regression in tgamma?

regression in tgamma?

[Paul Zimmermann]

       Hi,

while updating my comparison from Newlib 3.3.0 to Newlib 4.0.0, I found a
regression in tgamma on x86_64/Linux:

$ cat test3.c
#include <stdio.h>
#include <math.h>

#ifdef NEWLIB
int errno;
int* __errno () { return &errno; }
#endif

int main()
{
  double x = -0x1.53f198fe3b278p+7, y;
  y = tgamma (x);
  printf ("x=%a y=%a\n", x, y);
}

Newlib 3.3.0:
$ gcc -DNEWLIB -no-pie test3.c /localdisk/zimmerma/newlib-3.3.0/libm.a; ./a.out
x=-0x1.53f198fe3b278p+7 y=0x1.fd6d312572d9cp-1015

Newlib 4.0.0:
$ gcc -DNEWLIB -no-pie test3.c /localdisk/zimmerma/newlib-4.0.0/libm.a; ./a.out
x=-0x1.53f198fe3b278p+7 y=0x1p+0

Please can someone confirm?

Best regards,
Paul

Re: regression in tgamma?

[Jeff Johnston]

Hi Paul,

I copied the newlib/libm/common directory and header files from master to a
separate directory and called __ieee754_tgamma directly in your test (adding
a declaration of __ieee754_tgamma as well).  I then built as follows:

gcc -g -O0 -I. -DNEWLIB  testtgamma.c e_tgamma.c e_exp.c k_sin.c math_err.c
er_lgamma.c k_cos.c s_floor.c e_log.c w_log.c

I get the same result as you get for 3.3.0.

Can you possibly recheck to see if your 4.0.0 library has been built
correctly?  Perhaps I have somehow by-passed the cause of the problem by
compiling directly.

-- Jeff J.

On Tue, Dec 15, 2020 at 8:36 AM Paul Zimmermann <Paul.Zimmermann@inria.fr>
wrote:

>        Hi,
>
> while updating my comparison from Newlib 3.3.0 to Newlib 4.0.0, I found a
> regression in tgamma on x86_64/Linux:
>
> $ cat test3.c
> #include <stdio.h>
> #include <math.h>
>
> #ifdef NEWLIB
> int errno;
> int* __errno () { return &errno; }
> #endif
>
> int main()
> {
>   double x = -0x1.53f198fe3b278p+7, y;
>   y = tgamma (x);
>   printf ("x=%a y=%a\n", x, y);
> }
>
> Newlib 3.3.0:
> $ gcc -DNEWLIB -no-pie test3.c /localdisk/zimmerma/newlib-3.3.0/libm.a;
> ./a.out
> x=-0x1.53f198fe3b278p+7 y=0x1.fd6d312572d9cp-1015
>
> Newlib 4.0.0:
> $ gcc -DNEWLIB -no-pie test3.c /localdisk/zimmerma/newlib-4.0.0/libm.a;
> ./a.out
> x=-0x1.53f198fe3b278p+7 y=0x1p+0
>
> Please can someone confirm?
>
> Best regards,
> Paul
>
>

Re: regression in tgamma?

[Paul Zimmermann]

       Hi Jeff,

thank you for your answer. Here are the steps I followed to compile Newlib
(thanks Alexei Sibidanov):

# from newlib-3.3.0 or newlib-4.0.0
mkdir build
cd build
../configure --prefix=`pwd` --target=x86_64
# perform the following in $HOME/bin
ln -sf `which ar` x86_64-ar
ln -sf `which as` x86_64-as
ln -sf `which cc` x86_64-cc
ln -sf `which gcc` x86_64-gcc
ln -sf `which c++` x86_64-c++
ln -sf `which ld` x86_64-ld
ln -sf `which nm` x86_64-nm
ln -sf `which objdump` x86_64-objdump
ln -sf `which objcopy` x86_64-objcopy
ln -sf `which ranlib` x86_64-ranlib
ln -sf `which readelf` x86_64-readelf
ln -sf `which strip` x86_64-strip
make # it will fail but it is ok
cd x86_64/newlib/libm
make # fails too
cd tmp
ar cr libm.a *.o

