diff options
Diffstat (limited to 'tests/benchmarks/corelib/tools/qstring/main.cpp')
| -rw-r--r-- | tests/benchmarks/corelib/tools/qstring/main.cpp | 1262 |
1 files changed, 1261 insertions, 1 deletions
diff --git a/tests/benchmarks/corelib/tools/qstring/main.cpp b/tests/benchmarks/corelib/tools/qstring/main.cpp index 12826eb..9616052 100644 --- a/tests/benchmarks/corelib/tools/qstring/main.cpp +++ b/tests/benchmarks/corelib/tools/qstring/main.cpp @@ -40,7 +40,7 @@ ****************************************************************************/ #include <QStringList> #include <QFile> -#include <qtest.h> +#include <QtTest/QtTest> #ifdef Q_OS_SYMBIAN // In Symbian OS test data is located in applications private dir @@ -48,12 +48,27 @@ #define SRCDIR "" #endif +#ifdef Q_OS_UNIX +#include <sys/mman.h> +#include <unistd.h> +#endif + +#include <private/qsimd_p.h> + +#include "data.h" + class tst_QString: public QObject { Q_OBJECT +public: + tst_QString(); private slots: void equals() const; void equals_data() const; + void equals2_data() const; + void equals2() const; + void ucstrncmp_data() const; + void ucstrncmp() const; void fromUtf8() const; }; @@ -67,6 +82,10 @@ void tst_QString::equals() const } } +tst_QString::tst_QString() +{ +} + void tst_QString::equals_data() const { static const struct { @@ -126,6 +145,1247 @@ void tst_QString::equals_data() const << QString::fromRawData(ptr + 1, 58) << QString::fromRawData(ptr + 3, 58); } +static bool equals2_memcmp_call(const ushort *p1, const ushort *p2, int len) +{ + return memcmp(p1, p2, len * 2) == 0; +} + +static bool equals2_bytewise(const ushort *p1, const ushort *p2, int len) +{ + if (p1 == p2 || !len) + return true; + uchar *b1 = (uchar *)p1; + uchar *b2 = (uchar *)p2; + len *= 2; + while (len--) + if (*b1++ != *b2++) + return false; + return true; +} + +static bool equals2_shortwise(const ushort *p1, const ushort *p2, int len) +{ + if (p1 == p2 || !len) + return true; +// for (register int counter; counter < len; ++counter) +// if (p1[counter] != p2[counter]) +// return false; + while (len--) { + if (p1[len] != p2[len]) + return false; + } + return true; +} + +static bool equals2_intwise(const ushort *p1, const ushort *p2, int length) +{ + if (p1 == p2 || !length) + return true; + register union { + const quint16 *w; + const quint32 *d; + quintptr value; + } sa, sb; + sa.w = p1; + sb.w = p2; + + // check alignment + if ((sa.value & 2) == (sb.value & 2)) { + // both addresses have the same alignment + if (sa.value & 2) { + // both addresses are not aligned to 4-bytes boundaries + // compare the first character + if (*sa.w != *sb.w) + return false; + --length; + ++sa.w; + ++sb.w; + + // now both addresses are 4-bytes aligned + } + + // both addresses are 4-bytes aligned + // do a fast 32-bit comparison + register const quint32 *e = sa.d + (length >> 1); + for ( ; sa.d != e; ++sa.d, ++sb.d) { + if (*sa.d != *sb.d) + return false; + } + + // do we have a tail? + return (length & 1) ? *sa.w == *sb.w : true; + } else { + // one of the addresses isn't 4-byte aligned but the other is + register const quint16 *e = sa.w + length; + for ( ; sa.w != e; ++sa.w, ++sb.w) { + if (*sa.w != *sb.w) + return false; + } + } + return true; +} + +static inline bool equals2_short_tail(const ushort *p1, const ushort *p2, int len) +{ + if (len) { + if (*p1 != *p2) + return false; + if (--len) { + if (p1[1] != p2[1]) + return false; + if (--len) { + if (p1[2] != p2[2]) + return false; + if (--len) { + if (p1[3] != p2[3]) + return false; + if (--len) { + if (p1[4] != p2[4]) + return false; + if (--len) { + if (p1[5] != p2[5]) + return false; + if (--len) { + if (p1[6] != p2[6]) + return false; + return p1[7] == p2[7]; + } + } + } + } + } + } + } + return true; +} + +//#pragma GCC optimize("no-unroll-loops") +#ifdef __SSE2__ +static bool equals2_sse2_aligned(const ushort *p1, const ushort *p2, int len) +{ + if (len >= 8) { + qptrdiff counter = 0; + while (len > 8) { + __m128i q1 = _mm_load_si128((__m128i *)(p1 + counter)); + __m128i q2 = _mm_load_si128((__m128i *)(p2 + counter)); + __m128i cmp = _mm_cmpeq_epi16(q1, q2); + if (ushort(_mm_movemask_epi8(cmp)) != ushort(0xffff)) + return false; + + len -= 8; + counter += 8; + } + p1 += counter; + p2 += counter; + } + + return equals2_short_tail(p1, p2, len); +} + +static bool __attribute__((optimize("no-unroll-loops"))) equals2_sse2(const