From 8b954ab9b808a51d8cf6a90c7d1d46a3366e3274 Mon Sep 17 00:00:00 2001 From: "David S. Miller" Date: Mon, 14 Jan 2013 21:47:29 -0800 Subject: Optimize sparc {ceil,floor}{,f} using vis2 'siam' instruction. * sysdeps/sparc/sparc-ifunc.h (SPARC_ASM_IFUNC2): New macro. (SPARC_ASM_VIS2_IFUNC): Likewise. (SPARC_ASM_VIS3_VIS2_IFUNC): Likewise. * sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceil-vis3.S: Make use of 'siam' instruction. * sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceilf-vis3.S: Likewise. * sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floor-vis3.S: Likewise. * sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floorf-vis3.S: Likewise. * sysdeps/sparc/sparc64/fpu/multiarch/s_ceil-vis3.S: Likewise. * sysdeps/sparc/sparc64/fpu/multiarch/s_ceilf-vis3.S: Likewise. * sysdeps/sparc/sparc64/fpu/multiarch/s_floor-vis3.S: Likewise. * sysdeps/sparc/sparc64/fpu/multiarch/s_floorf-vis3.S: Likewise. * sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceil-vis2.S: New file. * sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceilf-vis2.S: New file. * sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floor-vis2.S: New file. * sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floorf-vis2.S: New file. * sysdeps/sparc/sparc64/fpu/multiarch/s_ceil-vis2.S: New file. * sysdeps/sparc/sparc64/fpu/multiarch/s_ceilf-vis2.S: New file. * sysdeps/sparc/sparc64/fpu/multiarch/s_floor-vis2.S: New file. * sysdeps/sparc/sparc64/fpu/multiarch/s_floorf-vis2.S: New file. * sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceil.S: Hook in new VIS2 routines. * sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceilf.S: Likewise. * sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floor.S: Likewise. * sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floorf.S: Likewise. * sysdeps/sparc/sparc64/fpu/multiarch/s_ceil.S: Likewise. * sysdeps/sparc/sparc64/fpu/multiarch/s_ceilf.S: Likewise. * sysdeps/sparc/sparc64/fpu/multiarch/s_floor.S: Likewise. * sysdeps/sparc/sparc64/fpu/multiarch/s_floorf.S: Likewise. * sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/Makefile: Add new VIS2 routines to libm-sysdep_routines. * sysdeps/sparc/sparc64/fpu/multiarch/Makefile: Likewise. --- .../sparc/sparc32/sparcv9/fpu/multiarch/Makefile | 2 + .../sparc32/sparcv9/fpu/multiarch/s_ceil-vis2.S | 61 ++++++++++++++++++++++ .../sparc32/sparcv9/fpu/multiarch/s_ceil-vis3.S | 40 +++++--------- .../sparc/sparc32/sparcv9/fpu/multiarch/s_ceil.S | 2 +- .../sparc32/sparcv9/fpu/multiarch/s_ceilf-vis2.S | 58 ++++++++++++++++++++ .../sparc32/sparcv9/fpu/multiarch/s_ceilf-vis3.S | 37 ++++--------- .../sparc/sparc32/sparcv9/fpu/multiarch/s_ceilf.S | 2 +- .../sparc32/sparcv9/fpu/multiarch/s_floor-vis2.S | 61 ++++++++++++++++++++++ .../sparc32/sparcv9/fpu/multiarch/s_floor-vis3.S | 40 +++++--------- .../sparc/sparc32/sparcv9/fpu/multiarch/s_floor.S | 2 +- .../sparc32/sparcv9/fpu/multiarch/s_floorf-vis2.S | 58 ++++++++++++++++++++ .../sparc32/sparcv9/fpu/multiarch/s_floorf-vis3.S | 37 ++++--------- .../sparc/sparc32/sparcv9/fpu/multiarch/s_floorf.S | 2 +- 13 files changed, 290 insertions(+), 112 deletions(-) create mode 100644 sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceil-vis2.S create mode 100644 sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceilf-vis2.S create mode 100644 sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floor-vis2.S create mode 100644 sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floorf-vis2.S (limited to 'sysdeps/sparc/sparc32') diff --git