#include "prx/libc/include/Api.hpp" #include "prx/libc/include/general/VabiMacros.hpp" #include #include #include #include #include #include #include #include #include #include #ifdef _WIN32 #include #else #include #include #include #endif namespace { struct Environment { std::array x87; std::uint32_t sse; }; Environment ReadEnvironment() { Environment state; asm volatile("fnstenv %0\n\tfldenv %0" : "=m"(state.x87)); state.sse = _mm_getcsr(); return state; } void WriteEnvironment(const Environment& state) { asm volatile("fldenv %0" : : "m"(state.x87)); _mm_setcsr(state.sse); } bool SameControlState(const Environment& left, const Environment& right) { return left.x87[0] == right.x87[0] && left.x87[1] == right.x87[1] && left.sse == right.sse; } void Require(bool value, const char* message) { if (!value) { std::fprintf(stderr, "floating-point environment: %s\n", message); std::abort(); } } Environment SeedFlags(unsigned x87, unsigned sse) { auto state = ReadEnvironment(); state.x87[0] |= 0x3f; state.x87[1] = (state.x87[1] & ~0xffu) | x87; state.sse = (state.sse & ~0x3fu) | (0x3fu << 7) | sse; WriteEnvironment(state); return ReadEnvironment(); } void CheckEnabledExceptions() { const auto seeded = SeedFlags(0, 0); for (unsigned x87 = 0; x87 <= 0x3f; ++x87) { for (unsigned sse = 0; sse <= 0x3f; ++sse) { auto state = seeded; state.x87[0] = (state.x87[0] & ~0xc3fu) | 0x800 | x87; state.x87[1] = (state.x87[1] & ~0x3fu) | (x87 & 0x15u); state.sse = (state.sse & ~0x7fbfu) | 0x2000 | (sse << 7) | (sse & 0x2au); WriteEnvironment(state); *__error_nid_postfix() = 123; const int enabled = fegetexcept_nid_postfix(); const int error = *__error_nid_postfix(); const auto after = ReadEnvironment(); WriteEnvironment(seeded); Require(enabled == static_cast(0x3f ^ x87), "enabled exceptions use the x87 masks independently of SSE"); Require(after.x87[0] == state.x87[0] && after.x87[1] == state.x87[1] && after.sse == state.sse, "enabled exception query preserves control modes and flags"); Require(error == 123, "enabled exception query preserves errno"); } } } void CheckRounding() { const auto seeded = SeedFlags(0x02, 0x10); constexpr int modes[] = {0, 0x400, 0x800, 0xc00}; constexpr int positive[] = {2, 1, 2, 1}; constexpr int negative[] = {-2, -2, -1, -1}; for (unsigned index = 0; index < 4; ++index) { const auto before = ReadEnvironment(); Require(fesetround_nid_postfix(modes[index]) == 0, "set rounding mode"); Require(fegetround_nid_postfix() == modes[index], "query rounding mode"); const auto state = ReadEnvironment(); Require((state.x87[0] & 0xc00) == static_cast(modes[index]), "x87 rounding bits"); Require((state.sse & 0x6000) == static_cast(modes[index] << 3), "SSE rounding bits"); Require((state.x87[0] & ~0xc00u) == (seeded.x87[0] & ~0xc00u), "preserve x87 control bits"); Require((state.sse & ~0x603fu) == (seeded.sse & ~0x603fu), "preserve SSE control bits"); Require(state.x87[1] == before.x87[1] && (state.sse & 0x3f) == (before.sse & 0x3f), "rounding preserves exceptions"); for (const double input : {1.75, -1.75}) { const int expected = input > 0 ? positive[index] : negative[index]; int x87; asm volatile("fldl %1\n\tfistpl %0" : "=m"(x87) : "m"(input) : "st"); Require(x87 == expected, "x87 arithmetic observes rounding"); int sse; asm volatile("cvtsd2si %1, %0" : "=r"(sse) : "x"(input)); Require(sse == expected, "SSE arithmetic observes rounding"); } } for (const int invalid : {-1, 1, 0x200, 0x1000, 0x7fffffff}) { const auto before = ReadEnvironment(); Require(fesetround_nid_postfix(invalid) == -1, "reject invalid rounding mode"); const auto after = ReadEnvironment(); Require(before.x87[0] == after.x87[0] && before.sse == after.sse, "invalid mode leaves state unchanged"); } auto distinct = ReadEnvironment(); distinct.x87[0] = (distinct.x87[0] & ~0xc00u) | 0x400; distinct.sse = (distinct.sse & ~0x6000u) | 0x4000; WriteEnvironment(distinct); Require(fegetround_nid_postfix() == 