[llvm-branch-commits] [libc] [libc][math] Use multiply_sub and sub_multiply across math functions. (PR #243290)
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Tue Nov 17 16:11:37 PST 2026
https://github.com/lntue updated https://github.com/llvm/llvm-project/pull/243290
>From 69f5e684dd7fedb203a6754b5b9e14167f25240b Mon Sep 17 00:00:00 2001
From: Tue Ly <lntue.h at gmail.com>
Date: Mon, 16 Nov 2026 23:21:04 +0000
Subject: [PATCH] [libc][math] Use multiply_sub and sub_multiply across math
headers.
Replace negated multiply_add calls (multiply_add(a, b, -c), multiply_add(-b, c, a),
and multiply_add(b, -c, a)) across FPUtil and math implementation headers with
multiply_sub and sub_multiply.
---
libc/src/__support/FPUtil/double_double.h | 12 ++++++------
libc/src/__support/math/acos.h | 19 +++++++++----------
libc/src/__support/math/acosbf16.h | 4 ++--
libc/src/__support/math/acosf.h | 8 ++++----
libc/src/__support/math/acosf16.h | 6 +++---
libc/src/__support/math/acoshbf16.h | 2 +-
libc/src/__support/math/acoshf.h | 4 ++--
libc/src/__support/math/acoshf16.h | 2 +-
libc/src/__support/math/acoshf_utils.h | 2 +-
libc/src/__support/math/acospif.h | 2 +-
libc/src/__support/math/acospif16.h | 6 +++---
libc/src/__support/math/asin.h | 15 +++++++--------
libc/src/__support/math/asin_utils.h | 2 +-
libc/src/__support/math/asinbf16.h | 4 ++--
libc/src/__support/math/asinf.h | 2 +-
libc/src/__support/math/asinf16.h | 4 ++--
libc/src/__support/math/asinpi.h | 13 ++++++-------
libc/src/__support/math/asinpif.h | 2 +-
libc/src/__support/math/asinpif16.h | 4 ++--
libc/src/__support/math/atan.h | 4 ++--
libc/src/__support/math/atan2f.h | 12 ++++++------
libc/src/__support/math/atan2f128.h | 2 +-
libc/src/__support/math/atan2f16.h | 2 +-
libc/src/__support/math/atan2f_float.h | 2 +-
libc/src/__support/math/atan_utils.h | 10 +++++-----
libc/src/__support/math/atanbf16.h | 2 +-
libc/src/__support/math/atanf.h | 8 ++++----
libc/src/__support/math/atanf16.h | 2 +-
libc/src/__support/math/atanf_float.h | 6 +++---
libc/src/__support/math/atanpif16.h | 2 +-
libc/src/__support/math/cbrt.h | 4 ++--
libc/src/__support/math/cosf16.h | 2 +-
libc/src/__support/math/cosf_double_eval.h | 8 ++++----
libc/src/__support/math/cospif.h | 8 ++++----
libc/src/__support/math/cospif16.h | 2 +-
libc/src/__support/math/exp10f16_utils.h | 2 +-
libc/src/__support/math/exp10f_float_eval.h | 6 +++---
libc/src/__support/math/exp10m1f.h | 2 +-
libc/src/__support/math/exp10m1f16.h | 2 +-
libc/src/__support/math/exp2_double_eval.h | 2 +-
libc/src/__support/math/exp2bf16.h | 2 +-
libc/src/__support/math/exp2f_double_eval.h | 2 +-
libc/src/__support/math/exp2f_float_eval.h | 2 +-
libc/src/__support/math/exp2m1f.h | 4 ++--
libc/src/__support/math/exp2m1f16.h | 2 +-
libc/src/__support/math/expbf16.h | 2 +-
libc/src/__support/math/expf16_utils.h | 2 +-
libc/src/__support/math/expf_double_eval.h | 2 +-
libc/src/__support/math/expf_float_eval.h | 8 ++++----
libc/src/__support/math/expm1f.h | 4 ++--
libc/src/__support/math/expm1f16.h | 2 +-
libc/src/__support/math/expxf16_utils.h | 4 ++--
libc/src/__support/math/hypot.h | 2 +-
libc/src/__support/math/hypotf.h | 4 ++--
libc/src/__support/math/hypotf16.h | 2 +-
libc/src/__support/math/lgammaf.h | 12 ++++++------
libc/src/__support/math/log.h | 2 +-
libc/src/__support/math/log10.h | 2 +-
libc/src/__support/math/log10f16.h | 2 +-
libc/src/__support/math/log10f_double_eval.h | 6 +++---
libc/src/__support/math/log10p1bf16.h | 2 +-
libc/src/__support/math/log10p1f16.h | 2 +-
libc/src/__support/math/log1p.h | 2 +-
libc/src/__support/math/log2.h | 2 +-
libc/src/__support/math/log2f16.h | 2 +-
libc/src/__support/math/log2f_double_eval.h | 4 ++--
libc/src/__support/math/log2p1f16.h | 2 +-
libc/src/__support/math/logf16.h | 2 +-
libc/src/__support/math/logf_double_eval.h | 4 ++--
libc/src/__support/math/pow.h | 2 +-
libc/src/__support/math/pow_fast.h | 2 +-
libc/src/__support/math/powf_double_eval.h | 8 ++++----
libc/src/__support/math/powf_float_eval.h | 4 ++--
libc/src/__support/math/rsqrtf16.h | 2 +-
libc/src/__support/math/sin.h | 6 +++---
libc/src/__support/math/sincos.h | 6 +++---
libc/src/__support/math/sincos_eval.h | 14 +++++++-------
libc/src/__support/math/sincosf.h | 14 +++++++-------
libc/src/__support/math/sincosf16_utils.h | 2 +-
libc/src/__support/math/sincosf_float_eval.h | 6 +++---
libc/src/__support/math/sincosf_utils.h | 2 +-
libc/src/__support/math/sinf_double_eval.h | 4 ++--
libc/src/__support/math/sinhcosh_utils.h | 2 +-
libc/src/__support/math/sinhcoshbf16_utils.h | 4 ++--
libc/src/__support/math/tanbf16.h | 4 ++--
libc/src/__support/math/tanf16.h | 2 +-
libc/src/__support/math/tanf_double_eval.h | 2 +-
libc/src/__support/math/tanf_float_eval.h | 2 +-
libc/src/__support/math/tanh_double_eval.h | 10 +++++-----
libc/src/__support/math/tanhf.h | 2 +-
libc/src/__support/math/tanhf16.h | 2 +-
libc/src/__support/math/tanpif.h | 6 +++---
libc/src/__support/math/tanpif16.h | 4 ++--
93 files changed, 201 insertions(+), 204 deletions(-)
diff --git a/libc/src/__support/FPUtil/double_double.h b/libc/src/__support/FPUtil/double_double.h
index 3c1c052352a0a9..c74d5a9a038394 100644
--- a/libc/src/__support/FPUtil/double_double.h
+++ b/libc/src/__support/FPUtil/double_double.h
@@ -208,11 +208,11 @@ LIBC_INLINE LIBC_CONSTEXPR NumberPair<T> exact_mult(T a, T b) {
if constexpr (TargetHasFmaInstruction<T>::VALUE) {
r.hi = a * b;
- // Prevent `-r.hi` below from being rewritten into `a * (-b)` when `a * b`
- // is duplicated (e.g. by SLP vectorization), which breaks directed
- // rounding modes (`RU(a * (-b)) != -RU(a * b)`).
+ // Prevent `-r.hi` inside `multiply_sub` from being rewritten into
+ // `a * (-b)` when `a * b` is duplicated (e.g. by SLP vectorization), which
+ // breaks directed rounding modes (`RU(a * (-b)) != -RU(a * b)`).
LIBC_DO_NOT_OPTIMIZE_ROUNDING(r.hi);
- r.lo = fputil::multiply_add(a, b, -r.hi);
+ r.lo = fputil::multiply_sub(a, b, r.hi);
} else {
// Dekker's Product.
NumberPair<T> as = split(a);
@@ -289,8 +289,8 @@ LIBC_INLINE NumberPair<T> div(const NumberPair<T> &a, const NumberPair<T> &b) {
LIBC_DO_NOT_OPTIMIZE_ROUNDING(r.hi);
#ifdef LIBC_TARGET_CPU_HAS_FMA
- T e_hi = fputil::multiply_add(b.hi, -r.hi, a.hi);
- T e_lo = fputil::multiply_add(b.lo, -r.hi, a.lo);
+ T e_hi = fputil::sub_multiply(a.hi, b.hi, r.hi);
+ T e_lo = fputil::sub_multiply(a.lo, b.lo, r.hi);
#else
NumberPair<T> b_hi_r_hi = fputil::exact_mult(b.hi, -r.hi);
NumberPair<T> b_lo_r_hi = fputil::exact_mult(b.lo, -r.hi);
diff --git a/libc/src/__support/math/acos.h b/libc/src/__support/math/acos.h
index dbcee8f0df6f15..481d4d9e8768c1 100644
--- a/libc/src/__support/math/acos.h
+++ b/libc/src/__support/math/acos.h
@@ -52,7 +52,7 @@ LIBC_INLINE double acos(double x) {
// = pi/2 - x - x^3 * P1(x^2)
double xsq = x * x;
return PI_OVER_TWO.hi +
- fputil::multiply_add(-x * xsq, asin_eval(xsq), PI_OVER_TWO.lo - x);
+ fputil::sub_multiply(PI_OVER_TWO.lo - x, x * xsq, asin_eval(xsq));
#else
unsigned idx = 0;
DoubleDouble x_sq = fputil::exact_mult(x, x);
@@ -65,10 +65,10 @@ LIBC_INLINE double acos(double x) {
// acos(x) = pi/2 - asin(x)
// ~ pi/2 - x * p
// = pi/2 - x * (p.hi + p.lo)
- double r_hi = fputil::multiply_add(-x, p.hi, PI_OVER_TWO.hi);
+ double r_hi = fputil::sub_multiply(PI_OVER_TWO.hi, x, p.hi);
// Use Dekker's 2SUM algorithm to compute the lower part.
double r_lo = ((PI_OVER_TWO.hi - r_hi) - r0.hi) - r0.lo;
- r_lo = fputil::multiply_add(-x, p.lo, r_lo + PI_OVER_TWO.lo);
+ r_lo = fputil::sub_multiply(r_lo + PI_OVER_TWO.lo, x, p.lo);
// Ziv's accuracy test.
@@ -91,7 +91,7 @@ LIBC_INLINE double acos(double x) {
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
// u = x^2 - idx/64
DFloat128 u_hi(
- fputil::multiply_add(static_cast<double>(idx), -0x1.0p-6, x_sq.hi));
+ fputil::sub_multiply(x_sq.hi, static_cast<double>(idx), 0x1.0p-6));
DFloat128 u = fputil::quick_add(u_hi, DFloat128(x_sq.lo));
#else
DFloat128 x_sq_f128 = fputil::quick_mul(x_f128, x_f128);
@@ -121,7 +121,7 @@ LIBC_INLINE double acos(double x) {
if (x_abs == 1.0) {
// x = 1, acos(x) = 0,
// x = -1, acos(x) = pi
- return x == 1.0 ? 0.0 : fputil::multiply_add(-x_sign, PI.hi, PI.lo);
+ return x == 1.0 ? 0.0 : fputil::sub_multiply(PI.lo, x_sign, PI.hi);
}
// |x| > 1, return NaN.
if (xbits.is_quiet_nan())
@@ -164,7 +164,7 @@ LIBC_INLINE double acos(double x) {
// where u = (1 - |x|)/2.
// u = (1 - |x|)/2
- double u = fputil::multiply_add(x_abs, -0.5, 0.5);
+ double u = fputil::sub_multiply(0.5, x_abs, 0.5);
// v_hi + v_lo ~ sqrt(u).
// Let:
// h = u - v_hi^2 = (sqrt(u) - v_hi) * (sqrt(u) + v_hi)
@@ -186,7 +186,7 @@ LIBC_INLINE double acos(double x) {
#else
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
- double h = fputil::multiply_add(v_hi, -v_hi, u);
+ double h = fputil::sub_multiply(u, v_hi, v_hi);
#else
DoubleDouble v_hi_sq = fputil::exact_mult(v_hi, v_hi);
double h = (u - v_hi_sq.hi) - v_hi_sq.lo;
@@ -251,7 +251,7 @@ LIBC_INLINE double acos(double x) {
// Get the rounding error of vl = 2 * v_lo ~ h / vh
// Get full product of vh * vl
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
- double vl_lo = fputil::multiply_add(-v_hi, vl, h) / v_hi;
+ double vl_lo = fputil::sub_multiply(h, v_hi, vl) / v_hi;
#else
DoubleDouble vh_vl = fputil::exact_mult(v_hi, vl);
double vl_lo = ((h - vh_vl.hi) - vh_vl.lo) / v_hi;
@@ -267,8 +267,7 @@ LIBC_INLINE double acos(double x) {
// Perform computations in DFloat128:
// acos(x) = (v_hi + v_lo + vll) * P(u) , when 0.5 <= x < 1,
// = pi - (v_hi + v_lo + vll) * P(u) , when -1 < x <= -0.5.
- DFloat128 y_f128(
- fputil::multiply_add(static_cast<double>(idx), -0x1.0p-6, u));
+ DFloat128 y_f128(fputil::sub_multiply(u, static_cast<double>(idx), 0x1.0p-6));
DFloat128 p_f128 = asin_eval(y_f128, idx);
DFloat128 r_f128 = fputil::quick_mul(m_v, p_f128);
diff --git a/libc/src/__support/math/acosbf16.h b/libc/src/__support/math/acosbf16.h
index e0fb861accb475..4f6b73f20637cc 100644
--- a/libc/src/__support/math/acosbf16.h
+++ b/libc/src/__support/math/acosbf16.h
@@ -67,12 +67,12 @@ LIBC_INLINE bfloat16 acosbf16(bfloat16 x) {
// using reduction for acos:
// acos(|x|) = 2*asin(sqrt((1 - |x|)/2)),
// and acos(x) = acos(|x|) for x >= 0, pi - acos(|x|) for x < 0
- float t = fputil::multiply_add<float>(xf_abs, -0.5f, 0.5f);
+ float t = fputil::sub_multiply<float>(0.5f, xf_abs, 0.5f);
float t_sqrt = fputil::sqrt<float>(t);
float asin_sqrt_t = t_sqrt * asin_internal::asin_eval_float(t);
return fputil::cast<bfloat16>(
- (sign) ? fputil::multiply_add(asin_sqrt_t, -2.0f, PI)
+ (sign) ? fputil::sub_multiply(PI, asin_sqrt_t, 2.0f)
: 2.0f * asin_sqrt_t);
}
// case 3: NaN or |x| > 1
diff --git a/libc/src/__support/math/acosf.h b/libc/src/__support/math/acosf.h
index 2f19ed4d99b292..6a047b0e8bda37 100644
--- a/libc/src/__support/math/acosf.h
+++ b/libc/src/__support/math/acosf.h
@@ -74,8 +74,8 @@ LIBC_INLINE float acosf(float x) {
#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
double xd = static_cast<double>(x);
- return static_cast<float>(fputil::multiply_add(
- -0x1.5555555555555p-3 * xd, xd * xd, M_MATH_PI_2 - xd));
+ return static_cast<float>(fputil::sub_multiply(
+ M_MATH_PI_2 - xd, 0x1.5555555555555p-3 * xd, xd * xd));
}
// For |x| <= 0.5, we approximate acosf(x) by:
@@ -88,7 +88,7 @@ LIBC_INLINE float acosf(float x) {
double xsq = xd * xd;
double x3 = xd * xsq;
double r = asin_eval(xsq);
- return static_cast<float>(fputil::multiply_add(-x3, r, M_MATH_PI_2 - xd));
+ return static_cast<float>(fputil::sub_multiply(M_MATH_PI_2 - xd, x3, r));
}
// |x| >= 1, return 0, 2pi, or NaNs.
