* libguile/vm-i-system.c (long-local-ref, long-local-set)
(make-variable): New intructions, for handling nlocs >= 256.
* module/language/glil/compile-assembly.scm (glil->assembly): Compile
<glil-lexical> with support for nlocs >= 256.
* libguile/objcodes.h (struct scm_objcode): Remove the "unused" field --
the old "nexts" -- and expand nlocs to 16 bits.
* module/language/assembly/compile-bytecode.scm (write-bytecode): Write
the nlocs as a uint16.
* module/language/assembly/decompile-bytecode.scm (decode-load-program):
Decompile 16-bit nlocs. It seems this decompilation is little-endian
:-/
* test-suite/tests/asm-to-bytecode.test: Fix up to understand nlocs as a
little-endian value. The test does the right thing regarding
endianness.
* libguile/programs.h:
* libguile/programs.c: (SCM_PROGRAM_FREE_VARIABLES): Rename from
SCM_PROGRAM_FREE_VARS. Callers changed.
* libguile/programs.c (scm_make_program): Rename arg to
"free_variables".
(scm_program_free_variables): Rename from program-free-vars.
* libguile/vm-engine.h:
* libguile/vm-engine.c (VM_CHECK_FREE_VARIABLES): Rename from
VM_CHECK_CLOSURE.
(vm_engine, CACHE_PROGRAM): Rename closure and closure_count to free_vars and
free_vars_vount.
* libguile/vm-i-system.c (FREE_VARIABLE_REF): Rename from CLOSURE_REF.
(free-ref, free-boxed-ref, free-boxed-set): Rename from closure-ref,
closure-boxed-ref, closure-boxed-set.
(make-closure): Renamed from make-closure2.
* module/language/glil/compile-assembly.scm (glil->assembly): Hack to
never write out the the old "make-closure" instruction. Will fix
better later. Change to emit free-ref etc instead of closure-ref.
* module/language/tree-il/compile-glil.scm (flatten): Emit make-closure
instead of make-closure2, now that the old make-closure is gone.
* module/system/vm/program.scm (system): Rename program-free-vars to
program-free-variables.
* test-suite/tests/tree-il.test ("lambda"): Update for make-closure.
* libguile/frames.c (scm_frame_external_link): Removed.
* libguile/frames.h: No need to have the "external link" in the stack
frame -- update macros to take the new situation into account.
* libguile/objcodes.h (struct scm_objcode): Rename the nexts field to
"unused". In the future we can use it for nlocs, I think.
(SCM_OBJCODE_NEXTS): removed.
* libguile/programs.h:
* libguile/programs.c (scm_make_program): Expect the third argument to
be a vector of free variables, not a list of free variables.
SCM_BOOL_F indicates no free variables, not SCM_EOL.
(program_mark): Adapt.
(scm_program_arity): No more nexts.
(scm_program_free_vars): Replaces scm_program_externals.
* libguile/vm-engine.c (VM_CHECK_EXTERNAL)
(vm_engine): No need for the "external" var.
* libguile/vm-engine.h (CACHE_PROGRAM): Update for SCM_PROGRAM_FREE_VARS
instead of SCM_PROGRAM_EXTERNALS.
(NEW_FRAME): Update for new frame size, and no need to cons up
externals. Yay :)
* libguile/vm-i-loader.c (load-program): Update for scm_make_program.
* libguile/vm-i-system.c (external-ref, external-set): No more.
(make-closure): No more.
(goto/args): No need to re-cons externals here. Update for new stack
frame size.
(mv-call, return, return/values): Update for new frame size. No need
to reinstate externals on return.
* libguile/vm.c (really_make_boot_program, scm_load_compiled_with_vm):
Update for scm_make_program.
* module/language/objcode/spec.scm (objcode-env-externals): Treat '() as
#f, for the externals. Need to clean this up later...
* module/system/vm/program.scm (arity:nexts): Remove. Rename
program-external to program-free-vars.
* libguile/vm-i-system.c (box, empty-box): Boxing values and storing
them in local variables.
(local-boxed-ref, local-boxed-set): A combination of local-ref then
variable-ref/set.
