AArch64: Add support for roundeven[f]

Message ID VE1PR08MB5599F0396F6F6A483CDEB2CB833C9@VE1PR08MB5599.eurprd08.prod.outlook.com
State New
Headers show
Series
  • AArch64: Add support for roundeven[f]
Related show

Commit Message

Szabolcs Nagy via Libc-alpha June 3, 2021, 2:51 p.m.
Add inline assembler versions of the roundeven functions.

Passes GLIBC regress. OK for commit?

---

Comments

Szabolcs Nagy via Libc-alpha June 8, 2021, 8:49 a.m. | #1
The 06/03/2021 14:51, Wilco Dijkstra via Libc-alpha wrote:
> 

> Add inline assembler versions of the roundeven functions.

> 

> Passes GLIBC regress. OK for commit?


i checked if we can do this via builtins, but

- __builtin_roundeven is only available since gcc-10
- gcc as of now is not able to inline it as single instruction.

i think you should open a gcc bug about that (inlining
is more useful than optimizing the code in libm.so)

then reference the gcc bug in the commit.

with that change it is OK for master, thanks.

> 

> ---

> 

> diff --git a/sysdeps/aarch64/fpu/s_roundeven.c b/sysdeps/aarch64/fpu/s_roundeven.c

> new file mode 100644

> index 0000000000000000000000000000000000000000..d74b40daf5d02257f059c73eb4929e5c3824595c

> --- /dev/null

> +++ b/sysdeps/aarch64/fpu/s_roundeven.c

> @@ -0,0 +1,29 @@

> +/* Copyright (C) 2021 Free Software Foundation, Inc.

> +

> +   This file is part of the GNU C Library.

> +

> +   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.

> +

> +   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 <math.h>

> +#include <libm-alias-double.h>

> +

> +double

> +__roundeven (double x)

> +{

> +  asm volatile ("frintn \t%d0, %d1" : "=w" (x) : "w" (x));

> +  return x;

> +}

> +hidden_def (__roundeven)

> +libm_alias_double (__roundeven, roundeven)

> diff --git a/sysdeps/aarch64/fpu/s_roundevenf.c b/sysdeps/aarch64/fpu/s_roundevenf.c

> new file mode 100644

> index 0000000000000000000000000000000000000000..dfc492c2f8855d491896ace1669a69a4cab98492

> --- /dev/null

> +++ b/sysdeps/aarch64/fpu/s_roundevenf.c

> @@ -0,0 +1,28 @@

> +/* Copyright (C) 2021 Free Software Foundation, Inc.

> +

> +   This file is part of the GNU C Library.

> +

> +   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.

> +

> +   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 <math.h>

> +#include <libm-alias-float.h>

> +

> +float

> +__roundevenf (float x)

> +{

> +  asm volatile ("frintn \t%s0, %s1" : "=w" (x) : "w" (x));

> +  return x;

> +}

> +libm_alias_float (__roundeven, roundeven)

> 


--
Szabolcs Nagy via Libc-alpha June 8, 2021, 12:40 p.m. | #2
Hi Szabolcs,

> i checked if we can do this via builtins, but
>
> - __builtin_roundeven is only available since gcc-10
> - gcc as of now is not able to inline it as single instruction.

I've created https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100966,
but even if it did, we can't use the builtin until the minimum GCC that
GLIBC supports inlines the builtin. So that may be quite a few years away...

Cheers,
Wilco
Szabolcs Nagy via Libc-alpha June 8, 2021, 5:11 p.m. | #3
On 08/06/2021 09:40, Wilco Dijkstra via Libc-alpha wrote:
> Hi Szabolcs,

> 

>> i checked if we can do this via builtins, but

>>

>> - __builtin_roundeven is only available since gcc-10

>> - gcc as of now is not able to inline it as single instruction.

> 

> I've created https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100966,

> but even if it did, we can't use the builtin until the minimum GCC that

> GLIBC supports inlines the builtin. So that may be quite a few years away...


I was thinking more in laying out the required internal code, by
adding the math-use-builtin-roundeven.h and the required code on
the generic implementation.  However, due current compiler support
I am not sure it is really worth.
Joseph Myers June 8, 2021, 8:33 p.m. | #4
On Tue, 8 Jun 2021, Szabolcs Nagy via Libc-alpha wrote:

> The 06/03/2021 14:51, Wilco Dijkstra via Libc-alpha wrote:

> > 

> > Add inline assembler versions of the roundeven functions.

> > 

> > Passes GLIBC regress. OK for commit?

> 

> i checked if we can do this via builtins, but

> 

> - __builtin_roundeven is only available since gcc-10

> - gcc as of now is not able to inline it as single instruction.


It can inline it on x86_64 (with -msse4.1 or later), but it only has the 
instruction patterns for x86 and not other architectures at present.  
Effectively this is one bug per architecture with relevant instructions 
that aren't used at present.

-- 
Joseph S. Myers
joseph@codesourcery.com

Patch

diff --git a/sysdeps/aarch64/fpu/s_roundeven.c b/sysdeps/aarch64/fpu/s_roundeven.c
new file mode 100644
index 0000000000000000000000000000000000000000..d74b40daf5d02257f059c73eb4929e5c3824595c
--- /dev/null
+++ b/sysdeps/aarch64/fpu/s_roundeven.c
@@ -0,0 +1,29 @@ 
+/* Copyright (C) 2021 Free Software Foundation, Inc.
+
+   This file is part of the GNU C Library.
+
+   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.
+
+   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 <math.h>
+#include <libm-alias-double.h>
+
+double
+__roundeven (double x)
+{
+  asm volatile ("frintn \t%d0, %d1" : "=w" (x) : "w" (x));
+  return x;
+}
+hidden_def (__roundeven)
+libm_alias_double (__roundeven, roundeven)
diff --git a/sysdeps/aarch64/fpu/s_roundevenf.c b/sysdeps/aarch64/fpu/s_roundevenf.c
new file mode 100644
index 0000000000000000000000000000000000000000..dfc492c2f8855d491896ace1669a69a4cab98492
--- /dev/null
+++ b/sysdeps/aarch64/fpu/s_roundevenf.c
@@ -0,0 +1,28 @@ 
+/* Copyright (C) 2021 Free Software Foundation, Inc.
+
+   This file is part of the GNU C Library.
+
+   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.
+
+   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 <math.h>
+#include <libm-alias-float.h>
+
+float
+__roundevenf (float x)
+{
+  asm volatile ("frintn \t%s0, %s1" : "=w" (x) : "w" (x));
+  return x;
+}
+libm_alias_float (__roundeven, roundeven)