[COMMITTED] math: Sync lgammaf with CORE-MATH
Paul Zimmermann
Paul.Zimmermann@inria.fr
Sat Mar 21 07:01:06 GMT 2026
Hi Adhemerval,
I checked again with the below patch that you have committed, and all tests
now pass (tested with FMA only). Thank you!
Paul
> Date: Fri, 20 Mar 2026 14:09:52 -0300
> Cc: libc-alpha@sourceware.org
> From: Adhemerval Zanella Netto <adhemerval.zanella@linaro.org>
> Organization: Linaro
>
>
>
> On 20/03/26 07:42, Paul Zimmermann wrote:
> > Hi Adhemerval,
> >
> > with the CORE-MATH test suite, I get:
> >
> > zimmerma@thym:~/svn/core-math$ CORE_MATH_CHECK_STD=true CORE_MATH_LAUNCHER="/tmp/lib/ld-linux-x86-64.so.2 --library-path /tmp/lib:/usr/lib/x86_64-linux-gnu" LDFLAGS="-L /tmp/lib" ./check.sh lgammaf
> > Running exhaustive check in --rndn mode...
> > all ok
> > Running exhaustive check in --rndz mode...
> > Spurious overflow exception for x=0x1.895f1cp+121 (y=0x1.fffffep+127)
> >
> > Please can you check?
>
> It seems to be a pre-existing issue, on 5c28a3cf9f (pre-sync) if I add
> 0x1.895f1cp+121 on math/auto-libm-test-in and regenerating
> math/auto-libm-test-out-lgamma (as below) I see:
>
> $ math/test-float-lgamma
> testing float (without inline functions)
> Failure: lgamma (0x3.12be38p+120): errno set to 0, expected 34 (ERANGE)
> Failure: lgamma_upward (0x3.12be38p+120): errno set to 0, expected 34 (ERANGE)
>
> Test suite completed:
> Maximum error found of `0' ulp
> 8 max error test cases,
> 7368 input tests,
> - with 29488 tests for exception flags,
> - with 3688 tests for errno executed.
> 2 errors occurred.
>
> I forgot to sync the some ERANGE fixes like 20d51f2ee1e945cdc910aa3fbd40c7d5959ff3d0.
> I will add the following patch:
>
> diff --git a/math/auto-libm-test-in b/math/auto-libm-test-in
> index 4c3cd23ffd..2574957ada 100644
> --- a/math/auto-libm-test-in
> +++ b/math/auto-libm-test-in
> @@ -6994,6 +6994,7 @@ lgamma -0x3.ec4298p+0
> lgamma 0x1.ecf3fep-73
> lgamma 0x1.58ace8p+112
> lgamma -0x1.efc2a2p+14
> +lgamma 0x1.895f1cp+121
>
> # Values +/- 10ulp from overflow threshold. (Values very close to
> # overflow threshold produce results very close of that threshold,
> diff --git a/math/auto-libm-test-out-lgamma b/math/auto-libm-test-out-lgamma
> index d27c186639..fed3f86c6c 100644
> --- a/math/auto-libm-test-out-lgamma
> +++ b/math/auto-libm-test-out-lgamma
> @@ -2301,6 +2301,31 @@ lgamma -0x1.efc2a2p+14
> = lgamma tonearest ibm128 -0x7.bf0a88p+12 : -0x4.88b6f00000005971e29c3b8dd8p+16 -1 : inexact-ok
> = lgamma towardzero ibm128 -0x7.bf0a88p+12 : -0x4.88b6f00000005971e29c3b8dd8p+16 -1 : inexact-ok
> = lgamma upward ibm128 -0x7.bf0a88p+12 : -0x4.88b6f00000005971e29c3b8dd8p+16 -1 : inexact-ok
> +lgamma 0x1.895f1cp+121
> += lgamma downward binary32 0x3.12be38p+120 : 0xf.fffffp+124 1 : inexact-ok
> += lgamma tonearest binary32 0x3.12be38p+120 : plus_infty 1 : inexact-ok overflow errno-erange
> += lgamma towardzero binary32 0x3.12be38p+120 : 0xf.fffffp+124 1 : inexact-ok
> += lgamma upward binary32 0x3.12be38p+120 : plus_infty 1 : inexact-ok overflow errno-erange