@@ -136,7 +136,7 @@ LIBC_INLINE float acosf(float x) {
xbits.set_sign(Sign::POS);
double xd = static_cast<double>(xbits.get_val());
- double u = fputil::multiply_add(-0.5, xd, 0.5);
+ double u = fputil::sub_multiply(0.5, 0.5, xd);
double cv = 2 * fputil::sqrt<double>(u);
double r3 = asin_eval(u);
diff --git a/libc/src/__support/math/acosf16.h b/libc/src/__support/math/acosf16.h
index b7bbbd75d06afa..70a0bab9b8fbac 100644
--- a/libc/src/__support/math/acosf16.h
+++ b/libc/src/__support/math/acosf16.h
@@ -98,7 +98,7 @@ LIBC_INLINE constexpr float16 acosf16(float16 x) {
float interm =
fputil::polyeval(xsq, 0x1.000002p0f, 0x1.554c2ap-3f, 0x1.3541ccp-4f,
0x1.43b2d6p-5f, 0x1.a0d73ep-5f);
- return fputil::cast<float16>(fputil::multiply_add(-xf, interm, PI_OVER_2));
+ return fputil::cast<float16>(fputil::sub_multiply(PI_OVER_2, xf, interm));
}
// When |x| > 0.5, assume that 0.5 < |x| <= 1
@@ -142,7 +142,7 @@ LIBC_INLINE constexpr float16 acosf16(float16 x) {
// with a positive x value instead (rhs).
float xf_abs = (xf < 0 ? -xf : xf);
- float u = fputil::multiply_add(-0.5f, xf_abs, 0.5f);
+ float u = fputil::sub_multiply(0.5f, 0.5f, xf_abs);
float sqrt_u = fputil::sqrt<float>(u);
// Degree-6 minimax polynomial of asin(x) generated by Sollya with:
@@ -152,7 +152,7 @@ LIBC_INLINE constexpr float16 acosf16(float16 x) {
0x1.3541ccp-4f, 0x1.43b2d6p-5f, 0x1.a0d73ep-5f);
return fputil::cast<float16>(
- x_sign ? fputil::multiply_add(-2.0f, asin_sqrt_u, PI) : 2 * asin_sqrt_u);
+ x_sign ? fputil::sub_multiply(PI, 2.0f, asin_sqrt_u) : 2 * asin_sqrt_u);
}
} // namespace math
diff --git a/libc/src/__support/math/acoshbf16.h b/libc/src/__support/math/acoshbf16.h
index aaf9b0c05f7868..008a69baad0bb0 100644
--- a/libc/src/__support/math/acoshbf16.h
+++ b/libc/src/__support/math/acoshbf16.h
@@ -136,7 +136,7 @@ LIBC_INLINE LIBC_CONSTEXPR bfloat16 acoshbf16(bfloat16 x) {
// this avoids overflow when squaring the largest finite BFloat16 inputs.
// Convert directly to BFloat16 to avoid an intermediate float rounding.
double xd = static_cast<double>(xf);
- double sqrt_term = fputil::sqrt<double>(fputil::multiply_add(xd, xd, -1.0));
+ double sqrt_term = fputil::sqrt<double>(fputil::multiply_sub(xd, xd, 1.0));
double result = log_eval(xd + sqrt_term);
return fputil::cast<bfloat16>(result);
diff --git a/libc/src/__support/math/acoshf.h b/libc/src/__support/math/acoshf.h
index 02b2a3d5a513db..c68eaafede7b38 100644
--- a/libc/src/__support/math/acoshf.h
+++ b/libc/src/__support/math/acoshf.h
@@ -79,7 +79,7 @@ LIBC_INLINE float acoshf(float x) {
#endif // LIBC_MATH_HAS_SKIP_ACCURATE_PASS
double y_inv = 0.5 / x_d;
return static_cast<float>(
- fputil::multiply_add(y_inv, -y_inv, log_eval(y)));
+ fputil::sub_multiply(log_eval(y), y_inv, y_inv));
} else {
// x > 2^26
@@ -114,7 +114,7 @@ LIBC_INLINE float acoshf(float x) {
// For 1 < x < 2^12, we use the formula:
// acosh(x) = log(x + sqrt(x^2 - 1))
return static_cast<float>(log_eval(
- x_d + fputil::sqrt<double>(fputil::multiply_add(x_d, x_d, -1.0))));
+ x_d + fputil::sqrt<double>(fputil::multiply_sub(x_d, x_d, 1.0))));
}
} // namespace math
diff --git a/libc/src/__support/math/acoshf16.h b/libc/src/__support/math/acoshf16.h
index eeb9e9f7afa9d5..567dd1a2d7bdaa 100644
--- a/libc/src/__support/math/acoshf16.h
+++ b/libc/src/__support/math/acoshf16.h
@@ -108,7 +108,7 @@ LIBC_INLINE constexpr float16 acoshf16(float16 x) {
}
// acosh(x) = log(x + sqrt(x^2 - 1))
- float sqrt_term = fputil::sqrt<float>(fputil::multiply_add(xf, xf, -1.0f));
+ float sqrt_term = fputil::sqrt<float>(fputil::multiply_sub(xf, xf, 1.0f));
float result = static_cast<float>(log_eval(xf + sqrt_term));
return fputil::cast<float16>(result);
diff --git a/libc/src/__support/math/acoshf_utils.h b/libc/src/__support/math/acoshf_utils.h
index 3d760f98352490..5706d5a8652558 100644
--- a/libc/src/__support/math/acoshf_utils.h
+++ b/libc/src/__support/math/acoshf_utils.h
@@ -103,7 +103,7 @@ LIBC_INLINE LIBC_CONSTEXPR double log_eval(double x) {
// Perform exact range reduction
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
- dx = fputil::multiply_add(r, m, -1.0); // exact
+ dx = fputil::multiply_sub(r, m, 1.0); // exact
#else
uint64_t c_m = x_m & 0x3FFF'C000'0000'0000ULL;
double c = FPBits(c_m).get_val();
diff --git a/libc/src/__support/math/acospif.h b/libc/src/__support/math/acospif.h
index f193a34f95316e..ba937614beead6 100644
--- a/libc/src/__support/math/acospif.h
+++ b/libc/src/__support/math/acospif.h
@@ -78,7 +78,7 @@ LIBC_INLINE float acospif(float x) {
constexpr double C0_MINUS_1OVERPI =
(asin_internal::ASINPIF_COEFFS[0] - ONE_OVER_PI_HI) + ONE_OVER_PI_LO;
- double u = fputil::multiply_add(-0.5, x_abs, 0.5);
+ double u = fputil::sub_multiply(0.5, 0.5, x_abs);
double sqrt_u = fputil::sqrt<double>(u);
double neg2_sqrt_u = -2.0 * sqrt_u;
diff --git a/libc/src/__support/math/acospif16.h b/libc/src/__support/math/acospif16.h
index 4b5c28084f7094..40bf3b8b04387f 100644
--- a/libc/src/__support/math/acospif16.h
+++ b/libc/src/__support/math/acospif16.h
@@ -79,7 +79,7 @@ LIBC_INLINE constexpr float16 acospif16(float16 x) {
fputil::polyeval(xsq, POLY_COEFFS[0], POLY_COEFFS[1], POLY_COEFFS[2],
POLY_COEFFS[3], POLY_COEFFS[4]);
- return fputil::cast<float16>(fputil::multiply_add(-xf, interm, 0.5f));
+ return fputil::cast<float16>(fputil::sub_multiply(0.5f, xf, interm));
}
// When |x| > 0.5, assume that 0.5 < |x| <= 1
@@ -125,7 +125,7 @@ LIBC_INLINE constexpr float16 acospif16(float16 x) {
// with a positive x value instead (rhs).
float xf_abs = (xf < 0 ? -xf : xf);
- float u = fputil::multiply_add(-0.5f, xf_abs, 0.5f);
+ float u = fputil::sub_multiply(0.5f, 0.5f, xf_abs);
float sqrt_u = fputil::sqrt<float>(u);
float asin_sqrt_u =
@@ -134,7 +134,7 @@ LIBC_INLINE constexpr float16 acospif16(float16 x) {
// Same as acos(x), but devided the expression with pi
return fputil::cast<float16>(
- x_sign ? fputil::multiply_add(-2.0f, asin_sqrt_u, 1.0f)
+ x_sign ? fputil::sub_multiply(1.0f, 2.0f, asin_sqrt_u)
: 2.0f * asin_sqrt_u);
}
diff --git a/libc/src/__support/math/asin.h b/libc/src/__support/math/asin.h
index cf4c2736c3e209..3e66db92f17dde 100644
--- a/libc/src/__support/math/asin.h
+++ b/libc/src/__support/math/asin.h
@@ -67,7 +67,7 @@ LIBC_INLINE double asin(double x) {
double normalize_const = (x_exp == 0) ? eps : 0.0;
double scaled_normal =
fputil::multiply_add(x + normalize_const, 0x1.0p54, eps);
- return fputil::multiply_add(scaled_normal, 0x1.0p-54, -normalize_const);
+ return fputil::multiply_sub(scaled_normal, 0x1.0p-54, normalize_const);
#endif // LIBC_MATH_HAS_SKIP_ACCURATE_PASS
}
@@ -110,7 +110,7 @@ LIBC_INLINE double asin(double x) {
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
// u = x^2 - idx/64
DFloat128 u_hi(
- fputil::multiply_add(static_cast<double>(idx), -0x1.0p-6, x_sq.hi));
+ fputil::sub_multiply(x_sq.hi, static_cast<double>(idx), 0x1.0p-6));
DFloat128 u = fputil::quick_add(u_hi, DFloat128(x_sq.lo));
#else
DFloat128 x_sq_f128 = fputil::quick_mul(x_f128, x_f128);
@@ -176,7 +176,7 @@ LIBC_INLINE double asin(double x) {
// asin(x) ~ pi/2 - 2 * sqrt(u) * P(u),
// u = (1 - |x|)/2
- double u = fputil::multiply_add(x_abs, -0.5, 0.5);
+ double u = fputil::sub_multiply(0.5, x_abs, 0.5);
// v_hi + v_lo ~ sqrt(u).
// Let:
// h = u - v_hi^2 = (sqrt(u) - v_hi) * (sqrt(u) + v_hi)
@@ -197,7 +197,7 @@ LIBC_INLINE double asin(double x) {
#else
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
- double h = fputil::multiply_add(v_hi, -v_hi, u);
+ double h = fputil::sub_multiply(u, v_hi, v_hi);
#else
DoubleDouble v_hi_sq = fputil::exact_mult(v_hi, v_hi);
double h = (u - v_hi_sq.hi) - v_hi_sq.lo;
@@ -229,7 +229,7 @@ LIBC_INLINE double asin(double x) {
double r_upper = fputil::multiply_add(
r.hi, x_sign, fputil::multiply_add(r_lo, x_sign, err));
double r_lower = fputil::multiply_add(
- r.hi, x_sign, fputil::multiply_add(r_lo, x_sign, -err));
+ r.hi, x_sign, fputil::multiply_sub(r_lo, x_sign, err));
#else
r_lo *= x_sign;
r.hi *= x_sign;
@@ -264,7 +264,7 @@ LIBC_INLINE double asin(double x) {
// Get the rounding error of vl = 2 * v_lo ~ h / vh
// Get full product of vh * vl
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
- double vl_lo = fputil::multiply_add(-v_hi, vl, h) / v_hi;
+ double vl_lo = fputil::sub_multiply(h, v_hi, vl) / v_hi;
#else
DoubleDouble vh_vl = fputil::exact_mult(v_hi, vl);
double vl_lo = ((h - vh_vl.hi) - vh_vl.lo) / v_hi;
@@ -279,8 +279,7 @@ LIBC_INLINE double asin(double x) {
// Perform computations in DFloat128:
// asin(x) = pi/2 - (v_hi + v_lo + vll) * P(u).
- DFloat128 y_f128(
- fputil::multiply_add(static_cast<double>(idx), -0x1.0p-6, u));
+ DFloat128 y_f128(fputil::sub_multiply(u, static_cast<double>(idx), 0x1.0p-6));
DFloat128 p_f128 = asin_eval(y_f128, idx);
DFloat128 r0_f128 = fputil::quick_mul(m_v, p_f128);
diff --git a/libc/src/__support/math/asin_utils.h b/libc/src/__support/math/asin_utils.h
index 59d3df2e32b161..f0c81b32b6195d 100644
--- a/libc/src/__support/math/asin_utils.h
+++ b/libc/src/__support/math/asin_utils.h
@@ -237,7 +237,7 @@ LIBC_INLINE DoubleDouble asin_eval(const DoubleDouble &u, unsigned &idx,
double k = fputil::nearest_integer(u.hi * 0x1.0p5);
idx = static_cast<unsigned>(k);
// y = u - k/32.
