This commit changes to use <operand> structures to hold the context
needed to visit lexical bindings lazily, in context, instead of eagerly
visiting them for value. This laziness enables inlining of mutually
recursive bindings.
* module/language/tree-il/peval.scm (<var>): Remove comment about copy
propagation having to run build-var-table; things don't work like that
any more.
(build-var-table): Build <var> entries for all variables, even
unreferenced variables.
(alpha-rename): Remove. We will rename bindings on-demand now.
(peval lookup-var): New helper, to fetch the <var> of a gensym.
(peval fresh-gensyms): Fold here, under peval, and in it, handle
updating the store to record a mapping between new names and <var>
entries from the source program.
(peval record-source-expression): Don't call build-var-table on the
new expression, as alpha-renaming happens on-demand now.
(peval prune-bindings): Rewrite to work with mutually-recursive
bindings, while optionally preserving binding order.
(peval extend-env): New helper.
(peval loop): OK, here goes... Remove the `operand' context, as now we
visit operands lazily. Add a `call' context, which does not
copy-propagate lambda expressions, used to residualize a call after
aborting an inlining attempt. Change the `env' to be a mapping of
gensym to <operand>. Instead of looking up the operand's binding then
alpha-renaming it, just rely on the fact that visiting the operand
will rename it if necessary.
If we residualize a lexical, do so with the fresh name from the
environment. If we visit an operand and it doesn't turn out to be
constant, we will never be able to copy it, and so cache that fact in
the operand. If we residualize a binding and we know what the value
should be, record that binding so that prune-bindings won't have to
visit it again. If the operand folds to a constant, cache that too,
to save effort when unrolling loops.
For let, letrec, fix, and lambda-case, instead of visiting the
bindings eagerly for value, simply record the source expressions and
environments in an <operand> and rely on copy-propagation to visit
them later in the right context. In the case of letrec and fix, this
allows mutually-recursive bindings to be inlined.
Refactor folding of "constructors" (which still need renaming) to
avoid visiting operands twice in some contexts.
For applications, if we have to abort, process the procedure in call
context, which allows some folding but avoids copying lambdas. If we
find a recursive procedure, mark intervening counters as recursive
too, to allow for mutual recursion at the top level.
For lambdas, if we are processing for value, record the source
expression so we can detect recursion. This was previously done in
the lexical-ref copy propagator.
* test-suite/tests/tree-il.test ("partial evaluation"): Remove unused
recursive lexicals in a couple of cases. Add a couple test cases for
pruning. Add a few recursive binding cases.
* libguile/read.c (scm_read_r6rs_block_comment):
* test-suite/tests/reader.test ("reading"): Fix reading of #||||#,
originally reported in bug debbugs.gnu.org/9672, by Bruno Haible.
Thanks, Bruno!
This was a pretty big merge involving a fair amount of porting,
especially to peval and its tests. I did not update psyntax-pp.scm,
that comes in the next commit.
Conflicts:
module/ice-9/boot-9.scm
module/ice-9/psyntax-pp.scm
module/language/ecmascript/compile-tree-il.scm
module/language/tree-il.scm
module/language/tree-il/analyze.scm
module/language/tree-il/inline.scm
test-suite/tests/tree-il.test
* module/language/tree-il/optimize.scm (peval): If it's a lambda in the
operator position, inline without a nested counter, as it's not
possible to increase code size.
* module/language/tree-il/optimize.scm (peval): Factor prune-bindings
out of `let' and company. Have it process unreferenced bindings in
effect context instead of always residualizing non-constant
expressions.
* module/language/tree-il/optimize.scm (types-check?): New helper, to
determine if a primcall will apply without throwing an exception.
(peval): constant-expression? returns #f for expressions that don't
types-check?. Effect-free primitives that type-check are void.
* module/language/tree-il/optimize.scm (peval): Fix a duplicate
traversal for constructors in effect or test context. Add support for
eliding make-prompt-tag.