If you have access to the GCC Compile Farm [1], you can reproduce on gcc67:

$ cd /home/zimmerma
$ gcc -DNEWLIB -no-pie test3.c newlib-3.3.0/build/x86_64/newlib/libm/tmp/libm.a
$ ./a.out 
x=-0x1.53f198fe3b278p+7 y=0x1.fd6d312572d9cp-1015
$ gcc -DNEWLIB -no-pie test3.c newlib-4.0.0/build/x86_64/newlib/libm/tmp/libm.a
$ ./a.out 
x=-0x1.53f198fe3b278p+7 y=0x1p+0

If you have a simpler way to compile Newlib, please tell me.

Best regards,
Paul

[1] https://gcc.gnu.org/wiki/CompileFarm#List_of_machines

> From: Jeff Johnston <jjohnstn@redhat.com>
> Date: Tue, 15 Dec 2020 12:11:25 -0500
> 
> Hi Paul,
> 
> I copied the newlib/libm/common directory and header files from master to a
> separate directory and called __ieee754_tgamma directly in your test (adding
> a declaration of __ieee754_tgamma as well).  I then built as follows:
> 
> gcc -g -O0 -I. -DNEWLIB  testtgamma.c e_tgamma.c e_exp.c k_sin.c math_err.c
> er_lgamma.c k_cos.c s_floor.c e_log.c w_log.c
> 
> I get the same result as you get for 3.3.0. 
> 
> Can you possibly recheck to see if your 4.0.0 library has been built
> correctly?  Perhaps I have somehow by-passed the cause of the problem by
> compiling directly.
> 
> -- Jeff J.
> 
> On Tue, Dec 15, 2020 at 8:36 AM Paul Zimmermann <Paul.Zimmermann@inria.fr>
> wrote:
> 
>         Hi,
> 
>  while updating my comparison from Newlib 3.3.0 to Newlib 4.0.0, I found
>  a
>  regression in tgamma on x86_64/Linux:
> 
>  $ cat test3.c
>  #include <stdio.h>
>  #include <math.h>
> 
>  #ifdef NEWLIB
>  int errno;
>  int* __errno () { return &errno; }
>  #endif
> 
>  int main()
>  {
>    double x = -0x1.53f198fe3b278p+7, y;
>    y = tgamma (x);
>    printf ("x=%a y=%a\n", x, y);
>  }
> 
>  Newlib 3.3.0:
>  $ gcc -DNEWLIB -no-pie test3.c /localdisk/zimmerma/newlib-3.3.0/libm.a;
>  ./a.out
>  x=-0x1.53f198fe3b278p+7 y=0x1.fd6d312572d9cp-1015
> 
>  Newlib 4.0.0:
>  $ gcc -DNEWLIB -no-pie test3.c /localdisk/zimmerma/newlib-4.0.0/libm.a;
>  ./a.out
>  x=-0x1.53f198fe3b278p+7 y=0x1p+0
> 
>  Please can someone confirm?
> 
>  Best regards,
>  Paul

regression in tgamma?

[Paul Zimmermann]

       Hi,

the regression I observed also concerns the binary32 code.
Consider the following code:

$ cat /tmp/test_tgamma.c
#include <stdio.h>
#include <math.h>

#ifdef NEWLIB
int errno;
int* __errno () { return &errno; }
#endif

int main()
{
  float x = -0.0f;
  float y = tgammaf (x);
  printf ("x=%a y=%a\n", x, y);
  {
    double x = -0x1.53f198fe3b278p+7, y;
    y = tgamma (x);
    printf ("x=%a y=%a\n", x, y);
  }
}

With commit a0d7982ff486292540078c0121a435013c5ee1ea it gives on x86_64/linux
(Intel(R) Core(TM) i5-4590):

$ gcc -no-pie -DNEWLIB /tmp/test_tgamma.c build/x86_64/newlib/libm/tmp/libm.a
$ ./a.out
x=-0x0p+0 y=inf
x=-0x1.53f198fe3b278p+7 y=0x1.fd6d312572d9cp-1015

and with commit 1f8e5847dff27e504949cd21bfeadb987d36ad19:

$ gcc -no-pie -DNEWLIB /tmp/test_tgamma.c build/x86_64/newlib/libm/tmp/libm.a
$ ./a.out 
x=-0x0p+0 y=0x1p+0
x=-0x1.53f198fe3b278p+7 y=0x1p+0

Paul

Re: regression in tgamma?

