From 347a5b592c91f0230d52a5117fd83e1f75028207 Mon Sep 17 00:00:00 2001 From: Szabolcs Nagy Date: Fri, 31 Dec 2021 09:50:50 +0000 Subject: [PATCH] math: Fix float conversion regressions with gcc-12 [BZ #28713] Converting double precision constants to float is now affected by the runtime dynamic rounding mode instead of being evaluated at compile time with default rounding mode (except static object initializers). This can change the computed result and cause performance regression. The known correctness issues (increased ulp errors) are already fixed, this patch fixes remaining cases of unnecessary runtime conversions. Add float M_* macros to math.h as new GNU extension API. To avoid conversions the new M_* macros are used and instead of casting double literals to float, use float literals (only required if the conversion is inexact). The patch was tested on aarch64 where the following symbols had new spurious conversion instructions that got fixed: __clog10f __gammaf_r_finite@GLIBC_2.17 __j0f_finite@GLIBC_2.17 __j1f_finite@GLIBC_2.17 __jnf_finite@GLIBC_2.17 __kernel_casinhf __lgamma_negf __log1pf __y0f_finite@GLIBC_2.17 __y1f_finite@GLIBC_2.17 cacosf cacoshf casinhf catanf catanhf clogf gammaf_positive Fixes bug 28713. Reviewed-by: Paul Zimmermann --- NEWS | 5 +++++ manual/math.texi | 7 ++++--- math/k_casinh_template.c | 2 +- math/math.h | 17 +++++++++++++++++ math/s_cacos_template.c | 2 +- math/s_catan_template.c | 2 +- math/s_catanh_template.c | 2 +- math/s_clog10_template.c | 8 ++++---- math/s_clog_template.c | 4 ++-- sysdeps/generic/math-type-macros-float.h | 5 ++--- sysdeps/ieee754/flt-32/e_gammaf_r.c | 12 ++++++------ sysdeps/ieee754/flt-32/e_j0f.c | 4 ++-- sysdeps/ieee754/flt-32/e_j1f.c | 4 ++-- sysdeps/ieee754/flt-32/e_jnf.c | 2 +- sysdeps/ieee754/flt-32/lgamma_negf.c | 8 ++++---- sysdeps/ieee754/flt-32/s_log1pf.c | 2 +- 16 files changed, 54 insertions(+), 32 deletions(-) diff --git a/NEWS b/NEWS index eb68bb1..a957b19 100644 --- a/NEWS +++ b/NEWS @@ -43,6 +43,11 @@ Major new features: fminimum_mag, fminimum_mag_num and corresponding functions for float, long double, _FloatN and _FloatNx. +* macros for single-precision float constants are added as a + GNU extension: M_Ef, M_LOG2Ef, M_LOG10Ef, M_LN2f, M_LN10f, M_PIf, + M_PI_2f, M_PI_4f, M_1_PIf, M_2_PIf, M_2_SQRTPIf, M_SQRT2f and + M_SQRT1_2f. + * The __STDC_IEC_60559_BFP__ and __STDC_IEC_60559_COMPLEX__ macros are predefined as specified in TS 18661-1:2014. diff --git a/manual/math.texi b/manual/math.texi index 376306c..e00871c 100644 --- a/manual/math.texi +++ b/manual/math.texi @@ -131,9 +131,10 @@ defined. The default set of features includes these constants. @xref{Feature Test Macros}. All values are of type @code{double}. As an extension, @theglibc{} -also defines these constants with type @code{long double}. The -@code{long double} macros have a lowercase @samp{l} appended to their -names: @code{M_El}, @code{M_PIl}, and so forth. These are only +also defines these constants with type @code{long double} and +@code{float}. The @code{long double} macros have a lowercase @samp{l} +while the @code{float} macros have a lowercase @samp{f} appended to +their names: @code{M_El}, @code{M_PIl}, and so forth. These are only available if @code{_GNU_SOURCE} is defined. Likewise, @theglibc{} also defines these constants with the