double
ppc_estimate_reciprocal(double v)
{
auto bits = get_float_bits(v);
if (bits.mantissa == 0 && bits.exponent == 0) {
return std::copysign(std::numeric_limits<double>::infinity(), v);
}
if (bits.exponent == bits.exponent_max) {
if (bits.mantissa == 0) {
return std::copysign(0.0, v);
}
return static_cast<float>(v);
}
if (bits.exponent < 895) {
return std::copysign(std::numeric_limits<float>::max(), v);
}
if (bits.exponent > 1149) {
return std::copysign(0.0, v);
}
int idx = (int)(bits.mantissa >> 37);
bits.exponent = 0x7FD - bits.exponent;
bits.mantissa = (int64_t)(fres_expected_base[idx / 1024] - (fres_expected_dec[idx / 1024] * (idx % 1024) + 1) / 2) << 29;
return bits.v;
}
static void
mergeGeneric(cpu::Core *state, Instruction instr)
{
float d0, d1;
if (flags & MergeValue0) {
if (!is_signalling_nan(state->fpr[instr.frA].paired1)) {
d0 = static_cast<float>(state->fpr[instr.frA].paired1);
} else {
d0 = truncate_double(state->fpr[instr.frA].paired1);
}
} else {
if (!is_signalling_nan(state->fpr[instr.frA].paired0)) {
d0 = static_cast<float>(state->fpr[instr.frA].paired0);
} else {
d0 = truncate_double(state->fpr[instr.frA].paired0);
}
}
// When inserting a double-precision value into slot 1, the value is
// truncated rather than rounded.
double d1_double;
if (flags & MergeValue1) {
d1_double = state->fpr[instr.frB].paired1;
} else {
d1_double = state->fpr[instr.frB].paired0;
}
auto d1_bits = get_float_bits(d1_double);
if (d1_bits.exponent >= 1151 && d1_bits.exponent < 2047) {
d1 = std::numeric_limits<float>::max();
} else {
d1 = truncate_double(d1_double);
}
state->fpr[instr.frD].paired0 = extend_float(d0);
state->fpr[instr.frD].paired1 = extend_float(d1);
// Don't leak any exceptions (inexact, overflow etc.) to later instructions.
std::feclearexcept(FE_ALL_EXCEPT);
if (instr.rc) {
updateFloatConditionRegister(state);
}
}
// Reciprocal Square Root
static void
ps_rsqrte(cpu::Core *state, Instruction instr)
{
const double b0 = state->fpr[instr.frB].paired0;
const double b1 = state->fpr[instr.frB].paired1;
const bool vxsnan0 = is_signalling_nan(b0);
const bool vxsnan1 = is_signalling_nan(b1);
const bool vxsqrt0 = !vxsnan0 && std::signbit(b0) && !is_zero(b0);
const bool vxsqrt1 = !vxsnan1 && std::signbit(b1) && !is_zero(b1);
const bool zx0 = is_zero(b0);
const bool zx1 = is_zero(b1);
const uint32_t oldFPSCR = state->fpscr.value;
state->fpscr.vxsnan |= vxsnan0 || vxsnan1;
state->fpscr.vxsqrt |= vxsqrt0 || vxsqrt1;
state->fpscr.zx |= zx0 || zx1;
double d0, d1;
bool write = true;
if (((vxsnan0 || vxsqrt0) && state->fpscr.ve) || (zx0 && state->fpscr.ze)) {
write = false;
} else {
d0 = ppc_estimate_reciprocal_root(b0);
updateFPRF(state, d0);
}
if (((vxsnan1 || vxsqrt1) && state->fpscr.ve) || (zx1 && state->fpscr.ze)) {
write = false;
} else {
d1 = ppc_estimate_reciprocal_root(b1);
}
if (write) {
// ps_rsqrte behaves strangely when the result's magnitude is out of
// range: ps0 keeps its double-precision exponent, while ps1 appears
// to get an arbitrary value from the floating-point circuitry. The
// details of how ps1's exponent is affected are unknown, but the
// logic below works for double-precision inputs 0x7FE...FFF (maximum
// normal) and 0x000...001 (minimum denormal).
auto bits0 = get_float_bits(d0);
bits0.mantissa &= UINT64_C(0xFFFFFE0000000);
state->fpr[instr.frD].paired0 = bits0.v;
auto bits1 = get_float_bits(d1);
if (bits1.exponent == 0) {
// Leave as zero (reciprocal square root can never be a denormal).