* test-suite/tests/tree-il.test ("partial evaluation"): Update the test
for make-prompt-tag elision.
* module/language/tree-il/optimize.scm (alpha-rename, peval): Handle
<prompt> and <abort>. Attempt to remove the prompt if the tag is
otherwise unreferenced.
* module/language/tree-il/primitives.scm (*primitive-constructors*): Add
make-prompt-tag as a constructor.
* test-suite/tests/tree-il.test ("partial evaluation"): Add a test that
an prompt whose tag is unreferenced is removed.
* module/language/tree-il/optimize.scm (transfer!, make-nested-counter):
(make-recursive-counter, peval): Limit the algorithm's time to be
strictly O(N) by transferring effort and size counters of recursive
inlining attempts from containing counters.
* test-suite/tests/tree-il.test ("partial evaluation"): Update
expectations for the ((lambda (x) (x x)) (lambda (x) (x x))) case, as
the new accounting policy will cause the entire inlining attempt to
abort.
* module/language/tree-il/optimize.scm (peval): Fix calculation of how
many init expressions to drop when inlining lambdas.
* test-suite/tests/tree-il.test ("partial evaluation"): Add tests.
* module/language/tree-il/optimize.scm (peval): The old approach of
optimistically producing constants and then de-constifying them at
their uses was not only cumbersome but incorrect: it both failed to
preserve identity in some cases and failed to retain immutable
constant values. Instead, now we only produce constants if they
really are constant and immutable. The constant folder has to have a
few more algebraic cases to be as effective as it was, to destructure
(car (cons _ _)) appropriately. On the plus side, now constructors
and deconstructors can handle impure cases more generally.
* test-suite/tests/tree-il.test ("partial evaluation"): Add constructor
and destructuring tests. Adapt other tests to new expectations.
* module/language/tree-il/optimize.scm (code-contains-calls?): Remove
this helper, we will deal with recursion when it happens, not after
the fact.
(peval): Add keyword args for various size and effort limits. Instead
of keeping a call stack, keep a chain of <counter> records, each with
an abort continuation. If ever an inlining attempt is taking too
long, measured in terms of number of trips through the main loop, the
counter will abort. Add new contexts, `operator' and `operand'. They
have different default size limits. In the future we should actually
use the size counter, instead of these heuristics.
The <lexical-ref> case is smarter now, and tries to avoid propagating
too much data. Perhaps it should be dumber though, and use a
counter. That would require changes to the environment structure.
Inline <lambda> applications to <let>, so that we allow residual
lexical references to have bindings. Add a `for-operand' helper, and
use it for the RHS of `let' expressions. A `let' is an inlined
`lambda'.
`Let' and company no longer elide bindings if the result is a
constant, as the arguments could have effects. Peval will still do as
much as it can, though.
* test-suite/tests/tree-il.test ("partial evaluation"): Update the tests
for the new expectations. They are uniformly awesomer, with the
exception of two cases in which pure but not constant data is not
propagated.
* module/language/tree-il/optimize.scm (peval): Rename `record-lexicals'
to `record-lexical-bindings'. Record residualized lexical
references. Record lexical references in maybe-unlambda.
Unfortunately this has the disadvantage that the speculative mapping
of lambda expressions to lexical references records that reference,
even if we are not going to residualize it. After processing a `let',
prune pure unreferenced bindings. (We can do better than this in the
future: we can simply process them for effect.)
* test-suite/tests/tree-il.test (pass-if-peval): More debugging.
("partial evaluation"): Update to reflect the fact that the `y'
binding won't be emitted.
* module/language/tree-il/optimize.scm (peval): Rename
`pure-expression?' to `constant-expression?', in the sense of GCC's
`pure' and `const'. A <toplevel-ref> is not constant, because it can
be mutated. A <dynref> isn't constant either, for the same reason.
* test-suite/tests/tree-il.test ("partial evaluation"): Add a test, and
update existing tests that assumed that toplevel-ref would propagate.
