math: Remove the SVID error handling from tgammaf

It improves latency for about 1.5% and throughput for about 2-4%. Tested on x86_64-linux-gnu and i686-linux-gnu. Reviewed-by: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
2026-10-24 01:33:52 +03:00 · 2025-11-05 08:10:01 -03:00
parent de0e623434
commit 3078358ac6
32 changed files with 65 additions and 33 deletions
--- a/math/Versions
+++ b/math/Versions
@@ -707,5 +707,6 @@ libm {
    y0f;
    y1f;
    ynf;
+    tgammaf;
  }
 }
--- a/math/w_tgammaf_compat.c
+++ b/math/w_tgammaf_compat.c
@@ -18,12 +18,13 @@
 #include <math_private.h>
 #include <math-svid-compat.h>
 #include <libm-alias-float.h>
+#include <shlib-compat.h>

-#if LIBM_SVID_COMPAT
+#if LIBM_SVID_COMPAT && SHLIB_COMPAT (libm, GLIBC_2_1, GLIBC_2_43)
 float
-__tgammaf(float x)
+__tgammaf_svid(float x)
 {
-	float y = __ieee754_gammaf_r(x, NULL);
+	float y = __tgammaf(x);

 	if(__glibc_unlikely (!isfinite (y) || y == 0)
 	   && (isfinite (x) || (isinf (x) && x < 0.0f))
@@ -43,5 +44,5 @@ __tgammaf(float x)
 	}
 	return y;
 }
-libm_alias_float (__tgamma, tgamma)
+compat_symbol (libm, __tgammaf_svid, tgammaf, GLIBC_2_1);
 #endif
--- a/sysdeps/generic/math_private_calls.h
+++ b/sysdeps/generic/math_private_calls.h
@@ -36,7 +36,9 @@ extern _Mdouble_ __MSUF (__ieee754_exp10) (_Mdouble_);
 extern _Mdouble_ __MSUF (__ieee754_exp2) (_Mdouble_);
 extern _Mdouble_ __MSUF (__ieee754_fmod) (_Mdouble_, _Mdouble_);
 extern _Mdouble_ __MSUF (__ieee754_gamma) (_Mdouble_);
+#if !defined __MATH_DECLARING_FLOAT
 extern _Mdouble_ __MSUF_R (__ieee754_gamma) (_Mdouble_, int *);
+#endif
 extern _Mdouble_ __MSUF (__ieee754_hypot) (_Mdouble_, _Mdouble_);
 extern _Mdouble_ __MSUF (__ieee754_j0) (_Mdouble_);
 extern _Mdouble_ __MSUF (__ieee754_j1) (_Mdouble_);
--- a/sysdeps/ieee754/flt-32/e_gammaf_r.c
+++ b/sysdeps/ieee754/flt-32/e_gammaf_r.c
@@ -25,31 +25,23 @@ SOFTWARE.
 */

 /* Changes with respect to the original CORE-MATH code:
-   - removed the dealing with errno
-     (this is done in the wrapper math/w_tgammaf_compat.c)
   - usage of math_narrow_eval to deal with underflow/overflow
-   - deal with signgamp
 */

+#include <array_length.h>
 #include <math.h>
 #include <float.h>
 #include <stdint.h>
 #include <stddef.h>
+#include <libm-alias-float.h>
+#include <math-svid-compat.h>
 #include <libm-alias-finite.h>
 #include <math-narrow-eval.h>
 #include "math_config.h"

