| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| ... | |
| |
|
|
|
|
|
|
|
|
|
|
|
| |
* parser.c (listener_auto_compound_s): New symbol variable.
(repl): If *listener-auto-compound-p* is true, then evaluate
multiple forms directly as a compound expression, without
inserting progn at the head.
(parse_init): Initialize symbol variable and register
the *listener-auto-compound-p* special.
* txr.1: Documented.
* stdlib/doc-syms.tl: Updated.
|
| |
|
|
|
|
| |
* parser.c (parse_init): Set the initial value
of *listener-sel-inclusive-p* to t instead of nil,
since terminals tend to have block cursors by default.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
This problem was reported by Void Linux package maintainer
adigitoleo.
* configure: in glibc, _TIME_BITS and _FILE_OFFSET_BITS
are not independent. You cannot make the former 64 if
the latter isn't. I'm rewriting the test to check for
these together, in all combinations. We succeed if
we detect a combination that makes both time_t and
off_t 64 bits. If we don't find such a combination then
we at least enable 64 bit off_t alone or 64 bit time_t
alone, if we encountered an option which does that.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
* Makefile (COMPILE_C_WITH_DEPS, WINDRES): Do not call mkdir.
(win/%.res): Express dependency on existence of win
subdirectory in the build directory. We use an order-only
prerequisite, so that timestamps are not compared.
We don't want to be rebuilding objects if the directory
they are in has a newer timestamp, which is often the case.
(OBJS): Extract the relative directory paths from $(OBJS),
and make each directory the target of a rule which creates it.
Then, make each target in $(OBJS) dependent (order-only) on
the directory into which it will be placed, so before that
object is built, the directory gets created with mkdir -p.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
We get rid of the sed-based processing which produces .v
files from .d files. The only purpose of the variable
assignments in the .v files ended up being a filter
expression in the ABBREVN macro. I think the idea here
was to show all the direct prerequisites of the target,
suppressing the ones computed by generated dependency rules.
* Makefile (ABBREVN): Just use $^ instead filtering
out $^ using the $(DEP_$@) computed variable that holds
all the dependencies. I don't see an issue. This is
only used for linking and it correctly shows the .o
files.
(DEPGEN): Macro removed.
(COMPILE_C_WITH_DEPS): Remove call to DEPGEN, removing
an ugly sed step from the compilation of each file.
(NL, CM, DEP): Macros removed.
(OBJS, EXTRA_OBJS): Directly write rule which makes
all objects depend on config.make.
(opt/lex.yy.o, opt/txr.o, opt/match.o, opt/parser.o,
opt/y.tab.o, dbg/lex.yy.o, dbg/txr.o, dbg/match.o,
dbg/parser.o, dbg/y.tab.o): Explicitly write direct
rules for these so the parser generation is correctly
hooked into the dependency graph.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
* lib.c (unuse_sym): A used symbol may now appear in a package
under a different name. So if we don't find a symbol under
the symbol's name, or find a different symbol, we must try
a reverse hash search before giving up.
* txr.1: Add notes to use-sym-as that unuse-sym must be
used to undo its effect. Add notes to unuse-sym discussing
similarities and differences versus unintern.
* tests/012/use-as.tl: New test cases.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The new function use-sym-as can bring a foreign
symbol into a package under a different name,
which is not that symbol's name. This is also
featured in a new defpackage clause, :use-syms-as.
With this simple relaxation in the package system,
we don't require package local nicknames, which is
more complicated to implement and less ergonomic,
because it doesn't actually vanquish the use of
ugly package prefixes on clashing symbols.
* eval.c (eval_init): Register use-syms-as.
* lib.c (use_sym_as): New function, made out of
use_sym.
(use_sym): Now a wrapper for use_sym_as.
* lib.h (use_sym_as): Declared.
* stdlib/package.tl (defpackage): Implement :use-syms-as
clause.
* tests/012/use-as.tl: New file.
* txr.1: Documented,
* stdlib/doc-syms.tl: Updated.
|
| |
|
|
|
|
|
|
|
| |
* stdlib/compiler.tl (compiler comp-fun-form): Recognize
the pattern (subtypep (typeof x) y) and rewrite it to
(typep x y).
* stdlib/match.tl (compile-struct-match): Don't generate
the (subtype (typeof x) y) pattern, but (typeof x y).