- double y_hi = multiply_add(k, -0x1.0p-5, u.hi); // Exact
+ double y_hi = fputil::sub_multiply(u.hi, k, 0x1.0p-5); // Exact
DoubleDouble y = fputil::exact_add(y_hi, u.lo);
double y2 = y.hi * y.hi;
// Add double-double errors in addition to the relative errors from y2.
diff --git a/libc/src/__support/math/asinbf16.h b/libc/src/__support/math/asinbf16.h
index 9083ca9adcce91..dc67f642d602ab 100644
--- a/libc/src/__support/math/asinbf16.h
+++ b/libc/src/__support/math/asinbf16.h
@@ -67,10 +67,10 @@ LIBC_INLINE LIBC_CONSTEXPR bfloat16 asinbf16(bfloat16 x) {
return fputil::cast<bfloat16>(x_sign * PI_2);
}
- float t = fputil::multiply_add<float>(xf_abs, -0.5f, 0.5f);
+ float t = fputil::sub_multiply<float>(0.5f, xf_abs, 0.5f);
float t_sqrt = fputil::sqrt<float>(t);
float asin_sqrt_t = t_sqrt * asin_internal::asin_eval_float(t);
- float result = fputil::multiply_add<float>(-2.0f, asin_sqrt_t, PI_2);
+ float result = fputil::sub_multiply<float>(PI_2, 2.0f, asin_sqrt_t);
return fputil::cast<bfloat16>(x_sign * result);
}
diff --git a/libc/src/__support/math/asinf.h b/libc/src/__support/math/asinf.h
index fbcac68dfb55c7..9c452579b34b44 100644
--- a/libc/src/__support/math/asinf.h
+++ b/libc/src/__support/math/asinf.h
@@ -122,7 +122,7 @@ LIBC_INLINE float asinf(float x) {
xbits.set_sign(Sign::POS);
double sign_two = TWO[x_sign]; // sign * (-2)
- double uf = fputil::multiply_add(-0.5f, xbits.get_val(), 0.5f);
+ double uf = fputil::sub_multiply(0.5f, 0.5f, xbits.get_val());
double u = static_cast<double>(uf);
double c1 = sign_two * fputil::sqrt<double>(u);
double c2 = fputil::multiply_add(sign_two, M_PI_OVER_4, c1);
diff --git a/libc/src/__support/math/asinf16.h b/libc/src/__support/math/asinf16.h
index 367a034521146d..943faa5e9f665d 100644
--- a/libc/src/__support/math/asinf16.h
+++ b/libc/src/__support/math/asinf16.h
@@ -126,7 +126,7 @@ LIBC_INLINE constexpr float16 asinf16(float16 x) {
// 0x1p-1 < |x| <= 0x1p0, 0.5 < |x| <= 1.0
float xf_abs = (xf < 0 ? -xf : xf);
float sign = (xbits.uintval() >> 15 == 1 ? -1.0 : 1.0);
- float u = fputil::multiply_add(-0.5f, xf_abs, 0.5f);
+ float u = fputil::sub_multiply(0.5f, 0.5f, xf_abs);
float u_sqrt = fputil::sqrt<float>(u);
// Degree-6 minimax odd polynomial of asin(x) generated by Sollya with:
@@ -136,7 +136,7 @@ LIBC_INLINE constexpr float16 asinf16(float16 x) {
0x1.3541ccp-4f, 0x1.43b2d6p-5f, 0x1.a0d73ep-5f);
return fputil::cast<float16>(sign *
- fputil::multiply_add(-2.0f, asin_sqrt_u, PI_2));
+ fputil::sub_multiply(PI_2, 2.0f, asin_sqrt_u));
}
} // namespace math
diff --git a/libc/src/__support/math/asinpi.h b/libc/src/__support/math/asinpi.h
index 9ed3cc419dd5ea..e9fd001e5d5bb6 100644
--- a/libc/src/__support/math/asinpi.h
+++ b/libc/src/__support/math/asinpi.h
@@ -147,7 +147,7 @@ LIBC_INLINE double asinpi(double x) {
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
DFloat128 u_hi(
- fputil::multiply_add(static_cast<double>(idx), -0x1.0p-6, x_sq.hi));
+ fputil::sub_multiply(x_sq.hi, static_cast<double>(idx), 0x1.0p-6));
DFloat128 u = fputil::quick_add(u_hi, DFloat128(x_sq.lo));
#else
DFloat128 x_sq_f128 = fputil::quick_mul(x_f128, x_f128);
@@ -200,7 +200,7 @@ LIBC_INLINE double asinpi(double x) {
// where u = (1 - |x|) / 2.
// u = (1 - |x|)/2
- double u = fputil::multiply_add(x_abs, -0.5, 0.5);
+ double u = fputil::sub_multiply(0.5, x_abs, 0.5);
// v_hi ~ sqrt(u).
double v_hi = fputil::sqrt<double>(u);
@@ -215,7 +215,7 @@ LIBC_INLINE double asinpi(double x) {
using DoubleDouble = fputil::DoubleDouble;
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
- double h = fputil::multiply_add(v_hi, -v_hi, u);
+ double h = fputil::sub_multiply(u, v_hi, v_hi);
#else
DoubleDouble v_hi_sq = fputil::exact_mult(v_hi, v_hi);
double h = (u - v_hi_sq.hi) - v_hi_sq.lo;
@@ -247,7 +247,7 @@ LIBC_INLINE double asinpi(double x) {
double r_upper = fputil::multiply_add(
r.hi, x_sign, fputil::multiply_add(r_lo, x_sign, err));
double r_lower = fputil::multiply_add(
- r.hi, x_sign, fputil::multiply_add(r_lo, x_sign, -err));
+ r.hi, x_sign, fputil::multiply_sub(r_lo, x_sign, err));
#else
r_lo *= x_sign;
r.hi *= x_sign;
@@ -270,7 +270,7 @@ LIBC_INLINE double asinpi(double x) {
// Get the rounding error of vl = 2 * v_lo ~ h / vh
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
- double vl_lo = fputil::multiply_add(-v_hi, vl, h) / v_hi;
+ double vl_lo = fputil::sub_multiply(h, v_hi, vl) / v_hi;
#else
DoubleDouble vh_vl = fputil::exact_mult(v_hi, vl);
double vl_lo = ((h - vh_vl.hi) - vh_vl.lo) / v_hi;
@@ -285,8 +285,7 @@ LIBC_INLINE double asinpi(double x) {
// Perform computations in DFloat128:
// asinpi(x) = 0.5 - (v_hi + v_lo + vll) * P_pi(u).
- DFloat128 y_f128(
- fputil::multiply_add(static_cast<double>(idx), -0x1.0p-6, u));
+ DFloat128 y_f128(fputil::sub_multiply(u, static_cast<double>(idx), 0x1.0p-6));
DFloat128 p_f128 = asinpi_eval(y_f128, idx);
DFloat128 r0_f128 = fputil::quick_mul(m_v, p_f128);
diff --git a/libc/src/__support/math/asinpif.h b/libc/src/__support/math/asinpif.h
index 0a823c2bce627d..cd3096940121c6 100644
--- a/libc/src/__support/math/asinpif.h
+++ b/libc/src/__support/math/asinpif.h
@@ -76,7 +76,7 @@ LIBC_INLINE float asinpif(float x) {
constexpr double C0_MINUS_1OVERPI =
(asin_internal::ASINPIF_COEFFS[0] - ONE_OVER_PI_HI) + ONE_OVER_PI_LO;
- double u = fputil::multiply_add(-0.5, x_abs, 0.5);
+ double u = fputil::sub_multiply(0.5, 0.5, x_abs);
double sqrt_u = fputil::sqrt<double>(u);
double neg2_sqrt_u = -2.0 * sqrt_u;
diff --git a/libc/src/__support/math/asinpif16.h b/libc/src/__support/math/asinpif16.h
index 7aefc35f82d01f..fcee574850c51f 100644
--- a/libc/src/__support/math/asinpif16.h
+++ b/libc/src/__support/math/asinpif16.h
@@ -124,9 +124,9 @@ LIBC_INLINE float16 asinpif16(float16 x) {
// = 0.5 - 0.5 * x
// = multiply_add(-0.5, x, 0.5)
- double u = fputil::multiply_add(-0.5, x_abs, 0.5);
+ double u = fputil::sub_multiply(0.5, 0.5, x_abs);
double asinpi_sqrt_u = asinpi_polyeval(fputil::sqrt<double>(u));
- double result = fputil::multiply_add(-2.0, asinpi_sqrt_u, 0.5);
+ double result = fputil::sub_multiply(0.5, 2.0, asinpi_sqrt_u);
return fputil::cast<float16>(signed_result(result));
}
diff --git a/libc/src/__support/math/atan.h b/libc/src/__support/math/atan.h
index 8a11167a6268ef..27c412accb555a 100644
--- a/libc/src/__support/math/atan.h
+++ b/libc/src/__support/math/atan.h
@@ -79,7 +79,7 @@ LIBC_INLINE double atan(double x) {
if (x == 0.0)
return x;
// |x| < 2^-26
- return fputil::multiply_add(-0x1.0p-54, x, x);
+ return fputil::sub_multiply(x, 0x1.0p-54, x);
#endif // LIBC_MATH_HAS_SKIP_ACCURATE_PASS
}
@@ -153,7 +153,7 @@ LIBC_INLINE double atan(double x) {
DoubleDouble den = fputil::exact_add(x_d, k);
// numerator = 1 - k * |x|
DoubleDouble num;
- num.hi = fputil::multiply_add(-x_d, k, 1.0);
+ num.hi = fputil::sub_multiply(1.0, x_d, k);
DoubleDouble prod = fputil::exact_mult(x_d, k);
// Using Dekker's 2SUM algorithm to compute the lower part.
num.lo = ((1.0 - num.hi) - prod.hi) - prod.lo;
diff --git a/libc/src/__support/math/atan2f.h b/libc/src/__support/math/atan2f.h
index ff2419e69c6aed..2721f165fc7315 100644
--- a/libc/src/__support/math/atan2f.h
+++ b/libc/src/__support/math/atan2f.h
@@ -127,8 +127,8 @@ LIBC_INLINE float atan2f_double_double(double num_d, double den_d, double q_d,
// atan(n/d) - atan(idx/16) = atan((n/d - idx/16) / (1 + (n/d) * (idx/16)))
// = atan((n - d*(idx/16)) / (d + n*idx/16))
k_d *= 0x1.0p-4;
- num_r = fputil::multiply_add(k_d, -den_d, num_d); // Exact
- den_r = fputil::multiply_add(k_d, num_d, den_d); // Exact
+ num_r = fputil::sub_multiply(num_d, k_d, den_d); // Exact
+ den_r = fputil::multiply_add(k_d, num_d, den_d); // Exact
q.hi = num_r / den_r;
} else {
// For 0 < n/d < 1/16, we just need to calculate the lower part of their
@@ -138,13 +138,13 @@ LIBC_INLINE float atan2f_double_double(double num_d, double den_d, double q_d,
den_r = den_d;
}
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
- q.lo = fputil::multiply_add(q.hi, -den_r, num_r) / den_r;
+ q.lo = fputil::sub_multiply(num_r, q.hi, den_r) / den_r;
#else
// Compute `(num_r - q.hi * den_r) / den_r` accurately without FMA
// instructions.
fputil::DoubleDouble q_hi_dd = split_d(q.hi);
- double t1 = fputil::multiply_add(q_hi_dd.hi, -den_r, num_r); // Exact
- double t2 = fputil::multiply_add(q_hi_dd.lo, -den_r, t1);
+ double t1 = fputil::sub_multiply(num_r, q_hi_dd.hi, den_r); // Exact
+ double t2 = fputil::sub_multiply(t1, q_hi_dd.lo, den_r);
q.lo = t2 / den_r;
#endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE
@@ -314,7 +314,7 @@ LIBC_INLINE float atan2f(float y, float x) {
double p = atan_eval_no_table(num_d, den_d, k_d * 0x1.0p-4);
r = final_sign * (p + (const_term.hi + ATAN_K_OVER_16[idx]));
#else
- q_d = fputil::multiply_add(k_d, -0x1.0p-4, q_d);
+ q_d = fputil::sub_multiply(q_d, k_d, 0x1.0p-4);
double p = atan_eval(q_d, idx);
r = final_sign *
diff --git a/libc/src/__support/math/atan2f128.h b/libc/src/__support/math/atan2f128.h
index 5df9071acc51bc..f17e8d113a3eda 100644
--- a/libc/src/__support/math/atan2f128.h
+++ b/libc/src/__support/math/atan2f128.h
@@ -187,7 +187,7 @@ LIBC_INLINE Float128 atan2f128(Float128 y, Float128 x) {
// atan(n/d) - atan(k) = atan((n/d - k/64) / (1 + (n/d) * (k/64)))
// = atan((n - d * k/64)) / (d + n * k/64))
// num_f128 = n - d * k/64
- DFloat128 num_f128 = fputil::multiply_add(den, -k_f128, num);
+ DFloat128 num_f128 = fputil::sub_multiply(num, den, k_f128);
// den_f128 = d + n * k/64
DFloat128 den_f128 = fputil::multiply_add(num, k_f128, den);
diff --git a/libc/src/__support/math/atan2f16.h b/libc/src/__support/math/atan2f16.h
index 0621ab4fd05477..c7570bd544e2d0 100644
--- a/libc/src/__support/math/atan2f16.h
+++ b/libc/src/__support/math/atan2f16.h
@@ -91,7 +91,7 @@ LIBC_INLINE float16 atan2f16(float16 y, float16 x) {
double k_d = fputil::nearest_integer(q_d * 0x1.0p4);
int idx = static_cast<int>(k_d);
- q_d = fputil::multiply_add(k_d, -0x1.0p-4, q_d);
+ q_d = fputil::sub_multiply(q_d, k_d, 0x1.0p-4);
double p = atan_eval(q_d, static_cast<unsigned>(idx));
double r = final_sign *
diff --git a/libc/src/__support/math/atan2f_float.h b/libc/src/__support/math/atan2f_float.h
index 744783e855cef6..58cd97878fa5b2 100644
--- a/libc/src/__support/math/atan2f_float.h
+++ b/libc/src/__support/math/atan2f_float.h
@@ -67,7 +67,7 @@ LIBC_INLINE constexpr FloatFloat atan_eval(const FloatFloat &x) {
// c0 ~ - x_hi^2 / 3
float c0 = -0x1.555556p-2f * x_hi_sq;
// c1 ~ x_lo * (1 - x_hi^2)
- float c1 = fputil::multiply_add(x_hi_sq, -x.lo, x.lo);
+ float c1 = fputil::sub_multiply(x.lo, x_hi_sq, x.lo);
// p.lo ~ - x_hi^3 / 3 + x_lo * (1 - x_hi*2)
p.lo = fputil::multiply_add(x.hi, c0, c1);
return p;
diff --git a/libc/src/__support/math/atan_utils.h b/libc/src/__support/math/atan_utils.h
index 60d25e7d7374f9..ff857402dfa9dd 100644
--- a/libc/src/__support/math/atan_utils.h
+++ b/libc/src/__support/math/atan_utils.h
@@ -115,11 +115,11 @@ LIBC_INLINE_VAR constexpr DoubleDouble ATAN_I[65] = {
p.hi = x.hi;
double x_hi_sq = x.hi * x.hi;
// c0 ~ x_hi^2 * 1/5 - 1/3
- double c0 = fputil::multiply_add(x_hi_sq, 0x1.999999999999ap-3,
- -0x1.5555555555555p-2);
+ double c0 =
+ fputil::multiply_sub(x_hi_sq, 0x1.999999999999ap-3, 0x1.5555555555555p-2);
// c1 ~ x_hi^2 * 1/9 - 1/7
- double c1 = fputil::multiply_add(x_hi_sq, 0x1.c71c71c71c71cp-4,
- -0x1.2492492492492p-3);
+ double c1 =
+ fputil::multiply_sub(x_hi_sq, 0x1.c71c71c71c71cp-4, 0x1.2492492492492p-3);
// x_hi^3
double x_hi_3 = x_hi_sq * x.hi;
// x_hi^4
@@ -340,7 +340,7 @@ LIBC_INLINE double atan_eval(double x, unsigned i) {
// atan(q) ~ q - q^3/3 + q^5/5 - q^7/7 + q^9/9
LIBC_INLINE double atan_eval_no_table(double num, double den,
double k_over_16) {
- double num_r = fputil::multiply_add(den, -k_over_16, num);
+ double num_r = fputil::sub_multiply(num, den, k_over_16);
double den_r = fputil::multiply_add(num, k_over_16, den);
double q = num_r / den_r;
diff --git a/libc/src/__support/math/atanbf16.h b/libc/src/__support/math/atanbf16.h
index 796420a6cf4b62..a1e4605269072d 100644
--- a/libc/src/__support/math/atanbf16.h
+++ b/libc/src/__support/math/atanbf16.h
@@ -69,7 +69,7 @@ LIBC_INLINE bfloat16 atanbf16(bfloat16 x) {
if (LIBC_UNLIKELY(x_abs <= 0x3db8)) {
// Keep the small correction normal even for the smallest bfloat16 input.
double xd = xf;
- return fputil::cast<bfloat16>(fputil::multiply_add(xd, -0x1p-9, xd));
+ return fputil::cast<bfloat16>(fputil::sub_multiply(xd, xd, 0x1p-9));
}
float x_sq = xf * xf;
diff --git a/libc/src/__support/math/atanf.h b/libc/src/__support/math/atanf.h
index 4f185650322517..fa3d7d2ea62bfd 100644
--- a/libc/src/__support/math/atanf.h
+++ b/libc/src/__support/math/atanf.h
@@ -65,9 +65,9 @@ LIBC_INLINE float atanf(float x) {
// x <= 2^-12;
if (LIBC_UNLIKELY(x_abs < 0x3980'0000)) {
#if defined(LIBC_TARGET_CPU_HAS_FMA_FLOAT)
- return fputil::multiply_add(x, -0x1.0p-25f, x);
+ return fputil::sub_multiply(x, x, 0x1.0p-25f);
#else
- return static_cast<float>(fputil::multiply_add(x_d, -0x1.0p-25, x_d));
+ return static_cast<float>(fputil::sub_multiply(x_d, x_d, 0x1.0p-25));
#endif // LIBC_TARGET_CPU_HAS_FMA_FLOAT
}
// Use Taylor polynomial:
@@ -101,7 +101,7 @@ LIBC_INLINE float atanf(float x) {
// double precision.
x_d = 1.0 / static_cast<double>(x_bits.get_val());
double k_d = fputil::nearest_integer(x_d * 0x1.0p4);
- x_d = fputil::multiply_add(k_d, -0x1.0p-4, x_d);
+ x_d = fputil::sub_multiply(x_d, k_d, 0x1.0p-4);
idx = static_cast<int>(k_d);
final_sign = FINAL_SIGN[sign.is_pos()];
// Adjust constant term of the polynomial by +- pi/2.