[Jeff Johnston]

Hi Paul,

I believe I found the problem.

The __ieee754_tgamma functions are not declared in fdlibm.h so the calls in
tgamma and tgammaf are using
default parms and getting things wrong.  My test calling __ieee754_tgamma
directly only worked after I added
a declaration to __ieee754_tgamma locally.

I have pushed a change to master.  Please try it out.

-- Jeff J.


On Wed, Dec 16, 2020 at 8:49 AM Paul Zimmermann <Paul.Zimmermann@inria.fr>
wrote:

>        Hi,
>
> the regression I observed also concerns the binary32 code.
> Consider the following code:
>
> $ cat /tmp/test_tgamma.c
> #include <stdio.h>
> #include <math.h>
>
> #ifdef NEWLIB
> int errno;
> int* __errno () { return &errno; }
> #endif
>
> int main()
> {
>   float x = -0.0f;
>   float y = tgammaf (x);
>   printf ("x=%a y=%a\n", x, y);
>   {
>     double x = -0x1.53f198fe3b278p+7, y;
>     y = tgamma (x);
>     printf ("x=%a y=%a\n", x, y);
>   }
> }
>
> With commit a0d7982ff486292540078c0121a435013c5ee1ea it gives on
> x86_64/linux
> (Intel(R) Core(TM) i5-4590):
>
> $ gcc -no-pie -DNEWLIB /tmp/test_tgamma.c
> build/x86_64/newlib/libm/tmp/libm.a
> $ ./a.out
> x=-0x0p+0 y=inf
> x=-0x1.53f198fe3b278p+7 y=0x1.fd6d312572d9cp-1015
>
> and with commit 1f8e5847dff27e504949cd21bfeadb987d36ad19:
>
> $ gcc -no-pie -DNEWLIB /tmp/test_tgamma.c
> build/x86_64/newlib/libm/tmp/libm.a
> $ ./a.out
> x=-0x0p+0 y=0x1p+0
> x=-0x1.53f198fe3b278p+7 y=0x1p+0
>
> Paul
>
>
>
>
>
>
>
>

Re: regression in tgamma?

[Paul Zimmermann]

       Hi Jeff,

> From: Jeff Johnston <jjohnstn@redhat.com>
> Date: Wed, 16 Dec 2020 15:31:47 -0500
> 
> Hi Paul,
> 
> I believe I found the problem.
> 
> The __ieee754_tgamma functions are not declared in fdlibm.h so the calls in
> tgamma and tgammaf are using
> default parms and getting things wrong.  My test calling __ieee754_tgamma
> directly only worked after I added
> a declaration to __ieee754_tgamma locally.
> 
> I have pushed a change to master.  Please try it out.
> 
> -- Jeff J.

with revision 865cd30 I now get:

zimmerma@tomate:/tmp/newlib-cygwin$ ./a.out 
x=-0x0p+0 y=inf
x=-0x1.53f198fe3b278p+7 y=0x1.fd6d312572d9cp-1015

The double value is ok, however the float result should be -inf.
This should have been fixed by Keith Packard in commit 4641693.

Paul

Re: regression in tgamma?

[Keith Packard]

[-- Attachment #1: Type: text/plain, Size: 478 bytes --]

Paul Zimmermann <Paul.Zimmermann@inria.fr> writes:

> zimmerma@tomate:/tmp/newlib-cygwin$ ./a.out 
> x=-0x0p+0 y=inf
> x=-0x1.53f198fe3b278p+7 y=0x1.fd6d312572d9cp-1015
>
> The double value is ok, however the float result should be -inf.
> This should have been fixed by Keith Packard in commit 4641693.

And it remains fixed in my code here (which is in picolibc instead of
newlib)?

x=-0 y=-inf
x=-169.972 y=5.66756e-306

(using %g instead of %a)
-- 
-keith

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 832 bytes --]

Re: regression in tgamma?