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* stdlib/op.tl (opip-expand): Take arguments which specify
the op and do operators to be inserted. Pass these
through the recursive calls.
(opip, oand): Pass op and do for the new arguments.
(lopip, loand): New macros like opip and oand, but
passing lop and ldo to the expander.
(lflow): New macro.
* autoload.c (op_set_entries): Add autoload entries
for lopip, loand and lflow.
* tests/012/op.tl: A few new tests.
* txr.1: Documented.
* stdlib/doc-syms.tl: Regenerated.
|
| |
|
|
|
| |
* txr.1: Mention that @nil can be used in pattern matching
to suppress unused variable warnings.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
* stream.c (close_stream): If the underlying method
returns the colon symbol :, then keep the cached
close_result as nil, so that the method can be called
again, but return t to the caller to indicate success.
* tests/018/close-delegate.tl: Test case added.
* tests/018/close-delegate.expected: Updated.
* txr.1: Documented.
|
| |
|
|
|
|
|
|
| |
* socket.c (dgram_close): Return t when a
descriptor is closed, returning nil only when
the object is already in a closed state.
* stream.c (dev_null_close, dir_close): Likewise.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is motivated by trying to implement a struct delegate
stream which performs reference counting in close, in
order to close the real stream when the count hits zero.
The caching behavior of close-stream is a problem.
* stream.c (strm_base_init): Initialize close_result to nil,
rather than nao.
(strm_base_mark): Don't check close_result for nao.
(close_stream): Suppress the call to op->close if close_result
has a non-nil value, rather than a value other than nao.
* tests/018/close-delegate.tl,
* tests/018/close-delegate.expected: New files.
* txr.1: Document that only a non-nil return is cached
by close-stream.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
* stdlib/compiler.tl (dump-to-tlo): To ensure numbers are
externalized in such a way that they will be loaded back
exactly, we need to set a few special variables. For integers,
we want *print-base* to be 10. Numbers printed in other bases
cannot be read back correctly. Octal, hex and binary could be,
but they would need to be printed with the correct prefixes.
For floating-point values, we want to switch to the default
print format, and use flo-max-dig for the precision. That one
s not not the default value; the default is flo-dig.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The compilation of stdlib from source code (an initial state
when no .tlo files exist obtained by "make clean" or
"make clean-tlo") is very sensitive to the file order.
If the compiler and optimize modules are compiled late,
it slows down the build greatly.
The compiler mainly executes the code in the files
compiler.tl, param.tl, optimize.tl and asm.tl. These
should be compiled first, and I empirically determined
the best order.
* Makefile (STDLIB_EARLY_PATS): Remove error.tlo. The circular
dependency involving error.tl doesn't seem to be an issue
any more. A while ago I addressed some circular dependencies
in stdlib that prevented it from being usable uncompiled;
make tests didn't pass with a source-only stdlib.
We instead repurpose STDLIB_EARLY_PATS to specify the four
modules that should compile first, in their desired order,
at least if parallel make is not being used.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
* RELNOTES: Updated.
* configure (txr_ver): Bumped version.
* stdlib/ver.tl (lib-version): Bumped.
* txr.1: Bumped version and date.
* txr.vim, tl.vim: Regenerated.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The optimizer eliminates calls to pure library functions when
all their arguments are D-registers. The call is made at
compiled time and its value is inserted into the program
as a constant (in a newly allocated D register).
The bug is that we can't do this for a D register that
is linked to a load-time value, because we don't know its
value until run-time.
* stdlib/optimize.tl (basic-blocks do-peephole-block): Add
a constraint that none of the D registers can be a member
of bb.lt-dregs, which holds the list of D registers that
are used for load-time values.
|
| |
|
|
| |
* txr.1: Fix misspelled "elements" in 289 compatibility notes.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
* stdlib/op.tl (sys:opip-single-let-p,
sys:opip-let-p): New functions.
(sys:opip-expand): Restructure from collect loop
to car/cdr recursive form, because the new let operators
in opip need access to the rest of the pipeline.
Implement let operators.
* tests/012/op.tl: New tests.
* txr.1: Documented.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is a regression due to a March 2016 commit which
introduced the ability for :vars in an output-side @(repeat)
block to have initial values.
The bug has the effect that all arguments in @(repeat)
which are conses/lists get duplicated, which messes up
the property list structure.