* module/ice-9/match.scm (match): Always introduce a lexical binding, to
avoid http://debbugs.gnu.org/9567. Real fix ongoing. Patch and
original report by Stefan Israelsson Tampe.
* test-suite/tests/match.test: Add test.
* module/language/tree-il/primitives.scm (*primitive-constructors*):
Record car, cdr, vector-ref, and struct-ref as "constructors".
Comment to come later.
(*effect-free-primitives*): Update.
* test-suite/tests/tree-il.test ("partial evaluation"): Add tests.
* module/language/tree-il/optimize.scm (code-contains-calls?): New
procedure.
(peval): Use it and abort inlining if the residual code of a procedure
application contains recursive calls. Suggested by Wingo, Waddell,
and Dybvig. Fixes <http://debbugs.gnu.org/9542>.
* test-suite/tests/tree-il.test ("partial evaluation"): Update 2 tests
that relied on the previous behavior. Add 1 another test.
* test-suite/tests/gc.test (stack-cleanup): New procedure.
("Unused modules are removed"): Use it.
* test-suite/tests/threads.test (stack-cleanup): Likewise.
("mutex with owner not retained (bug #27450)"): Use it.
* module/language/tree-il/optimize.scm (peval)[maybe-unlambda]: New
procedures.
Use it to de-duplicate named lambdas. This fixes the scoping bug
described at <https://lists.gnu.org/archive/html/bug-guile/2011-09/msg00019.html>.
* test-suite/tests/tree-il.test ("partial evaluation"): Add tests to
reproduce the bug.
* module/language/tree-il/optimize.scm (peval): Propagate ARGS to BODY
only when all of ARGS are pure. Change APP to use `maybe-unconst' for
its arguments.
* test-suite/tests/tree-il.test ("partial evaluation"): Add tests for
mutability preservation and non-propagation of non-constant arguments
to lambdas.
* module/language/tree-il/optimize.scm (peval)[make-values]: Distinguish
between 1 or another number of values.
[mutable?, make-value-construction, maybe-unconst]: New procedures.
Use it in <let>, <letrec>, <toplevel-define>, and <lambda-case>.
* test-suite/tests/tree-il.test ("partial evaluation"): Add tests
for mutability preservation.
This is a followup to f9c1b8278d ("Tweak
`statprof.test' for faster machines.").
* test-suite/tests/statprof.test ("statistical sample counts within
expected range"): Compile with `#:partial-eval? #f'.
* test-suite/tests/statprof.test ("statistical sample counts within
expected range"): Increase NUM-CALLS and the frequency so that they
are at least a few samples on my new 2.6 GHz laptop.
Thanks to William R. Cook for his excellent tutorial,
<http://softlang.uni-koblenz.de/dsl11/>.
* module/language/tree-il/optimize.scm (optimize!): Call `peval' unless
the #:partial-eval? option asks otherwise.
(peval): New procedure.
* module/language/tree-il/inline.scm: Add comment.
* module/language/tree-il/primitives.scm (*primitive-constructors*): New
variable.
(*effect-free-primitives*): Use it.
(constructor-primitive?): New primitive.
* test-suite/tests/tree-il.test (assert-tree-il->glil): Extend to
support `with-partial-evaluation', `without-partial-evaluation', and
`with-options'.
(peval): New binding.
(pass-if-peval): New macro.
("lexical refs"): Run tests without partial evaluation.
("letrec"): Likewise.
("the or hack"): Likewise.
("conditional"): Likewise, for some tests.
("sequence"): Adjust to new generated code.
("partial evaluation"): New test prefix.
* module/ice-9/match.scm (slot-ref, slot-set!, is-a?): New macros.
* module/ice-9/match.upstream.scm: Update from Chibi-Scheme.
* test-suite/Makefile.am (SCM_TESTS): Add `tests/match.test.upstream'.
* test-suite/tests/match.test (rtd-2-slots, rtd-3-slots): New record
types.
("matches")["records"]: New test prefix.
