+2013-10-01 Joern Rennecke <joern.rennecke@embecosm.com>
+ Brendan Kehoe <brendan@zen.org>
+ Simon Cook <simon.cook@embecosm.com>
+
+ * config.host (arc*-*-elf*, arc*-*-linux-uclibc*): New configurations.
+ * config/arc: New directory.
+ * longlong.h [__arc__] (umul_ppmm): Remove.
+ [__arc__] (__umulsidi3): Define.
+ [__arc__ && __ARC_NORM__] (count_leading_zeroes): Define.
+ [__arc__ && __ARC_NORM__] (COUNT_LEADING_ZEROS_0): Likewise.
+
2013-09-17 Jacek Caban <jacek@codeweavers.com>
* config/i386/gthr-win32.c: CreateSemaphoreW instead of
am33_2.0-*-linux*)
cpu_type=mn10300
;;
+arc*-*-*)
+ cpu_type=arc
+ ;;
arm*-*-*)
cpu_type=arm
;;
extra_parts="$extra_parts vms-dwarf2.o vms-dwarf2eh.o"
md_unwind_header=alpha/vms-unwind.h
;;
+arc*-*-elf*)
+ tmake_file="arc/t-arc-newlib arc/t-arc"
+ extra_parts="crti.o crtn.o crtend.o crtbegin.o crtendS.o crtbeginS.o libgmon.a crtg.o crtgend.o"
+ ;;
+arc*-*-linux-uclibc*)
+ tmake_file="${tmake_file} t-slibgcc-libgcc t-slibgcc-nolc-override arc/t-arc700-uClibc arc/t-arc"
+ extra_parts="crti.o crtn.o crtend.o crtbegin.o crtendS.o crtbeginS.o libgmon.a crtg.o crtgend.o"
+ ;;
arm-wrs-vxworks)
tmake_file="$tmake_file arm/t-arm arm/t-vxworks t-softfp-sfdf t-softfp-excl arm/t-softfp t-softfp"
extra_parts="$extra_parts crti.o crtn.o"
--- /dev/null
+/* Assembler macros for the Synopsys DesignWare ARC CPU.
+
+ Copyright (C) 1994, 1995, 1997, 2004, 2007-2012
+ Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#define FUNC(X) .type X,@function
+#define ENDFUNC(X) .size X, .-X
--- /dev/null
+/* Code to start and stop profiling for the Synopsys DesignWare ARC CPU.
+
+ Copyright (C) 1994, 1995, 1997, 2004, 2007-2012
+ Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+ .section .init
+ .global _init
+ .global _fini
+ .global __monstartup
+ mov_s r0,_init
+ mov_s r1,_fini
+ jl __monstartup
+
+ .section .__arc_profile_desc, "a"
+ .global __arc_profile_desc_secstart
+ .balign 4
+__arc_profile_desc_secstart:
+ .section .__arc_profile_forward, "a"
+ .global __arc_profile_forward_secstart
+ .balign 4
+__arc_profile_forward_secstart:
+ .section .__arc_profile_counters, "aw"
+ .global __arc_profile_counters_secstart
+ .balign 4
+__arc_profile_counters_secstart:
+
+ .section .fini
+ .global _mcleanup
+ jl _mcleanup
--- /dev/null
+/* Code to start and stop profiling for the Synopsys DesignWare ARC CPU.
+
+ Copyright (C) 1994, 1995, 1997, 2004, 2007-2012
+ Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+ .section .__arc_profile_desc, "a"
+ .global __arc_profile_desc_secend
+__arc_profile_desc_secend:
+ .section .__arc_profile_forward, "a"
+ .global __arc_profile_forward_secend
+__arc_profile_forward_secend:
--- /dev/null
+/* .fini/.init stack frame setup for the Synopsys DesignWare ARC CPU.
+
+ Copyright (C) 1994, 1995, 1997, 2004, 2007-2012
+ Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+# This file contains the stack frame setup for contents of the .fini and
+# .init sections.
+
+ .section .init
+ .global _init
+ .word 0
+_init:
+ push_s blink
+
+ .section .fini
+ .global _fini
+ .word 0
+_fini:
+ push_s blink
--- /dev/null
+/* Ensure .fini/.init return for the Synopsys DesignWare ARC CPU.
+
+ Copyright (C) 1994, 1995, 1997, 2004, 2007-2012
+ Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+# This file just makes sure that the .fini and .init sections do in
+# fact return. This file is the last thing linked into any executable.
+
+ .section .init
+ pop_s blink
+ j_s [blink]
+
+
+ .section .fini
+ pop_s blink
+ j_s [blink]
--- /dev/null
+/* Copyright (C) 2004, 2006, 2007-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is free software; you can redistribute it and/or modify it
+under the terms of the GNU General Public License as published by the
+Free Software Foundation; either version 3, or (at your option) any
+later version.
+
+In addition to the permissions in the GNU General Public License, the
+Free Software Foundation gives you unlimited permission to link the
+compiled version of this file into combinations with other programs,
+and to distribute those combinations without any restriction coming
+from the use of this file. (The General Public License restrictions
+do apply in other respects; for example, they cover modification of
+the file, and distribution when not linked into a combine
+executable.)
+
+This file 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
+General Public License for more details.
+
+You should have received a copy of the GNU General Public License
+along with this program; see the file COPYING3. If not see
+<http://www.gnu.org/licenses/>. */
+
+/* Calculate division table for ARC700 integer division
+ Contributed by Joern Rennecke
+ joern.rennecke@arc.com */
+
+#include <stdio.h>
+#include <math.h>
+
+int
+main ()
+{
+ int i, j;
+ unsigned x;
+ double q, r, err, max_err = -1;
+
+ puts("/* This table has been generated by divtab-arc700.c. */");
+ puts("\
+/* 1/512 .. 1/256, normalized. There is a leading 1 in bit 31.\n\
+ For powers of two, we list unnormalized numbers instead. The values\n\
+ for powers of 2 are loaded, but not used. The value for 1 is actually\n\
+ the first instruction after .Lmuldiv. */\n\
+ .balign 4");
+ puts (".Ldivtab:\n");
+ for (i = 256; i >= 2; --i)
+ {
+ j = i < 0 ? -i : i;
+ if (j & (j-1))
+ while (j < 128)
+ j += j;
+ else
+ /* Power of two. */
+ j *= 128;
+ q = 4.*(1<<30)*128/j;
+ r = ceil (q);
+ printf ("\t.long\t0x%X\n", (unsigned) r);
+ err = r - q;
+ if (err > max_err)
+ max_err = err;
+ }
+#if 0
+ printf ("\t/* maximum error: %f */\n", max_err);
+#endif
+ exit (0);
+}
--- /dev/null
+/* Copyright (C) 2007-2013 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+/* This file selects the double-precision parts of fp-bit.c that are
+ still needed for some ARC hardware variants; it also renames functions
+ that duplicate asm-coded functionality so that their results can be
+ used to compare with the optimized versions for debugging. */
+
+#define FINE_GRAINED_LIBRARIES
+#define ARC_DP_DEBUG 1
+#if !defined (__ARC_NORM__) || ARC_DP_DEBUG
+#define L_pack_df
+#define L_unpack_df
+#define L_make_df
+#define L_thenan_df
+#define L_sf_to_df
+#endif
+#ifndef __ARC_NORM__
+#define L_addsub_df
+#elif ARC_DP_DEBUG
+#define L_addsub_df
+#define __adddf3 __adddf3_c
+#define __subdf3 __subdf3_c
+#endif
+#ifndef __ARC_NORM__
+#define L_mul_df
+#define L_div_df
+#elif (!defined (__ARC700__) && !defined (__ARC_MUL64__) \
+ && !defined(__ARC_MUL32BY16__))
+#define L_mul_df
+#define L_div_df
+#undef QUIET_NAN
+#define QUIET_NAN 0xfffffffffffffLL
+#elif ARC_DP_DEBUG
+#define L_mul_df
+#define __muldf3 __muldf3_c
+#define L_div_df
+#define __divdf3 __divdf3_c
+#endif
+#ifndef __ARC_NORM__
+#define L_df_to_sf
+#define L_si_to_df
+#define L_df_to_si
+#define L_tf_to_usi /* need to defined this instead of df_to_usi */
+#define L_usi_to_df
+#elif ARC_DP_DEBUG
+#define L_df_to_sf
+#define __truncdfsf2 __truncdfsf2_c
+#define L_si_to_df
+#define __floatsidf __floatsidf_c
+#define L_df_to_si
+#define __fixdfsi __fixdfsi_c
+#define L_tf_to_usi
+#define __fixunsdfsi __fixunsdfsi_c
+#define L_usi_to_df
+#define __floatunsidf __floatunsidf_c
+#endif
+#ifndef __ARC_NORM__
+#define L_fpcmp_parts_df
+#define L_compare_df
+#define L_eq_df
+#define L_ne_df
+#define L_gt_df
+#define L_ge_df
+#define L_lt_df
+#define L_le_df
+#define L_unord_df
+#define L_negate_df
+#elif ARC_DP_DEBUG
+#define L_fpcmp_parts_df
+#define L_eq_df
+#define __eqdf2 __eqdf2_c
+#define L_gt_df
+#define __gtdf2 __gtdf2_c
+#define L_ge_df
+#define __gedf2 __gedf2_c
+#define L_unord_df
+#define __unorddf2 __unorddf2_c
+#endif
--- /dev/null
+/* Copyright (C) 2007-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+/* This file selects the single-precision parts of fp-bit.c that are
+ still needed for some ARC hardware variants; it also renames functions
+ that duplicate asm-coded functionality so that their results can be
+ used to compare with the optimized versions for debugging. */
+
+#define ARC_FP_DEBUG 1
+#define FINE_GRAINED_LIBRARIES
+#if !defined (__ARC_NORM__) || ARC_FP_DEBUG
+#define L_pack_sf
+#define L_unpack_sf
+#define L_make_sf
+#define L_thenan_sf
+#endif
+#ifndef __ARC_NORM__
+#define L_addsub_sf
+#define L_mul_sf
+#define L_div_sf
+#define L_sf_to_df
+#define L_si_to_sf
+#define L_sf_to_si
+#define L_usi_to_sf
+#elif ARC_FP_DEBUG
+#define L_addsub_sf
+#define __addsf3 __addsf3_c
+#define __subsf3 __subsf3_c
+#define L_mul_sf
+#define __mulsf3 __mulsf3_c
+#define L_div_sf
+#define __divsf3 __divsf3_c
+#define L_sf_to_df
+#define __extendsfdf2 __extendsfdf2_c
+#define L_si_to_sf
+#define __floatsisf __floatsisf_c
+#define L_sf_to_si
+#define __fixsfsi __fixsfsi_c
+#define L_usi_to_sf
+#define __floatunsisf __floatunsisf_c
+#endif
+#ifndef __ARC_NORM__
+#define L_fpcmp_parts_sf
+#define L_compare_sf
+#define L_eq_sf
+#define L_ne_sf
+#define L_gt_sf
+#define L_ge_sf
+#define L_lt_sf
+#define L_le_sf
+#define L_unord_sf
+#define L_negate_sf
+#elif ARC_FP_DEBUG
+#define L_fpcmp_parts_sf
+#define L_eq_sf
+#define __eqsf2 __eqsf2_c
+#define L_gt_sf
+#define __gtsf2 __gtsf2_c
+#define L_ge_sf
+#define __gesf2 __gesf2_c
+#define L_unord_sf
+#define __unordsf2 __unordsf2_c
+#endif
--- /dev/null
+/* Copyright (C) 2007-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+/* File deliberately left blank. */
--- /dev/null
+/* Copyright (C) 2007-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#define LP_START 0x02
+#define LP_END 0x03
+#define IDENTITY 0x04
+#define STATUS32 0x0a
+#define COUNT0 0x21 /* Timer 0 count */
+#define CONTROL0 0x22 /* Timer 0 control */
+#define LIMIT0 0x23 /* Timer 0 limit */
+#define INT_VECTOR_BASE 0x25
+#define D_CACHE_BUILD 0x72
+#define DC_FLDL 0x4c
--- /dev/null
+/* This file contains code to do profiling.
+
+ Copyright (C) 2007-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "../asm.h"
+#include "auxreg.h"
+/* This file contains code to do profiling. */
+ .weak __profile_timer_cycles
+ .global __profile_timer_cycles
+ .set __profile_timer_cycles, 200
+ .text
+ ; For Arctangent-A5, if no data cache is present, a read of the
+ ; cache build register returns the ID register. For ARC600 and
+ ; later, the version field will be zero.
+ .global __dcache_linesz
+ .balign 4
+__dcache_linesz:
+ lr r12,[D_CACHE_BUILD]
+ extb_s r0,r12
+ breq_s r0,0,.Lsz_nocache
+ brge r0,0x20,.Lsz_havecache
+ lr r0,[IDENTITY]
+ breq r12,r0,.Lsz_nocache
+.Lsz_havecache:
+ lsr_s r12,r12,16
+ mov_s r0,16
+ bmsk_s r12,r12,3
+ asl_s r0,r0,r12
+ j_s [blink]
+.Lsz_nocache:
+ mov_s r0,1
+ j_s [blink]
--- /dev/null
+/*-
+ * Copyright (c) 1983, 1992, 1993
+ * The Regents of the University of California. All rights reserved.
+ * Copyright (C) 2007-2012 Free Software Foundation, Inc.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ * 1. Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimer in the
+ * documentation and/or other materials provided with the distribution.
+ * 4. Neither the name of the University nor the names of its contributors
+ * may be used to endorse or promote products derived from this software
+ * without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
+ * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+ * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
+ * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+ * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
+ * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
+ * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+ * SUCH DAMAGE.
+ */
+#if 0
+#include <sys/param.h>
+#include <sys/time.h>
+#endif
+#include <sys/gmon.h>
+#include <sys/gmon_out.h>
+
+#include <stddef.h>
+#include <errno.h>
+#include <stdio.h>
+#include <fcntl.h>
+#include <unistd.h>
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+#if 0
+#include <libc-internal.h>
+#include <not-cancel.h>
+
+#ifdef USE_IN_LIBIO
+# include <wchar.h>
+#endif
+#endif
+#define internal_function
+#define weak_alias(fun,aliasid) extern __typeof(fun) aliasid __attribute__ ((weak, alias (#fun)));
+#define __libc_enable_secure 0
+
+/* Head of basic-block list or NULL. */
+struct __bb *__bb_head attribute_hidden;
+
+struct gmonparam _gmonparam attribute_hidden = { GMON_PROF_OFF };
+
+/*
+ * See profil(2) where this is described:
+ */
+static int s_scale;
+#define SCALE_1_TO_1 0x10000L
+
+#define ERR(s) write (STDERR_FILENO, s, sizeof (s) - 1)
+
+void moncontrol (int mode);
+void __moncontrol (int mode);
+static void write_hist (int fd) internal_function;
+static void write_call_graph (int fd) internal_function;
+static void write_bb_counts (int fd) internal_function;
+
+/*
+ * Control profiling
+ * profiling is what mcount checks to see if
+ * all the data structures are ready.
+ */
+void
+__moncontrol (int mode)
+{
+ struct gmonparam *p = &_gmonparam;
+
+ /* Don't change the state if we ran into an error. */
+ if (p->state == GMON_PROF_ERROR)
+ return;
+
+ if (mode)
+ {
+ /* start */
+ __profil((void *) p->kcount, p->kcountsize, p->lowpc, s_scale);
+ p->state = GMON_PROF_ON;
+ }
+ else
+ {
+ /* stop */
+ __profil(NULL, 0, 0, 0);
+ p->state = GMON_PROF_OFF;
+ }
+}
+weak_alias (__moncontrol, moncontrol)
+
+
+void
+__monstartup (u_long lowpc, u_long highpc)
+{
+ register int o;
+ char *cp;
+ struct gmonparam *p = &_gmonparam;
+ int linesz;
+
+ /*
+ * round lowpc and highpc to multiples of the density we're using
+ * so the rest of the scaling (here and in gprof) stays in ints.
+ */
+ p->lowpc = ROUNDDOWN(lowpc, HISTFRACTION * sizeof(HISTCOUNTER));
+ if (sizeof *p->froms % sizeof(HISTCOUNTER) != 0)
+ {
+ p->highpc = ROUNDUP(highpc, HISTFRACTION * sizeof(HISTCOUNTER));
+ p->textsize = p->highpc - p->lowpc;
+ p->kcountsize = ROUNDUP((p->textsize + HISTFRACTION - 1) / HISTFRACTION,
+ sizeof (*p->froms));
+ }
+ else
+ {
+ /* Avoid odd scales by rounding up highpc to get kcountsize rounded. */
+ p->textsize = ROUNDUP (highpc - p->lowpc,
+ HISTFRACTION * sizeof (*p->froms));
+ p->highpc = p->lowpc + p->textsize;
+ p->kcountsize = p->textsize / HISTFRACTION;
+ }
+ p->hashfraction = HASHFRACTION;
+ p->log_hashfraction = -1;
+ /* The following test must be kept in sync with the corresponding
+ test in mcount.c. */
+ if ((HASHFRACTION & (HASHFRACTION - 1)) == 0) {
+ /* if HASHFRACTION is a power of two, mcount can use shifting
+ instead of integer division. Precompute shift amount. */
+ p->log_hashfraction = ffs(p->hashfraction * sizeof(*p->froms)) - 1;
+ }
+ p->tolimit = p->textsize * ARCDENSITY / 100;
+ if (p->tolimit < MINARCS)
+ p->tolimit = MINARCS;
+ else if (p->tolimit > MAXARCS)
+ p->tolimit = MAXARCS;
+ p->tossize = p->tolimit * sizeof(struct tostruct);
+
+ /* p->kcount must not share cache lines with the adjacent data, because
+ we use uncached accesses while profiling. */
+ linesz = __dcache_linesz ();
+ cp = calloc (ROUNDUP (p->kcountsize, linesz) + p->tossize
+ + (linesz - 1), 1);
+ if (! cp)
+ {
+ ERR("monstartup: out of memory\n");
+ p->tos = NULL;
+ p->state = GMON_PROF_ERROR;
+ /* In case we loose the error state due to a race,
+ prevent invalid writes also by clearing tolimit. */
+ p->tolimit = 0;
+ return;
+ }
+ p->tos = (struct tostruct *)cp;
+ cp += p->tossize;
+ cp = (char *) ROUNDUP ((ptrdiff_t) cp, linesz);
+ p->kcount = (HISTCOUNTER *)cp;
+ cp += ROUNDUP (p->kcountsize, linesz);
+
+ p->tos[0].link = 0;
+
+ o = p->highpc - p->lowpc;
+ if (p->kcountsize < (u_long) o)
+ {
+#ifndef hp300
+ s_scale = ((float)p->kcountsize / o ) * SCALE_1_TO_1;
+#else
+ /* avoid floating point operations */
+ int quot = o / p->kcountsize;
+
+ if (quot >= 0x10000)
+ s_scale = 1;
+ else if (quot >= 0x100)
+ s_scale = 0x10000 / quot;
+ else if (o >= 0x800000)
+ s_scale = 0x1000000 / (o / (p->kcountsize >> 8));
+ else
+ s_scale = 0x1000000 / ((o << 8) / p->kcountsize);
+#endif
+ } else
+ s_scale = SCALE_1_TO_1;
+
+ __moncontrol(1);
+}
+weak_alias (__monstartup, monstartup)
+
+
+static void
+internal_function
+write_hist (int fd)
+{
+ u_char tag = GMON_TAG_TIME_HIST;
+ struct arc_gmon_hist_hdr thdr __attribute__ ((aligned (__alignof__ (char *))));
+ int r;
+
+ if (_gmonparam.kcountsize > 0)
+ {
+ *(char **) thdr.low_pc = (char *) _gmonparam.lowpc;
+ *(char **) thdr.high_pc = (char *) _gmonparam.highpc;
+ *(int32_t *) thdr.hist_size = (_gmonparam.kcountsize
+ / sizeof (HISTCOUNTER));
+ *(int32_t *) thdr.prof_rate = __profile_frequency ();
+ strncpy (thdr.dimen, "seconds", sizeof (thdr.dimen));
+ thdr.dimen_abbrev = 's';
+
+ r = write (fd, &tag, sizeof tag);
+ if (r != sizeof tag)
+ return;
+ r = write (fd, &thdr, sizeof thdr);
+ if (r != sizeof thdr)
+ return;
+ r = write (fd,_gmonparam.kcount, _gmonparam.kcountsize);
+ if ((unsigned) r != _gmonparam.kcountsize)
+ return;
+ }
+}
+
+
+static void
+internal_function
+write_call_graph (int fd)
+{
+#define NARCS_PER_WRITE 64
+#define BYTES_PER_ARC (1 + sizeof (struct gmon_cg_arc_record))
+#define BYTES_PER_WRITE (BYTES_PER_ARC * NARCS_PER_WRITE)
+ ARCINDEX to_index;
+ u_long frompc, selfpc, count;
+ char buffer[BYTES_PER_WRITE], *p;
+ u_long *prof_desc = __arc_profile_desc_secstart;
+ u_long *prof_count = __arc_profile_counters_secstart;
+ u_long *prof_desc_end = __arc_profile_desc_secend;
+ u_long *prof_forward = __arc_profile_forward_secstart;
+
+ for (p = buffer; p < buffer + BYTES_PER_WRITE; p += BYTES_PER_ARC)
+ *p = GMON_TAG_CG_ARC;
+ p = buffer;
+ frompc = *prof_desc++ & -2;
+ while (prof_desc < prof_desc_end)
+ {
+ selfpc = *prof_desc++;
+ if (selfpc & 1)
+ {
+ frompc = selfpc & -2;
+ selfpc = *prof_desc++;
+ }
+ count = *prof_count++;
+ if (selfpc)
+ {
+ struct arc
+ {
+ char *frompc;
+ char *selfpc;
+ int32_t count;
+ }
+ arc;
+
+ if (!count)
+ continue;
+ arc.frompc = (char *) frompc;
+ arc.selfpc = (char *) selfpc;
+ arc.count = count;
+ memcpy (p + 1, &arc, sizeof arc);
+ p += 1 + sizeof arc;
+
+ if (p == buffer + BYTES_PER_WRITE)
+ {
+ write (fd, buffer, BYTES_PER_WRITE);
+ p = buffer;
+ }
+ }
+ else
+ {
+ for (to_index = count;
+ to_index != 0;
+ to_index = _gmonparam.tos[to_index].link)
+ {
+ struct arc
+ {
+ char *frompc;
+ char *selfpc;
+ int32_t count;
+ }
+ arc;
+
+ arc.frompc = (char *) frompc;
+ arc.selfpc = (char *) _gmonparam.tos[to_index].selfpc;
+ arc.count = _gmonparam.tos[to_index].count;
+ memcpy (p + 1, &arc, sizeof arc);
+ p += 1 + sizeof arc;
+
+ if (p == buffer + BYTES_PER_WRITE)
+ {
+ write (fd, buffer, BYTES_PER_WRITE);
+ p = buffer;
+ }
+ }
+ }
+ }
+ while (prof_forward < __arc_profile_forward_secend)
+ {
+ /* ??? The 'call count' is actually supposed to be a fixed point
+ factor, with 16 bits each before and after the point.
+ It would be much nicer if we figured out the actual number
+ of calls to the caller, and multiplied that with the fixed point
+ factor to arrive at the estimated calls for the callee. */
+ memcpy (p + 1, prof_forward, 3 * sizeof *prof_forward);
+ prof_forward += 3;
+ p += 1 + 3 * sizeof *prof_forward;
+ if (p == buffer + BYTES_PER_WRITE)
+ {
+ write (fd, buffer, BYTES_PER_WRITE);
+ p = buffer;
+ }
+ }
+ if (p != buffer)
+ write (fd, buffer, p - buffer);
+}
+
+
+static void
+internal_function
+write_bb_counts (int fd)
+{
+ struct __bb *grp;
+ u_char tag = GMON_TAG_BB_COUNT;
+ size_t ncounts;
+ size_t i;
+
+ struct { unsigned long address; long count; } bbbody[8];
+ size_t nfilled;
+
+ /* Write each group of basic-block info (all basic-blocks in a
+ compilation unit form a single group). */
+
+ for (grp = __bb_head; grp; grp = grp->next)
+ {
+ ncounts = grp->ncounts;
+ write (fd, &tag, 1);
+ write (fd, &ncounts, sizeof ncounts);
+ for (nfilled = i = 0; i < ncounts; ++i)
+ {
+ if (nfilled == sizeof (bbbody) / sizeof (bbbody[0]))
+ {
+ write (fd, bbbody, sizeof bbbody);
+ nfilled = 0;
+ }
+
+ bbbody[nfilled].address = grp->addresses[i];
+ bbbody[nfilled++].count = grp->counts[i];
+ }
+ if (nfilled > 0)
+ write (fd, bbbody, nfilled * sizeof bbbody[0]);
+ }
+}
+
+
+static void
+write_gmon (void)
+{
+ struct gmon_hdr ghdr __attribute__ ((aligned (__alignof__ (int))));
+ int fd = -1;
+ char *env;
+
+#ifndef O_NOFOLLOW
+# define O_NOFOLLOW 0
+#endif
+
+ env = getenv ("GMON_OUT_PREFIX");
+ if (env != NULL && !__libc_enable_secure)
+ {
+ size_t len = strlen (env);
+ char buf[len + 20];
+ snprintf (buf, sizeof (buf), "%s.%u", env, getpid ());
+ fd = open (buf, O_CREAT|O_TRUNC|O_WRONLY|O_NOFOLLOW, 0666);
+ }
+
+ if (fd == -1)
+ {
+ fd = open ("gmon.out", O_CREAT|O_TRUNC|O_WRONLY|O_NOFOLLOW,
+ 0666);
+ if (fd < 0)
+ {
+ perror ("_mcleanup: gmon.out");
+ return;
+ }
+ }
+
+ /* write gmon.out header: */
+ memset (&ghdr, '\0', sizeof (struct gmon_hdr));
+ memcpy (&ghdr.cookie[0], GMON_MAGIC, sizeof (ghdr.cookie));
+ *(int32_t *) ghdr.version = GMON_VERSION;
+ write (fd, &ghdr, sizeof (struct gmon_hdr));
+
+ /* write PC histogram: */
+ write_hist (fd);
+
+ /* write call-graph: */
+ write_call_graph (fd);
+
+ /* write basic-block execution counts: */
+ write_bb_counts (fd);
+
+ close (fd);
+}
+
+
+void
+__write_profiling (void)
+{
+ int save = _gmonparam.state;
+ _gmonparam.state = GMON_PROF_OFF;
+ if (save == GMON_PROF_ON)
+ write_gmon ();
+ _gmonparam.state = save;
+}
+#ifndef SHARED
+/* This symbol isn't used anywhere in the DSO and it is not exported.
+ This would normally mean it should be removed to get the same API
+ in static libraries. But since profiling is special in static libs
+ anyway we keep it. But not when building the DSO since some
+ quality assurance tests will otherwise trigger. */
+weak_alias (__write_profiling, write_profiling)
+#endif
+
+
+void
+_mcleanup (void)
+{
+ __moncontrol (0);
+
+ if (_gmonparam.state != GMON_PROF_ERROR)
+ write_gmon ();
+
+ /* free the memory. */
+ if (_gmonparam.tos != NULL)
+ free (_gmonparam.tos);
+}
--- /dev/null
+/* Copyright (C) 2007-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#ifndef MACHINE_GMON_H
+#define MACHINE_GMON_H
+
+/* We can't fake out own <sys/types.h> header because the newlib / uclibc
+ headers in GCC_FOR_TARGET take precedence. */
+
+#define __BEGIN_DECLS
+#define __END_DECLS
+
+#define __THROW
+
+extern int __dcache_linesz (void);
+
+#define _MCOUNT_DECL(countp, selfpc) \
+ static inline void _mcount_internal (void *countp, u_long selfpc)
+
+extern void _mcount (void);
+extern void _mcount_call (void);
+
+/* N.B.: the calling point might be a sibcall, thus blink does not necessarily
+ hold the caller's address. r8 doesn't hold the caller's address, either,
+ but rather a pointer to the counter data structure associated with the
+ caller.
+ This function must be compiled with optimization turned on in order to
+ enable a sibcall for the final call to selfpc; this is important when trying
+ to profile a program with deep tail-recursion that would get a stack
+ overflow otherwise. */
+#define MCOUNT \
+void \
+_mcount_call (void) \
+{ \
+ register void *countp __asm("r8"); \
+ register u_long selfpc __asm("r9"); \
+ _mcount_internal (countp, selfpc); \
+ ((void (*)(void)) selfpc) (); \
+}
+
+extern int __profil (u_short *,size_t, size_t, u_int);
+
+#endif /* MACHINE_GMON_H */
--- /dev/null
+/*-
+ * Copyright (c) 1983, 1992, 1993
+ * The Regents of the University of California. All rights reserved.
+ *
+ * Copyright (C) 2007-2012 Free Software Foundation, Inc.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ * 1. Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimer in the
+ * documentation and/or other materials provided with the distribution.
+ * 4. Neither the name of the University nor the names of its contributors
+ * may be used to endorse or promote products derived from this software
+ * without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
+ * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+ * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
+ * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+ * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
+ * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
+ * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+ * SUCH DAMAGE.
+ */
+
+#if !defined(lint) && !defined(KERNEL) && defined(LIBC_SCCS)
+static char sccsid[] = "@(#)mcount.c 8.1 (Berkeley) 6/4/93";
+#endif
+
+#if 0
+#include <unistd.h>
+#include <sys/param.h>
+#endif
+#include <sys/gmon.h>
+
+/* This file provides the machine-dependent definitions of the _MCOUNT_DECL
+ and MCOUNT macros. */
+#include <machine-gmon.h>
+
+#include <atomic.h>
+
+/*
+ * mcount is called on entry to each function compiled with the profiling
+ * switch set. _mcount(), which is declared in a machine-dependent way
+ * with _MCOUNT_DECL, does the actual work and is either inlined into a
+ * C routine or called by an assembly stub. In any case, this magic is
+ * taken care of by the MCOUNT definition in <machine/profile.h>.
+ *
+ * _mcount updates data structures that represent traversals of the
+ * program's call graph edges. frompc and selfpc are the return
+ * address and function address that represents the given call graph edge.
+ *
+ * Note: the original BSD code used the same variable (frompcindex) for
+ * both frompcindex and frompc. Any reasonable, modern compiler will
+ * perform this optimization.
+ */
+_MCOUNT_DECL(count_ptr, selfpc) /* _mcount; may be static, inline, etc */
+{
+ register ARCINDEX *frompcindex;
+ register struct tostruct *top, *prevtop;
+ register struct gmonparam *p;
+ register ARCINDEX toindex;
+
+ /* Check for nested function trampoline. */
+ if (selfpc & 2)
+ selfpc = *(u_long *) (selfpc + 10);
+
+ p = &_gmonparam;
+ /*
+ * check that we are profiling
+ * and that we aren't recursively invoked.
+ */
+#if 0
+ if (catomic_compare_and_exchange_bool_acq (&p->state, GMON_PROF_BUSY,
+ GMON_PROF_ON))
+ return;
+#elif defined (__ARC700__)
+/* ??? This could temporrarily loose the ERROR / OFF condition in a race,
+ but doing an actual compare_and_exchange would be too costly. It would
+ be better if we had a semaphore independent of the 'sticky' state, but
+ then we could run into ABI compatibility problems with the size of struct
+ gmonparam. */
+ {
+ u_long old_state;
+
+ __asm ("ex %0,%1": "=r" (old_state), "+m" (p->state)
+ : "0" (GMON_PROF_BUSY));
+ if (old_state != GMON_PROF_ON)
+ {
+ switch (old_state)
+ {
+ case GMON_PROF_OFF:
+ __asm ("ex %0,%1": "+r" (old_state), "+m" (p->state));
+ if (old_state == GMON_PROF_BUSY
+ /* Switching off while we say we are busy while profiling
+ was actually already switched off is all right. */
+ || old_state == GMON_PROF_OFF)
+ break;
+ /* It is not clear if we should allow switching on
+ profiling at this point, and how to handle further races.
+ For now, record an error in this case. */
+ /* Fall through. */
+ default: /* We expect here only GMON_PROF_ERROR. */
+ p->state = GMON_PROF_ERROR;
+ break;
+ case GMON_PROF_BUSY: break;
+ }
+ return;
+ }
+ }
+#else /* ??? No semaphore primitives available. */
+ if (p->state != GMON_PROF_ON)
+ return;
+ p->state = GMON_PROF_BUSY;
+#endif
+
+ frompcindex = count_ptr;
+ toindex = *frompcindex;
+ if (toindex == 0) {
+ /*
+ * first time traversing this arc
+ */
+ toindex = ++p->tos[0].link;
+ if (toindex >= (ARCINDEX) p->tolimit)
+ /* halt further profiling */
+ goto overflow;
+
+ *frompcindex = toindex;
+ top = &p->tos[toindex];
+ top->selfpc = selfpc;
+ top->count = 1;
+ top->link = 0;
+ goto done;
+ }
+ top = &p->tos[toindex];
+ if (top->selfpc == selfpc) {
+ /*
+ * arc at front of chain; usual case.
+ */
+ top->count++;
+ goto done;
+ }
+ /*
+ * have to go looking down chain for it.
+ * top points to what we are looking at,
+ * prevtop points to previous top.
+ * we know it is not at the head of the chain.
+ */
+ for (; /* goto done */; ) {
+ if (top->link == 0) {
+ /*
+ * top is end of the chain and none of the chain
+ * had top->selfpc == selfpc.
+ * so we allocate a new tostruct
+ * and link it to the head of the chain.
+ */
+ toindex = ++p->tos[0].link;
+ if (toindex >= (ARCINDEX) p->tolimit)
+ goto overflow;
+
+ top = &p->tos[toindex];
+ top->selfpc = selfpc;
+ top->count = 1;
+ top->link = *frompcindex;
+ *frompcindex = toindex;
+ goto done;
+ }
+ /*
+ * otherwise, check the next arc on the chain.
+ */
+ prevtop = top;
+ top = &p->tos[top->link];
+ if (top->selfpc == selfpc) {
+ /*
+ * there it is.
+ * increment its count
+ * move it to the head of the chain.
+ */
+ top->count++;
+ toindex = prevtop->link;
+ prevtop->link = top->link;
+ top->link = *frompcindex;
+ *frompcindex = toindex;
+ goto done;
+ }
+
+ }
+done:
+ p->state = GMON_PROF_ON;
+ return;
+overflow:
+ p->state = GMON_PROF_ERROR;
+ return;
+}
+
+/*
+ * Actual definition of mcount function. Defined in <machine/profile.h>,
+ * which is included by <sys/gmon.h>.
+ */
+MCOUNT
--- /dev/null
+/* This file contains code to do profiling.
+
+ Copyright (C) 2007-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "../asm.h"
+/* This file contains code to do profiling. */
+ .weak __profile_frequency_value
+ .global __profile_frequency_value
+ .set __profile_frequency_value, 1000
+ .text
+ .balign 4
+ .global __profile_frequency
+ FUNC(__profile_frequency)
+__profile_frequency:
+ mov_s r0,__profile_frequency_value
+ j_s [blink]
+ ENDFUNC(__profile_frequency)
--- /dev/null
+/* Return frequency of ticks reported by profil. Generic version. */
+/*-
+ * Copyright (c) 1983, 1992, 1993
+ * The Regents of the University of California. All rights reserved.
+ *
+ * Copyright (C) 2007-2012 Free Software Foundation, Inc.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ * 1. Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimer in the
+ * documentation and/or other materials provided with the distribution.
+ * 4. Neither the name of the University nor the names of its contributors
+ * may be used to endorse or promote products derived from this software
+ * without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
+ * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+ * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
+ * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+ * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
+ * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
+ * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+ * SUCH DAMAGE.
+ */
+
+
+#include <sys/types.h>
+#include <sys/time.h>
+#if 0
+#include <libc-internal.h>
+#else
+#include "sys/gmon.h"
+#endif
+
+int
+__profile_frequency (void)
+{
+ /*
+ * Discover the tick frequency of the machine if something goes wrong,
+ * we return 0, an impossible hertz.
+ */
+ struct itimerval tim;
+
+ tim.it_interval.tv_sec = 0;
+ tim.it_interval.tv_usec = 1;
+ tim.it_value.tv_sec = 0;
+ tim.it_value.tv_usec = 0;
+ setitimer(ITIMER_REAL, &tim, 0);
+ setitimer(ITIMER_REAL, 0, &tim);
+ if (tim.it_interval.tv_usec < 2)
+ return 0;
+ return (1000000 / tim.it_interval.tv_usec);
+}
--- /dev/null
+/* This file contains code to do profiling.
+
+ Copyright (C) 2007-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "../asm.h"
+#include "auxreg.h"
+/* This file contains code to do profiling. */
+ .weak __profile_timer_cycles
+ .global __profile_timer_cycles
+ .set __profile_timer_cycles, 200
+
+ .section .bss
+ .global __profil_offset
+ .align 4
+ .type __profil_offset, @object
+ .size __profil_offset, 4
+__profil_offset:
+ .zero 4
+
+ .text
+ .global __dcache_linesz
+ .global __profil
+ FUNC(__profil)
+.Lstop_profiling:
+ sr r0,[CONTROL0]
+ j_s [blink]
+ .balign 4
+__profil:
+.Lprofil:
+ breq_s r0,0,.Lstop_profiling
+ ; r0: buf r1: bufsiz r2: offset r3: scale
+ bxor.f r3,r3,15; scale must be 0x8000, i.e. 1/2; generate 0.
+ push_s blink
+ lsr_s r2,r2,1
+ mov_s r8,r0
+ flag.ne 1 ; halt if wrong scale
+ sub_s r0,r0,r2
+ st r0,[__profil_offset]
+ bl __dcache_linesz
+ pop_s blink
+ bbit1.d r0,0,nocache
+ mov_s r0,r8
+#ifdef __ARC700__
+ add_s r1,r1,31
+ lsr.f lp_count,r1,5
+ lpne 2f
+ sr r0,[DC_FLDL]
+ add_s r0,r0,32
+#else /* !__ARC700__ */
+# FIX ME: set up loop according to cache line size
+ lr r12,[D_CACHE_BUILD]
+ sub_s r0,r0,16
+ sub_s r1,r1,1
+ lsr_s r12,r12,16
+ asr_s r1,r1,4
+ bmsk_s r12,r12,3
+ asr_s r1,r1,r12
+ add.f lp_count,r1,1
+ mov_s r1,16
+ asl_s r1,r1,r12
+ lpne 2f
+ add r0,r0,r1
+ sr r0,[DC_FLDL]
+#endif /* __ARC700__ */
+2: b_s .Lcounters_cleared
+nocache:
+.Lcounters_cleared:
+ lr r1,[INT_VECTOR_BASE] ; disable timer0 interrupts
+ sr r3,[CONTROL0]
+ sr r3,[COUNT0]
+0: ld_s r0,[pcl,1f-0b+((0b-.Lprofil) & 2)] ; 1f@GOTOFF
+0: ld_s r12,[pcl,1f+4-0b+((0b-.Lprofil) & 2)] ; 1f@GOTOFF + 4
+ st_s r0,[r1,24]; timer0 uses vector3
+ st_s r12,[r1,24+4]; timer0 uses vector3
+ ;sr 10000,[LIMIT0]
+ sr __profile_timer_cycles,[LIMIT0]
+ mov_s r12,3 ; enable timer interrupts; count only when not halted.
+ sr r12,[CONTROL0]
+ lr r12,[STATUS32]
+ bset_s r12,r12,1 ; allow level 1 interrupts
+ flag r12
+ mov_s r0,0
+ j_s [blink]
+ .balign 4
+1: j __profil_irq
+ ENDFUNC(__profil)
+
+ FUNC(__profil_irq)
+ .balign 4 ; make final jump unaligned to avoid delay penalty
+ .balign 32,0,12 ; make sure the code spans no more that two cache lines
+ nop_s
+__profil_irq:
+ push_s r0
+ ld r0,[__profil_offset]
+ push_s r1
+ lsr r1,ilink1,2
+ push_s r2
+ ldw.as.di r2,[r0,r1]
+ add1 r0,r0,r1
+ ld_s r1,[sp,4]
+ add_s r2,r2,1
+ bbit1 r2,16,nostore
+ stw.di r2,[r0]
+nostore:ld.ab r2,[sp,8]
+ pop_s r0
+ j.f [ilink1]
+ ENDFUNC(__profil_irq)
+
+; could save one cycle if the counters were allocated at link time and
+; the contents of __profil_offset were pre-computed at link time, like this:
+#if 0
+; __profil_offset needs to be PROVIDEd as __profile_base-text/4
+ .global __profil_offset
+ .balign 4
+__profil_irq:
+ push_s r0
+ lsr r0,ilink1,2
+ add1 r0,__profil_offset,r0
+ push_s r1
+ ldw.di r1,[r0]
+
+
+ add_s r1,r1,1
+ bbit1 r1,16,nostore
+ stw.di r1,[r0]
+nostore:pop_s r1
+ pop_s r0
+ j [ilink1]
+#endif /* 0 */
--- /dev/null
+/*-
+ * Copyright (c) 1982, 1986, 1992, 1993
+ * The Regents of the University of California. All rights reserved.
+ * Copyright (C) 2007-2012 Free Software Foundation, Inc.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ * 1. Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimer in the
+ * documentation and/or other materials provided with the distribution.
