Files
Ryan Houdek 8ab00758be Merge pull request #5448 from bylaws/claudefix6
X87: Fix FXTRACT with Inf and NaN inputs
2026-04-30 14:42:08 -07:00

674 lines
17 KiB
C++
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
// SPDX-License-Identifier: MIT
#pragma once
#include <FEXCore/Utils/LogManager.h>
#include <FEXCore/fextl/sstream.h>
#include <FEXCore/fextl/string.h>
#include "cephes_128bit.h"
#include <bit>
#include <cmath>
#include <cstring>
#include <stdint.h>
#include "Common/VectorRegType.h"
extern "C" {
#include "SoftFloat-3e/platform.h"
#include "SoftFloat-3e/softfloat.h"
}
struct FEX_PACKED X80SoftFloat {
#ifdef ARCHITECTURE_x86_64
// Define this to push some operations to x87
// Only useful to see if precision loss is killing something
// #define DEBUG_X86_FLOAT
#ifdef DEBUG_X86_FLOAT
#define BIGFLOAT long double
#define BIGFLOATSIZE 10
#else
#define BIGFLOAT float128_t
#define BIGFLOATSIZE 16
#endif
#elif defined(ARCHITECTURE_arm64)
#define BIGFLOAT float128_t
#define BIGFLOATSIZE 16
#else
#error No 128bit float for this target!
#endif
uint64_t Significand;
union {
uint16_t Raw;
struct {
uint16_t Exponent : 15;
uint16_t Sign : 1;
};
} Top;
X80SoftFloat() {
memset(this, 0, sizeof(*this));
}
X80SoftFloat(uint16_t _Sign, uint16_t _Exponent, uint64_t _Significand)
: Significand {_Significand}
, Top {.Raw = static_cast<uint16_t>((_Exponent & 0x7FFF) | (_Sign << 15))} {}
fextl::string str() const {
fextl::ostringstream string;
string << std::hex << Top.Sign;
string << "_" << Top.Exponent;
string << "_" << (Significand >> 63);
string << "_" << (Significand & ((1ULL << 63) - 1));
return string.str();
}
// Ops
FEXCORE_PRESERVE_ALL_ATTR static X80SoftFloat FADD(softfloat_state* state, const X80SoftFloat& lhs, const X80SoftFloat& rhs) {
#ifdef DEBUG_X86_FLOAT
BIGFLOAT Result;
asm(R"(
fninit;
fldt %[rhs]; # st1
fldt %[lhs]; # st0
faddp;
fstpt %[result];
)"
: [result] "=m"(Result)
: [lhs] "m"(lhs), [rhs] "m"(rhs)
: "st", "st(1)");
return Result;
#else
return extF80_add(state, lhs, rhs);
#endif
}
FEXCORE_PRESERVE_ALL_ATTR static X80SoftFloat FSUB(softfloat_state* state, const X80SoftFloat& lhs, const X80SoftFloat& rhs) {
#ifdef DEBUG_X86_FLOAT
BIGFLOAT Result;
asm(R"(
fninit;
fldt %[rhs]; # st1
fldt %[lhs]; # st0
fsubp;
fstpt %[result];
)"
: [result] "=m"(Result)
: [lhs] "m"(lhs), [rhs] "m"(rhs)
: "st", "st(1)");
return Result;
#else
return extF80_sub(state, lhs, rhs);
#endif
}
FEXCORE_PRESERVE_ALL_ATTR static X80SoftFloat FMUL(softfloat_state* state, const X80SoftFloat& lhs, const X80SoftFloat& rhs) {
#ifdef DEBUG_X86_FLOAT
BIGFLOAT Result;
asm(R"(
fninit;
fldt %[rhs]; # st1
fldt %[lhs]; # st0
fmulp;
fstpt %[result];
)"
: [result] "=m"(Result)
: [lhs] "m"(lhs), [rhs] "m"(rhs)
: "st", "st(1)");
return Result;
#else
return extF80_mul(state, lhs, rhs);
#endif
}
FEXCORE_PRESERVE_ALL_ATTR static X80SoftFloat FDIV(softfloat_state* state, const X80SoftFloat& lhs, const X80SoftFloat& rhs) {
#ifdef DEBUG_X86_FLOAT
BIGFLOAT Result;
asm(R"(
fninit;
fldt %[rhs]; # st1
fldt %[lhs]; # st0
fdivp;
fstpt %[result];
)"
: [result] "=m"(Result)
: [lhs] "m"(lhs), [rhs] "m"(rhs)
: "st", "st(1)");
return Result;
#else
return extF80_div(state, lhs, rhs);
#endif
}
FEXCORE_PRESERVE_ALL_ATTR static X80SoftFloat FREM(softfloat_state* state, const X80SoftFloat& lhs, const X80SoftFloat& rhs) {
#if defined(DEBUG_X86_FLOAT)
BIGFLOAT Result;
asm(R"(
fninit;
fldt %[rhs]; # st1
fldt %[lhs]; # st0
fprem;
fstpt %[result];
ffreep %%st(0);
)"
: [result] "=m"(Result)
: [lhs] "m"(lhs), [rhs] "m"(rhs)
: "st", "st(1)");
return Result;
#else
/*
* Check for invalid operation cases first - Intel FPREM sets Invalid Operation
* for several cases including infinity dividend and zero divisor.
