From ca34d7a7c77fa8dc823fd05439870e5a96610c5a Mon Sep 17 00:00:00 2001 From: Ulrich Drepper Date: Thu, 14 Aug 1997 01:54:13 +0000 Subject: Update. 1997-08-14 03:14 Ulrich Drepper * elf/dl-minimal.c: Don't use relative #include paths. (malloc): Don't try to initialize _dl_pagesize, it already is. (__sigjmp_save, longjmp): Reformat. * sysdeps/generic/dl-sysdep.c: Implement _dl_show_auxv function to show content of auxiliary array. (_dl_sysdep_start): Remember start of auxiliary vector. * sysdeps/mach/hurd/dl-sysdep.c: (_dl_show_auxv): New dummy function. * elf/link.h: Add prototype for _dl_show_auxv. * elf/rtld.c (dl_main): Call _dl_show_auxv if LD_SHOW_AUXV envvar is present. * libio/iofdopen.c: Better prepare for use in libstdc++. * libio/iofflush.c: Likewise. * libio/iofgetpos.c: Likewise. * libio/iofgets.c: Likewise. * libio/iofopen.c: Likewise. * libio/iofputs.c: Likewise. * libio/iofread.c: Likewise. * libio/iofsetpos.c: Likewise. * libio/ioftell.c: Likewise. * libio/iofwrite.c: Likewise. * libio/iogetdelim.c: Likewise. * libio/iogets.c: Likewise. * libio/iopopen.c: Likewise. * libio/ioputs.c: Likewise. * libio/iosetbuffer.c: Likewise. * libio/iosetvbuf.c: Likewise. * libio/ioungetc.c: Likewise. * libio/iovsprintf.c: Likewise. * libio/iovsscanf.c: Likewise. * manual/creature.texi: Mention _XOPEN_SOURCE_EXTENDED in correct place. * manual/math.texi: Add comment to sincos decsription to say it's a GNU extension. Fix description of random function. * posix/fnmatch.h: Recognize Windog by _WINDOWS32 not WIN32. * posix/glob.c: Likewise. Handle stupid system headers on SunOS. Add casts for all __alloca calls. * posix/glob.h: Recognize Windog by _WINDOWS32 not WIN32. Declare additional GLOB_* constants also if _GNU_SOURCE is defined. Patches by Paul D. Smith . * stdio-common/_itoa.c: Don't use relateive #include paths. * sysdeps/i386/fpu/bits/mathinline.h: Enable optimizations even for gcc 2.7. (atan): Loading 1.0 from memory does not pay off. (asinh): Fix typo. 1997-08-13 09:44 Philip Blundell * sysdeps/stub/accept.c: Use socklen_t where necessary. * sysdeps/stub/bind.c: Likewise. * sysdeps/stub/connect.c: Likewise. * sysdeps/stub/getsockname.c: Likewise. * sysdeps/stub/getsockopt.c: Likewise. * sysdeps/stub/recvfrom.c: Likewise. * sysdeps/stub/sendto.c: Likewise. * sysdeps/stub/setsockopt.c: Likewise. 1997-08-11 Andreas Schwab * io/ftw.c (object_compare): Avoid memcmp when comparing scalar types. 1997-08-12 Andreas Schwab * math/libm-test.c (cbrt_test): Add epsilon for long double to `cbrt (-0.001)'. 1997-08-11 Andreas Schwab * sysdeps/stub/bits/string.h: Fix comment. 1997-08-01 Andreas Schwab * elf/Makefile (routines): Remove dl-version. (dl-routines): Add it here instead. This removes dead code from libc.so. 1997-08-12 17:17 Wolfram Gloger * malloc/thread-m.h (thread_atfork, thread_atfork_static): New macros to encapsulate pthread_atfork's functionality. * malloc/malloc.c (ptmalloc_lock_all, ptmalloc_unlock_all): New functions, registered via thread_atfork in the initialization routine or via thread_atfork_static in global scope. This ensures a consistent state of all locks across fork(). 1997-08-11 23:00:32 1997 Mark Kettenis * login/programs/xtmp.c (compare_entry): Partially undo patch made by Klaus Espenlaub. There was no xtmp/utmp typo. Added comment to cclarify this. * login/programs/xtmp.h: Likewise. * posix/getopt.c: Recognize Windog by _WINDOWS32 not WIN32. Patch by Paul D. Smith . 