65816-llvm-mos/runtime/src/math.c
Scott Duensing bb3aad3911 Checkpoint
2026-05-03 23:10:41 -05:00

753 lines
20 KiB
C

// Minimal math.h subset for the W65816 runtime.
//
// All operations work directly on the IEEE-754 bit representation
// rather than going through long sequences of soft-float libcalls.
// fabs/copysign/floor/ceil/fmod cover the common needs of routine
// numeric code; transcendentals are intentionally out of scope.
//
// Layout assumed:
// double = sign(1) | exp(11) | frac(52) stored little-endian
// (high word at +6, low word at +0)
// float = sign(1) | exp(8) | frac(23)
//
// We type-pun via __builtin_memcpy + uint64_t / uint32_t to stay
// strict-aliasing-clean.
typedef unsigned long uint32_t;
typedef unsigned long long uint64_t;
typedef long long int64_t;
static uint64_t dToBits(double v) {
uint64_t b;
__builtin_memcpy(&b, &v, sizeof(b));
return b;
}
static double dFromBits(uint64_t b) {
double v;
__builtin_memcpy(&v, &b, sizeof(v));
return v;
}
static uint32_t fToBits(float v) {
uint32_t b;
__builtin_memcpy(&b, &v, sizeof(b));
return b;
}
static float fFromBits(uint32_t b) {
float v;
__builtin_memcpy(&v, &b, sizeof(v));
return v;
}
double fabs(double x) {
return dFromBits(dToBits(x) & ~((uint64_t)1 << 63));
}
float fabsf(float x) {
return fFromBits(fToBits(x) & ~((uint32_t)1 << 31));
}
double copysign(double x, double y) {
uint64_t mask = (uint64_t)1 << 63;
return dFromBits((dToBits(x) & ~mask) | (dToBits(y) & mask));
}
float copysignf(float x, float y) {
uint32_t mask = (uint32_t)1 << 31;
return fFromBits((fToBits(x) & ~mask) | (fToBits(y) & mask));
}
// floor(x) = round toward -inf. Truncate the fraction bits below
// the integer position; if any non-zero bit was discarded for a
// negative value, subtract one (round more negative).
double floor(double x) {
uint64_t b = dToBits(x);
int e = (int)((b >> 52) & 0x7FF) - 1023;
if (e < 0) {
// |x| < 1: floor(positive) = 0, floor(negative) = -1.
return (b >> 63) ? -1.0 : 0.0;
}
if (e >= 52) {
return x; // already integer (or NaN/Inf)
}
uint64_t fracMask = ((uint64_t)1 << (52 - e)) - 1;
if ((b & fracMask) == 0) {
return x; // already integer
}
uint64_t truncBits = b & ~fracMask;
double truncVal = dFromBits(truncBits);
// For negatives with a non-zero discarded fraction, subtract 1.
if (b >> 63) {
truncVal = truncVal - 1.0;
}
return truncVal;
}
double ceil(double x) {
return -floor(-x);
}
float floorf(float x) {
return (float)floor((double)x);
}
float ceilf(float x) {
return (float)ceil((double)x);
}
// fmod(x, y) = x - trunc(x / y) * y. trunc(q) = round-toward-zero.
double fmod(double x, double y) {
if (y == 0.0) {
return 0.0; // implementation-defined; we return 0
}
double q = x / y;
// trunc: floor for positives, ceil for negatives.
double t;
if (q < 0.0) {
t = -floor(-q);
} else {
t = floor(q);
}
return x - t * y;
}
float fmodf(float x, float y) {
return (float)fmod((double)x, (double)y);
}
// sqrt via Newton iteration with a bit-trick initial guess that
// uses the IEEE-754 layout: y0 = bits >> 1 + bias_shift. Then ~6
// Newton iterations land within ~1 ULP for double precision.
//
// We dodge the snprintf-style libcall pitfalls by NOT comparing
// double-to-double via `<`/`>=`; instead we test the IEEE-754
// sign bit and exponent directly.
double sqrt(double x) {
uint64_t b;
__builtin_memcpy(&b, &x, sizeof(b));
// Check zero first (positive or negative) — IEEE-754 says
// sqrt(+0)=+0 and sqrt(-0)=-0; both lower 63 bits are zero.
if ((b & ~((uint64_t)1 << 63)) == 0) {
return x;
}
if (b & ((uint64_t)1 << 63)) {
return 0.0 / 0.0; // NaN for negatives
}
// Initial guess: halve the exponent. IEEE-754 trick gives a
// surprisingly good starting point — within 2x of the true value.
uint64_t guess_b = ((b - ((uint64_t)1023 << 52)) >> 1)
+ ((uint64_t)1023 << 52);
double y;
__builtin_memcpy(&y, &guess_b, sizeof(y));
// Newton: y_{n+1} = (y_n + x / y_n) / 2. 6 iterations for double.
for (int i = 0; i < 6; i++) {
y = (y + x / y) * 0.5;
}
return y;
}
float sqrtf(float x) {
return (float)sqrt((double)x);
}
// pow(x, y) supports two restricted regimes:
// 1. integer y (positive, negative, or zero) — repeated multiply.