@@ -116,7 +116,7 @@ LIBC_INLINE float atanf(float x) {
// Perform range reduction in single precision.
float x_f = x_bits.get_val();
float k_f = fputil::nearest_integer(x_f * 0x1.0p4f);
- x_f = fputil::multiply_add(k_f, -0x1.0p-4f, x_f);
+ x_f = fputil::sub_multiply(x_f, k_f, 0x1.0p-4f);
x_d = static_cast<double>(x_f);
idx = static_cast<int>(k_f);
final_sign = FINAL_SIGN[sign.is_neg()];
diff --git a/libc/src/__support/math/atanf16.h b/libc/src/__support/math/atanf16.h
index 01eb8770591ebd..92c9a4de581172 100644
--- a/libc/src/__support/math/atanf16.h
+++ b/libc/src/__support/math/atanf16.h
@@ -107,7 +107,7 @@ LIBC_INLINE constexpr float16 atanf16(float16 x) {
-0x1.e65492p-5f, 0x1.8c0c36p-6f, -0x1.32316ep-8f);
return fputil::cast<float16>(sign *
- fputil::multiply_add(x_inv, -interm, PI_2));
+ fputil::sub_multiply(PI_2, x_inv, interm));
}
} // namespace math
diff --git a/libc/src/__support/math/atanf_float.h b/libc/src/__support/math/atanf_float.h
index 56db2ad2955e61..f3a10380a5256c 100644
--- a/libc/src/__support/math/atanf_float.h
+++ b/libc/src/__support/math/atanf_float.h
@@ -70,12 +70,12 @@ LIBC_INLINE float atanf_eval(float u, float k_over_16) {
// (k/16)^2
float c2 = k_over_16 * k_over_16;
// -(k/16)^3
- float c3 = fputil::multiply_add(-k_over_16, c2, k_over_16);
+ float c3 = fputil::sub_multiply(k_over_16, k_over_16, c2);
float u2 = u * u;
// (k^2/256 - 1/3) + u * (k/16 - (k/16)^3)
float a0 = fputil::multiply_add(c3, u, c2 - 0x1.555556p-2f);
// -k/16 + u*(k^2/256 - 1/3) + u^2 * (k/16 - (k/16)^3)
- float a1 = fputil::multiply_add(u, a0, -k_over_16);
+ float a1 = fputil::multiply_sub(u, a0, k_over_16);
// u - u^2 * k/16 + u^3 * ((k^2/256 - 1/3) + u^4 * (k/16 - (k/16)^3))
return fputil::multiply_add(u2, a1, u);
}
@@ -115,7 +115,7 @@ LIBC_INLINE float atanf(float x) {
// atan(x) = sign(x) * pi/2 - atan(1/x)
// ~ sign(x) * pi/2 - 1/x
if (LIBC_UNLIKELY(x_abs >= 0x4200'0000))
- return fputil::multiply_add(sign, PI_OVER_2.hi, -1.0f / x);
+ return fputil::multiply_sub(sign, PI_OVER_2.hi, 1.0f / x);
// x <= 2^-11:
// atan(x) ~ x
return x;
diff --git a/libc/src/__support/math/atanpif16.h b/libc/src/__support/math/atanpif16.h
index e0eaff6678fc58..faac2b62fb1cb5 100644
--- a/libc/src/__support/math/atanpif16.h
+++ b/libc/src/__support/math/atanpif16.h
@@ -151,7 +151,7 @@ LIBC_INLINE float16 atanpif16(float16 x) {
double x_recip_sq = x_recip * x_recip;
double sqrt_term = fputil::sqrt<double>(1.0 + x_recip_sq);
double x_prime = x_recip / (1.0 + sqrt_term);
- result = fputil::multiply_add(-2.0, atanpi_eval(x_prime), 0.5);
+ result = fputil::sub_multiply(0.5, 2.0, atanpi_eval(x_prime));
} else {
// direct evaluation since 1/|x| <= 0.5
result = 0.5 - atanpi_eval(x_recip);
diff --git a/libc/src/__support/math/cbrt.h b/libc/src/__support/math/cbrt.h
index c6631e6f67e2cb..0e12f887d4ac19 100644
--- a/libc/src/__support/math/cbrt.h
+++ b/libc/src/__support/math/cbrt.h
@@ -61,7 +61,7 @@ LIBC_INLINE double initial_approximation(double x) {
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
LIBC_INLINE double get_error(const DoubleDouble &x_3,
const DoubleDouble &a_sq) {
- return fputil::multiply_add(x_3.hi, a_sq.hi, -1.0) +
+ return fputil::multiply_sub(x_3.hi, a_sq.hi, 1.0) +
fputil::multiply_add(x_3.lo, a_sq.hi, x_3.hi * a_sq.lo);
}
#else
@@ -288,7 +288,7 @@ LIBC_INLINE double cbrt(double x) {
DoubleDouble r2 = fputil::quick_mult(a, x2);
double r2_upper = r2.hi + fputil::multiply_add(a, 0x1.0p-102, r2.lo);
- double r2_lower = r2.hi + fputil::multiply_add(a, -0x1.0p-102, r2.lo);
+ double r2_lower = r2.hi + fputil::sub_multiply(r2.lo, a, 0x1.0p-102);
// Ziv's accuracy test.
if (LIBC_LIKELY(r2_upper == r2_lower))
diff --git a/libc/src/__support/math/cosf16.h b/libc/src/__support/math/cosf16.h
index f7485a074ac639..2965ffa1f8581d 100644
--- a/libc/src/__support/math/cosf16.h
+++ b/libc/src/__support/math/cosf16.h
@@ -94,7 +94,7 @@ LIBC_INLINE constexpr float16 cosf16(float16 x) {
// = cos_k * (cos_y - 1 + 1) - sin_k * sin_y
// = cos_k * cosm1_y - sin_k * sin_y + cos_k
return fputil::cast<float16>(fputil::multiply_add(
- cos_k, cosm1_y, fputil::multiply_add(-sin_k, sin_y, cos_k)));
+ cos_k, cosm1_y, fputil::sub_multiply(cos_k, sin_k, sin_y)));
}
} // namespace math
diff --git a/libc/src/__support/math/cosf_double_eval.h b/libc/src/__support/math/cosf_double_eval.h
index aa7132df937869..d44e795e92b195 100644
--- a/libc/src/__support/math/cosf_double_eval.h
+++ b/libc/src/__support/math/cosf_double_eval.h
@@ -115,10 +115,10 @@ LIBC_INLINE float cosf(float x) {
// double for intermediate results as it is more efficient than using an
// emulated version of FMA.
#if defined(LIBC_TARGET_CPU_HAS_FMA_FLOAT)
- return fputil::multiply_add(xbits.get_val(), -0x1.0p-25f, 1.0f);
+ return fputil::sub_multiply(1.0f, xbits.get_val(), 0x1.0p-25f);
#else // !LIBC_TARGET_CPU_HAS_FMA_FLOAT
double xd = static_cast<double>(xbits.get_val());
- return static_cast<float>(fputil::multiply_add(xd, -0x1.0p-25, 1.0));
+ return static_cast<float>(fputil::sub_multiply(1.0, xd, 0x1.0p-25));
#endif // LIBC_TARGET_CPU_HAS_FMA_FLOAT
}
@@ -150,8 +150,8 @@ LIBC_INLINE float cosf(float x) {
sincosf_eval(xd, x_abs, sin_k, cos_k, sin_y, cosm1_y);
- return static_cast<float>(fputil::multiply_add(
- sin_y, -sin_k, fputil::multiply_add(cosm1_y, cos_k, cos_k)));
+ return static_cast<float>(fputil::sub_multiply(
+ fputil::multiply_add(cosm1_y, cos_k, cos_k), sin_y, sin_k));
}
} // namespace double_eval
diff --git a/libc/src/__support/math/cospif.h b/libc/src/__support/math/cospif.h
index 49100feba32308..7f9c81b22f8555 100644
--- a/libc/src/__support/math/cospif.h
+++ b/libc/src/__support/math/cospif.h
@@ -56,9 +56,9 @@ LIBC_INLINE float cospif(float x) {
if (LIBC_UNLIKELY(x_abs < 0x38A2'F984U)) {
#if defined(LIBC_TARGET_CPU_HAS_FMA_FLOAT)
- return fputil::multiply_add(xbits.get_val(), -0x1.0p-25f, 1.0f);
+ return fputil::sub_multiply(1.0f, xbits.get_val(), 0x1.0p-25f);
#else
- return static_cast<float>(fputil::multiply_add(xd, -0x1.0p-25, 1.0));
+ return static_cast<float>(fputil::sub_multiply(1.0, xd, 0x1.0p-25));
#endif // LIBC_TARGET_CPU_HAS_FMA_FLOAT
}
@@ -99,8 +99,8 @@ LIBC_INLINE float cospif(float x) {
return 0.0f;
}
- return static_cast<float>(fputil::multiply_add(
- sin_y, -sin_k, fputil::multiply_add(cosm1_y, cos_k, cos_k)));
+ return static_cast<float>(fputil::sub_multiply(
+ fputil::multiply_add(cosm1_y, cos_k, cos_k), sin_y, sin_k));
}
} // namespace math
diff --git a/libc/src/__support/math/cospif16.h b/libc/src/__support/math/cospif16.h
index 8cdbe108a5bbae..02bc38569c83d6 100644
--- a/libc/src/__support/math/cospif16.h
+++ b/libc/src/__support/math/cospif16.h
@@ -95,7 +95,7 @@ LIBC_INLINE constexpr float16 cospif16(float16 x) {
// Since, cosm1_y = cos_y - 1, therefore:
// cos(x * pi) = cos_k(cosm1_y) + cos_k - sin_k * sin_y
return fputil::cast<float16>(fputil::multiply_add(
- cos_k, cosm1_y, fputil::multiply_add(-sin_k, sin_y, cos_k)));
+ cos_k, cosm1_y, fputil::sub_multiply(cos_k, sin_k, sin_y)));
}
} // namespace math
diff --git a/libc/src/__support/math/exp10f16_utils.h b/libc/src/__support/math/exp10f16_utils.h
index a0fca6399a78fa..87d81d060f2eda 100644
--- a/libc/src/__support/math/exp10f16_utils.h
+++ b/libc/src/__support/math/exp10f16_utils.h
@@ -40,7 +40,7 @@ LIBC_INLINE ExpRangeReduction exp10_range_reduction(float16 x) {
unsigned x_hi = static_cast<unsigned>(x_hi_mid) >> 3;
unsigned x_mid = static_cast<unsigned>(x_hi_mid) & 0x7;
// lo = x - (hi + mid) = round(x * 2^3 * log2(10)) * log10(2) * (-2^(-3)) + x
- float lo = fputil::multiply_add(kf, LOG10F_2 * -0x1.0p-3f, xf);
+ float lo = fputil::sub_multiply(xf, kf, LOG10F_2 * 0x1.0p-3f);
uint32_t exp2_hi_mid_bits =
EXP2_MID_BITS[x_mid] +
diff --git a/libc/src/__support/math/exp10f_float_eval.h b/libc/src/__support/math/exp10f_float_eval.h
index 700397a85bd2b9..045ea008868c24 100644
--- a/libc/src/__support/math/exp10f_float_eval.h
+++ b/libc/src/__support/math/exp10f_float_eval.h
@@ -105,7 +105,7 @@ LIBC_INLINE float exp10f(float x) {
float kf = fputil::nearest_integer(x * LOG2_10.hi);
int k = static_cast<int>(kf);
- float u_hi = fputil::multiply_add(x, LOG2_10.hi, -kf);
+ float u_hi = fputil::multiply_sub(x, LOG2_10.hi, kf);
float u = fputil::multiply_add(x, LOG2_10.lo, u_hi);
#else // !LIBC_TARGET_CPU_HAS_FMA_FLOAT
// Cody-Waite reduction for non-FMA targets:
@@ -121,8 +121,8 @@ LIBC_INLINE float exp10f(float x) {
float kf = fputil::nearest_integer(x * LOG2_10);
int k = static_cast<int>(kf);
- float v_hi = fputil::multiply_add(-kf, LOG10_2_HI, x);
- float v = fputil::multiply_add(-kf, LOG10_2_LO, v_hi);
+ float v_hi = fputil::sub_multiply(x, kf, LOG10_2_HI);
+ float v = fputil::sub_multiply(v_hi, kf, LOG10_2_LO);
// Convert reduced argument to base-2: u = v * log2(10)
float u = v * LOG2_10;
diff --git a/libc/src/__support/math/exp10m1f.h b/libc/src/__support/math/exp10m1f.h
index 4039b708cc09f2..a4604565f4db59 100644
--- a/libc/src/__support/math/exp10m1f.h
+++ b/libc/src/__support/math/exp10m1f.h
@@ -244,7 +244,7 @@ LIBC_INLINE float exp10m1f(float x) {
double exp10_lo = fputil::polyeval(lo_sq, c0, c1, c2);
// 10^x - 1 = 2^(mid + hi) * 10^lo - 1
// ~ mh * exp10_lo - 1
- return static_cast<float>(fputil::multiply_add(exp10_lo, rr.mh, -1.0));
+ return static_cast<float>(fputil::multiply_sub(exp10_lo, rr.mh, 1.0));
}
} // namespace math
diff --git a/libc/src/__support/math/exp10m1f16.h b/libc/src/__support/math/exp10m1f16.h
index 5d31ca6da90e4e..e14fb233cb9659 100644
--- a/libc/src/__support/math/exp10m1f16.h
+++ b/libc/src/__support/math/exp10m1f16.h
@@ -172,7 +172,7 @@ LIBC_INLINE constexpr float16 exp10m1f16(float16 x) {
auto [exp2_hi_mid, exp10_lo] = exp10_range_reduction(x);
// exp10m1(x) = exp2((hi + mid) * log2(lo)) * exp10(lo) - 1
return fputil::cast<float16>(
- fputil::multiply_add(exp2_hi_mid, exp10_lo, -1.0f));
+ fputil::multiply_sub(exp2_hi_mid, exp10_lo, 1.0f));
}
} // namespace math
diff --git a/libc/src/__support/math/exp2_double_eval.h b/libc/src/__support/math/exp2_double_eval.h
index a462e2038e96aa..12eca4774c5b1b 100644
--- a/libc/src/__support/math/exp2_double_eval.h
+++ b/libc/src/__support/math/exp2_double_eval.h
@@ -82,7 +82,7 @@ LIBC_ALWAYS_INLINE double exp2(double x) {
int hi = static_cast<int32_t>(k_bits >> 19) >> (13 + EXP_N_BITS);
unsigned idx = static_cast<unsigned>(k_bits >> 32) & EXP_MASK;
- double dx = fputil::multiply_add(kd, -EXP_SCALE_INV, x); // exact
+ double dx = fputil::sub_multiply(x, kd, EXP_SCALE_INV); // exact
#ifdef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
double u = dx * LOG_2_DD.hi;
diff --git a/libc/src/__support/math/exp2bf16.h b/libc/src/__support/math/exp2bf16.h
index 1dcd05b82f3e1d..2990a261ebc2dd 100644
--- a/libc/src/__support/math/exp2bf16.h
+++ b/libc/src/__support/math/exp2bf16.h
@@ -144,7 +144,7 @@ LIBC_INLINE bfloat16 exp2bf16(bfloat16 x) {
// lo = x - (hi + mid) = x - k / 2^n
// = k * (-2^(-n)) + x
- float lo = fputil::multiply_add(k_float, -0.0625f, x_float);
+ float lo = fputil::sub_multiply(x_float, k_float, 0.0625f);
float exp2_mid = exp2bf16_internal::EXP2_MID[j];
diff --git a/libc/src/__support/math/exp2f_double_eval.h b/libc/src/__support/math/exp2f_double_eval.h
index b69dbf65ef18d7..85efd0bec0cb7f 100644
--- a/libc/src/__support/math/exp2f_double_eval.h
+++ b/libc/src/__support/math/exp2f_double_eval.h
@@ -54,7 +54,7 @@ LIBC_INLINE float exp2f(float x) {
#endif // LIBC_TARGET_CPU_HAS_NEAREST_INT
// dx = lo = x - (hi + mid) = x - kf * 2^(-5)
- double dx = fputil::multiply_add(-0x1.0p-5f, kf, x);
+ double dx = fputil::sub_multiply(x, 0x1.0p-5f, kf);
// hi = floor(kf * 2^(-4))
// exp_hi = shift hi to the exponent field of double precision.