0x400, "query uses the x87 rounding mode"); } void CheckFeraise() { for (unsigned requested = 0; requested <= 0x3f; ++requested) { const auto seeded = SeedFlags(0, 0); *__error_nid_postfix() = 0; _Feraise_nid_postfix(static_cast(requested)); const unsigned raised = (requested & 0x18) != 0 ? requested | 0x20 : requested; const auto state = ReadEnvironment(); Require((state.x87[1] & 0x3f) == raised && (state.sse & 0x3f) == raised, "_Feraise raises the flags on both units"); Require(state.x87[0] == seeded.x87[0] && (state.sse & ~0x3fu) == (seeded.sse & ~0x3fu), "_Feraise preserves control bits"); const int error = (raised & 0x01) != 0 || (raised & 0x1c) == 0 ? 33 : 34; Require(*__error_nid_postfix() == error, "_Feraise sets EDOM or ERANGE"); } SeedFlags(0, 0); } void CheckFlags() { const auto initial = SeedFlags(0x03, 0x24); Require(fetestexcept_nid_postfix(0) == 0, "empty exception query"); Require(fetestexcept_nid_postfix(0x3f) == 0x27, "combine x87 and SSE flags, including denormals"); Require(fetestexcept_nid_postfix(0x0c) == 0x04, "query only requested flags"); struct { std::uint16_t before, value, after; } saved{0x1234, 0xffff, 0x5678}; Require(fegetexceptflag_nid_postfix(&saved.value, 0x07) == 0 && saved.value == 0x07, "save selected flags"); Require(saved.before == 0x1234 && saved.after == 0x5678, "fexcept_t writes exactly two bytes"); auto state = ReadEnvironment(); Require(state.x87[1] == initial.x87[1] && state.sse == initial.sse, "queries do not change flags"); const std::uint16_t flags = 0x22; Require(fesetexceptflag_nid_postfix(&flags, 0x0a) == 0, "restore selected flags"); state = ReadEnvironment(); Require((state.x87[1] & 0x3f) == 0x03 && (state.sse & 0x3f) == 0x26, "restore only selected flags on both units"); Require(fesetexceptflag_nid_postfix(&flags, 0) == 0, "empty restore"); Require(feclearexcept_nid_postfix(0x03) == 0, "clear selected flags"); state = ReadEnvironment(); Require((state.x87[1] & 0x3f) == 0 && (state.sse & 0x3f) == 0x24, "clear flags on both units"); Require(state.x87[0] == initial.x87[0], "flag operations preserve x87 control"); Require((state.sse & ~0x3fu) == (initial.sse & ~0x3fu), "flag operations preserve SSE control"); Require(feclearexcept_nid_postfix(0) == 0 && fetestexcept_nid_postfix(0x3f) == 0x24, "empty clear"); SeedFlags(0x3f, 0x3f); Require(feclearexcept_nid_postfix(0x3f) == 0 && fetestexcept_nid_postfix(0x3f) == 0, "clear every exception"); Require(fegetexceptflag_nid_postfix(&saved.value, 0) == 0 && saved.value == 0, "save empty mask"); auto traps = ReadEnvironment(); traps.x87[0] &= ~0x08u; traps.sse &= ~(0x08u << 7); WriteEnvironment(traps); traps = ReadEnvironment(); const std::uint16_t none = 0; Require(fesetexceptflag_nid_postfix(&none, 0x01) == 0, "restore with an unmasked exception"); Require(feclearexcept_nid_postfix(0x02) == 0, "clear with an unmasked exception"); state = ReadEnvironment(); Require(state.x87[0] == traps.x87[0] && state.sse == traps.sse, "flag updates preserve exception masks"); SeedFlags(0, 0); for (const bool get : {false, true}) { bool threw = false; try { if (get) fegetexceptflag_nid_postfix(nullptr, 0x3f); else fesetexceptflag_nid_postfix(nullptr, 0x3f); } catch (const std::invalid_argument&) { threw = true; } Require(threw, "reject null exception storage"); } CheckFeraise(); } void CheckThreads() { Require(fesetround_nid_postfix(0) == 0, "reset main rounding mode"); Require(feclearexcept_nid_postfix(0x3f) == 0, "reset main exceptions"); std::barrier ready(2); const auto worker = [&](int rounding, std::uint16_t flags) { Require(fesetround_nid_postfix(rounding) == 0, "thread rounding mode"); Require(fesetexceptflag_nid_postfix(&flags, 0x3f) == 0, "thread exceptions"); ready.arrive_and_wait(); Require(fegetround_nid_postfix() == rounding, "thread-local rounding mode"); Require(fetestexcept_nid_postfix(0x3f) == flags, "thread-local exception