a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/Makefile b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/Makefile index 6349681..561b0ee 100644 --- a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/Makefile +++ b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/Makefile @@ -1,4 +1,6 @@ ifeq ($(subdir),math) +libm-sysdep_routines += s_ceil-vis2 s_ceilf-vis2 \ + s_floor-vis2 s_floorf-vis2 ifeq ($(have-as-vis3),yes) libm-sysdep_routines += m_copysignf-vis3 m_copysign-vis3 s_ceilf-vis3 \ s_ceil-vis3 s_fabs-vis3 s_fabsf-vis3 s_floor-vis3 \ diff --git a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceil-vis2.S b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceil-vis2.S new file mode 100644 index 0000000..9438800 --- /dev/null +++ b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceil-vis2.S @@ -0,0 +1,61 @@ +/* ceil function, sparc32 v9 vis2 version. + Copyright (C) 2013 Free Software Foundation, Inc. + This file is part of the GNU C Library. + Contributed by David S. Miller , 2013. + + 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 + . */ + +#include + + /* 'siam' (Set Interval Arithmetic Mode) is used to quickly override + the rounding mode during this routine. + + We add then subtract (or subtract than add if the initial + value was negative) 2**23 to the value, then subtract it + back out. + + This will clear out the fractional portion of the value and, + with suitable 'siam' initiated rouding mode settings, round + the final result in the proper direction. */ + +#define TWO_FIFTYTWO 0x43300000 /* 2**52 */ + +#define ZERO %f10 /* 0.0 */ +#define SIGN_BIT %f12 /* -0.0 */ + +ENTRY (__ceil_vis2) + sethi %hi(TWO_FIFTYTWO), %o2 + sllx %o0, 32, %o0 + or %o0, %o1, %o0 + stx %o0, [%sp + 72] + sllx %o2, 32, %o2 + fzero ZERO + ldd [%sp + 72], %f0 + fnegd ZERO, SIGN_BIT + stx %o2, [%sp + 72] + fabsd %f0, %f14 + ldd [%sp + 72], %f16 + fcmpd %fcc3, %f14, %f16 + fmovduge %fcc3, ZERO, %f16 + fand %f0, SIGN_BIT, SIGN_BIT + for %f16, SIGN_BIT, %f16 + siam (1 << 2) | 2 + faddd %f0, %f16, %f18 + siam (1 << 2) | 0 + fsubd %f18, %f16, %f18 + siam (0 << 2) + retl + for %f18, SIGN_BIT, %f0 +END (__ceil_vis2) diff --git a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceil-vis3.S b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceil-vis3.S index 0c2140d..aebff5c 100644 --- a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceil-vis3.S +++ b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceil-vis3.S @@ -19,27 +19,21 @@ #include - /* Since changing the rounding mode is extremely expensive, we - try to round up using a method that is rounding mode - agnostic. + /* 'siam' (Set Interval Arithmetic Mode) is used to quickly override + the rounding mode during this routine. We add then subtract (or subtract than add if the initial value was negative) 2**23 to the value, then subtract it back out. - This will clear out the fractional portion of the value. - One of two things will happen for non-whole initial values. - Either the rounding mode will round it up, or it will be - rounded down. If the value started out whole, it will be - equal after the addition and subtraction. This means we - can accurately detect with one test whether we need to add - another 1.0 to round it up properly. + This will clear out the fractional portion of the value and, + with suitable 'siam' initiated rouding mode settings, round + the final result in the proper direction. - VIS instructions are used to facilitate the formation of - easier constants, and the propagation of the sign