} else if (bits1.exponent < 1151) {
int8_t exponent8 = (bits1.exponent - 1023) & 0xFF;
bits1.exponent = 1023 + exponent8;
} else if (bits1.exponent < 2047) {
bits1.exponent = 1022;
}
bits1.mantissa &= UINT64_C(0xFFFFFE0000000);
state->fpr[instr.frD].paired1 = bits1.v;
}
updateFPSCR(state, oldFPSCR);
if (instr.rc) {
updateFloatConditionRegister(state);
}
}
bool runTests(const std::string &path)
{
uint32_t testsFailed = 0, testsPassed = 0;
uint32_t baseAddress = mem::MEM2Base;
Instruction bclr = encodeInstruction(InstructionID::bclr);
bclr.bo = 0x1f;
mem::write(baseAddress + 4, bclr.value);
fs::FileSystem filesystem;
fs::FolderEntry entry;
fs::HostPath base = path;
filesystem.mountHostFolder("/tests", base, fs::Permissions::Read);
auto folder = filesystem.openFolder("/tests");
while (folder->read(entry)) {
std::ifstream file(base.join(entry.name).path(), std::ifstream::in | std::ifstream::binary);
cereal::BinaryInputArchive cerealInput(file);
TestFile testFile;
// Parse test file with cereal
testFile.name = entry.name;
cerealInput(testFile);
// Run tests
gLog->info("Checking {}", testFile.name);
for (auto &test : testFile.tests) {
bool failed = false;
if (!TEST_FMADDSUB) {
auto data = espresso::decodeInstruction(test.instr);
switch (data->id) {
case InstructionID::fmadd:
case InstructionID::fmadds:
case InstructionID::fmsub:
case InstructionID::fmsubs:
case InstructionID::fnmadd:
case InstructionID::fnmadds:
case InstructionID::fnmsub:
case InstructionID::fnmsubs:
failed = true;
break;
}
if (failed) {
continue;
}
}
// Setup core state from test input
cpu::CoreRegs *state = cpu::this_core::state();
memset(state, 0, sizeof(cpu::CoreRegs));
state->cia = 0;
state->nia = baseAddress;
state->xer = test.input.xer;
state->cr = test.input.cr;
state->fpscr = test.input.fpscr;
state->ctr = test.input.ctr;
for (auto i = 0; i < 4; ++i) {
state->gpr[i + 3] = test.input.gpr[i];
state->fpr[i + 1].paired0 = test.input.fr[i];
}
// Execute test
mem::write(baseAddress, test.instr.value);
cpu::jit::clearCache();
cpu::this_core::executeSub();
// Check XER (all bits)
if (state->xer.value != test.output.xer.value) {
gLog->error("Test failed, xer expected {:08X} found {:08X}", test.output.xer.value, state->xer.value);
failed = true;
}
// Check Condition Register (all bits)
if (state->cr.value != test.output.cr.value) {
gLog->error("Test failed, cr expected {:08X} found {:08X}", test.output.cr.value, state->cr.value);
failed = true;
}
// Check FPSCR
if (TEST_FPSCR) {
if (!TEST_FPSCR_FR) {
state->fpscr.fr = 0;
test.output.fpscr.fr = 0;
}
if (!TEST_FPSCR_UX) {
state->fpscr.ux = 0;
test.output.fpscr.ux = 0;
}
auto state_fpscr = state->fpscr.value;
auto test_fpscr = test.output.fpscr.value;
if (state_fpscr != test_fpscr) {
gLog->error("Test failed, fpscr {:08X} found {:08X}", test.output.fpscr.value, state->fpscr.value);
compareFPSCR(test.input.fpscr, state->fpscr, test.output.fpscr);
failed = true;
}
}
// Check CTR
if (state->ctr != test.output.ctr) {
gLog->error("Test failed, ctr expected {:08X} found {:08X}", test.output.ctr, state->ctr);
failed = true;
}
// Check all GPR
for (auto i = 0; i < 4; ++i) {
auto reg = i + hwtest::GPR_BASE;