flags"); }; std::thread first(worker, 0x400, 0x01); std::thread second(worker, 0x800, 0x04); first.join(); second.join(); Require(fegetround_nid_postfix() == 0 && fetestexcept_nid_postfix(0x3f) == 0, "worker state does not change main"); } } unsigned X87Masks() { return ReadEnvironment().x87[0] & 0x3fu; } unsigned SseMasks() { return (ReadEnvironment().sse >> 7) & 0x3fu; } bool HostKeepsSseTraps() { const unsigned saved = _mm_getcsr(); _mm_setcsr(saved & ~(0x01u << 7)); const bool kept = (_mm_getcsr() & (0x01u << 7)) == 0; _mm_setcsr(saved); return kept; } void CheckTraps() { const auto seeded = SeedFlags(0, 0); const bool sseKept = HostKeepsSseTraps(); const auto sse = [&](unsigned masks) { return sseKept ? masks : 0x3fu; }; const auto both = [&](unsigned traps) { return static_cast(0xffffffc0u | (sseKept ? traps : 0u)); }; Require(feenableexcept_nid_postfix(0x49) == static_cast(0xffffffc0u), "enable reports no previous trap as the console does"); Require(X87Masks() == 0x36 && fegettrapenable_nid_postfix() == 0x09, "enable unmasks the x87 exceptions"); Require(SseMasks() == sse(0x36), "enable unmasks the SSE exceptions"); Require(fedisableexcept_nid_postfix(0x41) == both(0x09), "disable returns the traps enabled on both units"); Require(X87Masks() == 0x37 && fegettrapenable_nid_postfix() == 0x08, "disable masks the x87 exception"); Require(SseMasks() == sse(0x37), "disable masks the SSE exception"); Require(__feenableexcept_nid_postfix(0x02) == both(0x08) && X87Masks() == 0x35, "the internal enable differs"); Require(__fedisableexcept_nid_postfix(0x3f) == both(0x0a) && X87Masks() == 0x3f && fegettrapenable_nid_postfix() == 0, "the internal disable differs"); Require(fesettrapenable_nid_postfix(0x40) == 1 && X87Masks() == 0x3f, "an out-of-range trap set was accepted"); Require(fesettrapenable_nid_postfix(0x10004) == 0 && fegettrapenable_nid_postfix() == 0x04 && SseMasks() == sse(0x3b), "a trap set did not use the low 16 bits"); Require(fesettrapenable_nid_postfix(0x01) == 0 && fegettrapenable_nid_postfix() == 0x05 && SseMasks() == sse(0x3a), "a trap set masked enabled traps"); Require(fesettrapenable_nid_postfix(0) == 0 && fegettrapenable_nid_postfix() == 0x05 && SseMasks() == sse(0x3a), "an empty trap set masked enabled traps"); auto mixed = ReadEnvironment(); mixed.x87[0] = (mixed.x87[0] & ~0x3fu) | 0x36; mixed.sse = (mixed.sse & ~(0x3fu << 7)) | (0x3au << 7); WriteEnvironment(mixed); Require(feenableexcept_nid_postfix(0) == both(0x01), "enable reports only the traps enabled on both units"); Require(fegettrapenable_nid_postfix() == 0x09, "the trap query reads only the x87 control word"); WriteEnvironment(seeded); } using HoldExcept = int (APS5_VABI*)(Environment*); void CheckHold(HoldExcept hold, Environment* saved) { const auto before = ReadEnvironment(); *__error_nid_postfix() = 123; Require(hold(saved) == 0, "feholdexcept succeeds"); const auto after = ReadEnvironment(); Require(std::memcmp(saved, &before, sizeof(before)) == 0, "feholdexcept saves the full original environment"); Require((after.x87[0] & 0x3f) == 0x3f && (after.sse & 0x1f80) == 0x1f80, "feholdexcept masks both units"); Require((after.x87[1] & 0x80ff) == 0 && (after.sse & 0x3f) == 0, "feholdexcept clears pending exceptions"); Require((after.x87[1] & ~0x80ffu) == (before.x87[1] & ~0x80ffu), "feholdexcept preserves other x87 status bits"); Require((after.x87[0] & ~0x3fu) == (before.x87[0] & ~0x3fu), "feholdexcept preserves x87 control settings"); Require((after.sse & ~0x1fbfu) == (before.sse & ~0x1fbfu), "feholdexcept preserves SSE control settings"); Require(after.x87[2] == before.x87[2] && std::memcmp(after.x87.data() + 3, before.x87.data() + 3, 16) == 0, "feholdexcept preserves x87 tags and instruction/data state"); Require(*__error_nid_postfix() == 123, "feholdexcept preserves errno"); } void CheckHoldExcept(HoldExcept hold) { Require(hold != nullptr, "feholdexcept export is