* parser.y (expand_repeat_rep_args): Do not unconditionally
add reg to the output at the bottom of the loop. A few
cases above in the consp(arg) case handle that themselves, and
do not continue the loop, so control ends up at the bottom,
adding a spurious item. By removing this list_collect,
we have to introduce it to just one case which relies on it.
* tests/008/repeat.txr,
* tests/008/repeat.expected: New files.
* y.tab.c.shipped: Updated.
|
| |
|
|
|
|
|
|
|
|
|
| |
Discovered while experimenting with new optimizations.
* stdlib/optimize.tl (basic-blocks join-block): When we
join the following block into the current block, we must
propagate the nojoin property of the following block.
The nojoin property has to do with the last instruction
being xend. The joined block has that last instruction
and so must be nojoin.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
Discovered while experimenting with new optimizations.
* stdlib/optimize.tl (basic-blocks :postinit): Pass t argument
to new parameter of basic-blocks link-graph.
(basic-blocks link-graph): New parameter indicating whether
this is the first call; if false, we reset all the links.
(basic-blocks elim-dead-code): This no longer has to reset
the links before calling link-graph. But now calls link-graph
one more time after the dead code removal so that no dead
blocks appear in the graph.
|
| |
|
|
|
|
|
|
|
| |
* stdlib/place.tl (sys:get-fun-getter-setter): The check added
recently for a non-bindable sym at the top is completely
bogus, and makes it impossible to use trace for methods.
The function handles various kinds of function names that
are not symbols. The check must be done in the fallback
case, where plain symbols are handled.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
* RELNOTES: Updated.
* configure (txr_ver): Bumped version.
* stdlib/ver.tl (lib-version): Bumped.
* txr.1: Bumped version and date.
* txr.vim, tl.vim: Regenerated.
* protsym.c: Regenerated.
|
| |
|
|
|
|
| |
* eval.c (error_trace): Push a new dynamic environment and
bind *print-circle* to t. More could be done here like
setting object limits.
|
| |
|
|
|
|
|
|
| |
* stdlib/optimize.tl (basic-blocks :postinit): Calculate the
basic block partitions more directly using partition-if,
eliminating the calculation of two sequences of indices
that have to be merged and then passed to the partition
function.
|
| |
|
|
|
|
|
| |
* stdlib/place.tl (sys:get-fun-setter-getter):
Throw error if sym isn't a bindable symbol, so
that nonsense like (set (symbol-function 3) ...)
isn't allowed.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
When a pattern variable match like @foo references a global
symbol macro, that's treated as an existing expression to
match, and not a new binding. However, local symbol macros
are not treated this way; they are invisible to variable
patterns. That is an unintended inconsistency.
* stdlib/match.tl (var-list exists): Use lexical-binding-kind
rather than lexical-var-p. This returns true for lexical
symbol macros also.
* tests/011/patmatch.tl: New test cases.
* txr.1: Documentation revised to clarify that both global
and local symbol macros are considered to be existing variable
bindings by pattern matching.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* evalc (macro_k): New keyword symbol variable.
(lexical_binding_kind, lexical_fun_binding_kind)
New functions.
(lexical_var_p): Bugfix: if the symbol is a special
variable, do not short-circuit to a nil answer.
Special variables can be shadowed by symbol macros.
The function is now defined in terms of lexical_binding_kind.
(lexical_symacro_p, lexical_macro_p): New
functions.
(lexical_fun_p): Now defined using lexical_fun_binding_kind.
(lexical_lisp1_binding): Bugfix: check for special
variables; do not report special variables as :var.
(eval_init): Initialize macro_k. Register new intrinsics:
lexical-binding-kind, lexical-fun-binding-kind,
lexical-symacro-p, lexical-macro-p.
* txr.1: Documented.
* stdlib/doc-syms.tl: Updated.
|
| |
|
|
|
|
| |
* tests/011/patmatch.tl: Move the form which compiles the
entire file to the end of the file, so that all the
interpreted test cases complete before we compile.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
When functions are optimized away due to constant folding,
instead of replacing them with a nil, we now compact the
table to close the gaps and renumber the references in the
code.
* stdlib/compiler.tl (compiler null-stab): Method removed.
(compiler compact-dregs): Renamed to compact-dregs-and-syms.