ushort *p1, const ushort *p2, int len) +{ + if (p1 == p2 || !len) + return true; + + if (len >= 8) { + qptrdiff counter = 0; + while (len >= 8) { + __m128i q1 = _mm_loadu_si128((__m128i *)(p1 + counter)); + __m128i q2 = _mm_loadu_si128((__m128i *)(p2 + counter)); + __m128i cmp = _mm_cmpeq_epi16(q1, q2); + if (ushort(_mm_movemask_epi8(cmp)) != 0xffff) + return false; + + len -= 8; + counter += 8; + } + p1 += counter; + p2 += counter; + } + + return equals2_short_tail(p1, p2, len); +} + +//static bool equals2_sse2(const ushort *p1, const ushort *p2, int len) +//{ +// register int val1 = quintptr(p1) & 0xf; +// register int val2 = quintptr(p2) & 0xf; +// if (false && val1 + val2 == 0) +// return equals2_sse2_aligned(p1, p2, len); +// else +// return equals2_sse2_unaligned(p1, p2, len); +//} + +static bool equals2_sse2_aligning(const ushort *p1, const ushort *p2, int len) +{ + if (len < 8) + return equals2_short_tail(p1, p2, len); + + qptrdiff counter = 0; + + // which one is easier to align, p1 or p2 ? + register int val1 = quintptr(p1) & 0xf; + register int val2 = quintptr(p2) & 0xf; + if (val1 && val2) { +#if 0 + // we'll align the one which requires the least number of steps + if (val1 > val2) { + qSwap(p1, p2); + val1 = val2; + } + + // val1 contains the number of bytes past the 16-aligned mark + // we must read 16-val1 bytes to align + val1 = 16 - val1; + if (val1 & 0x2) { + if (*p1 != *p2) + return false; + --len; + ++counter; + } + while (val1 & 12) { + if (*(uint*)p1 != *(uint*)p2) + return false; + --len; + counter += 2; + val1 -= 4; + } +#else + // we'll align the one closest to the 16-byte mark + if (val1 > val2) { + qSwap(p1, p2); + val1 = val2; + } + + // we're reading val1 bytes too many + __m128i q2 = _mm_loadu_si128((__m128i *)(p2 - val1/2)); + __m128i cmp = _mm_cmpeq_epi16(*(__m128i *)(p1 - val1/2), q2); + if (short(_mm_movemask_epi8(cmp)) >> val1 != short(-1)) + return false; + + counter = 8 - val1/2; + len -= 8 - val1/2; +#endif + } else if (!val2) { + // p2 is already aligned + qSwap(p1, p2); + } + + // p1 is aligned + + while (len >= 8) { + __m128i q1 = _mm_load_si128((__m128i *)(p1 + counter)); + __m128i q2 = _mm_loadu_si128((__m128i *)(p2 + counter)); + __m128i cmp = _mm_cmpeq_epi16(q1, q2); + if (ushort(_mm_movemask_epi8(cmp)) != ushort(0xffff)) + return false; + + len -= 8; + counter += 8; + } + + // tail + return equals2_short_tail(p1 + counter, p2 + counter, len); +} + +#ifdef __SSE3__ +static bool __attribute__((optimize("no-unroll-loops"))) equals2_sse3(const ushort *p1, const ushort *p2, int len) +{ + if (p1 == p2 || !len) + return true; + + if (len >= 8) { + qptrdiff counter = 0; + while (len >= 8) { + __m128i q1 = _mm_lddqu_si128((__m128i *)(p1 + counter)); + __m128i q2 = _mm_lddqu_si128((__m128i *)(p2 + counter)); + __m128i cmp = _mm_cmpeq_epi16(q1, q2); + if (ushort(_mm_movemask_epi8(cmp)) != 0xffff) + return false; + + len -= 8; + counter += 8; + } + p1 += counter; + p2 += counter; + } + + return equals2_short_tail(p1, p2, len); +} + +#ifdef __SSSE3__ +template<int N> static __attribute__((optimize("unroll-loops"))) inline bool equals2_ssse3_alignr(__m128i *m1, __m128i *m2, int len) +{ + __m128i lower = _mm_load_si128(m1); + while (len >= 8) { + __m128i upper = _mm_load_si128(m1 + 1); + __m128i correct; + correct = _mm_alignr_epi8(upper, lower, N); + + __m128i q2 = _mm_lddqu_si128(m2); + __m128i cmp = _mm_cmpeq_epi16(correct, q2); + if (ushort(_mm_movemask_epi8(cmp)) != 0xffff) + return false; + + len -= 8; + ++m2; + ++m1; + lower = upper; + } + + // tail + return len == 0 || equals2_short_tail((const ushort *)m1 + N / 2, (const ushort*)m2, len); +} + +static inline __attribute__((optimize("unroll-loops"))) bool equals2_ssse3_aligned(__m128i *m1, __m128i *m2, int len) +{ + while (len >= 8) { + __m128i q2 = _mm_lddqu_si128(m2); + __m128i cmp = _mm_cmpeq_epi16(*m1, q2); + if (ushort(_mm_movemask_epi8(cmp)) != 0xffff) + return false; + + len -= 8; + ++m1; + ++m2; + } + return len == 0 || equals2_short_tail((const ushort *)m1, (const ushort *)m2, len); +} + +static bool equals2_ssse3(const ushort *p1, const ushort *p2, int len) +{ + // p1 & 0xf can be: + // 0, 2, 4, 6, 8, 10, 12, 14 + // If it's 0, we're aligned + // If it's not, then we're interested in the 16 - (p1 & 0xf) bytes only + + if (len >= 8) { + // find the last aligned position below the p1 memory + __m128i *m1 = (__m128i *)(quintptr(p1) & ~0xf); + __m128i *m2 = (__m128i *)p2; + qptrdiff diff = quintptr(p1) - quintptr(m1); + + // diff contains the number of extra bytes + if (diff == 10) + return equals2_ssse3_alignr<10>(m1, m2, len); + else if (diff == 2) + return equals2_ssse3_alignr<2>(m1, m2, len); + if (diff < 8) { + if (diff < 4) { + return equals2_ssse3_aligned(m1, m2, len); + } else { + if (diff == 4) + return equals2_ssse3_alignr<4>(m1, m2, len); + else // diff == 6 + return equals2_ssse3_alignr<6>(m1, m2, len); + } + } else { + if (diff < 12) { + return equals2_ssse3_alignr<8>(m1, m2, len); + } else { + if (diff == 12) + return equals2_ssse3_alignr<12>(m1, m2, len); + else // diff == 14 + return equals2_ssse3_alignr<14>(m1, m2, len); + } + } + } + + // tail + return equals2_short_tail(p1, p2, len); +} + +template<int N> static inline bool equals2_ssse3_aligning_alignr(__m128i *m1, __m128i *m2, int len) +{ + __m128i lower = _mm_load_si128(m1); + while (len >= 8) { + __m128i upper = _mm_load_si128(m1 + 1); + __m128i correct; + correct = _mm_alignr_epi8(upper, lower, N); + + __m128i cmp = _mm_cmpeq_epi16(correct, *m2); + if (ushort(_mm_movemask_epi8(cmp)) != 0xffff) + return false; + + len -= 8; + ++m2; + ++m1; + lower = upper; + } + + // tail + return len == 0 || equals2_short_tail((const ushort *)m1 + N / 2, (const ushort*)m2, len); +} + +static bool equals2_ssse3_aligning(const ushort *p1, const ushort *p2, int len) +{ + if (len < 8) + return equals2_short_tail(p1, p2, len); + qptrdiff counter = 0; + + // which one is easier to align, p1 or p2 ? + { + register int val1 = quintptr(p1) & 0xf; + register int val2 = quintptr(p2) & 0xf; + if (val1 && val2) { + // we'll align the one closest to the 16-byte mark + if (val1 < val2) { + qSwap(p1, p2); + val2 = val1; + } + + // we're reading val1 bytes too many + __m128i q1 = _mm_lddqu_si128((__m128i *)(p1 - val2/2)); + __m128i cmp = _mm_cmpeq_epi16(q1, *(__m128i *)(p2 - val2/2)); + if (short(_mm_movemask_epi8(cmp)) >> val1 != short(-1)) + return false; + + counter = 8 - val2/2; + len -= 8 - val2/2; + } else if (!val1) { + // p1 is already aligned + qSwap(p1, p2); + } + } + + // p2 is aligned now + // we want to use palignr in the mis-alignment of p1 + __m128i *m1 = (__m128i *)(quintptr(p1 + counter) & ~0xf); + __m128i *m2 = (__m128i *)(p2 + counter); + register int val1 = quintptr(p1 + counter) - quintptr(m1); + + // val1 contains the number of extra bytes + if (val1 == 8) + return equals2_ssse3_aligning_alignr<8>(m1, m2, len); + if (val1 == 0) + return equals2_sse2_aligned(p1 + counter, p2 + counter, len); + if (val1 < 8) { + if (val1 < 4) { + return equals2_ssse3_aligning_alignr<2>(m1, m2, len); + } else { + if (val1 == 4) + return equals2_ssse3_aligning_alignr<4>(m1, m2, len); + else // diff == 6 + return equals2_ssse3_aligning_alignr<6>(m1, m2, len); + } + } else { + if (val1 < 12) { + return equals2_ssse3_aligning_alignr<10>(m1, m2, len); + } else { + if (val1 == 12) + return equals2_ssse3_aligning_alignr<12>(m1, m2, len); + else // diff == 14 + return equals2_ssse3_aligning_alignr<14>(m1, m2, len); + } + } +} + +#ifdef __SSE4_1__ +static bool equals2_sse4(const ushort *p1, const ushort *p2, int len) +{ + // We use the pcmpestrm instruction searching for differences (negative polarity) + // it will reset CF if it's all equal + // it will reset OF if the first char is equal + // it will set ZF & SF if the length is less than 8 (which means we've done the last operation) + // the three possible conditions are: + // difference found: CF = 1 + // all equal, not finished: CF = ZF = SF = 0 + // all equal, finished: CF = 0, ZF = SF = 1 + // We use the JA instruction that jumps if ZF = 0 and CF = 0 + if (p1 == p2 || !len) + return true; + + // This function may read some bytes past the end of p1 or p2 + // It is safe to do that, as long as those extra bytes (beyond p1+len and p2+len) + // are on the same page as the last valid byte. + // If len is a multiple of 8, we'll never load invalid bytes. + if (len & 7) { + // The last load would load (len & ~7) valid bytes and (8 - (len & ~7)) invalid bytes. + // So we can't do the last load if any of those bytes is in a different + // page. That is, if: + // pX + len is on a different page from pX + (len & ~7) + 8 + // + // that is, if second-to-last load ended up less than 16 bytes from the page end: + // pX + (len & ~7) is the last ushort read in the second-to-last load + if (len < 8) + return equals2_short_tail(p1, p2, len); + if ((quintptr(p1 + (len & ~7)) & 0xfff) > 0xff0 || + (quintptr(p2 + (len & ~7)) & 0xfff) > 0xff0) { + + // yes, so we mustn't do the final 128-bit load + bool result; + asm ( + "sub %[p1], %[p2]\n\t" + "sub $16, %[p1]\n\t" + "add $8, %[len]\n\t" + + // main loop: + "0:\n\t" + "add $16, %[p1]\n\t" + "sub $8, %[len]\n\t" + "jz 1f\n\t" + "lddqu (%[p1]), %%xmm0\n\t" + "mov %[len], %%edx\n\t" + "pcmpestri %[mode], (%[p2],%[p1]), %%xmm0\n\t" + + "jna 1f\n\t" + "add $16, %[p1]\n\t" + "sub $8, %[len]\n\t" + "jz 1f\n\t" + "lddqu (%[p1]), %%xmm0\n\t" + "mov %[len], %%edx\n\t" + "pcmpestri %[mode], (%[p2],%[p1]), %%xmm0\n\t" + + "ja 0b\n\t" + "1:\n\t" + "setnc %[result]\n\t" + : [result] "=a" (result), + [p1] "+r" (p1), + [p2] "+r" (p2) + : [len] "0" (len & ~7), + [mode] "i" (_SIDD_UWORD_OPS | _SIDD_CMP_EQUAL_EACH | _SIDD_NEGATIVE_POLARITY) + : "%edx", "%ecx", "%xmm0" + ); + return result && equals2_short_tail(p1, (const ushort *)(quintptr(p1) + quintptr(p2)), len & 7); + } + } + +// const qptrdiff disp = p2 - p1; +// p1 -= 8; +// len += 8; +// while (true) { +// enum { Mode = _SIDD_UWORD_OPS | _SIDD_CMP_EQUAL_EACH | _SIDD_NEGATIVE_POLARITY }; + +// p1 += 8; +// len -= 8; +// if (!len) +// return true; + +// __m128i q1 = _mm_lddqu_si128((__m128i *)(p1 + disp)); +// __m128i *m2 = (__m128i *)p1; + +// bool cmp_a = _mm_cmpestra(q1, len, *m2, len, Mode); +// if (cmp_a) +// continue; +// return !_mm_cmpestrc(q1, len, *m2, len, Mode); +// } +// return true; + bool result; + asm ( + "sub %[p1], %[p2]\n\t" + "sub $16, %[p1]\n\t" + "add $8, %[len]\n\t" + + "0:\n\t" + "add $16, %[p1]\n\t" + "sub $8, %[len]\n\t" + "jz 1f\n\t" + "lddqu (%[p2],%[p1]), %%xmm0\n\t" + "mov %[len], %%edx\n\t" + "pcmpestri %[mode], (%[p1]), %%xmm0\n\t" + + "jna 1f\n\t" + "add $16, %[p1]\n\t" + "sub $8, %[len]\n\t" + "jz 1f\n\t" + "lddqu (%[p2],%[p1]), %%xmm0\n\t" + "mov %[len], %%edx\n\t" + "pcmpestri %[mode], (%[p1]), %%xmm0\n\t" + + "ja 0b\n\t" + + "1:\n\t" + "setnc %[result]\n\t" + : [result] "=a" (result) + : [len] "0" (len), + [p1] "r" (p1), + [p2] "r" (p2), + [mode] "i" (_SIDD_UWORD_OPS | _SIDD_CMP_EQUAL_EACH | _SIDD_NEGATIVE_POLARITY) + : "%edx", "%ecx", "%xmm0" + ); + return result; +} + +#endif +#endif +#endif +#endif + +typedef bool (* FuncPtr)(const ushort *, const ushort *, int); +static const FuncPtr func[] = { + equals2_memcmp_call, // 0 + equals2_bytewise, // 1 + equals2_shortwise, // 1 + equals2_intwise, // 3 +#ifdef __SSE2__ + equals2_sse2, // 4 + equals2_sse2_aligning, // 5 +#ifdef __SSE3__ + equals2_sse3, // 6 +#ifdef __SSSE3__ + equals2_ssse3, // 7 + equals2_ssse3, // 8 +#ifdef __SSE4_1__ + equals2_sse4, // 9 +#endif +#endif +#endif +#endif + 0 +}; +static const int functionCount = sizeof(func)/sizeof(func[0]) - 1; + +void tst_QString::equals2_data() const +{ + QTest::addColumn<int>("algorithm"); + QTest::newRow("selftest") << -1; + QTest::newRow("memcmp_call") << 0; + QTest::newRow("bytewise") << 1; + QTest::newRow("shortwise") << 2; + QTest::newRow("intwise") << 3; +#ifdef __SSE2__ + QTest::newRow("sse2") << 4; + QTest::newRow("sse2_aligning") << 5; +#ifdef __SSE3__ + QTest::newRow("sse3") << 6; +#ifdef __SSSE3__ + QTest::newRow("ssse3") << 7; + QTest::newRow("ssse3_aligning") << 8; +#ifdef __SSE4_1__ + QTest::newRow("sse4.2") << 9; +#endif +#endif +#endif +#endif +} + +static void __attribute__((noinline)) equals2_selftest() +{ +#ifdef Q_OS_UNIX + const long pagesize = sysconf(_SC_PAGESIZE); + void *page1, *page3; + ushort *page2; + page1 = mmap(0, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); + page2 = (ushort *)mmap(0, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS | MAP_POPULATE, -1, 0); + page3 = mmap(0, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); + + Q_ASSERT(quintptr(page2) == quintptr(page1) + pagesize || quintptr(page2) == quintptr(page1) - pagesize); + Q_ASSERT(quintptr(page3) == quintptr(page2) + pagesize || quintptr(page3) == quintptr(page2) - pagesize); + munmap(page1, pagesize); + munmap(page3, pagesize); + + // populate our page + for (uint i = 0; i < pagesize / sizeof(long long); ++i) + ((long long *)page2)[i] = Q_INT64_C(0x0041004100410041); + + // the following should crash: + //page2[-1] = 0xdead; + //page2[pagesize / sizeof(ushort) + 1] = 0xbeef; + + static const ushort