(make-closure2, closure-ref, closure-boxed-ref, closure-boxed-set):
New ops. The idea is to migrate Guile over to using flat dispay
closures. See the paper "Three Implementation Models for Scheme" by
Kent Dybvig for more details; this is the "stack-based" model.
* libguile/vm-engine.c:
* libguile/vm-engine.h: Add the necessary infrastructure to keep track
of a "closure" variable, like our "externals" in semantics, but
minimal, flat, and O(1) in implementation.
* libguile/objcodes.c (OBJCODE_COOKIE): Bump.
* libguile/vm-i-loader.c:
* libguile/vm-i-scheme.c:
* libguile/vm-i-system.c: Renumber instructions, so I can have a bit
more space to work.
* libguile/vm.c (struct t_32bit_aligned): New.
(really_make_boot_program)[bytes]: Use it. This fixes possibly
unaligned accesses, which cause a "bus error" on some platforms (e.g.,
sparc-*).
* libguile/objcodes.c: Fix `sizeof (OBJCODE_COOKIE)' assertion: the
trailing 0 must not be taken into account, and multiple of 8 means the
3 LSBs are clear.
* libguile/objcodes.c (OBJCODE_ENDIANNESS, _OBJCODE_STRINGIFY,
OBJCODE_STRINGIFY, OBJCODE_WORD_SIZE): New macros.
(OBJCODE_COOKIE): Use them. The intent is that `.go' files compiled
for a different endianness or word size are detected.
* libguile/gen-scmconfig.c (main): Don't emit typedefs for `long_long'
and `ulong_long'. This was already deprecated in 1.8 and known to
cause conflicts with other libraries such as HDF5, as reported by Mark
Patterson <mpatterson@physics.queensu.ca>
(http://lists.gnu.org/archive/html/bug-guile/2009-02/msg00003.html).
* libguile/instructions.c: In loops, replace scm_op_last with
SCM_VM_NUM_INSTRUCTIONS.
(fetch_instruction_table): Protect the instruction symbols from
collection. Before they were only marked by the name->opcode hash
table, leading to races in which they could be collected.
(scm_lookup_instruction_by_name): Protect the hash table earlier, as
it's not actually a stack variable, since it's static.
* libguile/vm-i-scheme.c (BV_FIXABLE_INT_REF, BV_INT_REF):
(BV_FLOAT_REF, BV_FIXABLE_INT_SET, BV_INT_SET, BV_FLOAT_SET): Fix the
bounds check for the last element.
* libguile/vm-i-scheme.c (vector_ref, vector_set, BV_FIXABLE_INT_REF,
BV_INT_REF, BV_FLOAT_REF, BV_FIXABLE_INT_SET, BV_INT_SET,
BV_FLOAT_SET): Explicitly initialize all locals, to make some versions
of GCC happier. Patch by Dale P. Smith <dsmich@roadrunner.com>.
Thanks to Bill Schottstaedt for reporting this problem!
* libguile/numbers.c (mem2ureal): Don't be misled by *p_exactness
being INEXACT on entry (as is possible when reading a complex
number): use local exactness variable x which starts as EXACT.
Call mem2decimal_from_point () with &x instead of p_exactness.
* test-suite/tests/numbers.test ("string->number"): Add complex number
tests suggested by Bill.
* libguile/instructions.c (scm_instruction_list): Fix a longstanding bug
in this humble function.
* libguile/vm-i-scheme.c (BV_FIXABLE_INT_SET, BV_INT_SET, BV_FLOAT_SET):
Fix some bugs in these macros -- now the bytevector ops work.
* module/language/tree-il/compile-glil.scm (*primcall-ops*): Compile
bytevector calls to VM ops.
* module/language/tree-il/primitives.scm
(*interesting-primitive-names*): Resolve bytevector calls to primitive
calls.
* module/language/tree-il/primitives.scm
(*interesting-primitive-names*): Resolve vector-ref and vector-set!.
* module/language/tree-il/compile-glil.scm (*primcall-ops*): And compile
vector-ref and vector-set! to their opcodes.