Now compacts the symbol table also. This is combined with
D-reg compacting because it makes just two passes through
the instruction: a pass to identify the used D registers
and symbol indices, and then another pass to edit the
instructions with the renamed D registers and renumbered
symbol indices.
(compiler optimize): Remove the call to the null-unused-data
on the basic-blocks object; nulling out D regs and symbol
table entries is no longer required. Fllow the rename of
compact-dregs to compact-dregs-and-syms which is called
the same way otherwise.
* stdlib/optimize.tl (basic-blocks null-unused-data):
No longer used method removed.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
We now have some constant folding in the optimizer too, not
just in the front end compiler pass. This is leaving behind
dead D registers that are not referenced in the code.
Let's compact the D register table to close the gap.
* stdlib/compiler.tl (compiler get-dreg): In this function
we no longer check that we have allocated too many D
registers. We let the counter blow past %lev-size%.
Because this creates the fighting chance that the compaction
of D regs will reduce their number to %lev-size% or less.
By doing this, we allow code to be compilable that otherwise
would not be: code that allocates too many D regs which
are then optimized away.
(compiler compact-dregs): New function. Does all the work.
(compiler optimize): Compact the D regs at optimization
level 5 or higher.
(compile-toplevel): Check for an overflowing D reg count
here, after optimization.
* stdlib/optimize.tl (basic-blocks null-unused-data):
Here, we no longer have to do anything with the D registers.
|
| |
|
|
|
|
|
|
| |
* stdlib/compiler.tl (compiler get-dreg): Fix
indentation proble.
* stdlib/optimize.tl (basic-block fill-treg-compacting-map):
Likewise.
|
| |
|
|
|
|
|
| |
* hash.c (hash_iter_peek): The loop here must be a top-of-test
while loop, not a bottom-test do loop. In the chained hashing
implementation, this was a do loop, but it also had a test
with a break for the index.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
* stdlib/path-test.tl (path-volume): Don't return
:abs for a path whose empty first component
isn't followed by any more items. Otherwise
we return :abs for a path formed by splitting
the empty string, and then calls like (rel-path "" "a")
complain about a mixture of absolute and relative.
With this change, empty paths given to rel-path
behave as if they were ".".
* tests/018/rel-path.tl: New test cases.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This commit does two things. The replace function, implemented
under the hood by four specializations: replace-list, replace-vec,
replace-str and replace-buf, will handle the index-list case
a little differently. This is needed to fix the ability of the
del macro work on place designated by an index list, such as:
(del [sequence '(1 3 5 6)]
which now deletes elements 1, 3, 5 and 6 from the sequence,
and returns a sequence of those items. The underlying
implementation uses replace with an index-list, which is now
capable of deleting items. Previously, replace would stop
processing the index list when the replacement-sequence
corresponding to the index list ran out of items. Now,
when the replacement-sequence runs out of items, the
remaining index-list sequence elements specify items to
be deleted. For instance if str holds "abcdefg" then:
(set [str '(1 3 5)] "xy")
will change str to "axcyeg". Elements 1 and 3 are replaced
by x and y, respectively. Element 5, the letter f, is
deleted, because the replacement "xy" has no element
corresponding to 5.
* lib.c (replace_list, replace_str, replace_vec): Implement
new deleteion semantics for the case when the replacement
sequence runs out of items.
* buf.c (replace_buf): Likewise.
* tests/010/seq.txr: Some new test cases here for
deletion.
* tests/010/seq.expected: Updated.
* txr.1: Documented new semantics of replace, including
a new restriction that if elements are being deleted,
the indices should be monotonically increasing regardless
of the type of the sequence (not only list).
A value of 289 for the -C option documented, which restores
the previous behavior of replace (breaking deletion by
index-list, unfortunately: you don't always get to
simulate an old version of TXR while using new features.)
|
| |
|
|
|
|
| |
* txr.1: fix misspelling of Arithmetic and algorithm
introduced by recent updates: new math functions,
and documentation of CRC-32 parameters.
|
| |
|
|
|
|
|
| |
* stdlib/optimize.tl (basic-blocks do-peephole-block): Extend
constant-folding case to recognize gapply as well as gcall.
