigset_t*)&oldMask))
result = false;
if (setjmp(s_jmpNoAES)) if (setjmp(s_jmpNoAES))
result = false; result = false;
else else
{ {
static const uint8x16_t data = vdupq_n_u8(0), key = vdupq_n_u8(0); uint8x16_t data = vdupq_n_u8(0), key = vdupq_n_u8(0);
uint8x16_t r1 = vaeseq_u8(data, key); uint8x16_t r1 = vaeseq_u8(data, key);
uint8x16_t r2 = vaesdq_u8(data, key); uint8x16_t r2 = vaesdq_u8(data, key);
CRYPTOPP_UNUSED(r1), CRYPTOPP_UNUSED(r2);
// Hack... GCC optimizes away the code and returns true
result = !!(vgetq_lane_u8(r1,0) | vgetq_lane_u8(r2,7));
} }
sigprocmask(SIG_SETMASK, (sigset_t*)&oldMask, NULL);
signal(SIGILL, oldHandler); signal(SIGILL, oldHandler);
return result; return result;
# endif # endif
@ -503,8 +533,8 @@ static bool TrySHA1()
# if defined(CRYPTOPP_MS_STYLE_INLINE_ASSEMBLY) # if defined(CRYPTOPP_MS_STYLE_INLINE_ASSEMBLY)
__try __try
{ {
static const uint32x4_t data = vdupq_n_u32(0); uint32x4_t data = {0,0,0,0};
static const uint32_t hash = 0x0; uint32_t hash = 0x0;
uint32x4_t r1 = vsha1cq_u32 (data, hash, data); uint32x4_t r1 = vsha1cq_u32 (data, hash, data);
uint32x4_t r2 = vsha1mq_u32 (data, hash, data); uint32x4_t r2 = vsha1mq_u32 (data, hash, data);
@ -521,23 +551,30 @@ static bool TrySHA1()
// http://github.com/weidai11/cryptopp/issues/24 and http://stackoverflow.com/q/7721854 // http://github.com/weidai11/cryptopp/issues/24 and http://stackoverflow.com/q/7721854
volatile bool result = true; volatile bool result = true;
SigHandler oldHandler = signal(SIGILL, SigIllHandlerSHA1); volatile SigHandler oldHandler = signal(SIGILL, SigIllHandlerSHA1);
if (oldHandler == SIG_ERR) if (oldHandler == SIG_ERR)
result = false; result = false;
volatile sigset_t oldMask;
if (sigprocmask(0, NULL, (sigset_t*)&oldMask))
result = false;
if (setjmp(s_jmpNoSHA1)) if (setjmp(s_jmpNoSHA1))
result = false; result = false;
else else
{ {
static const uint32x4_t data = vdupq_n_u32(0); uint32x4_t data = {0,0,0,0};
static const uint32_t hash = 0x0; uint32_t hash = 0x0;
uint32x4_t r1 = vsha1cq_u32 (data, hash, data); uint32x4_t r1 = vsha1cq_u32 (data, hash, data);
uint32x4_t r2 = vsha1mq_u32 (data, hash, data); uint32x4_t r2 = vsha1mq_u32 (data, hash, data);
uint32x4_t r3 = vsha1pq_u32 (data, hash, data); uint32x4_t r3 = vsha1pq_u32 (data, hash, data);
CRYPTOPP_UNUSED(r1), CRYPTOPP_UNUSED(r2), CRYPTOPP_UNUSED(r3);
// Hack... GCC optimizes away the code and returns true
result = !!(vgetq_lane_u32(r1,0) | vgetq_lane_u32(r2,1) | vgetq_lane_u32(r3,2));
} }
sigprocmask(SIG_SETMASK, (sigset_t*)&oldMask, NULL);
signal(SIGILL, oldHandler); signal(SIGILL, oldHandler);
return result; return result;
# endif # endif
@ -552,8 +589,8 @@ static bool TrySHA2()
# if defined(CRYPTOPP_MS_STYLE_INLINE_ASSEMBLY) # if defined(CRYPTOPP_MS_STYLE_INLINE_ASSEMBLY)
__try __try
{ {
static const uint32x4_t data = vdupq_n_u32(0); uint32x4_t data = {0,0,0,0};
static const uint32x4_t hash = vdupq_n_u32(0); uint32x4_t hash = {0,0,0,0};
uint32x4_t r1 = vsha256hq_u32 (hash, hash, data); uint32x4_t r1 = vsha256hq_u32 (hash, hash, data);
uint32x4_t r2 = vsha256h2q_u32 (hash, hash, data); uint32x4_t r2 = vsha256h2q_u32 (hash, hash, data);
@ -571,24 +608,31 @@ static bool TrySHA2()
// http://github.com/weidai11/cryptopp/issues/24 and http://stackoverflow.com/q/7721854 // http://github.com/weidai11/cryptopp/issues/24 and http://stackoverflow.com/q/7721854
volatile bool result = true; volatile bool result = true;
SigHandler oldHandler = signal(SIGILL, SigIllHandlerSHA2); volatile SigHandler oldHandler = signal(SIGILL, SigIllHandlerSHA2);
if (oldHandler == SIG_ERR) if (oldHandler == SIG_ERR)
result = false; result = false;
volatile sigset_t oldMask;
if (sigprocmask(0, NULL, (sigset_t*)&oldMask))
result = false;
if (setjmp(s_jmpNoSHA2)) if (setjmp(s_jmpNoSHA2))
result = false; result = false;
else else
{ {
static const uint32x4_t data = vdupq_n_u32(0); uint32x4_t data = {0,0,0,0};
static const uint32x4_t hash = vdupq_n_u32(0); uint32x4_t hash = {0,0,0,0};
uint32x4_t r1 = vsha256hq_u32 (hash, hash, data); uint32x4_t r1 = vsha256hq_u32 (hash, hash, data);
uint32x4_t r2 = vsha256h2q_u32 (hash, hash, data); uint32x4_t r2 = vsha256h2q_u32 (hash, hash, data);
uint32x4_t r3 = vsha256su0q_u32 (data, data); uint32x4_t r3 = vsha256su0q_u32 (data, data);
uint32x4_t r4 = vsha256su1q_u32 (data, data, data); uint32x4_t r4 = vsha256su1q_u32 (data, data, data);
CRYPTOPP_UNUSED(r1), CRYPTOPP_UNUSED(r2), CRYPTOPP_UNUSED(r3), CRYPTOPP_UNUSED(r4);
// Hack... GCC optimizes away the code and returns true
result = !!(vgetq_lane_u32(r1,0) | vgetq_lane_u32(r2,1) | vgetq_lane_u32(r3,2) | vgetq_lane_u32(r4,3));
} }
sigprocmask(SIG_SETMASK, (sigset_t*)&oldMask, NULL);
signal(SIGILL, oldHandler); signal(SIGILL, oldHandler);
return result; return result;
# endif # endif