*/
X80SoftFloat result = 0;
if (HandleInfinityOp(state, lhs, result)) {
return result;
} else if (lhs.Top.Exponent == 0x7FFF && (lhs.Significand & 0x7FFFFFFFFFFFFFFFULL)) { // NaN
// propagate NaN
state->exceptionFlags |= softfloat_flag_invalid;
return lhs;
}
// Check for zero divisor - fprem(x, 0) is invalid operation
if (rhs.Top.Exponent == 0 && rhs.Significand == 0) {
state->exceptionFlags |= softfloat_flag_invalid;
// Return QNaN
result.Top.Sign = 0;
result.Top.Exponent = 0x7FFF;
result.Significand = 0xC000000000000000ULL;
return result;
}
/*
* FPREM is not an IEEE-754 remainder. From the Intel spec:
*
* Computes the remainder obtained from dividing the value in the ST(0)
* register (the dividend) by the value in the ST(1) register (the divisor
* or modulus), and stores the result in ST(0). The remainder represents the
* following value:
*
* Remainder := ST(0) − (Q * ST(1))
*
* Here, Q is an integer value that is obtained by truncating the
* floating-point number quotient of [ST(0) / ST(1)] toward zero.
*
* We implement this sequence literally. softfloat_round_minMag means
* "truncate towards zero".
*/
extFloat80_t quotient = extF80_div(state, lhs, rhs);
extFloat80_t Q = extF80_roundToInt(state, quotient, softfloat_round_minMag, true);
bool Q_zero = Q.signif == 0 && (Q.signExp & ~(1 << 15)) == 0;
if (Q_zero) {
return lhs;
} else {
return extF80_sub(state, lhs, extF80_mul(state, Q, rhs));
}
#endif
}
FEXCORE_PRESERVE_ALL_ATTR static X80SoftFloat FREM1(softfloat_state* state, const X80SoftFloat& lhs, const X80SoftFloat& rhs) {
#if defined(DEBUG_X86_FLOAT)
BIGFLOAT Result;
asm(R"(
fninit;
fldt %[rhs]; # st1
fldt %[lhs]; # st0
fprem1;
fstpt %[result];
ffreep %%st(0);
)"
: [result] "=m"(Result)
: [lhs] "m"(lhs), [rhs] "m"(rhs)
: "st", "st(1)");
return Result;
#else
return extF80_rem(state, lhs, rhs);
#endif
}
FEXCORE_PRESERVE_ALL_ATTR static X80SoftFloat FRNDINT(softfloat_state* state, const X80SoftFloat& lhs) {
return extF80_roundToInt(state, lhs, state->roundingMode, false);
}
FEXCORE_PRESERVE_ALL_ATTR static X80SoftFloat FRNDINT(softfloat_state* state, const X80SoftFloat& lhs, uint_fast8_t RoundMode) {
return extF80_roundToInt(state, lhs, RoundMode, false);
}
FEXCORE_PRESERVE_ALL_ATTR static X80SoftFloat FXTRACT_SIG(const X80SoftFloat& lhs) {
#if defined(DEBUG_X86_FLOAT)
BIGFLOAT Result;
asm(R"(
fninit;
fldt %[lhs]; # st0
fxtract;
fstpt %[result];
ffreep %%st(0);
)"
: [result] "=m"(Result)
: [lhs] "m"(lhs)
: "st", "st(1)");
return Result;
#else
// Zero is a special case, the significand for +/- 0 is +/- zero.
if (lhs.Top.Exponent == 0x0 && lhs.Significand == 0x0) {
return lhs;
}
// Inf/NaN pass through unchanged in the significand slot.