diff --git a/libc/src/__support/math/exp2f_float_eval.h b/libc/src/__support/math/exp2f_float_eval.h
index fdd8167a098661..f32a65cc5c605f 100644
--- a/libc/src/__support/math/exp2f_float_eval.h
+++ b/libc/src/__support/math/exp2f_float_eval.h
@@ -68,7 +68,7 @@ LIBC_INLINE float exp2f(float x) {
float kf = fputil::nearest_integer(x);
int k = static_cast<int>(kf);
- float u = fputil::multiply_add(kf, -1.0f, x);
+ float u = fputil::sub_multiply(x, kf, 1.0f);
return exp2f_eval(u, k);
}
diff --git a/libc/src/__support/math/exp2m1f.h b/libc/src/__support/math/exp2m1f.h
index 3b5cfb05918f2f..bed44e36066833 100644
--- a/libc/src/__support/math/exp2m1f.h
+++ b/libc/src/__support/math/exp2m1f.h
@@ -154,7 +154,7 @@ LIBC_INLINE float exp2m1f(float x) {
#endif // LIBC_TARGET_CPU_HAS_NEAREST_INT
// lo = x - (hi + mid) = x - kf * 2^(-5)
- double lo = fputil::multiply_add(-0x1.0p-5f, kf, x);
+ double lo = fputil::sub_multiply(x, 0x1.0p-5f, kf);
// hi = floor(kf * 2^(-4))
// exp2_hi = shift hi to the exponent field of double precision.
@@ -181,7 +181,7 @@ LIBC_INLINE float exp2m1f(float x) {
// = 2^(hi + mid) * 2^lo - 1
// ~ mh * (1 + lo * P(lo)) - 1
// = mh * exp2_lo - 1
- return static_cast<float>(fputil::multiply_add(exp2_lo, mh, -1.0));
+ return static_cast<float>(fputil::multiply_sub(exp2_lo, mh, 1.0));
}
} // namespace math
diff --git a/libc/src/__support/math/exp2m1f16.h b/libc/src/__support/math/exp2m1f16.h
index 09f349e15bb3a1..61adb79790b13f 100644
--- a/libc/src/__support/math/exp2m1f16.h
+++ b/libc/src/__support/math/exp2m1f16.h
@@ -174,7 +174,7 @@ LIBC_INLINE constexpr float16 exp2m1f16(float16 x) {
auto [exp2_hi_mid, exp2_lo] = exp2_range_reduction(x);
// exp2m1(x) = exp2(hi + mid) * exp2(lo) - 1
return fputil::cast<float16>(
- fputil::multiply_add(exp2_hi_mid, exp2_lo, -1.0f));
+ fputil::multiply_sub(exp2_hi_mid, exp2_lo, 1.0f));
}
} // namespace math
diff --git a/libc/src/__support/math/expbf16.h b/libc/src/__support/math/expbf16.h
index 2fdb5c3192583c..c186b9beea7776 100644
--- a/libc/src/__support/math/expbf16.h
+++ b/libc/src/__support/math/expbf16.h
@@ -174,7 +174,7 @@ LIBC_INLINE bfloat16 expbf16(bfloat16 x) {
int x_hi = x_hi_mid >> 5;
int x_mid = x_hi_mid & 0b11111;
// lo = x - (hi + mid) = round(x * 4) / (-4) + x
- float lo = fputil::multiply_add(kf, -0.25f, xf);
+ float lo = fputil::sub_multiply(xf, kf, 0.25f);
float exp_hi = expbf16_internal::EXP_HI[x_hi + 12];
float exp_mid = expbf16_internal::EXP_MID[x_mid];
diff --git a/libc/src/__support/math/expf16_utils.h b/libc/src/__support/math/expf16_utils.h
index de7add4ecf7317..15441b091f31d9 100644
--- a/libc/src/__support/math/expf16_utils.h
+++ b/libc/src/__support/math/expf16_utils.h
@@ -68,7 +68,7 @@ struct ExpRangeReduction {
int x_hi = x_hi_mid >> 3;
int x_mid = x_hi_mid & 0x7;
// lo = x - (hi + mid) = round(x * 2^3) * (-2^(-3)) + x
- float lo = fputil::multiply_add(kf, -0x1.0p-3f, xf);
+ float lo = fputil::sub_multiply(xf, kf, 0x1.0p-3f);
float exp_hi = EXP_HI[x_hi + 18];
float exp_mid = EXP_MID[x_mid];
diff --git a/libc/src/__support/math/expf_double_eval.h b/libc/src/__support/math/expf_double_eval.h
index 2c5c947edbe191..4404cb3d4f507a 100644
--- a/libc/src/__support/math/expf_double_eval.h
+++ b/libc/src/__support/math/expf_double_eval.h
@@ -88,7 +88,7 @@ LIBC_INLINE float expf(float x) {
// x_hi = (hi + mid) * 2^7 = round(x * 2^7).
float kf = fputil::nearest_integer(x * 0x1.0p7f);
// Subtract (hi + mid) from x to get lo.
- double xd = static_cast<double>(fputil::multiply_add(kf, -0x1.0p-7f, x));
+ double xd = static_cast<double>(fputil::sub_multiply(x, kf, 0x1.0p-7f));
int x_hi = static_cast<int>(kf);
x_hi += 104 << 7;
// hi = x_hi >> 7
diff --git a/libc/src/__support/math/expf_float_eval.h b/libc/src/__support/math/expf_float_eval.h
index 02cd0444958790..dceba0929e43b1 100644
--- a/libc/src/__support/math/expf_float_eval.h
+++ b/libc/src/__support/math/expf_float_eval.h
@@ -82,7 +82,7 @@ LIBC_INLINE float expf(float x) {
float kf = fputil::nearest_integer(x * LOG2_E.hi);
int k = static_cast<int>(kf);
- float u_hi = fputil::multiply_add(x, LOG2_E.hi, -kf);
+ float u_hi = fputil::multiply_sub(x, LOG2_E.hi, kf);
float u = fputil::multiply_add(x, LOG2_E.lo, u_hi);
#else // !LIBC_TARGET_CPU_HAS_FMA_FLOAT
// Cody-Waite reduction for non-FMA targets:
@@ -98,8 +98,8 @@ LIBC_INLINE float expf(float x) {
float kf = fputil::nearest_integer(x * LOG2_E);
int k = static_cast<int>(kf);
- float v_hi = fputil::multiply_add(-kf, LOG_2_HI, x);
- float v = fputil::multiply_add(-kf, LOG_2_LO, v_hi);
+ float v_hi = fputil::sub_multiply(x, kf, LOG_2_HI);
+ float v = fputil::sub_multiply(v_hi, kf, LOG_2_LO);
// Convert reduced argument to base-2: u = v * log2(e)
float u = v * LOG2_E;
@@ -161,7 +161,7 @@ LIBC_INLINE float expf(float x) {
float kf = fputil::nearest_integer(x_safe * LOG2_E);
int k = static_cast<int>(kf);
- float u_hi = fputil::multiply_add(-kf, LOG_2_HI, x_safe);
+ float u_hi = fputil::sub_multiply(x_safe, kf, LOG_2_HI);
fputil::FloatFloat u_lo = fputil::quick_mult(-kf, LOG_2_LO);
fputil::FloatFloat u = fputil::exact_add(u_hi, u_lo.hi);
diff --git a/libc/src/__support/math/expm1f.h b/libc/src/__support/math/expm1f.h
index 8505d7e72740b5..ec58e74de9bb4b 100644
--- a/libc/src/__support/math/expm1f.h
+++ b/libc/src/__support/math/expm1f.h
@@ -159,7 +159,7 @@ LIBC_INLINE float expm1f(float x) {
float kf = fputil::nearest_integer(x * 0x1.0p7f);
int x_hi = static_cast<int>(kf);
// Subtract (hi + mid) from x to get lo.
- double xd = static_cast<double>(fputil::multiply_add(kf, -0x1.0p-7f, x));
+ double xd = static_cast<double>(fputil::sub_multiply(x, kf, 0x1.0p-7f));
x_hi += 104 << 7;
// hi = x_hi >> 7
double exp_hi = EXP_M1[x_hi >> 7];
@@ -174,7 +174,7 @@ LIBC_INLINE float expm1f(float x) {
double exp_lo =
fputil::polyeval(xd, 0x1.0p0, 0x1.ffffffffff777p-1, 0x1.000000000071cp-1,
0x1.555566668e5e7p-3, 0x1.55555555ef243p-5);
- return static_cast<float>(fputil::multiply_add(exp_hi_mid, exp_lo, -1.0));
+ return static_cast<float>(fputil::multiply_sub(exp_hi_mid, exp_lo, 1.0));
}
} // namespace math
diff --git a/libc/src/__support/math/expm1f16.h b/libc/src/__support/math/expm1f16.h
index ccc73c88b7f9e8..3c3d49ee990fd5 100644
--- a/libc/src/__support/math/expm1f16.h
+++ b/libc/src/__support/math/expm1f16.h
@@ -132,7 +132,7 @@ LIBC_INLINE constexpr float16 expm1f16(float16 x) {
// exp(x) = exp(hi + mid) * exp(lo)
auto [exp_hi_mid, exp_lo] = exp_range_reduction(x);
// expm1(x) = exp(hi + mid) * exp(lo) - 1
- return fputil::cast<float16>(fputil::multiply_add(exp_hi_mid, exp_lo, -1.0f));
+ return fputil::cast<float16>(fputil::multiply_sub(exp_hi_mid, exp_lo, 1.0f));
}
} // namespace math
diff --git a/libc/src/__support/math/expxf16_utils.h b/libc/src/__support/math/expxf16_utils.h
index 33f66a93df6317..b9852db6b097c3 100644
--- a/libc/src/__support/math/expxf16_utils.h
+++ b/libc/src/__support/math/expxf16_utils.h
@@ -50,7 +50,7 @@ LIBC_INLINE ExpRangeReduction exp2_range_reduction(float16 x) {
unsigned x_hi = static_cast<unsigned>(x_hi_mid) >> 3;
unsigned x_mid = static_cast<unsigned>(x_hi_mid) & 0x7;
// lo = x - (hi + mid) = round(x * 2^3) * (-2^(-3)) + x
- float lo = fputil::multiply_add(kf, -0x1.0p-3f, xf);
+ float lo = fputil::sub_multiply(xf, kf, 0x1.0p-3f);
uint32_t exp2_hi_mid_bits =
EXP2_MID_BITS[x_mid] +
@@ -154,7 +154,7 @@ LIBC_INLINE constexpr float16 eval_sinh_or_cosh(float16 x) {
float exp2_hi_mid_diff = exp2_hi_mid_p - exp2_hi_mid_m;
// lo = x - (hi + mid) = round(x * log2(e) * 2^5) * log(2) * (-2^(-5)) + x
- float lo = fputil::multiply_add(kf, LOGF_2 * -0x1.0p-5f, xf);
+ float lo = fputil::sub_multiply(xf, kf, LOGF_2 * 0x1.0p-5f);
float lo_sq = lo * lo;
// Degree-3 minimax polynomial generated by Sollya with the following
diff --git a/libc/src/__support/math/hypot.h b/libc/src/__support/math/hypot.h
index 56516de0c80e89..1417420f9c257f 100644
--- a/libc/src/__support/math/hypot.h
+++ b/libc/src/__support/math/hypot.h
@@ -274,7 +274,7 @@ LIBC_INLINE double hypot(double x, double y) {
// Ziv's rounding test.
double upper = r_hi + fputil::multiply_add(r_hi, ERR, r_lo);
- double lower = r_hi + fputil::multiply_add(r_hi, -ERR, r_lo);
+ double lower = r_hi + fputil::sub_multiply(r_lo, r_hi, ERR);
if (LIBC_LIKELY(upper == lower)) {
return upper * scale_back;
diff --git a/libc/src/__support/math/hypotf.h b/libc/src/__support/math/hypotf.h
index 5d877db190acef..2a9d0012b35939 100644
--- a/libc/src/__support/math/hypotf.h
+++ b/libc/src/__support/math/hypotf.h
@@ -87,7 +87,7 @@ LIBC_INLINE float hypotf(float x, float y) {
double a = y_a_smaller ? x_sq : y_sq;
double b = y_a_smaller ? y_sq : x_sq;
- double e = b - fputil::multiply_add(r_d, r_d, -a);
+ double e = b - fputil::multiply_sub(r_d, r_d, a);
if (e == 0.0)
return r_f;
@@ -98,7 +98,7 @@ LIBC_INLINE float hypotf(float x, float y) {
// We use FastTwoSum algorithm to compute those errors and then combine.