missing"); static_assert(sizeof(Environment) == 32 && alignof(Environment) == 4 && offsetof(Environment, sse) == 28); const auto original = ReadEnvironment(); struct { std::array before; Environment value; std::array after; } output; for (unsigned rounding = 0; rounding < 4; ++rounding) { for (unsigned modes = 0; modes < 4; ++modes) { for (unsigned mask = 0; mask < 64; ++mask) { auto state = original; const unsigned sseMask = (mask * 13 + 7) & 0x3f; state.x87[0] = (state.x87[0] & ~0xc3fu) | (rounding << 10) | mask; state.x87[1] = (state.x87[1] & ~0x80ffu) | (mask & 0x25); state.sse = (state.sse & ~0xffffu) | (sseMask << 7) | (sseMask & 0x1a) | (((rounding + 1) & 3) << 13) | ((modes & 1) << 6) | ((modes & 2) << 14); WriteEnvironment(state); std::memset(&output, 0xa5, sizeof(output)); CheckHold(hold, &output.value); for (const auto byte : output.before) Require(byte == 0xa5, "feholdexcept underwrites output"); for (const auto byte : output.after) Require(byte == 0xa5, "feholdexcept overwrites output"); } } } WriteEnvironment(original); bool threw = false; try { hold(nullptr); } catch (const std::invalid_argument&) { threw = true; } Require(threw, "feholdexcept rejects null storage"); const auto afterNull = ReadEnvironment(); Require(SameControlState(original, afterNull), "null feholdexcept preserves the environment"); #ifdef _WIN32 SYSTEM_INFO info{}; GetSystemInfo(&info); const std::size_t page = info.dwPageSize; auto* memory = static_cast(VirtualAlloc(nullptr, page * 2, MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE)); Require(memory != nullptr, "allocate feholdexcept guard pages"); DWORD oldProtection; Require(VirtualProtect(memory + page, page, PAGE_NOACCESS, &oldProtection) != 0, "protect feholdexcept trailing page"); #else const auto page = static_cast(sysconf(_SC_PAGESIZE)); auto* memory = static_cast(mmap(nullptr, page * 2, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0)); Require(memory != MAP_FAILED, "allocate feholdexcept guard pages"); Require(mprotect(memory + page, page, PROT_NONE) == 0, "protect feholdexcept trailing page"); #endif CheckHold(hold, reinterpret_cast(memory + page - sizeof(Environment))); #ifdef _WIN32 Require(VirtualFree(memory, 0, MEM_RELEASE) != 0, "release feholdexcept guard pages"); #else Require(munmap(memory, page * 2) == 0, "release feholdexcept guard pages"); #endif WriteEnvironment(original); std::barrier ready(2); const auto worker = [&](unsigned mode) { const auto saved = ReadEnvironment(); auto state = saved; state.x87[0] = (state.x87[0] & ~0xc00u) | (mode << 10) | 0x3f; state.x87[1] &= ~0x80ffu; state.sse |= 0x1f80; state.sse &= ~0x3fu; WriteEnvironment(state); ready.arrive_and_wait(); Environment output; CheckHold(hold, &output); ready.arrive_and_wait(); Require(fegetround_nid_postfix() == static_cast(mode << 10), "feholdexcept keeps thread-local rounding"); WriteEnvironment(saved); }; std::thread first(worker, 1); std::thread second(worker, 2); first.join(); second.join(); const auto afterThreads = ReadEnvironment(); Require(SameControlState(original, afterThreads), "feholdexcept workers preserve the main environment"); } int main(int argc, char** argv) { Require(argc == 2, "expected libc path"); #ifdef _WIN32 const auto library = LoadLibraryA(argv[1]); Require(library != nullptr, "load libc for feholdexcept"); const auto hold = reinterpret_cast(GetProcAddress(library, "feholdexcept_nid_postfix")); #else const auto library = dlopen(argv[1], RTLD_NOW); Require(library != nullptr, "load libc for feholdexcept"); const auto hold = reinterpret_cast(dlsym(library, "feholdexcept_nid_postfix")); #endif const auto original = ReadEnvironment(); CheckRounding(); CheckEnabledExceptions(); CheckFlags(); CheckTraps(); CheckThreads(); CheckHoldExcept(hold); WriteEnvironment(original); #ifdef _WIN32 FreeLibrary(library); #else dlclose(library); #endif std::puts("floating-point environment checks passed"); }