if (lhs.Top.Exponent == 0x7FFF) {
return lhs;
}
X80SoftFloat Tmp = lhs;
Tmp.Top.Exponent = 0x3FFF;
Tmp.Top.Sign = lhs.Top.Sign;
return Tmp;
#endif
}
FEXCORE_PRESERVE_ALL_ATTR static X80SoftFloat FXTRACT_EXP(const X80SoftFloat& lhs) {
#if defined(DEBUG_X86_FLOAT)
BIGFLOAT Result;
asm(R"(
fninit;
fldt %[lhs]; # st0
fxtract;
ffreep %%st(0);
fstpt %[result];
)"
: [result] "=m"(Result)
: [lhs] "m"(lhs)
: "st", "st(1)");
return Result;
#else
// Zero is a special case, the exponent is always -inf
if (lhs.Top.Exponent == 0x0 && lhs.Significand == 0x0) {
X80SoftFloat Result(1, 0x7FFFUL, 0x8000'0000'0000'0000UL);
return Result;
}
// +/-Inf returns +Inf in the exponent slot; NaN propagates.
if (lhs.Top.Exponent == 0x7FFF) {
if ((lhs.Significand & 0x7FFFFFFFFFFFFFFFULL) == 0) {
X80SoftFloat Result(0, 0x7FFFUL, 0x8000'0000'0000'0000UL);
return Result;
}
return lhs;
}
int32_t TrueExp = lhs.Top.Exponent - ExponentBias;
return i32_to_extF80(TrueExp);
#endif
}
FEXCORE_PRESERVE_ALL_ATTR static void
FCMP(softfloat_state* state, const X80SoftFloat& lhs, const X80SoftFloat& rhs, bool* eq, bool* lt, bool* nan) {
*eq = extF80_eq(state, lhs, rhs);
*lt = extF80_lt(state, lhs, rhs);
// Use IEEE 754 semantics: unordered if neither <, =, nor > is true
// This is more reliable than custom NaN detection
bool gt = !(*eq) && !(*lt) && extF80_le(state, rhs, lhs);
*nan = !(*eq) && !(*lt) && !gt;
}
FEXCORE_PRESERVE_ALL_ATTR static X80SoftFloat FSCALE(softfloat_state* state, const X80SoftFloat& lhs, const X80SoftFloat& rhs) {
#ifdef DEBUG_X86_FLOAT
BIGFLOAT Result;
asm(R"(
fninit;
fldt %[rhs]; # st1
fldt %[lhs]; # st0
fscale; # st0 = st0 * 2^(rdint(st1))
fstpt %[result];
ffreep %%st(0);
)"
: [result] "=m"(Result)
: [lhs] "m"(lhs), [rhs] "m"(rhs)
: "st", "st(1)");
return Result;
#else
extFloat80_t Zero {0, 0};
if (extF80_eq(state, lhs, Zero)) {
// FSCALE(0, +Inf) is 0 * Inf, which is invalid. FSCALE(0, anything
// else) is still 0.
if (rhs.Top.Exponent == 0x7FFF && rhs.Top.Sign == 0 && (rhs.Significand & 0x7FFFFFFFFFFFFFFFULL) == 0) {
state->exceptionFlags |= softfloat_flag_invalid;
X80SoftFloat QNaN(0, 0x7FFFUL, 0xC000000000000000ULL);
return QNaN;
}
return lhs;
}
X80SoftFloat Int = FRNDINT(state, rhs, softfloat_round_minMag);
BIGFLOAT Src2_d = Int.ToFMax(state);
Src2_d = FEXCore::cephes_128bit::exp2l(Src2_d);
X80SoftFloat Src2_X80(state, Src2_d);
X80SoftFloat Result = extF80_mul(state, lhs, Src2_X80);
return Result;
#endif
}
FEXCORE_PRESERVE_ALL_ATTR static X80SoftFloat F2XM1(softfloat_state* state, const X80SoftFloat& lhs) {
#ifdef DEBUG_X86_FLOAT
BIGFLOAT Result;
asm(R"(
fninit;
fldt %[lhs]; # st0
f2xm1; # st0 = 2^st(0) - 1
fstpt %[result];
)"
: [result] "=m"(Result)
: [lhs] "m"(lhs)
: "st");
return Result;
#else
auto Src1_d = lhs.ToFMax(state);
auto Result = FEXCore::cephes_128bit::exp2l(Src1_d);
static const float128_t one {0x0ULL, 0x3fff000000000000ULL};
return X80SoftFloat(state, f128_sub(state, Result, one));
#endif
}
FEXCORE_PRESERVE_ALL_ATTR static X80SoftFloat FYL2X(softfloat_state* state, const X80SoftFloat& lhs, const X80SoftFloat& rhs) {
#ifdef DEBUG_X86_FLOAT
BIGFLOAT Result;
asm(R"(
fninit;
fldt %[rhs]; # st(1)
fldt %[lhs]; # st(0)
fyl2x; # st(1) * log2l(st(0))
fstpt %[result];
)"
: [result] "=m"(Result)
: [lhs] "m"(lhs), [rhs] "m"(rhs)
: "st", "st(1)");
return Result;
#else