[Jeff Johnston]

I am also seeing inf instead of -inf for both tgamma and tgammaf in my
testing.

-- Jeff J.


On Thu, Dec 17, 2020 at 11:13 AM Keith Packard <keithp@keithp.com> wrote:

> Paul Zimmermann <Paul.Zimmermann@inria.fr> writes:
>
> > zimmerma@tomate:/tmp/newlib-cygwin$ ./a.out
> > x=-0x0p+0 y=inf
> > x=-0x1.53f198fe3b278p+7 y=0x1.fd6d312572d9cp-1015
> >
> > The double value is ok, however the float result should be -inf.
> > This should have been fixed by Keith Packard in commit 4641693.
>
> And it remains fixed in my code here (which is in picolibc instead of
> newlib)?
>
> x=-0 y=-inf
> x=-169.972 y=5.66756e-306
>
> (using %g instead of %a)
> --
> -keith
>

Re: regression in tgamma?

[Keith Packard]

[-- Attachment #1: Type: text/plain, Size: 809 bytes --]

Jeff Johnston <jjohnstn@redhat.com> writes:

> I am also seeing inf instead of -inf for both tgamma and tgammaf in my
> testing.

Yeah, looks like there are quite a few fixes in picolibc that I haven't
had time to send over to newlib.

I've also fixed a bunch of exception problems when compiling with clang,
which is happy to evaluate expressions known at compile time, even
though that doesn't generate exceptions at runtime. (things like 'return
one/zero;')

I've also generated tests for return values, exceptions and errnos for
the math functions to make sure they match the relevant specs and also
match glibc (mostly as a way to check the tests).

Anyone is welcome to pull those fixes out of picolibc and get them into
newlib; I probably won't have time for a while.

-- 
-keith

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 832 bytes --]

Re: regression in tgamma?

[Jeff Johnston]

Thanks Keith,

I have modified er_lgamma.c, erf_lgamma.c, w_tgamma.c, wf_tgamma.c based on
code changes in picolibc.
I did not add the new ___ieee754_ sub-functions and instead modified the
existing code assuming that the function
was called with *signgam = 0.  We can always do a merge from picolibc in
the future.

With the new code, the test runs as expected.

Paul, please confirm before tomorrow.

Thanks,

-- Jeff J.

On Thu, Dec 17, 2020 at 2:17 PM Keith Packard <keithp@keithp.com> wrote:

> Jeff Johnston <jjohnstn@redhat.com> writes:
>
> > I am also seeing inf instead of -inf for both tgamma and tgammaf in my
> > testing.
>
> Yeah, looks like there are quite a few fixes in picolibc that I haven't
> had time to send over to newlib.
>
> I've also fixed a bunch of exception problems when compiling with clang,
> which is happy to evaluate expressions known at compile time, even
> though that doesn't generate exceptions at runtime. (things like 'return
> one/zero;')
>
> I've also generated tests for return values, exceptions and errnos for
> the math functions to make sure they match the relevant specs and also
> match glibc (mostly as a way to check the tests).
>
> Anyone is welcome to pull those fixes out of picolibc and get them into
> newlib; I probably won't have time for a while.
>
> --
> -keith
>

Re: regression in tgamma?

[Paul Zimmermann]

       Hi Jeff,

> I have modified er_lgamma.c, erf_lgamma.c, w_tgamma.c, wf_tgamma.c based on
> code changes in picolibc.
> I did not add the new ___ieee754_ sub-functions and instead modified the
> existing code assuming that the function
> was called with *signgam = 0.  We can always do a merge from picolibc in the
> future.
> 
> With the new code, the test runs as expected.
> 
> Paul, please confirm before tomorrow.

version b2f3d59 fails to compile on my x86_64 machine:

../../../../../newlib/libm/math/w_tgamma.c: In function ‘tgamma’:
../../../../../newlib/libm/math/w_tgamma.c:38:6: error: ‘errno’ undeclared (first use in this function)
   38 |      errno = EDOM;
      |      ^~~~~
../../../../../newlib/libm/math/w_tgamma.c:19:1: note: ‘errno’ is defined in header ‘<errno.h>’; did you forget to ‘#include <errno.h>’?
   18 | #include "fdlibm.h"
  +++ |+#include <errno.h>