1997-08-12 13:51 Ulrich Drepper * login/getutid.c: Rename to __getutid and make getutid and getutxid weak aliases. Patch by ir. Mark M._Kettenis . 1997-08-11 23:55 Ulrich Drepper * sysdeps/unix/sysv/linux/Makefile [subdir=misc] (sysdep_headers): Add bits/mman.h. 1997-08-11 08:51:21 1997 H.J. Lu * stdlib/fmtmsg.c (fmtmsg): Add static to __libc_once_define macro. * libc.map (__progname_full, __progname): Added. 1997-08-04 08:31 H.J. Lu * sysdeps/unix/sysv/linux/if_index.c (if_nametoindex): Return 0 if SIOGIFINDEX is not defined. (if_indextoname): Initialize "result" to NULL. (if_nametoindex, if_nameindex): Fix ENOSYS setting. 1997-08-11 11:58 Richard Henderson Sparc ISO C 9x exception handling: * sysdeps/sparc/fpu/bits/mathdef.h: New file. * sysdeps/sparc/fpu/fclrexcpt.c: New file. * sysdeps/sparc/fpu/fegetenv.c: New file. * sysdeps/sparc/fpu/fegetround.c: New file. * sysdeps/sparc/fpu/feholdexcpt.c: New file. * sysdeps/sparc/fpu/fesetenv.c: New file. * sysdeps/sparc/fpu/fesetround.c: New file. * sysdeps/sparc/fpu/feupdateenv.c: New file. * sysdeps/sparc/fpu/fgetexcptflg.c: New file. * sysdeps/sparc/fpu/fraiseexcpt.c: New file. * sysdeps/sparc/fpu/fsetexcptflg.c: New file. * sysdeps/sparc/fpu/ftestexcept.c: New file. * sysdeps/sparc/sparc32/fpu/bits/fenv.h: New file. * sysdeps/sparc/sparc64/fpu/bits/fenv.h: New file. 1997-08-11 11:58 Richard Henderson Sparc64 merge: * configure.in: Change up subdirectories, rename sparc8->sparcv8. * csu/Makefile: Nuke need-nopic-initfini bits. The problem was that sparc64 ld did not recognize pic relocs at all. * elf/dl-deps.c: Include . * elf/do-rel.h (elf_dynamic_do_rel): Take reloc address and size directly instead of the DT names to get them from. * elf/dynamic-link.h (_ELF_DYNAMIC_DO_RELOC): New macro distilled from ELF_DYNAMIC_DO_REL{,A}. Define a second version that accounts for the possibility that DT_RELA and DT_JMPPLT overlap, as such is the case on the Sparc. * elf/elf.h: Stub out Elf64_Ver*, and make them the same as the Elf32_Ver* bits. The linker does not even know to differentiate between the word sizes. Someone should examine this and figure out what the proper sizes of things should be for 64-bit hosts before the sparc64 libraries are distributed. * shlib-versions: Rename sparc64 ld.so to not conflict with sparc32 on the same machine. Move old sysdeps/sparc contents to sysdeps/sparc/sparc32, rename sparc8 subdir to sparcv8, move sysdeps/sparc64 content to sysdeps/sparc/sparc64. Same for sysdeps/unix/sysv/linux/sparc. * sysdeps/sparc/Implies: Do ieee754 here. * sysdeps/sparc/configure.in: Move from sparc32/. * sysdeps/sparc/sparc32/Implies: Do wordsize here. * sysdeps/sparc/sparc64/Implies: Likewise. * sysdeps/sparc/sparc32/dl-machine.h (elf_machine_load_address): Use uninitialized .got to find load address instead of AT_BASE, as the later does not exist when running ld.so directly. * sysdeps/sparc/sparc32/fpu/fpu_control.h: Move from ../. * sysdeps/sparc/sparc64/fpu/fpu_control.h: Likewise. * sysdeps/sparc/sparc32/udiv_qrnnd.S: Clean up PIC support. * sysdeps/sparc/sparc64/Makefile (sysdep_routines): Add s_finitel and s_signbitl for stdio. * sysdeps/sparc/sparc64/configure.in: Remove, we no longer need nopic_initfini. * sysdeps/sparc/sparc64/bsd-_setjmp.S: Remove. * sysdeps/sparc/sparc64/bsd-setjmp.S: Remove. * sysdeps/unix/sysv/linux/sparc/sparc64/bsd-_setjmp.S: New empty file. * sysdeps/unix/sysv/linux/sparc/sparc64/bsd-setjmp.S: New empty file. * sysdeps/unix/sysv/linux/sparc/sparc64/longjmp.S: Add __longjmp alias. * sysdeps/unix/sysv/linux/sparc/sparc64/setjmp.S: Add weak _setjmp and setjmp entry points. * sysdeps/sparc/sparc64/dl-machine.h: Finish real implementation. * sysdeps/sparc/sparc64/elf/crtbegin.S: Add copyright, fix PIC. * sysdeps/sparc/sparc64/elf/crtend.S: Likewise. * sysdeps/unix/sysv/linux/sparc/bits/sigaction.h (struct sigaction): Fix sa_flags size for sparc64; add sa_restorer for kernel. * sysdeps/unix/sysv/linux/sparc/sparc64/brk.S: Fix PIC. * sysdeps/unix/sysv/linux/sparc/sparc64/clone.S: Likewise. * sysdeps/unix/sysv/linux/sparc/sparc64/init-first.h: Likewise. * sysdeps/unix/sysv/linux/sparc/sparc64/sysdep.h: Likewise. * sysdeps/unix/sysv/linux/sparc/sparc64/syscall.S: New file. 1997-08-10 18:29:08 1997 Mark Kettenis * login/login.c (login): Check for correct return value of utmpname. 1997-08-11 16:49 Ulrich Drepper * time/tzset.c (__tzset_internal): Correctly handle TZ strings following incorrect old POSIX specs. Patch by Paul Eggert . * sysdeps/generic/bits/sigaction.h: Correct comment for SA_RESTART. * sysdeps/unix/bsd/osf/bits/sigaction.h: Likewise. * sysdeps/unix/sysv/linux/alpha/bits/sigaction.h: Likewise. * sysdeps/unix/sysv/linux/bits/sigaction.h: Likewise. * sysdeps/unix/sysv/linux/mips/bits/sigaction.h: Likewise. * sysdeps/unix/sysv/linux/sparc/bits/sigaction.h: Likewise. * sysdeps/unix/sysv/minix/bits/sigaction.h: Likewise. * sysdeps/unix/sysv/sysv4/bits/sigaction.h: Likewise. Reported by NIIBE Yutaka . 1997-08-09 Andreas Schwab * manual/math.texi: Define the macros @mul and @infinity and use them to format the multiplication and infinity signs. (FP Comparison Functions): Use @code, not @math, for C code examples. * elf/dl-profile.c: Add descriptions and implement reading of --- sysdeps/sparc/sparc32/umul.S | 153 +++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 153 insertions(+) create mode 100644 sysdeps/sparc/sparc32/umul.S (limited to 'sysdeps/sparc/sparc32/umul.S') diff --git a/sysdeps/sparc/sparc32/umul.S b/sysdeps/sparc/sparc32/umul.S new file mode 100644 index 0000000000..7a26c295cb --- /dev/null +++ b/sysdeps/sparc/sparc32/umul.S @@ -0,0 +1,153 @@ +/* + * Unsigned multiply. Returns %o0 * %o1 in %o1%o0 (i.e., %o1 holds the + * upper 32 bits of the 64-bit product). + * + * This code optimizes short (less than 13-bit) multiplies. Short + * multiplies require 25 instruction cycles, and long ones require + * 45 instruction cycles. + * + * On return, overflow has occurred (%o1 is not zero) if and only if + * the Z condition code is clear, allowing, e.g., the following: + * + * call .umul + * nop + * bnz overflow (or tnz) + */ + +#include "DEFS.h" +FUNC(.umul) + or %o0, %o1, %o4 + mov %o0, %y ! multiplier -> Y + andncc %o4, 0xfff, %g0 ! test bits 12..31 of *both* args + be Lmul_shortway ! if zero, can do it the short way + andcc %g0, %g0, %o4 ! zero the partial product and clear N and V + + /* + * Long multiply. 32 steps, followed by a final shift step. + */ + mulscc %o4, %o1, %o4 ! 1 + mulscc %o4, %o1, %o4 ! 2 + mulscc %o4, %o1, %o4 ! 3 + mulscc %o4, %o1, %o4 ! 4 + mulscc %o4, %o1, %o4 ! 5 + mulscc %o4, %o1, %o4 ! 6 + mulscc %o4, %o1, %o4 ! 7 + mulscc %o4, %o1, %o4 ! 8 + mulscc %o4, %o1, %o4 ! 9 + mulscc %o4, %o1, %o4 ! 10 + mulscc %o4, %o1, %o4 ! 11 + mulscc %o4, %o1, %o4 ! 12 + mulscc %o4, %o1, %o4 ! 13 + mulscc %o4, %o1, %o4 ! 