if this code is compiled without optimization the not-inlined functions
will be included (statically) in the object files, and now show up
as undefined references.
with the syscall numbers changed.
Since 4.4BSD vfork(2) did not share address space, Ovfork.S could be changed
to be identical to fork.S (modulo syscall numbers). __vfork14.S, however,
needs to remain separate since with a shared address space, the stub needs to
pre-decrement the stack and do an indirect jump to the return address, so that
the shared stackframe is still there when the parent returns. (Obviously,
this does not apply to systems which to not build stackframes on leaf calls.)
undefined". The sigaltstack structure layout didn't actually change on
the sparc, so just change the syscall name. A better fix would probably
be to replace the direct system call with a function call to the libc
sigaltstack().
specified by including a Makefile.inc from the appropriate MD directory.
stdlib doesn't do that, but there's no reason that it shouldn't (and
it'd be nice to eliminate the 'if's from the MI stdlib Makefile.inc).
incorrectly internally on certain divides.
Handle the case where the internal result appears negative. This condition
can occur when dividing MIN_INT by certain values.
Raise an exception on divide by zero.
68040 and 68060 owners.
With ``make M68040=1'' or ``make M68060=1'', we build a libc where the ldexp()
doesn't use excessive amounts of system cpu time.
Without any of those defines, the normal assembler version using FSCALE is
included, for the benefit of M68020 and M68030 users, or M68040/M68060 users
where the ldexp performance is no bottleneck.
[Its not yet worked out how to include both library versions in a binary
distribution, but you can build your own from source now if you need it.]
For details, look at PR 3105.
Examining the .so output from binutils 2.8 and 2.8.1 shows the
li.d $f2, 4503599627370496e0
is (incorrectly?) emitted as GP-relative loads when PIC is in effect.
This code uses subb instead of cmpb to compare chars. When they are
not equal, the function's return value is computed by sign extending
the difference. The basic idea was contributed by Hiroshi Horitomo
<horimoto@cs-aoi.cs.sist.ac.jp> in PR #3223, but his implementation
used additional instructions to compute the return value.
Also, the series of insns used to compare one pair of characters has
been unrolled 4 times. This helps the 680[23]0, where the cost of a
taken branch is significantly more than that of a not-taken branch.
The implementation of these functions is losely based on the original
BSD code plus versions I was working on which would yeild code tuned
for a particular m68k or coldfire family cpu, code space or run time
efficiency, etc. depending on the values of various macros.
My original code was intended for use in OS-less embedded systems.
Because NetBSD can not benefit from the considerable complexity, I've
removed the extraneous preprocessor goop. The resulting code is tuned
for the '020 - '060. It will not function on the '000 or coldfire.
* Remove $Id$ headers.
* Support LIB_SCCS in all files.
* Use local labels where ever possible.
* Use cpufunc.h instead of __asm__ when possible.
* Access labels with _C_LABEL and _ASM_LABEL.
* Always set errno by calling cerror.
* No more FRAME/EMARF.
* No more SVC, use SYSTRAP instead.
* Use "$Masqueraded: as XXX $" to show e.g. memcpy compiled as bcopy.
* Rewrite vfork to work even whith a "real" vfork system call.
* Make cerror take care of _REENTRANT.
faster) byte-displacement form of the instruction if the offset will
fit. This happens to be all occurances in this case, so I could have
used bCCs. However, a quick survey of the rest of NetBSD's m68k code
showed that jCC was used in similar cases. I did the same to be
consistant.
memory regions are "potentially overlapping" to a test that determines
that the regions are actually overlapping. Because the code for the
overlapping case is seven instructions longer, this signifcantly
improves performance in the average case.
for the address of a variable used to store the error number instead
of writing to the global errno.
In a multi-threaded program, __errno() will return a pointer to a
thread-specific variable.
SYSCALL_NOERROR(x): Like SYSCALL except that "x" is a syscall
that can never fail.
RSYSCALL_NOERROR(x): Like RSYSCALL except that "x" is a syscall
that can never fail.
These macros simply call SYSCALL / RSYSCALL, and serve as placeholders
until an optimized implementation is done.
Changed all instances of ENTRY() to match the new calling convention
(it takes a second argument ).
Added new macros:
SYSTRAP(x): Expands to the code used to call syscall x.
This is used to simplify other macros.
SYSCALL_NOERROR(x): Like SYSCALL except that "x" is a syscall
that can never fail.
RSYSCALL_NOERROR(x): Like RSYSCALL except that "x" is a syscall
that can never fail.
SYSTRAP(x): Expands to the code used to call syscall x.
This is used to simplify other macros.
SYSCALL_NOERROR(x): Like SYSCALL except that "x" is a syscall
that can never fail.
RSYSCALL_NOERROR(x): Like RSYSCALL except that "x" is a syscall
that can never fail.
SYSTRAP(x): Expands to the code used to call syscall x.
This is used to simplify other macros.
SYSCALL_NOERROR(x): Like SYSCALL except that "x" is a syscall
that can never fail.
RSYSCALL_NOERROR(x): Like RSYSCALL except that "x" is a syscall
that can never fail.
SYSTRAP(x): Expands to the code used to call syscall x.
This is used to simplify other macros.
SYSCALL_NOERROR(x): Like SYSCALL except that "x" is a syscall
that can never fail.
RSYSCALL_NOERROR(x): Like RSYSCALL except that "x" is a syscall
that can never fail.
rename SYSCALL_NOLABEL as CALLSYS_ERROR (for consistency with CALLSYS_NOERROR
now in asm.h), and implement it using CALLSYS_NOERROR.
reimplement SYSCALL, RSYSCALL, and PSEUDO to use the appropriate macros.
implement SYSCALL_NOERROR and RSYSCALL_NOERROR to do the same thing as
SYSCALL and RSYSCALL, but without the error checking/handling.