Master/upx
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

src: sort cleanups

Browse Source
This commit is contained in:
Markus F.X.J. Oberhumer
2023-09-04 07:28:48 +02:00
parent 9331ed39d4
commit 62dbf8485f
13 changed files with 194 additions and 109 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/util/util.cpp
+145 -85
View File
@@ -42,10 +42,9 @@
// assert sane memory buffer sizes to protect against integer overflows
// and malicious header fields
// see C 11 standard, Annex K
//
// this limits uncompressed_size to about 682 MiB (715_128_832 bytes)
**************************************************************************/
// this limits uncompressed_size to about 682 MiB (715_128_832 bytes)
ACC_COMPILE_TIME_ASSERT_HEADER(UPX_RSIZE_MAX_MEM == UPX_RSIZE_MAX)
ACC_COMPILE_TIME_ASSERT_HEADER(UPX_RSIZE_MAX_STR <= UPX_RSIZE_MAX / 256)
ACC_COMPILE_TIME_ASSERT_HEADER(2ull * UPX_RSIZE_MAX * 9 / 8 + 256 * 1024 * 1024 < INT_MAX)
@@ -54,23 +53,6 @@ ACC_COMPILE_TIME_ASSERT_HEADER(5ull * UPX_RSIZE_MAX < UINT_MAX)
ACC_COMPILE_TIME_ASSERT_HEADER(UPX_RSIZE_MAX >= 8192 * 65536)
ACC_COMPILE_TIME_ASSERT_HEADER(UPX_RSIZE_MAX_STR >= 1024)
upx_rsize_t mem_size(upx_uint64_t element_size, upx_uint64_t n, upx_uint64_t extra1,
upx_uint64_t extra2) {
assert(element_size > 0);
if very_unlikely (element_size == 0 || element_size > UPX_RSIZE_MAX)
throwCantPack("mem_size 1; take care");
if very_unlikely (n > UPX_RSIZE_MAX)
throwCantPack("mem_size 2; take care");
if very_unlikely (extra1 > UPX_RSIZE_MAX)
throwCantPack("mem_size 3; take care");
if very_unlikely (extra2 > UPX_RSIZE_MAX)
throwCantPack("mem_size 4; take care");
upx_uint64_t bytes = element_size * n + extra1 + extra2; // cannot overflow
if very_unlikely (bytes > UPX_RSIZE_MAX)
throwCantPack("mem_size 5; take care");
return ACC_ICONV(upx_rsize_t, bytes);
}
bool mem_size_valid(upx_uint64_t element_size, upx_uint64_t n, upx_uint64_t extra1,
upx_uint64_t extra2) noexcept {
assert_noexcept(element_size > 0);
@@ -88,6 +70,23 @@ bool mem_size_valid(upx_uint64_t element_size, upx_uint64_t n, upx_uint64_t extr
return true;
}
upx_rsize_t mem_size(upx_uint64_t element_size, upx_uint64_t n, upx_uint64_t extra1,
upx_uint64_t extra2) {
assert(element_size > 0);
if very_unlikely (element_size == 0 || element_size > UPX_RSIZE_MAX)
throwCantPack("mem_size 1; take care");
if very_unlikely (n > UPX_RSIZE_MAX)
throwCantPack("mem_size 2; take care");
if very_unlikely (extra1 > UPX_RSIZE_MAX)
throwCantPack("mem_size 3; take care");
if very_unlikely (extra2 > UPX_RSIZE_MAX)
throwCantPack("mem_size 4; take care");
upx_uint64_t bytes = element_size * n + extra1 + extra2; // cannot overflow
if very_unlikely (bytes > UPX_RSIZE_MAX)
throwCantPack("mem_size 5; take care");
return ACC_ICONV(upx_rsize_t, bytes);
}
TEST_CASE("mem_size") {
CHECK(mem_size_valid(1, 0));
CHECK(mem_size_valid(1, 0x30000000));
@@ -277,18 +276,18 @@ void upx_memswap(void *a, void *b, size_t n) {
}
}
// somewhat better memswap(), optimized for our use cases in sort functions
// much better memswap(), optimized for our use case in sort functions below
static void memswap_no_overlap(char *a, char *b, size_t n) {
#if defined(__clang__) && __clang_major__ < 15 && 1
// work around a clang ICE (Internal Compiler Error); sigh
#if defined(__clang__) && __clang_major__ < 15
// work around a clang < 15 ICE (Internal Compiler Error)
upx_memswap(a, b, n);
#else // clang bug
alignas(16) char tmpbuf[16];