14 + mulscc %o4, %o1, %o4 ! 15 + mulscc %o4, %o1, %o4 ! 16 + mulscc %o4, %o1, %o4 ! 17 + mulscc %o4, %o1, %o4 ! 18 + mulscc %o4, %o1, %o4 ! 19 + mulscc %o4, %o1, %o4 ! 20 + mulscc %o4, %o1, %o4 ! 21 + mulscc %o4, %o1, %o4 ! 22 + mulscc %o4, %o1, %o4 ! 23 + mulscc %o4, %o1, %o4 ! 24 + mulscc %o4, %o1, %o4 ! 25 + mulscc %o4, %o1, %o4 ! 26 + mulscc %o4, %o1, %o4 ! 27 + mulscc %o4, %o1, %o4 ! 28 + mulscc %o4, %o1, %o4 ! 29 + mulscc %o4, %o1, %o4 ! 30 + mulscc %o4, %o1, %o4 ! 31 + mulscc %o4, %o1, %o4 ! 32 + mulscc %o4, %g0, %o4 ! final shift + + + /* + * Normally, with the shift-and-add approach, if both numbers are + * positive you get the correct result. With 32-bit two's-complement + * numbers, -x is represented as + * + * x 32 + * ( 2 - ------ ) mod 2 * 2 + * 32 + * 2 + * + * (the `mod 2' subtracts 1 from 1.bbbb). To avoid lots of 2^32s, + * we can treat this as if the radix point were just to the left + * of the sign bit (multiply by 2^32), and get + * + * -x = (2 - x) mod 2 + * + * Then, ignoring the `mod 2's for convenience: + * + * x * y = xy + * -x * y = 2y - xy + * x * -y = 2x - xy + * -x * -y = 4 - 2x - 2y + xy + * + * For signed multiplies, we subtract (x << 32) from the partial + * product to fix this problem for negative multipliers (see mul.s). + * Because of the way the shift into the partial product is calculated + * (N xor V), this term is automatically removed for the multiplicand, + * so we don't have to adjust. + * + * But for unsigned multiplies, the high order bit wasn't a sign bit, + * and the correction is wrong. So for unsigned multiplies where the + * high order bit is one, we end up with xy - (y << 32). To fix it + * we add y << 32. + */ +#if 0 + tst %o1 + bl,a 1f ! if %o1 < 0 (high order bit = 1), + add %o4, %o0, %o4 ! %o4 += %o0 (add y to upper half) +1: rd %y, %o0 ! get lower half of product + retl + addcc %o4, %g0, %o1 ! put upper half in place and set Z for %o1==0 +#else + /* Faster code from tege@sics.se. */ + sra %o1, 31, %o2 ! make mask from sign bit + and %o0, %o2, %o2 ! %o2 = 0 or %o0, depending on sign of %o1 + rd %y, %o0 ! get lower half of product + retl + addcc %o4, %o2, %o1 ! add compensation and put upper half in place +#endif + +Lmul_shortway: + /* + * Short multiply. 12 steps, followed by a final shift step. + * The resulting bits are off by 12 and (32-12) = 20 bit positions, + * but there is no problem with %o0 being negative (unlike above), + * and overflow is impossible (the answer is at most 24 bits long). + */ + mulscc %o4, %o1, %o4 ! 1 + mulscc %o4, %o1, %o4 ! 2 + mulscc %o4, %o1, %o4 ! 3 + mulscc %o4, %o1, %o4 ! 4 + mulscc %o4, %o1, %o4 ! 5 + mulscc %o4, %o1, %o4 ! 6 + mulscc %o4, %o1, %o4 ! 7 + mulscc %o4, %o1, %o4 ! 8 + mulscc %o4, %o1, %o4 ! 9 + mulscc %o4, %o1, %o4 ! 10 + mulscc %o4, %o1, %o4 ! 11 + mulscc %o4, %o1, %o4 ! 12 + mulscc %o4, %g0, %o4 ! final shift + + /* + * %o4 has 20 of the bits that should be in the result; %y has + * the bottom 12 (as %y's top 12). That is: + * + * %o4 %y + * +----------------+----------------+ + * | -12- | -20- | -12- | -20- | + * +------(---------+------)---------+ + * -----result----- + * + * The 12 bits of %o4 left of the `result' area are all zero; + * in fact, all top 20 bits of %o4 are zero. + */ + + rd %y, %o5 + sll %o4, 12, %o0 ! shift middle bits left 12 + srl %o5, 20, %o5 ! shift low bits right 20 + or %o5, %o0, %o0 + retl + addcc %g0, %g0, %o1 ! %o1 = zero, and set Z -- cgit 1.4.1