+ * 4. Neither the name of the University nor the names of its contributors
+ * may be used to endorse or promote products derived from this software
+ * without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
+ * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+ * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
+ * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+ * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
+ * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
+ * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+ * SUCH DAMAGE.
+ *
+ * @(#)gmon.h 8.2 (Berkeley) 1/4/94
+ */
+
+#ifndef _SYS_GMON_H
+#define _SYS_GMON_H 1
+
+#if 0
+#include <features.h>
+#include <sys/types.h>
+#else
+#include <sys/types.h>
+#include "machine-gmon.h"
+#define attribute_hidden __attribute__ ((visibility("hidden")))
+#endif
+
+#include <stdint.h>
+
+/*
+ * See gmon_out.h for gmon.out format.
+ */
+
+/* structure emitted by "gcc -a". This must match struct bb in
+ gcc/libgcc2.c. It is OK for gcc to declare a longer structure as
+ long as the members below are present. */
+struct __bb
+{
+ long zero_word;
+ const char *filename;
+ long *counts;
+ long ncounts;
+ struct __bb *next;
+ const unsigned long *addresses;
+};
+
+extern struct __bb *__bb_head;
+
+/*
+ * histogram counters are unsigned shorts (according to the kernel).
+ */
+#define HISTCOUNTER unsigned short
+
+/*
+ * fraction of text space to allocate for histogram counters here, 1/2
+ */
+#define HISTFRACTION 2
+
+/*
+ * Fraction of text space to allocate for from hash buckets.
+ * The value of HASHFRACTION is based on the minimum number of bytes
+ * of separation between two subroutine call points in the object code.
+ * Given MIN_SUBR_SEPARATION bytes of separation the value of
+ * HASHFRACTION is calculated as:
+ *
+ * HASHFRACTION = MIN_SUBR_SEPARATION / (2 * sizeof(short) - 1);
+ *
+ * For example, on the VAX, the shortest two call sequence is:
+ *
+ * calls $0,(r0)
+ * calls $0,(r0)
+ *
+ * which is separated by only three bytes, thus HASHFRACTION is
+ * calculated as:
+ *
+ * HASHFRACTION = 3 / (2 * 2 - 1) = 1
+ *
+ * Note that the division above rounds down, thus if MIN_SUBR_FRACTION
+ * is less than three, this algorithm will not work!
+ *
+ * In practice, however, call instructions are rarely at a minimal
+ * distance. Hence, we will define HASHFRACTION to be 2 across all
+ * architectures. This saves a reasonable amount of space for
+ * profiling data structures without (in practice) sacrificing
+ * any granularity.
+ */
+#define HASHFRACTION 2
+
+/*
+ * Percent of text space to allocate for tostructs.
+ * This is a heuristic; we will fail with a warning when profiling programs
+ * with a very large number of very small functions, but that's
+ * normally OK.
+ * 2 is probably still a good value for normal programs.
+ * Profiling a test case with 64000 small functions will work if
+ * you raise this value to 3 and link statically (which bloats the
+ * text size, thus raising the number of arcs expected by the heuristic).
+ */
+#define ARCDENSITY 3
+
+/*
+ * Always allocate at least this many tostructs. This
+ * hides the inadequacy of the ARCDENSITY heuristic, at least
+ * for small programs.
+ */
+#define MINARCS 50
+
+/*
+ * The type used to represent indices into gmonparam.tos[].
+ */
+#define ARCINDEX u_long
+
+/*
+ * Maximum number of arcs we want to allow.
+ * Used to be max representable value of ARCINDEX minus 2, but now
+ * that ARCINDEX is a long, that's too large; we don't really want
+ * to allow a 48 gigabyte table.
+ * The old value of 1<<16 wasn't high enough in practice for large C++
+ * programs; will 1<<20 be adequate for long? FIXME
+ */
+#define MAXARCS (1 << 20)
+
+struct tostruct {
+ u_long selfpc;
+ long count;
+ ARCINDEX link;
+};
+
+/*
+ * a raw arc, with pointers to the calling site and
+ * the called site and a count.
+ */
+struct rawarc {
+ u_long raw_frompc;
+ u_long raw_selfpc;
+ long raw_count;
+};
+
+/*
+ * general rounding functions.
+ */
+#define ROUNDDOWN(x,y) (((x)/(y))*(y))
+#define ROUNDUP(x,y) ((((x)+(y)-1)/(y))*(y))
+
+/*
+ * The profiling data structures are housed in this structure.
+ */
+struct gmonparam {
+ long int state;
+ u_short *kcount;
+ u_long kcountsize;
+ ARCINDEX *froms;
+ u_long fromssize;
+ struct tostruct *tos;
+ u_long tossize;
+ long tolimit;
+ u_long lowpc;
+ u_long highpc;
+ u_long textsize;
+ u_long hashfraction;
+ long log_hashfraction;
+};
+extern struct gmonparam _gmonparam;
+
+/*
+ * Possible states of profiling.
+ */
+#define GMON_PROF_ON 0
+#define GMON_PROF_BUSY 1
+#define GMON_PROF_ERROR 2
+#define GMON_PROF_OFF 3
+
+/*
+ * Sysctl definitions for extracting profiling information from the kernel.
+ */
+#define GPROF_STATE 0 /* int: profiling enabling variable */
+#define GPROF_COUNT 1 /* struct: profile tick count buffer */
+#define GPROF_FROMS 2 /* struct: from location hash bucket */
+#define GPROF_TOS 3 /* struct: destination/count structure */
+#define GPROF_GMONPARAM 4 /* struct: profiling parameters (see above) */
+
+__BEGIN_DECLS
+
+/* Set up data structures and start profiling. */
+extern void __monstartup (u_long __lowpc, u_long __highpc) __THROW;
+extern void monstartup (u_long __lowpc, u_long __highpc) __THROW;
+
+/* Clean up profiling and write out gmon.out. */
+extern void _mcleanup (void) __THROW;
+
+extern void __write_profiling (void);
+extern int attribute_hidden __profile_frequency (void);
+
+extern u_long __arc_profile_desc_secstart[], __arc_profile_desc_secend[];
+extern u_long __arc_profile_forward_secstart[], __arc_profile_forward_secend[];
+extern u_long __arc_profile_counters_secstart[];
+
+__END_DECLS
+
+#endif /* sys/gmon.h */
--- /dev/null
+/* Copyright (C) 2007-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#define GMON_TAG_TIME_HIST 0
+#define GMON_TAG_CG_ARC 1
+#define GMON_TAG_BB_COUNT 2
+
+#define GMON_MAGIC "gmon"
+#define GMON_VERSION 1
+
+struct arc_gmon_hist_hdr
+{
+ char low_pc[4];
+ char high_pc[4];
+ char hist_size[4];
+ char prof_rate[4];
+ char dimen[15];
+ char dimen_abbrev;
+};
+
+struct gmon_cg_arc_record
+{
+ char afrompc[4];
+ char selfpc[4];
+ char count[4];
+};
+
+struct gmon_hdr
+{
+ char cookie[4];
+ char version[4];
+ char c[12];
+};
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "arc-ieee-754.h"
+#if 0 /* DEBUG */
+ .global __adddf3
+ .balign 4
+__adddf3:
+ push_s blink
+ push_s r2
+ push_s r3
+ push_s r0
+ bl.d __adddf3_c
+ push_s r1
+ ld_s r2,[sp,12]
+ ld_s r3,[sp,8]
+ st_s r0,[sp,12]
+ st_s r1,[sp,8]
+ pop_s r1
+ bl.d __adddf3_asm
+ pop_s r0
+ pop_s r3
+ pop_s r2
+ pop_s blink
+ cmp r0,r2
+ cmp.eq r1,r3
+ jeq_s [blink]
+ bl abort
+ .global __subdf3
+ .balign 4
+__subdf3:
+ push_s blink
+ push_s r2
+ push_s r3
+ push_s r0
+ bl.d __subdf3_c
+ push_s r1
+ ld_s r2,[sp,12]
+ ld_s r3,[sp,8]
+ st_s r0,[sp,12]
+ st_s r1,[sp,8]
+ pop_s r1
+ bl.d __subdf3_asm
+ pop_s r0
+ pop_s r3
+ pop_s r2
+ pop_s blink
+ cmp r0,r2
+ cmp.eq r1,r3
+ jeq_s [blink]
+ bl abort
+#define __adddf3 __adddf3_asm
+#define __subdf3 __subdf3_asm
+#endif /* DEBUG */
+/* N.B. This is optimized for ARC700.
+ ARC600 has very different scheduling / instruction selection criteria. */
+
+/* inputs: DBL0, DBL1 (r0-r3)
+ output: DBL0 (r0, r1)
+ clobber: r2-r10, r12, flags
+ All NaN highword bits must be 1. NaN low word is random. */
+
+ .balign 4
+ .global __adddf3
+ .global __subdf3
+ .long 0x7ff00000 ; exponent mask
+ FUNC(__adddf3)
+ FUNC(__subdf3)
+__subdf3:
+ bxor_l DBL1H,DBL1H,31
+__adddf3:
+ ld r9,[pcl,-8]
+ bmsk r4,DBL0H,30
+ xor r10,DBL0H,DBL1H
+ and r6,DBL1H,r9
+ sub.f r12,r4,r6
+ asr_s r12,r12,20
+ blo .Ldbl1_gt
+ brhs r4,r9,.Linf_nan
+ brhs r12,32,.Large_shift
+ brne r12,0,.Lsmall_shift
+ brge r10,0,.Ladd_same_exp ; r12 == 0
+
+/* After subtracting, we need to normalize; when shifting to place the
+ leading 1 into position for the implicit 1 and adding that to DBL0H,
+ we increment the exponent. Thus, we have to subtract one more than
+ the shift count from the exponent beforehand. Iff the exponent drops thus
+ below zero (before adding in the fraction with the leading one), we have
+ generated a denormal number. Denormal handling is basicallly reducing the
+ shift count so that we produce a zero exponent instead; however, this way
+ the shift count can become zero (if we started out with exponent 1).
+ Therefore, a simple min operation is not good enough, since we don't
+ want to handle a zero normalizing shift in the main path.
+ On the plus side, we don't need to check for denorm input, the result
+ of subtracing these looks just the same as denormals generated during
+ subtraction. */
+ bmsk r7,DBL1H,30
+ cmp r4,r7
+ cmp.eq DBL0L,DBL1L
+ blo .L_rsub_same_exp
+ sub.f DBL0L,DBL0L,DBL1L
+ bmsk r12,DBL0H,19
+ bic DBL1H,DBL0H,r12
+ sbc.f r4,r4,r7
+ beq_l .Large_cancel
+ norm DBL1L,r4
+ b.d .Lsub_done_same_exp
+ sub r12,DBL1L,9
+
+ .balign 4
+.Linf_nan:
+ ; If both inputs are inf, but with different signs, the result is NaN.
+ asr r12,r10,31
+ or_s DBL1H,DBL1H,r12
+ j_s.d [blink]
+ or.eq DBL0H,DBL0H,DBL1H
+
+ .balign 4
+.L_rsub_same_exp:
+ rsub.f DBL0L,DBL0L,DBL1L
+ bmsk r12,DBL1H,19
+ bic_s DBL1H,DBL1H,r12
+ sbc.f r4,r7,r4
+ beq_l .Large_cancel
+ norm DBL1L,r4
+
+ sub r12,DBL1L,9
+.Lsub_done_same_exp:
+ asl_s r12,r12,20
+ sub_s DBL1L,DBL1L,10
+ sub DBL0H,DBL1H,r12
+ xor.f 0,DBL0H,DBL1H
+ bmi .Ldenorm
+.Lpast_denorm:
+ neg_s r12,DBL1L
+ lsr r7,DBL0L,r12
+ asl r12,r4,DBL1L
+ asl_s DBL0L,DBL0L,DBL1L
+ add_s r12,r12,r7
+ j_s.d [blink]
+ add_l DBL0H,DBL0H,r12
+ .balign 4
+.Ladd_same_exp:
+ /* This is a special case because we can't test for need to shift
+ down by checking if bit 20 of DBL0H changes. OTOH, here we know
+ that we always need to shift down. */
+ ; The implicit 1 of DBL0 is not shifted together with the
+ ; fraction, thus effectively doubled, compensating for not setting
+ ; implicit1 for DBL1
+ add_s r12,DBL0L,DBL1L
+ lsr.f 0,r12,2 ; round to even
+ breq r6,0,.Ldenorm_add
+ adc.f DBL0L,DBL0L,DBL1L
+ sub r7,DBL1H,DBL0H
+ sub1 r7,r7,r9 ; boost exponent by 2/2
+ rrc DBL0L,DBL0L
+ asr.f r7,r7 ; DBL1.fraction/2 - DBL0.fraction/2 ; exp++
+ add.cs.f DBL0L,DBL0L,0x80000000
+ add_l DBL0H,DBL0H,r7 ; DBL0.implicit1 not shifted for DBL1.implicit1
+ add.cs DBL0H,DBL0H,1
+ bic.f 0,r9,DBL0H ; check for overflow -> infinity.
+ jne_l [blink]
+ and DBL0H,DBL0H,0xfff00000
+ j_s.d [blink]
+ mov_s DBL0L,0
+ .balign 4
+.Large_shift:
+ brhs r12,55,.Lret_dbl0
+ bmsk_s DBL1H,DBL1H,19
+ brne r6,0,.Lno_denorm_large_shift
+ brhi.d r12,33,.Lfixed_denorm_large_shift
+ sub_s r12,r12,1
+ breq r12,31, .Lfixed_denorm_small_shift
+.Lshift32:
+ mov_s r12,DBL1L
+ mov_s DBL1L,DBL1H
+ brlt.d r10,0,.Lsub
+ mov_s DBL1H,0
+ b_s .Ladd
+.Ldenorm_add:
+ cmp_s r12,DBL1L
+ mov_s DBL0L,r12
+ j_s.d [blink]
+ adc DBL0H,r4,DBL1H
+
+.Lret_dbl0:
+ j_s [blink]
+ .balign 4
+.Lsmall_shift:
+ breq.d r6,0,.Ldenorm_small_shift
+ bmsk_s DBL1H,DBL1H,19
+ bset_s DBL1H,DBL1H,20
+.Lfixed_denorm_small_shift:
+ neg r8,r12
+ asl r4,DBL1H,r8
+ lsr_l DBL1H,DBL1H,r12
+ lsr r5,DBL1L,r12
+ asl r12,DBL1L,r8
+ brge.d r10,0,.Ladd
+ or DBL1L,r4,r5
+/* subtract, abs(DBL0) > abs(DBL1) */
+/* DBL0H, DBL0L: original values
+ DBL1H, DBL1L: fraction with explicit leading 1, shifted into place
+ r4: orig. DBL0H & 0x7fffffff
+ r6: orig. DBL1H & 0x7ff00000
+ r9: 0x7ff00000
+ r10: orig. DBL0H ^ DBL1H
+ r12: guard bits */
+ .balign 4
+.Lsub:
+ neg.f r12,r12
+ mov_s r7,DBL1H
+ bmsk r5,DBL0H,19
+ sbc.f DBL0L,DBL0L,DBL1L
+ bic DBL1H,DBL0H,r5
+ bset r5,r5,20
+ sbc.f r4,r5,r7
+ beq_l .Large_cancel_sub
+ norm DBL1L,r4
+ bmsk r6,DBL1H,30
+.Lsub_done:
+ sub_s DBL1L,DBL1L,9
+ breq DBL1L,1,.Lsub_done_noshift
+ asl r5,DBL1L,20
+ sub_s DBL1L,DBL1L,1
+ brlo r6,r5,.Ldenorm_sub
+ sub DBL0H,DBL1H,r5
+.Lpast_denorm_sub:
+ neg_s DBL1H,DBL1L
+ lsr r6,r12,DBL1H
+ asl_s r12,r12,DBL1L
+ and r8,r6,1
+ add1.f 0,r8,r12
+ add.ne.f r12,r12,r12
+ asl r8,DBL0L,DBL1L
+ lsr r12,DBL0L,DBL1H
+ adc.f DBL0L,r8,r6
+ asl r5,r4,DBL1L
+ add_s DBL0H,DBL0H,r12
+ j_s.d [blink]
+ adc DBL0H,DBL0H,r5
+
+ .balign 4
+.Lno_denorm_large_shift:
+ breq.d r12,32,.Lshift32
+ bset_l DBL1H,DBL1H,20
+.Lfixed_denorm_large_shift:
+ neg r8,r12
+ asl r4,DBL1H,r8
+ lsr r5,DBL1L,r12
+ asl.f 0,DBL1L,r8
+ lsr DBL1L,DBL1H,r12
+ or r12,r4,r5
+ tst.eq r12,1
+ or.ne r12,r12,2
+ brlt.d r10,0,.Lsub
+ mov_s DBL1H,0
+ b_l .Ladd
+
+ ; If a denorm is produced without shifting, we have an exact result -
+ ; no need for rounding.
+ .balign 4
+.Ldenorm_sub:
+ lsr DBL1L,r6,20
+ xor DBL0H,r6,DBL1H
+ brne.d DBL1L,1,.Lpast_denorm_sub
+ sub_s DBL1L,DBL1L,1
+.Lsub_done_noshift:
+ add.f 0,r12,r12
+ btst.eq DBL0L,0
+ cmp.eq r12,r12
+ add.cs.f DBL0L,DBL0L,1
+ bclr r4,r4,20
+ j_s.d [blink]
+ adc DBL0H,DBL1H,r4
+
+ .balign 4
+.Ldenorm_small_shift:
+ brne.d r12,1,.Lfixed_denorm_small_shift
+ sub_l r12,r12,1
+ brlt r10,0,.Lsub
+.Ladd: ; bit 20 of DBL1H is clear and bit 0 of r12 does not matter
+ add.f DBL0L,DBL0L,DBL1L
+ add_s DBL1H,DBL1H,DBL0H
+ add.cs DBL1H,DBL1H,1
+ xor_l DBL0H,DBL0H,DBL1H
+ bbit0 DBL0H,20,.Lno_shiftdown
+ lsr.f DBL0H,DBL1H
+ and r4,DBL0L,2
+ bmsk DBL0H,DBL0H,18
+ sbc DBL0H,DBL1H,DBL0H
+ rrc.f DBL0L,DBL0L
+ or.f r12,r12,r4
+ cmp.eq r12,r12
+ add.cs.f DBL0L,DBL0L,1
+ bic.f 0,r9,DBL0H ; check for generating infinity with possible ...
+ jne.d [blink] ; ... non-zero fraction
+ add.cs DBL0H,DBL0H,1
+ mov_s DBL0L,0
+ bmsk DBL1H,DBL0H,19
+ j_s.d [blink]
+ bic_s DBL0H,DBL0H,DBL1H
+.Lno_shiftdown:
+ mov_s DBL0H,DBL1H
+ add.f 0,r12,r12
+ btst.eq DBL0L,0
+ cmp.eq r12,r12
+ add.cs.f DBL0L,DBL0L,1
+ j_s.d [blink]
+ add.cs DBL0H,DBL0H,1
+ .balign 4
+.Ldenorm:
+ bmsk DBL0H,DBL1H,30
+ lsr r12,DBL0H,20
+ xor_s DBL0H,DBL0H,DBL1H
+ sub_l DBL1L,r12,1
+ bgt .Lpast_denorm
+ j_s.d [blink]
+ add_l DBL0H,DBL0H,r4
+
+ .balign 4
+.Large_cancel:
+ ;DBL0L: mantissa DBL1H: sign & exponent
+ norm.f DBL1L,DBL0L
+ bmsk DBL0H,DBL1H,30
+ add_s DBL1L,DBL1L,22
+ mov.mi DBL1L,21
+ add_s r12,DBL1L,1
+ asl_s r12,r12,20
+ beq_s .Lret0
+ brhs.d DBL0H,r12,.Lpast_denorm_large_cancel
+ sub DBL0H,DBL1H,r12
+ bmsk DBL0H,DBL1H,30
+ lsr r12,DBL0H,20
+ xor_s DBL0H,DBL0H,DBL1H
+ sub.f DBL1L,r12,1
+ jle [blink]
+.Lpast_denorm_large_cancel:
+ rsub.f r7,DBL1L,32
+ lsr r7,DBL0L,r7
+ asl_s DBL0L,DBL0L,DBL1L
+ mov.ls r7,DBL0L
+ add_s DBL0H,DBL0H,r7
+ j_s.d [blink]
+ mov.ls DBL0L,0
+.Lret0:
+ j_s.d [blink]
+ mov_l DBL0H,0
+
+/* r4:DBL0L:r12 : unnormalized result fraction
+ DBL1H: result sign and exponent */
+/* When seeing large cancellation, only the topmost guard bit might be set. */
+ .balign 4
+.Large_cancel_sub:
+ norm.f DBL1L,DBL0L
+ bpnz.d 0f
+ bmsk DBL0H,DBL1H,30
+ mov r5,22<<20
+ bne.d 1f
+ mov_s DBL1L,21
+ bset r5,r5,5+20
+ add_s DBL1L,DBL1L,32
+ brne r12,0,1f
+ j_s.d [blink]
+ mov_l DBL0H,0
+ .balign 4
+0: add r5,DBL1L,23
+ asl r5,r5,20
+ add_s DBL1L,DBL1L,22
+1: brlo DBL0H,r5,.Ldenorm_large_cancel_sub
+ sub DBL0H,DBL1H,r5
+.Lpast_denorm_large_cancel_sub:
+ rsub.f r7,DBL1L,32
+ lsr r12,r12,r7
+ lsr r7,DBL0L,r7
+ asl_s DBL0L,DBL0L,DBL1L
+ add.ge DBL0H,DBL0H,r7
+ add_s DBL0L,DBL0L,r12
+ add.lt DBL0H,DBL0H,DBL0L
+ mov.eq DBL0L,r12
+ j_s.d [blink]
+ mov.lt DBL0L,0
+ .balign 4
+.Ldenorm_large_cancel_sub:
+ lsr r5,DBL0H,20
+ xor_s DBL0H,DBL0H,DBL1H
+ brgt.d r5,1,.Lpast_denorm_large_cancel_sub
+ sub DBL1L,r5,1
+ j_l [blink] ; denorm, no shift -> no rounding needed.
+
+/* r4: DBL0H & 0x7fffffff
+ r6: DBL1H & 0x7ff00000
+ r9: 0x7ff00000
+ r10: sign difference
+ r12: shift count (negative) */
+ .balign 4
+.Ldbl1_gt:
+ brhs r6,r9,.Lret_dbl1 ; inf or NaN
+ neg r8,r12
+ brhs r8,32,.Large_shift_dbl0
+.Lsmall_shift_dbl0:
+ breq.d r6,0,.Ldenorm_small_shift_dbl0
+ bmsk_s DBL0H,DBL0H,19
+ bset_s DBL0H,DBL0H,20
+.Lfixed_denorm_small_shift_dbl0:
+ asl r4,DBL0H,r12
+ lsr DBL0H,DBL0H,r8
+ lsr r5,DBL0L,r8
+ asl r12,DBL0L,r12
+ brge.d r10,0,.Ladd_dbl1_gt
+ or DBL0L,r4,r5
+/* subtract, abs(DBL0) < abs(DBL1) */
+/* DBL0H, DBL0L: fraction with explicit leading 1, shifted into place
+ DBL1H, DBL1L: original values
+ r6: orig. DBL1H & 0x7ff00000
+ r9: 0x7ff00000
+ r12: guard bits */
+ .balign 4
+.Lrsub:
+ neg.f r12,r12
+ bmsk r7,DBL1H,19
+ mov_s r5,DBL0H
+ sbc.f DBL0L,DBL1L,DBL0L
+ bic DBL1H,DBL1H,r7
+ bset r7,r7,20
+ sbc.f r4,r7,r5
+ beq_l .Large_cancel_sub
+ norm DBL1L,r4
+ b_l .Lsub_done ; note: r6 is already set up.
+
+.Lret_dbl1:
+ mov_s DBL0H,DBL1H
+ j_s.d [blink]
+ mov_l DBL0L,DBL1L
+ .balign 4
+.Ldenorm_small_shift_dbl0:
+ sub.f r8,r8,1
+ bne.d .Lfixed_denorm_small_shift_dbl0
+ add_s r12,r12,1
+ brlt r10,0,.Lrsub
+.Ladd_dbl1_gt: ; bit 20 of DBL0H is clear and bit 0 of r12 does not matter
+ add.f DBL0L,DBL0L,DBL1L
+ add_s DBL0H,DBL0H,DBL1H
+ add.cs DBL0H,DBL0H,1
+ xor DBL1H,DBL0H,DBL1H
+ bbit0 DBL1H,20,.Lno_shiftdown_dbl1_gt
+ lsr.f DBL1H,DBL0H
+ and r4,DBL0L,2
+ bmsk DBL1H,DBL1H,18
+ sbc DBL0H,DBL0H,DBL1H
+ rrc.f DBL0L,DBL0L
+ or.f r12,r12,r4
+ cmp.eq r12,r12
+ add.cs.f DBL0L,DBL0L,1
+ bic.f 0,r9,DBL0H ; check for generating infinity with possible ...
+ jne.d [blink] ; ... non-zero fraction
+ add.cs DBL0H,DBL0H,1
+ mov_s DBL0L,0
+ bmsk DBL1H,DBL0H,19
+ j_s.d [blink]
+ bic_s DBL0H,DBL0H,DBL1H
+.Lno_shiftdown_dbl1_gt:
+ add.f 0,r12,r12
+ btst.eq DBL0L,0
+ cmp.eq r12,r12
+ add.cs.f DBL0L,DBL0L,1
+ j_s.d [blink]
+ add.cs DBL0H,DBL0H,1
+
+ .balign 4
+.Large_shift_dbl0:
+ brhs r8,55,.Lret_dbl1
+ bmsk_s DBL0H,DBL0H,19
+ brne r6,0,.Lno_denorm_large_shift_dbl0
+ add_s r12,r12,1
+ brne.d r8,33,.Lfixed_denorm_large_shift_dbl0
+ sub r8,r8,1
+ bset_s DBL0H,DBL0H,20
+.Lshift32_dbl0:
+ mov_s r12,DBL0L
+ mov_s DBL0L,DBL0H
+ brlt.d r10,0,.Lrsub
+ mov_s DBL0H,0
+ b_s .Ladd_dbl1_gt
+
+ .balign 4
+.Lno_denorm_large_shift_dbl0:
+ breq.d r8,32,.Lshift32_dbl0
+ bset_l DBL0H,DBL0H,20
+.Lfixed_denorm_large_shift_dbl0:
+ asl r4,DBL0H,r12
+ lsr r5,DBL0L,r8
+ asl.f 0,DBL0L,r12
+ lsr DBL0L,DBL0H,r8
+ or r12,r4,r5
+ tst.eq r12,1
+ or.ne r12,r12,2
+ brlt.d r10,0,.Lrsub
+ mov_s DBL0H,0
+ b_l .Ladd_dbl1_gt
+ ENDFUNC(__adddf3)
+ ENDFUNC(__subdf3)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "arc-ieee-754.h"
+#if 0 /* DEBUG */
+ .global __addsf3
+ FUNC(__addsf3)
+ .balign 4
+__addsf3:
+ push_s blink
+ push_s r1
+ bl.d __addsf3_c
+ push_s r0
+ ld_s r1,[sp,4]
+ st_s r0,[sp,4]
+ bl.d __addsf3_asm
+ pop_s r0
+ pop_s r1
+ pop_s blink
+ cmp r0,r1
+ jeq_s [blink]
+ bl abort
+ ENDFUNC(__addsf3)
+ .global __subsf3
+ FUNC(__subsf3)
+ .balign 4
+__subsf3:
+ push_s blink
+ push_s r1
+ bl.d __subsf3_c
+ push_s r0
+ ld_s r1,[sp,4]
+ st_s r0,[sp,4]
+ bl.d __subsf3_asm
+ pop_s r0
+ pop_s r1
+ pop_s blink
+ cmp r0,r1
+ jeq_s [blink]
+ bl abort
+ ENDFUNC(__subsf3)
+#define __addsf3 __addsf3_asm
+#define __subsf3 __subsf3_asm
+#endif /* DEBUG */
+/* N.B. This is optimized for ARC700.
+ ARC600 has very different scheduling / instruction selection criteria. */
+
+/* inputs: r0, r1
+ output: r0
+ clobber: r1-r10, r12, flags */
+
+ .balign 4
+ .global __addsf3
+ .global __subsf3
+ FUNC(__addsf3)
+ FUNC(__subsf3)
+ .long 0x7f800000 ; exponent mask
+__subsf3:
+ bxor_l r1,r1,31
+__addsf3:
+ ld r9,[pcl,-8]
+ bmsk r4,r0,30
+ xor r10,r0,r1
+ and r6,r1,r9
+ sub.f r12,r4,r6
+ asr_s r12,r12,23
+ blo .Ldbl1_gt
+ brhs r4,r9,.Linf_nan
+ brne r12,0,.Lsmall_shift
+ brge r10,0,.Ladd_same_exp ; r12 == 0
+/* After subtracting, we need to normalize; when shifting to place the
+ leading 1 into position for the implicit 1 and adding that to DBL0,
+ we increment the exponent. Thus, we have to subtract one more than
+ the shift count from the exponent beforehand. Iff the exponent drops thus
+ below zero (before adding in the fraction with the leading one), we have
+ generated a denormal number. Denormal handling is basicallly reducing the
+ shift count so that we produce a zero exponent instead; FWIW, this way
+ the shift count can become zero (if we started out with exponent 1).
+ On the plus side, we don't need to check for denorm input, the result
+ of subtracing these looks just the same as denormals generated during
+ subtraction. */
+ bmsk r7,r1,30
+ breq r4,r7,.Lret0
+ sub.f r5,r4,r7
+ lsr r12,r4,23
+ neg.cs r5,r5
+ norm r3,r5
+ bmsk r2,r0,22
+ sub_s r3,r3,6
+ min r12,r12,r3
+ bic r1,r0,r2
+ sub_s r3,r12,1
+ asl_s r12,r12,23
+ asl r2,r5,r3
+ sub_s r1,r1,r12
+ add_s r0,r1,r2
+ j_s.d [blink]
+ bxor.cs r0,r0,31
+ .balign 4
+.Linf_nan:
+ ; If both inputs are inf, but with different signs, the result is NaN.
+ asr r12,r10,31
+ or_s r1,r1,r12
+ j_s.d [blink]
+ or.eq r0,r0,r1
+ .balign 4
+.Ladd_same_exp:
+ /* This is a special case because we can't test for need to shift
+ down by checking if bit 23 of DBL0 changes. OTOH, here we know
+ that we always need to shift down. */
+ ; adding the two floating point numbers together makes the sign
+ ; cancel out and apear as carry; the exponent is doubled, and the
+ ; fraction also in need of shifting left by one. The two implicit
+ ; ones of the sources make an implicit 1 of the result, again
+ ; non-existent in a place shifted by one.
+ add.f r0,r0,r1
+ btst_s r0,1
+ breq r6,0,.Ldenorm_add
+ add.ne r0,r0,1 ; round to even.
+ rrc r0,r0
+ bmsk r1,r9,23
+ add r0,r0,r1 ; increment exponent
+ bic.f 0,r9,r0; check for overflow -> infinity.
+ jne_l [blink]
+ mov_s r0,r9
+ j_s.d [blink]
+ bset.cs r0,r0,31
+
+.Ldenorm_add:
+ j_s.d [blink]
+ add r0,r4,r1
+
+.Lret_dbl0:
+ j_s [blink]
+
+ .balign 4
+.Lsmall_shift:
+ brhi r12,25,.Lret_dbl0
+ breq.d r6,0,.Ldenorm_small_shift
+ bmsk_s r1,r1,22
+ bset_s r1,r1,23
+.Lfixed_denorm_small_shift:
+ neg r8,r12
+ asl r5,r1,r8
+ brge.d r10,0,.Ladd
+ lsr_l r1,r1,r12
+/* subtract, abs(DBL0) > abs(DBL1) */
+/* DBL0: original values
+ DBL1: fraction with explicit leading 1, shifted into place
+ r4: orig. DBL0 & 0x7fffffff
+ r6: orig. DBL1 & 0x7f800000
+ r9: 0x7f800000
+ r10: orig. DBL0H ^ DBL1H
+ r5 : guard bits */
+ .balign 4
+.Lsub:
+ neg.f r12,r5
+ bmsk r3,r0,22
+ bset r5,r3,23
+ sbc.f r4,r5,r1
+ beq.d .Large_cancel_sub
+ bic r7,r0,r3
+ norm r3,r4
+ bmsk r6,r7,30
+.Lsub_done:
+ sub_s r3,r3,6
+ breq r3,1,.Lsub_done_noshift
+ asl r5,r3,23
+ sub_l r3,r3,1
+ brlo r6,r5,.Ldenorm_sub
+ sub r0,r7,r5
+ neg_s r1,r3
+ lsr.f r2,r12,r1
+ asl_s r12,r12,r3
+ btst_s r2,0
+ bmsk.eq.f r12,r12,30
+ asl r5,r4,r3
+ add_s r0,r0,r2
+ adc.ne r0,r0,0
+ j_s.d [blink]
+ add_l r0,r0,r5
+
+.Lret0:
+ j_s.d [blink]
+ mov_l r0,0
+
+ .balign 4
+.Ldenorm_small_shift:
+ brne.d r12,1,.Lfixed_denorm_small_shift
+ sub_s r12,r12,1
+ brlt.d r10,0,.Lsub
+ mov_s r5,r12 ; zero r5, and align following code
+.Ladd: ; Both bit 23 of DBL1 and bit 0 of r5 are clear.
+ bmsk r2,r0,22
+ add_s r2,r2,r1
+ bbit0.d r2,23,.Lno_shiftdown
+ add_s r0,r0,r1
+ bic.f 0,r9,r0; check for overflow -> infinity; eq : infinity
+ bmsk r1,r2,22
+ lsr.ne.f r2,r2,2; cc: even ; hi: might round down
+ lsr.ne r1,r1,1
+ rcmp.hi r5,1; hi : round down
+ bclr.hi r0,r0,0
+ j_l.d [blink]
+ sub_s r0,r0,r1
+
+/* r4: DBL0H & 0x7fffffff
+ r6: DBL1H & 0x7f800000
+ r9: 0x7f800000
+ r10: sign difference
+ r12: shift count (negative) */
+ .balign 4
+.Ldbl1_gt:
+ brhs r6,r9,.Lret_dbl1 ; inf or NaN
+ neg r8,r12
+ brhi r8,25,.Lret_dbl1
+.Lsmall_shift_dbl0:
+ breq.d r6,0,.Ldenorm_small_shift_dbl0
+ bmsk_s r0,r0,22
+ bset_s r0,r0,23
+.Lfixed_denorm_small_shift_dbl0:
+ asl r5,r0,r12
+ brge.d r10,0,.Ladd_dbl1_gt
+ lsr r0,r0,r8
+/* subtract, abs(DBL0) < abs(DBL1) */
+/* DBL0: fraction with explicit leading 1, shifted into place
+ DBL1: original value
+ r6: orig. DBL1 & 0x7f800000
+ r9: 0x7f800000
+ r5: guard bits */
+ .balign 4
+.Lrsub:
+ neg.f r12,r5
+ bmsk r5,r1,22
+ bic r7,r1,r5
+ bset r5,r5,23
+ sbc.f r4,r5,r0
+ bne.d .Lsub_done ; note: r6 is already set up.
+ norm r3,r4
+ /* Fall through */
+
+/* r4:r12 : unnormalized result fraction
+ r7: result sign and exponent */
+/* When seeing large cancellation, only the topmost guard bit might be set. */
+ .balign 4
+.Large_cancel_sub:
+ breq_s r12,0,.Lret0
+ sub r0,r7,24<<23
+ xor.f 0,r0,r7 ; test if exponent is negative
+ tst.pl r9,r0 ; test if exponent is zero
+ jpnz [blink] ; return if non-denormal result
+ bmsk r6,r7,30
+ lsr r3,r6,23
+ xor r0,r6,r7
+ sub_s r3,r3,24-22
+ j_s.d [blink]
+ bset r0,r0,r3
+
+ ; If a denorm is produced, we have an exact result -
+ ; no need for rounding.
+ .balign 4
+.Ldenorm_sub:
+ sub r3,r6,1
+ lsr.f r3,r3,23
+ xor r0,r6,r7
+ neg_s r1,r3
+ asl.ne r4,r4,r3
+ lsr_s r12,r12,r1
+ add_s r0,r0,r4
+ j_s.d [blink]
+ add.ne r0,r0,r12
+
+ .balign 4
+.Lsub_done_noshift:
+ add.f 0,r12,r12
+ btst.eq r4,0
+ bclr r4,r4,23
+ add r0,r7,r4
+ j_s.d [blink]
+ adc.ne r0,r0,0
+
+ .balign 4
+.Lno_shiftdown:
+ add.f 0,r5,r5
+ btst.eq r0,0
+ cmp.eq r5,r5
+ j_s.d [blink]
+ add.cs r0,r0,1
+
+.Lret_dbl1:
+ j_s.d [blink]
+ mov_l r0,r1
+ .balign 4
+.Ldenorm_small_shift_dbl0:
+ sub.f r8,r8,1
+ bne.d .Lfixed_denorm_small_shift_dbl0
+ add_s r12,r12,1
+ brlt.d r10,0,.Lrsub
+ mov r5,0
+.Ladd_dbl1_gt: ; both bit 23 of DBL0 and bit 0 of r5 are clear.
+ bmsk r2,r1,22
+ add_s r2,r2,r0
+ bbit0.d r2,23,.Lno_shiftdown_dbl1_gt
+ add_s r0,r1,r0
+ bic.f 0,r9,r0; check for overflow -> infinity; eq : infinity
+ bmsk r1,r2,22
+ lsr.ne.f r2,r2,2; cc: even ; hi: might round down
+ lsr.ne r1,r1,1
+ rcmp.hi r5,1; hi : round down
+ bclr.hi r0,r0,0
+ j_l.d [blink]
+ sub_s r0,r0,r1
+
+ .balign 4
+.Lno_shiftdown_dbl1_gt:
+ add.f 0,r5,r5
+ btst.eq r0,0
+ cmp.eq r5,r5
+ j_s.d [blink]
+ add.cs r0,r0,1
+ ENDFUNC(__addsf3)
+ ENDFUNC(__subsf3)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#ifdef __LITTLE_ENDIAN__
+#define DBL0L r0
+#define DBL0H r1
+#define DBL1L r2
+#define DBL1H r3
+#else
+#define DBL0L r1
+#define DBL0H r0
+#define DBL1L r3
+#define DBL1H r2
+#endif
+#define add_l add
+#define asr_l asr
+#define j_l j
+#define jne_l jne
+#define jeq_l jeq
+#define or_l or
+#define mov_l mov
+#define b_l b
+#define beq_l beq
+#define bne_l bne
+#define brne_l brne
+#define bset_l bset
+#define sub_l sub
+#define sub1_l sub1
+#define lsr_l lsr
+#define xor_l xor
+#define bic_l bic
+#define bmsk_l bmsk
+#define bxor_l bxor
+#define bcs_s blo_s
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+/*
+ to calculate a := b/x as b*y, with y := 1/x:
+ - x is in the range [1..2)
+ - calculate 15..18 bit inverse y0 using a table of approximating polynoms.
+ Precision is higher for polynoms used to evaluate input with larger
+ value.
+ - Do one newton-raphson iteration step to double the precision,
+ then multiply this with the divisor
+ -> more time to decide if dividend is subnormal
+ - the worst error propagation is on the side of the value range
+ with the least initial defect, thus giving us about 30 bits precision.
+ The truncation error for the either is less than 1 + x/2 ulp.
+ A 31 bit inverse can be simply calculated by using x with implicit 1
+ and chaining the multiplies. For a 32 bit inverse, we multiply y0^2
+ with the bare fraction part of x, then add in y0^2 for the implicit
+ 1 of x.
+ - If calculating a 31 bit inverse, the systematic error is less than
+ -1 ulp; likewise, for 32 bit, it is less than -2 ulp.
+ - If we calculate our seed with a 32 bit fraction, we can archive a
+ tentative result strictly better than -2 / +2.5 (1) ulp/128, i.e. we
+ only need to take the step to calculate the 2nd stage rest and
+ rounding adjust 1/32th of the time. However, if we use a 20 bit
+ fraction for the seed, the negative error can exceed -2 ulp/128, (2)
+ thus for a simple add / tst check, we need to do the 2nd stage
+ rest calculation/ rounding adjust 1/16th of the time.
+ (1): The inexactness of the 32 bit inverse contributes an error in the
+ range of (-1 .. +(1+x/2) ) ulp/128. Leaving out the low word of the
+ rest contributes an error < +1/x ulp/128 . In the interval [1,2),
+ x/2 + 1/x <= 1.5 .
+ (2): Unless proven otherwise. I have not actually looked for an
+ example where -2 ulp/128 is exceeded, and my calculations indicate
+ that the excess, if existent, is less than -1/512 ulp.
+ ??? The algorithm is still based on the ARC700 optimized code.
+ Maybe we could make better use of 32x16 bit multiply, or 64 bit multiply
+ results.