* libguile/vm-i-scheme.c (vector-ref, vector-set): New opcodes, placed
before the bytevector ops. The renumbering shouldn't affect anyone,
given that the bytevector ops were not yet used. Fix a few bugs in the
bytevector ops.
* libguile/instructions.h (SCM_VM_NUM_INSTRUCTIONS): Enlarge to 255. Not
sure what performance effects this will have.
* libguile/vm-engine.c: Add new error case, vm_error_not_a_bytevector.
* libguile/vm-engine.h: Don't assign specific registers for i386. Having
added the new VM vector ops, GCC 4.4 is erroring for me now.
* libguile/vm-i-scheme.c: Add bytevector-specific ops to the VM.
We don't actually use them yet, though.
* libguile/bytevectors.h (scm_i_native_endianness): Allow the VM to use
scm_i_native_endianness, but still keep it marked as internal.
* libguile/bytevectors.c: Adjust to use scm_i_native_endianness instead
of native_endianness. Define it at bootstrap time.
* libguile/gen-scmconfig.c (main): Produce a definition for
`scm_t_off'.
* libguile/ports.h (scm_t_port)[read_buf_size, saved_read_buf_size,
write_buf_size, seek, truncate]: Use `scm_t_off' instead of `off_t' so
that the layout and size of the structure does not depend on the
application's `_FILE_OFFSET_BITS' value. Reported by Bill
Schottstaedt, see
http://lists.gnu.org/archive/html/bug-guile/2009-06/msg00018.html.
(scm_set_port_seek, scm_set_port_truncate): Update.
* libguile/ports.c (scm_set_port_seek, scm_set_port_truncate): Use
`scm_t_off' and `off_t_or_off64_t'.
* libguile/fports.c (fport_seek, fport_truncate): Use `scm_t_off'
instead of `off_t'.
* libguile/r6rs-ports.c (bip_seek, cbp_seek, bop_seek): Use `scm_t_off'
instead of `off_t'.
* libguile/rw.c (scm_write_string_partial): Likewise.
* libguile/strports.c (st_resize_port, st_seek, st_truncate): Likewise.
* doc/ref/api-io.texi (Port Implementation): Update prototype of
`scm_set_port_seek ()' and `scm_set_port_truncate ()'.
* NEWS: Update.
* libguile/objcodes.c (OBJCODE_COOKIE): Bump the objcode cookie. We'll
be doing this on incompatible changes until 2.0.
* libguile/vm-i-scheme.c (set_car, set_cdr, slot_set): These
instructions don't have natural return values -- so declare them that
way, that they push 0 values.
* module/language/tree-il/compile-glil.scm (flatten): When compiling
primitive calls, check `(instruction-pushes op)' to see how many
values that instruction will push, and do something appropriate,
instead of just assuming that all primcall ops push 1 value.
As the Autoconf documentation says, "These days portable programs
[...] should not rely on `HAVE_RESTARTABLE_SYSCALLS', since nowadays
whether a system call is restartable is a dynamic issue, not a
configuration-time issue."
In other words, if we ever rely on HAVE_RESTARTABLE_SYSCALLS, we are
at the mercy of any code that Guile happens to be linked with, because
that code could install a signal handler without the SA_RESTART flag,
and then a Guile system call could unexpectedly return EINTR.
The readline part of this goes back to this problem report:
http://sources.redhat.com/ml/guile/2000-05/msg00177.html; and is an
excellent example of the above paragraph. It was noted during the
discussion that undefining HAVE_RESTARTABLE_SYSCALLS would fix the
problem, but that solution wasn't adopted - I guess because Guile was
still using cooperative threads then (not pthreads) and so there was a
significant concern (whether founded or not) that not using
restartable syscalls (where available) could lead to a loss of
performance.
Now Guile's default mode of operation is with pthreads, where we
already don't assume that HAVE_RESTARTABLE_SYSCALLS is reliable, so
there is no possible further performance loss. And in any case we
really have no choice, if we want correct operation.
Thanks to Sylvain Beucler for reporting this and suggesting the fix.