upx_alignas_max char tmp_buf[16];
#define SWAP(x) \
ACC_BLOCK_BEGIN \
upx_memcpy_inline(tmpbuf, a, x); \
upx_memcpy_inline(tmp_buf, a, x); \
upx_memcpy_inline(a, b, x); \
upx_memcpy_inline(b, tmpbuf, x); \
upx_memcpy_inline(b, tmp_buf, x); \
a += x; \
b += x; \
ACC_BLOCK_END
@@ -310,25 +309,9 @@ static void memswap_no_overlap(char *a, char *b, size_t n) {
#endif // clang bug
}
// simple Shell sort using Knuth's gap; NOT stable
void upx_shellsort(void *array, size_t n, size_t element_size, upx_compare_func_t compare) {
mem_size_assert(element_size, n); // check size
size_t gap = 0;
while (gap * 3 + 1 < n) // cannot overflow
gap = gap * 3 + 1;
for (; gap > 0; gap = (gap - 1) / 3) {
const size_t gap_bytes = element_size * gap;
char *const gbase = (char *) array + gap_bytes; // gbase := &array[gap]
char *ii = gbase;
for (size_t i = gap; i < n; i += gap, ii += gap_bytes)
for (char *a = ii; a >= gbase && compare(a - gap_bytes, a) > 0; a -= gap_bytes)
memswap_no_overlap(a - gap_bytes, a, element_size);
}
}
// extremely simple (and beautiful) stable sort: Gnomesort
// WARNING: O(n^2) and thus very inefficient for large n
void upx_stable_sort(void *array, size_t n, size_t element_size, upx_compare_func_t compare) {
void upx_gnomesort(void *array, size_t n, size_t element_size, upx_compare_func_t compare) {
for (size_t i = 1; i < n; i++) {
char *a = (char *) array + element_size * i; // a := &array[i]
if (i != 0 && compare(a - element_size, a) > 0) { // if a[-1] > a[0] then
@@ -338,36 +321,100 @@ void upx_stable_sort(void *array, size_t n, size_t element_size, upx_compare_fun
}
}
#if !defined(DOCTEST_CONFIG_DISABLE) && DEBUG
TEST_CASE("basic upx_stable_sort") {
{
unsigned a[] = {0, 1};
upx_stable_sort(a, 2, sizeof(*a), ne32_compare);
CHECK((a[0] == 0 && a[1] == 1));
}
{
unsigned a[] = {1, 0};
upx_stable_sort(a, 2, sizeof(*a), ne32_compare);
CHECK((a[0] == 0 && a[1] == 1));
}
{
BE64 a[3];
a[0] = 257;
a[1] = 256;
a[2] = 255;
upx_stable_sort(a, 3, sizeof(*a), be64_compare);
CHECK((a[0] == 255 && a[1] == 256 && a[2] == 257));
// simple Shell sort using Knuth's gap; NOT stable; uses memswap()
// cannot compete with modern sort algorithms, but not too bad as a generic fallback
void upx_shellsort_memswap(void *array, size_t n, size_t element_size, upx_compare_func_t compare) {
mem_size_assert(element_size, n); // check size
size_t gap = 0; // 0, 1, 4, 13, 40, 121, 364, 1093, ...
while (gap * 3 + 1 < n) // cannot overflow because of size check above
gap = gap * 3 + 1;
for (; gap > 0; gap = (gap - 1) / 3) {
const size_t gap_bytes = element_size * gap;
char *p = (char *) array + gap_bytes;
for (size_t i = gap; i < n; i += gap, p += gap_bytes) // invariant: p == &array[i]
for (char *a = p; a != array && compare(a - gap_bytes, a) > 0; a -= gap_bytes)
memswap_no_overlap(a - gap_bytes, a, element_size);
}
}
// simple Shell sort using Knuth's gap; NOT stable; uses memcpy()
// should be faster than memswap() in theory, but benchmarks are inconsistent
void upx_shellsort_memcpy(void *array, size_t n, size_t element_size, upx_compare_func_t compare) {
mem_size_assert(element_size, n); // check size
constexpr size_t MAX_INLINE_ELEMENT_SIZE = 256;
upx_alignas_max char tmp_buf[MAX_INLINE_ELEMENT_SIZE]; // buffer for one element
char *tmp = tmp_buf;
if (element_size > MAX_INLINE_ELEMENT_SIZE) {
tmp = (char *) malloc(element_size);
assert(tmp != nullptr);
}
size_t gap = 0; // 0, 1, 4, 13, 40, 121, 364, 1093, ...