+ */
+#include "../arc-ieee-754.h"
+#define mlo acc2
+#define mhi acc1
+#define mul64(b,c) mullw 0,b,c` machlw 0,b,c
+#define mulu64(b,c) mululw 0,b,c` machulw 0,b,c
+
+/* N.B. fp-bit.c does double rounding on denormal numbers. */
+#if 0 /* DEBUG */
+ .global __divdf3
+ FUNC(__divdf3)
+ .balign 4
+__divdf3:
+ push_s blink
+ push_s r2
+ push_s r3
+ push_s r0
+ bl.d __divdf3_c
+ push_s r1
+ ld_s r2,[sp,12]
+ ld_s r3,[sp,8]
+ st_s r0,[sp,12]
+ st_s r1,[sp,8]
+ pop_s r1
+ bl.d __divdf3_asm
+ pop_s r0
+ pop_s r3
+ pop_s r2
+ pop_s blink
+ cmp r0,r2
+ cmp.eq r1,r3
+ jeq_s [blink]
+ and r12,DBL0H,DBL1H
+ bic.f 0,0x7ff80000,r12 ; both NaN -> OK
+ jeq_s [blink]
+ bl abort
+ ENDFUNC(__divdf3)
+#define __divdf3 __divdf3_asm
+#endif /* DEBUG */
+
+ FUNC(__divdf3)
+ .balign 4
+.L7ff00000:
+ .long 0x7ff00000
+.Ldivtab:
+ .long 0xfc0fffe1
+ .long 0xf46ffdfb
+ .long 0xed1ffa54
+ .long 0xe61ff515
+ .long 0xdf7fee75
+ .long 0xd91fe680
+ .long 0xd2ffdd52
+ .long 0xcd1fd30c
+ .long 0xc77fc7cd
+ .long 0xc21fbbb6
+ .long 0xbcefaec0
+ .long 0xb7efa100
+ .long 0xb32f92bf
+ .long 0xae8f83b7
+ .long 0xaa2f7467
+ .long 0xa5ef6479
+ .long 0xa1cf53fa
+ .long 0x9ddf433e
+ .long 0x9a0f3216
+ .long 0x965f2091
+ .long 0x92df0f11
+ .long 0x8f6efd05
+ .long 0x8c1eeacc
+ .long 0x88eed876
+ .long 0x85dec615
+ .long 0x82eeb3b9
+ .long 0x800ea10b
+ .long 0x7d3e8e0f
+ .long 0x7a8e7b3f
+ .long 0x77ee6836
+ .long 0x756e5576
+ .long 0x72fe4293
+ .long 0x709e2f93
+ .long 0x6e4e1c7f
+ .long 0x6c0e095e
+ .long 0x69edf6c5
+ .long 0x67cde3a5
+ .long 0x65cdd125
+ .long 0x63cdbe25
+ .long 0x61ddab3f
+ .long 0x600d991f
+ .long 0x5e3d868c
+ .long 0x5c6d7384
+ .long 0x5abd615f
+ .long 0x590d4ecd
+ .long 0x576d3c83
+ .long 0x55dd2a89
+ .long 0x545d18e9
+ .long 0x52dd06e9
+ .long 0x516cf54e
+ .long 0x4ffce356
+ .long 0x4e9cd1ce
+ .long 0x4d3cbfec
+ .long 0x4becae86
+ .long 0x4aac9da4
+ .long 0x496c8c73
+ .long 0x483c7bd3
+ .long 0x470c6ae8
+ .long 0x45dc59af
+ .long 0x44bc4915
+ .long 0x43ac3924
+ .long 0x428c27fb
+ .long 0x418c187a
+ .long 0x407c07bd
+
+__divdf3_support: /* This label makes debugger output saner. */
+ .balign 4
+.Ldenorm_dbl1:
+ brge r6, \
+ 0x43500000,.Linf_NaN ; large number / denorm -> Inf
+ bmsk.f r12,DBL1H,19
+ mov.eq r12,DBL1L
+ mov.eq DBL1L,0
+ sub.eq r7,r7,32
+ norm.f r11,r12 ; flag for x/0 -> Inf check
+ beq_s .Linf_NaN
+ mov.mi r11,0
+ add.pl r11,r11,1
+ add_s r12,r12,r12
+ asl r8,r12,r11
+ rsub r12,r11,31
+ lsr r12,DBL1L,r12
+ tst_s DBL1H,DBL1H
+ or r8,r8,r12
+ lsr r4,r8,26
+ lsr DBL1H,r8,12
+ ld.as r4,[r10,r4]
+ bxor.mi DBL1H,DBL1H,31
+ sub r11,r11,11
+ asl DBL1L,DBL1L,r11
+ sub r11,r11,1
+ mulu64 (r4,r8)
+ sub r7,r7,r11
+ b.d .Lpast_denorm_dbl1
+ asl r7,r7,20
+
+.Linf_NaN:
+ tst_s DBL0L,DBL0L ; 0/0 -> NaN
+ xor_s DBL1H,DBL1H,DBL0H
+ bclr.eq.f DBL0H,DBL0H,31
+ bmsk DBL0H,DBL1H,30
+ xor_s DBL0H,DBL0H,DBL1H
+ sub.eq DBL0H,DBL0H,1
+ mov_s DBL0L,0
+ j_s.d [blink]
+ or DBL0H,DBL0H,r9
+ .balign 4
+.Lret0_2:
+ xor_s DBL1H,DBL1H,DBL0H
+ mov_s DBL0L,0
+ bmsk DBL0H,DBL1H,30
+ j_s.d [blink]
+ xor_s DBL0H,DBL0H,DBL1H
+ .balign 4
+ .global __divdf3
+/* N.B. the spacing between divtab and the sub3 to get its address must
+ be a multiple of 8. */
+__divdf3:
+ asl r8,DBL1H,12
+ lsr r4,r8,26
+ sub3 r10,pcl,51;(.-.Ldivtab) >> 3
+ ld.as r9,[pcl,-104]; [pcl,(-((.-.L7ff00000) >> 2))] ; 0x7ff00000
+ ld.as r4,[r10,r4]
+ lsr r12,DBL1L,20
+ and.f r7,DBL1H,r9
+ or r8,r8,r12
+ mulu64 (r4,r8)
+ beq.d .Ldenorm_dbl1
+.Lpast_denorm_dbl1:
+ and.f r6,DBL0H,r9
+ breq.d r7,r9,.Linf_nan_dbl1
+ asl r4,r4,12
+ sub r4,r4,mhi
+ mululw 0,r4,r4
+ machulw r5,r4,r4
+ bne.d .Lnormal_dbl0
+ lsr r8,r8,1
+
+ .balign 4
+.Ldenorm_dbl0:
+ bmsk.f r12,DBL0H,19
+ ; wb stall
+ mov.eq r12,DBL0L
+ sub.eq r6,r6,32
+ norm.f r11,r12 ; flag for 0/x -> 0 check
+ brge r7, \
+ 0x43500000, .Lret0_2 ; denorm/large number -> 0
+ beq_s .Lret0_2
+ mov.mi r11,0
+ add.pl r11,r11,1
+ asl r12,r12,r11
+ sub r6,r6,r11
+ add.f 0,r6,31
+ lsr r10,DBL0L,r6
+ mov.mi r10,0
+ add r6,r6,11+32
+ neg.f r11,r6
+ asl DBL0L,DBL0L,r11
+ mov.pl DBL0L,0
+ sub r6,r6,32-1
+ b.d .Lpast_denorm_dbl0
+ asl r6,r6,20
+
+ .balign 4
+.Linf_nan_dbl1: ; 0/Inf -> NaN Inf/Inf -> NaN x/Inf-> 0 x/NaN -> NaN
+ or.f 0,r6,DBL0L
+ cmp.ne r6,r9
+ not_s DBL0L,DBL1H
+ sub_s.ne DBL0L,DBL0L,DBL0L
+ tst_s DBL0H,DBL0H
+ add_s DBL0H,DBL1H,DBL0L
+ j_s.d [blink]
+ bxor.mi DBL0H,DBL0H,31
+
+ .balign 4
+.Lnormal_dbl0:
+ breq.d r6,r9,.Linf_nan_dbl0
+ asl r12,DBL0H,11
+ lsr r10,DBL0L,21
+.Lpast_denorm_dbl0:
+ bset r8,r8,31
+ mulu64 (r5,r8)
+ add_s r12,r12,r10
+ bset r5,r12,31
+ cmp r5,r8
+ cmp.eq DBL0L,DBL1L
+ lsr.cc r5,r5,1
+ sub r4,r4,mhi ; u1.31 inverse, about 30 bit
+ mululw 0,r5,r4
+ machulw r11,r5,r4 ; result fraction highpart
+ lsr r8,r8,2 ; u3.29
+ add r5,r6, /* wait for immediate */ \
+ 0x3fe00000
+ mulu64 (r11,r8) ; u-28.31
+ asl_s DBL1L,DBL1L,9 ; u-29.23:9
+ sbc r6,r5,r7
+ mov r12,mlo ; u-28.31
+ mulu64 (r11,DBL1L) ; mhi: u-28.23:9
+ add.cs DBL0L,DBL0L,DBL0L
+ asl_s DBL0L,DBL0L,6 ; u-26.25:7
+ asl r10,r11,23
+ sub_l DBL0L,DBL0L,r12
+ lsr r7,r11,9
+ sub r5,DBL0L,mhi ; rest msw ; u-26.31:0
+ mul64 (r5,r4) ; mhi: result fraction lowpart
+ xor.f 0,DBL0H,DBL1H
+ and DBL0H,r6,r9
+ add_s DBL0H,DBL0H,r7
+ bclr r12,r9,20 ; 0x7fe00000
+ brhs.d r6,r12,.Linf_denorm
+ bxor.mi DBL0H,DBL0H,31
+ add.f r12,mhi,0x11
+ asr r9,r12,5
+ sub.mi DBL0H,DBL0H,1
+ add.f DBL0L,r9,r10
+ tst r12,0x1c
+ jne.d [blink]
+ add.cs DBL0H,DBL0H,1
+ /* work out exact rounding if we fall through here. */
+ /* We know that the exact result cannot be represented in double
+ precision. Find the mid-point between the two nearest
+ representable values, multiply with the divisor, and check if
+ the result is larger than the dividend. Since we want to know
+ only the sign bit, it is sufficient to calculate only the
+ highpart of the lower 64 bits. */
+ mulu64 (r11,DBL1L) ; rest before considering r12 in r5 : -mlo
+ sub.f DBL0L,DBL0L,1
+ asl r12,r9,2 ; u-22.30:2
+ sub.cs DBL0H,DBL0H,1
+ sub.f r12,r12,2
+ mov r10,mlo ; rest before considering r12 in r5 : -r10
+ mululw 0,r12,DBL1L
+ machulw r7,r12,DBL1L ; mhi: u-51.32
+ asl r5,r5,25 ; s-51.7:25
+ lsr r10,r10,7 ; u-51.30:2
+ mulu64 (r12,r8) ; mlo: u-51.31:1
+ sub r5,r5,r10
+ add.mi r5,r5,DBL1L ; signed multiply adjust for r12*DBL1L
+ bset r7,r7,0 ; make sure that the result is not zero, and that
+ sub r5,r5,r7 ; a highpart zero appears negative
+ sub.f r5,r5,mlo ; rest msw
+ add.pl.f DBL0L,DBL0L,1
+ j_s.d [blink]
+ add.eq DBL0H,DBL0H,1
+
+.Linf_nan_dbl0:
+ tst_s DBL1H,DBL1H
+ j_s.d [blink]
+ bxor.mi DBL0H,DBL0H,31
+ .balign 4
+.Linf_denorm:
+ lsr r12,r6,28
+ brlo.d r12,0xc,.Linf
+.Ldenorm:
+ asr r6,r6,20
+ neg r9,r6
+ mov_s DBL0H,0
+ brhs.d r9,54,.Lret0
+ bxor.mi DBL0H,DBL0H,31
+ add r12,mhi,1
+ and r12,r12,-4
+ rsub r7,r6,5
+ asr r10,r12,28
+ bmsk r4,r12,27
+ min r7,r7,31
+ asr DBL0L,r4,r7
+ add DBL1H,r11,r10
+ abs.f r10,r4
+ sub.mi r10,r10,1
+ add.f r7,r6,32-5
+ asl r4,r4,r7
+ mov.mi r4,r10
+ add.f r10,r6,23
+ rsub r7,r6,9
+ lsr r7,DBL1H,r7
+ asl r10,DBL1H,r10
+ or.pnz DBL0H,DBL0H,r7
+ or.mi r4,r4,r10
+ mov.mi r10,r7
+ add.f DBL0L,r10,DBL0L
+ add.cs.f DBL0H,DBL0H,1 ; carry clear after this point
+ bxor.f 0,r4,31
+ add.pnz.f DBL0L,DBL0L,1
+ add.cs.f DBL0H,DBL0H,1
+ jne_s [blink]
+ /* Calculation so far was not conclusive; calculate further rest. */
+ mulu64 (r11,DBL1L) ; rest before considering r12 in r5 : -mlo
+ asr.f r12,r12,3
+ asl r5,r5,25 ; s-51.7:25
+ mov r11,mlo ; rest before considering r12 in r5 : -r11
+ mulu64 (r12,r8) ; u-51.31:1
+ and r9,DBL0L,1 ; tie-breaker: round to even
+ lsr r11,r11,7 ; u-51.30:2
+ mov DBL1H,mlo ; u-51.31:1
+ mulu64 (r12,DBL1L) ; u-51.62:2
+ sub.mi r11,r11,DBL1L ; signed multiply adjust for r12*DBL1L
+ add_s DBL1H,DBL1H,r11
+ sub DBL1H,DBL1H,r5 ; -rest msw
+ add_s DBL1H,DBL1H,mhi ; -rest msw
+ add.f 0,DBL1H,DBL1H ; can't ror.f by 32 :-(
+ tst_s DBL1H,DBL1H
+ cmp.eq mlo,r9
+ add.cs.f DBL0L,DBL0L,1
+ j_s.d [blink]
+ add.cs DBL0H,DBL0H,1
+
+.Lret0:
+ /* return +- 0 */
+ j_s.d [blink]
+ mov_s DBL0L,0
+.Linf:
+ mov_s DBL0H,r9
+ mov_s DBL0L,0
+ j_s.d [blink]
+ bxor.mi DBL0H,DBL0H,31
+ ENDFUNC(__divdf3)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+/*
+ - calculate 15..18 bit inverse using a table of approximating polynoms.
+ precision is higher for polynoms used to evaluate input with larger
+ value.
+ - do one newton-raphson iteration step to double the precision,
+ then multiply this with the divisor
+ -> more time to decide if dividend is subnormal
+ - the worst error propagation is on the side of the value range
+ with the least initial defect, thus giving us about 30 bits precision.
+ */
+#include "../arc-ieee-754.h"
+#define mlo acc2
+#define mhi acc1
+#define mul64(b,c) mullw 0,b,c` machlw 0,b,c
+#define mulu64(b,c) mululw 0,b,c` machulw 0,b,c
+
+#if 0 /* DEBUG */
+ .global __divsf3
+ FUNC(__divsf3)
+ .balign 4
+__divsf3:
+ push_s blink
+ push_s r1
+ bl.d __divsf3_c
+ push_s r0
+ ld_s r1,[sp,4]
+ st_s r0,[sp,4]
+ bl.d __divsf3_asm
+ pop_s r0
+ pop_s r1
+ pop_s blink
+ cmp r0,r1
+#if 1
+ bne abort
+ jeq_s [blink]
+ b abort
+#else
+ bne abort
+ j_s [blink]
+#endif
+ ENDFUNC(__divsf3)
+#define __divsf3 __divsf3_asm
+#endif /* DEBUG */
+
+ FUNC(__divsf3)
+ .balign 4
+.Ldivtab:
+ .long 0xfc0ffff0
+ .long 0xf46ffefd
+ .long 0xed1ffd2a
+ .long 0xe627fa8e
+ .long 0xdf7ff73b
+ .long 0xd917f33b
+ .long 0xd2f7eea3
+ .long 0xcd1fe986
+ .long 0xc77fe3e7
+ .long 0xc21fdddb
+ .long 0xbcefd760
+ .long 0xb7f7d08c
+ .long 0xb32fc960
+ .long 0xae97c1ea
+ .long 0xaa27ba26
+ .long 0xa5e7b22e
+ .long 0xa1cfa9fe
+ .long 0x9ddfa1a0
+ .long 0x9a0f990c
+ .long 0x9667905d
+ .long 0x92df878a
+ .long 0x8f6f7e84
+ .long 0x8c27757e
+ .long 0x88f76c54
+ .long 0x85df630c
+ .long 0x82e759c5
+ .long 0x8007506d
+ .long 0x7d3f470a
+ .long 0x7a8f3da2
+ .long 0x77ef341e
+ .long 0x756f2abe
+ .long 0x72f7212d
+ .long 0x709717ad
+ .long 0x6e4f0e44
+ .long 0x6c1704d6
+ .long 0x69e6fb44
+ .long 0x67cef1d7
+ .long 0x65c6e872
+ .long 0x63cedf18
+ .long 0x61e6d5cd
+ .long 0x6006cc6d
+ .long 0x5e36c323
+ .long 0x5c76b9f3
+ .long 0x5abeb0b7
+ .long 0x5916a79b
+ .long 0x57769e77
+ .long 0x55de954d
+ .long 0x54568c4e
+ .long 0x52d6834d
+ .long 0x51667a7f
+ .long 0x4ffe71b5
+ .long 0x4e9e68f1
+ .long 0x4d466035
+ .long 0x4bf65784
+ .long 0x4aae4ede
+ .long 0x496e4646
+ .long 0x48363dbd
+ .long 0x47063547
+ .long 0x45de2ce5
+ .long 0x44be2498
+ .long 0x43a61c64
+ .long 0x4296144a
+ .long 0x41860c0e
+ .long 0x407e03ee
+.L7f800000:
+ .long 0x7f800000
+ .balign 4
+ .global __divsf3_support
+__divsf3_support:
+.Linf_NaN:
+ bclr.f 0,r0,31 ; 0/0 -> NaN
+ xor_s r0,r0,r1
+ bmsk r1,r0,30
+ bic_s r0,r0,r1
+ sub.eq r0,r0,1
+ j_s.d [blink]
+ or r0,r0,r9
+.Lret0:
+ xor_s r0,r0,r1
+ bmsk r1,r0,30
+ j_s.d [blink]
+ bic_s r0,r0,r1
+/* N.B. the spacing between divtab and the sub3 to get its address must
+ be a multiple of 8. */
+__divsf3:
+ ld.as r9,[pcl,-9]; [pcl,(-((.-.L7f800000) >> 2))] ; 0x7f800000
+ sub3 r3,pcl,37;(.-.Ldivtab) >> 3
+ lsr r2,r1,17
+ and.f r11,r1,r9
+ bmsk r5,r2,5
+ beq.d .Ldenorm_fp1
+ asl r6,r1,8
+ and.f r2,r0,r9
+ ld.as r5,[r3,r5]
+ asl r4,r1,9
+ bset r6,r6,31
+ breq.d r11,r9,.Linf_nan_fp1
+.Lpast_denorm_fp1:
+ mululw 0,r5,r4
+ machulw r8,r5,r4
+ breq.d r2,r9,.Linf_nan_fp0
+ asl r5,r5,13
+ sub r7,r5,r8
+ mululw 0,r7,r6
+ machulw r8,r7,r6
+ beq.d .Ldenorm_fp0
+ asl r12,r0,8
+ mulu64 (r8,r7)
+ bset r3,r12,31
+.Lpast_denorm_fp0:
+ cmp_s r3,r6
+ lsr.cc r3,r3,1
+ add_s r2,r2, /* wait for immediate */ \
+ 0x3f000000
+ sub r7,r7,mhi ; u1.31 inverse, about 30 bit
+ mulu64 (r3,r7)
+ sbc r2,r2,r11
+ xor.f 0,r0,r1
+ and r0,r2,r9
+ bclr r3,r9,23 ; 0x7f000000
+ brhs.d r2,r3,.Linf_denorm
+ bxor.mi r0,r0,31
+.Lpast_denorm:
+ add r3,mhi,0x22 ; round to nearest or higher
+ tst r3,0x3c ; check if rounding was unsafe
+ lsr r3,r3,6
+ jne.d [blink] ; return if rounding was safe.
+ add_s r0,r0,r3
+ /* work out exact rounding if we fall through here. */
+ /* We know that the exact result cannot be represented in single
+ precision. Find the mid-point between the two nearest
+ representable values, multiply with the divisor, and check if
+ the result is larger than the dividend. */
+ add_s r3,r3,r3
+ sub_s r3,r3,1
+ mulu64 (r3,r6)
+ asr.f 0,r0,1 ; for round-to-even in case this is a denorm
+ rsub r2,r9,25
+ asl_s r12,r12,r2
+ sub.f 0,r12,mlo
+ j_s.d [blink]
+ sub.mi r0,r0,1
+.Linf_nan_fp1:
+ lsr_s r0,r0,31
+ bmsk.f 0,r1,22
+ asl_s r0,r0,31
+ bne_s 0f ; inf/inf -> nan
+ brne r2,r9,.Lsigned0 ; x/inf -> 0, but x/nan -> nan
+0: j_s.d [blink]
+ mov r0,-1
+.Lsigned0:
+.Linf_nan_fp0:
+ tst_s r1,r1
+ j_s.d [blink]
+ bxor.mi r0,r0,31
+ .balign 4
+ .global __divsf3
+/* For denormal results, it is possible that an exact result needs
+ rounding, and thus the round-to-even rule has to come into play. */
+.Linf_denorm:
+ brlo r2,0xc0000000,.Linf
+.Ldenorm:
+ asr_s r2,r2,23
+ bic r0,r0,r9
+ neg r9,r2
+ brlo.d r9,25,.Lpast_denorm
+ lsr r3,mlo,r9
+ /* Fall through: return +- 0 */
+ j_s [blink]
+.Linf:
+ j_s.d [blink]
+ or r0,r0,r9
+ .balign 4
+.Ldenorm_fp1:
+ norm.f r12,r6 ; flag for x/0 -> Inf check
+ add r6,r6,r6
+ rsub r5,r12,16
+ ror r5,r1,r5
+ bmsk r5,r5,5
+ bic.ne.f 0, \
+ 0x60000000,r0 ; large number / denorm -> Inf
+ ld.as r5,[r3,r5]
+ asl r6,r6,r12
+ beq.d .Linf_NaN
+ and.f r2,r0,r9
+ add r4,r6,r6
+ asl_s r12,r12,23
+ bne.d .Lpast_denorm_fp1
+ add_s r2,r2,r12
+.Ldenorm_fp0:
+ mulu64 (r8,r7)
+ bclr r12,r12,31
+ norm.f r3,r12 ; flag for 0/x -> 0 check
+ bic.ne.f 0,0x60000000,r1 ; denorm/large number -> 0
+ beq_s .Lret0
+ asl_s r12,r12,r3
+ asl_s r3,r3,23
+ add_s r12,r12,r12
+ add r11,r11,r3
+ b.d .Lpast_denorm_fp0
+ mov_s r3,r12
+ ENDFUNC(__divsf3)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "../arc-ieee-754.h"
+
+#if 0 /* DEBUG */
+ .global __muldf3
+ .balign 4
+__muldf3:
+ push_s blink
+ push_s r2
+ push_s r3
+ push_s r0
+ bl.d __muldf3_c
+ push_s r1
+ ld_s r2,[sp,12]
+ ld_s r3,[sp,8]
+ st_s r0,[sp,12]
+ st_s r1,[sp,8]
+ pop_s r1
+ bl.d __muldf3_asm
+ pop_s r0
+ pop_s r3
+ pop_s r2
+ pop_s blink
+ cmp r0,r2
+ cmp.eq r1,r3
+ jeq_s [blink]
+ b abort
+#define __muldf3 __muldf3_asm
+#endif /* DEBUG */
+
+__muldf3_support: /* This label makes debugger output saner. */
+ .balign 4
+ FUNC(__muldf3)
+.Ldenorm_2:
+ breq.d DBL1L,0,.Lret0_2 ; 0 input -> 0 output
+ norm.f r12,DBL1L
+ mov.mi r12,21
+ add.pl r12,r12,22
+ neg r11,r12
+ asl_s r12,r12,20
+ lsr.f DBL1H,DBL1L,r11
+ ror DBL1L,DBL1L,r11
+ sub_s DBL0H,DBL0H,r12
+ mov.eq DBL1H,DBL1L
+ sub_l DBL1L,DBL1L,DBL1H
+ /* Fall through. */
+ .global __muldf3
+ .balign 4
+__muldf3:
+ mululw 0,DBL0L,DBL1L
+ machulw r4,DBL0L,DBL1L
+ ld.as r9,[pcl,0x67] ; ((.L7ff00000-.+2)/4)]
+ bmsk r6,DBL0H,19
+ bset r6,r6,20
+ mov r8,acc2
+ mululw 0,r4,1
+ and r11,DBL0H,r9
+ breq.d r11,0,.Ldenorm_dbl0
+ and r12,DBL1H,r9
+ breq.d r12,0,.Ldenorm_dbl1
+ maclw 0,r6,DBL1L
+ machulw 0,r6,DBL1L
+ breq.d r11,r9,.Linf_nan
+ bmsk r10,DBL1H,19
+ breq.d r12,r9,.Linf_nan
+ bset r10,r10,20
+ maclw 0,r10,DBL0L
+ machulw r5,r10,DBL0L
+ add_s r12,r12,r11 ; add exponents
+ mov r4,acc2
+ mululw 0,r5,1
+ maclw 0,r6,r10
+ machulw r7,r6,r10 ; fraction product in r7:acc2:r4:r8
+ tst r8,r8
+ bclr r8,r9,30 ; 0x3ff00000
+ bset.ne r4,r4,0 ; put least significant word into sticky bit
+ bclr r6,r9,20 ; 0x7fe00000
+ lsr.f r10,r7,9
+ rsub.eq r8,r8,r9 ; 0x40000000
+ sub r12,r12,r8 ; subtract bias + implicit 1
+ brhs.d r12,r6,.Linf_denorm
+ rsub r10,r10,12
+.Lshift_frac:
+ neg r8,r10
+ asl r6,r4,r10
+ lsr DBL0L,r4,r8
+ add.f 0,r6,r6
+ btst.eq DBL0L,0
+ cmp.eq r4,r4 ; round to nearest / round to even
+ asl r4,acc2,r10
+ lsr r5,acc2,r8
+ adc.f DBL0L,DBL0L,r4
+ xor.f 0,DBL0H,DBL1H
+ asl r7,r7,r10
+ add_s r12,r12,r5
+ adc DBL0H,r12,r7
+ j_s.d [blink]
+ bset.mi DBL0H,DBL0H,31
+
+/* N.B. This is optimized for ARC700.
+ ARC600 has very different scheduling / instruction selection criteria. */
+
+/* If one number is denormal, subtract some from the exponent of the other
+ one (if the other exponent is too small, return 0), and normalize the
+ denormal. Then re-run the computation. */
+.Lret0_2:
+ lsr_s DBL0H,DBL0H,31
+ asl_s DBL0H,DBL0H,31
+ j_s.d [blink]
+ mov_s DBL0L,0
+ .balign 4
+.Ldenorm_dbl0:
+ mov_s r12,DBL0L
+ mov_s DBL0L,DBL1L
+ mov_s DBL1L,r12
+ mov_s r12,DBL0H
+ mov_s DBL0H,DBL1H
+ mov_s DBL1H,r12
+ and r11,DBL0H,r9
+.Ldenorm_dbl1:
+ brhs r11,r9,.Linf_nan
+ brhs 0x3ca00001,r11,.Lret0
+ sub_s DBL0H,DBL0H,DBL1H
+ bmsk.f DBL1H,DBL1H,30
+ add_s DBL0H,DBL0H,DBL1H
+ beq.d .Ldenorm_2
+ norm r12,DBL1H
+ sub_s r12,r12,10
+ asl r5,r12,20
+ asl_s DBL1H,DBL1H,r12
+ sub DBL0H,DBL0H,r5
+ neg r5,r12
+ lsr r6,DBL1L,r5
+ asl_s DBL1L,DBL1L,r12
+ b.d __muldf3
+ add_s DBL1H,DBL1H,r6
+
+.Lret0: xor_s DBL0H,DBL0H,DBL1H
+ bclr DBL1H,DBL0H,31
+ xor_s DBL0H,DBL0H,DBL1H
+ j_s.d [blink]
+ mov_s DBL0L,0
+
+ .balign 4
+.Linf_nan:
+ bclr r12,DBL1H,31
+ xor_s DBL1H,DBL1H,DBL0H
+ bclr_s DBL0H,DBL0H,31
+ max r8,DBL0H,r12 ; either NaN -> NaN ; otherwise inf
+ or.f 0,DBL0H,DBL0L
+ mov_s DBL0L,0
+ or.ne.f DBL1L,DBL1L,r12
+ not_s DBL0H,DBL0L ; inf * 0 -> NaN
+ mov.ne DBL0H,r8
+ tst_s DBL1H,DBL1H
+ j_s.d [blink]
+ bset.mi DBL0H,DBL0H,31
+
+/* We have checked for infinitey / NaN input before, and transformed
+ denormalized inputs into normalized inputs. Thus, the worst case
+ exponent overflows are:
+ 1 + 1 - 0x400 == 0xc02 : maximum underflow
+ 0x7fe + 0x7fe - 0x3ff == 0xbfd ; maximum overflow
+ N.B. 0x7e and 0x7f are also values for overflow.
+
+ If (r12 <= -54), we have an underflow to zero. */
+ .balign 4
+.Linf_denorm:
+ lsr r6,r12,28
+ brlo.d r6,0xc,.Linf
+ asr r6,r12,20
+ add.f r10,r10,r6
+ brgt.d r10,0,.Lshift_frac
+ mov_s r12,0
+ beq.d .Lround_frac
+ add r10,r10,32
+.Lshift32_frac:
+ tst r4,r4
+ mov r4,acc2
+ bset.ne r4,r4,1
+ mululw 0,r7,1
+ brge.d r10,1,.Lshift_frac
+ mov r7,0
+ breq.d r10,0,.Lround_frac
+ add r10,r10,32
+ brgt r10,21,.Lshift32_frac
+ b_s .Lret0
+
+.Lround_frac:
+ add.f 0,r4,r4
+ btst.eq acc2,0
+ mov_s DBL0L,acc2
+ mov_s DBL0H,r7
+ adc.eq.f DBL0L,DBL0L,0
+ j_s.d [blink]
+ adc.eq DBL0H,DBL0H,0
+
+.Linf: mov_s DBL0L,0
+ xor.f DBL1H,DBL1H,DBL0H
+ mov_s DBL0H,r9
+ j_s.d [blink]
+ bset.mi DBL0H,DBL0H,31
+ ENDFUNC(__muldf3)
+
+ .balign 4
+.L7ff00000:
+ .long 0x7ff00000
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "../arc-ieee-754.h"
+
+#if 0 /* DEBUG */
+ .global __mulsf3
+ FUNC(__mulsf3)
+ .balign 4
+__mulsf3:
+ push_s blink
+ push_s r1
+ bl.d __mulsf3_c
+ push_s r0
+ ld_s r1,[sp,4]
+ st_s r0,[sp,4]
+ bl.d __mulsf3_asm
+ pop_s r0
+ pop_s r1
+ pop_s blink
+ cmp r0,r1
+ jeq_s [blink]
+ and r12,r0,r1
+ bic.f 0,0x7f800000,r12
+ bne 0f
+ bmsk.f 0,r0,22
+ bmsk.ne.f r1,r1,22
+ jne_s [blink] ; both NaN -> OK
+0: bl abort
+ ENDFUNC(__mulsf3)
+#define __mulsf3 __mulsf3_asm
+#endif /* DEBUG */
+
+ .balign 4
+ .global __mulsf3
+ FUNC(__mulsf3)
+__mulsf3:
+ ld.as r9,[pcl,80]; [pcl,((.L7f800000-.+2)/4)]
+ bmsk r4,r1,22
+ bset r2,r0,23
+ asl_s r2,r2,8
+ bset r3,r4,23
+ and r11,r0,r9
+ breq.d r11,0,.Ldenorm_dbl0
+ and r12,r1,r9
+ breq.d r12,0,.Ldenorm_dbl1
+ xor_s r0,r0,r1
+ mululw 0,r2,r3
+ machulw r6,r2,r3
+ breq.d r11,r9,.Linf_nan_dbl0
+ ld.as r4,[pcl,69]; [pcl,((.L7fffffff-.+2)/4)]
+ breq.d r12,r9,.Linf_nan_dbl1
+.Lpast_denorm:
+ asl.f 0,r6,8
+ mov r7,acc2
+ add.pl r6,r6,r6
+ bclr.pl r6,r6,23
+ add.pl.f r7,r7,r7
+ add.cs r6,r6,1
+ lsr.f 0,r6,1
+ add_s r12,r12,r11
+ adc.f 0,r7,r4
+ add_s r12,r12, \
+ -0x3f800000
+ adc.f r8,r6,r12
+ tst.pl r8,r9
+ bic r0,r0,r4
+ min r3,r8,r9
+ jpnz.d [blink]
+ add.pnz r0,r0,r3
+; infinity or denormal number
+ add.ne.f r3,r3,r3
+ asr_s r3,r3,23+1
+ bset r6,r6,23
+ bpnz.d .Linfinity
+ sub_s r3,r3,1
+ neg_s r2,r3
+ brhi.d r2,24,.Lret_r0 ; right shift shift > 24 -> return +-0
+ lsr r2,r6,r2
+ asl r9,r6,r3
+ lsr.f 0,r2,1
+ tst r7,r7
+ add_s r0,r0,r2
+ bset.ne r9,r9,0
+ adc.f 0,r9,r4
+ j_s.d [blink]
+ add.cs r0,r0,1
+.Linfinity:
+ j_s.d [blink]
+ add_s r0,r0,r9
+
+.Lret_r0: j_s [blink]
+
+ .balign 4
+.Ldenorm_dbl0:
+ bclr_s r2,r2,31
+ norm.f r4,r2
+ add_s r2,r2,r2
+ asl r2,r2,r4
+ breq.d r12,r9,.Ldenorm_dbl0_inf_nan_dbl1
+ asl r4,r4,23
+ mululw 0,r2,r3
+ machulw r6,r2,r3
+ sub.ne.f r12,r12,r4
+ ld.as r4,[pcl,28]; [pcl,((.L7fffffff-.+2)/4)]
+ bhi.d .Lpast_denorm
+ xor_s r0,r0,r1
+ bmsk r1,r0,30
+ j_s.d [blink]
+ bic_s r0,r0,r1
+
+ .balign 4
+.Ldenorm_dbl0_inf_nan_dbl1:
+ bmsk.f 0,r0,30
+ mov.eq r1,-1
+.Linf_nan_dbl1:
+ xor_s r1,r1,r0
+.Linf_nan_dbl0:
+ bclr_s r1,r1,31
+ j_s.d [blink]
+ xor_s r0,r0,r1
+
+ .balign 4
+.Ldenorm_dbl1:
+ breq.d r11,r9,.Linf_nan_dbl0_2
+ norm.f r3,r4
+ sub_s r3,r3,7
+ asl r4,r4,r3
+ mululw 0,r2,r4
+ machulw r6,r2,r4
+ sub_s r3,r3,1
+ asl_s r3,r3,23
+ sub.ne.f r11,r11,r3
+ ld.as r4,[pcl,11]; [pcl,((.L7fffffff-.+2)/4)]
+ bhi.d .Lpast_denorm
+ bmsk r8,r0,30
+ j_s.d [blink]
+ bic r0,r0,r8
+
+ .balign 4
+.Linf_nan_dbl0_2:
+ bclr_s r1,r1,31
+ xor_s r0,r0,r1
+ sub.eq r1,r1,1 ; inf/nan * 0 -> nan
+ bic.f 0,r9,r1
+ j_s.d [blink]
+ or.eq r0,r0,r1 ; r1 nan -> result nan
+
+ .balign 4
+.L7f800000:
+ .long 0x7f800000
+.L7fffffff:
+ .long 0x7fffffff
+ ENDFUNC(__mulsf3)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+/*
+ to calculate a := b/x as b*y, with y := 1/x:
+ - x is in the range [1..2)
+ - calculate 15..18 bit inverse y0 using a table of approximating polynoms.
+ Precision is higher for polynoms used to evaluate input with larger
+ value.
+ - Do one newton-raphson iteration step to double the precision,
+ then multiply this with the divisor
+ -> more time to decide if dividend is subnormal
+ - the worst error propagation is on the side of the value range
+ with the least initial defect, thus giving us about 30 bits precision.
+ The truncation error for the either is less than 1 + x/2 ulp.
+ A 31 bit inverse can be simply calculated by using x with implicit 1
+ and chaining the multiplies. For a 32 bit inverse, we multiply y0^2
+ with the bare fraction part of x, then add in y0^2 for the implicit
+ 1 of x.
+ - If calculating a 31 bit inverse, the systematic error is less than
+ -1 ulp; likewise, for 32 bit, it is less than -2 ulp.
+ - If we calculate our seed with a 32 bit fraction, we can archive a
+ tentative result strictly better than -2 / +2.5 (1) ulp/128, i.e. we
+ only need to take the step to calculate the 2nd stage rest and
+ rounding adjust 1/32th of the time. However, if we use a 20 bit
+ fraction for the seed, the negative error can exceed -2 ulp/128, (2)
+ thus for a simple add / tst check, we need to do the 2nd stage
+ rest calculation/ rounding adjust 1/16th of the time.
+ (1): The inexactness of the 32 bit inverse contributes an error in the
+ range of (-1 .. +(1+x/2) ) ulp/128. Leaving out the low word of the
+ rest contributes an error < +1/x ulp/128 . In the interval [1,2),
+ x/2 + 1/x <= 1.5 .
+ (2): Unless proven otherwise. I have not actually looked for an
+ example where -2 ulp/128 is exceeded, and my calculations indicate
+ that the excess, if existent, is less than -1/512 ulp.
+ ??? The algorithm is still based on the ARC700 optimized code.
+ Maybe we could make better use of 64 bit multiply results and/or mmed .
