| Message ID | 1482258824-5741-1-git-send-email-tuliom@linux.vnet.ibm.com |
|---|---|
| State | New |
| Headers | show |
On Tue, 20 Dec 2016, Tulio Magno Quites Machado Filho wrote: > Provides a POWER8-specific implementation for GET_FLOAT_WORD and > SET_FLOAT_WORD that is able to extract or set a float using only VSR and > GPR. Why doesn't the compiler handle this automatically when building for POWER8? My comments in <https://sourceware.org/ml/libc-alpha/2016-06/msg01180.html> apply equally here (but reinterpretation of a bit pattern between integer and float is more generic than bitwise masking of a float, so a wider range of code is likely to benefit from such a compiler optimization).
On 20/12/2016 16:46, Joseph Myers wrote: > On Tue, 20 Dec 2016, Tulio Magno Quites Machado Filho wrote: > >> Provides a POWER8-specific implementation for GET_FLOAT_WORD and >> SET_FLOAT_WORD that is able to extract or set a float using only VSR and >> GPR. > > Why doesn't the compiler handle this automatically when building for > POWER8? My comments in > <https://sourceware.org/ml/libc-alpha/2016-06/msg01180.html> apply equally > here (but reinterpretation of a bit pattern between integer and float is > more generic than bitwise masking of a float, so a wider range of code is > likely to benefit from such a compiler optimization). > Looks like GCC, at least master (7.0.0), does generate awful code for this conversion. It does use a xscvdpuxws+mfvsrd, but still store and load the float memory as pre-ISA 2.07 expectation plus a lot of superfluous precision instructions. Double precision to 64 bits integers is somewhat better (mtvsrd plus 2 fcfidu), but the contrary (64 bits integer to double precision) is also far from optimal. And I agree with Joseph here, these optimization should be handled by the compiler.
diff --git a/sysdeps/powerpc/powerpc64/power8/fpu/math_private.h b/sysdeps/powerpc/powerpc64/power8/fpu/math_private.h new file mode 100644 index 0000000..74fa4c6 --- /dev/null +++ b/sysdeps/powerpc/powerpc64/power8/fpu/math_private.h @@ -0,0 +1,48 @@ +/* Private inline math functions for POWER8. + Copyright (C) 2016 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 + <http://www.gnu.org/licenses/>. */ + +/* Move a float to a GPR using VSX instructions available in the POWER + ISA 2.07. */ +#define GET_FLOAT_WORD(i,d) \ + do { \ + float tmpd = d; \ + double tmp; \ + long tmpi; \ + __asm__ ("xscvdpspn %x1, %x2\n\t" \ + "mfvsrd %0, %x1\n\t" \ + : "=wr" (tmpi), \ + "=wa" (tmp) \ + : "wa" (tmpd) ); \ + i = tmpi >> 32; \ + } while(0) + +/* Move a float from GPR using VSX instructions available in the POWER + ISA 2.07. */ +#define SET_FLOAT_WORD(d,i) \ + do { \ + long tmpi = i; \ + float tmpd; \ + tmpi = tmpi << 32; \ + __asm__ ("mtvsrd %x0, %1\n\t" \ + "xscvspdpn %x0, %x0\n\t" \ + : "=wa" (tmpd) \ + : "wr" (tmpi) ); \ + d = tmpd; \ + } while(0) + +#include_next <math_private.h>