Julian Seward [Thu, 2 Dec 2004 16:16:11 +0000 (16:16 +0000)]
x86 guest/host: do SSE comparisons.
To make this easier, add to IRConst a new kind of literal --
a 128-bit literal specified by a 16-bit value. Each bit of the
latter is defines one byte of the 128-bit value, either 0x00 or
0xFF.
Changed message at the top of files, and the startup message, and the
string in valgrind.pc.in, so that they describe Valgrind as a "dynamic
binary instrumentation framework", and don't mention platforms at all.
I had to tweak the regtest filters a bit for this.
Clear a few more AMD64 start-up hurdles:
- implemented VG_(clone)()
- implemented PLATFORM_DO_MMAP()
- implemented VG_(init_thread1state)() [will need to be updated as the
Vex AMD64 guest state is updated]
- implemented OYNK, because it's useful
Julian Seward [Tue, 30 Nov 2004 13:40:29 +0000 (13:40 +0000)]
Loop unroller: do not unroll loops in which the conditional branch is
not Ijk_Boring, as that will cause events that should be passed to the
caller's scheduler, not to be.
64-bit cleanness: fix nasty bug... in the memory allocators, redzones
were being kept as UInts. They were sometimes negated and then used as
indices to arrays, and on 64-bit platforms they get converted
incorrectly (using zero-extension instead of sign extension, I think)
before the indexing happens, giving a totally bogus index.
The fix was to convert all these types to SizeT. It would be nice to be
able to avoid this automatically somehow in the future.
Get AMD64 slightly further before dying:
- implemented VG_(do_syscall)()
- fixed a problem in ume.c with mapping when loading, which was causing
stage2's memory to be trashed
- fixed stage2.lds substitution so stage2 goes in the right address
Overhauled the docs. Removed all the HTML files, put in XML files as
converted by Donna. Hooked it into the build system so they are only
built when specifically asked for, and when doing "make dist".
They're not perfect; in particular, there are the following problems:
- The plain-text FAQ should be built from FAQ.xml, but this is not
currently done. (The text FAQ has been left in for now.)
- The PS/PDF building doesn't work -- it fails with an incomprehensible
error message which I haven't yet deciphered.
Nonetheless, I'm putting it in so others can see it.
Julian Seward [Mon, 29 Nov 2004 19:57:54 +0000 (19:57 +0000)]
Create a new mechanism: "emulation warnings", which is a way for Vex
to report to whatever is using it that it cannot emulate precisely.
Net result is that a bb can exit with the guest state pointer set to
VEX_TRC_EMWARN. In this case, the (mandatory) guest state psuedo-reg
called "guest_EMWARN" holds a value of type VexEmWarn, indicating the
kind of problem encounted.
Use this to warn about approximations in the x87 FPU simulation:
unmasked exceptions not supported, round to +inf/-inf not supported,
precisions other than 80-bit not supported.
Reason being that sometimes you could get incorrect args passed, when
passing 32-bit ints on 64-bit platforms. I also added macros
VG_(do_syscall[123456]) to make life easier, and converted all the
relevant calls.
Fix post-amd64-commit x86 breakage:
- remove warnings from vg_scheduler.c by using (UWord) casts rather than
64-bit-assuming (ULong) casts.
- move deref_Addr() to x86-linux/syscalls.c, where it's now used.
Also got rid of the ancient SIGNAL_SIMULATION flag, which is a remnant of very
early days -- things now only work with signal simulation, so no point in
keeping it around.
Also make addrcheck/tests/fprw work again now that --single-step is
dead.
Added beginnings of an AMD64 port, so lots of new files and directories.
It compiles, but aborts immediately if you try to run it.
I didn't include ldt.c; I'm not sure how the LDT is used on AMD64. It can be
added later if necessary.
While doing this, did some 64-bit cleanness fixes:
- Added necessary intermediate casts to ULong to avoid warnings when converting
ThreadId to void* and vice versa, in vg_scheduler.c.
- Fixed VALGRIND_NON_SIMD_CALL[0123] to use 'long' as the return type.
- Fixed VALGRIND_PRINTF{,BACKTRACE} to use unsigned longs instead of unsigned
ints, as needed.
- Converted some offsets in vg_symtab2.h from "Int" to "OffT".
- Made strlen, strncat, etc, use SizeT instead of 'unsigned int' for the length
parameter.
- Couple of other minor things.
I had to insert some "#ifdef __amd64__" and "#ifndef __amd64__" guards in
places. In particular, in vg_mylibc.c, some of our syscall wrappers aren't
appropriate for AMD64 because the syscall numbering is a bit different in
places. This difference will have to be abstracted out somehow.
Also rewrote the sys_fcntl and sys_fcntl64 wrappers, as required for AMD64.