+ */
+#include "../arc-ieee-754.h"
+
+/* N.B. fp-bit.c does double rounding on denormal numbers. */
+#if 0 /* DEBUG */
+ .global __divdf3
+ FUNC(__divdf3)
+ .balign 4
+__divdf3:
+ push_s blink
+ push_s r2
+ push_s r3
+ push_s r0
+ bl.d __divdf3_c
+ push_s r1
+ ld_s r2,[sp,12]
+ ld_s r3,[sp,8]
+ st_s r0,[sp,12]
+ st_s r1,[sp,8]
+ pop_s r1
+ bl.d __divdf3_asm
+ pop_s r0
+ pop_s r3
+ pop_s r2
+ pop_s blink
+ cmp r0,r2
+ cmp.eq r1,r3
+ jeq_s [blink]
+ and r12,DBL0H,DBL1H
+ bic.f 0,0x7ff80000,r12 ; both NaN -> OK
+ jeq_s [blink]
+ bl abort
+ ENDFUNC(__divdf3)
+#define __divdf3 __divdf3_asm
+#endif /* DEBUG */
+
+ FUNC(__divdf3)
+ .balign 4
+.L7ff00000:
+ .long 0x7ff00000
+.Ldivtab:
+ .long 0xfc0fffe1
+ .long 0xf46ffdfb
+ .long 0xed1ffa54
+ .long 0xe61ff515
+ .long 0xdf7fee75
+ .long 0xd91fe680
+ .long 0xd2ffdd52
+ .long 0xcd1fd30c
+ .long 0xc77fc7cd
+ .long 0xc21fbbb6
+ .long 0xbcefaec0
+ .long 0xb7efa100
+ .long 0xb32f92bf
+ .long 0xae8f83b7
+ .long 0xaa2f7467
+ .long 0xa5ef6479
+ .long 0xa1cf53fa
+ .long 0x9ddf433e
+ .long 0x9a0f3216
+ .long 0x965f2091
+ .long 0x92df0f11
+ .long 0x8f6efd05
+ .long 0x8c1eeacc
+ .long 0x88eed876
+ .long 0x85dec615
+ .long 0x82eeb3b9
+ .long 0x800ea10b
+ .long 0x7d3e8e0f
+ .long 0x7a8e7b3f
+ .long 0x77ee6836
+ .long 0x756e5576
+ .long 0x72fe4293
+ .long 0x709e2f93
+ .long 0x6e4e1c7f
+ .long 0x6c0e095e
+ .long 0x69edf6c5
+ .long 0x67cde3a5
+ .long 0x65cdd125
+ .long 0x63cdbe25
+ .long 0x61ddab3f
+ .long 0x600d991f
+ .long 0x5e3d868c
+ .long 0x5c6d7384
+ .long 0x5abd615f
+ .long 0x590d4ecd
+ .long 0x576d3c83
+ .long 0x55dd2a89
+ .long 0x545d18e9
+ .long 0x52dd06e9
+ .long 0x516cf54e
+ .long 0x4ffce356
+ .long 0x4e9cd1ce
+ .long 0x4d3cbfec
+ .long 0x4becae86
+ .long 0x4aac9da4
+ .long 0x496c8c73
+ .long 0x483c7bd3
+ .long 0x470c6ae8
+ .long 0x45dc59af
+ .long 0x44bc4915
+ .long 0x43ac3924
+ .long 0x428c27fb
+ .long 0x418c187a
+ .long 0x407c07bd
+
+__divdf3_support: /* This label makes debugger output saner. */
+ .balign 4
+.Ldenorm_dbl1:
+ brge r6, \
+ 0x43500000,.Linf_NaN ; large number / denorm -> Inf
+ bmsk.f r12,DBL1H,19
+ mov.eq r12,DBL1L
+ mov.eq DBL1L,0
+ sub.eq r7,r7,32
+ norm.f r11,r12 ; flag for x/0 -> Inf check
+ beq_s .Linf_NaN
+ mov.mi r11,0
+ add.pl r11,r11,1
+ add_s r12,r12,r12
+ asl r8,r12,r11
+ rsub r12,r11,31
+ lsr r12,DBL1L,r12
+ tst_s DBL1H,DBL1H
+ or r8,r8,r12
+ lsr r4,r8,26
+ lsr DBL1H,r8,12
+ ld.as r4,[r10,r4]
+ bxor.mi DBL1H,DBL1H,31
+ sub r11,r11,11
+ asl DBL1L,DBL1L,r11
+ sub r11,r11,1
+ mulu64 r4,r8
+ sub r7,r7,r11
+ b.d .Lpast_denorm_dbl1
+ asl r7,r7,20
+
+ .balign 4
+.Ldenorm_dbl0:
+ bmsk.f r12,DBL0H,19
+ ; wb stall
+ mov.eq r12,DBL0L
+ sub.eq r6,r6,32
+ norm.f r11,r12 ; flag for 0/x -> 0 check
+ brge r7, \
+ 0x43500000, .Lret0_2 ; denorm/large number -> 0
+ beq_s .Lret0_2
+ mov.mi r11,0
+ add.pl r11,r11,1
+ asl r12,r12,r11
+ sub r6,r6,r11
+ add.f 0,r6,31
+ lsr r10,DBL0L,r6
+ mov.mi r10,0
+ add r6,r6,11+32
+ neg.f r11,r6
+ asl DBL0L,DBL0L,r11
+ mov.pl DBL0L,0
+ sub r6,r6,32-1
+ b.d .Lpast_denorm_dbl0
+ asl r6,r6,20
+
+.Linf_NaN:
+ tst_s DBL0L,DBL0L ; 0/0 -> NaN
+ xor_s DBL1H,DBL1H,DBL0H
+ bclr.eq.f DBL0H,DBL0H,31
+ bmsk DBL0H,DBL1H,30
+ xor_s DBL0H,DBL0H,DBL1H
+ sub.eq DBL0H,DBL0H,1
+ mov_s DBL0L,0
+ j_s.d [blink]
+ or DBL0H,DBL0H,r9
+ .balign 4
+.Lret0_2:
+ xor_s DBL1H,DBL1H,DBL0H
+ mov_s DBL0L,0
+ bmsk DBL0H,DBL1H,30
+ j_s.d [blink]
+ xor_s DBL0H,DBL0H,DBL1H
+ .balign 4
+ .global __divdf3
+/* N.B. the spacing between divtab and the sub3 to get its address must
+ be a multiple of 8. */
+__divdf3:
+ asl r8,DBL1H,12
+ lsr r4,r8,26
+ sub3 r10,pcl,61; (.-.Ldivtab) >> 3
+ ld.as r9,[pcl,-124]; [pcl,(-((.-.L7ff00000) >> 2))] ; 0x7ff00000
+ ld.as r4,[r10,r4]
+ lsr r12,DBL1L,20
+ and.f r7,DBL1H,r9
+ or r8,r8,r12
+ mulu64 r4,r8
+ beq.d .Ldenorm_dbl1
+.Lpast_denorm_dbl1:
+ and.f r6,DBL0H,r9
+ breq.d r7,r9,.Linf_nan_dbl1
+ asl r4,r4,12
+ sub r4,r4,mhi
+ mulu64 r4,r4
+ beq.d .Ldenorm_dbl0
+ lsr r8,r8,1
+ breq.d r6,r9,.Linf_nan_dbl0
+ asl r12,DBL0H,11
+ lsr r10,DBL0L,21
+.Lpast_denorm_dbl0:
+ bset r8,r8,31
+ mulu64 mhi,r8
+ add_s r12,r12,r10
+ bset r5,r12,31
+ cmp r5,r8
+ cmp.eq DBL0L,DBL1L
+ lsr.cc r5,r5,1
+ sub r4,r4,mhi ; u1.31 inverse, about 30 bit
+ mulu64 r5,r4 ; result fraction highpart
+ lsr r8,r8,2 ; u3.29
+ add r5,r6, /* wait for immediate */ \
+ 0x3fe00000
+ mov r11,mhi ; result fraction highpart
+ mulu64 r11,r8 ; u-28.31
+ asl_s DBL1L,DBL1L,9 ; u-29.23:9
+ sbc r6,r5,r7
+ mov r12,mlo ; u-28.31
+ mulu64 r11,DBL1L ; mhi: u-28.23:9
+ add.cs DBL0L,DBL0L,DBL0L
+ asl_s DBL0L,DBL0L,6 ; u-26.25:7
+ asl r10,r11,23
+ sub_l DBL0L,DBL0L,r12
+ lsr r7,r11,9
+ sub r5,DBL0L,mhi ; rest msw ; u-26.31:0
+ mul64 r5,r4 ; mhi: result fraction lowpart
+ xor.f 0,DBL0H,DBL1H
+ and DBL0H,r6,r9
+ add_s DBL0H,DBL0H,r7
+ bclr r12,r9,20 ; 0x7fe00000
+ brhs.d r6,r12,.Linf_denorm
+ bxor.mi DBL0H,DBL0H,31
+ add.f r12,mhi,0x11
+ asr r9,r12,5
+ sub.mi DBL0H,DBL0H,1
+ add.f DBL0L,r9,r10
+ tst r12,0x1c
+ jne.d [blink]
+ add.cs DBL0H,DBL0H,1
+ /* work out exact rounding if we fall through here. */
+ /* We know that the exact result cannot be represented in double
+ precision. Find the mid-point between the two nearest
+ representable values, multiply with the divisor, and check if
+ the result is larger than the dividend. Since we want to know
+ only the sign bit, it is sufficient to calculate only the
+ highpart of the lower 64 bits. */
+ mulu64 r11,DBL1L ; rest before considering r12 in r5 : -mlo
+ sub.f DBL0L,DBL0L,1
+ asl r12,r9,2 ; u-22.30:2
+ sub.cs DBL0H,DBL0H,1
+ sub.f r12,r12,2
+ mov r10,mlo ; rest before considering r12 in r5 : -r10
+ mulu64 r12,DBL1L ; mhi: u-51.32
+ asl r5,r5,25 ; s-51.7:25
+ lsr r10,r10,7 ; u-51.30:2
+ mov r7,mhi ; u-51.32
+ mulu64 r12,r8 ; mlo: u-51.31:1
+ sub r5,r5,r10
+ add.mi r5,r5,DBL1L ; signed multiply adjust for r12*DBL1L
+ bset r7,r7,0 ; make sure that the result is not zero, and that
+ sub r5,r5,r7 ; a highpart zero appears negative
+ sub.f r5,r5,mlo ; rest msw
+ add.pl.f DBL0L,DBL0L,1
+ j_s.d [blink]
+ add.eq DBL0H,DBL0H,1
+
+.Linf_nan_dbl1: ; 0/Inf -> NaN Inf/Inf -> NaN x/Inf-> 0 x/NaN -> NaN
+ or.f 0,r6,DBL0L
+ cmp.ne r6,r9
+ not_s DBL0L,DBL1H
+ sub_s.ne DBL0L,DBL0L,DBL0L
+ tst_s DBL0H,DBL0H
+ add_s DBL0H,DBL1H,DBL0L
+ j_s.d [blink]
+ bxor.mi DBL0H,DBL0H,31
+.Linf_nan_dbl0:
+ tst_s DBL1H,DBL1H
+ j_s.d [blink]
+ bxor.mi DBL0H,DBL0H,31
+ .balign 4
+.Linf_denorm:
+ lsr r12,r6,28
+ brlo.d r12,0xc,.Linf
+.Ldenorm:
+ asr r6,r6,20
+ neg r9,r6
+ mov_s DBL0H,0
+ brhs.d r9,54,.Lret0
+ bxor.mi DBL0H,DBL0H,31
+ add r12,mhi,1
+ and r12,r12,-4
+ rsub r7,r6,5
+ asr r10,r12,28
+ bmsk r4,r12,27
+ min r7,r7,31
+ asr DBL0L,r4,r7
+ add DBL1H,r11,r10
+ abs.f r10,r4
+ sub.mi r10,r10,1
+ add.f r7,r6,32-5
+ asl r4,r4,r7
+ mov.mi r4,r10
+ add.f r10,r6,23
+ rsub r7,r6,9
+ lsr r7,DBL1H,r7
+ asl r10,DBL1H,r10
+ or.pnz DBL0H,DBL0H,r7
+ or.mi r4,r4,r10
+ mov.mi r10,r7
+ add.f DBL0L,r10,DBL0L
+ add.cs.f DBL0H,DBL0H,1 ; carry clear after this point
+ bxor.f 0,r4,31
+ add.pnz.f DBL0L,DBL0L,1
+ add.cs.f DBL0H,DBL0H,1
+ jne_s [blink]
+ /* Calculation so far was not conclusive; calculate further rest. */
+ mulu64 r11,DBL1L ; rest before considering r12 in r5 : -mlo
+ asr.f r12,r12,3
+ asl r5,r5,25 ; s-51.7:25
+ mov r11,mlo ; rest before considering r12 in r5 : -r11
+ mulu64 r12,r8 ; u-51.31:1
+ and r9,DBL0L,1 ; tie-breaker: round to even
+ lsr r11,r11,7 ; u-51.30:2
+ mov DBL1H,mlo ; u-51.31:1
+ mulu64 r12,DBL1L ; u-51.62:2
+ sub.mi r11,r11,DBL1L ; signed multiply adjust for r12*DBL1L
+ add_s DBL1H,DBL1H,r11
+ sub DBL1H,DBL1H,r5 ; -rest msw
+ add_s DBL1H,DBL1H,mhi ; -rest msw
+ add.f 0,DBL1H,DBL1H ; can't ror.f by 32 :-(
+ tst_s DBL1H,DBL1H
+ cmp.eq mlo,r9
+ add.cs.f DBL0L,DBL0L,1
+ j_s.d [blink]
+ add.cs DBL0H,DBL0H,1
+
+.Lret0:
+ /* return +- 0 */
+ j_s.d [blink]
+ mov_s DBL0L,0
+.Linf:
+ mov_s DBL0H,r9
+ mov_s DBL0L,0
+ j_s.d [blink]
+ bxor.mi DBL0H,DBL0H,31
+ ENDFUNC(__divdf3)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+/*
+ - calculate 15..18 bit inverse using a table of approximating polynoms.
+ precision is higher for polynoms used to evaluate input with larger
+ value.
+ - do one newton-raphson iteration step to double the precision,
+ then multiply this with the divisor
+ -> more time to decide if dividend is subnormal
+ - the worst error propagation is on the side of the value range
+ with the least initial defect, thus giving us about 30 bits precision.
+ */
+#include "../arc-ieee-754.h"
+
+#if 0 /* DEBUG */
+ .global __divsf3
+ FUNC(__divsf3)
+ .balign 4
+__divsf3:
+ push_s blink
+ push_s r1
+ bl.d __divsf3_c
+ push_s r0
+ ld_s r1,[sp,4]
+ st_s r0,[sp,4]
+ bl.d __divsf3_asm
+ pop_s r0
+ pop_s r1
+ pop_s blink
+ cmp r0,r1
+#if 1
+ bne abort
+ jeq_s [blink]
+ b abort
+#else
+ bne abort
+ j_s [blink]
+#endif
+ ENDFUNC(__divsf3)
+#define __divsf3 __divsf3_asm
+#endif /* DEBUG */
+
+ FUNC(__divsf3)
+ .balign 4
+.Ldivtab:
+ .long 0xfc0ffff0
+ .long 0xf46ffefd
+ .long 0xed1ffd2a
+ .long 0xe627fa8e
+ .long 0xdf7ff73b
+ .long 0xd917f33b
+ .long 0xd2f7eea3
+ .long 0xcd1fe986
+ .long 0xc77fe3e7
+ .long 0xc21fdddb
+ .long 0xbcefd760
+ .long 0xb7f7d08c
+ .long 0xb32fc960
+ .long 0xae97c1ea
+ .long 0xaa27ba26
+ .long 0xa5e7b22e
+ .long 0xa1cfa9fe
+ .long 0x9ddfa1a0
+ .long 0x9a0f990c
+ .long 0x9667905d
+ .long 0x92df878a
+ .long 0x8f6f7e84
+ .long 0x8c27757e
+ .long 0x88f76c54
+ .long 0x85df630c
+ .long 0x82e759c5
+ .long 0x8007506d
+ .long 0x7d3f470a
+ .long 0x7a8f3da2
+ .long 0x77ef341e
+ .long 0x756f2abe
+ .long 0x72f7212d
+ .long 0x709717ad
+ .long 0x6e4f0e44
+ .long 0x6c1704d6
+ .long 0x69e6fb44
+ .long 0x67cef1d7
+ .long 0x65c6e872
+ .long 0x63cedf18
+ .long 0x61e6d5cd
+ .long 0x6006cc6d
+ .long 0x5e36c323
+ .long 0x5c76b9f3
+ .long 0x5abeb0b7
+ .long 0x5916a79b
+ .long 0x57769e77
+ .long 0x55de954d
+ .long 0x54568c4e
+ .long 0x52d6834d
+ .long 0x51667a7f
+ .long 0x4ffe71b5
+ .long 0x4e9e68f1
+ .long 0x4d466035
+ .long 0x4bf65784
+ .long 0x4aae4ede
+ .long 0x496e4646
+ .long 0x48363dbd
+ .long 0x47063547
+ .long 0x45de2ce5
+ .long 0x44be2498
+ .long 0x43a61c64
+ .long 0x4296144a
+ .long 0x41860c0e
+ .long 0x407e03ee
+.L7f800000:
+ .long 0x7f800000
+ .balign 4
+ .global __divsf3_support
+__divsf3_support:
+.Linf_NaN:
+ bclr.f 0,r0,31 ; 0/0 -> NaN
+ xor_s r0,r0,r1
+ bmsk r1,r0,30
+ bic_s r0,r0,r1
+ sub.eq r0,r0,1
+ j_s.d [blink]
+ or r0,r0,r9
+.Lret0:
+ xor_s r0,r0,r1
+ bmsk r1,r0,30
+ j_s.d [blink]
+ bic_s r0,r0,r1
+/* N.B. the spacing between divtab and the sub3 to get its address must
+ be a multiple of 8. */
+__divsf3:
+ lsr r2,r1,17
+ sub3 r3,pcl,37 ; (.-.Ldivtab) >> 3
+ bmsk_s r2,r2,5
+ ld.as r5,[r3,r2]
+ asl r4,r1,9
+ ld.as r9,[pcl,-13]; [pcl,(-((.-.L7f800000) >> 2))] ; 0x7f800000
+ mulu64 r5,r4
+ and.f r11,r1,r9
+ asl r6,r1,8
+ bset r6,r6,31
+ beq.d .Ldenorm_fp1
+ asl r5,r5,13
+ breq.d r11,r9,.Linf_nan_fp1
+ and.f r2,r0,r9
+ sub r7,r5,mhi
+ mulu64 r7,r6
+ beq.d .Ldenorm_fp0
+ asl r12,r0,8
+ breq.d r2,r9,.Linf_nan_fp0
+ mulu64 mhi,r7
+.Lpast_denorm_fp1:
+ bset r3,r12,31
+.Lpast_denorm_fp0:
+ cmp_s r3,r6
+ lsr.cc r3,r3,1
+ add_s r2,r2, /* wait for immediate */ \
+ 0x3f000000
+ sub r7,r7,mhi ; u1.31 inverse, about 30 bit
+ mulu64 r3,r7
+ sbc r2,r2,r11
+ xor.f 0,r0,r1
+ and r0,r2,r9
+ bclr r3,r9,23 ; 0x7f000000
+ brhs.d r2,r3,.Linf_denorm
+ bxor.mi r0,r0,31
+.Lpast_denorm:
+ add r3,mhi,0x22 ; round to nearest or higher
+ tst r3,0x3c ; check if rounding was unsafe
+ lsr r3,r3,6
+ jne.d [blink] ; return if rounding was safe.
+ add_s r0,r0,r3
+ /* work out exact rounding if we fall through here. */
+ /* We know that the exact result cannot be represented in single
+ precision. Find the mid-point between the two nearest
+ representable values, multiply with the divisor, and check if
+ the result is larger than the dividend. */
+ add_s r3,r3,r3
+ sub_s r3,r3,1
+ mulu64 r3,r6
+ asr.f 0,r0,1 ; for round-to-even in case this is a denorm
+ rsub r2,r9,25
+ asl_s r12,r12,r2
+ sub.f 0,r12,mlo
+ j_s.d [blink]
+ sub.mi r0,r0,1
+.Linf_nan_fp1:
+ lsr_s r0,r0,31
+ bmsk.f 0,r1,22
+ asl_s r0,r0,31
+ bne_s 0f ; inf/inf -> nan
+ brne r2,r9,.Lsigned0 ; x/inf -> 0, but x/nan -> nan
+0: j_s.d [blink]
+ mov r0,-1
+.Lsigned0:
+.Linf_nan_fp0:
+ tst_s r1,r1
+ j_s.d [blink]
+ bxor.mi r0,r0,31
+ .balign 4
+ .global __divsf3
+/* For denormal results, it is possible that an exact result needs
+ rounding, and thus the round-to-even rule has to come into play. */
+.Linf_denorm:
+ brlo r2,0xc0000000,.Linf
+.Ldenorm:
+ asr_s r2,r2,23
+ bic r0,r0,r9
+ neg r9,r2
+ brlo.d r9,25,.Lpast_denorm
+ lsr r3,mlo,r9
+ /* Fall through: return +- 0 */
+ j_s [blink]
+.Linf:
+ j_s.d [blink]
+ or r0,r0,r9
+ .balign 4
+.Ldenorm_fp1:
+ bclr r6,r6,31
+ norm.f r12,r6 ; flag for x/0 -> Inf check
+ add r6,r6,r6
+ rsub r5,r12,16
+ ror r5,r1,r5
+ asl r6,r6,r12
+ bmsk r5,r5,5
+ ld.as r5,[r3,r5]
+ add r4,r6,r6
+ ; load latency
+ mulu64 r5,r4
+ bic.ne.f 0, \
+ 0x60000000,r0 ; large number / denorm -> Inf
+ asl r5,r5,13
+ sub r7,r5,mhi
+ beq.d .Linf_NaN
+ mulu64 r7,r6
+ asl_s r12,r12,23
+ and.f r2,r0,r9
+ add_s r2,r2,r12
+ asl r12,r0,8
+ bne.d .Lpast_denorm_fp1
+.Ldenorm_fp0: mulu64 mhi,r7
+ bclr r12,r12,31
+ norm.f r3,r12 ; flag for 0/x -> 0 check
+ bic.ne.f 0,0x60000000,r1 ; denorm/large number -> 0
+ beq_s .Lret0
+ asl_s r12,r12,r3
+ asl_s r3,r3,23
+ add_s r12,r12,r12
+ add r11,r11,r3
+ b.d .Lpast_denorm_fp0
+ mov_s r3,r12
+ ENDFUNC(__divsf3)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "../arc-ieee-754.h"
+
+#if 0 /* DEBUG */
+ .global __muldf3
+ .balign 4
+__muldf3:
+ push_s blink
+ push_s r2
+ push_s r3
+ push_s r0
+ bl.d __muldf3_c
+ push_s r1
+ ld_s r2,[sp,12]
+ ld_s r3,[sp,8]
+ st_s r0,[sp,12]
+ st_s r1,[sp,8]
+ pop_s r1
+ bl.d __muldf3_asm
+ pop_s r0
+ pop_s r3
+ pop_s r2
+ pop_s blink
+ cmp r0,r2
+ cmp.eq r1,r3
+ jeq_s [blink]
+ and r12,DBL0H,DBL1H
+ bic.f 0,0x7ff80000,r12 ; both NaN -> OK
+ jeq_s [blink]
+ b abort
+#define __muldf3 __muldf3_asm
+#endif /* DEBUG */
+
+__muldf3_support: /* This label makes debugger output saner. */
+ .balign 4
+ FUNC(__muldf3)
+.Ldenorm_2:
+ breq.d DBL1L,0,.Lret0_2 ; 0 input -> 0 output
+ norm.f r12,DBL1L
+ mov.mi r12,21
+ add.pl r12,r12,22
+ neg r11,r12
+ asl_s r12,r12,20
+ lsr.f DBL1H,DBL1L,r11
+ ror DBL1L,DBL1L,r11
+ sub_s DBL0H,DBL0H,r12
+ mov.eq DBL1H,DBL1L
+ sub_l DBL1L,DBL1L,DBL1H
+ /* Fall through. */
+ .global __muldf3
+ .balign 4
+__muldf3:
+ mulu64 DBL0L,DBL1L
+ ld.as r9,[pcl,0x68] ; ((.L7ff00000-.+2)/4)]
+ bmsk r6,DBL0H,19
+ bset r6,r6,20
+ and r11,DBL0H,r9
+ breq.d r11,0,.Ldenorm_dbl0
+ and r12,DBL1H,r9
+ breq.d r12,0,.Ldenorm_dbl1
+ mov r8,mlo
+ mov r4,mhi
+ mulu64 r6,DBL1L
+ breq.d r11,r9,.Linf_nan
+ bmsk r10,DBL1H,19
+ breq.d r12,r9,.Linf_nan
+ bset r10,r10,20
+ add.f r4,r4,mlo
+ adc r5,mhi,0
+ mulu64 r10,DBL0L
+ add_s r12,r12,r11 ; add exponents
+ add.f r4,r4,mlo
+ adc r5,r5,mhi
+ mulu64 r6,r10
+ tst r8,r8
+ bclr r8,r9,30 ; 0x3ff00000
+ bset.ne r4,r4,0 ; put least significant word into sticky bit
+ bclr r6,r9,20 ; 0x7fe00000
+ add.f r5,r5,mlo
+ adc r7,mhi,0 ; fraction product in r7:r5:r4
+ lsr.f r10,r7,9
+ rsub.eq r8,r8,r9 ; 0x40000000
+ sub r12,r12,r8 ; subtract bias + implicit 1
+ brhs.d r12,r6,.Linf_denorm
+ rsub r10,r10,12
+.Lshift_frac:
+ neg r8,r10
+ asl r6,r4,r10
+ lsr DBL0L,r4,r8
+ add.f 0,r6,r6
+ btst.eq DBL0L,0
+ cmp.eq r4,r4 ; round to nearest / round to even
+ asl r4,r5,r10
+ lsr r5,r5,r8
+ adc.f DBL0L,DBL0L,r4
+ xor.f 0,DBL0H,DBL1H
+ asl r7,r7,r10
+ add_s r12,r12,r5
+ adc DBL0H,r12,r7
+ j_s.d [blink]
+ bset.mi DBL0H,DBL0H,31
+
+/* N.B. This is optimized for ARC700.
+ ARC600 has very different scheduling / instruction selection criteria. */
+
+/* If one number is denormal, subtract some from the exponent of the other
+ one (if the other exponent is too small, return 0), and normalize the
+ denormal. Then re-run the computation. */
+.Lret0_2:
+ lsr_s DBL0H,DBL0H,31
+ asl_s DBL0H,DBL0H,31
+ j_s.d [blink]
+ mov_s DBL0L,0
+ .balign 4
+.Ldenorm_dbl0:
+ mov_s r12,DBL0L
+ mov_s DBL0L,DBL1L
+ mov_s DBL1L,r12
+ mov_s r12,DBL0H
+ mov_s DBL0H,DBL1H
+ mov_s DBL1H,r12
+ and r11,DBL0H,r9
+.Ldenorm_dbl1:
+ brhs r11,r9,.Linf_nan
+ brhs 0x3ca00001,r11,.Lret0
+ sub_s DBL0H,DBL0H,DBL1H
+ bmsk.f DBL1H,DBL1H,30
+ add_s DBL0H,DBL0H,DBL1H
+ beq.d .Ldenorm_2
+ norm r12,DBL1H
+ sub_s r12,r12,10
+ asl r5,r12,20
+ asl_s DBL1H,DBL1H,r12
+ sub DBL0H,DBL0H,r5
+ neg r5,r12
+ lsr r6,DBL1L,r5
+ asl_s DBL1L,DBL1L,r12
+ b.d __muldf3
+ add_s DBL1H,DBL1H,r6
+
+.Lret0: xor_s DBL0H,DBL0H,DBL1H
+ bclr DBL1H,DBL0H,31
+ xor_s DBL0H,DBL0H,DBL1H
+ j_s.d [blink]
+ mov_s DBL0L,0
+
+ .balign 4
+.Linf_nan:
+ bclr r12,DBL1H,31
+ xor_s DBL1H,DBL1H,DBL0H
+ bclr_s DBL0H,DBL0H,31
+ max r8,DBL0H,r12 ; either NaN -> NaN ; otherwise inf
+ or.f 0,DBL0H,DBL0L
+ mov_s DBL0L,0
+ or.ne.f DBL1L,DBL1L,r12
+ not_s DBL0H,DBL0L ; inf * 0 -> NaN
+ mov.ne DBL0H,r8
+ tst_s DBL1H,DBL1H
+ j_s.d [blink]
+ bset.mi DBL0H,DBL0H,31
+
+/* We have checked for infinitey / NaN input before, and transformed
+ denormalized inputs into normalized inputs. Thus, the worst case
+ exponent overflows are:
+ 1 + 1 - 0x400 == 0xc02 : maximum underflow
+ 0x7fe + 0x7fe - 0x3ff == 0xbfd ; maximum overflow
+ N.B. 0x7e and 0x7f are also values for overflow.
+
+ If (r12 <= -54), we have an underflow to zero. */
+ .balign 4
+.Linf_denorm:
+ lsr r6,r12,28
+ brlo.d r6,0xc,.Linf
+ asr r6,r12,20
+ add.f r10,r10,r6
+ brgt.d r10,0,.Lshift_frac
+ mov_s r12,0
+ beq.d .Lround_frac
+ add r10,r10,32
+.Lshift32_frac:
+ tst r4,r4
+ mov r4,r5
+ bset.ne r4,r4,1
+ mov r5,r7
+ brge.d r10,1,.Lshift_frac
+ mov r7,0
+ breq.d r10,0,.Lround_frac
+ add r10,r10,32
+ brgt r10,21,.Lshift32_frac
+ b_s .Lret0
+
+.Lround_frac:
+ add.f 0,r4,r4
+ btst.eq r5,0
+ mov_s DBL0L,r5
+ mov_s DBL0H,r7
+ adc.eq.f DBL0L,DBL0L,0
+ j_s.d [blink]
+ adc.eq DBL0H,DBL0H,0
+
+.Linf: mov_s DBL0L,0
+ xor.f DBL1H,DBL1H,DBL0H
+ mov_s DBL0H,r9
+ j_s.d [blink]
+ bset.mi DBL0H,DBL0H,31
+ ENDFUNC(__muldf3)
+
+ .balign 4
+.L7ff00000:
+ .long 0x7ff00000
--- /dev/null
+/* Copyright (C) 2008-2013 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "../arc-ieee-754.h"
+
+#if 0 /* DEBUG */
+ .global __mulsf3
+ FUNC(__mulsf3)
+ .balign 4
+__mulsf3:
+ push_s blink
+ push_s r1
+ bl.d __mulsf3_c
+ push_s r0
+ ld_s r1,[sp,4]
+ st_s r0,[sp,4]
+ bl.d __mulsf3_asm
+ pop_s r0
+ pop_s r1
+ pop_s blink
+ cmp r0,r1
+ jeq_s [blink]
+ and r12,r0,r1
+ bic.f 0,0x7f800000,r12
+ bne 0f
+ bmsk.f 0,r0,22
+ bmsk.ne.f r1,r1,22
+ jne_s [blink] ; both NaN -> OK
+0: bl abort
+ ENDFUNC(__mulsf3)
+#define __mulsf3 __mulsf3_asm
+#endif /* DEBUG */
+
+ .balign 4
+ .global __mulsf3
+ FUNC(__mulsf3)
+__mulsf3:
+ ld.as r9,[pcl,80]; [pcl,((.L7f800000-.+2)/4)]
+ bmsk r4,r1,22
+ bset r2,r0,23
+ asl_s r2,r2,8
+ bset r3,r4,23
+ mulu64 r2,r3
+ and r11,r0,r9
+ breq.d r11,0,.Ldenorm_dbl0
+ and r12,r1,r9
+ breq.d r12,0,.Ldenorm_dbl1
+ xor_s r0,r0,r1
+ breq.d r11,r9,.Linf_nan_dbl0
+ ld.as r4,[pcl,70]; [pcl,((.L7fffffff-.+2)/4)]
+ breq.d r12,r9,.Linf_nan_dbl1
+.Lpast_denorm:
+ asl.f 0,mhi,8
+ mov r6,mhi
+ mov r7,mlo
+ add.pl r6,r6,r6
+ bclr.pl r6,r6,23
+ add.pl.f r7,r7,r7
+ add.cs r6,r6,1
+ lsr.f 0,r6,1
+ add_s r12,r12,r11
+ adc.f 0,r7,r4
+ add_s r12,r12, \
+ -0x3f800000
+ adc.f r8,r6,r12
+ tst.pl r8,r9
+ bic r0,r0,r4
+ min r3,r8,r9
+ jpnz.d [blink]
+ add.pnz r0,r0,r3
+; infinity or denormal number
+ add.ne.f r3,r3,r3
+ asr_s r3,r3,23+1
+ bset r6,r6,23
+ bpnz.d .Linfinity
+ sub_s r3,r3,1
+ neg_s r2,r3
+ brhi.d r2,24,.Lret_r0 ; right shift shift > 24 -> return +-0
+ lsr r2,r6,r2
+ asl r9,r6,r3
+ lsr.f 0,r2,1
+ tst r7,r7
+ add_s r0,r0,r2
+ bset.ne r9,r9,0
+ adc.f 0,r9,r4
+ j_s.d [blink]
+ add.cs r0,r0,1
+.Linfinity:
+ j_s.d [blink]
+ add_s r0,r0,r9
+
+.Lret_r0: j_s [blink]
+
+ .balign 4
+.Ldenorm_dbl0:
+ bclr_s r2,r2,31
+ norm.f r4,r2
+ add_s r2,r2,r2
+ asl r2,r2,r4
+ mulu64 r2,r3
+ breq.d r12,r9,.Ldenorm_dbl0_inf_nan_dbl1
+ asl r4,r4,23
+ sub.ne.f r12,r12,r4
+ ld.as r4,[pcl,29]; [pcl,((.L7fffffff-.+2)/4)]
+ bhi.d .Lpast_denorm
+ xor_s r0,r0,r1
+ bmsk r1,r0,30
+ j_s.d [blink]
+ bic_s r0,r0,r1
+
+ .balign 4
+.Ldenorm_dbl0_inf_nan_dbl1:
+ bmsk.f 0,r0,30
+ beq_s .Lretnan
+ xor_s r0,r0,r1
+.Linf_nan_dbl1:
+ xor_s r1,r1,r0
+.Linf_nan_dbl0:
+ bclr_s r1,r1,31
+ cmp_s r1,r9
+ jls.d [blink]
+ xor_s r0,r0,r1
+; r1 NaN -> result NaN
+.Lretnan:
+ j_s.d [blink]
+ mov r0,-1
+
+ .balign 4
+.Ldenorm_dbl1:
+ breq.d r11,r9,.Linf_nan_dbl0_2
+ norm.f r3,r4
+ sub_s r3,r3,7
+ asl r4,r4,r3
+ mulu64 r2,r4
+ sub_s r3,r3,1
+ asl_s r3,r3,23
+ sub.ne.f r11,r11,r3
+ ld.as r4,[pcl,11]; [pcl,((.L7fffffff-.+2)/4)]
+ bhi.d .Lpast_denorm
+ bmsk r8,r0,30
+ j_s.d [blink]
+ bic r0,r0,r8
+
+ .balign 4
+.Linf_nan_dbl0_2:
+ bclr_s r1,r1,31
+ xor_s r0,r0,r1
+ sub.eq r1,r1,1 ; inf/nan * 0 -> nan
+ bic.f 0,r9,r1
+ j_s.d [blink]
+ or.eq r0,r0,r1 ; r1 nan -> result nan
+
+ .balign 4
+.L7f800000:
+ .long 0x7f800000
+.L7fffffff:
+ .long 0x7fffffff
+ ENDFUNC(__mulsf3)
--- /dev/null
+/* Copyright (C) 2008-2013 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "../arc-ieee-754.h"
+
+#if 0 /* DEBUG */
+ .global __divsf3
+ FUNC(__divsf3)
+ .balign 4
+__divsf3:
+ push_s blink
+ push_s r1
+ bl.d __divsf3_c
+ push_s r0
+ ld_s r1,[sp,4]
+ st_s r0,[sp,4]
+ bl.d __divsf3_asm
+ pop_s r0
+ pop_s r1
+ pop_s blink
+ cmp r0,r1
+ jeq_s [blink]
+ and r12,r0,r1
+ bic.f 0,0x7f800000,r12 ; both NaN -> OK
+ jeq_s [blink]
+ bl abort
+ ENDFUNC(__divsf3)
+#define __divsf3 __divsf3_asm
+#endif /* DEBUG */
+
+ .balign 4
+__divdf3_support: /* This label makes debugger output saner. */
+ FUNC(__divsf3)
+.Ldenorm_fp0:
+ norm.f r12,r2 ; flag for 0/x -> 0 check
+ bic.ne.f 0,0x60000000,r1 ; denorm/large number -> 0
+ beq_s .Lret0_NaN
+ tst r1,r9
+ add_s r2,r2,r2
+ sub_s r12,r12,8
+ asl_s r2,r2,r12
+ asl_l r12,r12,23
+ bne.d .Lpast_denorm_fp0
+ add r5,r5,r12
+/* r0 is subnormal, r1 is subnormal or 0. */
+
+ .balign 4
+.Ldenorm_fp1:
+ norm.f r12,r3 ; flag for x/0 -> Inf check
+ bic.ne.f 0,0x60000000,r0 ; large number/denorm -> Inf
+ beq_s .Linf
+ add_s r3,r3,r3
+ sub_s r12,r12,8
+ asl_s r3,r3,r12
+ asl_s r12,r12,23
+ b.d .Lpast_denorm_fp1
+ add r4,r4,r12
+
+.Lret0_NaN:
+ bclr.f 0,r1,31 ; 0/0 -> NaN
+ bic r0,r10,r9
+ j_s.d [blink]
+ sub.eq r0,r0,1
+
+ .balign 4
+.Linf_nan_fp0:
+ bic.f 0,r9,r1 ; fp1 Inf -> result NaN
+ bic r1,r5,r9 ; fp1 sign
+ sub.eq r1,r1,1
+ j_s.d [blink]
+ xor_s r0,r0,r1
+.Linf_nan_fp1:
+ bic r0,r4,r9 ; fp0 sign
+ bmsk.f 0,r1,22 ; x/inf -> 0, x/nan -> nan
+ xor.eq r1,r1,r9
+ j_s.d [blink]
+ xor_s r0,r0,r1
+
+ .global __divsf3
+ .balign 4
+ .long 0x7f800000 ; exponent mask
+__divsf3:
+ ld r9,[pcl,-4]
+ bmsk r2,r0,22
+ xor r4,r0,r2
+ bmsk r3,r1,22
+ xor r5,r1,r3
+ and r11,r0,r9
+ breq.d r11,0,.Ldenorm_fp0
+ xor r10,r4,r5
+ breq r11,r9,.Linf_nan_fp0
+ bset_s r2,r2,23
+ and r11,r1,r9
+ breq r11,0,.Ldenorm_fp1
+ breq r11,r9,.Linf_nan_fp1
+.Lpast_denorm_fp0:
+ bset_s r3,r3,23
+.Lpast_denorm_fp1:
+ cmp r2,r3
+ asl_s r2,r2,6+1
+ asl_s r3,r3,7
+ add.lo r2,r2,r2
+ bclr r8,r9,30 ; exponent bias
+ bclr.lo r8,r8,23 ; reduce exp by one if fraction is shifted
+ sub r4,r4,r5
+ add r4,r4,r8
+ xor.f 0,r10,r4
+ bmi .Linf_denorm
+ and.f r12,r4,r9
+ beq .Ldenorm
+ sub_s r2,r2,r3 ; discard implicit 1
+ rsub r3,r3,1 ; prime r3 for two-insn divide-step use
+.Ldiv_23bit:
+ .rep 6
+ add1.f r2,r3,r2
+ sub.cc r2,r2,r3
+ .endr
+ breq r12,r9,.Linf
+ bmsk r0,r2,6
+ xor_s r2,r2,r0
+.Ldiv_17bit:
+ .rep 7
+ add1.f r2,r3,r2
+ sub.cc r2,r2,r3
+ .endr
+ asl_s r0,r0,7
+ bmsk r1,r2,6
+ xor_s r2,r2,r1
+ or_s r0,r0,r1
+.Ldiv_10bit:
+ .rep 7
+ add1.f r2,r3,r2
+ sub.cc r2,r2,r3
+ .endr
+ asl_s r0,r0,7
+ bmsk r1,r2,6
+ xor_s r2,r2,r1
+ or_s r0,r0,r1
+.Ldiv_3bit:
+ .rep 3
+ add1.f r2,r3,r2
+ sub.cc r2,r2,r3
+ .endr
+ asl_s r0,r0,3
+.Ldiv_0bit:
+ add1.f r1,r3,r2
+ sub.cc r1,r1,r3
+ bmsk_s r2,r2,2
+ tst r1,-0x7e ; 0xffffff82, test for rest or odd
+ bmsk_s r1,r1,0
+ add_s r0,r0,r2 ; assemble fraction
+ add_s r0,r0,r4 ; add in sign & exponent
+ j_s.d [blink]
+ add.ne r0,r0,r1 ; round to nearest / even
+
+ .balign 4
+.Linf:
+ j_s.d [blink]
+ or r0,r10,r9
+
+.Lret_r4:
+ j_s.d [blink]
+ mov_s r0,r4
+ .balign 4
+.Linf_denorm:
+ add.f r12,r4,r4
+ asr_l r12,r12,24
+ bpl .Linf
+ max r12,r12,-24
+.Ldenorm:
+ rsub r3,r3,1
+ add r1,pcl,68; .Ldenorm_tab-.
+ ldw.as r12,[r1,r12]
+ mov_s r0,0
+ lsr_s r2,r2
+ sub_s r1,r1,r12
+ j_s.d [r1]
+ bic r4,r10,r9
+ .short .Ldenorm_tab-.Lret_r4
+ .short .Ldenorm_tab-.Ldiv_0bit
+ .short .Ldenorm_tab-.Ldiv_3bit-2*8
+ .short .Ldenorm_tab-.Ldiv_3bit-1*8
+ .short .Ldenorm_tab-.Ldiv_3bit
+ .short .Ldenorm_tab-.Ldiv_10bit-6*8
+ .short .Ldenorm_tab-.Ldiv_10bit-5*8
+ .short .Ldenorm_tab-.Ldiv_10bit-3*8
+ .short .Ldenorm_tab-.Ldiv_10bit-3*8
+ .short .Ldenorm_tab-.Ldiv_10bit-2*8
+ .short .Ldenorm_tab-.Ldiv_10bit-1*8
+ .short .Ldenorm_tab-.Ldiv_10bit
+ .short .Ldenorm_tab-.Ldiv_17bit-6*8
+ .short .Ldenorm_tab-.Ldiv_17bit-5*8
+ .short .Ldenorm_tab-.Ldiv_17bit-4*8
+ .short .Ldenorm_tab-.Ldiv_17bit-3*8
+ .short .Ldenorm_tab-.Ldiv_17bit-2*8
+ .short .Ldenorm_tab-.Ldiv_17bit-1*8
+ .short .Ldenorm_tab-.Ldiv_17bit
+ .short .Ldenorm_tab-.Ldiv_23bit-5*8
+ .short .Ldenorm_tab-.Ldiv_23bit-4*8
+ .short .Ldenorm_tab-.Ldiv_23bit-3*8
+ .short .Ldenorm_tab-.Ldiv_23bit-2*8
+ .short .Ldenorm_tab-.Ldiv_23bit-1*8
+.Ldenorm_tab:
+ .short .Ldenorm_tab-.Ldiv_23bit
+ ENDFUNC(__divsf3)
--- /dev/null
+/* Copyright (C) 2008-2013 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "../arc-ieee-754.h"
+
+#if 0 /* DEBUG */
+ .global __mulsf3
+ FUNC(__mulsf3)
+ .balign 4
+__mulsf3:
+ push_s blink
+ push_s r1
+ bl.d __mulsf3_c
+ push_s r0
+ ld_s r1,[sp,4]
+ st_s r0,[sp,4]
+ bl.d __mulsf3_asm
+ pop_s r0
+ pop_s r1
+ pop_s blink
+ cmp r0,r1
+ jeq_s [blink]
+ and r12,r0,r1
+ bic.f 0,0x7f800000,r12
+ bne 0f
+ bmsk.f 0,r0,22
+ bmsk.ne.f r1,r1,22
+ jne_s [blink] ; both NaN -> OK
+0: bl abort
+ ENDFUNC(__mulsf3)
+#define __mulsf3 __mulsf3_asm
+#endif /* DEBUG */
+
+ .balign 4
+ .global __mulsf3
+ FUNC(__mulsf3)
+__mulsf3:
+ ld.as r9,[pcl,76]; [pcl,((.L7f800000-.+2)/4)]
+ bmsk r4,r1,22
+ bset r3,r4,23
+ bmsk r2,r0,22
+ and r11,r0,r9
+ breq.d r11,0,.Ldenorm_dbl0
+ and r12,r1,r9
+ xor_s r0,r0,r1
+ breq.d r11,r9,.Linf_nan_dbl0
+ bset_s r2,r2,23
+ breq r12,0,.Ldenorm_dbl1
+ breq r12,r9,.Linf_nan_dbl1
+.Lpast_denorm:
+ mov r6,0
+ lsr.f r7,r2
+; We could so this a bit faster here with a 32 bit shift register and
+; inserting the r2 factor / retrieving the low result a byte at a time,
+; but that'd increase code size.