types diff --git a/math/k_casinh_template.c b/math/k_casinh_template.c index 98980dc..e9fdc97 100644 --- a/math/k_casinh_template.c +++ b/math/k_casinh_template.c @@ -56,7 +56,7 @@ M_DECL_FUNC (__kernel_casinh) (CFLOAT x, int adj) } res = M_SUF (__clog) (y); - __real__ res += (FLOAT) M_MLIT (M_LN2); + __real__ res += M_MLIT (M_LN2); } else if (rx >= M_LIT (0.5) && ix < M_EPSILON / 8) { diff --git a/math/math.h b/math/math.h index 7bd56c9..27963ef 100644 --- a/math/math.h +++ b/math/math.h @@ -1158,6 +1158,23 @@ iszero (__T __val) # define M_SQRT1_2 0.70710678118654752440 /* 1/sqrt(2) */ #endif +/* GNU extension to provide float constants with similar names. */ +#ifdef __USE_GNU +# define M_Ef 2.7182818284590452354f /* e */ +# define M_LOG2Ef 1.4426950408889634074f /* log_2 e */ +# define M_LOG10Ef 0.43429448190325182765f /* log_10 e */ +# define M_LN2f 0.69314718055994530942f /* log_e 2 */ +# define M_LN10f 2.30258509299404568402f /* log_e 10 */ +# define M_PIf 3.14159265358979323846f /* pi */ +# define M_PI_2f 1.57079632679489661923f /* pi/2 */ +# define M_PI_4f 0.78539816339744830962f /* pi/4 */ +# define M_1_PIf 0.31830988618379067154f /* 1/pi */ +# define M_2_PIf 0.63661977236758134308f /* 2/pi */ +# define M_2_SQRTPIf 1.12837916709551257390f /* 2/sqrt(pi) */ +# define M_SQRT2f 1.41421356237309504880f /* sqrt(2) */ +# define M_SQRT1_2f 0.70710678118654752440f /* 1/sqrt(2) */ +#endif + /* The above constants are not adequate for computation using `long double's. Therefore we provide as an extension constants with similar names as a GNU extension. Provide enough digits for the 128-bit IEEE quad. */ diff --git a/math/s_cacos_template.c b/math/s_cacos_template.c index 644330f..eaf7c80 100644 --- a/math/s_cacos_template.c +++ b/math/s_cacos_template.c @@ -32,7 +32,7 @@ M_DECL_FUNC (__cacos) (CFLOAT x) { y = M_SUF (__casin) (x); - __real__ res = (FLOAT) M_MLIT (M_PI_2) - __real__ y; + __real__ res = M_MLIT (M_PI_2) - __real__ y; if (__real__ res == 0) __real__ res = 0; __imag__ res = -__imag__ y; diff --git a/math/s_catan_template.c b/math/s_catan_template.c index be3459c..6a89309 100644 --- a/math/s_catan_template.c +++ b/math/s_catan_template.c @@ -106,7 +106,7 @@ M_DECL_FUNC (__catan) (CFLOAT x) if (M_FABS (__imag__ x) == 1 && M_FABS (__real__ x) < M_EPSILON * M_EPSILON) __imag__ res = (M_COPYSIGN (M_LIT (0.5), __imag__ x) - * ((FLOAT) M_MLIT (M_LN2) + * (M_MLIT (M_LN2) - M_LOG (M_FABS (__real__ x)))); else { diff --git a/math/s_catanh_template.c b/math/s_catanh_template.c index 59c8e36..7b57dda 100644 --- a/math/s_catanh_template.c +++ b/math/s_catanh_template.c @@ -75,7 +75,7 @@ M_DECL_FUNC (__catanh) (CFLOAT x) if (M_FABS (__real__ x) == 1 && M_FABS (__imag__ x) < M_EPSILON * M_EPSILON) __real__ res = (M_COPYSIGN (M_LIT (0.5), __real__ x) - * ((FLOAT) M_MLIT (M_LN2) + * (M_MLIT (M_LN2) - M_LOG (M_FABS (__imag__ x)))); else { diff --git a/math/s_clog10_template.c b/math/s_clog10_template.c index f5a0873..e50bd6c 100644 --- a/math/s_clog10_template.c +++ b/math/s_clog10_template.c @@ -72,7 +72,7 @@ M_DECL_FUNC (__clog10) (CFLOAT x) if (absx == 1 && scale == 0) { __real__ result = (M_LOG1P (absy * absy) - * ((FLOAT) M_MLIT (M_LOG10E) / 2)); + * (M_MLIT (M_LOG10E) / 2)); math_check_force_underflow_nonneg (__real__ result); } else if (absx > 1 && absx < 2 && absy < 1 && scale == 0) @@ -80,7 +80,7 @@ M_DECL_FUNC (__clog10) (CFLOAT x) FLOAT d2m1 = (absx - 1) * (absx + 1); if (absy >= M_EPSILON) d2m1 += absy * absy; - __real__ result = M_LOG1P (d2m1) * ((FLOAT) M_MLIT (M_LOG10E) / 2); + __real__ result = M_LOG1P (d2m1) * (M_MLIT (M_LOG10E) / 2); } else if (absx < 1 && absx >= M_LIT (0.5) @@ -88,7 +88,7 @@ M_DECL_FUNC (__clog10) (CFLOAT x) && scale == 0) { FLOAT d2m1 = (absx - 1) * (absx + 1); - __real__ result = M_LOG1P (d2m1) * ((FLOAT) M_MLIT (M_LOG10E) / 2); + __real__ result = M_LOG1P (d2m1) * (M_MLIT (M_LOG10E) / 2); } else if (absx < 1 && absx >= M_LIT (0.5) @@ -96,7 +96,7 @@ M_DECL_FUNC (__clog10) (CFLOAT x) && absx * absx + absy * absy >= M_LIT (0.5)) { FLOAT d2m1 = M_SUF (__x2y2m1) (absx, absy); - __real__ result = M_LOG1P (d2m1) * ((FLOAT) M_MLIT (M_LOG10E) / 2); + __real__ result = M_LOG1P (d2m1) * (M_MLIT (M_LOG10E) / 2); } else { diff --git a/math/s_clog_template.c b/math/s_clog_template.c index 3b25bc9..eff3d48 100644 --- a/math/s_clog_template.c +++ b/math/s_clog_template.c @@ -32,7 +32,7 @@ M_DECL_FUNC (__clog) (CFLOAT x) if (__glibc_unlikely (rcls == FP_ZERO && icls == FP_ZERO)) { /* Real and imaginary part are 0.0. */ - __imag__ result = signbit (__real__ x) ? (FLOAT) M_MLIT (M_PI) : 0; + __imag__ result = signbit (__real__ x) ? M_MLIT (M_PI) : 0; __imag__ result = M_COPYSIGN (__imag__ result, __imag__ x); /* Yes, the following line raises an exception. */ __real__ result = -1 / M_FABS (__real__ x); @@ -94,7 +94,7 @@ M_DECL_FUNC (__clog) (CFLOAT x) else { FLOAT d = M_HYPOT (absx, absy); - __real__ result = M_LOG (d) - scale * (FLOAT) M_MLIT (M_LN2); + __real__ result = M_LOG (d) - scale * M_MLIT (M_LN2); } __imag__ result = M_ATAN2 (__imag__ x, __real__ x); diff --git a/sysdeps/generic/math-type-macros-float.h b/sysdeps/generic/math-type-macros-float.h index 5ee0ec3..3299abc 100644 --- a/sysdeps/generic/math-type-macros-float.h +++ b/sysdeps/generic/math-type-macros-float.h @@ -27,9 +27,8 @@ #define M_STRTO_NAN __strtof_nan #define M_USE_BUILTIN(c) USE_ ##c ##F_BUILTIN -/* Standard/GNU macro literals do not exist for the float type. Use - the double macro constants. */ -#define M_MLIT(c) c +/* GNU extension float constant macros. */ +#define M_MLIT(c) c ## f #include #include diff --git a/sysdeps/ieee754/flt-32/e_gammaf_r.c b/sysdeps/ieee754/flt-32/e_gammaf_r.c index 782beaa..3db8463 100644 --- a/sysdeps/ieee754/flt-32/e_gammaf_r.c +++ b/sysdeps/ieee754/flt-32/e_gammaf_r.c @@ -85,7 +85,7 @@ gammaf_positive (float x, int *exp2_adj) float x_adj_frac = x_adj - x_adj_int; int x_adj_log2; float x_adj_mant = __frexpf (x_adj, &x_adj_log2); - if (x_adj_mant < (float) M_SQRT1_2) + if (x_adj_mant < M_SQRT1_2f) { x_adj_log2--; x_adj_mant *= 2.0f; @@ -94,7 +94,7 @@ gammaf_positive (float x, int *exp2_adj) float ret = (__ieee754_powf (x_adj_mant, x_adj) * __ieee754_exp2f (x_adj_log2 * x_adj_frac) * __ieee754_expf (-x_adj) - * sqrtf (2 * (float) M_PI / x_adj) + * sqrtf (2 * M_PIf / x_adj) / prod); exp_adj += x_eps * __ieee754_logf (x_adj); float bsum = gamma_coeff[NCOEFF - 1]; @@ -176,11 +176,11 @@ __ieee754_gammaf_r (float x, int *signgamp) if (frac > 0.5f) frac = 1.0f - frac; float sinpix = (frac <= 0.25f - ? __sinf ((float) M_PI * frac) - : __cosf ((float) M_PI * (0.5f - frac))); + ? __sinf (M_PIf * frac) + : __cosf (M_PIf * (0.5f - frac))); int exp2_adj; - float