("doesn't match")["records"]: New test prefix.
Include `match.test.upstream'.
* test-suite/vm/t-match.scm (matches?): Fix `$' example.
* module/texinfo.scm (texi-command-specs, complete-start-command):
Upstream texinfo has aliased `url' to `uref'. Let's do the same.
* test-suite/tests/texinfo.test ("test-texinfo->stexinfo"): Add a test.
* libguile/srfi-1.h:
* libguile/srfi-1.c (scm_srfi1_drop_right, scm_srfi1_take_right): Remove
these internal functions, replacing with Scheme implementations.
* module/srfi/srfi-1.scm (take-right, drop-right): Add these impls from
the reference code. They do the right thing for improper lists,
according to the spec, but they diverge for circular lists. Oh well.
* test-suite/tests/srfi-1.test ("drop-right", "take-right"): Add more
tests.
* module/language/tree-il.scm (tree-il->scheme): Fix incorporation of
`lambda' in a `case-lambda'.
* test-suite/tests/tree-il.test ("tree-il->scheme"): Add a test.
* module/web/request.scm (build-request): Make sure that HTTP/1.1
requests have the Host header set, per RFC 2616 section 9.
* test-suite/tests/web-request.test ("example-1"): Add test.
* libguile/modules.c (resolve_duplicate_binding): Fix unbound -> #f
conversion for the imported bindings. Pass the existing entry in the
import obarray as the resolved var (7th arg), and properly pass #f as
the value (8th arg) if there is no such binding. Fixes
merge-generics; before, the <boolean> type test (indicating no
previous value) was not being triggered. This bug has been present
since 2007 at least, though it was not in 1.8.
* test-suite/tests/modules.test ("duplicate bindings"): Add a test that
the var and val are both #f. These types are used by GOOPS.
* libguile/bytevectors.c (COMPLEX_ACCESSOR_PROLOGUE, COMPLEX_NATIVE_REF,
COMPLEX_NATIVE_SET): New macros.
(bytevector_ref_c32, bytevector_ref_c64): Defined in terms of
`COMPLEX_NATIVE_REF'.
(bytevector_set_c32, bytevector_set_c64): Defined in terms of
`COMPLEX_NATIVE_SET'.
(bytevector_ref_fns): Make `static'.
* test-suite/tests/srfi-4.test ("c32 vectors")["generalized-vector-ref",
"generalized-vector-set!", "generalized-vector-ref, out-of-range",
"generalized-vector-set!, out-of-range"]: New tests.
("c64 vectors")["generalized-vector-ref", "generalized-vector-set!",
"generalized-vector-ref, out-of-range",
"generalized-vector-set!, out-of-range"]: New tests.
* libguile/read.c (scm_read_sexp): Don't confuse `#{.}#' with `.' for
the purpose of reading dotted pairs. Thanks to CRLF0710 for the
report.
* test-suite/tests/reader.test ("#{}#"): Add test.
* module/language/tree-il/primitives.scm (+, *, cons*): In the case of
just one argument (the identity case), expand to (values x) instead of
just x. Fixes values truncation in that case.
(values): Likewise remove (values x) -> x translation, as the compiler
will do it for us, and this fixes (values (values 1 2)).
* module/language/tree-il/compile-glil.scm (flatten-lambda-case): Handle
`values' in a push context here.
* test-suite/tests/tree-il.test ("values"): Add some tests.
http://savannah.gnu.org/bugs/?33362
* module/language/tree-il/compile-glil.scm (flatten-lambda-case): Rename
from flatten, as it really just takes a particular case. Instead of
iteratively compiling lambda cases through `comp', tail-call through
flatten-lambda-case. This allows code to see which case it's being
compiled in. Take advantage of that to limit the self-tail-call
optimization to self-calls to the same case -- otherwise we might be
jumping to a label without having reserved the right number of
locals.
(flatten-lambda): Adapt the caller.
* test-suite/tests/compiler.test ("case-lambda"): Add a test.