+ mov lp_count,24
+ .balign 4
+ lp 0f
+ add.cs r6,r6,r3
+ lsr.f r6,r6
+ rrc.f r7,r7
+0:
+ ld.as r4,[pcl,59]; [pcl,((.L7fffffff-.+2)/4)]
+ asl.f 0,r6,8
+ add.pl r6,r6,r6
+ bclr.pl r6,r6,23
+ add.pl.f r7,r7,r7
+ add.cs r6,r6,1
+ lsr.f 0,r6,1
+ add_s r12,r12,r11
+ adc.f 0,r7,r4
+ add_s r12,r12, \
+ -0x3f800000
+ adc.f r8,r6,r12
+ tst.pl r8,r9
+ bic r0,r0,r4
+ min r3,r8,r9
+ jpnz.d [blink]
+ add.pnz r0,r0,r3
+; infinity or denormal number
+ add.ne.f r3,r3,r3
+ asr_s r3,r3,23+1
+ bset r6,r6,23
+ bpnz.d .Linfinity
+ sub_s r3,r3,1
+ neg_s r2,r3
+ brhi.d r2,24,.Lret_r0 ; right shift shift > 24 -> return +-0
+ lsr r2,r6,r2
+ asl r9,r6,r3
+ lsr.f 0,r2,1
+ tst r7,r7
+ add_s r0,r0,r2
+ bset.ne r9,r9,0
+ adc.f 0,r9,r4
+ j_s.d [blink]
+ add.cs r0,r0,1
+.Linfinity:
+ j_s.d [blink]
+ add_s r0,r0,r9
+
+.Lret_r0: j_s [blink]
+
+ .balign 4
+.Ldenorm_dbl0:
+ asl_s r2,r2,8
+ norm.f r4,r2
+ lsr_s r2,r2,7
+ asl r2,r2,r4
+ breq.d r12,r9,.Ldenorm_dbl0_inf_nan_dbl1
+ asl r4,r4,23
+ sub.ne.f r12,r12,r4
+ bhi.d .Lpast_denorm
+ xor_s r0,r0,r1
+ bmsk r1,r0,30
+ j_s.d [blink]
+ bic_s r0,r0,r1
+
+ .balign 4
+.Ldenorm_dbl0_inf_nan_dbl1:
+ bmsk.f 0,r0,30
+ beq_s .Lretnan
+ xor_s r0,r0,r1
+.Linf_nan_dbl1:
+ xor_s r1,r1,r0
+ bclr_s r1,r1,31
+ j_s.d [blink]
+ xor_s r0,r0,r1
+.Linf_nan_dbl0:
+ sub_s r2,r1,1 ; inf/nan * 0 -> nan; inf * nan -> nan (use |r2| >= inf)
+ bic.f 0,r9,r2
+ xor_s r0,r0,r1
+ bclr_s r1,r1,31
+ xor_s r0,r0,r1
+ jne_s [blink]
+.Lretnan:
+ j_s.d [blink]
+ mov r0,-1
+ .balign 4
+.Ldenorm_dbl1:
+ norm.f r3,r4
+ sub_s r3,r3,7
+ asl r4,r4,r3
+ sub_s r3,r3,1
+ asl_s r3,r3,23
+ sub.ne.f r11,r11,r3
+ bhi.d .Lpast_denorm
+ mov_s r3,r4
+ bmsk r3,r0,30
+ j_s.d [blink]
+ bic_s r0,r0,r3
+
+ .balign 4
+.L7f800000:
+ .long 0x7f800000
+.L7fffffff:
+ .long 0x7fffffff
+ ENDFUNC(__mulsf3)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+/*
+ to calculate a := b/x as b*y, with y := 1/x:
+ - x is in the range [1..2)
+ - calculate 15..18 bit inverse y0 using a table of approximating polynoms.
+ Precision is higher for polynoms used to evaluate input with larger
+ value.
+ - Do one newton-raphson iteration step to double the precision,
+ then multiply this with the divisor
+ -> more time to decide if dividend is subnormal
+ - the worst error propagation is on the side of the value range
+ with the least initial defect, thus giving us about 30 bits precision.
+ The truncation error for the either is less than 1 + x/2 ulp.
+ A 31 bit inverse can be simply calculated by using x with implicit 1
+ and chaining the multiplies. For a 32 bit inverse, we multiply y0^2
+ with the bare fraction part of x, then add in y0^2 for the implicit
+ 1 of x.
+ - If calculating a 31 bit inverse, the systematic error is less than
+ -1 ulp; likewise, for 32 bit, it is less than -2 ulp.
+ - If we calculate our seed with a 32 bit fraction, we can archive a
+ tentative result strictly better than -2 / +2.5 (1) ulp/128, i.e. we
+ only need to take the step to calculate the 2nd stage rest and
+ rounding adjust 1/32th of the time. However, if we use a 20 bit
+ fraction for the seed, the negative error can exceed -2 ulp/128, (2)
+ thus for a simple add / tst check, we need to do the 2nd stage
+ rest calculation/ rounding adjust 1/16th of the time.
+ (1): The inexactness of the 32 bit inverse contributes an error in the
+ range of (-1 .. +(1+x/2) ) ulp/128. Leaving out the low word of the
+ rest contributes an error < +1/x ulp/128 . In the interval [1,2),
+ x/2 + 1/x <= 1.5 .
+ (2): Unless proven otherwise. I have not actually looked for an
+ example where -2 ulp/128 is exceeded, and my calculations indicate
+ that the excess, if existent, is less than -1/512 ulp.
+ */
+#include "arc-ieee-754.h"
+
+/* N.B. fp-bit.c does double rounding on denormal numbers. */
+#if 0 /* DEBUG */
+ .global __divdf3
+ FUNC(__divdf3)
+ .balign 4
+__divdf3:
+ push_s blink
+ push_s r2
+ push_s r3
+ push_s r0
+ bl.d __divdf3_c
+ push_s r1
+ ld_s r2,[sp,12]
+ ld_s r3,[sp,8]
+ st_s r0,[sp,12]
+ st_s r1,[sp,8]
+ pop_s r1
+ bl.d __divdf3_asm
+ pop_s r0
+ pop_s r3
+ pop_s r2
+ pop_s blink
+ cmp r0,r2
+ cmp.eq r1,r3
+ jeq_s [blink]
+ and r12,DBL0H,DBL1H
+ bic.f 0,0x7ff80000,r12 ; both NaN -> OK
+ jeq_s [blink]
+ bl abort
+ ENDFUNC(__divdf3)
+#define __divdf3 __divdf3_asm
+#endif /* DEBUG */
+
+ FUNC(__divdf3)
+__divdf3_support: /* This label makes debugger output saner. */
+ .balign 4
+.Ldenorm_dbl1:
+ brge r6, \
+ 0x43500000,.Linf_NaN ; large number / denorm -> Inf
+ bmsk.f r12,DBL1H,19
+ mov.eq r12,DBL1L
+ mov.eq DBL1L,0
+ sub.eq r7,r7,32
+ norm.f r11,r12 ; flag for x/0 -> Inf check
+ beq_s .Linf_NaN
+ mov.mi r11,0
+ add.pl r11,r11,1
+ add_s r12,r12,r12
+ asl r8,r12,r11
+ rsub r12,r11,31
+ lsr r12,DBL1L,r12
+ tst_s DBL1H,DBL1H
+ or r8,r8,r12
+ lsr r4,r8,26
+ lsr DBL1H,r8,12
+ ld.as r4,[r10,r4]
+ bxor.mi DBL1H,DBL1H,31
+ sub r11,r11,11
+ asl DBL1L,DBL1L,r11
+ sub r11,r11,1
+ mpyhu r5,r4,r8
+ sub r7,r7,r11
+ asl r4,r4,12
+ b.d .Lpast_denorm_dbl1
+ asl r7,r7,20
+ ; wb stall
+
+ .balign 4
+.Ldenorm_dbl0:
+ bmsk.f r12,DBL0H,19
+ ; wb stall
+ mov.eq r12,DBL0L
+ sub.eq r6,r6,32
+ norm.f r11,r12 ; flag for 0/x -> 0 check
+ brge r7, \
+ 0x43500000, .Lret0_NaN ; denorm/large number -> 0
+ beq_s .Lret0_NaN
+ mov.mi r11,0
+ add.pl r11,r11,1
+ asl r12,r12,r11
+ sub r6,r6,r11
+ add.f 0,r6,31
+ lsr r10,DBL0L,r6
+ mov.mi r10,0
+ add r6,r6,11+32
+ neg.f r11,r6
+ asl DBL0L,DBL0L,r11
+ mov.pl DBL0L,0
+ sub r6,r6,32-1
+ b.d .Lpast_denorm_dbl0
+ asl r6,r6,20
+
+.Linf_NaN:
+ tst_s DBL0L,DBL0L ; 0/0 -> NaN
+ xor_s DBL1H,DBL1H,DBL0H
+ bclr.eq.f DBL0H,DBL0H,31
+ bmsk DBL0H,DBL1H,30
+ xor_s DBL0H,DBL0H,DBL1H
+ sub.eq DBL0H,DBL0H,1
+ mov_s DBL0L,0
+ j_s.d [blink]
+ or DBL0H,DBL0H,r9
+ .balign 4
+.Lret0_NaN:
+ xor_s DBL1H,DBL1H,DBL0H
+ cmp_s r12,r9
+ mov_s DBL0L,0
+ bmsk DBL0H,DBL1H,30
+ xor_s DBL0H,DBL0H,DBL1H
+ j_s.d [blink]
+ sub.hi DBL0H,DBL0H,1
+.Linf_nan_dbl1: ; Inf/Inf -> NaN x/Inf-> 0 x/NaN -> NaN
+ not_s DBL0L,DBL1H
+ cmp r6,r9
+ sub_s.ne DBL0L,DBL0L,DBL0L
+ tst_s DBL0H,DBL0H
+ add_s DBL0H,DBL1H,DBL0L
+ j_s.d [blink]
+ bxor.mi DBL0H,DBL0H,31
+.Linf_nan_dbl0:
+ tst_s DBL1H,DBL1H
+ j_s.d [blink]
+ bxor.mi DBL0H,DBL0H,31
+ .balign 4
+ .global __divdf3
+/* N.B. the spacing between divtab and the add3 to get its address must
+ be a multiple of 8. */
+__divdf3:
+ asl r8,DBL1H,12
+ lsr r12,DBL1L,20
+ lsr r4,r8,26
+ add3 r10,pcl,59 ; (.Ldivtab-.) >> 3
+ ld.as r4,[r10,r4]
+ ld.as r9,[pcl,180]; [pcl,(-((.-.L7ff00000) >> 2))] ; 0x7ff00000
+ or r8,r8,r12
+ mpyhu r5,r4,r8
+ and.f r7,DBL1H,r9
+ asl r4,r4,12 ; having the asl here is a concession to the XMAC pipeline.
+ beq.d .Ldenorm_dbl1
+ and r6,DBL0H,r9
+.Lpast_denorm_dbl1: ; wb stall
+ sub r4,r4,r5
+ mpyhu r5,r4,r4
+ breq.d r6,0,.Ldenorm_dbl0
+ lsr r8,r8,1
+ asl r12,DBL0H,11
+ lsr r10,DBL0L,21
+.Lpast_denorm_dbl0: ; wb stall
+ bset r8,r8,31
+ mpyhu r11,r5,r8
+ add_s r12,r12,r10
+ bset r5,r12,31
+ cmp r5,r8
+ cmp.eq DBL0L,DBL1L
+ ; wb stall
+ lsr.cc r5,r5,1
+ sub r4,r4,r11 ; u1.31 inverse, about 30 bit
+ mpyhu r11,r5,r4 ; result fraction highpart
+ breq r7,r9,.Linf_nan_dbl1
+ lsr r8,r8,2 ; u3.29
+ add r5,r6, /* wait for immediate / XMAC wb stall */ \
+ 0x3fe00000
+ ; wb stall (not for XMAC)
+ breq r6,r9,.Linf_nan_dbl0
+ mpyu r12,r11,r8 ; u-28.31
+ asl_s DBL1L,DBL1L,9 ; u-29.23:9
+ sbc r6,r5,r7
+ ; resource conflict (not for XMAC)
+ mpyhu r5,r11,DBL1L ; u-28.23:9
+ add.cs DBL0L,DBL0L,DBL0L
+ asl_s DBL0L,DBL0L,6 ; u-26.25:7
+ asl r10,r11,23
+ sub_l DBL0L,DBL0L,r12
+ ; wb stall (before 'and' for XMAC)
+ lsr r7,r11,9
+ sub r5,DBL0L,r5 ; rest msw ; u-26.31:0
+ mpyh r12,r5,r4 ; result fraction lowpart
+ xor.f 0,DBL0H,DBL1H
+ and DBL0H,r6,r9
+ add_s DBL0H,DBL0H,r7 ; (XMAC wb stall)
+ bxor.mi DBL0H,DBL0H,31
+ brhs r6, /* wb stall / wait for immediate */ \
+ 0x7fe00000,.Linf_denorm
+ add.f r12,r12,0x11
+ asr r9,r12,5
+ sub.mi DBL0H,DBL0H,1
+ add.f DBL0L,r9,r10
+ tst r12,0x1c
+ jne.d [blink]
+ add.cs DBL0H,DBL0H,1
+ /* work out exact rounding if we fall through here. */
+ /* We know that the exact result cannot be represented in double
+ precision. Find the mid-point between the two nearest
+ representable values, multiply with the divisor, and check if
+ the result is larger than the dividend. Since we want to know
+ only the sign bit, it is sufficient to calculate only the
+ highpart of the lower 64 bits. */
+ sub.f DBL0L,DBL0L,1
+ asl r12,r9,2 ; u-22.30:2
+ mpyu r10,r11,DBL1L ; rest before considering r12 in r5 : -r10
+ sub.cs DBL0H,DBL0H,1
+ sub.f r12,r12,2
+ ; resource conflict (not for XMAC)
+ mpyhu r7,r12,DBL1L ; u-51.32
+ asl r5,r5,25 ; s-51.7:25
+ lsr r10,r10,7 ; u-51.30:2
+ ; resource conflict (not for XMAC)
+ ; resource conflict (not for XMAC)
+ mpyu r9,r12,r8 ; u-51.31:1
+ sub r5,r5,r10
+ add.mi r5,r5,DBL1L ; signed multiply adjust for r12*DBL1L
+ bset r7,r7,0 ; make sure that the result is not zero, and that
+ ; wb stall (one earlier for XMAC)
+ sub r5,r5,r7 ; a highpart zero appears negative
+ sub.f r5,r5,r9 ; rest msw
+ add.pl.f DBL0L,DBL0L,1
+ j_s.d [blink]
+ add.eq DBL0H,DBL0H,1
+
+ .balign 4
+.Linf_denorm:
+ brlo r6,0xc0000000,.Linf
+.Ldenorm:
+ asr r6,r6,20
+ neg r9,r6
+ mov_s DBL0H,0
+ brhs.d r9,54,.Lret0
+ bxor.mi DBL0H,DBL0H,31
+ add_l r12,r12,1
+ and r12,r12,-4
+ rsub r7,r6,5
+ asr r10,r12,28
+ bmsk r4,r12,27
+ asrs DBL0L,r4,r7
+ add DBL1H,r11,r10
+ add.f r7,r6,32-5
+ abss r10,r4
+ asl r4,r4,r7
+ mov.mi r4,r10
+ add.f r10,r6,23
+ rsub r7,r6,9
+ lsr r7,DBL1H,r7
+ asl r10,DBL1H,r10
+ or.pnz DBL0H,DBL0H,r7
+ or.mi r4,r4,r10
+ mov.mi r10,r7
+ add.f DBL0L,r10,DBL0L
+ add.cs.f DBL0H,DBL0H,1 ; carry clear after this point
+ bxor.f 0,r4,31
+ add.pnz.f DBL0L,DBL0L,1
+ add.cs.f DBL0H,DBL0H,1
+ jne_l [blink]
+ /* Calculation so far was not conclusive; calculate further rest. */
+ mpyu r11,r11,DBL1L ; rest before considering r12 in r5 : -r11
+ asr.f r12,r12,3
+ asl r5,r5,25 ; s-51.7:25
+ ; resource conflict (not for XMAC)
+ mpyu DBL1H,r12,r8 ; u-51.31:1
+ and r9,DBL0L,1 ; tie-breaker: round to even
+ lsr r11,r11,7 ; u-51.30:2
+ ; resource conflict (not for XMAC)
+ mpyhu r8,r12,DBL1L ; u-51.32
+ sub.mi r11,r11,DBL1L ; signed multiply adjust for r12*DBL1L
+ add_s DBL1H,DBL1H,r11
+ ; resource conflict (not for XMAC)
+ ; resource conflict (not for XMAC)
+ mpyu r12,r12,DBL1L ; u-83.30:2
+ sub DBL1H,DBL1H,r5 ; -rest msw
+ add_s DBL1H,DBL1H,r8 ; -rest msw
+ add.f 0,DBL1H,DBL1H ; can't ror.f by 32 :-(
+ ; wb stall (XMAC: Before add.f)
+ tst_s DBL1H,DBL1H
+ cmp.eq r12,r9
+ add.cs.f DBL0L,DBL0L,1
+ j_s.d [blink]
+ add.cs DBL0H,DBL0H,1
+
+.Lret0:
+ /* return +- 0 */
+ j_s.d [blink]
+ mov_s DBL0L,0
+.Linf:
+ mov_s DBL0H,r9
+ mov_s DBL0L,0
+ j_s.d [blink]
+ bxor.mi DBL0H,DBL0H,31
+
+ .balign 4
+.Ldivtab:
+ .long 0xfc0fffe1
+ .long 0xf46ffdfb
+ .long 0xed1ffa54
+ .long 0xe61ff515
+ .long 0xdf7fee75
+ .long 0xd91fe680
+ .long 0xd2ffdd52
+ .long 0xcd1fd30c
+ .long 0xc77fc7cd
+ .long 0xc21fbbb6
+ .long 0xbcefaec0
+ .long 0xb7efa100
+ .long 0xb32f92bf
+ .long 0xae8f83b7
+ .long 0xaa2f7467
+ .long 0xa5ef6479
+ .long 0xa1cf53fa
+ .long 0x9ddf433e
+ .long 0x9a0f3216
+ .long 0x965f2091
+ .long 0x92df0f11
+ .long 0x8f6efd05
+ .long 0x8c1eeacc
+ .long 0x88eed876
+ .long 0x85dec615
+ .long 0x82eeb3b9
+ .long 0x800ea10b
+ .long 0x7d3e8e0f
+ .long 0x7a8e7b3f
+ .long 0x77ee6836
+ .long 0x756e5576
+ .long 0x72fe4293
+ .long 0x709e2f93
+ .long 0x6e4e1c7f
+ .long 0x6c0e095e
+ .long 0x69edf6c5
+ .long 0x67cde3a5
+ .long 0x65cdd125
+ .long 0x63cdbe25
+ .long 0x61ddab3f
+ .long 0x600d991f
+ .long 0x5e3d868c
+ .long 0x5c6d7384
+ .long 0x5abd615f
+ .long 0x590d4ecd
+ .long 0x576d3c83
+ .long 0x55dd2a89
+ .long 0x545d18e9
+ .long 0x52dd06e9
+ .long 0x516cf54e
+ .long 0x4ffce356
+ .long 0x4e9cd1ce
+ .long 0x4d3cbfec
+ .long 0x4becae86
+ .long 0x4aac9da4
+ .long 0x496c8c73
+ .long 0x483c7bd3
+ .long 0x470c6ae8
+ .long 0x45dc59af
+ .long 0x44bc4915
+ .long 0x43ac3924
+ .long 0x428c27fb
+ .long 0x418c187a
+ .long 0x407c07bd
+.L7ff00000:
+ .long 0x7ff00000
+ ENDFUNC(__divdf3)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+/*
+ - calculate 15..18 bit inverse using a table of approximating polynoms.
+ precision is higher for polynoms used to evaluate input with larger
+ value.
+ - do one newton-raphson iteration step to double the precision,
+ then multiply this with the divisor
+ -> more time to decide if dividend is subnormal
+ - the worst error propagation is on the side of the value range
+ with the least initial defect, thus giving us about 30 bits precision.
+ */
+#include "arc-ieee-754.h"
+
+#if 0 /* DEBUG */
+ .global __divsf3
+ FUNC(__divsf3)
+ .balign 4
+__divsf3:
+ push_s blink
+ push_s r1
+ bl.d __divsf3_c
+ push_s r0
+ ld_s r1,[sp,4]
+ st_s r0,[sp,4]
+ bl.d __divsf3_asm
+ pop_s r0
+ pop_s r1
+ pop_s blink
+ cmp r0,r1
+#if 1
+ bne abort
+ jeq_s [blink]
+ b abort
+#else
+ bne abort
+ j_s [blink]
+#endif
+ ENDFUNC(__divsf3)
+#define __divsf3 __divsf3_asm
+#endif /* DEBUG */
+
+ FUNC(__divsf3)
+ .balign 4
+.L7f800000:
+ .long 0x7f800000
+.Ldivtab:
+ .long 0xfc0ffff0
+ .long 0xf46ffefd
+ .long 0xed1ffd2a
+ .long 0xe627fa8e
+ .long 0xdf7ff73b
+ .long 0xd917f33b
+ .long 0xd2f7eea3
+ .long 0xcd1fe986
+ .long 0xc77fe3e7
+ .long 0xc21fdddb
+ .long 0xbcefd760
+ .long 0xb7f7d08c
+ .long 0xb32fc960
+ .long 0xae97c1ea
+ .long 0xaa27ba26
+ .long 0xa5e7b22e
+ .long 0xa1cfa9fe
+ .long 0x9ddfa1a0
+ .long 0x9a0f990c
+ .long 0x9667905d
+ .long 0x92df878a
+ .long 0x8f6f7e84
+ .long 0x8c27757e
+ .long 0x88f76c54
+ .long 0x85df630c
+ .long 0x82e759c5
+ .long 0x8007506d
+ .long 0x7d3f470a
+ .long 0x7a8f3da2
+ .long 0x77ef341e
+ .long 0x756f2abe
+ .long 0x72f7212d
+ .long 0x709717ad
+ .long 0x6e4f0e44
+ .long 0x6c1704d6
+ .long 0x69e6fb44
+ .long 0x67cef1d7
+ .long 0x65c6e872
+ .long 0x63cedf18
+ .long 0x61e6d5cd
+ .long 0x6006cc6d
+ .long 0x5e36c323
+ .long 0x5c76b9f3
+ .long 0x5abeb0b7
+ .long 0x5916a79b
+ .long 0x57769e77
+ .long 0x55de954d
+ .long 0x54568c4e
+ .long 0x52d6834d
+ .long 0x51667a7f
+ .long 0x4ffe71b5
+ .long 0x4e9e68f1
+ .long 0x4d466035
+ .long 0x4bf65784
+ .long 0x4aae4ede
+ .long 0x496e4646
+ .long 0x48363dbd
+ .long 0x47063547
+ .long 0x45de2ce5
+ .long 0x44be2498
+ .long 0x43a61c64
+ .long 0x4296144a
+ .long 0x41860c0e
+ .long 0x407e03ee
+__divsf3_support: /* This label makes debugger output saner. */
+.Ldenorm_fp1:
+ bclr r6,r6,31
+ norm.f r12,r6 ; flag for x/0 -> Inf check
+ add r6,r6,r6
+ rsub r5,r12,16
+ ror r5,r1,r5
+ asl r6,r6,r12
+ bmsk r5,r5,5
+ ld.as r5,[r3,r5]
+ add r4,r6,r6
+ ; load latency
+ mpyhu r7,r5,r4
+ bic.ne.f 0, \
+ 0x60000000,r0 ; large number / denorm -> Inf
+ beq_s .Linf_NaN
+ asl r5,r5,13
+ ; wb stall
+ ; slow track
+ sub r7,r5,r7
+ mpyhu r8,r7,r6
+ asl_s r12,r12,23
+ and.f r2,r0,r9
+ add r2,r2,r12
+ asl r12,r0,8
+ ; wb stall
+ bne.d .Lpast_denorm_fp1
+.Ldenorm_fp0:
+ mpyhu r8,r8,r7
+ bclr r12,r12,31
+ norm.f r3,r12 ; flag for 0/x -> 0 check
+ bic.ne.f 0,0x60000000,r1 ; denorm/large number -> 0
+ beq_s .Lret0
+ asl_s r12,r12,r3
+ asl_s r3,r3,23
+ add_s r12,r12,r12
+ add r11,r11,r3
+ b.d .Lpast_denorm_fp0
+ mov_s r3,r12
+ .balign 4
+.Linf_NaN:
+ bclr.f 0,r0,31 ; 0/0 -> NaN
+ xor_s r0,r0,r1
+ bmsk r1,r0,30
+ bic_s r0,r0,r1
+ sub.eq r0,r0,1
+ j_s.d [blink]
+ or r0,r0,r9
+.Lret0:
+ xor_s r0,r0,r1
+ bmsk r1,r0,30
+ j_s.d [blink]
+ bic_s r0,r0,r1
+.Linf_nan_fp1:
+ lsr_s r0,r0,31
+ bmsk.f 0,r1,22
+ asl_s r0,r0,31
+ bne_s 0f ; inf/inf -> nan
+ brne r2,r9,.Lsigned0 ; x/inf -> 0, but x/nan -> nan
+0: j_s.d [blink]
+ mov r0,-1
+.Lsigned0:
+.Linf_nan_fp0:
+ tst_s r1,r1
+ j_s.d [blink]
+ bxor.mi r0,r0,31
+ .balign 4
+ .global __divsf3
+/* N.B. the spacing between divtab and the sub3 to get its address must
+ be a multiple of 8. */
+__divsf3:
+ lsr r2,r1,17
+ sub3 r3,pcl,55;(.-.Ldivtab) >> 3
+ bmsk_s r2,r2,5
+ ld.as r5,[r3,r2]
+ asl r4,r1,9
+ ld.as r9,[pcl,-114]; [pcl,(-((.-.L7f800000) >> 2))] ; 0x7f800000
+ mpyhu r7,r5,r4
+ asl r6,r1,8
+ and.f r11,r1,r9
+ bset r6,r6,31
+ asl r5,r5,13
+ ; wb stall
+ beq .Ldenorm_fp1
+ sub r7,r5,r7
+ mpyhu r8,r7,r6
+ breq.d r11,r9,.Linf_nan_fp1
+ and.f r2,r0,r9
+ beq.d .Ldenorm_fp0
+ asl r12,r0,8
+ ; wb stall
+ breq r2,r9,.Linf_nan_fp0
+ mpyhu r8,r8,r7
+.Lpast_denorm_fp1:
+ bset r3,r12,31
+.Lpast_denorm_fp0:
+ cmp_s r3,r6
+ lsr.cc r3,r3,1
+ add_s r2,r2, /* wait for immediate */ \
+ /* wb stall */ \
+ 0x3f000000
+ sub r7,r7,r8 ; u1.31 inverse, about 30 bit
+ mpyhu r3,r3,r7
+ sbc r2,r2,r11
+ xor.f 0,r0,r1
+ and r0,r2,r9
+ bxor.mi r0,r0,31
+ brhs r2, /* wb stall / wait for immediate */ \
+ 0x7f000000,.Linf_denorm
+.Lpast_denorm:
+ add_s r3,r3,0x22 ; round to nearest or higher
+ tst r3,0x3c ; check if rounding was unsafe
+ lsr r3,r3,6
+ jne.d [blink] ; return if rounding was safe.
+ add_s r0,r0,r3
+ /* work out exact rounding if we fall through here. */
+ /* We know that the exact result cannot be represented in single
+ precision. Find the mid-point between the two nearest
+ representable values, multiply with the divisor, and check if
+ the result is larger than the dividend. */
+ add_s r3,r3,r3
+ sub_s r3,r3,1
+ mpyu r3,r3,r6
+ asr.f 0,r0,1 ; for round-to-even in case this is a denorm
+ rsub r2,r9,25
+ asl_s r12,r12,r2
+ ; wb stall
+ ; slow track
+ sub.f 0,r12,r3
+ j_s.d [blink]
+ sub.mi r0,r0,1
+/* For denormal results, it is possible that an exact result needs
+ rounding, and thus the round-to-even rule has to come into play. */
+.Linf_denorm:
+ brlo r2,0xc0000000,.Linf
+.Ldenorm:
+ asr_s r2,r2,23
+ bic r0,r0,r9
+ neg r9,r2
+ brlo.d r9,25,.Lpast_denorm
+ lsr r3,r3,r9
+ /* Fall through: return +- 0 */
+ j_s [blink]
+.Linf:
+ j_s.d [blink]
+ or r0,r0,r9
+ ENDFUNC(__divsf3)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "arc-ieee-754.h"
+
+#if 0 /* DEBUG */
+ .global __divsf3
+ FUNC(__divsf3)
+ .balign 4
+__divsf3:
+ push_s blink
+ push_s r1
+ bl.d __divsf3_c
+ push_s r0
+ ld_s r1,[sp,4]
+ st_s r0,[sp,4]
+ bl.d __divsf3_asm
+ pop_s r0
+ pop_s r1
+ pop_s blink
+ cmp r0,r1
+#if 1
+ bne abort
+ jeq_s [blink]
+ b abort
+#else
+ bne abort
+ j_s [blink]
+#endif
+ ENDFUNC(__divsf3)
+#define __divsf3 __divsf3_asm
+#endif /* DEBUG */
+
+ .balign 4
+__divdf3_support: /* This label makes debugger output saner. */
+ FUNC(__divsf3)
+.Ldenorm_fp0:
+ norm.f r12,r2 ; flag for 0/x -> 0 check
+ bic.ne.f 0,0x60000000,r1 ; denorm/large number -> 0
+ beq_s .Lret0_NaN
+ tst r1,r9
+ add_s r2,r2,r2
+ sub_s r12,r12,8
+ asl_s r2,r2,r12
+ asl_l r12,r12,23
+ bne.d .Lpast_denorm_fp0
+ add r5,r5,r12
+/* r0 is subnormal, r1 is subnormal or 0. */
+
+ .balign 4
+.Ldenorm_fp1:
+ norm.f r12,r3 ; flag for x/0 -> Inf check
+ bic.ne.f 0,0x60000000,r0 ; large number/denorm -> Inf
+ beq_s .Linf
+ add_s r3,r3,r3
+ sub_s r12,r12,8
+ asl_s r3,r3,r12
+ asl_s r12,r12,23
+ b.d .Lpast_denorm_fp1
+ add r4,r4,r12
+
+.Lret0_NaN:
+ bclr.f 0,r1,31 ; 0/0 -> NaN
+ bic r0,r10,r9
+ j_s.d [blink]
+ sub.eq r0,r0,1
+
+ .global __divsf3
+ .balign 4
+ .long 0x7f800000 ; exponent mask
+__divsf3:
+ ld r9,[pcl,-4]
+ bmsk r2,r0,22
+ xor r4,r0,r2
+ bmsk r3,r1,22
+ xor r5,r1,r3
+ and r11,r0,r9
+ breq.d r11,0,.Ldenorm_fp0
+ xor r10,r4,r5
+ breq r11,r9,.Linf_nan_fp0
+ bset_s r2,r2,23
+ and r11,r1,r9
+ breq r11,0,.Ldenorm_fp1
+ breq r11,r9,.Linf_nan_fp1
+.Lpast_denorm_fp0:
+ bset_s r3,r3,23
+.Lpast_denorm_fp1:
+ cmp r2,r3
+ asl_s r2,r2,6+1
+ asl_s r3,r3,7
+ add.lo r2,r2,r2
+ bclr r8,r9,30 ; exponent bias
+ bclr.lo r8,r8,23 ; reduce exp by one if fraction is shifted
+ sub r4,r4,r5
+ add r4,r4,r8
+ xor.f 0,r10,r4
+ bmi .Linf_denorm
+ and r12,r4,r9
+ breq r12,0,.Ldenorm
+ sub_s r2,r2,r3 ; discard implicit 1
+.Ldiv_23bit:
+ .rep 6
+ divaw r2,r2,r3
+ .endr
+ breq r12,r9,.Linf
+ bmsk r0,r2,6
+ xor_s r2,r2,r0
+.Ldiv_17bit:
+ .rep 7
+ divaw r2,r2,r3
+ .endr
+ asl_s r0,r0,7
+ bmsk r1,r2,6
+ xor_s r2,r2,r1
+ or_s r0,r0,r1
+.Ldiv_10bit:
+ .rep 7
+ divaw r2,r2,r3
+ .endr
+ asl_s r0,r0,7
+ bmsk r1,r2,6
+ xor_s r2,r2,r1
+ or_s r0,r0,r1
+.Ldiv_3bit:
+ .rep 3
+ divaw r2,r2,r3
+ .endr
+ asl_s r0,r0,3
+.Ldiv_0bit:
+ divaw r1,r2,r3
+ bmsk_s r2,r2,2
+ tst r1,-0x7e ; 0xffffff82, test for rest or odd
+ bmsk_s r1,r1,0
+ add_s r0,r0,r2 ; assemble fraction
+ add_s r0,r0,r4 ; add in sign & exponent
+ j_s.d [blink]
+ add.ne r0,r0,r1 ; round to nearest / even
+
+ .balign 4
+.Linf_nan_fp0:
+ bic.f 0,r9,r1 ; fp1 Inf -> result NaN
+ bic r1,r5,r9 ; fp1 sign
+ sub.eq r1,r1,1
+ j_s.d [blink]
+ xor_s r0,r0,r1
+.Linf_nan_fp1:
+ bic r0,r4,r9 ; fp0 sign
+ bmsk.f 0,r1,22 ; x/inf -> 0, x/nan -> nan
+ xor.eq r1,r1,r9
+ j_s.d [blink]
+ xor_s r0,r0,r1
+.Linf:
+ j_s.d [blink]
+ or r0,r10,r9
+
+.Lret_r4:
+ j_s.d [blink]
+ mov_s r0,r4
+ .balign 4
+.Linf_denorm:
+ add.f r12,r4,r4
+ asr_l r12,r12,24
+ bpl .Linf
+ max r12,r12,-24
+.Ldenorm:
+ add r1,pcl,42; .Ldenorm_tab-.
+ ldb_s r12,[r12,r1]
+ mov_s r0,0
+ lsr_s r2,r2
+ sub_s r1,r1,r12
+ j_s.d [r1]
+ bic r4,r10,r9
+ .byte .Ldenorm_tab-.Lret_r4
+ .byte .Ldenorm_tab-.Ldiv_0bit
+ .byte .Ldenorm_tab-.Ldiv_3bit-8
+ .byte .Ldenorm_tab-.Ldiv_3bit-4
+ .byte .Ldenorm_tab-.Ldiv_3bit
+ .byte .Ldenorm_tab-.Ldiv_10bit-24
+ .byte .Ldenorm_tab-.Ldiv_10bit-20
+ .byte .Ldenorm_tab-.Ldiv_10bit-16
+ .byte .Ldenorm_tab-.Ldiv_10bit-12
+ .byte .Ldenorm_tab-.Ldiv_10bit-8
+ .byte .Ldenorm_tab-.Ldiv_10bit-4
+ .byte .Ldenorm_tab-.Ldiv_10bit
+ .byte .Ldenorm_tab-.Ldiv_17bit-24
+ .byte .Ldenorm_tab-.Ldiv_17bit-20
+ .byte .Ldenorm_tab-.Ldiv_17bit-16
+ .byte .Ldenorm_tab-.Ldiv_17bit-12
+ .byte .Ldenorm_tab-.Ldiv_17bit-8
+ .byte .Ldenorm_tab-.Ldiv_17bit-4
+ .byte .Ldenorm_tab-.Ldiv_17bit
+ .byte .Ldenorm_tab-.Ldiv_23bit-20
+ .byte .Ldenorm_tab-.Ldiv_23bit-16
+ .byte .Ldenorm_tab-.Ldiv_23bit-12
+ .byte .Ldenorm_tab-.Ldiv_23bit-8
+ .byte .Ldenorm_tab-.Ldiv_23bit-4
+.Ldenorm_tab:
+ .byte .Ldenorm_tab-.Ldiv_23bit
+ ENDFUNC(__divsf3)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+/* We use a polynom similar to a Tchebycheff polynom to get an initial
+ seed, and then use a newton-raphson iteration step to get an
+ approximate result
+ If this result can't be rounded to the exact result with confidence, we
+ round to the value between the two closest representable values, and
+ test if the correctly rounded value is above or below this value.
+
+ Because of the Newton-raphson iteration step, an error in the seed at X
+ is amplified by X. Therefore, we don't want a Tchebycheff polynom
+ or a polynom that is close to optimal according to the maximum norm
+ on the errro of the seed value; we want one that is close to optimal
+ according to the maximum norm on the error of the result, i.e. we
+ want the maxima of the polynom to increase linearily.
+ Given an interval [X0,X2) over which to approximate,
+ with X1 := (X0+X2)/2, D := X1-X0, F := 1/D, and S := D/X1 we have,
+ like for Tchebycheff polynoms:
+ P(0) := 1
+ but then we have:
+ P(1) := X + S*D
+ P(2) := 2 * X^2 + S*D * X - D^2
+ Then again:
+ P(n+1) := 2 * X * P(n) - D^2 * P (n-1)
+ */
+
+static long double merr = 42.;
+
+double
+err (long double a0, long double a1, long double x)
+{
+ long double y0 = a0 + (x-1)*a1;
+
+ long double approx = 2. * y0 - y0 * x * y0;
+ long double true = 1./x;
+ long double err = approx - true;
+
+ if (err <= -1./65536./16384.)
+ printf ("ERROR EXCEEDS 1 ULP %.15f %.15f %.15f\n",
+ (double)x, (double)approx, (double)true);
+ if (merr > err)
+ merr = err;
+ return err;
+}
+
+int
+main (void)
+{
+ long double T[5]; /* Taylor polynom */
+ long double P[5][5];
+ int i, j;
+ long double X0, X1, X2, S;
+ long double inc = 1./64;
+ long double D = inc*0.5;
+ long i0, i1, i2, io;
+
+ memset (P, 0, sizeof (P));
+ P[0][0] = 1.;
+ for (i = 1; i < 5; i++)
+ P[i][i] = 1 << i-1;
+ P[2][0] = -D*D;
+ for (X0 = 1.; X0 < 2.; X0 += inc)
+ {
+ X1 = X0 + inc * 0.5;
+ X2 = X0 + inc;
+ S = D / X1;
+ T[0] = 1./X1;
+ for (i = 1; i < 5; i++)
+ T[i] = T[i-1] * -T[0];
+#if 0
+ printf ("T %1.8f %f %f %f %f\n", (double)T[0], (double)T[1], (double)T[2],
+(double)T[3], (double)T[4]);
+#endif
+ P[1][0] = S*D;
+ P[2][1] = S*D;
+ for (i = 3; i < 5; i++)
+ {
+ P[i][0] = -D*D*P[i-2][0];
+ for (j = 1; j < i; j++)
+ P[i][j] = 2*P[i-1][j-1]-D*D*P[i-2][j];
+ }
+#if 0
+ printf ("P3 %1.8f %f %f %f %f\n", (double)P[3][0], (double)P[3][1], (double)P[3][2],
+(double)P[3][3], (double)P[3][4]);
+ printf ("P4 %1.8f %f %f %f %f\n", (double)P[4][0], (double)P[4][1], (double)P[4][2],
+(double)P[4][3], (double)P[4][4]);
+#endif
+ for (i = 4; i > 1; i--)
+ {
+ long double a = T[i]/P[i][i];
+
+ for (j = 0; j < i; j++)
+ T[j] -= a * P[i][j];
+ }
+#if 0
+ printf ("A %1.8f %f %f\n", (double)T[0], (double)T[1], (double)T[2]);
+#endif
+#if 0
+ i2 = T[2]*1024;
+ long double a = (T[2]-i/1024.)/P[2][2];
+ for (j = 0; j < 2; j++)
+ T[j] -= a * P[2][j];
+#else
+ i2 = 0;
+#endif
+ long double T0, Ti1;
+ for (i = 0, i0 = 0; i < 4; i++)
+ {
+
+ i1 = T[1]*4096. + i0 / (long double)(1 << 20) - 0.5;
+ i1 = - (-i1 & 0x0fff);
+ Ti1 = ((unsigned)(-i1 << 20) | i0) /-(long double)(1LL<<32LL);
+ T0 = T[0] - (T[1]-Ti1)/P[1][1] * P[1][0] - (X1 - 1) * Ti1;
+ i0 = T0 * 1024 * 1024 + 0.5;
+ i0 &= 0xfffff;
+ }
+#if 0
+ printf ("A %1.8f %f %f\n", (double)T[0], (double)T[1], (double)T[2]);
+#endif
+ io = (unsigned)(-i1 << 20) | i0;
+ long double A1 = (unsigned)io/-65536./65536.;
+ long double A0 = (unsigned)(io << 12)/65536./65536.;
+ long double Xm0 = 1./sqrt (-A1);
+ long double Xm1 = 0.5+0.5*-A0/A1;
+#if 0
+ printf ("%f %f %f %f\n", (double)A0, (double)A1, (double) Ti1, (double)X0);
+ printf ("%.12f %.12f %.12f\n",
+ err (A0, A1, X0), err (A0, A1, X1), err (A0, A1, X2));
+ printf ("%.12f %.12f\n", (double)Xm0, (double)Xm1);
+ printf ("%.12f %.12f\n", err (A0, A1, Xm0), err (A0, A1, Xm1));
+#endif
+ printf ("\t.long 0x%x\n", io);
+ }
+#if 0
+ printf ("maximum error: %.15f %x %f\n", (double)merr, (unsigned)(long long)(-merr * 65536 * 65536), (double)log(-merr)/log(2));
+#endif
+ return 0;
+}
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+/* We use a polynom similar to a Tchebycheff polynom to get an initial
+ seed, and then use a newton-raphson iteration step to get an
+ approximate result
+ If this result can't be rounded to the exact result with confidence, we
+ round to the value between the two closest representable values, and
+ test if the correctly rounded value is above or below this value.
