glibc/sysdeps/sparc/sparc32/dotmul.S
Ulrich Drepper f41c80910d Update.
1997-08-24 12:24  Ulrich Drepper  <drepper@cygnus.com>

	* configure.in (INSTALL): Quote `$'.

	* libc.map: Add __xpg_basename.

	* csu/Makefile (initfini.s): Disable optimization.

	* elf/dl-deps.c: Implement handling of DL_FILTER.

	* elf/dl-load.c (_dl_init_paths): Add error check.

	* intl/finddomain.c (_nl_find_domain): Correct comment.
	* intl/localealias.c: Include <bits/libc-lock.h> not <libc-lock.h>.

	* libio/stdio.h: Make {,v}snprintf available if __USE_BSD.
	Change extern inline functions to work correctly in C++.

	* locale/iso-4217.def: Update for more recent ISO 4217 version.

	* locale/loadlocale.c (_nl_load_locale): Add cast.

	* manual/message.texi: Finish gettext section.

	* posix/getopt_init.c: Don't use relative #include path.
	(__getopt_clean_environment): Change function to take pointer to
	environment as argument.  Optimize generation of test string a bit.
	* sysdeps/unix/sysv/linux/init-first.c: Call __getopt_clean_environment
	with additional argument.

	* poisx/glob.c: Add prototype for next_brace_sub.

	* sysdeps/generic/dl-sysdep.c: Recognize AT_BASE value on auxiliary
	vector.

	* sysdeps/i386/dl-machine.h (elf_machine_load_address): Rewrite
	to not generate relocation entry.  Suggested by Richard Henderson.
	(ELF_MACHINE_BEFORE_RTLD_RELOC): Removed.
	(elf_machine_runtime_setup): Add .aligns.

	* sysdeps/i386/fpu/fraiseexcpt.c: Add volatile to asms.

	* sysdeps/i386/fpu/bits/mathinline.h: Partially undo change of
	1997-08-14 03:14.  gcc 2.7.2* is really broken in some aspects.

	* sysdeps/standalone/i386/i386.h: Clean up asm statements a bit.
	* sysdeps/standalone/i960/i960ca.h: Likewise.

1997-08-22 19:04  Richard Henderson  <rth@cygnus.com>

	* elf/rtld.c (_dl_start): Init _dl_rtld_map.l_opencount due to
	undocumented test addition in _dl_map_object.

	Support ET_EXEC versions of ld.so, for debugging at least:

	* elf/dl-load.c (_dl_map_object): Add_name_to_object could get
	called despite the DT_SONAME != NULL test, segfaulting.  Simplify
	the code here as well.
	* elf/dl-lookup.c (do_lookup): Skip objects with no symtab.
	(_dl_setup_hash): Likewise for hash tables.
	* elf/dl-version.c (_dl_check_map_versions): Likewise for strtabs.
	* elf/rtld.c (_dl_start): Likewise for rpath.
	(_dl_rtld_libname2): New variable.
	(dl_main): Use it to add an soname for ourselves when we don't have
	one of our own.  Base it on the target's .interp.
	(dl_main): Again, skip printing of objects that don't have strtabs.

	Sparc 32 merge:

	* elf/dl-runtime.c (ELF_FIXUP_RETURN_VALUE): Provide default value.
	(fixup): Simplify code.  Use ELF_FIXUP_RETURN_VALUE.
	(profile_fixup): Likewise, though this still needs fixing for
	Sparc32 and PPC.
	* sysdeps/powerpc/dl-machine.h: Transmute ELF_FIXUP_RETURNS_ADDRESS
	to ELF_FIXUP_RETURN_VALUE.

	* sysdeps/sparc/sparc32/dl-machine.h: Implement lazy relocation.
	Fix up _dl_start_user to handle _dl_skip_args properly.
	Use _dl_hwcap to determine if "flush" is available/needed.

	* sysdeps/sparc/configure.in: Remove.  It doesn't actually do
	anything anymore, and what it did do is done somewhere else.
	* sysdeps/sparc/configure: Likewise.

	* sysdeps/sparc/fpu/bits/mathdef.h (FP_ILOGB0, FP_ILOGBNAN): New.

	* sysdeps/sparc/fpu/fraiseexcpt.c: Rearrange for smaller code.

	* sysdeps/sparc/sparc32/Makefile: Fix sparc->sparc/sparc32 bits
	in divrem expansions.

	* sysdeps/unix/sysv/linux/sparc/sparc32/sysdep.h (END, LOC): New
	definitions for assembly syntax differences.

	* sysdeps/sparc/sparc32/__longjmp.S: %g6,%g7 are reserved to the
	"system".  Use %g2,%g3 instead.  Use new local label macro.
	* sysdeps/sparc/sparc32/add_n.S: Use <sysdep.h> and ENTRY, END,
	and LOC for proper assembly headers/footers.
	* sysdeps/sparc/sparc32/addmul_1.S: Likewise.
	* sysdeps/sparc/sparc32/alloca.S: Likewise.
	* sysdeps/sparc/sparc32/dotmul.S: Likewise.
	* sysdeps/sparc/sparc32/lshift.S: Likewise.
	* sysdeps/sparc/sparc32/mul_1.S: Likewise.
	* sysdeps/sparc/sparc32/rshift.S: Likewise.
	* sysdeps/sparc/sparc32/sparcv8/addmul_1.S: Likewise.
	* sysdeps/sparc/sparc32/sparcv8/mul_1.S: Likewise.
	* sysdeps/sparc/sparc32/sparcv8/submul_1.S: Likewise.
	* sysdeps/sparc/sparc32/sparcv8/udiv_qrnnd.S: Likewise.
	* sysdeps/sparc/sparc32/sub_n.S: Likewise.
	* sysdeps/sparc/sparc32/submul_1.S: Likewise.
	* sysdeps/sparc/sparc32/udiv_qrnnd.S: Likewise.
	* sysdeps/sparc/sparc32/umul.S: Likewise.
	* sysdeps/sparc/sparc32/divrem.m4: Likewise.
	* sysdeps/sparc/sparc32/rem.S: Regenerate.
	* sysdeps/sparc/sparc32/sdiv.S: Regenerate.
	* sysdeps/sparc/sparc32/udiv.S: Regenerate.
	* sysdeps/sparc/sparc32/urem.S: Regenerate.