// 2. y == 0.5 — returns sqrt(x).
// Anything else (general real exponent) needs exp/log; not yet
// landed. Returning 0 is the documented behaviour for now.
double pow(double x, double y) {
if (y == 0.0) {
return 1.0;
}
if (y == 0.5) {
return sqrt(x);
}
// Test for integer y by floor-equal trick (no soft-fp comparison).
double yi = floor(y);
uint64_t yi_b, y_b;
__builtin_memcpy(&yi_b, &yi, sizeof(yi_b));
__builtin_memcpy(&y_b, &y, sizeof(y_b));
if (yi_b != y_b) {
return 0.0; // non-integer, non-0.5 y not supported yet
}
// y is a whole number; convert via __fixdfsi. Range -32768..32767
// covers any practical exponent. Use unsigned for the magnitude
// to avoid signed-overflow UB on INT_MIN.
int sn = (int)yi;
int neg = 0;
unsigned int n;
if (sn < 0) {
neg = 1;
n = 0u - (unsigned int)sn;
} else {
n = (unsigned int)sn;
}
double r = 1.0;
double base = x;
while (n > 0) {
if (n & 1) {
r = r * base;
}
base = base * base;
n = n >> 1;
}
if (neg) {
r = 1.0 / r;
}
return r;
}
float powf(float x, float y) {
return (float)pow((double)x, (double)y);
}
// ----- transcendentals --------------------------------------------------
// All implemented via range-reduction + Taylor series. Accuracy is in
// the ~1e-6 range — good enough for typical IIgs-era numeric work,
// well short of glibc-quality. The soft-double libcalls dominate the
// cost; each transcendental is dozens of milliseconds to seconds in
// MAME, so don't put these in tight inner loops.
#define M_PI 3.14159265358979323846
#define M_2PI 6.28318530717958647692
#define M_HALF_PI 1.57079632679489661923
#define M_LN2 0.69314718055994530942
// sin via Taylor: x - x^3/6 + x^5/120 - x^7/5040 + x^9/362880 - x^11/39916800
// Range-reduce first by subtracting multiples of 2*pi, then fold into
// [-pi/2, pi/2] using sin(pi - x) = sin(x), sin(-x) = -sin(x).
double sin(double x) {
// Reduce to [-pi, pi].
double k = floor((x + M_PI) / M_2PI);
x = x - k * M_2PI;
// Fold to [-pi/2, pi/2] via sign-bit-test (avoids __ltdf2).
uint64_t b;
__builtin_memcpy(&b, &x, sizeof(b));
int sign = (int)(b >> 63) & 1;
if (sign) {
// x < 0: work with -x and negate result at end.
b ^= ((uint64_t)1 << 63);
__builtin_memcpy(&x, &b, sizeof(x));
}
if (x > M_HALF_PI) {
x = M_PI - x; // reflect about pi/2
}
double x2 = x * x;
double term = x;
double s = term;
term = term * x2; s = s - term * (1.0 / 6.0);
term = term * x2; s = s + term * (1.0 / 120.0);
term = term * x2; s = s - term * (1.0 / 5040.0);
term = term * x2; s = s + term * (1.0 / 362880.0);
term = term * x2; s = s - term * (1.0 / 39916800.0);
if (sign) {
s = -s;
}
return s;
}
double cos(double x) {
return sin(x + M_HALF_PI);
}
// tan(x) = sin(x) / cos(x). No special handling for poles at pi/2
// + n*pi (where cos(x) == 0): the soft-double divide returns +/-Inf,
// which is the IEEE-754-correct answer. Accuracy follows sin/cos
// (~1e-6) but degrades fast as |x| approaches a pole.