Also moved the ipc wrapper into x86, since it's not applicable for
AMD64. However, it is applicable (I think) for ARM, so it would be nice
to work out a way to share syscall wrappers between some, but not all,
archs. Hmm. Also now using the real IPC constants rather than magic
numbers in the wrapper.
Other non-AMD64-related fixes:
- ARM: fixed syscall table by accounting for the fact that syscall
numbers don't start at 0, but rather at 0x900000.
- Converted a few places to use ThreadId instead of 'int' or 'Int' for
thread IDs.
- Added both AMD64 and ARM (which I'd forgotten) entries to valgrind.spec.in.
- Tweaked comments in various places.
Julian Seward [Sun, 28 Nov 2004 02:58:11 +0000 (02:58 +0000)]
Rewrite ensure_valgrind() in a slightly less stupid way. This
significantly reduces the number of client requests made by programs
which make intensive use of the pthread_ functions, such as
OpenOffice.
Introduced build variables VG_ARCH_ALL, VG_OS_ALL, and VG_PLATFORM_ALL,
which list all the arches/OSes/platforms supported. These are used by
several newly added DIST_SUBDIRS automake commands, which specify that
although when you are building you only want to build for the current
arch/OS/platform, when you do 'make dist' you want every
arch/OS/platform to get included.
Moved some duplicated macros out of arch/OS/platform-specific parts,
into the core. Renamed them with capital letters at the same time (eg.
arg1-->ARG1, res-->RES) to reduce likelihood of name conflicts, and also
it doesn't hurt to make it clearer that they're macros. The result is a
very big diff, but conceptually it's very simple.
Factored out some stuff duplicated across all archs, to do with syscall
wrappers. The management apologises for the excessive use of macros, but it's
hard to avoid and really does make the repetitive parts of the code (ie. the
parts that are repeated for each arch) much more concise.
Moved the variable VGOFF_(m_eip) from x86, and moved it (and its
computation) into the generic core, because it's needed by the dispatch
loop of all archs. Also renamed it VG_(instr_ptr_offset), which is
clearer. Since it was the last use of the VGOFF_ prefix, I removed that
prefix.
Also cleaned up the ARM stubs slightly, by removing some of the copied,
commented-out x86 code which has since fallen out of date with the real
x86 code and was thus misleading.
Added beginnings of an ARM port, to the point where it compiles. It does not
run, though. There are lots of stubs to be filled in. (The asm ones currently
just have "swi" in them, which seems to cause seg faults.)
Also, some of the macros are decided dubious, especially:
ARCH_* are bogus
SYSCALL_RET is bogus
PLATFORM_SET_SYSCALL_RESULT is bogus
not sure about SET_SYSCALL_RETVAL
FIRST_STACK_FRAME et al -- bogus?
VG_MAX_JUMPS ?
And in stage2.lds, the 0x8048000 is almost certainly wrong
This required some tweakings of the core:
- some of the vki_*.h kernel types were fixed up
- had to disable the AM_PROG_CC_C_O macro in configure.in, because automake
(autoconf?) didn't like it...
- some "#ifdef __x86__" guards were introduced, for nasty x86 things I don't
yet know how to factor out (trampoline page muck, sysinfo page muck).
- fixed a minor stupidity in vg_proxylwp.c.
- moved the ptrace wrapper into the x86-linux part
- had to change the intercept mangling scheme, to use 'J' instead of '$' as the
escape char because GCC didn't like '$'. This is all very dubious, and only
works because none of our intercepted symbols contains a 'J'. To be fixed up
ASAP.
Julian Seward [Fri, 26 Nov 2004 19:17:47 +0000 (19:17 +0000)]
When doing pessimistic lazy propagation through a dirty helper call,
do not consider inputs from those parts of the guest state marked as
read (or modified) which which are declared to be always-defined, and
dually do write outputs to those parts of the guest state written (or
modified) which are declared to be always-defined.
Simplified the way Cachegrind configures the caches; it was really
strange before. Also aborting if any command-line-provided cache
configuration is unacceptable, rather than falling back on defaults;
it's simpler and arguably better than just emitting a warning.
Julian Seward [Fri, 26 Nov 2004 15:04:24 +0000 (15:04 +0000)]
When creating a child, copy the parent's .vex/.vex_shadow into it so
as to give it a sane starting state. This may not be the Right
Answer, but it fixes regtest failure 'susphello' and doesn't break
anything else.
Disable core dumps, and remove the previous attempts at factoring out
the arch-specific bits, which were just plain wrong. This is a problem for a
later day.
Reinstated Helgrind and Cachegrind; at least, they are included in
compilation now. They both don't work, as I made the minimal changes
necessary to get them compiling (eg. both of the TL_(instrument)()
functions were commented out and replaced with a stub).
This means that 'make dist' should more or less work again, which is
necessary for my playing with the XML documentation.
projects / thirdparty / valgrind.git / log