auto Src1_d = lhs.ToFMax(state);
auto Src2_d = rhs.ToFMax(state);
auto Tmp = f128_mul(state, Src2_d, FEXCore::cephes_128bit::log2l(Src1_d));
return X80SoftFloat(state, Tmp);
#endif
}
FEXCORE_PRESERVE_ALL_ATTR static X80SoftFloat FATAN(softfloat_state* state, const X80SoftFloat& lhs, const X80SoftFloat& rhs) {
#ifdef DEBUG_X86_FLOAT
BIGFLOAT Result;
asm(R"(
fninit;
fldt %[lhs];
fldt %[rhs];
fpatan;
fstpt %[result];
)"
: [result] "=m"(Result)
: [lhs] "m"(lhs), [rhs] "m"(rhs)
: "st", "st(1)");
return Result;
#else
BIGFLOAT Src1_d = lhs.ToFMax(state);
BIGFLOAT Src2_d = rhs.ToFMax(state);
BIGFLOAT Tmp = FEXCore::cephes_128bit::atan2l(Src1_d, Src2_d);
return X80SoftFloat(state, Tmp);
#endif
}
FEXCORE_PRESERVE_ALL_ATTR static X80SoftFloat FTAN(softfloat_state* state, const X80SoftFloat& lhs) {
#ifdef DEBUG_X86_FLOAT
BIGFLOAT Result;
asm(R"(
fninit;
fldt %[lhs]; # st0
fptan;
ffreep %%st(0);
fstpt %[result];
)"
: [result] "=m"(Result)
: [lhs] "m"(lhs)
: "st");
return Result;
#else
X80SoftFloat result;
if (HandleInfinityOp(state, lhs, result)) {
return result;
}
BIGFLOAT Src_d = lhs.ToFMax(state);
Src_d = FEXCore::cephes_128bit::tanl(Src_d);
return X80SoftFloat(state, Src_d);
#endif
}
FEXCORE_PRESERVE_ALL_ATTR static X80SoftFloat FSIN(softfloat_state* state, const X80SoftFloat& lhs) {
#ifdef DEBUG_X86_FLOAT
BIGFLOAT Result;
asm(R"(
fninit;
fldt %[lhs]; # st0
fsin;
fstpt %[result];
)"
: [result] "=m"(Result)
: [lhs] "m"(lhs)
: "st");
return Result;
#else
X80SoftFloat result;
if (HandleInfinityOp(state, lhs, result)) {
return result;
}
BIGFLOAT Src_d = lhs.ToFMax(state);
Src_d = FEXCore::cephes_128bit::sinl(Src_d);
return X80SoftFloat(state, Src_d);
#endif
}
FEXCORE_PRESERVE_ALL_ATTR static X80SoftFloat FCOS(softfloat_state* state, const X80SoftFloat& lhs) {
#ifdef DEBUG_X86_FLOAT
BIGFLOAT Result;
asm(R"(
fninit;
fldt %[lhs]; # st0
fcos;
fstpt %[result];
)"
: [result] "=m"(Result)
: [lhs] "m"(lhs)
: "st");
return Result;
#else
X80SoftFloat result;
if (HandleInfinityOp(state, lhs, result)) {
return result;
}
BIGFLOAT Src_d = lhs.ToFMax(state);
Src_d = FEXCore::cephes_128bit::cosl(Src_d);
return X80SoftFloat(state, Src_d);
#endif
}
FEXCORE_PRESERVE_ALL_ATTR static X80SoftFloat FSQRT(softfloat_state* state, const X80SoftFloat& lhs) {
#ifdef DEBUG_X86_FLOAT
BIGFLOAT Result;
asm(R"(
fninit;
fldt %[lhs]; # st0
fsqrt;
fstpt %[result];
)"
: [result] "=m"(Result)
: [lhs] "m"(lhs)
: "st");
return Result;
#else
return extF80_sqrt(state, lhs);
#endif
}
float ToF32(softfloat_state* state) const {
const float32_t Result = extF80_to_f32(state, *this);
return std::bit_cast<float>(Result);
}
double ToF64(softfloat_state* state) const {
const float64_t Result = extF80_to_f64(state, *this);
return std::bit_cast<double>(Result);
}
FEXCore::VectorRegType ToVector() const {
FEXCore::VectorRegType Ret {};
memcpy(&Ret, this, sizeof(*this));
return Ret;
}
BIGFLOAT ToFMax(softfloat_state* state) const {
#if BIGFLOATSIZE == 16
const float128_t Result = extF80_to_f128(state, *this);
return std::bit_cast<BIGFLOAT>(Result);
#else
BIGFLOAT result {};
memcpy(&result, this, sizeof(result));
return result;
#endif
}
int16_t ToI16(softfloat_state* state) const {
auto rv = extF80_to_i32(state, *this, state->roundingMode, false);
if (rv > INT16_MAX || rv < INT16_MIN) {
///< Indefinite value for 16-bit conversions.