double tan(double x) {
return sin(x) / cos(x);
}
float sinf(float x) {
return (float)sin((double)x);
}
float cosf(float x) {
return (float)cos((double)x);
}
float tanf(float x) {
return (float)tan((double)x);
}
// exp via 2^k * e^r where x = k*ln2 + r, |r| < ln2/2. Then Taylor
// series for e^r converges in ~10 terms. k * 2 multiplication uses
// the IEEE-754 layout (add k to exponent field).
double exp(double x) {
double k_d = floor(x / M_LN2 + 0.5);
double r = x - k_d * M_LN2;
int k = (int)k_d;
// Taylor: 1 + r + r^2/2 + r^3/6 + ... + r^10/3628800
double term = 1.0;
double s = 1.0;
for (int n = 1; n <= 12; n++) {
term = term * r * (1.0 / (double)n);
s = s + term;
}
// Multiply by 2^k by adding k to the IEEE-754 exponent field.
uint64_t b;
__builtin_memcpy(&b, &s, sizeof(b));
uint64_t e = (b >> 52) & 0x7FF;
int en = (int)e + k;
if (en >= 0x7FF) {
return 1.0 / 0.0; // overflow -> inf
}
if (en <= 0) {
return 0.0; // underflow
}
b = (b & ~((uint64_t)0x7FF << 52)) | ((uint64_t)en << 52);
__builtin_memcpy(&s, &b, sizeof(s));
return s;
}
float expf(float x) {
return (float)exp((double)x);
}
// log: x = 2^k * m where m in [1, 2). log(x) = k*ln2 + log(m).
// log(m) for m in [1,2) computed via the substitution u = (m-1)/(m+1)
// giving log(m) = 2 * (u + u^3/3 + u^5/5 + ...) — converges fast for
// m in [1,2).
double log(double x) {
uint64_t b;
__builtin_memcpy(&b, &x, sizeof(b));
// log(±0) = -Infinity (pole error). Mask off the sign bit when
// testing for zero so -0.0 lands here instead of the negative path.
if ((b & ~((uint64_t)1 << 63)) == 0) {
return -1.0 / 0.0;
}
if (b & ((uint64_t)1 << 63)) {
return 0.0 / 0.0; // log(negative) = NaN (domain error)
}
int e = (int)((b >> 52) & 0x7FF) - 1023;
// Force the exponent field to 1023 so m lands in [1, 2).
b = (b & ~((uint64_t)0x7FF << 52)) | ((uint64_t)1023 << 52);
double m;
__builtin_memcpy(&m, &b, sizeof(m));
double u = (m - 1.0) / (m + 1.0);
double u2 = u * u;
double term = u;
double s = term;
for (int n = 1; n <= 8; n++) {
term = term * u2;
s = s + term * (1.0 / (double)(2 * n + 1));
}
return 2.0 * s + (double)e * M_LN2;
}
float logf(float x) {
return (float)log((double)x);
}
// atan: range-reduce |x| > 1 via atan(x) = pi/2 - atan(1/x), then
// further reduce via atan(x) = 2*atan(x / (1 + sqrt(1 + x*x))).
// For the reduced argument, Taylor: x - x^3/3 + x^5/5 - ...
double atan(double x) {
uint64_t b;
__builtin_memcpy(&b, &x, sizeof(b));
int sign = (int)(b >> 63) & 1;
if (sign) {
b ^= ((uint64_t)1 << 63);
__builtin_memcpy(&x, &b, sizeof(x));
}
int invert = 0;
// Test |x| > 1 via bit pattern: exponent > 1023 means value >= 1.
int e = (int)((b >> 52) & 0x7FF) - 1023;
if (e >= 0) {
x = 1.0 / x;
invert = 1;
}
// Taylor for |x| <= 1: x - x^3/3 + x^5/5 - x^7/7 + x^9/9 - ...
// Need many terms when |x| close to 1; cap at 30 for ~1e-6 accuracy.
double x2 = x * x;
double term = x;
double s = term;
for (int n = 1; n <= 25; n++) {
term = term * x2;
double denom = (double)(2 * n + 1);
if (n & 1) {
s = s - term / denom;
} else {
s = s + term / denom;
}
}
if (invert) {
s = M_HALF_PI - s;
}
if (sign) {
s = -s;
}
return s;
}
float atanf(float x) {
return (float)atan((double)x);
}
double atan2(double y, double x) {
// Quadrant adjustment.
uint64_t xb, yb;
__builtin_memcpy(&xb, &x, sizeof(xb));
__builtin_memcpy(&yb, &y, sizeof(yb));
int xNeg = (int)(xb >> 63) & 1;
int yNeg = (int)(yb >> 63) & 1;
// Strip sign of x for the quotient; restore quadrant later.