+
+ Because of the Newton-raphson iteration step, an error in the seed at X
+ is amplified by X. Therefore, we don't want a Tchebycheff polynom
+ or a polynom that is close to optimal according to the maximum norm
+ on the errro of the seed value; we want one that is close to optimal
+ according to the maximum norm on the error of the result, i.e. we
+ want the maxima of the polynom to increase linearily.
+ Given an interval [X0,X2) over which to approximate,
+ with X1 := (X0+X2)/2, D := X1-X0, F := 1/D, and S := D/X1 we have,
+ like for Tchebycheff polynoms:
+ P(0) := 1
+ but then we have:
+ P(1) := X + S*D
+ P(2) := 2 * X^2 + S*D * X - D^2
+ Then again:
+ P(n+1) := 2 * X * P(n) - D^2 * P (n-1)
+ */
+
+int
+main (void)
+{
+ long double T[5]; /* Taylor polynom */
+ long double P[5][5];
+ int i, j;
+ long double X0, X1, X2, S;
+ long double inc = 1./64;
+ long double D = inc*0.5;
+ long i0, i1, i2;
+
+ memset (P, 0, sizeof (P));
+ P[0][0] = 1.;
+ for (i = 1; i < 5; i++)
+ P[i][i] = 1 << i-1;
+ P[2][0] = -D*D;
+ for (X0 = 1.; X0 < 2.; X0 += inc)
+ {
+ X1 = X0 + inc * 0.5;
+ X2 = X1 + inc;
+ S = D / X1;
+ T[0] = 1./X1;
+ for (i = 1; i < 5; i++)
+ T[i] = T[i-1] * -T[0];
+#if 0
+ printf ("T %1.8f %f %f %f %f\n", (double)T[0], (double)T[1], (double)T[2],
+(double)T[3], (double)T[4]);
+#endif
+ P[1][0] = S*D;
+ P[2][1] = S*D;
+ for (i = 3; i < 5; i++)
+ {
+ P[i][0] = -D*D*P[i-2][0];
+ for (j = 1; j < i; j++)
+ P[i][j] = 2*P[i-1][j-1]-D*D*P[i-2][j];
+ }
+#if 0
+ printf ("P3 %1.8f %f %f %f %f\n", (double)P[3][0], (double)P[3][1], (double)P[3][2],
+(double)P[3][3], (double)P[3][4]);
+ printf ("P4 %1.8f %f %f %f %f\n", (double)P[4][0], (double)P[4][1], (double)P[4][2],
+(double)P[4][3], (double)P[4][4]);
+#endif
+ for (i = 4; i > 1; i--)
+ {
+ long double a = T[i]/P[i][i];
+
+ for (j = 0; j < i; j++)
+ T[j] -= a * P[i][j];
+ }
+#if 0
+ printf ("A %1.8f %f %f\n", (double)T[0], (double)T[1], (double)T[2]);
+#endif
+#if 0
+ i2 = T[2]*512;
+ long double a = (T[2]-i/512.)/P[2][2];
+ for (j = 0; j < 2; j++)
+ T[j] -= a * P[2][j];
+#else
+ i2 = 0;
+#endif
+ for (i = 0, i0 = 0; i < 4; i++)
+ {
+ long double T0, Ti1;
+
+ i1 = T[1]*8192. + i0 / (long double)(1 << 19) - 0.5;
+ i1 = - (-i1 & 0x1fff);
+ Ti1 = ((unsigned)(-i1 << 19) | i0) /-(long double)(1LL<<32LL);
+ T0 = T[0] - (T[1]-Ti1)/P[1][1] * P[1][0] - (X1 - 1) * Ti1;
+ i0 = T0 * 512 * 1024 + 0.5;
+ i0 &= 0x7ffff;
+ }
+#if 0
+ printf ("A %1.8f %f %f\n", (double)T[0], (double)T[1], (double)T[2]);
+#endif
+ printf ("\t.long 0x%x\n", (-i1 << 19) | i0);
+ }
+ return 0;
+}
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "arc-ieee-754.h"
+/* inputs: DBL0, DBL1
+ output: z flag
+ clobber: r12, flags
+ For NaNs, bit 19.. bit 30 of the high word must be set. */
+#if 0 /* DEBUG */
+ .global __eqdf2
+ .balign 4
+ FUNC(__eqdf2)
+__eqdf2:
+ st.a r11,[sp,-4]` push_s blink` st.a r10,[sp,-4]` st.a r9,[sp,-4]
+ st.a r8,[sp,-4]` st.a r7,[sp,-4]` st.a r6,[sp,-4]` st.a r5,[sp,-4]
+ st.a r4,[sp,-4]` push_s r3` push_s r2` push_s r1`
+ bl.d __eqdf2_c` push_s r0
+ mov r11,r0` pop_s r0` pop_s r1` pop_s r2` pop_s r3
+ ld.ab r4,[sp,4]` ld.ab r5,[sp,4]` ld.ab r6,[sp,4]`
+ ld.ab r7,[sp,4]` ld.ab r8,[sp,4]` ld.ab r9,[sp,4]
+ bl.d __eqdf2_asm` ld.ab r10,[sp,4]
+ pop_s blink
+ breq.d r11,0,0f
+ ld.ab r11,[sp,4]
+ jne_s [blink]
+ bl abort
+0: jeq_s [blink]
+ bl abort
+ ENDFUNC(__eqdf2)
+#define __eqdf2 __eqdf2_asm
+#endif /* DEBUG */
+ .global __eqdf2
+ .balign 4
+ HIDDEN_FUNC(__eqdf2)
+ /* Good performance as long as the difference in high word is
+ well predictable (as seen from the branch predictor). */
+__eqdf2:
+ brne.d DBL0H,DBL1H,.Lhighdiff
+ bmsk r12,DBL0H,20
+#ifdef DPFP_COMPAT
+ or.f 0,DBL0L,DBL1L
+ bset.ne r12,r12,21
+#endif /* DPFP_COMPAT */
+ add1.f r12,r12,DBL0H /* set c iff NaN; also, clear z if NaN. */
+ j_s.d [blink]
+ cmp.cc DBL0L,DBL1L
+ .balign 4
+.Lhighdiff:
+ or r12,DBL0H,DBL1H
+ or.f 0,DBL0L,DBL1L
+ j_s.d [blink]
+ bmsk.eq.f r12,r12,30
+ ENDFUNC(__eqdf2)
+/* ??? could we do better by speeding up some 'common' case of inequality? */
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "arc-ieee-754.h"
+/* inputs: r0, r1
+ output: z flag
+ clobber: r12, flags
+ For NaNs, bit 22 .. bit 30 must be set. */
+#if 0 /* DEBUG */
+ .global __eqsf2
+ .balign 4
+ FUNC(__eqsf2)
+__eqsf2:
+ st.a r11,[sp,-4]` push_s blink` st.a r10,[sp,-4]` st.a r9,[sp,-4]
+ st.a r8,[sp,-4]` st.a r7,[sp,-4]` st.a r6,[sp,-4]` st.a r5,[sp,-4]
+ st.a r4,[sp,-4]` push_s r3` push_s r2` push_s r1`
+ bl.d __eqsf2_c` push_s r0
+ mov r11,r0` pop_s r0` pop_s r1` pop_s r2` pop_s r3
+ ld.ab r4,[sp,4]` ld.ab r5,[sp,4]` ld.ab r6,[sp,4]`
+ ld.ab r7,[sp,4]` ld.ab r8,[sp,4]` ld.ab r9,[sp,4]
+ bl.d __eqsf2_asm` ld.ab r10,[sp,4]
+ pop_s blink
+ breq.d r11,0,0f
+ ld.ab r11,[sp,4]
+ jne_s [blink]
+ bl abort
+0: jeq_s [blink]
+ bl abort
+ ENDFUNC(__eqsf2)
+#define __eqsf2 __eqsf2_asm
+#endif /* DEBUG */
+ /* Good performance as long as the binary difference is
+ well predictable (as seen from the branch predictor). */
+ .global __eqsf2
+ .balign 4
+ HIDDEN_FUNC(__eqsf2)
+__eqsf2:
+ breq r0, r1,.Lno_bdiff
+ or r12,r0,r1
+ j_s.d [blink]
+ bmsk.f 0,r12,30
+.Lno_bdiff:
+ bmsk r12,r0,23
+ add1.f r12,r12,r0 /* set c iff NaN; also, clear z if NaN. */
+ j_s.d [blink]
+ cmp.cc r0,r1
+ ENDFUNC(__eqsf2)
--- /dev/null
+/* Copyright (C) 2006, 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "arc-ieee-754.h"
+
+#if 0 /* DEBUG */
+ .global __extendsfdf2
+ .balign 4
+ FUNC(__extendsfdf2)
+__extendsfdf2:
+ push_s blink
+ bl.d __extendsfdf2_c
+ push_s r0
+ ld_s r2,[sp]
+ st_s r1,[sp]
+ push_s r0
+ bl.d __extendsfdf2_asm
+ mov_s r0,r2
+ pop_s r2
+ pop_s r3
+ pop_s blink
+ cmp r0,r2
+ cmp.eq r1,r3
+ jeq_s [blink]
+ bl abort
+ ENDFUNC(__extendsfdf2)
+#define __extendsfdf2 __extendsfdf2_asm
+#endif /* DEBUG */
+#if 0 /* ARC600 */
+__extendsfdf2:
+ lsr r2,r0,23
+ tst r2,0xff
+ bic.ne.f r2,0xff
+ beq_s .Linf_nan_denorm_0
+..
+.Linf_nan_denorm:
+ bbit1 r0,30,.Linf_nan
+#endif
+ .global __extendsfdf2
+ .balign 4
+ FUNC(__extendsfdf2)
+__extendsfdf2:
+ add.f r1,r0,r0
+ norm r3,r1
+#ifdef __LITTLE_ENDIAN__
+ lsr_s DBL0H,r1,4
+ brhs r3,7,.Linf_nan_denorm_0
+ asl_s DBL0L,r0,29
+ add_s DBL0H,DBL0H, \
+ 0x38000000
+#else
+ lsr r2,r1,4
+ brhs r3,7,.Linf_nan_denorm_0
+ asl_s DBL0L,r1,28
+ add DBL0H,r2, \
+ 0x38000000
+#endif
+ j_s.d [blink]
+ bxor.cs DBL0H,DBL0H,31
+ .balign 4
+.Linf_nan_denorm_0:
+#ifdef __LITTLE_ENDIAN__
+ mov_s DBL0H,r0
+ jeq.d [blink]
+ mov.eq DBL0L,0
+#else
+ jeq_s [blink]
+#endif
+ bmi .Linf_nan
+ asl_s r0,r0,r3
+ rsub r3,r3,0x380+6
+#ifdef __LITTLE_ENDIAN__
+ asl_s r3,r3,20
+ lsr DBL0H,r0,9
+ asl_s DBL0L,r0,23
+ add_s DBL0H,DBL0H,r3
+ j_s.d [blink]
+ bxor.cs DBL0H,DBL0H,31
+#else
+ asl DBL0L,r0,23
+ lsr_s DBL0H,r0,9
+ asl_s r3,r3,20
+ bxor.cs DBL0H,DBL0H,31
+ j_s.d [blink]
+ add_l DBL0H,DBL0H,r3
+#endif
+.Linf_nan:
+#ifdef __LITTLE_ENDIAN__
+ lsr DBL0H,r0,3
+
+ or_s DBL0H,DBL0H,r0
+ j_s.d [blink]
+ mov_l DBL0L,0
+#else
+ lsr r3,r0,3
+ mov_s DBL0L,0
+ j_s.d [blink]
+ or_l DBL0H,r0,r3
+#endif
+ ENDFUNC(__extendsfdf2)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "arc-ieee-754.h"
+
+#if 0 /* DEBUG */
+ FUNC(__fixdfsi)
+ .global __fixdfsi
+ .balign 4
+__fixdfsi:
+ push_s blink
+ push_s r0
+ bl.d __fixdfsi_c
+ push_s r1
+ mov_s r2,r0
+ pop_s r1
+ ld r0,[sp]
+ bl.d __fixdfsi_asm
+ st r2,[sp]
+ pop_s r1
+ pop_s blink
+ cmp r0,r1
+ jeq_s [blink]
+ bl abort
+ ENDFUNC(__fixdfsi)
+#define __fixdfsi __fixdfsi_asm
+#endif /* DEBUG */
+
+/* If the fraction has to be shifted left by a positive non-zero amount,
+ we have to combine bits from DBL0L and DBL0H. If we shift right,
+ or shift by zero, we only want to have the bits from DBL0H in r0. */
+
+ .global __fixdfsi
+ FUNC(__fixdfsi)
+ .balign 4
+__fixdfsi:
+ bbit0 DBL0H,30,.Lret0or1
+ asr r2,DBL0H,20
+ bmsk_s DBL0H,DBL0H,19
+ sub_s r2,r2,19; 0x3ff+20-0x400
+ neg_s r3,r2
+ asr.f 0,r3,11
+ bset_s DBL0H,DBL0H,20
+#ifdef __LITTLE_ENDIAN__
+ mov.cs DBL0L,DBL0H
+ asl DBL0H,DBL0H,r2
+#else
+ asl.cc DBL0H,DBL0H,r2
+ lsr.cs DBL0H,DBL0H,r3
+#endif
+ lsr_s DBL0L,DBL0L,r3
+
+ add.cc r0,r0,r1
+ j_s.d [blink]
+ neg.pl r0,r0
+.Lret0or1:
+ add.f r0,DBL0H,0x100000
+ lsr_s r0,r0,30
+
+ bmsk_s r0,r0,0
+ j_s.d [blink]
+ neg.mi r0,r0
+ ENDFUNC(__fixdfsi)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "arc-ieee-754.h"
+
+#if 0 /* DEBUG */
+ .global __fixsfsi
+ FUNC(__fixsfsi)
+ .balign 4
+__fixsfsi:
+ push_s blink
+ bl.d __fixsfsi_c
+ push_s r0
+ ld_s r1,[sp]
+ st_s r0,[sp]
+ bl.d __fixsfsi_asm
+ mov_s r0,r1
+ pop_s r1
+ pop_s blink
+ cmp r0,r1
+ jeq_s [blink]
+ bl abort
+ ENDFUNC(__fixsfsi)
+#define __fixsfsi __fixsfsi_asm
+#endif /* DEBUG */
+
+ .global __fixsfsi
+ FUNC(__fixsfsi)
+ .balign 4
+__fixsfsi:
+ bbit0 r0,30,.Lret0or1
+ lsr r2,r0,23
+ bmsk_s r0,r0,22
+ bset_s r0,r0,23
+ sub_s r2,r2,22;0x7f+23-0x80
+ asl.f 0,r2,24
+ neg r3,r2
+ asl.mi r0,r0,r2
+ lsr.pl r0,r0,r3
+ j_s.d [blink]
+ neg.cs r0,r0
+.Lret0or1:
+ add.f r0,r0,0x800000
+ lsr_s r0,r0,30
+
+ bmsk_s r0,r0,0
+ j_s.d [blink]
+ neg.mi r0,r0
+ ENDFUNC(__fixsfsi)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "arc-ieee-754.h"
+
+#if 0 /* DEBUG */
+ FUNC(__fixunsdfsi)
+ .global __fixunsdfsi
+ .balign 4
+__fixunsdfsi:
+ push_s blink
+ push_s r0
+ bl.d __fixunsdfsi_c
+ push_s r1
+ mov_s r2,r0
+ pop_s r1
+ ld r0,[sp]
+ bl.d __fixunsdfsi_asm
+ st r2,[sp]
+ pop_s r1
+ pop_s blink
+ cmp r0,r1
+ jeq_s [blink]
+ bl abort
+ ENDFUNC(__fixunsdfsi)
+#define __fixunsdfsi __fixunsdfsi_asm
+#endif /* DEBUG */
+
+ .global __fixunsdfsi
+ FUNC(__fixunsdfsi)
+ .balign 4
+__fixunsdfsi:
+ bbit0 DBL0H,30,.Lret0or1
+ lsr r2,DBL0H,20
+ bmsk_s DBL0H,DBL0H,19
+ sub_s r2,r2,19; 0x3ff+20-0x400
+ neg_s r3,r2
+ btst_s r3,10
+ bset_s DBL0H,DBL0H,20
+#ifdef __LITTLE_ENDIAN__
+ mov.ne DBL0L,DBL0H
+ asl DBL0H,DBL0H,r2
+#else
+ asl.eq DBL0H,DBL0H,r2
+ lsr.ne DBL0H,DBL0H,r3
+#endif
+ lsr DBL0L,DBL0L,r3
+ j_s.d [blink]
+ add.eq r0,r0,r1
+.Lret0:
+ j_s.d [blink]
+ mov_l r0,0
+.Lret0or1:
+ add_s DBL0H,DBL0H,0x100000
+ lsr_s DBL0H,DBL0H,30
+ j_s.d [blink]
+ bmsk_l r0,DBL0H,0
+ ENDFUNC(__fixunsdfsi)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "arc-ieee-754.h"
+
+#if 0 /* DEBUG */
+ .global __floatsidf
+ .balign 4
+ FUNC(__floatsidf)
+__floatsidf:
+ push_s blink
+ bl.d __floatsidf_c
+ push_s r0
+ ld_s r2,[sp]
+ st_s r1,[sp]
+ push_s r0
+ bl.d __floatsidf_asm
+ mov_s r0,r2
+ pop_s r2
+ pop_s r3
+ pop_s blink
+ cmp r0,r2
+ cmp.eq r1,r3
+ jeq_s [blink]
+ bl abort
+ ENDFUNC(__floatsidf)
+#define __floatsidf __floatsidf_asm
+#endif /* DEBUG */
+
+ .global __floatsidf
+ .balign 4
+ FUNC(__floatsidf)
+__floatsidf:
+ abs.f r1,r0
+ jeq_s [blink]
+ lsr r2,r1
+ mov r12,-0x41d ; -(0x3ff+31-1)
+ norm r2,r2
+ bclr.cs r12,r12,11
+ rsub.f r3,r2,11
+ add_s r12,r2,r12
+ add_s r2,r2,21
+#ifdef __LITTLE_ENDIAN__
+ asl DBL0L,r1,r2
+ lsr_s DBL0H,r1,r3
+#else
+ lsr DBL0H,r1,r3
+ asl_s DBL0L,r1,r2
+#endif
+ asl_s r12,r12,20
+ mov.lo DBL0H,DBL0L
+ sub_s DBL0H,DBL0H,r12
+ j_s.d [blink]
+ mov.ls DBL0L,0
+ ENDFUNC(__floatsidf)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "arc-ieee-754.h"
+
+#if 0 /* DEBUG */
+ .global __floatsisf
+ FUNC(__floatsisf)
+ .balign 4
+__floatsisf:
+ push_s blink
+ bl.d __floatsisf_c
+ push_s r0
+ ld_s r1,[sp]
+ st_s r0,[sp]
+ bl.d __floatsisf_asm
+ mov_s r0,r1
+ pop_s r1
+ pop_s blink
+ cmp r0,r1
+ jeq_s [blink]
+ bl abort
+ ENDFUNC(__floatsisf)
+ .global __floatunsisf
+ FUNC(__floatunsisf)
+ .balign 4
+__floatunsisf:
+ push_s blink
+ bl.d __floatunsisf_c
+ push_s r0
+ ld_s r1,[sp]
+ st_s r0,[sp]
+ bl.d __floatunsisf_asm
+ mov_s r0,r1
+ pop_s r1
+ pop_s blink
+ cmp r0,r1
+ jeq_s [blink]
+ bl abort
+ ENDFUNC(__floatunsisf)
+#define __floatsisf __floatsisf_asm
+#define __floatunsisf __floatunsisf_asm
+#endif /* DEBUG */
+
+ .global __floatunsisf
+ .global __floatsisf
+ FUNC(__floatsisf)
+ FUNC(__floatunsisf)
+ .balign 4
+__floatunsisf:
+ lsr_s r2,r0
+ mov_l r12,0x9d ; 0x7f + 31 - 1
+ norm r2,r2
+ brne_l r0,0,0f
+ j_s [blink]
+ .balign 4
+__floatsisf:
+ abs.f r0,r0
+ jeq_s [blink]
+ lsr_s r2,r0
+ mov_s r12,0x9d ; 0x7f + 31 - 1
+ norm r2,r2
+ bset.cs r12,r12,8
+0: rsub.f r3,r2,8
+ bmsk r1,r0,r3
+ ror r1,r1,r3
+ lsr.pl r0,r0,r3
+ neg_s r3,r3
+ asl.mi r0,r0,r3
+ sub_s r12,r12,r2
+ asl_s r12,r12,23
+ bxor.pl.f r1,r1,31
+ add_s r0,r0,r12
+ j_s.d [blink]
+ add.pnz r0,r0,1
+ ENDFUNC(__floatunsisf)
+ ENDFUNC(__floatsisf)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "arc-ieee-754.h"
+
+#if 0 /* DEBUG */
+ .global __floatunsidf
+ .balign 4
+ FUNC(__floatunsidf)
+__floatunsidf:
+ push_s blink
+ bl.d __floatunsidf_c
+ push_s r0
+ ld_s r2,[sp]
+ st_s r1,[sp]
+ push_s r0
+ bl.d __floatunsidf_asm
+ mov_s r0,r2
+ pop_s r2
+ pop_s r3
+ pop_s blink
+ cmp r0,r2
+ cmp.eq r1,r3
+ jeq_s [blink]
+ bl abort
+ ENDFUNC(__floatunsidf)
+#define __floatunsidf __floatunsidf_asm
+#endif /* DEBUG */
+
+ .global __floatunsidf
+ .balign 4
+ FUNC(__floatunsidf)
+__floatunsidf:
+ lsr_s r1,r0
+ breq_s r0,0,.Lret0
+ norm r2,r1
+ mov r12,-0x41d ; -(0x3ff+31-1)
+ rsub.f r3,r2,11
+ add_s r12,r2,r12
+ add_s r2,r2,21
+#ifdef __LITTLE_ENDIAN__
+ lsr DBL0H,r0,r3
+ asl_s DBL0L,r0,r2
+#else
+ asl DBL0L,r0,r2
+ lsr_s DBL0H,r0,r3
+#endif
+ asl_s r12,r12,20
+ mov.lo DBL0H,DBL0L
+ sub_s DBL0H,DBL0H,r12
+.Lret0: j_s.d [blink]
+ mov.ls DBL0L,0
+ ENDFUNC(__floatunsidf)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "arc-ieee-754.h"
+/* inputs: DBL0, DBL1
+ output: c flags to be used for 'hs' condition
+ clobber: r12, flags */
+/* For NaNs, bit 19.. bit 30 of the high word must be set. */
+#if 0 /* DEBUG */
+ .global __gedf2
+ .balign 4
+ FUNC(__gedf2)
+__gedf2:
+ st.a r11,[sp,-4]` push_s blink` st.a r10,[sp,-4]` st.a r9,[sp,-4]
+ st.a r8,[sp,-4]` st.a r7,[sp,-4]` st.a r6,[sp,-4]` st.a r5,[sp,-4]
+ st.a r4,[sp,-4]` push_s r3` push_s r2` push_s r1`
+ bl.d __gedf2_c` push_s r0
+ mov r11,r0` pop_s r0` pop_s r1` pop_s r2` pop_s r3
+ ld.ab r4,[sp,4]` ld.ab r5,[sp,4]` ld.ab r6,[sp,4]`
+ ld.ab r7,[sp,4]` ld.ab r8,[sp,4]` ld.ab r9,[sp,4]
+ bl.d __gedf2_asm` ld.ab r10,[sp,4]
+ pop_s blink
+ brge.d r11,0,0f
+ ld.ab r11,[sp,4]
+ jlo [blink]
+ bl abort
+0: jhs [blink]
+ bl abort
+ ENDFUNC(__gedf2)
+#define __gedf2 __gedf2_asm
+#endif /* DEBUG */
+ .global __gedf2
+ .balign 4
+ HIDDEN_FUNC(__gedf2)
+__gedf2:
+ or.f r12,DBL0H,DBL1H
+ bmi.d .Lneg
+ bmsk_s r12,r12,20
+ add1.f 0,r12,DBL0H ; clear z; set c iff NaN
+ add1.cc.f r12,r12,DBL1H ; clear z; set c iff NaN
+ bbit1 DBL0H,31,.Lneg
+ cmp.cc DBL0H,DBL1H
+ j_s.d [blink]
+ cmp.eq DBL0L,DBL1L
+ .balign 4
+.Lneg: breq.d DBL1H,0,.L0
+ add1.f 0,r12,DBL0H
+ add1.cc.f r12,r12,DBL1H
+ cmp.cc DBL1H,DBL0H
+ j_s.d [blink]
+ cmp.eq DBL1L,DBL0L
+ .balign 4
+.L0:
+ bxor.f 0,DBL0H,31 ; check for high word of -0.
+ beq_s .Lcheck_0
+ cmp.cc DBL1H,DBL0H
+ j_s.d [blink]
+ cmp.eq DBL1L,DBL0L
+.Lcheck_0:
+ ; high words suggest DBL0 may be -0, DBL1 +0; check low words.
+ cmp_s DBL1H,DBL0L
+ j_s.d [blink]
+ cmp.cc DBL1H,DBL1L
+ ENDFUNC(__gedf2)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "arc-ieee-754.h"
+/* inputs: r0, r1
+ output: c flag to be used for 'hs' condition
+ clobber: r12,flags */
+/* For NaNs, bit 22.. bit 30 must be set. */
+#if 0 /* DEBUG */
+ .global __gesf2
+ .balign 4
+ FUNC(__gesf2)
+__gesf2:
+ st.a r11,[sp,-4]` push_s blink` st.a r10,[sp,-4]` st.a r9,[sp,-4]
+ st.a r8,[sp,-4]` st.a r7,[sp,-4]` st.a r6,[sp,-4]` st.a r5,[sp,-4]
+ st.a r4,[sp,-4]` push_s r3` push_s r2` push_s r1`
+ bl.d __gesf2_c` push_s r0
+ mov r11,r0` pop_s r0` pop_s r1` pop_s r2` pop_s r3
+ ld.ab r4,[sp,4]` ld.ab r5,[sp,4]` ld.ab r6,[sp,4]`
+ ld.ab r7,[sp,4]` ld.ab r8,[sp,4]` ld.ab r9,[sp,4]
+ bl.d __gesf2_asm` ld.ab r10,[sp,4]
+ pop_s blink
+ brge.d r11,0,0f
+ ld.ab r11,[sp,4]
+ jlo [blink]
+ bl abort
+0: jhs [blink]
+ bl abort
+ ENDFUNC(__gesf2)
+#define __gesf2 __gesf2_asm
+#endif /* DEBUG */
+ .global __gesf2
+ .balign 4
+ HIDDEN_FUNC(__gesf2)
+__gesf2:
+ or.f r12,r0,r1
+ bmi.d .Lneg
+ bmsk_s r12,r12,23
+ add1.f 0,r12,r0 ; check for NaN
+ add1.cc.f r12,r12,r1
+ j_s.d [blink]
+ cmp.cc r0,r1
+ .balign 4
+.Lneg: breq.d r1,0,.L0
+ add1.f 0,r12,r0 ; check for NaN
+ add1.cc.f r12,r12,r1
+ j_s.d [blink]
+ cmp.cc r1,r0
+ .balign 4
+.L0: bxor.f 0,r0,31 ; check for -0
+ j_s.d [blink]
+ cmp.hi r1,r0
+ ENDFUNC(__gesf2)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "arc-ieee-754.h"
+/* inputs: DBL0, DBL1
+ output: c,z flags to be used for 'hi' condition
+ clobber: r12, flags */
+/* For NaNs, bit 19.. bit 30 of the high word must be set. */
+#if 0 /* DEBUG */
+ .global __gtdf2
+ .balign 4
+ FUNC(__gtdf2)
+__gtdf2:
+ st.a r11,[sp,-4]` push_s blink` st.a r10,[sp,-4]` st.a r9,[sp,-4]
+ st.a r8,[sp,-4]` st.a r7,[sp,-4]` st.a r6,[sp,-4]` st.a r5,[sp,-4]
+ st.a r4,[sp,-4]` push_s r3` push_s r2` push_s r1`
+ bl.d __gtdf2_c` push_s r0
+ mov r11,r0` pop_s r0` pop_s r1` pop_s r2` pop_s r3
+ ld.ab r4,[sp,4]` ld.ab r5,[sp,4]` ld.ab r6,[sp,4]`
+ ld.ab r7,[sp,4]` ld.ab r8,[sp,4]` ld.ab r9,[sp,4]
+ bl.d __gtdf2_asm` ld.ab r10,[sp,4]
+ pop_s blink
+ brgt.d r11,0,0f
+ ld.ab r11,[sp,4]
+ jls [blink]
+ bl abort
+0: jhi [blink]
+ bl abort
+ ENDFUNC(__gtdf2)
+#define __gtdf2 __gtdf2_asm
+#endif /* DEBUG */
+ .global __gtdf2
+ .balign 4
+ HIDDEN_FUNC(__gtdf2)
+__gtdf2:
+ or.f r12,DBL0H,DBL1H
+ bmi.d .Lneg
+ bmsk_s r12,r12,20
+ add1.f 0,r12,DBL0H ; clear z; set c iff NaN
+ add1.cc.f r12,r12,DBL1H ; clear z; set c iff NaN
+ ; don't care: z may or may not be cleared if there is no NaN event
+ cmp.cc DBL0H,DBL1H
+ j_s.d [blink]
+ cmp.eq DBL0L,DBL1L
+ .balign 4
+.Lneg: breq.d DBL0H,0,.L0
+ add1.f 0,r12,DBL1H
+ add1.cc.f r12,r12,DBL0H
+ cmp.cc DBL1H,DBL0H
+ j_s.d [blink]
+ cmp.eq DBL1L,DBL0L
+ .balign 4
+.L0:
+ bxor.f 0,DBL1H,31
+ beq_s .Lcheck_0
+ cmp.cc DBL1H,DBL0H
+ j_s.d [blink]
+ cmp.eq DBL1L,DBL0L
+ .balign 4
+.Lcheck_0:
+ ; high words suggest DBL0 may be +0, DBL1 -0; check low words.
+ j_s.d [blink]
+ or.f 0,DBL0L,DBL1L
+ ENDFUNC(__gtdf2)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "arc-ieee-754.h"
+/* inputs: r0, r1
+ output: c, z flags to be used for 'hi' condition
+ clobber: r12,flags */
+/* For NaNs, bit 22.. bit 30 must be set. */
+#if 0 /* DEBUG */
+ .global __gtsf2
+ .balign 4
+ FUNC(__gtsf2)
+__gtsf2:
+ st.a r11,[sp,-4]` push_s blink` st.a r10,[sp,-4]` st.a r9,[sp,-4]
+ st.a r8,[sp,-4]` st.a r7,[sp,-4]` st.a r6,[sp,-4]` st.a r5,[sp,-4]
+ st.a r4,[sp,-4]` push_s r3` push_s r2` push_s r1`
+ bl.d __gtsf2_c` push_s r0
+ mov r11,r0` pop_s r0` pop_s r1` pop_s r2` pop_s r3
+ ld.ab r4,[sp,4]` ld.ab r5,[sp,4]` ld.ab r6,[sp,4]`
+ ld.ab r7,[sp,4]` ld.ab r8,[sp,4]` ld.ab r9,[sp,4]
+ bl.d __gtsf2_asm` ld.ab r10,[sp,4]
+ pop_s blink
+ brgt.d r11,0,0f
+ ld.ab r11,[sp,4]
+ jls [blink]
+ bl abort
+0: jhi [blink]
+ bl abort
+ ENDFUNC(__gtsf2)
+#define __gtsf2 __gtsf2_asm
+#endif /* DEBUG */
+ .global __gtsf2
+ .balign 4
+ HIDDEN_FUNC(__gtsf2)
+__gtsf2:
+ or.f r12,r0,r1
+ bmi.d .Lneg
+ bmsk_s r12,r12,23
+ add1.f 0,r12,r0 ; check for NaN
+ add1.cc.f r12,r12,r1
+ j_s.d [blink]
+ cmp.cc r0,r1
+ .balign 4
+.Lneg: breq.d r0,0,.L0
+ add1.f 0,r12,r0 ; check for NaN
+ add1.cc.f r12,r12,r1
+ j_s.d [blink]
+ cmp.cc r1,r0
+ .balign 4
+.L0: bxor.f 0,r1,31 ; check for -0
+ j_s.d [blink]
+ cmp.hi r1,r0
+ ENDFUNC(__gtsf2)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+/* XMAC schedule: directly back-to-back multiplies stall; the third
+ instruction after a multiply stalls unless it is also a multiply. */
+#include "arc-ieee-754.h"
+
+#if 0 /* DEBUG */
+ .global __muldf3
+ .balign 4
+__muldf3:
+ push_s blink
+ push_s r2
+ push_s r3
+ push_s r0
+ bl.d __muldf3_c
+ push_s r1
+ ld_s r2,[sp,12]
+ ld_s r3,[sp,8]
+ st_s r0,[sp,12]
+ st_s r1,[sp,8]
+ pop_s r1
+ bl.d __muldf3_asm
+ pop_s r0
+ pop_s r3
+ pop_s r2
+ pop_s blink
+ cmp r0,r2
+ cmp.eq r1,r3
+ jeq_s [blink]
+ b abort
+#define __muldf3 __muldf3_asm
+#endif /* DEBUG */
+/* N.B. This is optimized for ARC700.
+ ARC600 has very different scheduling / instruction selection criteria. */
+/* For the standard multiplier, instead of mpyu rx,DBL0L,DBL1L; tst rx,rx ,
+ we can do:
+ sub rx,DBL0L,1; bic rx,DBL0L,rx; lsr rx,rx; norm rx,rx; asl.f 0,DBL1L,rx */
+
+__muldf3_support: /* This label makes debugger output saner. */
+/* If one number is denormal, subtract some from the exponent of the other
+ one (if the other exponent is too small, return 0), and normalize the
+ denormal. Then re-run the computation. */
+ .balign 4
+ FUNC(__muldf3)
+.Ldenorm_dbl0:
+ mov_s r12,DBL0L
+ mov_s DBL0L,DBL1L
+ mov_s DBL1L,r12
+ mov_s r12,DBL0H
+ mov_s DBL0H,DBL1H
+ mov_s DBL1H,r12
+ and r11,DBL0H,r9
+.Ldenorm_dbl1:
+ brhs r11,r9,.Linf_nan
+ brhs 0x3ca00001,r11,.Lret0
+ sub_s DBL0H,DBL0H,DBL1H
+ bmsk_s DBL1H,DBL1H,30
+ add_s DBL0H,DBL0H,DBL1H
+ breq_s DBL1H,0,.Ldenorm_2
+ norm r12,DBL1H
+
+ sub_s r12,r12,10
+ asl r5,r12,20
+ asl_s DBL1H,DBL1H,r12
+ sub DBL0H,DBL0H,r5
+ neg r5,r12
+ lsr r6,DBL1L,r5
+ asl_s DBL1L,DBL1L,r12
+ b.d __muldf3
+ add_s DBL1H,DBL1H,r6
+
+ .balign 4
+.Linf_nan:
+ bclr r12,DBL1H,31
+ xor_s DBL1H,DBL1H,DBL0H
+ bclr_s DBL0H,DBL0H,31
+ max r8,DBL0H,r12 ; either NaN -> NaN ; otherwise inf
+ or.f 0,DBL0H,DBL0L
+ mov_s DBL0L,0
+ or.ne.f DBL1L,DBL1L,r12
+ not_s DBL0H,DBL0L ; inf * 0 -> NaN
+ mov.ne DBL0H,r8
+ tst_s DBL1H,DBL1H
+ j_s.d [blink]
+ bset.mi DBL0H,DBL0H,31
+
+.Lret0: xor_s DBL0H,DBL0H,DBL1H
+ bclr DBL1H,DBL0H,31
+ xor_s DBL0H,DBL0H,DBL1H
+ j_s.d [blink]
+ mov_l DBL0L,0
+
+ .balign 4
+.Ldenorm_2:
+ breq_s DBL1L,0,.Lret0 ; 0 input -> 0 output
+ norm.f r12,DBL1L
+
+ mov.mi r12,21
+ add.pl r12,r12,22
+ neg r11,r12
+ asl_s r12,r12,20
+ lsr.f DBL1H,DBL1L,r11
+ ror DBL1L,DBL1L,r11
+ sub_s DBL0H,DBL0H,r12
+ mov.eq DBL1H,DBL1L
+ sub_s DBL1L,DBL1L,DBL1H
+ /* Fall through. */
+ .global __muldf3
+ .balign 4
+__muldf3:
+ ld.as r9,[pcl,0x4b] ; ((.L7ff00000-.+2)/4)]
+ mpyhu r4,DBL0L,DBL1L
+ bmsk r6,DBL0H,19
+ bset r6,r6,20
+ mpyu r7,r6,DBL1L
+ and r11,DBL0H,r9
+ breq r11,0,.Ldenorm_dbl0
+ mpyhu r8,r6,DBL1L
+ bmsk r10,DBL1H,19
+ bset r10,r10,20
+ mpyhu r5,r10,DBL0L
+ add.f r4,r4,r7
+ and r12,DBL1H,r9
+ mpyhu r7,r6,r10
+ breq r12,0,.Ldenorm_dbl1
+ adc.f r5,r5,r8
+ mpyu r8,r10,DBL0L
+ breq r11,r9,.Linf_nan
+ breq r12,r9,.Linf_nan
+ mpyu r6,r6,r10
+ add.cs r7,r7,1
+ add.f r4,r4,r8
+ mpyu r10,DBL1L,DBL0L
+ bclr r8,r9,30 ; 0x3ff00000
+ adc.f r5,r5,r6
+ ; XMAC write-back stall / std. mult stall is one cycle later
+ bclr r6,r9,20 ; 0x7fe00000
+ add.cs r7,r7,1 ; fraction product in r7:r5:r4
+ tst r10,r10
+ bset.ne r4,r4,0 ; put least significant word into sticky bit
+ lsr.f r10,r7,9
+ add_l r12,r12,r11 ; add exponents
+ rsub.eq r8,r8,r9 ; 0x40000000
+ sub r12,r12,r8 ; subtract bias + implicit 1
+ brhs.d r12,r6,.Linf_denorm
+ rsub r10,r10,12
+.Lshift_frac:
+ neg r8,r10
+ asl r6,r4,r10
+ lsr DBL0L,r4,r8
+ add.f 0,r6,r6
+ btst.eq DBL0L,0
+ cmp.eq r4,r4 ; round to nearest / round to even
+ asl r4,r5,r10
+ lsr r5,r5,r8
+ adc.f DBL0L,DBL0L,r4
+ xor.f 0,DBL0H,DBL1H
+ asl r7,r7,r10
+ add_s r12,r12,r5
+ adc DBL0H,r12,r7
+ j_s.d [blink]
+ bset.mi DBL0H,DBL0H,31
+
+/* We have checked for infinitey / NaN input before, and transformed
+ denormalized inputs into normalized inputs. Thus, the worst case
+ exponent overflows are:
+ 1 + 1 - 0x400 == 0xc02 : maximum underflow
+ 0x7fe + 0x7fe - 0x3ff == 0xbfd ; maximum overflow
+ N.B. 0x7e and 0x7f are also values for overflow.