tret = (float) M_PI / (-x * sinpix - * gammaf_positive (-x, &exp2_adj)); + float tret = M_PIf / (-x * sinpix + * gammaf_positive (-x, &exp2_adj)); ret = __scalbnf (tret, -exp2_adj); math_check_force_underflow_nonneg (ret); } diff --git a/sysdeps/ieee754/flt-32/e_j0f.c b/sysdeps/ieee754/flt-32/e_j0f.c index 0453a30..9ae91a9 100644 --- a/sysdeps/ieee754/flt-32/e_j0f.c +++ b/sysdeps/ieee754/flt-32/e_j0f.c @@ -233,7 +233,7 @@ j0f_near_root (float x, float z) float index_f; int index; - index_f = roundf ((x - FIRST_ZERO_J0) / (float) M_PI); + index_f = roundf ((x - FIRST_ZERO_J0) / M_PIf); /* j0f_asympt fails to give an error <= 9 ulps for x=0x1.324e92p+7 (index 48) thus we can't reduce SMALL_SIZE below 49. */ if (index_f >= SMALL_SIZE) @@ -514,7 +514,7 @@ y0f_near_root (float x, float z) float index_f; int index; - index_f = roundf ((x - FIRST_ZERO_Y0) / (float) M_PI); + index_f = roundf ((x - FIRST_ZERO_Y0) / M_PIf); if (index_f >= SMALL_SIZE) return y0f_asympt (x); index = (int) index_f; diff --git a/sysdeps/ieee754/flt-32/e_j1f.c b/sysdeps/ieee754/flt-32/e_j1f.c index 052c661..ade1040 100644 --- a/sysdeps/ieee754/flt-32/e_j1f.c +++ b/sysdeps/ieee754/flt-32/e_j1f.c @@ -243,7 +243,7 @@ j1f_near_root (float x, float z) x = -x; sign = -1.0f; } - index_f = roundf ((x - FIRST_ZERO_J1) / (float) M_PI); + index_f = roundf ((x - FIRST_ZERO_J1) / M_PIf); if (index_f >= SMALL_SIZE) return sign * j1f_asympt (x); index = (int) index_f; @@ -525,7 +525,7 @@ y1f_near_root (float x, float z) float index_f; int index; - index_f = roundf ((x - FIRST_ZERO_Y1) / (float) M_PI); + index_f = roundf ((x - FIRST_ZERO_Y1) / M_PIf); if (index_f >= SMALL_SIZE) return y1f_asympt (x); index = (int) index_f; diff --git a/sysdeps/ieee754/flt-32/e_jnf.c b/sysdeps/ieee754/flt-32/e_jnf.c index 80b684d..ff0a9d7 100644 --- a/sysdeps/ieee754/flt-32/e_jnf.c +++ b/sysdeps/ieee754/flt-32/e_jnf.c @@ -134,7 +134,7 @@ __ieee754_jnf(int n, float x) tmp = n; v = two/x; tmp = tmp*__ieee754_logf(fabsf(v*tmp)); - if(tmp<(float)8.8721679688e+01) { + if(tmp<8.8721679688e+01f) { for(i=n-1,di=(float)(i+i);i>0;i--){ temp = b; b *= di; diff --git a/sysdeps/ieee754/flt-32/lgamma_negf.c b/sysdeps/ieee754/flt-32/lgamma_negf.c index 756742e..541b5a3 100644 --- a/sysdeps/ieee754/flt-32/lgamma_negf.c +++ b/sysdeps/ieee754/flt-32/lgamma_negf.c @@ -165,9 +165,9 @@ static float lg_sinpi (float x) { if (x <= 0.25f) - return __sinf ((float) M_PI * x); + return __sinf (M_PIf * x); else - return __cosf ((float) M_PI * (0.5f - x)); + return __cosf (M_PIf * (0.5f - x)); } /* Compute cos (pi * X) for -0.25 <= X <= 0.5. */ @@ -176,9 +176,9 @@ static float lg_cospi (float x) { if (x <= 0.25f) - return __cosf ((float) M_PI * x); + return __cosf (M_PIf * x); else - return __sinf ((float) M_PI * (0.5f - x)); + return __sinf (M_PIf * (0.5f - x)); } /* Compute cot (pi * X) for -0.25 <= X <= 0.5. */ diff --git a/sysdeps/ieee754/flt-32/s_log1pf.c b/sysdeps/ieee754/flt-32/s_log1pf.c index 6630f5a..555f0f8 100644 --- a/sysdeps/ieee754/flt-32/s_log1pf.c +++ b/sysdeps/ieee754/flt-32/s_log1pf.c @@ -92,7 +92,7 @@ __log1pf(float x) if(k==0) return zero; else {c += k*ln2_lo; return k*ln2_hi+c;} } - R = hfsq*((float)1.0-(float)0.66666666666666666*f); + R = hfsq*(1.0f-0.66666666666666666f*f); if(k==0) return f-R; else return k*ln2_hi-((R-(k*ln2_lo+c))-f); } -- 2.7.4