needle[] = { + 0x41, 0x41, 0x41, 0x41, 0x41, 0x41, 0x41, 0x41, + 0x41, 0x41, 0x41, 0x41, 0x41, 0x41, 0x41, 0x41, + 0x41 + }; + + for (int algo = 0; algo < functionCount; ++algo) { + // boundary condition test: + for (int i = 0; i < 8; ++i) { + (func[algo])(page2 + i, needle, sizeof needle / 2); + (func[algo])(page2 - i - 1 - sizeof(needle)/2 + pagesize/2, needle, sizeof needle/2); + } + } + + munmap(page2, pagesize); +#endif + + for (int algo = 0; algo < functionCount; ++algo) { + for (int i = 0; i < stringCollectionCount; ++i) { + const ushort *p1 = stringCollectionData + stringCollection[i].offset1; + const ushort *p2 = stringCollectionData + stringCollection[i].offset2; + bool expected = memcmp(p1, p2, stringCollection[i].len * 2) == 0; + + bool result = (func[algo])(p1, p2, stringCollection[i].len); + if (expected != result) + qWarning().nospace() + << "algo=" << algo + << " i=" << i + << " failed (" << result << "!=" << expected + << "); strings were " + << QByteArray((char*)p1, stringCollection[i].len).toHex() + << " and " + << QByteArray((char*)p2, stringCollection[i].len).toHex(); + } + } +} + +void tst_QString::equals2() const +{ + QFETCH(int, algorithm); + if (algorithm == -1) { + equals2_selftest(); + return; + } + + QBENCHMARK { + for (int i = 0; i < stringCollectionCount; ++i) { + const ushort *p1 = stringCollectionData + stringCollection[i].offset1; + const ushort *p2 = stringCollectionData + stringCollection[i].offset2; + bool result = (func[algorithm])(p1, p2, stringCollection[i].len); + Q_UNUSED(result); + } + } +} + +static int ucstrncmp_shortwise(const ushort *a, const ushort *b, int l) +{ + while (l-- && *a == *b) + a++,b++; + if (l==-1) + return 0; + return *a - *b; +} + +static int ucstrncmp_intwise(const ushort *a, const ushort *b, int len) +{ + // do both strings have the same alignment? + if ((quintptr(a) & 2) == (quintptr(b) & 2)) { + // are we aligned to 4 bytes? + if (quintptr(a) & 2) { + if (*a != *b) + return *a - *b; + ++a; + ++b; + --len; + } + + const uint *p1 = (const uint *)a; + const uint *p2 = (const uint *)b; + quintptr counter = 0; + for ( ; len > 1 ; len -= 2, ++counter) { + if (p1[counter] != p2[counter]) { + // which ushort isn't equal? + int diff = a[2*counter] - b[2*counter]; + return diff ? diff : a[2*counter + 1] - b[2*counter + 1]; + } + } + + return len ? a[2*counter] - b[2*counter] : 0; + } else { + while (len-- && *a == *b) + a++,b++; + if (len==-1) + return 0; + return *a - *b; + } +} + +#ifdef __SSE2__ +static inline int ucstrncmp_short_tail(const ushort *p1, const ushort *p2, int len) +{ + if (len) { + if (*p1 != *p2) + return *p1 - *p2; + if (--len) { + if (p1[1] != p2[1]) + return p1[1] - p2[1]; + if (--len) { + if (p1[2] != p2[2]) + return p1[2] - p2[2]; + if (--len) { + if (p1[3] != p2[3]) + return p1[3] - p2[3]; + if (--len) { + if (p1[4] != p2[4]) + return p1[4] - p2[4]; + if (--len) { + if (p1[5] != p2[5]) + return p1[5] - p2[5]; + if (--len) { + if (p1[6] != p2[6]) + return p1[6] - p2[6]; + return p1[7] - p2[7]; + } + } + } + } + } + } + } + return 0; +} + +static inline int bsf_nonzero(register long val) +{ + int result; +# ifdef Q_CC_GNU + // returns the first non-zero bit on a non-zero reg + asm ("bsf %1, %0" : "=r" (result) : "r" (val)); + return result; +# elif defined(Q_CC_MSVC) + _BitScanForward(&result, val); + return result; +# endif +} + +static __attribute__((optimize("no-unroll-loops"))) int ucstrncmp_sse2(const ushort *a, const ushort *b, int len) +{ + qptrdiff counter = 0; + while (len >= 8) { + __m128i m1 = _mm_loadu_si128((__m128i *)(a + counter)); + __m128i m2 = _mm_loadu_si128((__m128i *)(b + counter)); + __m128i cmp = _mm_cmpeq_epi16(m1, m2); + ushort mask = ~uint(_mm_movemask_epi8(cmp)); + if (mask) { + // which ushort isn't equal? + counter += bsf_nonzero(mask)/2; + return a[counter] - b[counter]; + } + + counter += 8; + len -= 8; + } + return ucstrncmp_short_tail(a + counter, b + counter, len); +} + +static __attribute__((optimize("no-unroll-loops"))) int ucstrncmp_sse2_aligning(const ushort *a, const ushort *b, int len) +{ + if (len >= 8) { + __m128i m1 = _mm_loadu_si128((__m128i *)a); + __m128i m2 = _mm_loadu_si128((__m128i *)b); + __m128i cmp = _mm_cmpeq_epi16(m1, m2); + ushort mask = ~uint(_mm_movemask_epi8(cmp)); + if (mask) { + // which ushort isn't equal? + int counter = bsf_nonzero(mask)/2; + return