* configure.in (AC_SYS_RESTARTABLE_SYSCALLS): Removed.
* doc/ref/posix.texi (Signals): Remove statement that Guile always
sets SA_RESTART flag.
* guile-readline/configure.in (GUILE_SIGWINCH_SA_RESTART_CLEARED):
Remove this setting, together with its test code.
(HAVE_RL_PRE_INPUT_HOOK): Remove this setting and its code, as no
longer needed.
* guile-readline/readline.c (sigwinch_enable_restart): Removed.
(scm_init_readline): Remove setting of rl_pre_input_hook.
* libguile/_scm.h (SCM_SYSCALL): Remove the definition that relies on
HAVE_RESTARTABLE_SYSCALLS.
* libguile/scmsigs.c (scm_sigaction_for_thread): Don't always set the
SA_RESTART flag if available. Update docstring accordingly.
(scm_init_scmsigs): Remove code that sets SA_RESTART flag for all
signals.
* THANKS: Add Sylvain.
* libguile/bytevectors.c (scm_tc16_bytevector, print_bytevector,
bytevector_equal_p, free_bytevector): Don't use the snarfing macros.
(scm_bootstrap_bytevectors): New.
(scm_init_bytevectors): No longer initialize SCM_NULL_BYTEVECTOR,
which is done by `scm_bootstrap_bytevectors ()'.
* libguile/bytevectors.h (scm_bootstrap_bytevectors): New declaration.
(scm_init_bytevectors): Made internal. This can be done because we
explicitly register it with `scm_c_register_extension ()' in
`scm_bootstrap_bytevectors ()'.
* libguile/init.c (scm_i_init_guile): Call `scm_bootstrap_bytevectors ()'.
This is so that expressions like "(generalized-vector-length #vu8())"
work even when `(rnrs bytevector)' hasn't been loaded.
As a side effect, this allows compilation of literal bytevectors
("#vu8(...)"), which gets done by the generic array handling
of the GLIL->assembly compiler.
* doc/ref/api-compound.texi (Generalized Vectors): Mention bytevectors.
(Arrays, Array Syntax): Likewise.
* doc/ref/api-data.texi (Bytevectors as Generalized Vectors): New node.
* libguile/bytevectors.c (scm_i_bytevector_generalized_set_x): New.
* libguile/bytevectors.h (scm_i_bytevector_generalized_set_x): New
declaration.
* libguile/srfi-4.c (scm_i_generalized_vector_type,
scm_array_handle_uniform_element_size,
scm_array_handle_uniform_writable_elements): Add support for
bytevectors.
* libguile/unif.c (type_creator_table): Add `vu8'.
(bytevector_ref, bytevector_set): New functions.
(memoize_ref, memoize_set): Add support for bytevectors.
* libguile/vectors.c (scm_is_generalized_vector,
scm_c_generalized_vector_length, scm_c_generalized_vector_ref,
scm_c_generalized_vector_set_x): Add support for bytevectors.
* test-suite/tests/bytevectors.test ("Generalized Vectors"): New test
set.
* libguile/r6rs-ports.c (cbp_mark): A closed port will have had its
stream destroyed, so don't dereference the stream in that case. Patch by
Mike Gran.
* libguile/filesys.h:
* libguile/filesys.c (scm_canonicalize_path): New function,
canonicalize-path.
* module/system/base/compile.scm (compiled-file-name): Canonicalize the
filename so that compiling e.g. ../foo.scm doesn't compile to
~/.guile-ccache/1.9/../foo.scm.
* libguile/read.c (scm_read_bytevector): New function.
(scm_read_sharp): Add `v' case for bytevectors.
* test-suite/lib.scm (exception:read-error): New variable.
* test-suite/tests/bytevectors.test ("Datum Syntax"): New test set.
* libguile/bytevectors.c (bytevector_equal_p): New function.
* test-suite/tests/bytevectors.test ("2.3 Operations on Bytes and
Octets")["equal?"]: New test.
* libguile/eval.c (scm_debug_opts): Up the default stack limit by a
factor of 4. Psyntax expansions currently bounce back and forth between
the VM and the interpreter, due to `map'. (Hopefully that won't be the
case in the future, when have map in scheme, and we get an inliner.)