bit. */ + We also use VIS3 moves to avoid using the stack to transfer + values between float and integer registers. */ #define TWO_FIFTYTWO 0x43300000 /* 2**52 */ -#define ONE_DOT_ZERO 0x3ff00000 /* 1.0 */ #define ZERO %f10 /* 0.0 */ #define SIGN_BIT %f12 /* -0.0 */ @@ -47,32 +41,22 @@ ENTRY (__ceil_vis3) sethi %hi(TWO_FIFTYTWO), %o2 sllx %o0, 32, %o0 - sethi %hi(ONE_DOT_ZERO), %o3 + sllx %o2, 32, %o2 or %o0, %o1, %o0 movxtod %o0, %f0 - sllx %o2, 32, %o2 fzero ZERO - sllx %o3, 32, %o3 - fnegd ZERO, SIGN_BIT - movxtod %o2, %f16 fabsd %f0, %f14 - fcmpd %fcc3, %f14, %f16 - fmovduge %fcc3, ZERO, %f16 fand %f0, SIGN_BIT, SIGN_BIT - for %f16, SIGN_BIT, %f16 + siam (1 << 2) | 2 faddd %f0, %f16, %f18 + siam (1 << 2) | 0 fsubd %f18, %f16, %f18 - fcmpd %fcc2, %f18, %f0 - movxtod %o3, %f20 - - fmovduge %fcc2, ZERO, %f20 - faddd %f18, %f20, %f0 - fabsd %f0, %f0 + siam (0 << 2) retl - for %f0, SIGN_BIT, %f0 + for %f18, SIGN_BIT, %f0 END (__ceil_vis3) diff --git a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceil.S b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceil.S index 835703f..efc8d49 100644 --- a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceil.S +++ b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceil.S @@ -1,7 +1,7 @@ #include #include -SPARC_ASM_VIS3_IFUNC(ceil) +SPARC_ASM_VIS3_VIS2_IFUNC(ceil) weak_alias (__ceil, ceil) diff --git a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceilf-vis2.S b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceilf-vis2.S new file mode 100644 index 0000000..bc51676 --- /dev/null +++ b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceilf-vis2.S @@ -0,0 +1,58 @@ +/* Float ceil function, sparc32 v9 vis2 version. + Copyright (C) 2013 Free Software Foundation, Inc. + This file is part of the GNU C Library. + Contributed by David S. Miller , 2013. + + 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 + . */ + +#include + + /* 'siam' (Set Interval Arithmetic Mode) is used to quickly override + the rounding mode during this routine. + + We add then subtract (or subtract than add if the initial + value was negative) 2**23 to the value, then subtract it + back out. + + This will clear out the fractional portion of the value and, + with suitable 'siam' initiated rouding mode settings, round + the final result in the proper direction. */ + +#define TWO_TWENTYTHREE 0x4b000000 /* 2**23 */ + +#define ZERO %f10 /* 0.0 */ +#define SIGN_BIT %f12 /* -0.0 */ + +ENTRY (__ceilf_vis2) + st %o0, [%sp + 68] + sethi %hi(TWO_TWENTYTHREE), %o2 + fzeros ZERO + ld [%sp + 68], %f0 + fnegs ZERO, SIGN_BIT + st %o2, [%sp + 68] + fabss %f0, %f14 + ld [%sp + 68], %f16 + fcmps %fcc3, %f14, %f16 + fmovsuge %fcc3, ZERO, %f16 + fands %f0, SIGN_BIT, SIGN_BIT + fors %f16, SIGN_BIT, %f16 + siam (1 << 2) | 2 + fadds %f0, %f16, %f1 + siam (1 << 2) | 0 + fsubs %f1, %f16, %f1 + siam (0 << 2) + retl + fors %f1, SIGN_BIT, %f0 +END (__ceilf_vis2) diff --git a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceilf-vis3.S b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceilf-vis3.S index 7d30c0b..0a6768c 100644 --- a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceilf-vis3.S +++ b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceilf-vis3.S @@ -19,27 +19,21 @@ #include - /* Since changing the rounding mode is extremely expensive, we - try to round up using a method that is rounding mode - agnostic. + /* 'siam' (Set Interval Arithmetic Mode) is used to quickly override + the rounding mode during this routine. We add then subtract (or subtract than add