a[counter] - b[counter]; + } + + + // now align to do 16-byte loads + int diff = 8 - (quintptr(a) & 0xf)/2; + len -= diff; + a += diff; + b += diff; + } + + qptrdiff counter = 0; + while (len >= 8) { + __m128i m1 = _mm_load_si128((__m128i *)(a + counter)); + __m128i m2 = _mm_loadu_si128((__m128i *)(b + counter)); + __m128i cmp = _mm_cmpeq_epi16(m1, m2); + ushort mask = ~uint(_mm_movemask_epi8(cmp)); + if (mask) { + // which ushort isn't equal? + counter += bsf_nonzero(mask)/2; + return a[counter] - b[counter]; + } + + counter += 8; + len -= 8; + } + return ucstrncmp_short_tail(a + counter, b + counter, len); +} + +static inline __attribute__((optimize("no-unroll-loops"))) int ucstrncmp_sse2_aligned(const ushort *a, const ushort *b, int len) +{ + quintptr counter = 0; + while (len >= 8) { + __m128i m1 = _mm_load_si128((__m128i *)(a + counter)); + __m128i m2 = _mm_load_si128((__m128i *)(b + counter)); + __m128i cmp = _mm_cmpeq_epi16(m1, m2); + ushort mask = ~uint(_mm_movemask_epi8(cmp)); + if (mask) { + // which ushort isn't equal? + counter += bsf_nonzero(mask)/2; + return a[counter] - b[counter]; + } + + counter += 8; + len -= 8; + } + return ucstrncmp_short_tail(a + counter, b + counter, len); +} + +static inline __attribute__((optimize("no-unroll-loops"))) int ucstrncmp_ssse3_alignr_aligned(const ushort *a, const ushort *b, int len) +{ + quintptr counter = 0; + while (len >= 8) { + __m128i m1 = _mm_load_si128((__m128i *)(a + counter)); + __m128i m2 = _mm_lddqu_si128((__m128i *)(b + counter)); + __m128i cmp = _mm_cmpeq_epi16(m1, m2); + ushort mask = ~uint(_mm_movemask_epi8(cmp)); + if (mask) { + // which ushort isn't equal? + counter += bsf_nonzero(mask)/2; + return a[counter] - b[counter]; + } + + counter += 8; + len -= 8; + } + return ucstrncmp_short_tail(a + counter, b + counter, len); +} + + +typedef __m128i (* MMLoadFunction)(const __m128i *); +template<int N, MMLoadFunction LoadFunction> +static inline __attribute__((optimize("no-unroll-loops"))) int ucstrncmp_ssse3_alignr(const ushort *a, const ushort *b, int len) +{ + qptrdiff counter = 0; + __m128i lower, upper; + upper = _mm_load_si128((__m128i *)a); + + do { + lower = upper; + upper = _mm_load_si128((__m128i *)(a + counter) + 1); + __m128i merged = _mm_alignr_epi8(upper, lower, N); + + __m128i m2 = LoadFunction((__m128i *)(b + counter)); + __m128i cmp = _mm_cmpeq_epi16(merged, m2); + ushort mask = ~uint(_mm_movemask_epi8(cmp)); + if (mask) { + // which ushort isn't equal? + counter += bsf_nonzero(mask)/2; + return a[counter + N/2] - b[counter]; + } + + counter += 8; + len -= 8; + } while (len >= 8); + + return ucstrncmp_short_tail(a + counter + N/2, b + counter, len); +} + +static int ucstrncmp_ssse3(const ushort *a, const ushort *b, int len) +{ + if (len >= 8) { + int val = quintptr(a) & 0xf; + a -= val/2; + + if (val == 10) + return ucstrncmp_ssse3_alignr<10, _mm_lddqu_si128>(a, b, len); + else if (val == 2) + return ucstrncmp_ssse3_alignr<2, _mm_lddqu_si128>(a, b, len); + if (val < 8) { + if (val < 4) + return ucstrncmp_ssse3_alignr_aligned(a, b, len); + else if (val == 4) + return ucstrncmp_ssse3_alignr<4, _mm_lddqu_si128>(a, b, len); + else + return ucstrncmp_ssse3_alignr<6, _mm_lddqu_si128>(a, b, len); + } else { + if (val < 12) + return ucstrncmp_ssse3_alignr<8, _mm_lddqu_si128>(a, b, len); + else if (val == 12) + return ucstrncmp_ssse3_alignr<12, _mm_lddqu_si128>(a, b, len); + else + return ucstrncmp_ssse3_alignr<14, _mm_lddqu_si128>(a, b, len); + } + } + return ucstrncmp_short_tail(a, b, len); +} + +static int ucstrncmp_ssse3_aligning(const ushort *a, const ushort *b, int len) +{ + if (len >= 8) { + __m128i m1 = _mm_loadu_si128((__m128i *)a); + __m128i m2 = _mm_loadu_si128((__m128i *)b); + __m128i cmp = _mm_cmpeq_epi16(m1, m2); + ushort mask = ~uint(_mm_movemask_epi8(cmp)); + if (mask) { + // which ushort isn't equal? + int counter = bsf_nonzero(mask)/2; + return a[counter] - b[counter]; + } + + + // now 'b' align to do 16-byte loads + int diff = 8 - (quintptr(b) & 0xf)/2; + len -= diff; + a += diff; + b += diff; + } + + if (len < 8) + return ucstrncmp_short_tail(a, b, len); + + // 'b' is aligned + int val = quintptr(a) & 0xf; + a -= val/2; + + if (val == 8) + return ucstrncmp_ssse3_alignr<8, _mm_load_si128>(a, b, len); + else if (val == 0) + return ucstrncmp_sse2_aligned(a, b, len); + if (val < 8) { + if (val < 4) + return