uint64_t xa = xb & ~((uint64_t)1 << 63);
if (xa == 0) {
// x == 0 (handles +0 and -0): return ±pi/2
return yNeg ? -M_HALF_PI : M_HALF_PI;
}
double a = atan(y / x);
if (!xNeg) {
return a;
}
return yNeg ? a - M_PI : a + M_PI;
}
float atan2f(float y, float x) {
return (float)atan2((double)y, (double)x);
}
// asin: asin(x) = atan(x / sqrt(1 - x*x))
double asin(double x) {
double r = 1.0 - x * x;
if (r <= 0.0) {
// |x| >= 1; clamp to ±pi/2.
uint64_t b;
__builtin_memcpy(&b, &x, sizeof(b));
return (b >> 63) ? -M_HALF_PI : M_HALF_PI;
}
return atan(x / sqrt(r));
}
float asinf(float x) {
return (float)asin((double)x);
}
double acos(double x) {
return M_HALF_PI - asin(x);
}
float acosf(float x) {
return (float)acos((double)x);
}
double sinh(double x) {
double e = exp(x);
return (e - 1.0 / e) * 0.5;
}
float sinhf(float x) {
return (float)sinh((double)x);
}
double cosh(double x) {
double e = exp(x);
return (e + 1.0 / e) * 0.5;
}
float coshf(float x) {
return (float)cosh((double)x);
}
double tanh(double x) {
double e2 = exp(x + x);
return (e2 - 1.0) / (e2 + 1.0);
}
float tanhf(float x) {
return (float)tanh((double)x);
}
// ---- Classification ------------------------------------------------
//
// Implemented as functions rather than the C99 macros so they can be
// referenced from outside math.h. The math.h header also exposes them
// as macros that expand to these calls.
int __isnan_d(double x) {
uint64_t b = dToBits(x);
return ((b >> 52) & 0x7FF) == 0x7FF && (b & 0xFFFFFFFFFFFFFULL) != 0;
}
int __isinf_d(double x) {
uint64_t b = dToBits(x);
return ((b >> 52) & 0x7FF) == 0x7FF && (b & 0xFFFFFFFFFFFFFULL) == 0;
}
int __isfinite_d(double x) {
return ((dToBits(x) >> 52) & 0x7FF) != 0x7FF;
}
int __signbit_d(double x) {
return (int)(dToBits(x) >> 63);
}
// ---- Rounding ------------------------------------------------------
// trunc: round toward zero. = floor for positive, ceil for negative.
double trunc(double x) {
return (dToBits(x) >> 63) ? ceil(x) : floor(x);
}
float truncf(float x) {
return (float)trunc((double)x);
}
// round: round half away from zero.
double round(double x) {
return (dToBits(x) >> 63) ? -floor(-x + 0.5) : floor(x + 0.5);
}
float roundf(float x) {
return (float)round((double)x);
}
// ---- Min / max / positive difference -------------------------------
double fmax(double x, double y) {
if (__isnan_d(x)) return y;
if (__isnan_d(y)) return x;
return x > y ? x : y;
}
double fmin(double x, double y) {
if (__isnan_d(x)) return y;
if (__isnan_d(y)) return x;
return x < y ? x : y;
}
double fdim(double x, double y) {
return x > y ? x - y : 0.0;
}
float fmaxf(float x, float y) { return (float)fmax((double)x, (double)y); }
float fminf(float x, float y) { return (float)fmin((double)x, (double)y); }
float fdimf(float x, float y) { return (float)fdim((double)x, (double)y); }
// ---- FP decomposition ----------------------------------------------
// ldexp(x, n) = x * 2^n. Implemented by adjusting the exponent field
// in place. Handles subnormals and overflow only crudely (overflow →
// infinity; underflow → zero).
double ldexp(double x, int n) {
uint64_t b = dToBits(x);
int e = (int)((b >> 52) & 0x7FF);
if (e == 0 || e == 0x7FF) {
// 0 or denorm or inf/nan — return as-is for inf/nan, else
// fall back to multiplication (handles subnormals correctly
// enough for typical use).
if (e == 0x7FF) return x;
// For zero / subnormal, multiply by 2^n via repeated *2.
if (n > 0) {
while (n--) x *= 2.0;
return x;
}
if (n < 0) {
while (n++) x *= 0.5;
return x;
}
return x;
}
int newE = e + n;
if (newE >= 0x7FF) {
// Overflow → ±infinity.