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
double sum_sq_lo = (b - (sum_sq - a));
- double err = sum_sq_lo - fputil::multiply_add(r, r, -sum_sq);
+ double err = sum_sq_lo - fputil::multiply_sub(r, r, sum_sq);
#else
fputil::DoubleDouble r_sq = fputil::exact_mult(r, r);
double sum_sq_lo = (b - (sum_sq - a));
diff --git a/libc/src/__support/math/hypotf16.h b/libc/src/__support/math/hypotf16.h
index f93500977f0e7f..6acc7b3cc2d3fd 100644
--- a/libc/src/__support/math/hypotf16.h
+++ b/libc/src/__support/math/hypotf16.h
@@ -80,7 +80,7 @@ LIBC_INLINE float16 hypotf16(float16 x, float16 y) {
// Perform rounding correction.
float sum_sq_lo = fputil::multiply_add(bf, bf, a_sq - sum_sq);
- float err = sum_sq_lo - fputil::multiply_add(r_d, r_d, -sum_sq);
+ float err = sum_sq_lo - fputil::multiply_sub(r_d, r_d, sum_sq);
if (err > 0) {
r_u |= 1;
diff --git a/libc/src/__support/math/lgammaf.h b/libc/src/__support/math/lgammaf.h
index 9d772916704205..486db3f699f5f5 100644
--- a/libc/src/__support/math/lgammaf.h
+++ b/libc/src/__support/math/lgammaf.h
@@ -126,7 +126,7 @@ LIBC_INLINE float lgammaf(float x) {
constexpr double EULER_GAMMA = 0x1.2788cfc6fb619p-1;
double sign_corr = xbits.is_neg() ? EULER_GAMMA : -EULER_GAMMA;
return fputil::cast<float>(
- fputil::multiply_add(sign_corr, abs_xd, -math::log(abs_xd)));
+ fputil::multiply_sub(sign_corr, abs_xd, math::log(abs_xd)));
}
if (x_abs < 0x3f290000u) {
@@ -173,7 +173,7 @@ LIBC_INLINE float lgammaf(float x) {
double s1625 = fputil::multiply_add(d8, p2425, r1623);
double poly_g = fputil::multiply_add(d16, s1625, s015);
// lgamma(x) = x*g(x) - log(-x) with single rounding via FMA.
- lgamma_val = fputil::multiply_add(xd, poly_g, -math::log(abs_xd));
+ lgamma_val = fputil::multiply_sub(xd, poly_g, math::log(abs_xd));
} else {
// x = 0x1.f8a754p-9f
if (LIBC_UNLIKELY(xbits.uintval() == 0x3b7c53aau))
@@ -209,7 +209,7 @@ LIBC_INLINE float lgammaf(float x) {
double s015 = fputil::multiply_add(d8, r815, r07);
double poly_h = fputil::multiply_add(d16, p1617, s015);
// poly_val - log(abs_xd) with single rounding via FMA.
- lgamma_val = fputil::multiply_add(abs_xd, poly_h, -math::log(abs_xd));
+ lgamma_val = fputil::multiply_sub(abs_xd, poly_h, math::log(abs_xd));
}
} else if (x_abs < 0x3f800000u) {
if (xbits.is_neg()) {
@@ -218,7 +218,7 @@ LIBC_INLINE float lgammaf(float x) {
// exact (Sterbenz); x(x+1) = (t-1)(t-2) is the M2 prefactor.
double w = xd * (xd + 1.0);
double poly = lgamma_m2_poly(xd + 0.5);
- lgamma_val = fputil::multiply_add(w, poly, -lg_ln(-w));
+ lgamma_val = fputil::multiply_sub(w, poly, lg_ln(-w));
} else {
// M1: t in [0.66015625, 1.0). lgamma(t) = (t-1) * P_M1(d), d = t - MID.
// Degree-14 monomial fit of lgamma(t)/(t-1), max error 2^-51.3.
@@ -254,7 +254,7 @@ LIBC_INLINE float lgammaf(float x) {
// are exact (Sterbenz); (x+1)(x+2) = (t-1)(t-2) is the M2 prefactor.
double u = (xd + 1.0) * (xd + 2.0);
double poly = lgamma_m2_poly(xd + 1.5);
- lgamma_val = fputil::multiply_add(u, poly, -lg_ln(xd * u));
+ lgamma_val = fputil::multiply_sub(u, poly, lg_ln(xd * u));
} else {
// M2: t in [1.0, 2.0). lgamma(t) = (t-1)*(t-2) * P_M2(t - 1.5).
double d = abs_xd - 0x1.8p+0;
@@ -376,7 +376,7 @@ LIBC_INLINE float lgammaf(float x) {
double pb = (xd + 2.0) * (xd + 3.0);
double prod = (pa * pb) * (xd + 4.0);
double aprod = prod < 0.0 ? -prod : prod;
- lgamma_val = fputil::multiply_add(xd + 3.0, poly, -lg_ln(aprod));
+ lgamma_val = fputil::multiply_sub(xd + 3.0, poly, lg_ln(aprod));
}
}
} else {
diff --git a/libc/src/__support/math/log.h b/libc/src/__support/math/log.h
index 0ccf347400217b..085bdd9560c7b0 100644
--- a/libc/src/__support/math/log.h
+++ b/libc/src/__support/math/log.h
@@ -809,7 +809,7 @@ LIBC_INLINE LIBC_CONSTEXPR double log(double x) {
// Perform exact range reduction
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
- u = fputil::multiply_add(r, m, -1.0); // exact
+ u = fputil::multiply_sub(r, m, 1.0); // exact
#else
uint64_t c_m = x_m & 0x3FFF'E000'0000'0000ULL;
double c = FPBits_t(c_m).get_val();
diff --git a/libc/src/__support/math/log10.h b/libc/src/__support/math/log10.h
index db6d345f98b71f..e79fe4cb91364a 100644
--- a/libc/src/__support/math/log10.h
+++ b/libc/src/__support/math/log10.h
@@ -810,7 +810,7 @@ LIBC_INLINE double log10(double x) {
// Perform exact range reduction
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
- double u = fputil::multiply_add(r, m, -1.0); // exact
+ double u = fputil::multiply_sub(r, m, 1.0); // exact
#else
uint64_t c_m = x_m & 0x3FFF'E000'0000'0000ULL;
double c = FPBits_t(c_m).get_val();
diff --git a/libc/src/__support/math/log10f16.h b/libc/src/__support/math/log10f16.h
index 424cf5f39a0872..1ba920f307f971 100644
--- a/libc/src/__support/math/log10f16.h
+++ b/libc/src/__support/math/log10f16.h
@@ -163,7 +163,7 @@ LIBC_INLINE float16 log10f16(float16 x) {
x_bits.set_biased_exponent(FPBits::EXP_BIAS);
float mant_f = x_bits.get_val();
// v = 1.mant * 1/f - 1 = d/f
- float v = fputil::multiply_add(mant_f, ONE_OVER_F_F[f], -1.0f);
+ float v = fputil::multiply_sub(mant_f, ONE_OVER_F_F[f], 1.0f);
// Degree-3 minimax polynomial generated by Sollya with the following
// commands:
diff --git a/libc/src/__support/math/log10f_double_eval.h b/libc/src/__support/math/log10f_double_eval.h
index b2695f33dca7eb..278ac1d6ddb61c 100644
--- a/libc/src/__support/math/log10f_double_eval.h
+++ b/libc/src/__support/math/log10f_double_eval.h
@@ -219,10 +219,10 @@ LIBC_INLINE float log10f(float x) {
float u = xbits.get_val();
#ifdef LIBC_TARGET_CPU_HAS_FMA_FLOAT
double v =
- static_cast<double>(fputil::multiply_add(u, R[index], -1.0f)); // Exact.
+ static_cast<double>(fputil::multiply_sub(u, R[index], 1.0f)); // Exact.
#else
- double v = fputil::multiply_add(static_cast<double>(u),
- static_cast<double>(R[index]), -1.0); // Exact
+ double v = fputil::multiply_sub(static_cast<double>(u),
+ static_cast<double>(R[index]), 1.0); // Exact
#endif // LIBC_TARGET_CPU_HAS_FMA_FLOAT
double v2 = v * v; // Exact
diff --git a/libc/src/__support/math/log10p1bf16.h b/libc/src/__support/math/log10p1bf16.h
index cf8aef01a28fd2..0af29eb2f15204 100644
--- a/libc/src/__support/math/log10p1bf16.h
+++ b/libc/src/__support/math/log10p1bf16.h
@@ -155,7 +155,7 @@ LIBC_INLINE bfloat16 log10p1bf16(bfloat16 x) {
float mant_f = y_bits.get_val();
int f = y_bits.get_mantissa() >> (FPBitsFloat::FRACTION_LEN - 5);
- float v = fputil::multiply_add(mant_f, ONE_OVER_F_F[f], -1.0f);
+ float v = fputil::multiply_sub(mant_f, ONE_OVER_F_F[f], 1.0f);
// Degree-3 minimax polynomial generated by Sollya with the following
// commands:
// > display = hexadecimal;
diff --git a/libc/src/__support/math/log10p1f16.h b/libc/src/__support/math/log10p1f16.h
index 6184a5795741b1..8add8fb1e6a69e 100644
--- a/libc/src/__support/math/log10p1f16.h
+++ b/libc/src/__support/math/log10p1f16.h
@@ -185,7 +185,7 @@ LIBC_INLINE float16 log10p1f16(float16 x) {
int f = y_bits.get_mantissa() >> (FPBitsFloat::FRACTION_LEN - 5);
// v = 1.mant * 1/f - 1 = d/f
- float v = fputil::multiply_add(mant_f, ONE_OVER_F_F[f], -1.0f);
+ float v = fputil::multiply_sub(mant_f, ONE_OVER_F_F[f], 1.0f);
// Degree-3 minimax polynomial generated by Sollya with the following
// commands:
diff --git a/libc/src/__support/math/log1p.h b/libc/src/__support/math/log1p.h
index 7dd7c66802c837..f1d3aeff8f3556 100644
--- a/libc/src/__support/math/log1p.h
+++ b/libc/src/__support/math/log1p.h
@@ -1024,7 +1024,7 @@ LIBC_INLINE double log1p(double x) {
// Perform exact range reduction
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
- double v_hi = fputil::multiply_add(r, m_dd.hi, -1.0); // Exact.
+ double v_hi = fputil::multiply_sub(r, m_dd.hi, 1.0); // Exact.
#else
// c = 1 + idx * 2^-7.
double c = FPBits_t((static_cast<uint64_t>(idx) << (FRACTION_LEN - 7)) +
diff --git a/libc/src/__support/math/log2.h b/libc/src/__support/math/log2.h
index 36e82e8f1e8dc3..5ceff7f88b846c 100644
--- a/libc/src/__support/math/log2.h
+++ b/libc/src/__support/math/log2.h
@@ -921,7 +921,7 @@ LIBC_INLINE double log2(double x) {
// Perform exact range reduction
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
- double u = fputil::multiply_add(r, m, -1.0); // exact
+ double u = fputil::multiply_sub(r, m, 1.0); // exact
#else
uint64_t c_m = x_m & 0x3FFF'E000'0000'0000ULL;
double c = FPBits_t(c_m).get_val();
diff --git a/libc/src/__support/math/log2f16.h b/libc/src/__support/math/log2f16.h
index 4a738b5d5b8d45..18e78082fb5648 100644
--- a/libc/src/__support/math/log2f16.h
+++ b/libc/src/__support/math/log2f16.h
@@ -151,7 +151,7 @@ LIBC_INLINE float16 log2f16(float16 x) {
x_bits.set_biased_exponent(FPBits::EXP_BIAS);
float mant_f = x_bits.get_val();
// v = 1.mant * 1/f - 1 = d/f
- float v = fputil::multiply_add(mant_f, ONE_OVER_F_F[f], -1.0f);
+ float v = fputil::multiply_sub(mant_f, ONE_OVER_F_F[f], 1.0f);
// Degree-3 minimax polynomial generated by Sollya with the following
// commands:
diff --git a/libc/src/__support/math/log2f_double_eval.h b/libc/src/__support/math/log2f_double_eval.h
index 91c6bd6ae87edb..0d7d8b6712d391 100644
--- a/libc/src/__support/math/log2f_double_eval.h
+++ b/libc/src/__support/math/log2f_double_eval.h
@@ -105,10 +105,10 @@ LIBC_INLINE float log2f(float x) {
float u = xbits.get_val();
#ifdef LIBC_TARGET_CPU_HAS_FMA_FLOAT
double v =
- static_cast<double>(fputil::multiply_add(u, R[index], -1.0f)); // Exact.
+ static_cast<double>(fputil::multiply_sub(u, R[index], 1.0f)); // Exact.
#else
double v =
- fputil::multiply_add(static_cast<double>(u), RD[index], -1.0); // Exact
+ fputil::multiply_sub(static_cast<double>(u), RD[index], 1.0); // Exact
#endif // LIBC_TARGET_CPU_HAS_FMA_FLOAT
double extra_factor = static_cast<double>(m) + LOG2_R[index];
diff --git a/libc/src/__support/math/log2p1f16.h b/libc/src/__support/math/log2p1f16.h
index c60c94dee24741..7e884aa0be89ff 100644
--- a/libc/src/__support/math/log2p1f16.h
+++ b/libc/src/__support/math/log2p1f16.h
@@ -185,7 +185,7 @@ LIBC_INLINE float16 log2p1f16(float16 x) {
int f = y_bits.get_mantissa() >> (FPBitsFloat::FRACTION_LEN - 5);
// v = 1.mant * 1/f - 1 = d/f
- float v = fputil::multiply_add(mant_f, ONE_OVER_F_F[f], -1.0f);
+ float v = fputil::multiply_sub(mant_f, ONE_OVER_F_F[f], 1.0f);
// Degree-3 minimax polynomial generated by Sollya with the following
// commands:
diff --git a/libc/src/__support/math/logf16.h b/libc/src/__support/math/logf16.h
index 8915f60e972a85..6a4218d5c68961 100644
--- a/libc/src/__support/math/logf16.h
+++ b/libc/src/__support/math/logf16.h
@@ -157,7 +157,7 @@ LIBC_INLINE float16 logf16(float16 x) {
x_bits.set_biased_exponent(FPBits::EXP_BIAS);
float mant_f = x_bits.get_val();
// v = 1.mant * 1/f - 1 = d/f
- float v = fputil::multiply_add(mant_f, ONE_OVER_F_F[f], -1.0f);
+ float v = fputil::multiply_sub(mant_f, ONE_OVER_F_F[f], 1.0f);
// Degree-3 minimax polynomial generated by Sollya with the following
// commands:
diff --git a/libc/src/__support/math/logf_double_eval.h b/libc/src/__support/math/logf_double_eval.h
index 1116aecab747d3..e1cf04bcc32446 100644
--- a/libc/src/__support/math/logf_double_eval.h
+++ b/libc/src/__support/math/logf_double_eval.h
@@ -172,9 +172,9 @@ LIBC_INLINE float logf(float x) {
float u = xbits.get_val();
double v = 0.0;
#ifdef LIBC_TARGET_CPU_HAS_FMA_FLOAT
- v = static_cast<double>(fputil::multiply_add(u, R[index], -1.0f)); // Exact.
+ v = static_cast<double>(fputil::multiply_sub(u, R[index], 1.0f)); // Exact.
#else
- v = fputil::multiply_add(static_cast<double>(u), RD[index], -1.0); // Exact
+ v = fputil::multiply_sub(static_cast<double>(u), RD[index], 1.0); // Exact
#endif // LIBC_TARGET_CPU_HAS_FMA_FLOAT
// Degree-5 polynomial approximation of log generated by Sollya with:
diff --git a/libc/src/__support/math/pow.h b/libc/src/__support/math/pow.h
index 143d58d2bb551a..2ecd9e02091baf 100644
--- a/libc/src/__support/math/pow.h
+++ b/libc/src/__support/math/pow.h
@@ -218,7 +218,7 @@ LIBC_INLINE double pow(double x, double y) {
// Perform exact range reduction.
