| Message ID | VE1PR08MB5599F0396F6F6A483CDEB2CB833C9@VE1PR08MB5599.eurprd08.prod.outlook.com |
|---|---|
| State | New |
| Headers | show |
| Series | AArch64: Add support for roundeven[f] | expand |
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) > --
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
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.
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.
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)