	* sysdeps/sparc/sparc32/sparcv8/dotmul.S: New file.
	* sysdeps/sparc/sparc32/sparcv8/rem.S: New file.
	* sysdeps/sparc/sparc32/sparcv8/sdiv.S: New file.
	* sysdeps/sparc/sparc32/sparcv8/udiv.S: New file.
	* sysdeps/sparc/sparc32/sparcv8/umul.S: New file.
	* sysdeps/sparc/sparc32/sparcv8/urem.S: New file.

	* sysdeps/sparc/sparc32/bsd-_setjmp.S: Dike out.
	* sysdeps/sparc/sparc32/bsd-setjmp.S: Likewise.
	* sysdeps/sparc/sparc32/setjmp.S: Add _setjmp and setjmp entry points.

	* sysdeps/unix/sysv/linux/sparc/sparc32/__sigtrampoline.S:
	Clean up PIC code.

	* sysdeps/sparc/sparc32/elf/start.S: New file, slightly modified
	from the sparc64 version.
	* sysdeps/sparc/sparc32/elf/start.c: Removed.

	* sysdeps/unix/sysv/linux/sparc/sparc32/init-first.h: Rewrite in
	assembly based on the sparc64 version.

	* sysdeps/sparc/sparc32/fpu/bits/fenv.h: Duh.  Use proper syntax
	for manipulating %fsr.
	* sysdeps/sparc/sparc32/fpu/fpu_control.h: Make IEEE conformance
	be the default.

	* elf/elf.h (HWCAP_SPARC_*): New definitions.
	* elf/rtld.c (_dl_hwcap): New variable.
	* sysdeps/generic/dl-sysdep.c (_dl_sysdep_start): Record AT_HWCAP.

	* sysdeps/unix/sysv/linux/sparc/sparc32/getpagesize.c: New file.
	Attempt to get hold of the page size based on what we might have
	been told at startup time in _dl_pagesize.  This will be obsolete
	when I finish the kernel hooks for a proper sysconf(), stay tuned.

	Sparc 64 merge:

	* sysdeps/sparc/sparc64/dl-machine.h (ELF_FIXUP_RETURN_VALUE): New.
	Figure out the right thing to return based on the .plt format.

	* sysdeps/sparc/sparc64/fpu/fpu_control.h: Update comment.

	* sysdeps/unix/sysv/linux/sparc/sparc64/bits/types.h (__dev_t):
	Should have been 64-bits wide.

	* sysdeps/unix/sysv/linux/sparc/sparc64/init-first.h: sll->sllx,
	optimize for branch delay slot usage.

1997-08-22  Andreas Schwab  <schwab@issan.informatik.uni-dortmund.de>

	* csu/Makefile ($(objpfx)crt%.o): Fix a missing *.so -> *.os
	change.

1997-08-20  Andreas Jaeger  <aj@arthur.rhein-neckar.de>

	* math/libm-test.c (identities): Change epsilon.

	* sysdeps/i386/fpu/bits/mathinline.h: Correct arguments to fabs,
	fabsf, fabsl, __fabsl.

	* sysdeps/libm-i387/e_remainderl.S: Pop extra value from FPU stack.
	* sysdeps/libm-ieee754/s_csinhl.c: Include <fenv.h>.
1997-08-24 10:55:18 +00:00

128 lines
3.2 KiB
ArmAsm

/*
* Signed multiply, from Appendix E of the Sparc Version 8
* Architecture Manual.
*/
/*
* 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.
*/
#include <sysdep.h>
ENTRY(.mul)
mov %o0, %y ! multiplier -> Y
andncc %o0, 0xfff, %g0 ! test bits 12..31
be LOC(mul_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
! If %o0 was negative, the result is
! (%o0 * %o1) + (%o1 << 32))
! We fix that here.
#if 0
tst %o0
bge 1f
rd %y, %o0
! %o0 was indeed negative; fix upper 32 bits of result by subtracting
! %o1 (i.e., return %o4 - %o1 in %o1).
retl
sub %o4, %o1, %o1
1:
retl
mov %o4, %o1
#else
/* Faster code adapted from tege@sics.se's code for umul.S. */
sra %o0, 31, %o2 ! make mask from sign bit
and %o1, %o2, %o2 ! %o2 = 0 or %o1, depending on sign of %o0
rd %y, %o0 ! get lower half of product
retl
sub %o4, %o2, %o1 ! subtract compensation
! and put upper half in place
#endif
LOC(mul_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).
*/
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 low part of the
* result; %y has the bottom 12 (as %y's top 12). That is:
*
* %o4 %y
* +----------------+----------------+
* | -12- | -20- | -12- | -20- |
* +------(---------+------)---------+
* --hi-- ----low-part----
*
* The upper 12 bits of %o4 should be sign-extended to form the
* high part of the product (i.e., highpart = %o4 >> 20).
*/
rd %y, %o5
sll %o4, 12, %o0 ! shift middle bits left 12
srl %o5, 20, %o5 ! shift low bits right 20, zero fill at left
or %o5, %o0, %o0 ! construct low part of result
retl
sra %o4, 20, %o1 ! ... and extract high part of result
END(.mul)