../../../../../newlib/libm/math/w_tgamma.c:38:6: note: each undeclared identifier is reported only once for each function it appears in
   38 |      errno = EDOM;
      |      ^~~~~
../../../../../newlib/libm/math/w_tgamma.c:38:14: error: ‘EDOM’ undeclared (first use in this function)
   38 |      errno = EDOM;
      |              ^~~~
../../../../../newlib/libm/math/w_tgamma.c:40:14: error: ‘ERANGE’ undeclared (first use in this function)
   40 |      errno = ERANGE;
      |              ^~~~~~

It does after the following changes:

diff --git a/newlib/libm/math/w_tgamma.c b/newlib/libm/math/w_tgamma.c
index 0f90dd4c6..52d5d71d3 100644
--- a/newlib/libm/math/w_tgamma.c
+++ b/newlib/libm/math/w_tgamma.c
@@ -16,6 +16,7 @@
  */
 
 #include "fdlibm.h"
+#include <errno.h>
 
 #ifndef _DOUBLE_IS_32BITS
 
diff --git a/newlib/libm/math/wf_tgamma.c b/newlib/libm/math/wf_tgamma.c
index 80aacf757..504cb4c29 100644
--- a/newlib/libm/math/wf_tgamma.c
+++ b/newlib/libm/math/wf_tgamma.c
@@ -15,6 +15,7 @@
 
 #include "math.h"
 #include "fdlibm.h"
+#include <errno.h>
 
 #ifdef __STDC__
        float tgammaf(float x)
@@ -34,7 +35,6 @@
            errno = EDOM;
          else if (finite(x))
            errno = ERANGE;
-       }
        return y;
 #endif
 }

and my small test program gives:

zimmerma@tomate:/tmp/newlib-cygwin$ ./a.out 
x=-0x0p+0 y=-inf
x=-0x1.53f198fe3b278p+7 y=0x1.fd6d312572d9cp-1015

which is ok. I'll do some further tests.

Paul

Re: regression in tgamma?

[Corinna Vinschen]

Paul,

could you please send patches in `git format-patch' format?
Even with `patch -p1', the below patch doesn't apply cleanly.


Thanks,
Corinna

On Dec 18 07:36, Paul Zimmermann wrote:
>        Hi Jeff,
> 
> > I have modified er_lgamma.c, erf_lgamma.c, w_tgamma.c, wf_tgamma.c based on
> > code changes in picolibc.
> > I did not add the new ___ieee754_ sub-functions and instead modified the
> > existing code assuming that the function
> > was called with *signgam = 0.  We can always do a merge from picolibc in the
> > future.
> > 
> > With the new code, the test runs as expected.
> > 
> > Paul, please confirm before tomorrow.
> 
> version b2f3d59 fails to compile on my x86_64 machine:
> 
> ../../../../../newlib/libm/math/w_tgamma.c: In function ‘tgamma’:
> ../../../../../newlib/libm/math/w_tgamma.c:38:6: error: ‘errno’ undeclared (first use in this function)
>    38 |      errno = EDOM;
>       |      ^~~~~
> ../../../../../newlib/libm/math/w_tgamma.c:19:1: note: ‘errno’ is defined in header ‘<errno.h>’; did you forget to ‘#include <errno.h>’?
>    18 | #include "fdlibm.h"
>   +++ |+#include <errno.h>
> 
> ../../../../../newlib/libm/math/w_tgamma.c:38:6: note: each undeclared identifier is reported only once for each function it appears in
>    38 |      errno = EDOM;
>       |      ^~~~~
> ../../../../../newlib/libm/math/w_tgamma.c:38:14: error: ‘EDOM’ undeclared (first use in this function)
>    38 |      errno = EDOM;
>       |              ^~~~
> ../../../../../newlib/libm/math/w_tgamma.c:40:14: error: ‘ERANGE’ undeclared (first use in this function)
>    40 |      errno = ERANGE;
>       |              ^~~~~~
> 
> It does after the following changes:
> 
> diff --git a/newlib/libm/math/w_tgamma.c b/newlib/libm/math/w_tgamma.c
> index 0f90dd4c6..52d5d71d3 100644
> --- a/newlib/libm/math/w_tgamma.c
> +++ b/newlib/libm/math/w_tgamma.c
> @@ -16,6 +16,7 @@
>   */
>  
>  #include "fdlibm.h"
> +#include <errno.h>
>  
>  #ifndef _DOUBLE_IS_32BITS
>  
> diff --git a/newlib/libm/math/wf_tgamma.c b/newlib/libm/math/wf_tgamma.c
> index 80aacf757..504cb4c29 100644
> --- a/newlib/libm/math/wf_tgamma.c
> +++ b/newlib/libm/math/wf_tgamma.c
> @@ -15,6 +15,7 @@
>  
>  #include "math.h"
>  #include "fdlibm.h"
> +#include <errno.h>
>  
>  #ifdef __STDC__
>         float tgammaf(float x)
> @@ -34,7 +35,6 @@
>             errno = EDOM;
>           else if (finite(x))
>             errno = ERANGE;
> -       }
>         return y;
>  #endif
>  }
> 
> and my small test program gives:
> 
> zimmerma@tomate:/tmp/newlib-cygwin$ ./a.out 
> x=-0x0p+0 y=-inf
> x=-0x1.53f198fe3b278p+7 y=0x1.fd6d312572d9cp-1015
> 
> which is ok. I'll do some further tests.
> 
> Paul