while (gap * 3 + 1 < n) // cannot overflow because of size check above
gap = gap * 3 + 1;
for (; gap > 0; gap = (gap - 1) / 3) {
const size_t gap_bytes = element_size * gap;
char *p = (char *) array + gap_bytes;
for (size_t i = gap; i < n; i += gap, p += gap_bytes) // invariant: p == &array[i]
if (compare(p - gap_bytes, p) > 0) {
char *a = p;
memcpy(tmp, a, element_size);
do {
memcpy(a, a - gap_bytes, element_size);
a -= gap_bytes;
} while (a != array && compare(a - gap_bytes, tmp) > 0);
memcpy(a, tmp, element_size);
}
}
if (element_size > MAX_INLINE_ELEMENT_SIZE)
free(tmp);
}
// wrap std::stable_sort()
template <size_t ElementSize>
void upx_std_stable_sort(void *array, size_t n, upx_compare_func_t compare) {
static_assert(ElementSize > 0 && ElementSize <= UPX_RSIZE_MAX);
mem_size_assert(ElementSize, n); // check size
#if 0
// just for testing
upx_gnomesort(array, n, ElementSize, compare);
#else
struct alignas(1) element_type { char data[ElementSize]; };
static_assert(sizeof(element_type) == ElementSize);
static_assert(alignof(element_type) == 1);
auto cmp = [compare](const element_type &a, const element_type &b) -> bool {
return compare(&a, &b) < 0;
};
std::stable_sort((element_type *) array, (element_type *) array + n, cmp);
#endif
}
#if UPX_QSORT_IS_STABLE_SORT
// instantiate function templates for all element sizes we need
// efficient, but code size bloat
template void upx_std_stable_sort<1>(void *, size_t, upx_compare_func_t);
template void upx_std_stable_sort<2>(void *, size_t, upx_compare_func_t);
template void upx_std_stable_sort<4>(void *, size_t, upx_compare_func_t);
template void upx_std_stable_sort<8>(void *, size_t, upx_compare_func_t);
template void upx_std_stable_sort<16>(void *, size_t, upx_compare_func_t);
template void upx_std_stable_sort<32>(void *, size_t, upx_compare_func_t);
template void upx_std_stable_sort<56>(void *, size_t, upx_compare_func_t);
template void upx_std_stable_sort<72>(void *, size_t, upx_compare_func_t);
#endif
#if !defined(DOCTEST_CONFIG_DISABLE) && DEBUG >= 1
#if __cplusplus >= 202002L // use C++20 std::next_permutation() to test all permutations
namespace {
template <class ElementType, upx_compare_func_t CompareFunc>
struct TestSortAllPermutations {
typedef ElementType element_type;
static noinline upx_uint64_t test(upx_sort_func_t sort, size_t n) {
constexpr size_t N = 16;
assert(n > 0 && n <= N);
assert_noexcept(n <= N);
ElementType perm[N];
if (n == 0) {
sort(perm, 0, sizeof(ElementType), CompareFunc); // check that n == 0 works
return 0;
}
for (size_t i = 0; i < n; i++)
perm[i] = 255 + i;
upx_uint64_t num_perms = 0;
@@ -376,40 +423,53 @@ struct TestSortAllPermutations {
memcpy(a, perm, sizeof(*a) * n);
sort(a, n, sizeof(*a), CompareFunc);
for (size_t i = 0; i < n; i++)
assert((a[i] == 255 + i));
assert_noexcept((a[i] == 255 + i));
num_perms += 1;
} while (std::next_permutation(perm, perm + n));
return num_perms;
}
static bool test_permutations(upx_sort_func_t sort) {
bool ok = true;
ok &= (test(sort, 0) == 0);
ok &= (test(sort, 1) == 1);
ok &= (test(sort, 2) == 2);
ok &= (test(sort, 3) == 6);
ok &= (test(sort, 4) == 24);
ok &= (test(sort, 5) == 120);
#if DEBUG >= 2
ok &= (test(sort, 6) == 720);
ok &= (test(sort, 7) == 5040);
ok &= (test(sort, 8) == 40320);
ok &= (test(sort, 9) == 362880);
ok &= (test(sort, 10) == 3628800);
// ok &= (test(sort, 11) == 39916800);