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
- double dx = fputil::multiply_add(RD[idx_x], m_x.get_val(), -1.0); // Exact
+ double dx = fputil::multiply_sub(RD[idx_x], m_x.get_val(), 1.0); // Exact
#else
double c = FPBits(m_x.uintval() & 0x3fff'e000'0000'0000).get_val();
double dx =
diff --git a/libc/src/__support/math/pow_fast.h b/libc/src/__support/math/pow_fast.h
index 2ef040c76f75b1..f4a121487f876f 100644
--- a/libc/src/__support/math/pow_fast.h
+++ b/libc/src/__support/math/pow_fast.h
@@ -111,7 +111,7 @@ LIBC_INLINE double pow(double x, double y) {
// Perform exact range reduction.
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
- double dx = fputil::multiply_add(RD[idx_x], m_x.get_val(), -1.0); // Exact
+ double dx = fputil::multiply_sub(RD[idx_x], m_x.get_val(), 1.0); // Exact
#else
double c = FPBits(m_x.uintval() & 0x3fff'e000'0000'0000).get_val();
double dx =
diff --git a/libc/src/__support/math/powf_double_eval.h b/libc/src/__support/math/powf_double_eval.h
index 7c8b8c57cb965d..cd4951e2fbf75a 100644
--- a/libc/src/__support/math/powf_double_eval.h
+++ b/libc/src/__support/math/powf_double_eval.h
@@ -48,9 +48,9 @@ LIBC_INLINE double powf_accurate(float x, float y, double e_x, uint64_t sign) {
int idx_x = static_cast<int>(x_mant >> (FloatBits::FRACTION_LEN - 7));
// Append hidden bit.
float m_x = cpp::bit_cast<float>(x_mant | 0x3f80'0000);
- double dx = fputil::multiply_add(
+ double dx = fputil::multiply_sub(
static_cast<double>(m_x),
- static_cast<double>(common_constants_internal::R[idx_x]), -1.0);
+ static_cast<double>(common_constants_internal::R[idx_x]), 1.0);
// Degree-11 minimax polynomial for log2(1 + x)/x on [-2^-8, 2^-7] in
// double-double, generated by Sollya with:
@@ -339,7 +339,7 @@ LIBC_INLINE float powf(float x, float y) {
// Since r has <= 7 bits and m_x has 24 bits, m_x * r - 1.0 is exact in
// double.
double dx =
- fputil::multiply_add(m_x, powf_internal::R_32_D[idx_x], -1.0); // Exact
+ fputil::multiply_sub(m_x, powf_internal::R_32_D[idx_x], 1.0); // Exact
// Degree-6 minimax polynomial approximation for log2(1 + dx)/dx on
// [-2^-6, 2^-5]. Generated by Sollya with:
@@ -378,7 +378,7 @@ LIBC_INLINE float powf(float x, float y) {
// y * log2(x) = k + f, with k integer and |f| <= 0.5.
double kd = fputil::nearest_integer(z);
int64_t k = static_cast<int64_t>(kd);
- double f = fputil::multiply_add(y_d, s, -kd);
+ double f = fputil::multiply_sub(y_d, s, kd);
// Degree-7 polynomial approximation P(f) ~ (2^f - 1)/f on [-0.5, 0.5]
// Generated by Sollya with:
diff --git a/libc/src/__support/math/powf_float_eval.h b/libc/src/__support/math/powf_float_eval.h
index d19effd17d0f63..49d15201f14a6c 100644
--- a/libc/src/__support/math/powf_float_eval.h
+++ b/libc/src/__support/math/powf_float_eval.h
@@ -147,13 +147,13 @@ LIBC_INLINE float powf(float x, float y) {
// The computation is exact: m_x = (1 + dx) / r.
float r = powf_internal::R_16[idx_x];
#if defined(LIBC_TARGET_CPU_HAS_FMA_FLOAT)
- float dx = fputil::multiply_add(r, m_x.get_val(), -1.0f); // Exact
+ float dx = fputil::multiply_sub(r, m_x.get_val(), 1.0f); // Exact
#else // !LIBC_TARGET_CPU_HAS_FMA_FLOAT
// Without FMA, split m_x = c + cd where c retains the high 14 bits
// to perform exact multiplication: r * c - 1.0f and r * cd.
float c = FPBits(m_x.uintval() & 0x3fff'e000U).get_val();
float cd = m_x.get_val() - c;
- float dx_hi = fputil::multiply_add(r, c, -1.0f);
+ float dx_hi = fputil::multiply_sub(r, c, 1.0f);
float dx = fputil::multiply_add(r, cd, dx_hi); // Exact
#endif // LIBC_TARGET_CPU_HAS_FMA_FLOAT
diff --git a/libc/src/__support/math/rsqrtf16.h b/libc/src/__support/math/rsqrtf16.h
index 9722975aa15bd6..69dedb49f8f279 100644
--- a/libc/src/__support/math/rsqrtf16.h
+++ b/libc/src/__support/math/rsqrtf16.h
@@ -304,7 +304,7 @@ LIBC_INLINE constexpr float16 rsqrtf16(float16 x) {
if (LIBC_UNLIKELY(half_mantissa == 0x011F)) {
result = fputil::multiply_add(result, 0x1.0p-21f, result);
} else if (LIBC_UNLIKELY(half_mantissa == 0x0313)) {
- result = fputil::multiply_add(result, -0x1.0p-21f, result);
+ result = fputil::sub_multiply(result, result, 0x1.0p-21f);
}
return fputil::cast<float16>(result);
diff --git a/libc/src/__support/math/sin.h b/libc/src/__support/math/sin.h
index 4ae04ae64a423d..cd3111ad21ad9c 100644
--- a/libc/src/__support/math/sin.h
+++ b/libc/src/__support/math/sin.h
@@ -97,7 +97,7 @@ LIBC_INLINE double sin(double x) {
return x + x; // Make sure it works with FTZ/DAZ.
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
- return fputil::multiply_add(x, -0x1.0p-54, x);
+ return fputil::sub_multiply(x, x, 0x1.0p-54);
#else
if (LIBC_UNLIKELY(x_e < 4)) {
#ifndef LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY
@@ -108,7 +108,7 @@ LIBC_INLINE double sin(double x) {
return FPBits(xbits.uintval() - 1).get_val();
#endif // !LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY
}
- return fputil::multiply_add(x, -0x1.0p-54, x);
+ return fputil::sub_multiply(x, x, 0x1.0p-54);
#endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE
}
// No range reduction needed.
@@ -136,7 +136,7 @@ LIBC_INLINE double sin(double x) {
// Overall errors <= ulp(x^3) + |x| * 2^-68.
double err = fputil::multiply_add(x_sq, 0x1.0p-52, 0x1.0p-68);
double r_lo_u = fputil::multiply_add(x, err, r_lo);
- double r_lo_l = fputil::multiply_add(-x, err, r_lo);
+ double r_lo_l = fputil::sub_multiply(r_lo, x, err);
double r_upper = x + r_lo_u;
double r_lower = x + r_lo_l;
diff --git a/libc/src/__support/math/sincos.h b/libc/src/__support/math/sincos.h
index 4d8cf96dc45115..99be10fc156962 100644
--- a/libc/src/__support/math/sincos.h
+++ b/libc/src/__support/math/sincos.h
@@ -61,8 +61,8 @@ LIBC_INLINE void sincos(double x, double *sin_x, double *cos_x) {
// For |x| < 2^-27, max(|sin(x) - x|, |cos(x) - 1|) < ulp(x)/2.
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
- *sin_x = fputil::multiply_add(x, -0x1.0p-54, x);
- *cos_x = fputil::multiply_add(x, -x, 1.0);
+ *sin_x = fputil::sub_multiply(x, x, 0x1.0p-54);
+ *cos_x = fputil::sub_multiply(1.0, x, x);
#else
*cos_x = fputil::round_result_slightly_down(1.0);
@@ -75,7 +75,7 @@ LIBC_INLINE void sincos(double x, double *sin_x, double *cos_x) {
*sin_x = FPBits(xbits.uintval() - 1).get_val();
#endif // !LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY
}
- *sin_x = fputil::multiply_add(x, -0x1.0p-54, x);
+ *sin_x = fputil::sub_multiply(x, x, 0x1.0p-54);
#endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE
return;
}
diff --git a/libc/src/__support/math/sincos_eval.h b/libc/src/__support/math/sincos_eval.h
index 4a3dd972861cfb..e33684db08e487 100644
--- a/libc/src/__support/math/sincos_eval.h
+++ b/libc/src/__support/math/sincos_eval.h
@@ -39,13 +39,13 @@ LIBC_INLINE double sincos_eval(const DoubleDouble &u, DoubleDouble &sin_u,
// + u_lo (1 + u_hi^2 * (-1/2 + u_hi^2 / 24))
double u_hi_sq = u.hi * u.hi; // Error < ulp(u_hi^2) < 2^(-6 - 52) = 2^-58.
// p1 ~ 1/120 + u_hi^2 / 5040.
- double p1 = fputil::multiply_add(u_hi_sq, -0x1.a01a01a01a01ap-13,
- 0x1.1111111111111p-7);
+ double p1 = fputil::sub_multiply(0x1.1111111111111p-7, u_hi_sq,
+ 0x1.a01a01a01a01ap-13);
// q1 ~ -1/2 + u_hi^2 / 24.
- double q1 = fputil::multiply_add(u_hi_sq, 0x1.5555555555555p-5, -0x1.0p-1);
+ double q1 = fputil::multiply_sub(u_hi_sq, 0x1.5555555555555p-5, 0x1.0p-1);
double u_hi_3 = u_hi_sq * u.hi;
// p2 ~ -1/6 + u_hi^2 (1/120 - u_hi^2 * 1/5040)
- double p2 = fputil::multiply_add(u_hi_sq, p1, -0x1.5555555555555p-3);
+ double p2 = fputil::multiply_sub(u_hi_sq, p1, 0x1.5555555555555p-3);
// q2 ~ 1 + u_hi^2 (-1/2 + u_hi^2 / 24)
double q2 = fputil::multiply_add(u_hi_sq, q1, 1.0);
double sin_lo = fputil::multiply_add(u_hi_3, p2, u.lo * q2);
@@ -78,10 +78,10 @@ LIBC_INLINE double sincos_eval(const DoubleDouble &u, DoubleDouble &sin_u,
#endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE
// r1 ~ -1/720 + u_hi^2 / 40320
- double r1 = fputil::multiply_add(u_hi_sq, 0x1.a01a01a01a01ap-16,
- -0x1.6c16c16c16c17p-10);
+ double r1 = fputil::multiply_sub(u_hi_sq, 0x1.a01a01a01a01ap-16,
+ 0x1.6c16c16c16c17p-10);
// s1 ~ -1 + u_hi^2 / 6
- double s1 = fputil::multiply_add(u_hi_sq, 0x1.5555555555555p-3, -1.0);
+ double s1 = fputil::multiply_sub(u_hi_sq, 0x1.5555555555555p-3, 1.0);
double u_hi_4 = u_hi_sq * u_hi_sq;
double u_hi_u_lo = u.hi * u.lo;
// r2 ~ 1/24 + u_hi^2 (-1/720 + u_hi^2 / 40320)
diff --git a/libc/src/__support/math/sincosf.h b/libc/src/__support/math/sincosf.h
index b5f700011bfd4c..b21b6603e94e78 100644
--- a/libc/src/__support/math/sincosf.h
+++ b/libc/src/__support/math/sincosf.h
@@ -142,12 +142,12 @@ LIBC_INLINE void sincosf(float x, float *sinp, float *cosp) {
// double for intermediate results as it is more efficient than using an
// emulated version of FMA.
#if defined(LIBC_TARGET_CPU_HAS_FMA_FLOAT)
- *sinp = fputil::multiply_add(x, -0x1.0p-25f, x);
- *cosp = fputil::multiply_add(FPBits(x_abs).get_val(), -0x1.0p-25f, 1.0f);
+ *sinp = fputil::sub_multiply(x, x, 0x1.0p-25f);
+ *cosp = fputil::sub_multiply(1.0f, FPBits(x_abs).get_val(), 0x1.0p-25f);
#else
- *sinp = static_cast<float>(fputil::multiply_add(xd, -0x1.0p-25, xd));
- *cosp = static_cast<float>(fputil::multiply_add(
- static_cast<double>(FPBits(x_abs).get_val()), -0x1.0p-25, 1.0));
+ *sinp = static_cast<float>(fputil::sub_multiply(xd, xd, 0x1.0p-25));
+ *cosp = static_cast<float>(fputil::sub_multiply(
+ 1.0, static_cast<double>(FPBits(x_abs).get_val()), 0x1.0p-25));
#endif // LIBC_TARGET_CPU_HAS_FMA_FLOAT
return;
}
@@ -221,8 +221,8 @@ LIBC_INLINE void sincosf(float x, float *sinp, float *cosp) {
*sinp = static_cast<float>(fputil::multiply_add(
sin_y, cos_k, fputil::multiply_add(cosm1_y, sin_k, sin_k)));
- *cosp = static_cast<float>(fputil::multiply_add(
- sin_y, -sin_k, fputil::multiply_add(cosm1_y, cos_k, cos_k)));
+ *cosp = static_cast<float>(fputil::sub_multiply(
+ fputil::multiply_add(cosm1_y, cos_k, cos_k), sin_y, sin_k));
}
} // namespace math
diff --git a/libc/src/__support/math/sincosf16_utils.h b/libc/src/__support/math/sincosf16_utils.h
index cd91eaf711df29..46c74a78c18bf8 100644
--- a/libc/src/__support/math/sincosf16_utils.h
+++ b/libc/src/__support/math/sincosf16_utils.h
@@ -44,7 +44,7 @@ LIBC_INLINE_VAR constexpr float SIN_K_PI_OVER_32[64] = {
LIBC_INLINE int32_t range_reduction_sincospif16(float x, float &y) {
float kf = fputil::nearest_integer(x * 32);
- y = fputil::multiply_add(x, 32.0f, -kf);
+ y = fputil::multiply_sub(x, 32.0f, kf);
return static_cast<int32_t>(kf);
}
diff --git a/libc/src/__support/math/sincosf_float_eval.h b/libc/src/__support/math/sincosf_float_eval.h
index d834f3d7cf23f1..4ca48b83f4d32a 100644
--- a/libc/src/__support/math/sincosf_float_eval.h
+++ b/libc/src/__support/math/sincosf_float_eval.h
@@ -197,9 +197,9 @@ template <bool IS_SIN> LIBC_INLINE float sincosf_eval(float x) {
// 0x1.aa8...p-29
// p1 = y^2 * 1/120 - 1/6
- float p1 = fputil::multiply_add(y_sq, 0x1.111112p-7f, -0x1.555556p-3f);
+ float p1 = fputil::multiply_sub(y_sq, 0x1.111112p-7f, 0x1.555556p-3f);
// q1 = y^2 * coeff(Q, 4) + coeff(Q, 2)
- float q1 = fputil::multiply_add(y_sq, 0x1.54b8bep-5f, -0x1.ffffc4p-2f);
+ float q1 = fputil::multiply_sub(y_sq, 0x1.54b8bep-5f, 0x1.ffffc4p-2f);
float y3 = y_sq * y;
// c1 ~ cos(y)
float c1 = fputil::multiply_add(y_sq, q1, 1.0f);
@@ -211,7 +211,7 @@ template <bool IS_SIN> LIBC_INLINE float sincosf_eval(float x) {
return fputil::multiply_add(sin_k, c1, cos_k * s1);
} else {
// cos(x) = cos(k * pi/8) * cos(y) - sin(k * pi/8) * sin(y).