We just have to take care in how we apply apply arguments
to the actual function to get the value.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Let's consider the DIM expression [a . b c]. Without
this change we get:
syms:
0: a
1: b
2: c
code:
0: 98020000 getf t2 0
1: 98030001 getf t3 1
2: 98040002 getf t4 2
3: 1C020002 apply t2 t2 t3 t4
4: 00030002
5: 00000004
6: 10000002 end t2
With this change:
syms:
0: a
1: b
2: c
code:
0: 98030001 getf t3 1
1: 98040002 getf t4 2
2: 24020002 gapply t2 0 t3 t4
3: 00030000
4: 00000004
5: 10000002 end t2
There are 17 hits for this optimization in optimize.tl
alone!
* stdlib/optimize.tl (basic-blocks do-peephole-block):
New pattern here. We recognize an instruction sequence
which begins with a (getf treg idx) and ends in
an (apply dest treg ...), without any instructions in
between accessing or redefining treg. Additionally,
the treg value must not be used after the apply,
unless the apply redefines it. In that case, we
rewrite this pattern to eliminate that initial getf
instruction, and substitute (gapply dest idsx ..)
for the apply.
|
| |
|
|
|
|
| |
* eval.c (op_defsymacro, rt_defsymacro, makunbound,
fmakunbound): Don't call vm_invalidate_binding if there
is no binding for the symbol.
|
| |
|
|
|
| |
* lib.c (obj_print_impl): Check for this case and handle.
The #nil syntax is not readable.
|
| |
|
|
|
|
|
|
| |
* eval.c (op_defsymacro, rt_defsymacro): We must call
vm_invalidate_binding so the VM forgets a cached binding
for this variable.
* tests/019/redef.tl: Test added.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
There was a bug in rt_defun in that it was not calling
vm_invalidate_binding. This mean that compiled functions
were not picking up redefinitions. This bug is fixed now
because rt_defun now calls sethash on the top_fb directly,
which modifies the existing binding cell; it is not
allocating a new cell.
We put in new test cases to confirm the proper redefinition
behaviors.
The proper redefinition behavior exposes an issue in
pattern matching.
* tests/019/redef.tl: New file.
* stdlib/match.tl (transform-quote): This function's compiled
image, when deposited into a .tlo file, becomes incorrect
because (sys:hash-lit) turns into #H() syntax, which reads
back as something else. In other words (sys:hash-lit)
deosn't have print-read consistency and so doesn't
externalize. To fix this right we would need a print mode
which ensures machine readability rather than human
readability, like in Common Lisp. For now, we just break up
the pattern so that it's not a literal match. This bug was
hidden due to theredefinition issue. When match.tl is
being compiled, it defines non-triv-pat-p twice. Due to
redefinitions not kicking in properly, the first definition
of non-triv-pat-p remains in effect for some functions.
When transform-qquote is being expanded, the (sys:hash-lit)
pattern is treated as non-trivial, even though it is
is trivial, and so it is turned into pattern matching code.
The code doesn't contain a (sys:hash-lit) literal and so
the issue doesn't occur.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Like was done with the function and variable top-level
environments, we simplify the macro ones.
* eval.c (func_get_name, lookup_mac, lookup_symac,
lookup_symac_lisp1, op_defsymacro, rt_defsymacro,
rt_defmacro, op_defmacro, reg_mac, reg_symacro):
Adjust to simpler representation where the hash cell
itself is the binding cell, rather than holding a
binding cell in its cdr.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Since their inception, the top_fb and top_fb hashes associated
symbols with bindings cells, causing an extra cons cell to be
allocated for each entry. I don't remember why this is. It
might have been that way so that gethash(top_fb, sym) would
return a cell when the variable exists, or else nil. This was
before we had functions like gethash_e and inhash that return
the hash cell itself. A hash cell is also a cons and can serve
as a binding just fine by itself.Let's make it so.
For now, the macro and symbol macro environments stay the
way they are; I will likely convert them also.
* eval.c (env_fbind, env_vbind, lookup_global_var, lookup_sym_lisp1,
lookup_fun, func_get_name, rt_defv, rt_defun, set_symbol_value,
reg_fun, reg_varl): Update all these functions so they treat the
hash cell from top_vb or top_fb as the binding cell, rather than
putting or expecting the cdr of that cell (i.e the hash value)
to be a binding cell.
* hash.[ch] (gethash_d): New function. Jus gethash_e without the
pesky self argument, that would only be needed for error
reporting if we pass an object that isn't a hash.
