* module/language/cps/closure-conversion.scm (convert-one): Build
closures with make-closure, cons, and so on; leave lowering to scm-ref
to the backend.
* module/language/cps/effects-analysis.scm (compute-known-allocations):
For a primcall that allocates, sometimes we will synthesize auxiliary
definitions as part of CSE, for example to indicate that if (cons x y)
is bound to z, that a later call to (car z) should give x unless there
might be an intervening set-car!. We had a bug in which aux definitions
attached to allocations were being incorrectly associated with the first
operand. Probably this is a bug in other contexts but it really starts
to hit with the high-level constructors, e.g. `box`.
* module/language/cps/lower-primcalls.scm (f64->scm): Move here...
* module/language/cps/reify-primitives.scm (reify-primitives): from
here. Seems a more fitting place.
* module/language/cps/optimize.scm (lower-cps/generic): Rename from
lower-cps; these are the lowerings that apply to everyone.
(select-opts-for-optimization-level): Factor out of make-cps-lowerer.
(make-backend-cps-lowerer): New procedure. For the Guile VM backend, we
have a few mandatory passes, including the new lower-primitives.
(make-cps-lowerer): Apply backend-specific lowering pass.
This pass will implement the specialized lowering of object accessors to
Guile's VM primitives.
* am/bootstrap.am (SOURCES): Add new file.
* module/language/cps/lower-primcalls.scm: New file.
* module/language/cps/utils.scm (compute-var-representations): For the
wasm target, these values are (ref $kvarargs), not i64. Will need to
distinguish from bytevector pointers at some point, though.
* module/language/cps/compile-bytecode.scm (compile-function):
The rsh/lsh patterns would never match. I think these would end up
dispatching through emit-text though.
* module/language/tree-il/peval.scm (peval): Reduce multiple case lambda
in <call> trees according to the number of arguments. Do not try to
reduce case-lambda using keyword arguments.
* test-suite/tests/peval.test: Tests.
* module/language/tree-il/analyze.scm (<module-info>): New record type.
(unused-module-analysis): New variable.
(make-unused-module-analysis): New analysis.
(make-analyzer): Add it.
* module/system/base/message.scm (%warning-types): Add 'unused-module'.
* test-suite/tests/tree-il.test (%opts-w-unused-module): New variable.
("warnings")["unused-module"]: New test prefix.
* NEWS: Update.
This reduces the amount of memory that needs to be allocated while
writing the ELF file to disk.
Note: We're abusing #:page-aligned? in 'link-elf' to choose whether to
return a bytevector or a procedure.
* module/system/vm/linker.scm (process-reloc): Subtract SECTION-OFFSET
when writing to BV.
(write-linker-object): Pass BV directly to the linker object writer.
(link-elf): When PAGE-ALIGNED? is false, call 'bytevector-slice' from
here. When it is true, return a procedure that takes a port and writes
to it, without having to allocate a bytevector for the whole ELF
container.
* module/language/bytecode/spec.scm (bytecode->value): Handle X being a
procedure instead of a bytevector.
(bytecode) <#:printer>: Likewise.
* test-suite/tests/linker.test (link-elf-with-one-main-section): Pass
#:page-aligned? #f.
Fixes <https://bugs.gnu.org/60522>.
Reported by Sascha Ziemann <ceving@gmail.com>.
* module/language/tree-il/primitives.scm (make-vector): Return #f when
passed an incorrect number of arguments.
* test-suite/tests/peval.test ("partial evaluation"): Add tests.
* module/language/tree-il/peval.scm
(augment-var-table-with-externally-introduced-lexicals): New helper.
* module/language/tree-il/peval.scm (peval): Augment store with any
lexicals introduced by expand-primitives.
* test-suite/tests/peval.test ("partial evaluation"): Add tests.
* module/language/tree-il/primitives.scm (bind-lexicals): New helper.
(expand-eq, expand-chained-comparisons): Ensure all arguments are
eagerly evaluated. Previously an intermediate #f result would shortcut
the evaluation.
* test-suite/tests/compiler.test ("size effects in multi-arg eq / <"):
Add test.
Avoiding systematic conversion from source vectors to property alists
saves 20% on the final heap size of a process doing:
(compile-file FILE #:optimization-level 1)
where FILE is large.