- return fputil::multiply_add(cos_k, c1, -sin_k * s1);
+ return fputil::multiply_sub(cos_k, c1, sin_k * s1);
}
}
diff --git a/libc/src/__support/math/sincosf_utils.h b/libc/src/__support/math/sincosf_utils.h
index 8fbc8384f9e325..88ee707c2d8232 100644
--- a/libc/src/__support/math/sincosf_utils.h
+++ b/libc/src/__support/math/sincosf_utils.h
@@ -119,7 +119,7 @@ LIBC_INLINE void sincosf_eval(double xd, uint32_t x_abs, double &sin_k,
// => pi * x = (k + y) * pi / 32
LIBC_INLINE int64_t range_reduction_sincospi(double x, double &y) {
double kd = fputil::nearest_integer(x * 32);
- y = fputil::multiply_add(x, 32.0, -kd);
+ y = fputil::multiply_sub(x, 32.0, kd);
return static_cast<int64_t>(kd);
}
diff --git a/libc/src/__support/math/sinf_double_eval.h b/libc/src/__support/math/sinf_double_eval.h
index 4d498109221955..0ba1769b3f0248 100644
--- a/libc/src/__support/math/sinf_double_eval.h
+++ b/libc/src/__support/math/sinf_double_eval.h
@@ -113,9 +113,9 @@ LIBC_INLINE float sinf(float x) {
// double for intermediate results as it is more efficient than using an
// emulated version of FMA.
#if defined(LIBC_TARGET_CPU_HAS_FMA_FLOAT)
- return fputil::multiply_add(x, -0x1.0p-25f, x);
+ return fputil::sub_multiply(x, x, 0x1.0p-25f);
#else
- return static_cast<float>(fputil::multiply_add(xd, -0x1.0p-25, xd));
+ return static_cast<float>(fputil::sub_multiply(xd, xd, 0x1.0p-25));
#endif // LIBC_TARGET_CPU_HAS_FMA_FLOAT
}
diff --git a/libc/src/__support/math/sinhcosh_utils.h b/libc/src/__support/math/sinhcosh_utils.h
index 6ca377241d0fb8..33b7258c143fa5 100644
--- a/libc/src/__support/math/sinhcosh_utils.h
+++ b/libc/src/__support/math/sinhcosh_utils.h
@@ -352,7 +352,7 @@ LIBC_MATH_SMALL_TABLES_INLINE double exp_pm_eval(double x, bool is_sinh) {
// even_lo ~ mp_lo + (s.hi * u^2) * P_even(u^2)
// poly_lo ~ even_lo + d.hi * odd_lo
double mp_lo = fputil::multiply_add(m_p.lo, t1, m_p.lo);
- double mm_lo = fputil::multiply_add(-m_m.lo, t1, m_m.lo);
+ double mm_lo = fputil::sub_multiply(m_m.lo, m_m.lo, t1);
double s_u2 = s.hi * u2;
double even_lo = fputil::multiply_add(s_u2, p_even, mp_lo);
double poly_lo = fputil::multiply_add(d.hi, odd_lo, even_lo);
diff --git a/libc/src/__support/math/sinhcoshbf16_utils.h b/libc/src/__support/math/sinhcoshbf16_utils.h
index 8e9a687a0579bb..32a1bcc8e0c1df 100644
--- a/libc/src/__support/math/sinhcoshbf16_utils.h
+++ b/libc/src/__support/math/sinhcoshbf16_utils.h
@@ -114,7 +114,7 @@ template <bool IsSinh> LIBC_INLINE float eval_sinh_or_cosh(float xf) {
float exp2_hi_mid_diff = exp2_hi_mid_p - exp2_hi_mid_m;
// lo = x - (hi + mid) = round(x * log2(e) * 2^5) * log(2) * (-2^(-5)) + x
- float lo = fputil::multiply_add(kf, LOGF_2 * -0x1.0p-5f, xf);
+ float lo = fputil::sub_multiply(xf, kf, LOGF_2 * 0x1.0p-5f);
float lo_sq = lo * lo;
// Degree-3 minimax polynomial generated by Sollya with the following
@@ -156,7 +156,7 @@ LIBC_INLINE float exp_half(float xf) {
float exp2_hi_mid_p = fputil::FPBits<float>(exp2_hi_mid_bits_p).get_val();
// lo = x - (hi + mid) = round(x * log2(e) * 2^5) * log(2) * (-2^(-5)) + x
- float lo = fputil::multiply_add(kf, LOGF_2 * -0x1.0p-5f, xf);
+ float lo = fputil::sub_multiply(xf, kf, LOGF_2 * 0x1.0p-5f);
float lo_sq = lo * lo;
constexpr cpp::array<float, 4> COEFFS = {0x1p+0f, 0x1p+0f, 0x1.0004p-1f,
diff --git a/libc/src/__support/math/tanbf16.h b/libc/src/__support/math/tanbf16.h
index 3dae149a4a643f..ee02b51d268653 100644
--- a/libc/src/__support/math/tanbf16.h
+++ b/libc/src/__support/math/tanbf16.h
@@ -72,8 +72,8 @@ LIBC_INLINE bfloat16 tanbf16(bfloat16 x) {
return fputil::cast<bfloat16>(
fputil::multiply_add(sin_y, cos_k,
fputil::multiply_add(cosm1_y, sin_k, sin_k)) /
- fputil::multiply_add(sin_y, -sin_k,
- fputil::multiply_add(cosm1_y, cos_k, cos_k)));
+ fputil::sub_multiply(fputil::multiply_add(cosm1_y, cos_k, cos_k), sin_y,
+ sin_k));
}
} // namespace math
diff --git a/libc/src/__support/math/tanf16.h b/libc/src/__support/math/tanf16.h
index 761990f73a90b2..6acd09348e0d5d 100644
--- a/libc/src/__support/math/tanf16.h
+++ b/libc/src/__support/math/tanf16.h
@@ -127,7 +127,7 @@ LIBC_INLINE float16 tanf16(float16 x) {
using fputil::multiply_add;
return fputil::cast<float16>(
multiply_add(sin_y, cos_k, multiply_add(cosm1_y, sin_k, sin_k)) /
- multiply_add(sin_y, -sin_k, multiply_add(cosm1_y, cos_k, cos_k)));
+ fputil::sub_multiply(multiply_add(cosm1_y, cos_k, cos_k), sin_y, sin_k));
}
} // namespace math
diff --git a/libc/src/__support/math/tanf_double_eval.h b/libc/src/__support/math/tanf_double_eval.h
index e5ef67ed0d157c..7f7ae0c65802e9 100644
--- a/libc/src/__support/math/tanf_double_eval.h
+++ b/libc/src/__support/math/tanf_double_eval.h
@@ -161,7 +161,7 @@ LIBC_INLINE float tanf(float x) {
using fputil::multiply_add;
return static_cast<float>(
multiply_add(sin_y, cos_k, multiply_add(cosm1_y, sin_k, sin_k)) /
- multiply_add(sin_y, -sin_k, multiply_add(cosm1_y, cos_k, cos_k)));
+ fputil::sub_multiply(multiply_add(cosm1_y, cos_k, cos_k), sin_y, sin_k));
}
} // namespace double_eval
diff --git a/libc/src/__support/math/tanf_float_eval.h b/libc/src/__support/math/tanf_float_eval.h
index 27f49c8e35bb6d..bcb215f636ef5e 100644
--- a/libc/src/__support/math/tanf_float_eval.h
+++ b/libc/src/__support/math/tanf_float_eval.h
@@ -95,7 +95,7 @@ LIBC_INLINE float tanf(float x) {
// exactly 0 respectively, so the remaining term is computed with a small
// relative error.
float num = fputil::multiply_add(cos_k, t, sin_k);
- float den = fputil::multiply_add(-sin_k, t, cos_k);
+ float den = fputil::sub_multiply(cos_k, sin_k, t);
return num / den;
}
diff --git a/libc/src/__support/math/tanh_double_eval.h b/libc/src/__support/math/tanh_double_eval.h
index 3c3b015030e568..7a17048f3d774f 100644
--- a/libc/src/__support/math/tanh_double_eval.h
+++ b/libc/src/__support/math/tanh_double_eval.h
@@ -338,8 +338,8 @@ LIBC_MATH_SMALL_TABLES_INLINE double tanh_eval(double x) {
double y_hi = 1.0 / d.hi;
double q_hi = two_m_hi * y_hi;
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
- double rem_hi = fputil::multiply_add(-q_hi, d.hi, two_m_hi);
- double rem_lo = fputil::multiply_add(-q_hi, d_lo, two_m_lo);
+ double rem_hi = fputil::sub_multiply(two_m_hi, q_hi, d.hi);
+ double rem_lo = fputil::sub_multiply(two_m_lo, q_hi, d_lo);
#else
// Split q_hi into 26 + 27 bits and d.hi into 27 + 26 bits via bit-masks so
// that q_hi_hi * d_hi_hi (53 bits) and q_hi_hi * d_hi_lo (52 bits) are exact
@@ -427,11 +427,11 @@ LIBC_INLINE double tanh_small_eval(double x, uint64_t x_abs) {
if (LIBC_UNLIKELY(x_abs == 0))
return x;
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
- double res = fputil::multiply_add(x, -0x1.0p-55, x);
+ double res = fputil::sub_multiply(x, x, 0x1.0p-55);
#else
// Avoid spurious FE_UNDERFLOW on x * -2^-55 when 2^-1022 < |x| < 2^-967,
// and preserve directed rounding for subnormal |x| <= 2^-1022:
- double x_scaled = fputil::multiply_add(x, 0x1.0p55, -x);
+ double x_scaled = fputil::multiply_sub(x, 0x1.0p55, x);
double res = x_scaled * 0x1.0p-55;
#endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE
if (LIBC_UNLIKELY(x_abs <= 0x0010'0000'0000'0000ULL)) {
@@ -485,7 +485,7 @@ LIBC_INLINE double tanh_small_eval(double x, uint64_t x_abs) {
// The error of y_hi + y_lo is bounded by |y_hi| * 0x1.8p-54, and the rounding
// error of r.lo + y_lo is bounded by |x| * 2^-103.
double x_err = x * 0x1.0p-103;
- double err = fputil::multiply_add(y_hi, -0x1.8p-54, x_err);
+ double err = fputil::sub_multiply(x_err, y_hi, 0x1.8p-54);
// Low tail bounds lo_{u,l} ~ r.lo + (y_lo +/- err):
double y_lo_u = y_lo + err;
diff --git a/libc/src/__support/math/tanhf.h b/libc/src/__support/math/tanhf.h
index 28d1c60a6912d1..9735750b2081b1 100644
--- a/libc/src/__support/math/tanhf.h
+++ b/libc/src/__support/math/tanhf.h
@@ -86,7 +86,7 @@ LIBC_INLINE float tanhf(float x) {
#else
const double half_way = xbits.is_neg() ? 0.5 : -0.5;
- mk = static_cast<int>(fputil::multiply_add(xd, -LOG2_E_EXP2_6, half_way));
+ mk = static_cast<int>(fputil::sub_multiply(half_way, xd, LOG2_E_EXP2_6));
k = static_cast<double>(-mk);
#endif // LIBC_TARGET_CPU_HAS_NEAREST_INT
// -hi = floor(-k * 2^(-MID_BITS))
diff --git a/libc/src/__support/math/tanhf16.h b/libc/src/__support/math/tanhf16.h
index 768ec6845fe946..e483d2a57189e5 100644
--- a/libc/src/__support/math/tanhf16.h
+++ b/libc/src/__support/math/tanhf16.h
@@ -149,7 +149,7 @@ LIBC_INLINE float16 tanhf16(float16 x) {
unsigned x_mid = static_cast<unsigned>(x_hi_mid) & 0x1f;
// lo = x - (hi + mid)
// = round(x * log2(e) * 2 * 2^5) * log(2) * 0.5 * (-2^(-5)) + x
- float lo = fputil::multiply_add(kf, LOGF_2 * 0.5f * -0x1.0p-5f, xf);
+ float lo = fputil::sub_multiply(xf, kf, LOGF_2 * 0.5f * 0x1.0p-5f);
uint32_t exp2_hi_mid_bits =
EXP2_MID_5_BITS[x_mid] +
diff --git a/libc/src/__support/math/tanpif.h b/libc/src/__support/math/tanpif.h
index 722189e7c57c3a..efbe87ad57984c 100644
--- a/libc/src/__support/math/tanpif.h
+++ b/libc/src/__support/math/tanpif.h
@@ -76,8 +76,8 @@ LIBC_INLINE float tanpif(float x) {
// Odd multiples of 1/4 have exact results of +1 or -1.
double xd = static_cast<double>(x);
if (LIBC_UNLIKELY(fputil::is_odd_integer(4.0 * xd)))
- return fputil::is_odd_integer(fputil::multiply_add(2.0, xd, -0.5)) ? -1.0f
- : 1.0f;
+ return fputil::is_odd_integer(fputil::multiply_sub(2.0, xd, 0.5)) ? -1.0f
+ : 1.0f;
#ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
// Handle exceptional values.
@@ -119,7 +119,7 @@ LIBC_INLINE float tanpif(float x) {
using fputil::multiply_add;
return fputil::cast<float>(
multiply_add(sin_y, cos_k, multiply_add(cosm1_y, sin_k, sin_k)) /
- multiply_add(sin_y, -sin_k, multiply_add(cosm1_y, cos_k, cos_k)));
+ fputil::sub_multiply(multiply_add(cosm1_y, cos_k, cos_k), sin_y, sin_k));
}
} // namespace math
diff --git a/libc/src/__support/math/tanpif16.h b/libc/src/__support/math/tanpif16.h
index db317ee9d3fbd3..5fb1e204c4cb47 100644
--- a/libc/src/__support/math/tanpif16.h
+++ b/libc/src/__support/math/tanpif16.h
@@ -68,7 +68,7 @@ LIBC_INLINE float16 tanpif16(float16 x) {
// Odd multiples of 1/4 have exact results of +1 or -1.
float xf = x;
if (LIBC_UNLIKELY(fputil::is_odd_integer(4.0f * xf)))
- return fputil::is_odd_integer(fputil::multiply_add(2.0f, xf, -0.5f))
+ return fputil::is_odd_integer(fputil::multiply_sub(2.0f, xf, 0.5f))
? -1.0f16
: 1.0f16;
@@ -129,7 +129,7 @@ LIBC_INLINE float16 tanpif16(float16 x) {
using fputil::multiply_add;
return fputil::cast<float16>(
multiply_add(sin_y, cos_k, multiply_add(cosm1_y, sin_k, sin_k)) /
- multiply_add(sin_y, -sin_k, multiply_add(cosm1_y, cos_k, cos_k)));
+ fputil::sub_multiply(multiply_add(cosm1_y, cos_k, cos_k), sin_y, sin_k));
}
} // namespace math
More information about the llvm-branch-commits
mailing list