ucstrncmp_ssse3_alignr<2, _mm_load_si128>(a, b, len); + else if (val == 4) + return ucstrncmp_ssse3_alignr<4, _mm_load_si128>(a, b, len); + else + return ucstrncmp_ssse3_alignr<6, _mm_load_si128>(a, b, len); + } else { + if (val < 12) + return ucstrncmp_ssse3_alignr<10, _mm_load_si128>(a, b, len); + else if (val == 12) + return ucstrncmp_ssse3_alignr<12, _mm_load_si128>(a, b, len); + else + return ucstrncmp_ssse3_alignr<14, _mm_load_si128>(a, b, len); + } +} + +static inline __attribute__((optimize("no-unroll-loops"))) +int ucstrncmp_ssse3_aligning2_aligned(const ushort *a, const ushort *b, int len, int garbage) +{ + // len >= 8 + __m128i m1 = _mm_load_si128((const __m128i *)a); + __m128i m2 = _mm_load_si128((const __m128i *)b); + __m128i cmp = _mm_cmpeq_epi16(m1, m2); + int mask = short(_mm_movemask_epi8(cmp)); // force sign extension + mask >>= garbage; + if (~mask) { + // which ushort isn't equal? + uint counter = (garbage + bsf_nonzero(~mask)); + return a[counter/2] - b[counter/2]; + } + + // the first 16-garbage bytes (8-garbage/2 ushorts) were equal + len -= 8 - garbage/2; + return ucstrncmp_sse2_aligned(a + 8, b + 8, len); +} + +template<int N> static inline __attribute__((optimize("no-unroll-loops"),always_inline)) +int ucstrncmp_ssse3_aligning2_alignr(const ushort *a, const ushort *b, int len, int garbage) +{ + // len >= 8 + __m128i lower, upper, merged; + lower = _mm_load_si128((const __m128i*)a); + upper = _mm_load_si128((const __m128i*)(a + 8)); + merged = _mm_alignr_epi8(upper, lower, N); + + __m128i m2 = _mm_load_si128((const __m128i*)b); + __m128i cmp = _mm_cmpeq_epi16(merged, m2); + int mask = short(_mm_movemask_epi8(cmp)); // force sign extension + mask >>= garbage; + if (~mask) { + // which ushort isn't equal? + uint counter = (garbage + bsf_nonzero(~mask)); + return a[counter/2 + N/2] - b[counter/2]; + } + + // the first 16-garbage bytes (8-garbage/2 ushorts) were equal + quintptr counter = 8; + len -= 8 - garbage/2; + while (len >= 8) { + lower = upper; + upper = _mm_load_si128((__m128i *)(a + counter) + 1); + merged = _mm_alignr_epi8(upper, lower, N); + + m2 = _mm_load_si128((__m128i *)(b + counter)); + cmp = _mm_cmpeq_epi16(merged, m2); + ushort mask = ~uint(_mm_movemask_epi8(cmp)); + if (mask) { + // which ushort isn't equal? + counter += bsf_nonzero(mask)/2; + return a[counter + N/2] - b[counter]; + } + + counter += 8; + len -= 8; + } + + return ucstrncmp_short_tail(a + counter + N/2, b + counter, len); +} + +static inline int conditional_invert(int result, bool invert) +{ + if (invert) + return -result; + return result; +} + +static int ucstrncmp_ssse3_aligning2(const ushort *a, const ushort *b, int len) +{ + // Different strategy from above: instead of doing two unaligned loads + // when trying to align, we'll only do aligned loads and round down the + // addresses of a and b. This means the first load will contain garbage + // in the beginning of the string, which we'll shift out of the way + // (after _mm_movemask_epi8) + + if (len < 8) + return ucstrncmp_intwise(a, b, len); + + // both a and b are misaligned + // we'll call the alignr function with the alignment *difference* between the two + int offset = (quintptr(a) & 0xf) - (quintptr(b) & 0xf); + if (offset >= 0) { + // from this point on, b has the shortest alignment + // and align(a) = align(b) + offset + // round down the alignment so align(b) == align(a) == 0 + int garbage = (quintptr(b) & 0xf); + a = (const ushort*)(quintptr(a) & ~0xf); + b = (const ushort*)(quintptr(b) & ~0xf); + + // now the first load of b will load 'garbage' extra bytes + // and the first load of a will load 'garbage + offset' extra bytes + if (offset == 8) + return ucstrncmp_ssse3_aligning2_alignr<8>(a, b, len, garbage); + if (offset == 0) + return ucstrncmp_ssse3_aligning2_aligned(a, b, len, garbage); + if (offset < 8) { + if (offset < 4) + return ucstrncmp_ssse3_aligning2_alignr<2>(a, b, len, garbage); + else if (offset == 4) + return ucstrncmp_ssse3_aligning2_alignr<4>(a, b, len, garbage); + else + return ucstrncmp_ssse3_aligning2_alignr<6>(a, b, len, garbage); + } else { + if (offset < 12) + return ucstrncmp_ssse3_aligning2_alignr<10>(a, b, len, garbage); + else if (offset == 12) + return ucstrncmp_ssse3_aligning2_alignr<12>(a, b, len, garbage); + else + return ucstrncmp_ssse3_aligning2_alignr<14>(a, b, len, garbage); + } + } else { + // same as above but inverted + int garbage = (quintptr(a) & 0xf); + a = (const ushort*)(quintptr(a) & ~0xf); + b = (const ushort*)(quintptr(b) & ~0xf); + + offset = -offset; + if (offset == 8) + return -ucstrncmp_ssse3_aligning2_alignr<8>(b, a, len, garbage); + if (offset < 8) { + if (offset < 4) + return -ucstrncmp_ssse3_aligning2_alignr<2>(b, a, len, garbage); + else if (offset == 4) + return -ucstrncmp_ssse3_aligning2_alignr<4>(b, a, len, garbage); + else + return -ucstrncmp_ssse3_aligning2_alignr<6>(b, a, len, garbage); + } else { + if (offset < 12) + return -ucstrncmp_ssse3_aligning2_alignr<10>(b, a, len, garbage); + else if (offset == 12) + return -ucstrncmp_ssse3_aligning2_alignr<12>(b, a, len, garbage); + else + return -ucstrncmp_ssse3_aligning2_alignr<14>(b, a, len, garbage); + } + } +} + +#endif + +typedef int (* UcstrncmpFunction)(const ushort *, const ushort *, int); +Q_DECLARE_METATYPE(UcstrncmpFunction) + +void tst_QString::ucstrncmp_data() const +{ + QTest::addColumn<UcstrncmpFunction>("function"); + QTest::newRow("selftest") << UcstrncmpFunction(0); + QTest::newRow("shortwise") << &ucstrncmp_shortwise; + QTest::newRow("intwise") << &ucstrncmp_intwise; +#ifdef __SSE2__ + QTest::newRow("sse2") << &ucstrncmp_sse2; + QTest::newRow("sse2_aligning") << &ucstrncmp_sse2_aligning; +#ifdef __SSSE3__ + QTest::newRow("ssse3") << &ucstrncmp_ssse3; + QTest::newRow("ssse3_aligning") << &ucstrncmp_ssse3_aligning; + QTest::newRow("ssse3_aligning2") << &ucstrncmp_ssse3_aligning2; +#endif +#endif +} + +void tst_QString::ucstrncmp() const +{ + QFETCH(UcstrncmpFunction, function); + if (!function) { + static const UcstrncmpFunction func[] = { + &ucstrncmp_shortwise, + &ucstrncmp_intwise, +#ifdef __SSE2__ + &ucstrncmp_sse2, + &ucstrncmp_sse2_aligning, +#ifdef __SSSE3__ + &ucstrncmp_ssse3, + &ucstrncmp_ssse3_aligning, + &ucstrncmp_ssse3_aligning2 +#endif +#endif + }; + static const int functionCount = sizeof func / sizeof func[0]; + +#ifdef Q_OS_UNIX + const long pagesize = sysconf(_SC_PAGESIZE); + void *page1, *page3; + ushort *page2; + page1 = mmap(0, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); + page2 = (ushort *)mmap(0, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS | MAP_POPULATE, -1, 0); + page3 = mmap(0, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); + + Q_ASSERT(quintptr(page2) == quintptr(page1) + pagesize || quintptr(page2) == quintptr(page1) - pagesize); + Q_ASSERT(quintptr(page3) == quintptr(page2) + pagesize || quintptr(page3) == quintptr(page2) - pagesize); + munmap(page1, pagesize); + munmap(page3, pagesize); + + // populate our page + for (uint i = 0; i < pagesize / sizeof(long long); ++i) + ((long long *)page2)[i] = Q_INT64_C(0x0041004100410041); + + // the following should crash: + //page2[-1] = 0xdead; + //page2[pagesize / sizeof(ushort) + 1] = 0xbeef; + + static const ushort needle[] = { + 0x41, 0x41, 0x41, 0x41, 0x41, 0x41, 0x41, 0x41, + 0x41, 0x41, 0x41, 0x41, 0x41, 0x41, 0x41, 0x41, + 0x41 + }; + + for (int algo = 0; algo < functionCount; ++algo) { + // boundary condition test: + for (int i = 0; i < 8; ++i) { + (func[algo])(page2 + i, needle, sizeof needle / 2); + (func[algo])(page2 - i - 1 - sizeof(needle)/2 + pagesize/2, needle, sizeof needle/2); + } + } + + munmap(page2, pagesize); +#endif + + for (int algo = 0; algo < functionCount; ++algo) { + for (int i = 0; i < stringCollectionCount; ++i) { + const ushort *p1 = stringCollectionData + stringCollection[i].offset1; + const ushort *p2 = stringCollectionData + stringCollection[i].offset2; + int expected = ucstrncmp_shortwise(p1, p2, stringCollection[i].len); + expected = qBound(-1, expected, 1); + + int result = (func[algo])(p1, p2, stringCollection[i].len); + result = qBound(-1, result, 1); + if (expected != result) + qWarning().nospace() + << "algo=" << algo + << " i=" << i + << " failed (" << result << "!=" << expected + << "); strings were " + << QByteArray((char*)p1, stringCollection[i].len).toHex() + << " and " + << QByteArray((char*)p2, stringCollection[i].len).toHex(); + } + } + return; + } + + QBENCHMARK { + for (int i = 0; i < stringCollectionCount; ++i) { + const ushort *p1 = stringCollectionData + stringCollection[i].offset1; + const ushort *p2 = stringCollectionData + stringCollection[i].offset2; + (function)(p1, p2, stringCollection[i].len); + } + } +} + void tst_QString::fromUtf8() const { QFile file(SRCDIR "utf-8.txt"); |