 float
-__ieee754_gammaf_r (float x, int *signgamp)
+__tgammaf (float x)
 {
-  /* The wrapper in math/w_tgamma_template.c expects *signgamp to be set to a
-     non-negative value if the returned value is gamma(x), and to a negative
-     value if it is -gamma(x).
-     Since the code here directly computes gamma(x), we set it to 1.
-  */
-  if (signgamp != NULL)
-    *signgamp = 1;
-
  /* List of exceptional cases. Each entry contains the 32-bit encoding u of x,
     a binary32 approximation f of gamma(x), and a correction term df.  */
  static const struct
@@ -92,27 +84,22 @@ __ieee754_gammaf_r (float x, int *signgamp)
      uint64_t rt = asuint64 (f);
      if (((rt + 2) & 0xfffffff) < 4)
 	{
-	  for (unsigned i = 0; i < sizeof (tb) / sizeof (tb[0]); i++)
+	  for (unsigned i = 0; i < array_length (tb); i++)
 	    if (t == tb[i].u)
 	      return tb[i].f + tb[i].df;
 	}
+      if (isinf (r))
+	return __math_erangef (r);
      return r;
    }
  float fx = floorf (x);
  if (__glibc_unlikely (x >= 0x1.18522p+5f))
-    {
-      /* Overflow case.  The original CORE-MATH code returns
-	 0x1p127f * 0x1p127f, but apparently some compilers replace this
-	 by +Inf.  */
-      return math_narrow_eval (x * 0x1p127f);
-    }
+    return __math_oflowf (0);
  /* compute k only after the overflow check, otherwise the case to integer
     might overflow */
  int k = fx;
  if (__glibc_unlikely (fx == x))
    { /* x is integer */
-      if (x == 0.0f)
-	return 1.0f / x;
      if (x < 0.0f)
 	return __math_invalidf (0.0f);
      double t0 = 1, x0 = 1;
@@ -121,12 +108,10 @@ __ieee754_gammaf_r (float x, int *signgamp)
      return t0;
    }
  if (__glibc_unlikely (x < -42.0f))
-    { /* negative non-integer */
-      /* For x < -42, x non-integer, |gamma(x)| < 2^-151.  */
-      static const float sgn[2] = { 0x1p-127f, -0x1p-127f };
-      /* Underflows always happens */
-      return math_narrow_eval (0x1p-127f * sgn[k & 1]);
-    }
+    /* negative non-integer */
+    /* For x < -42, x non-integer, |gamma(x)| < 2^-151.  */
+    /* Underflows always happens */
+    return __math_uflowf (k & 1);
  /* The array c[] stores a degree-15 polynomial approximation for
     gamma(x).  */
  static const double c[] =
@@ -164,13 +149,22 @@ __ieee754_gammaf_r (float x, int *signgamp)
  f *= w;
  uint64_t rt = asuint64 (f);
  float r = f;
+  if (__glibc_unlikely (r == 0))
+    return __math_uflowf (rt >> 63);
  /* Deal with exceptional cases.  */
  if (__glibc_unlikely (((rt + 2) & 0xfffffff) < 8))
    {
-      for (unsigned j = 0; j < sizeof (tb) / sizeof (tb[0]); j++)
+      for (unsigned j = 0; j < array_length (tb); j++)
 	if (t == tb[j].u)
 	  return tb[j].f + tb[j].df;
    }
  return r;
 }
-libm_alias_finite (__ieee754_gammaf_r, __gammaf_r)
+
+#if LIBM_SVID_COMPAT
+versioned_symbol (libm, __tgammaf, tgammaf, GLIBC_2_43);
+libm_alias_float_other (__tgamma, tgamma)
+#else
+libm_alias_float (__tgamma, tgamma)
+#endif
+libm_alias_finite (__tgammaf, __gammaf_r)
--- a/sysdeps/ieee754/flt-32/math_config.h
+++ b/sysdeps/ieee754/flt-32/math_config.h
@@ -225,6 +225,7 @@ attribute_hidden float __math_invalidf (float);
 attribute_hidden int __math_invalidf_i (int);
 attribute_hidden long int __math_invalidf_li (long int);
 attribute_hidden float __math_edomf (float x);
+attribute_hidden float __math_erangef (float x);

 /* Shared between expf, exp2f, exp10f, and powf.  */
 #define EXP2F_TABLE_BITS 5
--- a/sysdeps/ieee754/flt-32/math_errf.c
+++ b/sysdeps/ieee754/flt-32/math_errf.c
@@ -54,6 +54,12 @@ __math_edomf (float y)
  return with_errnof (y, EDOM);
 }

