diff mbox series

[3/4] math: Remove UB and optimize double ilogbf

Message ID 20250425205309.3866442-4-adhemerval.zanella@linaro.org
State New
Headers show
Series Remove UB and optimize ilogbf/ilogb | expand

Commit Message

Adhemerval Zanella April 25, 2025, 8:44 p.m. UTC 
The subnormal exponent calculation invokes UB by left shifting the
signed expoenent to find the first leading bit.

The patch reimplements ilogb using the math_config.h macros and
uses the new stdbit function to simplify the subnormal handling.

On aarch64 it generates better code:

* master:

0000000000000000 <__ieee754_ilogbf>:
   0:   1e260000        fmov    w0, s0
   4:   12007801        and     w1, w0, #0x7fffffff
   8:   72091c1f        tst     w0, #0x7f800000
   c:   54000141        b.ne    34 <__ieee754_ilogbf+0x34>  // b.any
  10:   34000201        cbz     w1, 50 <__ieee754_ilogbf+0x50>
  14:   53185c21        lsl     w1, w1, #8
  18:   12800fa0        mov     w0, #0xffffff82                 // #-126
  1c:   d503201f        nop
  20:   531f7821        lsl     w1, w1, #1
  24:   51000400        sub     w0, w0, #0x1
  28:   7100003f        cmp     w1, #0x0
  2c:   54ffffac        b.gt    20 <__ieee754_ilogbf+0x20>
  30:   d65f03c0        ret
  34:   13177c20        asr     w0, w1, #23
  38:   12b01002        mov     w2, #0x7f7fffff                 // #2139095039
  3c:   5101fc00        sub     w0, w0, #0x7f
  40:   6b02003f        cmp     w1, w2
  44:   12b00001        mov     w1, #0x7fffffff                 // #2147483647
  48:   1a819000        csel    w0, w0, w1, ls  // ls = plast
  4c:   d65f03c0        ret
  50:   320107e0        mov     w0, #0x80000001                 // #-2147483647
  54:   d65f03c0        ret

* patch:

0000000000000000 <__ieee754_ilogbf>:
   0:   1e260001        fmov    w1, s0
   4:   d3577820        ubfx    x0, x1, #23, #8
   8:   350000e0        cbnz    w0, 24 <__ieee754_ilogbf+0x24>
   c:   53175821        lsl     w1, w1, #9
  10:   34000141        cbz     w1, 38 <__ieee754_ilogbf+0x38>
  14:   5ac01021        clz     w1, w1
  18:   12800fc0        mov     w0, #0xffffff81                 // #-127
  1c:   4b010000        sub     w0, w0, w1
  20:   d65f03c0        ret
  24:   7103fc1f        cmp     w0, #0xff
  28:   5101fc00        sub     w0, w0, #0x7f
  2c:   12b00001        mov     w1, #0x7fffffff                 // #2147483647
  30:   1a811000        csel    w0, w0, w1, ne  // ne = any
  34:   d65f03c0        ret
  38:   320107e0        mov     w0, #0x80000001                 // #-2147483647
  3c:   d65f03c0        ret

Other architecture with support for stdc_leading_zeros and/or
__builtin_clzll should have similar improvements.

Checked on aarch64-linux-gnu and x86_64-linux-gnu.
---
 sysdeps/ieee754/flt-32/e_ilogbf.c | 68 +++++++++++++++----------------
 1 file changed, 33 insertions(+), 35 deletions(-)
diff mbox series

Patch

diff --git a/sysdeps/ieee754/flt-32/e_ilogbf.c b/sysdeps/ieee754/flt-32/e_ilogbf.c
index db24012eb4..024b114638 100644
--- a/sysdeps/ieee754/flt-32/e_ilogbf.c
+++ b/sysdeps/ieee754/flt-32/e_ilogbf.c
@@ -1,43 +1,41 @@ 
-/* s_ilogbf.c -- float version of s_ilogb.c.
- */
+/* Get integer exponent of a floating-point value.
+   Copyright (C) 1999-2025 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
 
-/*
- * ====================================================
- * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
- *
- * Developed at SunPro, a Sun Microsystems, Inc. business.
- * Permission to use, copy, modify, and distribute this
- * software is freely granted, provided that this notice
- * is preserved.
- * ====================================================
- */
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
 
-#if defined(LIBM_SCCS) && !defined(lint)
-static char rcsid[] = "$NetBSD: s_ilogbf.c,v 1.4 1995/05/10 20:47:31 jtc Exp $";
-#endif
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
 
 #include <limits.h>
 #include <math.h>
-#include <math_private.h>
+#include <stdbit.h>
+#include "math_config.h"
 
-int __ieee754_ilogbf(float x)
+int
+__ieee754_ilogbf (float x)
 {
-	int32_t hx,ix;
-
-	GET_FLOAT_WORD(hx,x);
-	hx &= 0x7fffffff;
-	if(hx<0x00800000) {
-	    if(hx==0)
-		return FP_ILOGB0;	/* ilogb(0) = FP_ILOGB0 */
-	    else			/* subnormal x */
-	        for (ix = -126,hx<<=8; hx>0; hx<<=1) ix -=1;
-	    return ix;
-	}
-	else if (hx<0x7f800000) return (hx>>23)-127;
-	else if (FP_ILOGBNAN != INT_MAX) {
-	    /* ISO C99 requires ilogbf(+-Inf) == INT_MAX.  */
-	    if (hx==0x7f800000)
-		return INT_MAX;
-	}
-	return FP_ILOGBNAN;
+  uint32_t ux = asuint (x);
+  int ex = (ux & ~SIGN_MASK) >> MANTISSA_WIDTH;
+  if (ex == 0) /* zero or subnormal */
+    {
+      /* Clear sign and exponent.  */
+      ux <<= 1 + EXPONENT_WIDTH;
+      if (ux == 0)
+	return FP_ILOGB0;
+      /* sbunormal */
+      return -127 - stdc_leading_zeros (ux);
+    }
+  if (ex == EXPONENT_MASK >> MANTISSA_WIDTH) /* NaN or Inf */
+    return ux << (1 + EXPONENT_WIDTH) ? FP_ILOGBNAN : INT_MAX;
+  return ex - 127;
 }