auto value = state->gpr[reg];
auto expected = test.output.gpr[i];
if (value != expected) {
gLog->error("Test failed, r{} expected {:08X} found {:08X}", reg, expected, value);
failed = true;
}
}
// Check all FPR
for (auto i = 0; i < 4; ++i) {
auto reg = i + hwtest::FPR_BASE;
auto value = state->fpr[reg].value;
auto expected = test.output.fr[i];
if (!is_nan(value) && !is_nan(expected) && !is_infinity(value) && !is_infinity(expected)) {
double dval = value / expected;
if (dval < 0.999 || dval > 1.001) {
gLog->error("Test failed, f{} expected {:16f} found {:16f}", reg, expected, value);
failed = true;
}
} else {
if (is_nan(value) && is_nan(expected)) {
auto bits = get_float_bits(value);
bits.sign = get_float_bits(expected).sign;
value = bits.v;
}
if (bit_cast<uint64_t>(value) != bit_cast<uint64_t>(expected)) {
gLog->error("Test failed, f{} expected {:16X} found {:16X}", reg, bit_cast<uint64_t>(expected), bit_cast<uint64_t>(value));
failed = true;
}
}
}
if (failed) {
Disassembly dis;
// Print disassembly
disassemble(test.instr, dis, baseAddress);
gLog->debug(dis.text);
// Print all test fields
gLog->debug("{:08x} Input Hardware Interp", test.instr.value);
for (auto field : dis.instruction->read) {
printTestField(field, test.instr, &test.input, &test.output, state);
}
for (auto field : dis.instruction->write) {
printTestField(field, test.instr, &test.input, &test.output, state);
}
for (auto field : dis.instruction->flags) {
printTestField(field, test.instr, &test.input, &test.output, state);
}
gLog->debug("");
++testsFailed;
} else {
++testsPassed;
}
}
}
gLog->info("Passed: {}, Failed: {}", testsPassed, testsFailed);
return true;
}
int runTests(const std::string &path)
{
uint32_t testsFailed = 0, testsPassed = 0;
auto baseAddress = cpu::VirtualAddress { 0x02000000u };
auto basePhysicalAddress = cpu::PhysicalAddress { 0x50000000u };
auto codeSize = 2048u;
cpu::allocateVirtualAddress(baseAddress, codeSize);
cpu::mapMemory(baseAddress, basePhysicalAddress, codeSize, cpu::MapPermission::ReadWrite);
Instruction bclr = encodeInstruction(InstructionID::bclr);
bclr.bo = 0x1f;
mem::write(baseAddress.getAddress() + 4, bclr.value);
auto ec = std::error_code { };
for (auto itr = std::filesystem::directory_iterator { "/tests", ec }; itr != end(itr); ++itr) {
std::ifstream file(itr->path().string(), std::ifstream::in | std::ifstream::binary);
cereal::BinaryInputArchive cerealInput(file);
TestFile testFile;
// Parse test file with cereal
testFile.name = itr->path().filename().string();
cerealInput(testFile);
// Run tests
for (auto &test : testFile.tests) {
bool failed = false;
if (!TEST_FMADDSUB) {
auto data = espresso::decodeInstruction(test.instr);
switch (data->id) {
case InstructionID::fmadd:
case InstructionID::fmadds:
case InstructionID::fmsub:
case InstructionID::fmsubs:
case InstructionID::fnmadd:
case InstructionID::fnmadds:
case InstructionID::fnmsub:
case InstructionID::fnmsubs:
failed = true;
break;
}
if (failed) {
continue;
}
}
// Setup core state from test input
cpu::CoreRegs *state = cpu::this_core::state();
memset(state, 0, sizeof(cpu::CoreRegs));
state->cia = 0;
state->nia = baseAddress.getAddress();
state->xer = test.input.xer;