return INT16_MIN;
} else {
return rv;
}
}
int32_t ToI32(softfloat_state* state) const {
return extF80_to_i32(state, *this, state->roundingMode, false);
}
int64_t ToI64(softfloat_state* state) const {
return extF80_to_i64(state, *this, state->roundingMode, false);
}
uint64_t ToUI64(softfloat_state* state) const {
return extF80_to_ui64(state, *this, state->roundingMode, false);
}
void operator=(const int16_t rhs) {
*this = i32_to_extF80(rhs);
}
void operator=(const int32_t rhs) {
*this = i32_to_extF80(rhs);
}
void operator=(const uint64_t rhs) {
*this = ui64_to_extF80(rhs);
}
#if BIGFLOATSIZE == 10
void operator=(const long double rhs) {
memcpy(this, &rhs, sizeof(rhs));
}
#endif
operator void*() {
return reinterpret_cast<void*>(this);
}
X80SoftFloat(extFloat80_t rhs) {
Significand = rhs.signif;
Top.Raw = rhs.signExp;
}
X80SoftFloat(softfloat_state* state, const float rhs) {
*this = f32_to_extF80(state, std::bit_cast<float32_t>(rhs));
}
X80SoftFloat(softfloat_state* state, const double rhs) {
*this = f64_to_extF80(state, std::bit_cast<float64_t>(rhs));
}
X80SoftFloat(softfloat_state* state, BIGFLOAT rhs) {
#if BIGFLOATSIZE == 16
*this = f128_to_extF80(state, std::bit_cast<float128_t>(rhs));
#else
*this = std::bit_cast<long double>(rhs);
#endif
}
X80SoftFloat(const int16_t rhs) {
*this = i32_to_extF80(rhs);
}
X80SoftFloat(const int32_t rhs) {
*this = i32_to_extF80(rhs);
}
X80SoftFloat(const FEXCore::VectorRegType rhs) {
memcpy(this, &rhs, sizeof(*this));
}
void operator=(extFloat80_t rhs) {
Significand = rhs.signif;
Top.Raw = rhs.signExp;
}
operator FEXCore::VectorRegType() const {
return ToVector();
}
operator extFloat80_t() const {
extFloat80_t Result {};
Result.signif = Significand;
Result.signExp = Top.Raw;
return Result;
}
static bool IsNan(const X80SoftFloat& lhs) {
return (lhs.Top.Exponent == 0x7FFF) && (lhs.Significand & IntegerBit) && (lhs.Significand & Bottom62Significand);
}
static bool SignBit(const X80SoftFloat& lhs) {
return lhs.Top.Sign;
}
private:
static constexpr uint64_t IntegerBit = (1ULL << 63);
static constexpr uint64_t Bottom62Significand = ((1ULL << 62) - 1);
static constexpr uint32_t ExponentBias = 16383;
// Helper function to check for infinity and set invalid operation flag.
// Returns true if infinity is dealt with, false otherwise.
FEXCORE_PRESERVE_ALL_ATTR static bool HandleInfinityOp(softfloat_state* state, const X80SoftFloat& arg, X80SoftFloat& result) {
if (arg.Top.Exponent == 0x7FFF && arg.Significand == 0x8000000000000000ULL) {
state->exceptionFlags |= softfloat_flag_invalid;
// Return QNaN.
result.Top.Sign = 0;
result.Top.Exponent = 0x7FFF;
result.Significand = 0xC000000000000000ULL;
return true;
}
return false;
}
};
static_assert(sizeof(X80SoftFloat) == 10, "tword must be 10bytes in size");