* module/language/tree-il.scm (tree-il-src/ensure-alist): New procedure
with setter. Export as 'tree-il-src'.
* module/ice-9/psyntax.scm (build-void, build-call)
(build-conditional, build-lexical-reference, build-lexical-assignment)
(build-global-reference, build-global-assignment)
(build-global-definition, build-simple-lambda, build-case-lambda)
(build-lambda-case, build-primcall, build-primref)
(build-data, build-sequence, build-let, build-named-let)
(build-letrec, expand-body): Remove (sourcev->alist src) calls.
* module/ice-9/psyntax-pp.scm: Regenerate.
* module/language/tree-il/analyze.scm (shadowed-toplevel-analysis): Use
'tree-il-src' instead of accessing the 'src' slot directly.
* module/system/vm/assembler.scm (link-debug): Adjust so PC can be
followed by a vector or an alist.
* module/language/tree-il/peval.scm (peval): The intention was for
resolve-interface to return an interface only if it exists, but actually
it throws if the interface isn't know. Fix to do what we intended to
do.
* module/language/tree-il/inlinable-exports.scm (compute-decoder): Fix
bug in which inlinable exports in a pure module would have a missing
make-struct/no-tail binding.
* module/language/cps/return-types.scm: New file.
* module/Makefile.am (SOURCES):
* am/bootstrap.am (SOURCES): Add new file.
* module/language/tree-il/compile-cps.scm (sanitize-meta): Strip
"noreturn" and "return-type" properties -- these should only be
computed by Guile.
* module/language/cps/verify.scm (check-arities): If a callk continues
to kargs, the caller knows the number of return values that the callee
provides and no number-of-values check is needed.
* module/language/cps/contification.scm (apply-contification): Allow
contification of known-return-values calls.
* module/language/cps/reify-primitives.scm (uniquify-receive)
(reify-primitives): No need for uniquify-receive any more as receive
shuffles are attached to the call, not the continuation.
* module/language/cps/compile-bytecode.scm (compile-function): Add kargs
case.
* module/language/cps/slot-allocation.scm (lookup-send-parallel-moves):
Rename from `lookup-parallel-moves'.
(lookup-receive-parallel-moves): New function. Now we attach "receive
moves" to call and prompt conts instead of to their continuations.
(compute-shuffles): Refactor to allow a continuation to have both send
and receive shuffles.
(compute-frame-size): Refactor for new shuffles mechanism
(allocate-slots): Allow calls to proceed directly to kargs.
* module/language/cps/effects-analysis.scm (compute-known-allocations):
(compute-clobber-map): Add "conts" parameter, and use it to compute
primcalls that access known allocations. A write to a known allocation
only clobbers a read to a known allocation if they are the same.
* module/language/cps/cse.scm (eliminate-common-subexpressions-in-fun):
Pass conts also to compute-clobber-map.
* module/language/cps/utils.scm (compute-var-representations): Use
'arg-representations from metadata for arg representations.
* module/language/tree-il/compile-cps.scm (sanitize-meta):
(convert): Make sure incoming terms have no arg representations.
* module/language/cps/graphs.scm (rename-keys, rename-intset)
(rename-graph, compute-reverse-control-flow-order)
(compute-live-variables): Move here from slot-allocation.
* module/language/cps/utils.scm: Remove duplicate compute-idoms
definition.
(compute-defs-and-uses, compute-var-representations): Move here from
slot-allocation.
* module/language/cps/slot-allocation.scm: Move routines out to utils
and graphs.
Reported by Marius Bakke <marius@gnu.org>
at <https://issues.guix.gnu.org/50696>.
Previously, the baseline compiler would incorrectly emit a right shift
when for, say, (ash x 2), and a left shift for (ash x -2).
* module/language/tree-il/compile-bytecode.scm (canonicalize): When Y is
negative, emit 'rsh', not 'lsh'.
* test-suite/tests/numbers.test ("ash at -O1"): New test.
Previously the pattern would only match when the two 'src' values were
the same, which is not the case for example when running on the
interpreter.
* module/language/tree-il/compile-bytecode.scm (canonicalize): In 'ash
primcall pattern, rename second 'src' to 'src*'.