+
+ If (r12 <= -54), we have an underflow to zero. */
+ .balign 4
+.Linf_denorm:
+ brlo r12,0xc0000000,.Linf
+ asr r6,r12,20
+ mov_s r12,0
+ add.f r10,r10,r6
+ brgt r10,0,.Lshift_frac
+ beq_s .Lround_frac
+ add.f r10,r10,32
+.Lshift32_frac:
+ tst r4,r4
+ mov r4,r5
+ bset.ne r4,r4,1
+ mov r5,r7
+ mov r7,0
+ brge r10,1,.Lshift_frac
+ breq r10,0,.Lround_frac
+ add.f r10,r10,32
+ brgt r10,21,.Lshift32_frac
+ b_s .Lret0
+
+.Lround_frac:
+ add.f 0,r4,r4
+ btst.eq r5,0
+ mov_s DBL0L,r5
+ mov_s DBL0H,r7
+ adc.eq.f DBL0L,DBL0L,0
+ j_s.d [blink]
+
+ adc.eq DBL0H,DBL0H,0
+
+.Linf: xor.f DBL1H,DBL1H,DBL0H
+ mov_s DBL0L,0
+ mov_s DBL0H,r9
+ j_s.d [blink]
+ bset.mi DBL0H,DBL0H,31
+ ENDFUNC(__muldf3)
+
+ .balign 4
+.L7ff00000:
+ .long 0x7ff00000
--- /dev/null
+/* Copyright (C) 2008-2013 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+/* XMAC schedule: directly back-to-back multiplies stall; the third
+ instruction after a multiply stalls unless it is also a multiply. */
+#include "arc-ieee-754.h"
+
+#if 0 /* DEBUG */
+ .global __mulsf3
+ FUNC(__mulsf3)
+ .balign 4
+__mulsf3:
+ push_s blink
+ push_s r1
+ bl.d __mulsf3_c
+ push_s r0
+ ld_s r1,[sp,4]
+ st_s r0,[sp,4]
+ bl.d __mulsf3_asm
+ pop_s r0
+ pop_s r1
+ pop_s blink
+ cmp r0,r1
+ jeq_s [blink]
+ and r12,r0,r1
+ bic.f 0,0x7f800000,r12
+ bne 0f
+ bmsk.f 0,r0,22
+ bmsk.ne.f r1,r1,22
+ jne_s [blink] ; both NaN -> OK
+0: bl abort
+ ENDFUNC(__mulsf3)
+#define __mulsf3 __mulsf3_asm
+#endif /* DEBUG */
+
+ .balign 4
+ .global __mulsf3
+ FUNC(__mulsf3)
+__mulsf3:
+ ld.as r9,[pcl,79]; [pcl,((.L7f800000-.+2)/4)]
+ bmsk r4,r1,22
+ bset r2,r0,23
+ asl_s r2,r2,8
+ bset r3,r4,23
+ mpyhu r6,r2,r3
+ and r11,r0,r9
+ breq r11,0,.Ldenorm_dbl0
+ mpyu r7,r2,r3
+ breq r11,r9,.Linf_nan_dbl0
+ and r12,r1,r9
+ asl.f 0,r6,8
+ breq r12,0,.Ldenorm_dbl1
+.Lpast_denorm:
+ xor_s r0,r0,r1
+.Lpast_denorm_dbl1:
+ add.pl r6,r6,r6
+ bclr.pl r6,r6,23
+ add.pl.f r7,r7,r7
+ ld.as r4,[pcl,64]; [pcl,((.L7fffffff-.+2)/4)]
+ add.cs r6,r6,1
+ lsr.f 0,r6,1
+ breq r12,r9,.Linf_nan_dbl1
+ add_s r12,r12,r11
+ adc.f 0,r7,r4
+ add_s r12,r12, \
+ -0x3f800000
+ adc.f r8,r6,r12
+ bic r0,r0,r4
+ tst.pl r8,r9
+ min r3,r8,r9
+ jpnz.d [blink]
+ add.pnz r0,r0,r3
+; infinity or denormal number
+ add.ne.f r3,r3,r3
+ bpnz .Linfinity
+ asr_s r3,r3,23+1
+ bset r6,r6,23
+ sub_s r3,r3,1
+ neg_s r2,r3
+ brhi r2,24,.Lret_r0 ; right shift shift > 24 -> return +-0
+ lsr r2,r6,r2
+ asl r9,r6,r3
+ lsr.f 0,r2,1
+ tst r7,r7
+ add_s r0,r0,r2
+ bset.ne r9,r9,0
+ adc.f 0,r9,r4
+ j_s.d [blink]
+ add.cs r0,r0,1
+.Linfinity:
+ j_s.d [blink]
+ add_s r0,r0,r9
+
+.Lret_r0: j_s [blink]
+
+ .balign 4
+.Linf_nan_dbl0:
+ sub_s r2,r1,1 ; inf/nan * 0 -> nan; inf * nan -> nan (use |r2| >= inf)
+ bic.f 0,r9,r2
+ xor_s r0,r0,r1
+ bclr_s r1,r1,31
+ xor_s r0,r0,r1
+ jne_s [blink]
+.Lretnan:
+ j_s.d [blink]
+ mov r0,-1
+.Ldenorm_dbl0_inf_nan_dbl1:
+ bmsk.f 0,r0,30
+ beq_s .Lretnan
+ xor_s r0,r0,r1
+.Linf_nan_dbl1:
+ xor_s r1,r1,r0
+ bclr_s r1,r1,31
+ j_s.d [blink]
+ xor_s r0,r0,r1
+
+ .balign 4
+.Ldenorm_dbl0:
+ bclr_s r2,r2,31
+ norm.f r4,r2
+ and r12,r1,r9
+ add_s r2,r2,r2
+ asl r2,r2,r4
+ asl r4,r4,23
+ mpyhu r6,r2,r3
+ breq r12,r9,.Ldenorm_dbl0_inf_nan_dbl1
+ sub.ne.f r12,r12,r4
+ mpyu r7,r2,r3
+ bhi.d .Lpast_denorm
+ asl.f 0,r6,8
+ xor_s r0,r0,r1
+ bmsk r1,r0,30
+ j_s.d [blink]
+ bic_l r0,r0,r1
+
+ .balign 4
+.Ldenorm_dbl1:
+ norm.f r3,r4
+ xor_s r0,r0,r1
+ sub_s r3,r3,7
+ asl r4,r4,r3
+ sub_s r3,r3,1
+ asl_s r3,r3,23
+ mpyhu r6,r2,r4
+ sub.ne.f r11,r11,r3
+ bmsk r8,r0,30
+ mpyu r7,r2,r4
+ bhi.d .Lpast_denorm_dbl1
+ asl.f 0,r6,8
+ j_s.d [blink]
+ bic r0,r0,r8
+
+ .balign 4
+.L7f800000:
+ .long 0x7f800000
+.L7fffffff:
+ .long 0x7fffffff
+ ENDFUNC(__mulsf3)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "arc-ieee-754.h"
+/* inputs: r0, r1
+ output: c flag
+ clobber: r12, flags
+ For NaNs, bit 19 .. bit 30 must be set. */
+#if 0 /* DEBUG */
+ .global __orddf2
+ .balign 4
+ FUNC(__orddf2)
+__orddf2:
+ st.a r11,[sp,-4]` push_s blink` st.a r10,[sp,-4]` st.a r9,[sp,-4]
+ st.a r8,[sp,-4]` st.a r7,[sp,-4]` st.a r6,[sp,-4]` st.a r5,[sp,-4]
+ st.a r4,[sp,-4]` push_s r3` push_s r2` push_s r1`
+ bl.d __unorddf2_c` push_s r0
+ mov r11,r0` pop_s r0` pop_s r1` pop_s r2` pop_s r3
+ ld.ab r4,[sp,4]` ld.ab r5,[sp,4]` ld.ab r6,[sp,4]`
+ ld.ab r7,[sp,4]` ld.ab r8,[sp,4]` ld.ab r9,[sp,4]
+ bl.d __orddf2_asm` ld.ab r10,[sp,4]
+ pop_s blink
+ brne.d r11,0,0f
+ ld.ab r11,[sp,4]
+ jcc [blink]
+ bl abort
+0: jcs [blink]
+ bl abort
+ ENDFUNC(__orddf2)
+#define __orddf2 __orddf2_asm
+#endif /* DEBUG */
+ .global __orddf2
+ .balign 4
+ HIDDEN_FUNC(__orddf2)
+__orddf2:
+ bmsk r12,DBL0H,20
+ add1.f r12,r12,DBL0H /* clear z; set c if NaN. */
+ bmsk r12,DBL1H,20
+ j_s.d [blink]
+ add1.cc.f r12,r12,DBL1H /* clear z; set c if NaN. */
+ ENDFUNC(__orddf2)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "arc-ieee-754.h"
+/* inputs: r0, r1
+ output: c flag
+ clobber: r12, flags
+ For NaNs, bit 22 .. bit 30 must be set. */
+#if 0 /* DEBUG */
+ .global __ordsf2
+ .balign 4
+ FUNC(__ordsf2)
+__ordsf2:
+ st.a r11,[sp,-4]` push_s blink` st.a r10,[sp,-4]` st.a r9,[sp,-4]
+ st.a r8,[sp,-4]` st.a r7,[sp,-4]` st.a r6,[sp,-4]` st.a r5,[sp,-4]
+ st.a r4,[sp,-4]` push_s r3` push_s r2` push_s r1`
+ bl.d __unordsf2_c` push_s r0
+ mov r11,r0` pop_s r0` pop_s r1` pop_s r2` pop_s r3
+ ld.ab r4,[sp,4]` ld.ab r5,[sp,4]` ld.ab r6,[sp,4]`
+ ld.ab r7,[sp,4]` ld.ab r8,[sp,4]` ld.ab r9,[sp,4]
+ bl.d __ordsf2_asm` ld.ab r10,[sp,4]
+ pop_s blink
+ brne.d r11,0,0f
+ ld.ab r11,[sp,4]
+ jcc [blink]
+ bl abort
+0: jcs [blink]
+ bl abort
+ ENDFUNC(__ordsf2)
+#define __ordsf2 __ordsf2_asm
+#endif /* DEBUG */
+ .global __ordsf2
+ .balign 4
+ HIDDEN_FUNC(__ordsf2)
+__ordsf2:
+ bmsk r12,r0,23
+ add1.f r12,r12,r0 /* clear z; set c if NaN. */
+ bmsk r12,r1,23
+ j_s.d [blink]
+ add1.cc.f r12,r12,r1 /* clear z; set c if NaN. */
+ ENDFUNC(__ordsf2)
--- /dev/null
+/* Copyright (C) 2006, 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "arc-ieee-754.h"
+
+#if 0 /* DEBUG */
+ FUNC(__truncdfsf2)
+ .global __truncdfsf2
+ .balign 4
+__truncdfsf2:
+ push_s blink
+ push_s r0
+ bl.d __truncdfsf2_c
+ push_s r1
+ mov_s r2,r0
+ pop_s r1
+ ld r0,[sp]
+ bl.d __truncdfsf2_asm
+ st r2,[sp]
+ pop_s r1
+ pop_s blink
+ cmp r0,r1
+ jeq_s [blink]
+ and r12,r0,r1
+ bic.f 0,0x7f800000,r12
+ bne 0f
+ bmsk.f 0,r0,22
+ bmsk.ne.f r1,r1,22
+ jne_s [blink] ; both NaN -> OK
+0: bl abort
+ ENDFUNC(__truncdfsf2)
+#define __truncdfsf2 __truncdfsf2_asm
+#endif /* DEBUG */
+
+ .global __truncdfsf2
+ .balign 4
+ FUNC(__truncdfsf2)
+__truncdfsf2:
+ lsr r2,DBL0H,20
+ asl_s DBL0H,DBL0H,12
+ sub r12,r2,0x380
+ bclr.f r3,r12,11
+ brhs r3,0xff,.Lill_exp
+ beq_l .Ldenorm0
+ asl_s r12,r12,23
+ tst DBL0L, \
+ 0x2fffffff /* Check if msb guard bit wants rounding up. */
+ lsr_s DBL0L,DBL0L,28
+ lsr_s DBL0H,DBL0H,8
+ add.ne DBL0L,DBL0L,1
+ add_s DBL0H,DBL0H,DBL0L
+ lsr_s DBL0H,DBL0H
+ btst_s r2,11
+ add_s r0,DBL0H,r12
+ j_s.d [blink]
+ bxor.ne r0,r0,31
+ .balign 4
+.Lill_exp:
+ bbit1 r2,10,.Linf_nan
+ bmsk_s r12,r12,9
+ rsub.f r12,r12,8+0x400-32 ; Go from 9 to 1 guard bit in MSW. */
+ bhs_s .Lzero
+ lsr r3,DBL0L,21
+ rrc DBL0H,DBL0H ; insert leading 1
+ asl.f 0,DBL0L,8 ; check lower 24 guard bits
+ add_s r3,DBL0H,r3
+ add.pnz r3,r3,1 ; assemble fraction with compressed guard bits.
+ lsr r0,r3,r12
+ neg_s r12,r12
+ btst_s r0,1
+ asl.eq.f r3,r3,r12
+ add.ne r0,r0,1
+ btst_s r2,11
+ lsr_s r0,r0
+ j_s.d [blink]
+ bxor.ne r0,r0,31
+.Lzero:
+ lsr_s r2,r2,11
+ j_s.d [blink]
+ asl r0,r2,31
+.Ldenorm0:
+ asl_s r12,r12,20
+ tst DBL0L, \
+ 0x5fffffff /* Check if msb guard bit wants rounding up. */
+ lsr_s DBL0L,DBL0L,29
+ lsr_s DBL0H,DBL0H,9
+ add.ne DBL0L,DBL0L,1
+ bset_s DBL0H,DBL0H,23
+ add_s DBL0H,DBL0H,DBL0L
+ lsr_s DBL0H,DBL0H
+ j_s.d [blink]
+ add_l r0,DBL0H,r12
+
+/* We would generally say that NaNs must have a non-zero high fraction part,
+ but to allow hardware double precision floating point to interoperate
+ with single precision software floating point, we make an exception here.
+ The cost is to replace a tst_s DBL0H with an or.f DBL0L,DBL0L,DBL0H .
+ As we start out unaligned, and there is an odd number of other short insns,
+ we have a choice of letting this cost us a misalign penalty or
+ 4 more bytes (if we align the code). We choose the former here because
+ infinity / NaN is not expected to be prevalent in time-critical code. */
+.Linf_nan:
+ or.f DBL0L,DBL0L,DBL0H
+ mov_s r0,1
+ add.ne r2,r2,1
+ tst r2,0x7ff
+ asl.ne r0,r0,23
+ btst_s r12,11
+ neg r0,r0
+ j_s.d [blink]
+ bxor.eq r0,r0,31
+ ENDFUNC(__truncdfsf2)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "arc-ieee-754.h"
+/* inputs: DBL0, DBL1
+ output: z flag
+ clobber: r12, flags
+ For NaNs, bit 19.. bit 30 of the high word must be set. */
+#if 0 /* DEBUG */
+ .global __uneqdf2
+ .balign 4
+ FUNC(__uneqdf2)
+__uneqdf2:
+ st.a r11,[sp,-4]` push_s blink` st.a r10,[sp,-4]` st.a r9,[sp,-4]
+ st.a r8,[sp,-4]` st.a r7,[sp,-4]` st.a r6,[sp,-4]` st.a r5,[sp,-4]
+ st.a r4,[sp,-4]` push_s r3` push_s r2` push_s r1`
+ bl.d __eqdf2_c` push_s r0
+ push_s r0` ld_s r0, [sp,4]` ld_s r1, [sp,8]` ld_s r2,[sp,12]
+ bl.d __unorddf2_c` ld_s r3,[sp,16]
+ ld.ab r11,[sp,4]` tst r0,r0` mov.ne r11,0
+ pop_s r0` pop_s r1` pop_s r2` pop_s r3
+ ld.ab r4,[sp,4]` ld.ab r5,[sp,4]` ld.ab r6,[sp,4]`
+ ld.ab r7,[sp,4]` ld.ab r8,[sp,4]` ld.ab r9,[sp,4]
+ bl.d __uneqdf2_asm` ld.ab r10,[sp,4]
+ pop_s blink
+ breq.d r11,0,0f
+ ld.ab r11,[sp,4]
+ jne_s [blink]
+ bl abort
+0: jeq_s [blink]
+ bl abort
+ ENDFUNC(__uneqdf2)
+#define __uneqdf2 __uneqdf2_asm
+#endif /* DEBUG */
+ .global __uneqdf2
+ .balign 4
+ HIDDEN_FUNC(__uneqdf2)
+__uneqdf2:
+ cmp_s DBL0H,DBL1H
+ cmp.eq DBL0L,DBL1L
+ jeq_s [blink]
+ or r12,DBL0H,DBL1H
+ or.f 0,DBL0L,DBL1L
+ bclr.eq.f r12,r12,31
+ jeq_s [blink]
+ mov_s r12, \
+ 0x7ff80000
+ bic.f 0,r12,DBL0H
+ j_s.d [blink]
+ bic.ne.f r12,r12,DBL1H
+ ENDFUNC(__uneqdf2)
--- /dev/null
+/* Copyright (C) 2008-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+#include "arc-ieee-754.h"
+/* inputs: r0, r1
+ output: z flag
+ clobber: r12, flags
+ For NaNs, bit 22 .. bit 30 must be set. */
+#if 0 /* DEBUG */
+ .global __uneqsf2
+ .balign 4
+ FUNC(__uneqsf2)
+__uneqsf2:
+ st.a r11,[sp,-4]` push_s blink` st.a r10,[sp,-4]` st.a r9,[sp,-4]
+ st.a r8,[sp,-4]` st.a r7,[sp,-4]` st.a r6,[sp,-4]` st.a r5,[sp,-4]
+ st.a r4,[sp,-4]` push_s r3` push_s r2` push_s r1`
+ bl.d __eqsf2_c` push_s r0
+ push_s r0` ld_s r0, [sp,4]
+ bl.d __unordsf2_c` ld_s r1,[sp,8]
+ ld.ab r11,[sp,4]` tst r0,r0` mov.ne r11,0
+ pop_s r0` pop_s r1` pop_s r2` pop_s r3
+ ld.ab r4,[sp,4]` ld.ab r5,[sp,4]` ld.ab r6,[sp,4]`
+ ld.ab r7,[sp,4]` ld.ab r8,[sp,4]` ld.ab r9,[sp,4]
+ bl.d __uneqsf2_asm` ld.ab r10,[sp,4]
+ pop_s blink
+ breq.d r11,0,0f
+ ld.ab r11,[sp,4]
+ jne_s [blink]
+ bl abort
+0: jeq_s [blink]
+ bl abort
+ ENDFUNC(__uneqsf2)
+#define __uneqsf2 __uneqsf2_asm
+#endif /* DEBUG */
+ .global __uneqsf2
+ .balign 4
+ HIDDEN_FUNC(__uneqsf2)
+__uneqsf2:
+ mov_s r12, \
+ 0x7fc00000
+ bic.f 0,r12,r0
+ bic.ne.f r12,r12,r1
+ or r12,r0,r1
+ bmsk.ne.f r12,r12,30
+ j_s.d [blink]
+ cmp.ne r0,r1
+ ENDFUNC(__uneqsf2)
--- /dev/null
+/* .init/.fini section handling + C++ global constructor/destructor handling.
+ This file is based on crtstuff.c, sol2-crti.asm, sol2-crtn.asm.
+
+ Copyright (C) 1995, 1997, 1998, 2007-2012 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 3, or (at your option)
+any later version.
+
+GCC 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 General Public License for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+/* Declare a pointer to void function type. */
+typedef void (*func_ptr) (void);
+
+#ifdef CRT_INIT
+
+/* NOTE: In order to be able to support SVR4 shared libraries, we arrange
+ to have one set of symbols { __CTOR_LIST__, __DTOR_LIST__, __CTOR_END__,
+ __DTOR_END__ } per root executable and also one set of these symbols
+ per shared library. So in any given whole process image, we may have
+ multiple definitions of each of these symbols. In order to prevent
+ these definitions from conflicting with one another, and in order to
+ ensure that the proper lists are used for the initialization/finalization
+ of each individual shared library (respectively), we give these symbols
+ only internal (i.e. `static') linkage, and we also make it a point to
+ refer to only the __CTOR_END__ symbol in crtfini.o and the __DTOR_LIST__
+ symbol in crtinit.o, where they are defined. */
+
+static func_ptr __CTOR_LIST__[1] __attribute__ ((section (".ctors")))
+ = { (func_ptr) (-1) };
+
+static func_ptr __DTOR_LIST__[1] __attribute__ ((section (".dtors")))
+ = { (func_ptr) (-1) };
+
+/* Run all the global destructors on exit from the program. */
+
+/* Some systems place the number of pointers in the first word of the
+ table. On SVR4 however, that word is -1. In all cases, the table is
+ null-terminated. On SVR4, we start from the beginning of the list and
+ invoke each per-compilation-unit destructor routine in order
+ until we find that null.
+
+ Note that this function MUST be static. There will be one of these
+ functions in each root executable and one in each shared library, but
+ although they all have the same code, each one is unique in that it
+ refers to one particular associated `__DTOR_LIST__' which belongs to the
+ same particular root executable or shared library file. */
+
+static void __do_global_dtors (void)
+asm ("__do_global_dtors") __attribute__ ((section (".text")));
+
+static void
+__do_global_dtors (void)
+{
+ func_ptr *p;
+ for (p = __DTOR_LIST__ + 1; *p; p++)
+ (*p) ();
+}
+
+/* .init section start.
+ This must appear at the start of the .init section. */
+
+asm ("\n\
+ .section .init\n\
+ .global init\n\
+ .word 0\n\
+init:\n\
+ st blink,[sp,4]\n\
+ st fp,[sp]\n\
+ mov fp,sp\n\
+ sub sp,sp,16\n\
+");
+
+/* .fini section start.
+ This must appear at the start of the .init section. */
+
+asm ("\n\
+ .section .fini\n\
+ .global fini\n\
+ .word 0\n\
+fini:\n\
+ st blink,[sp,4]\n\
+ st fp,[sp]\n\
+ mov fp,sp\n\
+ sub sp,sp,16\n\
+ bl.nd __do_global_dtors\n\
+");
+
+#endif /* CRT_INIT */
+
+#ifdef CRT_FINI
+
+/* Put a word containing zero at the end of each of our two lists of function
+ addresses. Note that the words defined here go into the .ctors and .dtors
+ sections of the crtend.o file, and since that file is always linked in
+ last, these words naturally end up at the very ends of the two lists
+ contained in these two sections. */
+
+static func_ptr __CTOR_END__[1] __attribute__ ((section (".ctors")))
+ = { (func_ptr) 0 };
+
+static func_ptr __DTOR_END__[1] __attribute__ ((section (".dtors")))
+ = { (func_ptr) 0 };
+
+/* Run all global constructors for the program.
+ Note that they are run in reverse order. */
+
+static void __do_global_ctors (void)
+asm ("__do_global_ctors") __attribute__ ((section (".text")));
+
+static void
+__do_global_ctors (void)
+{
+ func_ptr *p;
+ for (p = __CTOR_END__ - 1; *p != (func_ptr) -1; p--)
+ (*p) ();
+}
+
+/* .init section end.
+ This must live at the end of the .init section. */
+
+asm ("\n\
+ .section .init\n\
+ bl.nd __do_global_ctors\n\
+ ld blink,[fp,4]\n\
+ j.d blink\n\
+ ld.a fp,[sp,16]\n\
+");
+
+/* .fini section end.
+ This must live at the end of the .fini section. */
+
+asm ("\n\
+ .section .fini\n\
+ ld blink,[fp,4]\n\
+ j.d blink\n\
+ ld.a fp,[sp,16]\n\
+");
+
+#endif /* CRT_FINI */
--- /dev/null
+; libgcc1 routines for Synopsys DesignWare ARC cpu.
+
+/* Copyright (C) 1995, 1997, 2007-2013 Free Software Foundation, Inc.
+ Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+ on behalf of Synopsys Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free
+Software Foundation; either version 3, or (at your option) any later
+version.
+
+GCC 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 General Public License
+for more details.
+
+Under Section 7 of GPL version 3, you are granted additional
+permissions described in the GCC Runtime Library Exception, version
+3.1, as published by the Free Software Foundation.
+
+You should have received a copy of the GNU General Public License and
+a copy of the GCC Runtime Library Exception along with this program;
+see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+<http://www.gnu.org/licenses/>. */
+
+/* As a special exception, if you link this library with other files,
+ some of which are compiled with GCC, to produce an executable,
+ this library does not by itself cause the resulting executable
+ to be covered by the GNU General Public License.
+ This exception does not however invalidate any other reasons why
+ the executable file might be covered by the GNU General Public License. */
+
+
+ /* ANSI concatenation macros. */
+
+ #define CONCAT1(a, b) CONCAT2(a, b)
+ #define CONCAT2(a, b) a ## b
+
+ /* Use the right prefix for global labels. */
+
+ #define SYM(x) CONCAT1 (__USER_LABEL_PREFIX__, x)
+
+#ifndef WORKING_ASSEMBLER
+#define abs_l abs
+#define asl_l asl
+#define mov_l mov
+#endif
+
+#define FUNC(X) .type SYM(X),@function
+#define HIDDEN_FUNC(X) FUNC(X)` .hidden X
+#define ENDFUNC0(X) .Lfe_##X: .size X,.Lfe_##X-X
+#define ENDFUNC(X) ENDFUNC0(X)
+
+
+
+#ifdef L_mulsi3
+ .section .text
+ .align 4
+
+ .global SYM(__mulsi3)
+SYM(__mulsi3):
+
+/* This the simple version.
+
+ while (a)
+ {
+ if (a & 1)
+ r += b;
+ a >>= 1;
+ b <<= 1;
+ }
+*/
+
+#if defined (__ARC_MUL64__)
+ FUNC(__mulsi3)
+ mulu64 r0,r1
+ j_s.d [blink]
+ mov_s r0,mlo
+ ENDFUNC(__mulsi3)
+#elif defined (__ARC700__)
+ HIDDEN_FUNC(__mulsi3)
+ mpyu r0,r0,r1
+ nop_s
+ j_s [blink]
+ ENDFUNC(__mulsi3)
+#elif defined (__ARC_NORM__)
+ FUNC(__mulsi3)
+ norm.f r2,r0
+ rsub lp_count,r2,31
+ mov.mi lp_count,32
+ mov_s r2,r0
+ mov_s r0,0
+ lpnz @.Lend ; loop is aligned
+ lsr.f r2,r2
+ add.cs r0,r0,r1
+ add_s r1,r1,r1
+.Lend: j_s [blink]
+ ENDFUNC(__mulsi3)
+#elif !defined (__OPTIMIZE_SIZE__) && !defined(__ARC601__)
+ /* Up to 3.5 times faster than the simpler code below, but larger. */
+ FUNC(__mulsi3)
+ ror.f r2,r0,4
+ mov_s r0,0
+ add3.mi r0,r0,r1
+ asl.f r2,r2,2
+ add2.cs r0,r0,r1
+ jeq_s [blink]
+.Loop:
+ add1.mi r0,r0,r1
+ asl.f r2,r2,2
+ add.cs r0,r0,r1
+ asl_s r1,r1,4
+ ror.f r2,r2,8
+ add3.mi r0,r0,r1
+ asl.f r2,r2,2
+ bne.d .Loop
+ add2.cs r0,r0,r1
+ j_s [blink]
+ ENDFUNC(__mulsi3)
+#elif !defined (__OPTIMIZE_SIZE__) /* __ARC601__ */
+ FUNC(__mulsi3)
+ lsr.f r2,r0
+ mov_s r0,0
+ mov_s r3,0
+ add.cs r0,r0,r1
+.Loop:
+ lsr.f r2,r2
+ add1.cs r0,r0,r1
+ lsr.f r2,r2
+ add2.cs r0,r0,r1
+ lsr.f r2,r2
+ add3.cs r0,r0,r1
+ bne.d .Loop
+ add3 r1,r3,r1
+ j_s [blink]
+ ENDFUNC(__mulsi3)
+#else
+/********************************************************/
+ FUNC(__mulsi3)
+ mov_s r2,0 ; Accumulate result here.
+.Lloop:
+ bbit0 r0,0,@.Ly
+ add_s r2,r2,r1 ; r += b
+.Ly:
+ lsr_s r0,r0 ; a >>= 1
+ asl_s r1,r1 ; b <<= 1
+ brne_s r0,0,@.Lloop
+.Ldone:
+ j_s.d [blink]
+ mov_s r0,r2
+ ENDFUNC(__mulsi3)
+/********************************************************/
+#endif
+
+#endif /* L_mulsi3 */
+
+#ifdef L_umulsidi3
+ .section .text
+ .align 4
+
+ .global SYM(__umulsidi3)
+SYM(__umulsidi3):
+ HIDDEN_FUNC(__umulsidi3)
+/* We need ARC700 /ARC_MUL64 definitions of __umulsidi3 / __umulsi3_highpart
+ in case some code has been compiled without multiply support enabled,
+ but linked with the multiply-support enabled libraries.
+ For ARC601 (i.e. without a barrel shifter), we also use umuldisi3 as our
+ umulsi3_highpart implementation; the use of the latter label doesn't
+ actually benefit ARC601 platforms, but is useful when ARC601 code is linked
+ against other libraries. */
+#if defined (__ARC700__) || defined (__ARC_MUL64__) || defined (__ARC601__)
+ .global SYM(__umulsi3_highpart)
+SYM(__umulsi3_highpart):
+ HIDDEN_FUNC(__umulsi3_highpart)
+#endif
+
+/* This the simple version.
+
+ while (a)
+ {
+ if (a & 1)
+ r += b;
+ a >>= 1;
+ b <<= 1;
+ }
+*/
+#include "ieee-754/arc-ieee-754.h"
+
+#ifdef __ARC700__
+ mov_s r12,DBL0L
+ mpyu DBL0L,r12,DBL0H
+ j_s.d [blink]
+ mpyhu DBL0H,r12,DBL0H
+#elif defined (__ARC_MUL64__)
+/* Likewise for __ARC_MUL64__ */
+ mulu64 r0,r1
+ mov_s DBL0L,mlo
+ j_s.d [blink]
+ mov_s DBL0H,mhi
+#else /* !__ARC700__ && !__ARC_MUL64__ */
+/* Although it might look tempting to extend this to handle muldi3,
+ using mulsi3 twice with 2.25 cycles per 32 bit add is faster
+ than one loop with 3 or four cycles per 32 bit add. */
+ asl.f r12,0 ; Top part of b.
+ mov_s r2,0 ; Accumulate result here.
+ bbit1.d r0,0,@.Ladd
+ mov_s r3,0
+.Llooptst:
+ rlc r12,r12
+ breq r0,0,@.Ldone ; while (a)
+.Lloop:
+ asl.f r1,r1 ; b <<= 1
+ bbit0.d r0,1,@.Llooptst
+ lsr r0,r0 ; a >>= 1
+ rlc r12,r12
+.Ladd:
+ add.f r3,r3,r1 ; r += b
+ brne.d r0,0,@.Lloop ; while (a);
+ adc r2,r2,r12
+.Ldone:
+ mov_s DBL0L,r3
+ j_s.d [blink]
+ mov DBL0H,r2
+#endif /* !__ARC700__*/
+ ENDFUNC(__umulsidi3)
+#if defined (__ARC700__) || defined (__ARC_MUL64__) || defined (__ARC601__)
+ ENDFUNC(__umulsi3_highpart)
+#endif
+#endif /* L_umulsidi3 */
+
+#ifdef L_umulsi3_highpart
+#include "ieee-754/arc-ieee-754.h"
+/* For use without a barrel shifter, and for ARC700 / ARC_MUL64, the
+ mulsidi3 algorithms above look better, so for these, there is an
+ extra label up there. */
+#if !defined (__ARC700__) && !defined (__ARC_MUL64__) && !defined (__ARC601__)
+ .global SYM(__umulsi3_highpart)
+SYM(__umulsi3_highpart):
+ HIDDEN_FUNC(__umulsi3_highpart)
+ mov_s r2,0
+ mov_s r3,32
+.Loop:
+ lsr.f r0,r0
+ add.cs.f r2,r2,r1
+ sub_s r3,r3,1
+ brne.d r0,0,.Loop
+ rrc r2,r2
+ j_s.d [blink]
+/* Make the result register peephole-compatible with mulsidi3. */
+ lsr DBL0H,r2,r3
+ ENDFUNC(__umulsi3_highpart)
+#endif /* !__ARC700__ && !__ARC601__ */
+#endif /* L_umulsi3_highpart */
+
+#ifdef L_divmod_tools
+
+; Utilities used by all routines.
+
+ .section .text
+
+/*
+unsigned long
+udivmodsi4(int modwanted, unsigned long num, unsigned long den)
+{
+ unsigned long bit = 1;
+ unsigned long res = 0;
+
+ while (den < num && bit && !(den & (1L<<31)))
+ {
+ den <<=1;
+ bit <<=1;
+ }
+ while (bit)
+ {
+ if (num >= den)
+ {
+ num -= den;
+ res |= bit;
+ }
+ bit >>=1;
+ den >>=1;
+ }
+ if (modwanted) return num;
+ return res;
+}
+*/
+
+; inputs: r0 = numerator, r1 = denominator
+; outputs: r0 = quotient, r1 = remainder, r2/r3 trashed
+
+ .balign 4
+ .global SYM(__udivmodsi4)
+ FUNC(__udivmodsi4)
+SYM(__udivmodsi4):
+
+#if defined (__ARC700__)
+/* Normalize divisor and divident, and then use the appropriate number of
+ divaw (the number of result bits, or one more) to produce the result.
+ There are some special conditions that need to be tested:
+ - We can only directly normalize unsigned numbers that fit in 31 bit. For
+ the divisor, we test early on that it is not 'negative'.
+ - divaw can't corrrectly process a divident that is larger than the divisor.
+ We handle this be checking that the divident prior to normalization is
+ not larger than the normalized divisor. As we then already know then
+ that the divisor fits 31 bit, this check also makes sure that the
+ divident fits.
+ - ordinary normalization of the divident could make it larger than the
+ normalized divisor, which again would be unsuitable for divaw.
+ Thus, we want to shift left the divident by one less, except that we
+ want to leave it alone if it is already 31 bit. To this end, we
+ double the input to norm with adds.
+ - If the divident has less bits than the divisor, that would leave us
+ with a negative number of divaw to execute. Although we could use a
+ conditional loop to avoid excess divaw, and then the quotient could
+ be extracted correctly as there'd be more than enough zero bits, the
+ remainder would be shifted left too far, requiring a conditional shift
+ right. The cost of that shift and the possible mispredict on the
+ conditional loop cost as much as putting in an early check for a zero
+ result. */
+ bmsk r3,r0,29
+ brne.d r3,r0,.Large_dividend
+ norm.f r2,r1
+ brlo r0,r1,.Lret0
+ norm r3,r0
+ asl_s r1,r1,r2
+ sub_s r3,r3,1
+ asl_l r0,r0,r3 ; not short to keep loop aligned
+ sub lp_count,r2,r3
+ lp .Ldiv_end
+ divaw r0,r0,r1
+.Ldiv_end:sub_s r3,r2,1
+ lsr r1,r0,r2
+ j_s.d [blink]
+ bmsk r0,r0,r3
+
+ .balign 4
+.Large_dividend:
+ bmi .Ltrivial
+ asl_s r1,r1,r2
+ mov_s r3,0
+ sub1.f r4,r0,r1
+ mov.lo r4,r0
+ mov.hs r3,2
+ cmp r4,r1
+ sub.hs r4,r4,r1
+ add.hs r3,r3,1
+ mov.f lp_count,r2
+ lpne .Ldiv_end2
+ divaw r4,r4,r1
+.Ldiv_end2:asl r0,r3,r2
+ lsr r1,r4,r2
+ sub_s r2,r2,1
+ bmsk r4,r4,r2
+ j_s.d [blink]
+ or.ne r0,r0,r4
+
+.Lret0:
+ mov_s r1,r0
+ j_s.d [blink]
+ mov_l r0,0
+ .balign 4
+.Ltrivial:
+ sub.f r1,r0,r1
+ mov.c r1,r0
+ mov_s r0,1
+ j_s.d [blink]
+ mov.c r0,0
+#elif !defined (__OPTIMIZE_SIZE__)
+#ifdef __ARC_NORM__
+ lsr_s r2,r0
+ brhs.d r1,r2,.Lret0_3
+ norm r2,r2
+ norm r3,r1
+ sub_s r3,r3,r2
+ asl_s r1,r1,r3
+ sub1.f 0,r0,r1
+ lsr.cs r1,r1,1
+ sbc r2,r3,0
+ sub1 r0,r0,r1
+ cmp_s r0,r1
+ mov.f lp_count,r2
+#else /* ! __ARC_NORM__ */
+ lsr_s r2,r0
+ brhs.d r1,r2,.Lret0_3
+ mov lp_count,32
+.Lloop1:
+ asl_s r1,r1 ; den <<= 1
+ brls.d r1,r2,@.Lloop1
+ sub lp_count,lp_count,1
+ sub_s r0,r0,r1
+ lsr_s r1,r1
+ cmp_s r0,r1
+ xor.f r2,lp_count,31
+ mov_s lp_count,r2
+#endif /* !__ARC_NORM__ */
+ sub.cc r0,r0,r1
+ mov_s r3,3
+ sbc r3,r3,0
+#ifndef __ARC601__
+ asl_s r3,r3,r2
+ rsub r1,r1,1
+ lpne @.Lloop2_end
+ add1.f r0,r1,r0
+ sub.cc r0,r0,r1
+.Lloop2_end:
+ lsr r1,r0,r2
+#else
+ rsub r1,r1,1
+ lpne @.Lloop2_end
+ asl_s r3,r3
+ add1.f r0,r1,r0
+ sub.cc r0,r0,r1
+.Lloop2_end:
+ lsr_s r1,r0
+ lsr.f lp_count,r2
+ mov.cc r1,r0
+ lpnz 1f
+ lsr_s r1,r1
+ lsr_s r1,r1
+1:
+#endif
+ bmsk r0,r0,r2
+ bclr r0,r0,r2
+ j_s.d [blink]
+ or_s r0,r0,r3
+.Lret0_3:
+#if 0 /* Slightly shorter, but slower. */
+ lp .Loop3_end
+ brhi.d r1,r0,.Loop3_end
+ sub_s r0,r0,r1
+.Loop3_end
+ add_s r1,r1,r0
+ j_s.d [blink]
+ rsub r0,lp_count,32-1
+#else
+ mov_s r4,r1
+ sub.f r1,r0,r1
+ sbc r0,r0,r0
+ sub.cc.f r1,r1,r4
+ sbc r0,r0,0
+ sub.cc.f r1,r1,r4
+ sbc r0,r0,-3
+ j_s.d [blink]
+ add.cs r1,r1,r4
+#endif
+#else /* Arctangent-A5 */
+ breq_s r1,0,@.Ldivmodend
+ mov_s r2,1 ; bit = 1
+ mov_s r3,0 ; res = 0
+.Lloop1:
+ brhs r1,r0,@.Lloop2
+ bbit1 r1,31,@.Lloop2
+ asl_s r1,r1 ; den <<= 1
+ b.d @.Lloop1
+ asl_s r2,r2 ; bit <<= 1
+.Lloop2:
+ brlo r0,r1,@.Lshiftdown
+ sub_s r0,r0,r1 ; num -= den
+ or_s r3,r3,r2 ; res |= bit
+.Lshiftdown:
+ lsr_s r2,r2 ; bit >>= 1
+ lsr_s r1,r1 ; den >>= 1
+ brne_s r2,0,@.Lloop2
+.Ldivmodend:
+ mov_s r1,r0 ; r1 = mod
+ j.d [blink]
+ mov_s r0,r3 ; r0 = res
+/******************************************************/
+#endif
+ ENDFUNC(__udivmodsi4)
+
+#endif
+
+#ifdef L_udivsi3
+ .section .text
+ .align 4
+
+ .global SYM(__udivsi3)
+ FUNC(__udivsi3)
+SYM(__udivsi3):
+ b @SYM(__udivmodsi4)
+ ENDFUNC(__udivsi3)
+#if 0 /* interferes with linux loader */
+ .section .__arc_profile_forward, "a"
+ .long SYM(__udivsi3)
+ .long SYM(__udivmodsi4)
+ .long 65536
+#endif
+
+#endif /* L_udivsi3 */
+
+#ifdef L_divsi3
+ .section .text
+ .align 4
+
+ .global SYM(__divsi3)
+ FUNC(__divsi3)
+
+#ifndef __ARC700__
+SYM(__divsi3):
+ /* A5 / ARC60? */
+ mov r7,blink
+ xor r6,r0,r1
+ abs_s r0,r0
+ bl.d @SYM(__udivmodsi4)
+ abs_s r1,r1
+ tst r6,r6
+ j.d [r7]
+ neg.mi r0,r0
+#else /* !ifndef __ARC700__ */
+ ;; We can use the abs, norm, divaw and mpy instructions for ARC700
+#define MULDIV
+#ifdef MULDIV
+/* This table has been generated by divtab-arc700.c. */
+/* 1/512 .. 1/256, normalized. There is a leading 1 in bit 31.