Re: regression in tgamma?

[Paul Zimmermann]

       Dear Corinna,

sorry, here is the patch generated by "git format-patch".

Paul

From bb94a62d68012fc38a0b6b14cedb806ff86b5c29 Mon Sep 17 00:00:00 2001
From: Paul Zimmermann <Paul.Zimmermann@inria.fr>
Date: Fri, 18 Dec 2020 09:52:40 +0100
Subject: [PATCH] fixes to make compilation succeeds

---
 newlib/libm/math/w_tgamma.c  | 1 +
 newlib/libm/math/wf_tgamma.c | 2 +-
 2 files changed, 2 insertions(+), 1 deletion(-)

diff --git a/newlib/libm/math/w_tgamma.c b/newlib/libm/math/w_tgamma.c
index 0f90dd4c6..52d5d71d3 100644
--- a/newlib/libm/math/w_tgamma.c
+++ b/newlib/libm/math/w_tgamma.c
@@ -16,6 +16,7 @@
  */
 
 #include "fdlibm.h"
+#include <errno.h>
 
 #ifndef _DOUBLE_IS_32BITS
 
diff --git a/newlib/libm/math/wf_tgamma.c b/newlib/libm/math/wf_tgamma.c
index 80aacf757..504cb4c29 100644
--- a/newlib/libm/math/wf_tgamma.c
+++ b/newlib/libm/math/wf_tgamma.c
@@ -15,6 +15,7 @@
 
 #include "math.h"
 #include "fdlibm.h"
+#include <errno.h>
 
 #ifdef __STDC__
 	float tgammaf(float x)
@@ -34,7 +35,6 @@
 	    errno = EDOM;
 	  else if (finite(x))
 	    errno = ERANGE;
-	}
 	return y;
 #endif
 }
-- 
2.29.2

Re: regression in tgamma?

[Corinna Vinschen]

On Dec 18 09:54, Paul Zimmermann wrote:
>        Dear Corinna,
> 
> sorry, here is the patch generated by "git format-patch".

Thanks, I pushed the patch.