#endif
return ok;
}
};
} // namespace
TEST_CASE("upx_shellsort") {
TEST_CASE("upx_gnomesort") {
// typedef TestSortAllPermutations<BE64, be64_compare> TestSort;
typedef TestSortAllPermutations<LE16, le16_compare> TestSort;
CHECK(TestSort::test(upx_shellsort, 1) == 1);
CHECK(TestSort::test(upx_shellsort, 2) == 2);
CHECK(TestSort::test(upx_shellsort, 3) == 6);
CHECK(TestSort::test(upx_shellsort, 4) == 24);
CHECK(TestSort::test(upx_shellsort, 5) == 120);
// CHECK(TestSort::test(upx_shellsort, 6) == 720);
// CHECK(TestSort::test(upx_shellsort, 7) == 5040);
// CHECK(TestSort::test(upx_shellsort, 8) == 40320);
// CHECK(TestSort::test(upx_shellsort, 9) == 362880);
// CHECK(TestSort::test(upx_shellsort, 10) == 3628800);
CHECK(TestSort::test_permutations(upx_gnomesort));
}
TEST_CASE("upx_stable_sort") {
TEST_CASE("upx_shellsort_memswap") {
// typedef TestSortAllPermutations<BE64, be64_compare> TestSort;
typedef TestSortAllPermutations<LE16, le16_compare> TestSort;
CHECK(TestSort::test(upx_stable_sort, 1) == 1);
CHECK(TestSort::test(upx_stable_sort, 2) == 2);
CHECK(TestSort::test(upx_stable_sort, 3) == 6);
CHECK(TestSort::test(upx_stable_sort, 4) == 24);
CHECK(TestSort::test(upx_stable_sort, 5) == 120);
// CHECK(TestSort::test(upx_stable_sort, 6) == 720);
// CHECK(TestSort::test(upx_stable_sort, 7) == 5040);
// CHECK(TestSort::test(upx_stable_sort, 8) == 40320);
// CHECK(TestSort::test(upx_stable_sort, 9) == 362880);
// CHECK(TestSort::test(upx_stable_sort, 10) == 3628800);
CHECK(TestSort::test_permutations(upx_shellsort_memswap));
}
TEST_CASE("upx_shellsort_memcpy") {
// typedef TestSortAllPermutations<BE64, be64_compare> TestSort;
typedef TestSortAllPermutations<LE16, le16_compare> TestSort;
CHECK(TestSort::test_permutations(upx_shellsort_memcpy));
}
TEST_CASE("upx_std_stable_sort") {
// typedef TestSortAllPermutations<BE64, be64_compare> TestSort;
typedef TestSortAllPermutations<LE16, le16_compare> TestSort;
upx_sort_func_t wrap_stable_sort = [](void *a, size_t n, size_t, upx_compare_func_t compare) {
upx_std_stable_sort<sizeof(TestSort::element_type)>(a, n, compare);
};
CHECK(TestSort::test_permutations(wrap_stable_sort));
}
#endif // C++20
#endif // DEBUG
src/util/util.h
+15 -2
View File
@@ -130,9 +130,22 @@ void upx_memswap(void *a, void *b, size_t n);
typedef int(__acc_cdecl_qsort *upx_compare_func_t)(const void *, const void *);
typedef void (*upx_sort_func_t)(void *array, size_t n, size_t element_size, upx_compare_func_t);
void upx_shellsort(void *array, size_t n, size_t element_size, upx_compare_func_t compare);
void upx_gnomesort(void *array, size_t n, size_t element_size, upx_compare_func_t compare);
void upx_shellsort_memswap(void *array, size_t n, size_t element_size, upx_compare_func_t compare);
void upx_shellsort_memcpy(void *array, size_t n, size_t element_size, upx_compare_func_t compare);
void upx_stable_sort(void *array, size_t n, size_t element_size, upx_compare_func_t compare);
// this wraps std::stable_sort()
template <size_t ElementSize>
void upx_std_stable_sort(void *array, size_t n, upx_compare_func_t compare);
#if 1
// use libc qsort()
#define upx_qsort qsort
#else
// use std::stable_sort()
#define upx_qsort(a, b, c, d) upx_std_stable_sort<(c)>(a, b, d)
#define UPX_QSORT_IS_STABLE_SORT 1
#endif
/*************************************************************************
// misc. support functions