This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
[PATCH] Optimized sparc floor{,f} routines.
- From: David Miller <davem at davemloft dot net>
- To: libc-alpha at sourceware dot org
- Date: Mon, 27 Feb 2012 23:39:06 -0500 (EST)
- Subject: [PATCH] Optimized sparc floor{,f} routines.
These are straightforward transformations of the corresponding
ceil{,f} assembler routines. The final conditional move is fmovdule
instead of fmovduge, and the final addition is turned into a
subtraction.
Committed to master.
* sysdeps/sparc/sparc32/sparcv9/fpu/s_floor.S: New file.
* sysdeps/sparc/sparc32/sparcv9/fpu/s_floorf.S: New file.
* sysdeps/sparc/sparc64/fpu/s_floor.S: New file.
* sysdeps/sparc/sparc64/fpu/s_floorf.S: New file.
---
ChangeLog | 5 ++
sysdeps/sparc/sparc32/sparcv9/fpu/s_floor.S | 87 ++++++++++++++++++++++++++
sysdeps/sparc/sparc32/sparcv9/fpu/s_floorf.S | 83 ++++++++++++++++++++++++
sysdeps/sparc/sparc64/fpu/s_floor.S | 83 ++++++++++++++++++++++++
sysdeps/sparc/sparc64/fpu/s_floorf.S | 81 ++++++++++++++++++++++++
5 files changed, 339 insertions(+), 0 deletions(-)
create mode 100644 sysdeps/sparc/sparc32/sparcv9/fpu/s_floor.S
create mode 100644 sysdeps/sparc/sparc32/sparcv9/fpu/s_floorf.S
create mode 100644 sysdeps/sparc/sparc64/fpu/s_floor.S
create mode 100644 sysdeps/sparc/sparc64/fpu/s_floorf.S
diff --git a/ChangeLog b/ChangeLog
index df43a71..308a672 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,5 +1,10 @@
2012-02-27 David S. Miller <davem@davemloft.net>
+ * sysdeps/sparc/sparc32/sparcv9/fpu/s_floor.S: New file.
+ * sysdeps/sparc/sparc32/sparcv9/fpu/s_floorf.S: New file.
+ * sysdeps/sparc/sparc64/fpu/s_floor.S: New file.
+ * sysdeps/sparc/sparc64/fpu/s_floorf.S: New file.
+
* sysdeps/sparc/sparc32/sparcv9/fpu/s_ceil.S: Fix accidental use of
frame pointer instead of stack pointer relative arg slot.
* sysdeps/sparc/sparc32/sparcv9/fpu/s_ceilf.S: Likewise.
diff --git a/sysdeps/sparc/sparc32/sparcv9/fpu/s_floor.S b/sysdeps/sparc/sparc32/sparcv9/fpu/s_floor.S
new file mode 100644
index 0000000..bd94c7f
--- /dev/null
+++ b/sysdeps/sparc/sparc32/sparcv9/fpu/s_floor.S
@@ -0,0 +1,87 @@
+/* floor function, sparc32 v9 version.
+ Copyright (C) 2012 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+ Contributed by David S. Miller <davem@davemloft.net>, 2012.
+
+ 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/>. */
+
+#include <sysdep.h>
+
+ /* Since changing the rounding mode is extremely expensive, we
+ try to round up using a method that is rounding mode
+ agnostic.
+
+ 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.
+
+ We pop constants into the FPU registers using the incoming
+ argument stack slots, since this avoid having to use any
+ PIC references. We also thus avoid having to allocate a
+ register window.