+attribute_hidden float
+__math_erangef (float y)
+{
+  return with_errnof (y, ERANGE);
+}
+
 /* NOINLINE prevents fenv semantics breaking optimizations.  */
 NOINLINE static float
 xflowf (uint32_t sign, float y)
--- a/sysdeps/ieee754/flt-32/w_tgammaf.c
+++ b/sysdeps/ieee754/flt-32/w_tgammaf.c
@@ -0,0 +1 @@
+/* Not needed.  */
--- a/sysdeps/mach/hurd/i386/libm.abilist
+++ b/sysdeps/mach/hurd/i386/libm.abilist
@@ -1335,6 +1335,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/aarch64/libm.abilist
+++ b/sysdeps/unix/sysv/linux/aarch64/libm.abilist
@@ -1301,6 +1301,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/alpha/libm.abilist
+++ b/sysdeps/unix/sysv/linux/alpha/libm.abilist
@@ -1460,6 +1460,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/arm/be/libm.abilist
+++ b/sysdeps/unix/sysv/linux/arm/be/libm.abilist
@@ -966,6 +966,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/arm/le/libm.abilist
+++ b/sysdeps/unix/sysv/linux/arm/le/libm.abilist
@@ -966,6 +966,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/hppa/libm.abilist
+++ b/sysdeps/unix/sysv/linux/hppa/libm.abilist
@@ -966,6 +966,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/i386/libm.abilist
+++ b/sysdeps/unix/sysv/linux/i386/libm.abilist
@@ -1342,6 +1342,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/m68k/coldfire/libm.abilist
+++ b/sysdeps/unix/sysv/linux/m68k/coldfire/libm.abilist
@@ -966,6 +966,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/m68k/m680x0/libm.abilist
+++ b/sysdeps/unix/sysv/linux/m68k/m680x0/libm.abilist
@@ -997,6 +997,7 @@ GLIBC_2.43 lgammaf_r F
 GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/microblaze/be/libm.abilist
+++ b/sysdeps/unix/sysv/linux/microblaze/be/libm.abilist
@@ -966,6 +966,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/microblaze/le/libm.abilist
+++ b/sysdeps/unix/sysv/linux/microblaze/le/libm.abilist
@@ -966,6 +966,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/mips/mips32/libm.abilist
+++ b/sysdeps/unix/sysv/linux/mips/mips32/libm.abilist
@@ -966,6 +966,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/mips/mips64/libm.abilist
+++ b/sysdeps/unix/sysv/linux/mips/mips64/libm.abilist
@@ -1301,6 +1301,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/powerpc/powerpc32/fpu/libm.abilist
+++ b/sysdeps/unix/sysv/linux/powerpc/powerpc32/fpu/libm.abilist
@@ -1113,6 +1113,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/powerpc/powerpc32/nofpu/libm.abilist
+++ b/sysdeps/unix/sysv/linux/powerpc/powerpc32/nofpu/libm.abilist
@@ -1105,12 +1105,14 @@ GLIBC_2.43 coshf F
 GLIBC_2.43 j0f F
 GLIBC_2.43 j1f F
 GLIBC_2.43 jnf F
+GLIBC_2.43 lgammaf F
 GLIBC_2.43 lgammaf_r F
 GLIBC_2.43 log10f F
 GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/powerpc/powerpc64/be/libm.abilist
+++ b/sysdeps/unix/sysv/linux/powerpc/powerpc64/be/libm.abilist
@@ -1106,6 +1106,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/powerpc/powerpc64/le/libm.abilist
+++ b/sysdeps/unix/sysv/linux/powerpc/powerpc64/le/libm.abilist
@@ -1490,6 +1490,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/s390/s390-32/libm.abilist
+++ b/sysdeps/unix/sysv/linux/s390/s390-32/libm.abilist
@@ -1404,6 +1404,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/s390/s390-64/libm.abilist
+++ b/sysdeps/unix/sysv/linux/s390/s390-64/libm.abilist
@@ -1404,6 +1404,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/sh/be/libm.abilist
+++ b/sysdeps/unix/sysv/linux/sh/be/libm.abilist
@@ -966,6 +966,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/sh/le/libm.abilist
+++ b/sysdeps/unix/sysv/linux/sh/le/libm.abilist
@@ -966,6 +966,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/sparc/sparc32/libm.abilist
+++ b/sysdeps/unix/sysv/linux/sparc/sparc32/libm.abilist
@@ -1411,6 +1411,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/sparc/sparc64/libm.abilist
+++ b/sysdeps/unix/sysv/linux/sparc/sparc64/libm.abilist
@@ -1301,6 +1301,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/x86_64/64/libm.abilist
+++ b/sysdeps/unix/sysv/linux/x86_64/64/libm.abilist
@@ -1334,6 +1334,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F
--- a/sysdeps/unix/sysv/linux/x86_64/x32/libm.abilist
+++ b/sysdeps/unix/sysv/linux/x86_64/x32/libm.abilist
@@ -1334,6 +1334,7 @@ GLIBC_2.43 remainder F
 GLIBC_2.43 remainderf F
 GLIBC_2.43 sinhf F
 GLIBC_2.43 sqrtf F
+GLIBC_2.43 tgammaf F
 GLIBC_2.43 y0f F
 GLIBC_2.43 y1f F
 GLIBC_2.43 ynf F