state->cr = test.input.cr;
state->fpscr = test.input.fpscr;
state->ctr = test.input.ctr;
for (auto i = 0; i < 4; ++i) {
state->gpr[i + 3] = test.input.gpr[i];
state->fpr[i + 1].paired0 = test.input.fr[i];
}
// Execute test
mem::write(baseAddress.getAddress(), test.instr.value);
cpu::clearInstructionCache();
cpu::this_core::executeSub();
// Check XER (all bits)
if (state->xer.value != test.output.xer.value) {
gLog->error("Test failed, xer expected {:08X} found {:08X}", test.output.xer.value, state->xer.value);
failed = true;
}
// Check Condition Register (all bits)
if (state->cr.value != test.output.cr.value) {
gLog->error("Test failed, cr expected {:08X} found {:08X}", test.output.cr.value, state->cr.value);
failed = true;
}
// Check FPSCR
if (TEST_FPSCR) {
if (!TEST_FPSCR_FR) {
state->fpscr.fr = 0;
test.output.fpscr.fr = 0;
}
if (!TEST_FPSCR_UX) {
state->fpscr.ux = 0;
test.output.fpscr.ux = 0;
}
auto state_fpscr = state->fpscr.value;
auto test_fpscr = test.output.fpscr.value;
if (state_fpscr != test_fpscr) {
gLog->error("Test failed, fpscr {:08X} found {:08X}", test.output.fpscr.value, state->fpscr.value);
compareFPSCR(test.input.fpscr, state->fpscr, test.output.fpscr);
failed = true;
}
}
// Check CTR
if (state->ctr != test.output.ctr) {
gLog->error("Test failed, ctr expected {:08X} found {:08X}", test.output.ctr, state->ctr);
failed = true;
}
// Check all GPR
for (auto i = 0; i < 4; ++i) {
auto reg = i + hwtest::GPR_BASE;
auto value = state->gpr[reg];
auto expected = test.output.gpr[i];
if (value != expected) {
gLog->error("Test failed, r{} expected {:08X} found {:08X}", reg, expected, value);
failed = true;
}
}
// Check all FPR
for (auto i = 0; i < 4; ++i) {
auto reg = i + hwtest::FPR_BASE;
auto value = state->fpr[reg].value;
auto expected = test.output.fr[i];
if (!is_nan(value) && !is_nan(expected) && !is_infinity(value) && !is_infinity(expected)) {
double dval = value / expected;
if (dval < 0.999 || dval > 1.001) {
gLog->error("Test failed, f{} expected {:16f} found {:16f}", reg, expected, value);
failed = true;
}
} else {
if (is_nan(value) && is_nan(expected)) {
auto bits = get_float_bits(value);
bits.sign = get_float_bits(expected).sign;
value = bits.v;
}
if (bit_cast<uint64_t>(value) != bit_cast<uint64_t>(expected)) {
gLog->error("Test failed, f{} expected {:16X} found {:16X}", reg, bit_cast<uint64_t>(expected), bit_cast<uint64_t>(value));
failed = true;
}
}
}
if (failed) {
Disassembly dis;
// Print disassembly
disassemble(test.instr, dis, baseAddress.getAddress());
gLog->debug(dis.text);
// Print all test fields
gLog->debug("{:08x} Input Hardware Interp", test.instr.value);
for (auto field : dis.instruction->read) {
printTestField(field, test.instr, &test.input, &test.output, state);
}
for (auto field : dis.instruction->write) {
printTestField(field, test.instr, &test.input, &test.output, state);
}
for (auto field : dis.instruction->flags) {
printTestField(field, test.instr, &test.input, &test.output, state);
}
gLog->debug("");
++testsFailed;
} else {
++testsPassed;
}
}
}
if (testsFailed) {
gLog->error("Failed {} of {} tests.", testsFailed, testsFailed + testsPassed);
}
return testsFailed;
}