> += lgamma downward binary64 0x3.12be38p+120 : 0xf.fffffbf5af3fp+124 1 : inexact-ok
> += lgamma tonearest binary64 0x3.12be38p+120 : 0xf.fffffbf5af3f8p+124 1 : inexact-ok
> += lgamma towardzero binary64 0x3.12be38p+120 : 0xf.fffffbf5af3fp+124 1 : inexact-ok
> += lgamma upward binary64 0x3.12be38p+120 : 0xf.fffffbf5af3f8p+124 1 : inexact-ok
> += lgamma downward intel96 0x3.12be38p+120 : 0xf.fffffbf5af3f533p+124 1 : inexact-ok
> += lgamma tonearest intel96 0x3.12be38p+120 : 0xf.fffffbf5af3f534p+124 1 : inexact-ok
> += lgamma towardzero intel96 0x3.12be38p+120 : 0xf.fffffbf5af3f533p+124 1 : inexact-ok
> += lgamma upward intel96 0x3.12be38p+120 : 0xf.fffffbf5af3f534p+124 1 : inexact-ok
> += lgamma downward m68k96 0x3.12be38p+120 : 0xf.fffffbf5af3f533p+124 1 : inexact-ok
> += lgamma tonearest m68k96 0x3.12be38p+120 : 0xf.fffffbf5af3f534p+124 1 : inexact-ok
> += lgamma towardzero m68k96 0x3.12be38p+120 : 0xf.fffffbf5af3f533p+124 1 : inexact-ok
> += lgamma upward m68k96 0x3.12be38p+120 : 0xf.fffffbf5af3f534p+124 1 : inexact-ok
> += lgamma downward binary128 0x3.12be38p+120 : 0xf.fffffbf5af3f533cd835e6dc7ed8p+124 1 : inexact-ok
> += lgamma tonearest binary128 0x3.12be38p+120 : 0xf.fffffbf5af3f533cd835e6dc7ed8p+124 1 : inexact-ok
> += lgamma towardzero binary128 0x3.12be38p+120 : 0xf.fffffbf5af3f533cd835e6dc7ed8p+124 1 : inexact-ok
> += lgamma upward binary128 0x3.12be38p+120 : 0xf.fffffbf5af3f533cd835e6dc7eep+124 1 : inexact-ok
> += lgamma downward ibm128 0x3.12be38p+120 : 0xf.fffffbf5af3f533cd835e6dc7cp+124 1 : inexact-ok
> += lgamma tonearest ibm128 0x3.12be38p+120 : 0xf.fffffbf5af3f533cd835e6dc8p+124 1 : inexact-ok
> += lgamma towardzero ibm128 0x3.12be38p+120 : 0xf.fffffbf5af3f533cd835e6dc7cp+124 1 : inexact-ok
> += lgamma upward ibm128 0x3.12be38p+120 : 0xf.fffffbf5af3f533cd835e6dc8p+124 1 : inexact-ok
> lgamma 0x3.12be0cp+120
> = lgamma downward binary32 0x3.12be0cp+120 : 0xf.ffff1p+124 1 : inexact-ok
> = lgamma tonearest binary32 0x3.12be0cp+120 : 0xf.ffff1p+124 1 : inexact-ok
> diff --git a/sysdeps/ieee754/flt-32/e_lgammaf_r.c b/sysdeps/ieee754/flt-32/e_lgammaf_r.c
> index 7b947cd979..8efec19b48 100644
> --- a/sysdeps/ieee754/flt-32/e_lgammaf_r.c
> +++ b/sysdeps/ieee754/flt-32/e_lgammaf_r.c
> @@ -208,7 +208,19 @@ __lgammaf_r (float x, int *signgamp)
> if (ax > 0x1.afc1ap+1f)
> {
> if (__glibc_unlikely (x >= 0x1.895f1cp+121f))
> - return __math_oflowf (0);
> + {
> + /* for x=0x1.895f1cp+121, lgamma(x) < 2^128, thus there is no
> + overflow for rounding towards zero or downwards.
> + The following expression overflows for x > 0x1.895f1cp+121
> + or x = 0x1.895f1cp+121 and rounding to nearest or away,
> + and does not overflow for x = 0x1.895f1cp+121 and rounding
> + towards zero */
> + float r = fmaf (x, 0x1.4d3398p+6f, 0x1.10f35ep+103f);
> + if (x > 0x1.895f1cp+121f || (x == 0x1.895f1cp+121f &&
> + x * 5.0f >= 0x1.ebb6e4p+123))
> + return __math_oflowf (0);
> + return r;
> + }
> double lz = as_ln (z);
> f = (z - 0.5) * (lz - 1) + 0x1.acfe390c97d69p-2;
> if (ax < 0x1.0p+20f)
>
More information about the Libc-alpha
mailing list