+ For powers of two, we list unnormalized numbers instead. The values
+ for powers of 2 are loaded, but not used. The value for 1 is actually
+ the first instruction after .Lmuldiv. */
+ .balign 4
+.Ldivtab:
+
+ .long 0x1000000
+ .long 0x80808081
+ .long 0x81020409
+ .long 0x81848DA9
+ .long 0x82082083
+ .long 0x828CBFBF
+ .long 0x83126E98
+ .long 0x83993053
+ .long 0x84210843
+ .long 0x84A9F9C9
+ .long 0x85340854
+ .long 0x85BF3762
+ .long 0x864B8A7E
+ .long 0x86D90545
+ .long 0x8767AB60
+ .long 0x87F78088
+ .long 0x88888889
+ .long 0x891AC73B
+ .long 0x89AE408A
+ .long 0x8A42F871
+ .long 0x8AD8F2FC
+ .long 0x8B70344B
+ .long 0x8C08C08D
+ .long 0x8CA29C05
+ .long 0x8D3DCB09
+ .long 0x8DDA5203
+ .long 0x8E78356E
+ .long 0x8F1779DA
+ .long 0x8FB823EF
+ .long 0x905A3864
+ .long 0x90FDBC0A
+ .long 0x91A2B3C5
+ .long 0x92492493
+ .long 0x92F11385
+ .long 0x939A85C5
+ .long 0x94458095
+ .long 0x94F20950
+ .long 0x95A02569
+ .long 0x964FDA6D
+ .long 0x97012E03
+ .long 0x97B425EE
+ .long 0x9868C80A
+ .long 0x991F1A52
+ .long 0x99D722DB
+ .long 0x9A90E7DA
+ .long 0x9B4C6F9F
+ .long 0x9C09C09D
+ .long 0x9CC8E161
+ .long 0x9D89D89E
+ .long 0x9E4CAD24
+ .long 0x9F1165E8
+ .long 0x9FD809FE
+ .long 0xA0A0A0A1
+ .long 0xA16B312F
+ .long 0xA237C32C
+ .long 0xA3065E40
+ .long 0xA3D70A3E
+ .long 0xA4A9CF1E
+ .long 0xA57EB503
+ .long 0xA655C43A
+ .long 0xA72F053A
+ .long 0xA80A80A9
+ .long 0xA8E83F58
+ .long 0xA9C84A48
+ .long 0xAAAAAAAB
+ .long 0xAB8F69E3
+ .long 0xAC769185
+ .long 0xAD602B59
+ .long 0xAE4C415D
+ .long 0xAF3ADDC7
+ .long 0xB02C0B03
+ .long 0xB11FD3B9
+ .long 0xB21642C9
+ .long 0xB30F6353
+ .long 0xB40B40B5
+ .long 0xB509E68B
+ .long 0xB60B60B7
+ .long 0xB70FBB5B
+ .long 0xB81702E1
+ .long 0xB92143FB
+ .long 0xBA2E8BA3
+ .long 0xBB3EE722
+ .long 0xBC52640C
+ .long 0xBD691048
+ .long 0xBE82FA0C
+ .long 0xBFA02FE9
+ .long 0xC0C0C0C1
+ .long 0xC1E4BBD6
+ .long 0xC30C30C4
+ .long 0xC4372F86
+ .long 0xC565C87C
+ .long 0xC6980C6A
+ .long 0xC7CE0C7D
+ .long 0xC907DA4F
+ .long 0xCA4587E7
+ .long 0xCB8727C1
+ .long 0xCCCCCCCD
+ .long 0xCE168A78
+ .long 0xCF6474A9
+ .long 0xD0B69FCC
+ .long 0xD20D20D3
+ .long 0xD3680D37
+ .long 0xD4C77B04
+ .long 0xD62B80D7
+ .long 0xD79435E6
+ .long 0xD901B204
+ .long 0xDA740DA8
+ .long 0xDBEB61EF
+ .long 0xDD67C8A7
+ .long 0xDEE95C4D
+ .long 0xE070381D
+ .long 0xE1FC780F
+ .long 0xE38E38E4
+ .long 0xE525982B
+ .long 0xE6C2B449
+ .long 0xE865AC7C
+ .long 0xEA0EA0EB
+ .long 0xEBBDB2A6
+ .long 0xED7303B6
+ .long 0xEF2EB720
+ .long 0xF0F0F0F1
+ .long 0xF2B9D649
+ .long 0xF4898D60
+ .long 0xF6603D99
+ .long 0xF83E0F84
+ .long 0xFA232CF3
+ .long 0xFC0FC0FD
+ .long 0xFE03F810
+ .long 0x2000000
+ .long 0x81020409
+ .long 0x82082083
+ .long 0x83126E98
+ .long 0x84210843
+ .long 0x85340854
+ .long 0x864B8A7E
+ .long 0x8767AB60
+ .long 0x88888889
+ .long 0x89AE408A
+ .long 0x8AD8F2FC
+ .long 0x8C08C08D
+ .long 0x8D3DCB09
+ .long 0x8E78356E
+ .long 0x8FB823EF
+ .long 0x90FDBC0A
+ .long 0x92492493
+ .long 0x939A85C5
+ .long 0x94F20950
+ .long 0x964FDA6D
+ .long 0x97B425EE
+ .long 0x991F1A52
+ .long 0x9A90E7DA
+ .long 0x9C09C09D
+ .long 0x9D89D89E
+ .long 0x9F1165E8
+ .long 0xA0A0A0A1
+ .long 0xA237C32C
+ .long 0xA3D70A3E
+ .long 0xA57EB503
+ .long 0xA72F053A
+ .long 0xA8E83F58
+ .long 0xAAAAAAAB
+ .long 0xAC769185
+ .long 0xAE4C415D
+ .long 0xB02C0B03
+ .long 0xB21642C9
+ .long 0xB40B40B5
+ .long 0xB60B60B7
+ .long 0xB81702E1
+ .long 0xBA2E8BA3
+ .long 0xBC52640C
+ .long 0xBE82FA0C
+ .long 0xC0C0C0C1
+ .long 0xC30C30C4
+ .long 0xC565C87C
+ .long 0xC7CE0C7D
+ .long 0xCA4587E7
+ .long 0xCCCCCCCD
+ .long 0xCF6474A9
+ .long 0xD20D20D3
+ .long 0xD4C77B04
+ .long 0xD79435E6
+ .long 0xDA740DA8
+ .long 0xDD67C8A7
+ .long 0xE070381D
+ .long 0xE38E38E4
+ .long 0xE6C2B449
+ .long 0xEA0EA0EB
+ .long 0xED7303B6
+ .long 0xF0F0F0F1
+ .long 0xF4898D60
+ .long 0xF83E0F84
+ .long 0xFC0FC0FD
+ .long 0x4000000
+ .long 0x82082083
+ .long 0x84210843
+ .long 0x864B8A7E
+ .long 0x88888889
+ .long 0x8AD8F2FC
+ .long 0x8D3DCB09
+ .long 0x8FB823EF
+ .long 0x92492493
+ .long 0x94F20950
+ .long 0x97B425EE
+ .long 0x9A90E7DA
+ .long 0x9D89D89E
+ .long 0xA0A0A0A1
+ .long 0xA3D70A3E
+ .long 0xA72F053A
+ .long 0xAAAAAAAB
+ .long 0xAE4C415D
+ .long 0xB21642C9
+ .long 0xB60B60B7
+ .long 0xBA2E8BA3
+ .long 0xBE82FA0C
+ .long 0xC30C30C4
+ .long 0xC7CE0C7D
+ .long 0xCCCCCCCD
+ .long 0xD20D20D3
+ .long 0xD79435E6
+ .long 0xDD67C8A7
+ .long 0xE38E38E4
+ .long 0xEA0EA0EB
+ .long 0xF0F0F0F1
+ .long 0xF83E0F84
+ .long 0x8000000
+ .long 0x84210843
+ .long 0x88888889
+ .long 0x8D3DCB09
+ .long 0x92492493
+ .long 0x97B425EE
+ .long 0x9D89D89E
+ .long 0xA3D70A3E
+ .long 0xAAAAAAAB
+ .long 0xB21642C9
+ .long 0xBA2E8BA3
+ .long 0xC30C30C4
+ .long 0xCCCCCCCD
+ .long 0xD79435E6
+ .long 0xE38E38E4
+ .long 0xF0F0F0F1
+ .long 0x10000000
+ .long 0x88888889
+ .long 0x92492493
+ .long 0x9D89D89E
+ .long 0xAAAAAAAB
+ .long 0xBA2E8BA3
+ .long 0xCCCCCCCD
+ .long 0xE38E38E4
+ .long 0x20000000
+ .long 0x92492493
+ .long 0xAAAAAAAB
+ .long 0xCCCCCCCD
+ .long 0x40000000
+ .long 0xAAAAAAAB
+ .long 0x80000000
+__muldiv:
+ neg r4,r2
+ ld.as r5,[pcl,r4]
+ abs_s r12,r0
+ bic.f 0,r2,r4
+ mpyhu.ne r12,r12,r5
+ norm r3,r2
+ xor.f 0,r0,r1
+ ; write port allocation stall
+ rsub r3,r3,30
+ lsr r0,r12,r3
+ j_s.d [blink]
+ neg.mi r0,r0
+
+ .balign 4
+SYM(__divsi3):
+ norm r3,r1
+ abs_s r2,r1
+ brhs r3,23,__muldiv
+ norm r4,r0
+ abs_l r12,r0
+ brhs r4,r3,.Lonebit
+ asl_s r2,r2,r3
+ asl r12,r12,r4
+ sub lp_count,r3,r4
+ sub.f r12,r12,r2
+ brge.d r12,r2,.Lsbit
+ sub r4,r3,r4
+ add.lo r12,r12,r2
+ lp .Ldivend
+.Ldivstart:divaw r12,r12,r2
+.Ldivend:xor_s r1,r1,r0
+ sub r0,r4,1
+ bmsk r0,r12,r0
+ bset.hs r0,r0,r4
+ tst_s r1,r1
+ j_s.d [blink]
+ neg.mi r0,r0
+.Lonebit:
+ xor_s r1,r1,r0
+ asr_s r1,r1,31
+ sub1.f 0,r12,r2 ; special case: -2**(n+1) / 2**n
+ or r0,r1,1
+ add.eq r0,r0,r0
+ cmp_s r12,r2
+ j_s.d [blink]
+ mov.lo r0,0
+.Lsbit:
+ ; Need to handle special cases involving negative powers of two:
+ ; r12,r2 are normalized dividend / divisor;
+ ; divide anything by 0x80000000, or divide 0x80000000 by 0x40000000
+ add_s r12,r12,r2
+ xor_s r1,r1,r0
+ rsub r4,r4,-1
+ ror r0,r12,r4
+ tst_s r2,r2
+ bmsk r0,r0,r3
+ add.pl r0,r0,r0
+ tst_s r1,r1
+ j_s.d [blink]
+ neg.mi r0,r0
+#else /* !MULDIV */
+/* This version requires that divaw works with a divisor of 0x80000000U */
+ abs_s r2,r1
+ norm r4,r0
+ neg_s r3,r2
+ norm r3,r3
+ abs_s r12,r0
+ brhs r4,r3,.Lonebit
+ asl_s r2,r2,r3
+ asl r12,r12,r4
+ sub lp_count,r3,r4
+ cmp_s r12,r2
+ sub.hs r12,r12,r2
+ lp .Ldivend
+.Ldivstart:divaw r12,r12,r2
+.Ldivend:xor_s r1,r1,r0
+ sub_s r0,r3,1
+ bmsk r0,r12,r0
+ bset.hs r0,r0,r3
+ tst_s r1,r1
+ j_s.d [blink]
+ negmi r0,r0
+.Lonebit:
+ xor_s r1,r1,r0
+ asr_s r1,r1,31
+ cmp_s r12,r2
+ mov_s r0,0
+ j_s.d [blink]
+ orhs r0,r1,1
+#endif /* MULDIV */
+
+#endif /* ifndef __ARC700__ */
+ ENDFUNC(__divsi3)
+
+
+#endif /* L_divsi3 */
+
+#ifdef L_umodsi3
+ .section .text
+ .align 4
+
+ .global SYM(__umodsi3)
+ FUNC(__umodsi3)
+SYM(__umodsi3):
+ mov r7,blink
+ bl.nd @SYM(__udivmodsi4)
+ j.d [r7]
+ mov r0,r1
+ ENDFUNC(__umodsi3)
+#if 0 /* interferes with linux loader */
+ .section .__arc_profile_forward, "a"
+ .long SYM(__umodsi3)
+ .long SYM(__udivmodsi4)
+ .long 65536
+#endif
+
+#endif /* L_umodsi3 */
+
+#ifdef L_modsi3
+ .section .text
+ .align 4
+
+ .global SYM (__modsi3)
+ FUNC(__modsi3)
+SYM(__modsi3):
+#ifndef __ARC700__
+ /* A5 / ARC60? */
+ mov_s r12,blink
+ mov_s r6,r0
+ abs_s r0,r0
+ bl.d @SYM(__udivmodsi4)
+ abs_s r1,r1
+ tst r6,r6
+ neg_s r0,r1
+ j_s.d [r12]
+ mov.pl r0,r1
+#else /* __ARC700__ */
+ abs_s r2,r1
+ norm.f r4,r0
+ neg r5,r2
+ norm r3,r5
+ abs_l r12,r0
+ brhs r4,r3,.Lonebit
+ asl_s r2,r2,r3
+ asl r12,r12,r4
+ sub lp_count,r3,r4
+ cmp_s r12,r2
+ sub.hs r12,r12,r2
+ tst_s r0,r0
+ lp .Ldivend
+.Ldivstart:divaw r12,r12,r2
+.Ldivend:
+ lsr r0,r12,r3
+ j_s.d [blink]
+ neg.mi r0,r0
+ .balign 4
+.Lonebit:neg.pl r5,r5
+ cmp_s r12,r2
+ j_s.d [blink]
+ sub.hs r0,r0,r5
+#endif /* __ARC700__ */
+ ENDFUNC(__modsi3)
+
+#endif /* L_modsi3 */
+
+#ifdef L_clzsi2
+ .section .text
+ .align 4
+ .global SYM (__clzsi2)
+SYM(__clzsi2):
+#ifdef __ARC_NORM__
+ HIDDEN_FUNC(__clzsi2)
+ norm.f r0,r0
+ mov.n r0,0
+ j_s.d [blink]
+ add.pl r0,r0,1
+ ENDFUNC(__clzsi2)
+#elif defined (__ARC601__)
+ FUNC(__clzsi2)
+ mov lp_count,10
+ mov_l r1,0
+ bset r2,r1,29
+ lp .Loop_end
+ brhs r0,r2,.Loop_end
+ add3 r0,r1,r0
+.Loop_end:
+ asl.f 0,r0
+ sub2 r0,lp_count,lp_count
+ sub.cs.f r0,r0,1
+ add r0,r0,31
+ j_s.d [blink]
+ add.pl r0,r0,1
+ ENDFUNC(__clzsi2)
+#else
+ FUNC(__clzsi2)
+ asl.f 0,r0,2
+ mov r1,-1
+.Lcheck:
+ bbit1.d r0,31,.Ldone
+ asl.pl r0,r0,3
+ bcs.d .Ldone_1
+ add_s r1,r1,3
+ bpnz.d .Lcheck
+ asl.f 0,r0,2
+ mov_s r0,32
+ j_s.d [blink]
+ mov.ne r0,r1
+.Ldone:
+ j_s.d [blink]
+ add_s r0,r1,1
+.Ldone_1:
+ j_s.d [blink]
+ sub_s r0,r1,1
+ ENDFUNC(__clzsi2)
+#endif
+#endif /* L_clzsi2 */
+ .section .text
+
+
+;;; MILLICODE THUNK LIB ;***************
+
+;;; .macro push_regs from, to, offset
+;;; st_s "\from", [sp, \offset]
+;;; .if \to-\from
+;;; push_regs "(\from+1)", \to, "(\offset+4)"
+;;; .endif
+;;; .endm
+;;; push_regs 13, 18, 0
+;;;
+
+;;;; .macro sum from, to, three
+;;;; .long \from
+;;;; .long \three
+;;;; .local regno
+;;;; .set regno, \from+1
+;;;; .set shift, 32
+;;;; .set shift, shift - 1
+;;;; # st_s %shift @3 lsl #shift
+;;;; .if \to-\from
+;;;; sum "(\from+1)", \to, "(\three)"
+;;;; .endif
+;;;; .endm
+;;;;
+;;;; SUM 0,5, 9
+;;;;
+; .altmacro
+;; .macro push_regs from=0, to=3, offset
+;; st_s r\from, [sp, \offset]
+;; .if \to-\from
+;; push_regs "\from+1 ",\to,"(\offset+4)"
+;; .endif
+;; .endm
+;;
+;; .macro expand_to_push from=13, to
+;; ; .section .text
+;; ; .align 4
+;; ; .global st_
+;; ; .type foo,
+;; st_13_to_25:
+;; ; push_regs \from, \to, 0
+;; push_regs 0,3 ;
+;; .endm
+;;
+;; expand_to_push 13,18
+;;
+;#endif
+
+#ifdef L_millicodethunk_st
+ .section .text
+ .align 4
+ .global SYM(__st_r13_to_r15)
+ .global SYM(__st_r13_to_r16)
+ .global SYM(__st_r13_to_r17)
+ .global SYM(__st_r13_to_r18)
+ .global SYM(__st_r13_to_r19)
+ .global SYM(__st_r13_to_r20)
+ .global SYM(__st_r13_to_r21)
+ .global SYM(__st_r13_to_r22)
+ .global SYM(__st_r13_to_r23)
+ .global SYM(__st_r13_to_r24)
+ .global SYM(__st_r13_to_r25)
+ HIDDEN_FUNC(__st_r13_to_r15)
+ HIDDEN_FUNC(__st_r13_to_r16)
+ HIDDEN_FUNC(__st_r13_to_r17)
+ HIDDEN_FUNC(__st_r13_to_r18)
+ HIDDEN_FUNC(__st_r13_to_r19)
+ HIDDEN_FUNC(__st_r13_to_r20)
+ HIDDEN_FUNC(__st_r13_to_r21)
+ HIDDEN_FUNC(__st_r13_to_r22)
+ HIDDEN_FUNC(__st_r13_to_r23)
+ HIDDEN_FUNC(__st_r13_to_r24)
+ HIDDEN_FUNC(__st_r13_to_r25)
+ .align 4
+SYM(__st_r13_to_r25):
+ st r25, [sp,48]
+SYM(__st_r13_to_r24):
+ st r24, [sp,44]
+SYM(__st_r13_to_r23):
+ st r23, [sp,40]
+SYM(__st_r13_to_r22):
+ st r22, [sp,36]
+SYM(__st_r13_to_r21):
+ st r21, [sp,32]
+SYM(__st_r13_to_r20):
+ st r20, [sp,28]
+SYM(__st_r13_to_r19):
+ st r19, [sp,24]
+SYM(__st_r13_to_r18):
+ st r18, [sp,20]
+SYM(__st_r13_to_r17):
+ st r17, [sp,16]
+SYM(__st_r13_to_r16):
+ st r16, [sp,12]
+SYM(__st_r13_to_r15):
+#ifdef __ARC700__
+ st r15, [sp,8] ; minimum function size to avoid stall: 6 bytes.
+#else
+ st_s r15, [sp,8]
+#endif
+ st_s r14, [sp,4]
+ j_s.d [%blink]
+ st_s r13, [sp,0]
+ ENDFUNC(__st_r13_to_r15)
+ ENDFUNC(__st_r13_to_r16)
+ ENDFUNC(__st_r13_to_r17)
+ ENDFUNC(__st_r13_to_r18)
+ ENDFUNC(__st_r13_to_r19)
+ ENDFUNC(__st_r13_to_r20)
+ ENDFUNC(__st_r13_to_r21)
+ ENDFUNC(__st_r13_to_r22)
+ ENDFUNC(__st_r13_to_r23)
+ ENDFUNC(__st_r13_to_r24)
+ ENDFUNC(__st_r13_to_r25)
+#endif /* L_millicodethunk_st */
+
+
+#ifdef L_millicodethunk_ld
+ .section .text
+ .align 4
+; ==================================
+; the loads
+
+ .global SYM(__ld_r13_to_r15)
+ .global SYM(__ld_r13_to_r16)
+ .global SYM(__ld_r13_to_r17)
+ .global SYM(__ld_r13_to_r18)
+ .global SYM(__ld_r13_to_r19)
+ .global SYM(__ld_r13_to_r20)
+ .global SYM(__ld_r13_to_r21)
+ .global SYM(__ld_r13_to_r22)
+ .global SYM(__ld_r13_to_r23)
+ .global SYM(__ld_r13_to_r24)
+ .global SYM(__ld_r13_to_r25)
+ HIDDEN_FUNC(__ld_r13_to_r15)
+ HIDDEN_FUNC(__ld_r13_to_r16)
+ HIDDEN_FUNC(__ld_r13_to_r17)
+ HIDDEN_FUNC(__ld_r13_to_r18)
+ HIDDEN_FUNC(__ld_r13_to_r19)
+ HIDDEN_FUNC(__ld_r13_to_r20)
+ HIDDEN_FUNC(__ld_r13_to_r21)
+ HIDDEN_FUNC(__ld_r13_to_r22)
+ HIDDEN_FUNC(__ld_r13_to_r23)
+ HIDDEN_FUNC(__ld_r13_to_r24)
+ HIDDEN_FUNC(__ld_r13_to_r25)
+SYM(__ld_r13_to_r25):
+ ld r25, [sp,48]
+SYM(__ld_r13_to_r24):
+ ld r24, [sp,44]
+SYM(__ld_r13_to_r23):
+ ld r23, [sp,40]
+SYM(__ld_r13_to_r22):
+ ld r22, [sp,36]
+SYM(__ld_r13_to_r21):
+ ld r21, [sp,32]
+SYM(__ld_r13_to_r20):
+ ld r20, [sp,28]
+SYM(__ld_r13_to_r19):
+ ld r19, [sp,24]
+SYM(__ld_r13_to_r18):
+ ld r18, [sp,20]
+SYM(__ld_r13_to_r17):
+ ld r17, [sp,16]
+SYM(__ld_r13_to_r16):
+ ld r16, [sp,12]
+SYM(__ld_r13_to_r15):
+#ifdef __ARC700__
+ ld r15, [sp,8] ; minimum function size to avoid stall: 6 bytes.
+#else
+ ld_s r15, [sp,8]
+#endif
+ ld_s r14, [sp,4]
+ j_s.d [%blink]
+ ld_s r13, [sp,0]
+ ENDFUNC(__ld_r13_to_r15)
+ ENDFUNC(__ld_r13_to_r16)
+ ENDFUNC(__ld_r13_to_r17)
+ ENDFUNC(__ld_r13_to_r18)
+ ENDFUNC(__ld_r13_to_r19)
+ ENDFUNC(__ld_r13_to_r20)
+ ENDFUNC(__ld_r13_to_r21)
+ ENDFUNC(__ld_r13_to_r22)
+ ENDFUNC(__ld_r13_to_r23)
+ ENDFUNC(__ld_r13_to_r24)
+ ENDFUNC(__ld_r13_to_r25)
+
+#endif /* L_millicodethunk_ld */
+#ifdef L_millicodethunk_ret
+ .global SYM(__ld_r13_to_r14_ret)
+ .global SYM(__ld_r13_to_r15_ret)
+ .global SYM(__ld_r13_to_r16_ret)
+ .global SYM(__ld_r13_to_r17_ret)
+ .global SYM(__ld_r13_to_r18_ret)
+ .global SYM(__ld_r13_to_r19_ret)
+ .global SYM(__ld_r13_to_r20_ret)
+ .global SYM(__ld_r13_to_r21_ret)
+ .global SYM(__ld_r13_to_r22_ret)
+ .global SYM(__ld_r13_to_r23_ret)
+ .global SYM(__ld_r13_to_r24_ret)
+ .global SYM(__ld_r13_to_r25_ret)
+ HIDDEN_FUNC(__ld_r13_to_r14_ret)
+ HIDDEN_FUNC(__ld_r13_to_r15_ret)
+ HIDDEN_FUNC(__ld_r13_to_r16_ret)
+ HIDDEN_FUNC(__ld_r13_to_r17_ret)
+ HIDDEN_FUNC(__ld_r13_to_r18_ret)
+ HIDDEN_FUNC(__ld_r13_to_r19_ret)
+ HIDDEN_FUNC(__ld_r13_to_r20_ret)
+ HIDDEN_FUNC(__ld_r13_to_r21_ret)
+ HIDDEN_FUNC(__ld_r13_to_r22_ret)
+ HIDDEN_FUNC(__ld_r13_to_r23_ret)
+ HIDDEN_FUNC(__ld_r13_to_r24_ret)
+ HIDDEN_FUNC(__ld_r13_to_r25_ret)
+ .section .text
+ .align 4
+SYM(__ld_r13_to_r25_ret):
+ ld r25, [sp,48]
+SYM(__ld_r13_to_r24_ret):
+ ld r24, [sp,44]
+SYM(__ld_r13_to_r23_ret):
+ ld r23, [sp,40]
+SYM(__ld_r13_to_r22_ret):
+ ld r22, [sp,36]
+SYM(__ld_r13_to_r21_ret):
+ ld r21, [sp,32]
+SYM(__ld_r13_to_r20_ret):
+ ld r20, [sp,28]
+SYM(__ld_r13_to_r19_ret):
+ ld r19, [sp,24]
+SYM(__ld_r13_to_r18_ret):
+ ld r18, [sp,20]
+SYM(__ld_r13_to_r17_ret):
+ ld r17, [sp,16]
+SYM(__ld_r13_to_r16_ret):
+ ld r16, [sp,12]
+SYM(__ld_r13_to_r15_ret):
+ ld r15, [sp,8]
+SYM(__ld_r13_to_r14_ret):
+ ld blink,[sp,r12]
+ ld_s r14, [sp,4]
+ ld.ab r13, [sp,r12]
+ j_s.d [%blink]
+ add_s sp,sp,4
+ ENDFUNC(__ld_r13_to_r14_ret)
+ ENDFUNC(__ld_r13_to_r15_ret)
+ ENDFUNC(__ld_r13_to_r16_ret)
+ ENDFUNC(__ld_r13_to_r17_ret)
+ ENDFUNC(__ld_r13_to_r18_ret)
+ ENDFUNC(__ld_r13_to_r19_ret)
+ ENDFUNC(__ld_r13_to_r20_ret)
+ ENDFUNC(__ld_r13_to_r21_ret)
+ ENDFUNC(__ld_r13_to_r22_ret)
+ ENDFUNC(__ld_r13_to_r23_ret)
+ ENDFUNC(__ld_r13_to_r24_ret)
+ ENDFUNC(__ld_r13_to_r25_ret)
+
+#endif /* L_millicodethunk_ret */
+
+#ifdef L_adddf3
+#ifdef __ARC_NORM__
+#include "ieee-754/adddf3.S"
+#endif
+#endif
+
+#ifdef L_muldf3
+#ifdef __ARC700__
+#include "ieee-754/muldf3.S"
+#elif defined (__ARC_NORM__) && defined(__ARC_MUL64__)
+#include "ieee-754/arc600-mul64/muldf3.S"
+#elif defined (__ARC_NORM__) && defined(__ARC_MUL32BY16__)
+#include "ieee-754/arc600-dsp/muldf3.S"
+#endif
+#endif
+
+#ifdef L_addsf3
+#ifdef __ARC_NORM__
+#include "ieee-754/addsf3.S"
+#endif
+#endif
+
+#ifdef L_mulsf3
+#ifdef __ARC700__
+#include "ieee-754/mulsf3.S"
+#elif defined (__ARC_NORM__) && defined(__ARC_MUL64__)
+#include "ieee-754/arc600-mul64/mulsf3.S"
+#elif defined (__ARC_NORM__) && defined(__ARC_MUL32BY16__)
+#include "ieee-754/arc600-dsp/mulsf3.S"
+#elif defined (__ARC_NORM__)
+#include "ieee-754/arc600/mulsf3.S"
+#endif
+#endif
+
+#ifdef L_divdf3
+#ifdef __ARC700__
+#include "ieee-754/divdf3.S"
+#elif defined (__ARC_NORM__) && defined(__ARC_MUL64__)
+#include "ieee-754/arc600-mul64/divdf3.S"
+#elif defined (__ARC_NORM__) && defined(__ARC_MUL32BY16__)
+#include "ieee-754/arc600-dsp/divdf3.S"
+#endif
+#endif
+
+#ifdef L_divsf3
+#ifdef __ARC700__
+#include "ieee-754/divsf3-stdmul.S"
+#elif defined (__ARC_NORM__) && defined(__ARC_MUL64__)
+#include "ieee-754/arc600-mul64/divsf3.S"
+#elif defined (__ARC_NORM__) && defined(__ARC_MUL32BY16__)
+#include "ieee-754/arc600-dsp/divsf3.S"
+#elif defined (__ARC_NORM__)
+#include "ieee-754/arc600/divsf3.S"
+#endif
+#endif
+
+#ifdef L_extendsfdf2
+#ifdef __ARC_NORM__
+#include "ieee-754/extendsfdf2.S"
+#endif
+#endif
+
+#ifdef L_truncdfsf2
+#ifdef __ARC_NORM__
+#include "ieee-754/truncdfsf2.S"
+#endif
+#endif
+
+#ifdef L_floatsidf
+#ifdef __ARC_NORM__
+#include "ieee-754/floatsidf.S"
+#endif
+#endif
+
+#ifdef L_floatsisf
+#ifdef __ARC_NORM__
+#include "ieee-754/floatsisf.S"
+#endif
+#endif
+
+#ifdef L_floatunsidf
+#ifdef __ARC_NORM__
+#include "ieee-754/floatunsidf.S"
+#endif
+#endif
+
+#ifdef L_fixdfsi
+#ifdef __ARC_NORM__
+#include "ieee-754/fixdfsi.S"
+#endif
+#endif
+
+#ifdef L_fixsfsi
+#ifdef __ARC_NORM__
+#include "ieee-754/fixsfsi.S"
+#endif
+#endif
+
+#ifdef L_fixunsdfsi
+#ifdef __ARC_NORM__
+#include "ieee-754/fixunsdfsi.S"
+#endif
+#endif
+
+#ifdef L_eqdf2
+#ifdef __ARC_NORM__
+#include "ieee-754/eqdf2.S"
+#endif
+#endif
+
+#ifdef L_eqsf2
+#ifdef __ARC_NORM__
+#include "ieee-754/eqsf2.S"
+#endif
+#endif
+
+#ifdef L_gtdf2
+#ifdef __ARC_NORM__
+#include "ieee-754/gtdf2.S"
+#endif
+#endif
+
+#ifdef L_gtsf2
+#ifdef __ARC_NORM__
+#include "ieee-754/gtsf2.S"
+#endif
+#endif
+
+#ifdef L_gedf2
+#ifdef __ARC_NORM__
+#include "ieee-754/gedf2.S"
+#endif
+#endif
+
+#ifdef L_gesf2
+#ifdef __ARC_NORM__
+#include "ieee-754/gesf2.S"
+#endif
+#endif
+
+#ifdef L_uneqdf2
+#ifdef __ARC_NORM__
+#include "ieee-754/uneqdf2.S"
+#endif
+#endif
+
+#ifdef L_uneqsf2
+#ifdef __ARC_NORM__
+#include "ieee-754/uneqsf2.S"
+#endif
+#endif
+
+#ifdef L_orddf2
+#ifdef __ARC_NORM__
+#include "ieee-754/orddf2.S"
+#endif
+#endif
+
+#ifdef L_ordsf2
+#ifdef __ARC_NORM__
+#include "ieee-754/ordsf2.S"
+#endif
+#endif
--- /dev/null
+# Exclude libgcc.so symbols for the Synopsys DesignWare ARC CPU.
+
+# Copyright (C) 2007-2012 Free Software Foundation, Inc.
+# Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+# on behalf of Synopsys Inc.
+
+# This file is part of GCC.
+
+# GCC is free software; you can redistribute it and/or modify it under
+# the terms of the GNU General Public License as published by the Free
+# Software Foundation; either version 3, or (at your option) any later
+# version.
+
+# GCC 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 General Public License
+# for more details.
+
+# Under Section 7 of GPL version 3, you are granted additional
+# permissions described in the GCC Runtime Library Exception, version
+# 3.1, as published by the Free Software Foundation.
+
+# You should have received a copy of the GNU General Public License and
+# a copy of the GCC Runtime Library Exception along with this program;
+# see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+# <http://www.gnu.org/licenses/>. */
+
+# Exclude various symbols which should not be visible in libgcc.so for ARC.
+# Floating point comparisons use a special lightweight ABI which is not
+# compatible with calls via a plt. Moreover, the code is so compact that
+# it is better to include a separate copy in each dso.
+%exclude {
+ __eqsf2
+ __eqdf2
+ __gtsf2
+ __gtdf2
+ __gesf2
+ __gedf2
+ __uneqsf2
+ __uneqdf2
+ __ordsf2
+ __orddf2
+}
--- /dev/null
+# GCC Makefile fragment for Synopsys DesignWare ARC
+
+# Copyright (C) 2007-2013 Free Software Foundation, Inc.
+# Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+# on behalf of Synopsys Inc.
+
+# This file is part of GCC.
+
+# GCC is free software; you can redistribute it and/or modify it under the
+# terms of the GNU General Public License as published by the Free Software
+# Foundation; either version 3, or (at your option) any later version.
+
+# GCC 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 General Public License for more
+# details.
+
+# You should have received a copy of the GNU General Public License along
+# with GCC; see the file COPYING3. If not see
+# <http://www.gnu.org/licenses/>.
+
+CROSS_LIBGCC1 = libgcc1-asm.a
+LIB1ASMSRC = arc/lib1funcs.S
+LIB1ASMFUNCS = _mulsi3 _umulsidi3 _umulsi3_highpart \
+ _udivsi3 _divsi3 _umodsi3 _modsi3 \
+ _divmod_tools _clzsi2 \
+ _millicodethunk_st _millicodethunk_ld _millicodethunk_ret \
+ _adddf3 _muldf3 _addsf3 _mulsf3 _divsf3 _divdf3 _truncdfsf2 _extendsfdf2 \
+ _eqdf2 _eqsf2 _gedf2 _gesf2 _gtdf2 _gtsf2 _uneqdf2 _uneqsf2 _ordsf2 _orddf2 \
+ _fixdfsi _fixsfsi _floatsidf _floatsisf _fixunsdfsi _floatunsidf
+
+#LIBGCC2_CFLAGS = -g1 -O2 $(LIBGCC2_INCLUDES) $(GCC_CFLAGS)
+
+# For floating-point emulation, we mostly use hand-coded assembly.
+# We use fp-bit.c for debugging purposes, and some parts of it
+# as a fallback for hardware configurations for which the hand-coded
+# assembly support is incomplete, i.e., where there is no NORM and/or no
+# supported multiply instruction. Using floating point on such a
+# configuration is generally inadvisable, but we got to provide support
+# somehow so that we can run the testsuites.
+# fp-hack.h / dp-hack.h take care of slecting the parts that are needed,
+# and (for debugging) of renaming functions so that they can be
+# used in an asm wrapper.
+
+LIB2ADD = fp-bit.c dp-bit.c
+
+dp-bit.c: $(srcdir)/fp-bit.c
+ echo '#ifndef __big_endian__' > dp-bit.c
+ echo '#define FLOAT_BIT_ORDER_MISMATCH' >> dp-bit.c
+ echo '#endif' >> dp-bit.c
+ echo '#include "fp-bit.h"' >> dp-bit.c
+ echo '#include "config/arc/dp-hack.h"' >> dp-bit.c
+ grep -v 'include.*fp-bit.h' $(srcdir)/fp-bit.c >> dp-bit.c
+
+fp-bit.c: $(srcdir)/fp-bit.c
+ echo '#define FLOAT' > fp-bit.c
+ echo '#ifndef __big_endian__' >> fp-bit.c
+ echo '#define FLOAT_BIT_ORDER_MISMATCH' >> fp-bit.c
+ echo '#endif' >> fp-bit.c
+ echo '#include "config/arc/fp-hack.h"' >> fp-bit.c
+ cat $(srcdir)/fp-bit.c >> fp-bit.c
+
+# .init/.fini section routines
+
+crtg.o: $(srcdir)/config/arc/crtg.S
+ $(crt_compile) -c -x assembler-with-cpp $<
+
+crtgend.o: $(srcdir)/config/arc/crtgend.S
+ $(crt_compile) -c -x assembler-with-cpp $<
+
+mcount.o: $(srcdir)/config/arc/gmon/mcount.c
+ $(gcc_compile) -isystem $(srcdir)/config/arc/gmon -c $< \
+ -fcall-saved-r0 -fcall-saved-r1 -fcall-saved-r2 -fcall-saved-r3 \
+ -fcall-saved-r4 -fcall-saved-r5 -fcall-saved-r6 -fcall-saved-r7 \
+ -fomit-frame-pointer
+
+gmon.o: $(srcdir)/config/arc/gmon/gmon.c
+ $(gcc_compile) -isystem $(srcdir)/config/arc/gmon -mno-sdata -c $< \
+ -fno-strict-aliasing \
+ -Wno-extra # suppress inane warning about missing initializer.
+ # Adding initializers for the remaining elements of gmonparam would
+ # make the code more brittle.
+
+prof-freq-stub.o: $(srcdir)/config/arc/gmon/prof-freq-stub.S
+ $(gcc_compile) -isystem $(srcdir)/config/arc/gmon -c $<
+
+prof-freq.o: $(srcdir)/config/arc/gmon/prof-freq.c
+ $(gcc_compile) -isystem $(srcdir)/config/arc/gmon -c $<
+
+dcache_linesz.o: $(srcdir)/config/arc/gmon/dcache_linesz.S
+ $(gcc_compile) -isystem $(srcdir)/config/arc/gmon -c $<
+
+profil.o: $(srcdir)/config/arc/gmon/profil.S
+ $(gcc_compile) -isystem $(srcdir)/config/arc/gmon -c $<
+
+profil-uclibc.o: $(srcdir)/config/arc/gmon/profil-uclibc.c
+ $(gcc_compile) -isystem $(srcdir)/config/arc/gmon -c $<
+
+libgmon.a: mcount.o gmon.o dcache_linesz.o $(PROFILE_OSDEP)
+ $(AR_CREATE_FOR_TARGET) $@ $^
--- /dev/null
+# GCC Makefile fragment for the Synopsys DesignWare ARC CPU with newlib.
+
+# Copyright (C) 2007-2012 Free Software Foundation, Inc.
+# Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+# on behalf of Synopsys Inc.
+
+# This file is part of GCC.
+
+# GCC is free software; you can redistribute it and/or modify it under the
+# terms of the GNU General Public License as published by the Free Software
+# Foundation; either version 3, or (at your option) any later version.
+
+# GCC 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 General Public License for more
+# details.
+
+# You should have received a copy of the GNU General Public License along
+# with GCC; see the file COPYING3. If not see
+# <http://www.gnu.org/licenses/>.
+
+PROFILE_OSDEP = prof-freq-stub.o profil.o
--- /dev/null
+# GCC Makefile fragment for the Synopsys DesignWare ARC700 CPU with uClibc.
+
+# Copyright (C) 2007-2012 Free Software Foundation, Inc.
+# Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
+# on behalf of Synopsys Inc.
+
+# This file is part of GCC.
+
+# GCC is free software; you can redistribute it and/or modify it under
+# the terms of the GNU General Public License as published by the Free
+# Software Foundation; either version 3, or (at your option) any later
+# version.
+
+# GCC 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 General Public License
+# for more details.
+
+# Under Section 7 of GPL version 3, you are granted additional
+# permissions described in the GCC Runtime Library Exception, version
+# 3.1, as published by the Free Software Foundation.
+
+# You should have received a copy of the GNU General Public License and
+# a copy of the GCC Runtime Library Exception along with this program;
+# see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
+# <http://www.gnu.org/licenses/>. */
+
+CRTSTUFF_T_CFLAGS += -mno-sdata
+
+# Compile crtbeginS.o and crtendS.o with pic.
+CRTSTUFF_T_CFLAGS_S = $(CRTSTUFF_T_CFLAGS) -mA7 -fPIC
+
+# Compile libgcc2.a with pic.
+TARGET_LIBGCC2_CFLAGS = -mA7 -fPIC
+
+PROFILE_OSDEP = prof-freq.o
+
+# Override t-slibgcc-elf-ver to hide some lib1func
+# routines which should not be called via PLT.
+SHLIB_MAPFILES = libgcc-std.ver $(srcdir)/config/arc/libgcc-excl.ver
"rIJ" ((USItype) (bh)), \
"r" ((USItype) (al)), \
"rIJ" ((USItype) (bl)))
-/* Call libgcc routine. */
-#define umul_ppmm(w1, w0, u, v) \
-do { \
- DWunion __w; \
- __w.ll = __umulsidi3 (u, v); \
- w1 = __w.s.high; \
- w0 = __w.s.low; \
-} while (0)
-#define __umulsidi3 __umulsidi3
-UDItype __umulsidi3 (USItype, USItype);
+
+#define __umulsidi3(u,v) ((UDItype)(USItype)u*(USItype)v)
+#ifdef __ARC_NORM__
+#define count_leading_zeros(count, x) \
+ do \
+ { \
+ SItype c_; \
+ \
+ __asm__ ("norm.f\t%0,%1\n\tmov.mi\t%0,-1" : "=r" (c_) : "r" (x) : "cc");\
+ (count) = c_ + 1; \
+ } \
+ while (0)
+#define COUNT_LEADING_ZEROS_0 32
+#endif
#endif
#if defined (__arm__) && (defined (__thumb2__) || !defined (__thumb__)) \
projects / thirdparty / gcc.git / commitdiff