if the initial value was negative) 2**23 to the value, then subtract it back out. - This will clear out the fractional portion of the value. - One of two things will happen for non-whole initial values. - Either the rounding mode will round it up, or it will be - rounded down. If the value started out whole, it will be - equal after the addition and subtraction. This means we - can accurately detect with one test whether we need to add - another 1.0 to round it up properly. + This will clear out the fractional portion of the value and, + with suitable 'siam' initiated rouding mode settings, round + the final result in the proper direction. - VIS instructions are used to facilitate the formation of - easier constants, and the propagation of the sign bit. */ + We also use VIS3 moves to avoid using the stack to transfer + values between float and integer registers. */ #define TWO_TWENTYTHREE 0x4b000000 /* 2**23 */ -#define ONE_DOT_ZERO 0x3f800000 /* 1.0 */ #define ZERO %f10 /* 0.0 */ #define SIGN_BIT %f12 /* -0.0 */ @@ -47,28 +41,19 @@ ENTRY (__ceilf_vis3) movwtos %o0, %f0 sethi %hi(TWO_TWENTYTHREE), %o2 - sethi %hi(ONE_DOT_ZERO), %o3 fzeros ZERO - fnegs ZERO, SIGN_BIT - movwtos %o2, %f16 fabss %f0, %f14 - fcmps %fcc3, %f14, %f16 - fmovsuge %fcc3, ZERO, %f16 fands %f0, SIGN_BIT, SIGN_BIT - fors %f16, SIGN_BIT, %f16 + siam (1 << 2) | 2 fadds %f0, %f16, %f1 + siam (1 << 2) | 0 fsubs %f1, %f16, %f1 - fcmps %fcc2, %f1, %f0 - movwtos %o3, %f9 - - fmovsuge %fcc2, ZERO, %f9 - fadds %f1, %f9, %f0 - fabss %f0, %f0 + siam (0 << 2) retl - fors %f0, SIGN_BIT, %f0 + fors %f1, SIGN_BIT, %f0 END (__ceilf_vis3) diff --git a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceilf.S b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceilf.S index 3047dd8..1c72a57 100644 --- a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceilf.S +++ b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_ceilf.S @@ -1,6 +1,6 @@ #include -SPARC_ASM_VIS3_IFUNC(ceilf) +SPARC_ASM_VIS3_VIS2_IFUNC(ceilf) weak_alias (__ceilf, ceilf) diff --git a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floor-vis2.S b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floor-vis2.S new file mode 100644 index 0000000..3b5e8fd --- /dev/null +++ b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floor-vis2.S @@ -0,0 +1,61 @@ +/* floor function, sparc32 v9 vis2 version. + Copyright (C) 2013 Free Software Foundation, Inc. + This file is part of the GNU C Library. + Contributed by David S. Miller , 2013. + + 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 + . */ + +#include + + /* 'siam' (Set Interval Arithmetic Mode) is used to quickly override + the rounding mode during this routine. + + We add then subtract (or subtract than add if the initial + value was negative) 2**23 to the value, then subtract it + back out. + + This will clear out the fractional portion of the value and, + with suitable 'siam' initiated rouding mode settings, round + the final result in the proper direction. */ + +#define TWO_FIFTYTWO 0x43300000 /* 2**52 */ + +#define ZERO %f10 /* 0.0 */ +#define SIGN_BIT %f12 /* -0.0 */ + +ENTRY (__floor_vis2) + sethi %hi(TWO_FIFTYTWO), %o2 + sllx %o0, 32, %o0 + or %o0, %o1, %o0 + stx %o0, [%sp + 72] + sllx %o2, 32, %o2 + fzero ZERO + ldd [%sp + 72], %f0 + fnegd ZERO, SIGN_BIT + stx %o2, [%sp + 72] + fabsd %f0, %f14 + ldd [%sp + 72], %f16 + fcmpd %fcc3, %f14, %f16 + fmovduge %fcc3, ZERO, %f16 + fand %f0, SIGN_BIT, SIGN_BIT + for %f16, SIGN_BIT, %f16 + siam (1 << 2) | 3 + faddd %f0, %f16, %f18 + siam (1 << 2) | 0 + fsubd %f18, %f16, %f18 + siam (0 << 2) + retl + for %f18, SIGN_BIT, %f0 +END (__floor_vis2) diff --git a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floor-vis3.S b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floor-vis3.S index 8445f1d..41fdfac 100644 --- a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floor-vis3.S +++ b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floor-vis3.S @@ -19,27 +19,21 @@ #include - /* Since changing the rounding mode is extremely expensive, we - try to round up using a method that is rounding mode - agnostic. + /* 'siam' (Set Interval Arithmetic Mode) is used to quickly override + the rounding mode during this routine. We add then subtract (or subtract than add if the initial value was negative) 2**23 to the value, then subtract it back out. - This will clear out the fractional portion of the value. - One of two things will happen for non-whole initial values. - Either the rounding mode will round it up, or it will be - rounded down. If the value started out whole, it will be - equal after the addition and subtraction. This means we - can accurately detect with one test whether we need to add - another 1.0 to round it up properly. + This will clear out the fractional portion of the value and, + with suitable 'siam' initiated rouding mode settings, round + the final result in the proper direction. - VIS instructions are used to facilitate the formation of - easier constants, and the propagation of the sign bit. */ + We also use VIS3 moves to avoid using the stack to transfer + values between float and integer registers. */ #define TWO_FIFTYTWO 0x43300000 /* 2**52 */ -#define ONE_DOT_ZERO 0x3ff00000 /* 1.0 */ #define ZERO %f10 /* 0.0 */ #define SIGN_BIT %f12 /* -0.0 */ @@ -47,32 +41,22 @@ ENTRY (__floor_vis3) sethi %hi(TWO_FIFTYTWO), %o2 sllx %o0, 32, %o0 - sethi %hi(ONE_DOT_ZERO), %o3 + sllx %o2, 32, %o2 or %o0, %o1, %o0 movxtod %o0, %f0 - sllx %o2, 32, %o2 fzero ZERO - sllx %o3, 32, %o3 - fnegd ZERO, SIGN_BIT - movxtod %o2, %f16 fabsd %f0, %f14 - fcmpd %fcc3, %f14, %f16 - fmovduge %fcc3, ZERO, %f16 fand %f0, SIGN_BIT, SIGN_BIT - for %f16, SIGN_BIT, %f16 + siam (1 << 2) | 3 faddd %f0, %f16, %f18 + siam (1 << 2) | 0 fsubd %f18, %f16, %f18 - fcmpd %fcc2, %f18, %f0 - movxtod %o3, %f20 - - fmovdule %fcc2, ZERO, %f20 - fsubd %f18, %f20, %f0 - fabsd %f0, %f0 + siam (0 << 2) retl - for %f0, SIGN_BIT, %f0 + for %f18, SIGN_BIT, %f0 END (__floor_vis3) diff --git a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floor.S b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floor.S index 37aeb43..1fe4b95 100644 --- a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floor.S +++ b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floor.S @@ -1,7 +1,7 @@ #include #include -SPARC_ASM_VIS3_IFUNC(floor) +SPARC_ASM_VIS3_VIS2_IFUNC(floor) weak_alias (__floor, floor) diff --git a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floorf-vis2.S b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floorf-vis2.S new file mode 100644 index 0000000..4f73121 --- /dev/null +++ b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floorf-vis2.S @@ -0,0 +1,58 @@ +/* Float floor function, sparc32 v9 vis2 version. + Copyright (C) 2013 Free Software Foundation, Inc. + This file is part of the GNU C Library. + Contributed by David S. Miller , 2013. + + 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 + . */ + +#include + + /* 'siam' (Set Interval Arithmetic Mode) is used to quickly override + the rounding mode during this routine. + + We add then subtract (or subtract than add if the initial + value was negative) 2**23 to the value, then subtract it + back out. + + This will clear out the fractional portion of the value and, + with suitable 'siam' initiated rouding mode settings, round + the final result in the proper direction. */ + +#define TWO_TWENTYTHREE 0x4b000000 /* 2**23 */ + +#define ZERO %f10 /* 0.0 */ +#define SIGN_BIT %f12 /* -0.0 */ + +ENTRY (__floorf_vis2) + st %o0, [%sp + 68] + sethi %hi(TWO_TWENTYTHREE), %o2 + fzeros ZERO + ld [%sp + 68], %f0 + fnegs ZERO, SIGN_BIT + st %o2, [%sp + 68] + fabss %f0, %f14 + ld [%sp + 68], %f16 + fcmps %fcc3, %f14, %f16 + fmovsuge %fcc3, ZERO, %f16 + fands %f0, SIGN_BIT, SIGN_BIT + fors %f16, SIGN_BIT, %f16 + siam (1 << 2) | 3 + fadds %f0, %f16, %f1 + siam (1 << 2) | 0 + fsubs %f1, %f16, %f1 + siam (0 << 2) + retl + fors %f1, SIGN_BIT, %f0 +END (__floorf_vis2) diff --git a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floorf-vis3.S b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floorf-vis3.S index 133a0a4..fe2d2da 100644 --- a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floorf-vis3.S +++ b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floorf-vis3.S @@ -19,27 +19,21 @@ #include - /* Since changing the rounding mode is extremely expensive, we - try to round up using a method that is rounding mode - agnostic. + /* 'siam' (Set Interval Arithmetic Mode) is used to quickly override + the rounding mode during this routine. We add then subtract (or subtract than add if the initial value was negative) 2**23 to the value, then subtract it back out. - This will clear out the fractional portion of the value. - One of two things will happen for non-whole initial values. - Either the rounding mode will round it up, or it will be - rounded down. If the value started out whole, it will be - equal after the addition and subtraction. This means we - can accurately detect with one test whether we need to add - another 1.0 to round it up properly. + This will clear out the fractional portion of the value and, + with suitable 'siam' initiated rouding mode settings, round + the final result in the proper direction. - VIS instructions are used to facilitate the formation of - easier constants, and the propagation of the sign bit. */ + We also use VIS3 moves to avoid using the stack to transfer + values between float and integer registers. */ #define TWO_TWENTYTHREE 0x4b000000 /* 2**23 */ -#define ONE_DOT_ZERO 0x3f800000 /* 1.0 */ #define ZERO %f10 /* 0.0 */ #define SIGN_BIT %f12 /* -0.0 */ @@ -47,28 +41,19 @@ ENTRY (__floorf_vis3) movwtos %o0, %f0 sethi %hi(TWO_TWENTYTHREE), %o2 - sethi %hi(ONE_DOT_ZERO), %o3 fzeros ZERO - fnegs ZERO, SIGN_BIT - movwtos %o2, %f16 fabss %f0, %f14 - fcmps %fcc3, %f14, %f16 - fmovsuge %fcc3, ZERO, %f16 fands %f0, SIGN_BIT, SIGN_BIT - fors %f16, SIGN_BIT, %f16 + siam (1 << 2) | 3 fadds %f0, %f16, %f1 + siam (1 << 2) | 0 fsubs %f1, %f16, %f1 - fcmps %fcc2, %f1, %f0 - movwtos %o3, %f9 - - fmovsule %fcc2, ZERO, %f9 - fsubs %f1, %f9, %f0 - fabss %f0, %f0 + siam (0 << 2) retl - fors %f0, SIGN_BIT, %f0 + fors %f1, SIGN_BIT, %f0 END (__floorf_vis3) diff --git a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floorf.S b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floorf.S index 31cda38..d2a83cb 100644 --- a/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floorf.S +++ b/sysdeps/sparc/sparc32/sparcv9/fpu/multiarch/s_floorf.S @@ -1,6 +1,6 @@ #include -SPARC_ASM_VIS3_IFUNC(floorf) +SPARC_ASM_VIS3_VIS2_IFUNC(floorf) weak_alias (__floorf, floorf) -- cgit v1.1