+
+ VIS instructions are used to facilitate the formation of
+ easier constants, and the propagation of the sign bit. */
+
+#define TWO_FIFTYTWO 0x43300000 /* 2**52 */
+#define ONE_DOT_ZERO 0x3ff00000 /* 1.0 */
+
+#define ZERO %f10 /* 0.0 */
+#define SIGN_BIT %f12 /* -0.0 */
+
+ENTRY (__floor)
+ sethi %hi(TWO_FIFTYTWO), %o2
+ sllx %o0, 32, %o0
+ sethi %hi(ONE_DOT_ZERO), %o3
+ or %o0, %o1, %o0
+ stx %o0, [%sp + 72]
+ sllx %o2, 32, %o2
+ fzero ZERO
+ sllx %o3, 32, %o3
+
+ 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
+ faddd %f0, %f16, %f18
+ fsubd %f18, %f16, %f18
+ fcmpd %fcc2, %f18, %f0
+ stx %o3, [%sp + 72]
+
+ ldd [%sp + 72], %f20
+ fmovdule %fcc2, ZERO, %f20
+ fsubd %f18, %f20, %f0
+ fabsd %f0, %f0
+ retl
+ for %f0, SIGN_BIT, %f0
+END (__floor)
+weak_alias (__floor, floor)
diff --git a/sysdeps/sparc/sparc32/sparcv9/fpu/s_floorf.S b/sysdeps/sparc/sparc32/sparcv9/fpu/s_floorf.S
new file mode 100644
index 0000000..864644f
--- /dev/null
+++ b/sysdeps/sparc/sparc32/sparcv9/fpu/s_floorf.S
@@ -0,0 +1,83 @@
+/* Float floor function, sparc32 v9 version.
+ Copyright (C) 2012 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+ Contributed by David S. Miller <davem@davemloft.net>, 2012.
+
+ 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/>. */
+
+#include <sysdep.h>
+
+ /* Since changing the rounding mode is extremely expensive, we
+ try to round up using a method that is rounding mode
+ agnostic.
+
+ 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.
+
+ We pop constants into the FPU registers using the incoming
+ argument stack slots, since this avoid having to use any
+ PIC references. We also thus avoid having to allocate a
+ register window.
+
+ VIS instructions are used to facilitate the formation of
+ easier constants, and the propagation of the sign bit. */
+
+#define TWO_TWENTYTHREE 0x4b000000 /* 2**23 */
+#define ONE_DOT_ZERO 0x3f800000 /* 1.0 */
+
+#define ZERO %f10 /* 0.0 */
+#define SIGN_BIT %f12 /* -0.0 */
+
+ENTRY (__floorf)
+ st %o0, [%sp + 68]
+ sethi %hi(TWO_TWENTYTHREE), %o2
+ sethi %hi(ONE_DOT_ZERO), %o3
+ 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
+ fadds %f0, %f16, %f1
+ fsubs %f1, %f16, %f1
+ fcmps %fcc2, %f1, %f0
+ st %o3, [%sp + 68]
+
+ ld [%sp + 68], %f9
+ fmovsule %fcc2, ZERO, %f9
+ fsubs %f1, %f9, %f0
+ fabss %f0, %f0
+ retl
+ fors %f0, SIGN_BIT, %f0
+END (__floorf)
+weak_alias (__floorf, floorf)
diff --git a/sysdeps/sparc/sparc64/fpu/s_floor.S b/sysdeps/sparc/sparc64/fpu/s_floor.S
new file mode 100644
index 0000000..88d9c4d
--- /dev/null
+++ b/sysdeps/sparc/sparc64/fpu/s_floor.S
@@ -0,0 +1,83 @@
+/* floor function, sparc64 version.
+ Copyright (C) 2012 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+ Contributed by David S. Miller <davem@davemloft.net>, 2012.
+
+ 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/>. */
+
+#include <sysdep.h>
+
+ /* Since changing the rounding mode is extremely expensive, we
+ try to round up using a method that is rounding mode
+ agnostic.
+
+ 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.
+
+ We pop constants into the FPU registers using the incoming
+ argument stack slots, since this avoid having to use any
+ PIC references. We also thus avoid having to allocate a
+ register window.