* stdlib/place.tl (sys:get-fun-getter-setter, sys:get-vb):
These two functions must be adjusted since they refer to the
top-fb and top-vb variables. sys:get-vb isn't used anywhere;
it continues to exist in order to provide run-time support
to files that were compiled with an buggy version of the
symbol-value place.
|
| |
|
|
| |
* eval.c (rt_defv): Report as sys:rt-defv, not sys:defv.
|
| |
|
|
|
|
| |
* stdlib/optimize.tl (basic-blocks do-peephole-block): The
constant folding case should fire even if some of the
arguments of the call aren't D registers but T0.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
* stdlib/compiler.tl (%effect-free-funs%, %effect-free%,
%functional-funs%, %functional%): Move variables
into stdlib/constfun.tl
* stdlib/constfun.tl %effect-free-funs%, %effect-free%,
%functional-funs%, %functional%): Moved here.
* stdlib/optimize.tl: Use load-for to express dependency
on constfun module; don't depend on the compiler having
loaded it.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The compiler handles trivial constant folding over the
source code, as a source to source transformation.
However, there are more opportunities for constant folding
after data flow optimizations of the VM code.
Early constant folding will not fold, for instance,
(let ((a 2) (b 3)) (* a b))
but we can reduce this to an end instruction that returns
the value of a D register that holds 6. Data flow optimizations
will propagate the D registers for 2 and 3 into the gcall
instruction. We can then recognize that we have a gcall with
nothing but D register operands, calling a constant-foldable
function. We can allocate a new D register to hold the result
of that calculation and just move that D register's value
into the target register of the original gcall.
* stdlib/compiler.tl (compiler get-dreg): When allocating
a new D reg, we must invalidate the datavec slot which is
calculated from the data hash. This didn't matter before,
because until now, get-datavec was called after compilation,
at which point no new D regs will exist. That is changing;
the optimizer can allocate D regs.
(compiler null-dregs, compiler null-stab): New methods.
(compiler optimize): Pass self to constructor for basic-blocks.
basic-blocks now references back to the compiler.
At optimization level 5 or higher, constant folding can
now happen, so we call the new method in the optimizer to
null the unused data. This overwrites unused D registers
and unused parts of the symbol vector with nil.
* stdlib/optimize (basic-blocks): Boa constructor now takes
a new leftmost param, the compiler.
(basic-blocks do-peephole-block): New optimization case:
gcall instruction invoking const-foldable function, with
all arguments being dregs.
(basic-blocks null-unused-data): New method.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* configure: Detect all the new functions, with separate
tests for the unary and binary ones.
* arith.c (cbrt_s, erf_s, erfc_s, exp10_s, exp2_s,
expm1_s, gamma_s, j0_s, j1_s, lgamma_s, log1p_s, logb_s,
nearbyint_s, rint_s, significand_s, tgamma_s, y0_s, y1_s,
copysign_s, drem_s, fdim_s, fmax_s, fmin_s, hypot_s,
jn_s, ldexp_s, nextafter_s, remainder_s, scalb_s, scalbln_s,
yn_s, r_copysign_s, r_drem_s, r_fdim_s, r_fmax_s, r_fmin_s,
hypot_s, r_jn_s, r_ldexp_s, r_nextafter_s, r_remainder_s,
r_scalb_s, scalbln_s, r_yn_s): New symbol variables.
(not_available): New static function.
(cbrt_wrap, erf_wrap, erfc_wrap, exp10_wrap, exp2_wrap,
expm1_wrap, gamma_wrap, j0_wrap, j1_wrap, lgamma_wrap,
log1p_wrap, logb_wrap, nearbyint_wrap, rint_wrap,
significand_wrap, tgamma_wrap, y0_wrap, y1_wrap,
copysign_wrap, drem_wrap, fdim_wrap, fmax_wrap,
fmin_wrap, hypot_wrap, jn_wrap, ldexp_wrap,
nextafter_wrap, remainder_wrap, scalb_wrap, scalbln_wrap,
yn_wrap): New static functions.
(arith_set_entries, arith_instantiate): New static functions.
(arith_init): Initialize symbols and instantiate functions
via autoload mechanism. In a program that doesn't use the
functions, we suffer only the overhead of interning the symbols.
* lib.h (UNUSED): New macro for GCC unused attribute.
* txr.1: Documented.
* stdlib/doc-syms.tl: Updated.
|