Corinna


> 
> Paul
> 
> >From bb94a62d68012fc38a0b6b14cedb806ff86b5c29 Mon Sep 17 00:00:00 2001
> From: Paul Zimmermann <Paul.Zimmermann@inria.fr>
> Date: Fri, 18 Dec 2020 09:52:40 +0100
> Subject: [PATCH] fixes to make compilation succeeds
> 
> ---
>  newlib/libm/math/w_tgamma.c  | 1 +
>  newlib/libm/math/wf_tgamma.c | 2 +-
>  2 files changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/newlib/libm/math/w_tgamma.c b/newlib/libm/math/w_tgamma.c
> index 0f90dd4c6..52d5d71d3 100644
> --- a/newlib/libm/math/w_tgamma.c
> +++ b/newlib/libm/math/w_tgamma.c
> @@ -16,6 +16,7 @@
>   */
>  
>  #include "fdlibm.h"
> +#include <errno.h>
>  
>  #ifndef _DOUBLE_IS_32BITS
>  
> diff --git a/newlib/libm/math/wf_tgamma.c b/newlib/libm/math/wf_tgamma.c
> index 80aacf757..504cb4c29 100644
> --- a/newlib/libm/math/wf_tgamma.c
> +++ b/newlib/libm/math/wf_tgamma.c
> @@ -15,6 +15,7 @@
>  
>  #include "math.h"
>  #include "fdlibm.h"
> +#include <errno.h>
>  
>  #ifdef __STDC__
>  	float tgammaf(float x)
> @@ -34,7 +35,6 @@
>  	    errno = EDOM;
>  	  else if (finite(x))
>  	    errno = ERANGE;
> -	}
>  	return y;
>  #endif
>  }
> -- 
> 2.29.2

Re: regression in tgamma?

[Jeff Johnston]

Thanks Paul, Corinna.  I was testing locally in my directory and had a
work-around for errno already in place.

I will make the snapshot later this afternoon.

-- Jeff J.

On Fri, Dec 18, 2020 at 4:07 AM Corinna Vinschen via Newlib <
newlib@sourceware.org> wrote:

> On Dec 18 09:54, Paul Zimmermann wrote:
> >        Dear Corinna,
> >
> > sorry, here is the patch generated by "git format-patch".
>
> Thanks, I pushed the patch.
>
>
> Corinna
>
>
> >
> > Paul
> >
> > >From bb94a62d68012fc38a0b6b14cedb806ff86b5c29 Mon Sep 17 00:00:00 2001
> > From: Paul Zimmermann <Paul.Zimmermann@inria.fr>
> > Date: Fri, 18 Dec 2020 09:52:40 +0100
> > Subject: [PATCH] fixes to make compilation succeeds
> >
> > ---
> >  newlib/libm/math/w_tgamma.c  | 1 +
> >  newlib/libm/math/wf_tgamma.c | 2 +-
> >  2 files changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/newlib/libm/math/w_tgamma.c b/newlib/libm/math/w_tgamma.c
> > index 0f90dd4c6..52d5d71d3 100644
> > --- a/newlib/libm/math/w_tgamma.c
> > +++ b/newlib/libm/math/w_tgamma.c
> > @@ -16,6 +16,7 @@
> >   */
> >
> >  #include "fdlibm.h"
> > +#include <errno.h>
> >
> >  #ifndef _DOUBLE_IS_32BITS
> >
> > diff --git a/newlib/libm/math/wf_tgamma.c b/newlib/libm/math/wf_tgamma.c
> > index 80aacf757..504cb4c29 100644
> > --- a/newlib/libm/math/wf_tgamma.c
> > +++ b/newlib/libm/math/wf_tgamma.c
> > @@ -15,6 +15,7 @@
> >
> >  #include "math.h"
> >  #include "fdlibm.h"
> > +#include <errno.h>
> >
> >  #ifdef __STDC__
> >       float tgammaf(float x)
> > @@ -34,7 +35,6 @@
> >           errno = EDOM;
> >         else if (finite(x))
> >           errno = ERANGE;
> > -     }
> >       return y;
> >  #endif
> >  }
> > --
> > 2.29.2
>
>

Re: regression in tgamma?

[Paul Zimmermann]

       Hi again,

I have checked both tgammaf and tgamma with commit b2f3d59 (on x86_64).

tgammaf (exhaustive test of all the 2^32 inputs) no longer has the issue
of tgammaf(-0) = +inf instead of -inf, and similarly for -2^-128 <= x < 0.
The larger ulp error is 239 ulps as in 3.3.0 (for x=-0x1.01743cp+5).

For tgamma, by random testing I get the same maximal error as before:
2244.45 ulps for x=-0x1.53f198fe3b278p+7.

Thus the regression we had in 4.0.0 has gone.

Thanks Jeff,
Paul