+
+ VIS instructions are used to facilitate the formation of
+ easier constants, and the propagation of the sign bit. */
+
+#define TWO_FIFTYTWO 0x43300000 /* 2**52 */
+#define ONE_DOT_ZERO 0x3ff00000 /* 1.0 */
+
+#define ZERO %f10 /* 0.0 */
+#define SIGN_BIT %f12 /* -0.0 */
+
+ENTRY (__floor)
+ sethi %hi(TWO_FIFTYTWO), %o2
+ sethi %hi(ONE_DOT_ZERO), %o3
+ fzero ZERO
+
+ sllx %o2, 32, %o2
+ fnegd ZERO, SIGN_BIT
+
+ sllx %o3, 32, %o3
+ stx %o2, [%sp + STACK_BIAS + 128]
+ fabsd %f0, %f14
+
+ ldd [%sp + STACK_BIAS + 128], %f16
+ fcmpd %fcc3, %f14, %f16
+
+ fmovduge %fcc3, ZERO, %f16
+ fand %f0, SIGN_BIT, SIGN_BIT
+
+ for %f16, SIGN_BIT, %f16
+ faddd %f0, %f16, %f18
+ fsubd %f18, %f16, %f18
+ fcmpd %fcc2, %f18, %f0
+ stx %o3, [%sp + STACK_BIAS + 128]
+
+ ldd [%sp + STACK_BIAS + 128], %f20
+ fmovdule %fcc2, ZERO, %f20
+ fsubd %f18, %f20, %f0
+ fabsd %f0, %f0
+ retl
+ for %f0, SIGN_BIT, %f0
+END (__floor)
+weak_alias (__floor, floor)
diff --git a/sysdeps/sparc/sparc64/fpu/s_floorf.S b/sysdeps/sparc/sparc64/fpu/s_floorf.S
new file mode 100644
index 0000000..711cbe4
--- /dev/null
+++ b/sysdeps/sparc/sparc64/fpu/s_floorf.S
@@ -0,0 +1,81 @@
+/* Float floor function, sparc64 version.
+ Copyright (C) 2012 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+ Contributed by David S. Miller <davem@davemloft.net>, 2012.
+
+ 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/>. */
+
+#include <sysdep.h>
+
+ /* Since changing the rounding mode is extremely expensive, we
+ try to round up using a method that is rounding mode
+ agnostic.
+
+ 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.
+
+ We pop constants into the FPU registers using the incoming
+ argument stack slots, since this avoid having to use any
+ PIC references. We also thus avoid having to allocate a
+ register window.
+
+ VIS instructions are used to facilitate the formation of
+ easier constants, and the propagation of the sign bit. */
+
+#define TWO_TWENTYTHREE 0x4b000000 /* 2**23 */
+#define ONE_DOT_ZERO 0x3f800000 /* 1.0 */
+
+#define ZERO %f10 /* 0.0 */
+#define SIGN_BIT %f12 /* -0.0 */
+
+ENTRY (__floorf)
+ sethi %hi(TWO_TWENTYTHREE), %o2
+ sethi %hi(ONE_DOT_ZERO), %o3
+ fzeros ZERO
+
+ fnegs ZERO, SIGN_BIT
+
+ st %o2, [%sp + STACK_BIAS + 128]
+ fabss %f1, %f14
+
+ ld [%sp + STACK_BIAS + 128], %f16
+ fcmps %fcc3, %f14, %f16
+
+ fmovsuge %fcc3, ZERO, %f16
+ fands %f1, SIGN_BIT, SIGN_BIT
+
+ fors %f16, SIGN_BIT, %f16
+ fadds %f1, %f16, %f5
+ fsubs %f5, %f16, %f5
+ fcmps %fcc2, %f5, %f1
+ st %o3, [%sp + STACK_BIAS + 128]
+
+ ld [%sp + STACK_BIAS + 128], %f9
+ fmovsule %fcc2, ZERO, %f9
+ fsubs %f5, %f9, %f0
+ fabss %f0, %f0
+ retl
+ fors %f0, SIGN_BIT, %f0
+END (__floorf)
+weak_alias (__floorf, floorf)
--
1.7.6.401.g6a319