Anyway when expanding a big nested expression, as for example in
(language ecmascript compile-ghil) -- the pmatch code ends up being
super-nested -- we can consume loads o stack.
So given that on desktop machines, where rlimit is likely to be unset,
default rlimits are around 8 or 10 MB or so, let's bump up our default
limit to 640KB (on 32-bit). Should be enough for anyone.
See http://thread.gmane.org/gmane.lisp.guile.devel/8599/focus=8662 for
more info. Thanks to Mark H. Weaver for the diagnosis!
* libguile/vm-i-scheme.c (FUNC2): Use a signed value for the intermediate
result here. Not sure what the effect is, though.
* module/ice-9/psyntax.scm (chi-top): Toplevel definitions ensure that
variables are defined in the current module. Fixes the specific case of
guile-lib's md5.scm, which redefines + -- this code is needed so that
we don't incorrectly open-code +.
* module/language/tree-il/primitives.scm (resolve-primitives!): I think
there were some cases in which vars and names would not resolve
properly here. Fix those.
* doc/ref/vm.texi (Loading Instructions): Remove references to
load-integer and load-unsigned-integer -- they're still in the VM but
will be removed at some point.
(Data Control Instructions): Add make-int64 and make-uint64.
* libguile/vm-i-loader.c (load-unsigned-integer): Allow 8-byte values.
But this instruction is on its way out, yo.
* libguile/vm-i-system.c (make-int64, make-uint64): New instructions.
* module/language/assembly.scm (object->assembly): Write out make-int64
and make-uint64 instructions, using bytevectors to do the endianness
conversion.
(assembly->object): And pretty-print them back, for disassembly.
* module/language/glil/compile-assembly.scm: Don't generate load-integer
/ load-unsigned-integer instructions.
* module/rnrs/bytevector.scm (rnrs):
* libguile/bytevectors.h:
* libguile/bytevectors.c (scm_uniform_array_to_bytevector): New function.
* libguile/unif.h:
* libguile/unif.c (scm_from_contiguous_typed_array): New function.
* libguile/vm-i-loader.c (load-array): New instruction, for loading byte
data into uniform vectors. Currently it copies out the data, though in
the future we could avoid that.
* module/language/assembly.scm (align-code): New exported function,
aligns code on some boundary.
(align-program): Use align-code.
* module/language/assembly/compile-bytecode.scm (write-bytecode): Support
the load-array instruction.
* module/language/glil/compile-assembly.scm (dump-object): Dump uniform
arrays. Neat :)
* libguile/vm-i-system.c (long-object-ref, long-toplevel-ref)
(long-toplevel-set): Add new instructions, for accessing the object
table with a 16-bit offset. HTMLprag defines a test program that has
more than 256 constants, necessitating this addition.
* doc/ref/vm.texi: Mention the new instructions.
* module/language/glil/compile-assembly.scm: Emit long refs for object
tables bigger than 256 entries.
* libguile/vm-i-system.c (toplevel-ref, toplevel-set)
* libguile/vm-i-loader.c (link-now):
* libguile/vm.c (resolve_variable): Factor out common code to a static
method. The compiler can still inline it, so it shouldn't have a
significant performance effect.
* libguile/vm-engine.c (vm_error_no_such_module): Remove now-unused
label.
* libguile/load.c (compiled_is_fresh): Rename from compiled_is_newer.
Check that the mtines of the .go and .scm match exactly, so we don't
get fooled by rsync-like modifications of the filesystem.
* module/system/base/compile.scm (call-with-output-file/atomic): Add
optional arg, a reference file. If present we utime the output file to
match the source file, before the rename.
(compile-file): Stamp the .go file with the timestamp of the .scm.
* libguile/load.c (do_try_autocompile): Use module_variable, not
module_lookup, when resolving compile-file, so we get #f instead of an
exception if the compiler is in the process of being loaded.
(scm_primitive_load_path): In what I hope is the last patch to this
code, recheck the fallback path if we found a stale installed .go file.