return (b >> 63) ? -1e308 * 1e308 : 1e308 * 1e308;
}
if (newE <= 0) {
// Underflow → ±0 (skip subnormal handling).
return (b >> 63) ? -0.0 : 0.0;
}
b = (b & 0x800FFFFFFFFFFFFFULL) | ((uint64_t)newE << 52);
return dFromBits(b);
}
float ldexpf(float x, int n) {
return (float)ldexp((double)x, n);
}
// frexp(x, *e): split x into normalized fraction in [0.5, 1) (or 0)
// and integer exponent. x = frac * 2^e.
double frexp(double x, int *eOut) {
uint64_t b = dToBits(x);
int e = (int)((b >> 52) & 0x7FF);
if (e == 0) { *eOut = 0; return x; } // zero or subnormal
if (e == 0x7FF) { *eOut = 0; return x; } // inf or nan
*eOut = e - 1022; // bias adjustment for [0.5,1)
// Force exponent to 1022 (so result is in [0.5, 1)).
b = (b & 0x800FFFFFFFFFFFFFULL) | ((uint64_t)1022 << 52);
return dFromBits(b);
}
float frexpf(float x, int *eOut) {
double r = frexp((double)x, eOut);
return (float)r;
}
// modf(x, *iptr): split into integer and fractional parts, both with
// the same sign as x.
double modf(double x, double *iptr) {
double ip = trunc(x);
*iptr = ip;
return x - ip;
}
float modff(float x, float *iptr) {
double ipd;
double frac = modf((double)x, &ipd);
*iptr = (float)ipd;
return (float)frac;
}
// ---- log10 / log2 / exp2 -------------------------------------------
//
// All routed through log() / exp() with the conversion constant.
double log10(double x) {
// 1 / ln(10) = 0.43429448190325182765112891891660508229439700580366
return log(x) * 0.43429448190325182765;
}
double log2(double x) {
// 1 / ln(2) = 1.4426950408889634073599246810018921374266459541530
return log(x) * 1.4426950408889634074;
}
double exp2(double x) {
// ln(2) = 0.69314718055994530941723212145817656807550013436026
return exp(x * 0.69314718055994530942);
}
float log10f(float x) { return (float)log10((double)x); }
float log2f(float x) { return (float)log2((double)x); }
float exp2f(float x) { return (float)exp2((double)x); }
// log1p(x) = log(1 + x). For |x| small the naive form loses precision;
// we accept that for now since this is a standalone runtime, not a
// numerics library.
double log1p(double x) { return log(1.0 + x); }
float log1pf(float x) { return (float)log1p((double)x); }
// expm1(x) = exp(x) - 1. Same loss-of-precision caveat near 0.
double expm1(double x) { return exp(x) - 1.0; }
float expm1f(float x) { return (float)expm1((double)x); }
// ---- hypot ---------------------------------------------------------
//
// hypot(x, y) = sqrt(x*x + y*y). This implementation does NOT scale
// to avoid overflow — for |x|, |y| < ~1e150 the naive form is fine,
// past that you'd want the standard scale-by-max trick.
// hypot — naive sqrt(x*x + y*y). NO `volatile` on the temps —
// clang's codegen for volatile-double locals on this target generates
// stack-relative loads/stores that crash under the GS/OS Loader (the
// chain executes correctly under runInMame but not via Finder). The
// volatile-free version works in both contexts.
__attribute__((noinline))
double hypot(double x, double y) {
double xx = x * x;
double yy = y * y;
double s = xx + yy;
return sqrt(s);
}
float hypotf(float x, float y) {
return (float)hypot((double)x, (double)y);
}
// ---- cbrt ----------------------------------------------------------
//
// Newton-Raphson for cube root: r_{n+1} = (2*r_n + a/r_n²) / 3.
// Converges quadratically; 30 iters more than enough for double.
// Implemented WITHOUT calling pow because clang treats pow as a
// known builtin and either inlines it (with bad fold of pow(x,1/3))
// or DCEs the call entirely (cbrt body collapses to "return 0").
// This implementation has no pow dependency and is immune.
__attribute__((noinline))
double cbrt(double x) {
if (x == 0.0) return x;
int neg = (int)(dToBits(x) >> 63) & 1;
double a = neg ? -x : x;
double r = a; // crude initial guess
for (int i = 0; i < 30; i++) {
r = (2.0 * r + a / (r * r)) * (1.0 / 3.0);
}
return neg ? -r : r;
}
float cbrtf(float x) {
return (float)cbrt((double)x);
}