Repository navigation
Fixes from the pre-release review - #1368
Merged
Merged
Conversation
…kError The char writer behind io/writer, FileWriter and OutputStreamWriter rendered a char[] argument as #object[...] and ignored append's start/end; it now goes through the same writer-piece-range/append-text as the other writers. Files/delete and deleteIfExists threw away delete-file's #f, so a file in a read-only directory stayed put while the call succeeded. The failure is now raised with the JDK's class for the errno. PrintWriter.checkError answered false unconditionally; it now asks the stream it wraps, as PrintStream's does.
…nm versioned names
bigint/biginteger of a double went through the double's binary value; the JVM
uses BigDecimal.valueOf, the shortest decimal string, so (bigint 1e30) is 10^30.
A socket write region outside its array raised ex-info from ffi/write-array;
it checks first and raises IndexOutOfBoundsException with the JDK's message.
Chez on Windows rounds a timed wait to the nearest millisecond, so the exact
sub-millisecond deadlines spun through their last half millisecond. Waits
there round up to whole milliseconds.
GNU nm prints a versioned import as name@VERSION, which the :static {:lib}
symbol check never matched against the C name the app binds.
yogthos
changed the title
Apple's iconv decodes a double-byte character cut off at the end of the input as ASCII and drops the shift state, where glibc stops before it with EINVAL. A stream decode on ISO-2022-JP/KR/CN now tracks the shift state itself and passes iconv only up to the last whole character, so the macOS corpus row for the ISO-2022-JP reader passes and a KR stream read in 3-byte chunks no longer reads a가3*b.
…side the object evaluate read the whole stack on every call to have the caller's frames ready for an error, 2.6x the cost of a successful eval. It is read in the handler now, where the stack below is still evaluate's and its caller's. The flonum payload probe counted down from 8, but the payload sits 6 past the pointer on 64-bit Chez, so the first two probes read past the object's end. The bigint fix passes one more clojure-test-suite assertion; baseline updated.
Source mode installs no collector policy, so the stack limit's default fell back to 80% of physical RAM, per thread: a runaway recursion could take most of the machine before StackOverflowError. It now uses the ceiling a built program would get (JOLT_MAX_HEAP, else 25% of RAM), and RAM only when the ceiling is explicitly off.
It kept only the host string and resolved on first use inside =/hash, printed host:port, answered an InetAddress for a host that did not resolve, and had no createUnresolved. It now holds the InetAddress (or none), prints and compares the JDK's way, and connect/bind refuse an unresolved one. A literal-built InetAddress carries no name, so getByName "127.0.0.1" prints /127.0.0.1 and getHostName does the reverse lookup. Drops the two known-divergence entries this covered.
…the member PrintWriter let an IOException from the writer under it escape every write, flush and close. As the JDK's, it now records it for checkError (an InterruptedIOException re-asserts the interrupt) and lets anything else through. Whether the target is a Writer is decided once, at construction, so the guard costs nothing on a StringWriter or process stream. A static that a provided class lacks reported "No dependency provides" the class: (InetSocketAddress/createUnresolved ...) said so with jolt.socket loaded. A class with a registered constructor answers "No matching field or method".
The keyword, get and collection-call arms of emit-invoke bypass emit-call, so #1336's restore never reached them: a deftype valAt or a sorted map comparator that tail-called a jolt fn left that returned fn as the innermost frame of the next host throw. A non-tail lookup now restores the slot as other calls do; the keyword-lookup loop measures the same within noise. Reminted.
make -j builds release and debug at once, and both write build-boot.inputs and build-boot.so. On Windows a rename onto a name the other build just wrote raises, since MoveFile does not replace, which failed the build at the inputs rename. Both builds hash the same sources, so the second keeps the first's file and drops its own temp.
A callback that ended in a tail call and returned to untraced code (a core HOF runs the callbacks) left its site pair in the slot, and the next callback's throw from host code reported that returned fn as its innermost frame. A non-tail call now moves the slot's pair to an outer slot (vreg 12) and clears the slot while it runs, so a throw inside it has no innermost pair. The outer pair, the nearest erased caller further out, is captured with the throw and used only as a gap bridge or landed below the live frame its tail call went to; with no live frame confirming it, it is dropped. The apply-erased caller the stack-trace gate pins still comes back. ~1.5% on a non-tail-call-bound loop. Reminted.
A (lambda) literal list in its syntax-rules read to the checker as a lambda form with no body, and the file came back unreadable.
Merged
babashka.process adds the hook after .start returns, so a SIGTERM landing in between ended jolt with no hook, and :shutdown destroy-tree left the child running; process-test's fd-level INHERIT SIGTERM row caught it on a loaded CI runner. jolt.process wraps process*, the spawn every entry point uses, to register the hook first with a promise of the process, delivered when the spawn returns; a hook run inside the window waits for it and kills the tree. Upstream's drop-and-run on exit and drop on a failed spawn are kept.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Three bugs from the pre-release review of the changes since v0.8.21.
io/writer(since Take Clojure's options in clojure.java.io's openers #1320),FileWriterandOutputStreamWriterwrote achar[]as#object[...]and ignoredappend's start/end. It now uses writer-piece-range / append-text like the other writers.(.write w buf 0 n)is the reader-buffer copy pattern, so this broke real code.Files/delete/deleteIfExistsdiscardeddelete-file's#f: a file in a read-only directory stayed on disk and the call returned normally (deleteIfExistsanswered true). The failure is raised with the JDK's class for the errno (AccessDenied, DirectoryNotEmpty, else FileSystemException with strerror); Windows, which sets no errno, reads as AccessDenied.move's replace path relied on the silent failure and re-checked; it now lets nio-delete1 raise.PrintWriter.checkErroralways answered false; it asks the wrapped stream, as PrintStream's checkError does since Drop writes to a failed process stream instead of raising #1337.All three checked against Clojure 1.12 on the JVM. Unit rows for the first two were red before the fix; the closed-stdout smoke has a PrintWriter case.
Second commit, smaller findings from the same review:
bigint/bigintegerof a double used its binary value; the JVM goes throughBigDecimal.valueOf, so(bigint 1e30)is 10^30 (released bug, corpus rows certified).write(b, off, len)region outside the array raised ex-info (since Copy a socket write's bytes in one step #1318); it raisesIndexOutOfBoundsExceptionwith the JDK's message, before the write.:static {:lib}symbol check (Fail a build whose {:lib} native cannot supply what the app binds #1358) never matched GNU nm'sname@VERSIONimports against the C name.Third commit:
make teston macOS failed #1366's corpus row "an ISO-2022-JP reader keeps its shift state across reads" (CI runs the corpus on Linux only). Apple's iconv decodes a double-byte character cut off at the end of the input as ASCII and loses the shift state; glibc stops before it with EINVAL. The ISO-2022 decoders now track the shift state themselves and hand iconv only whole characters. New corpus row reads ISO-2022-JP and KR in 3-byte chunks; KR reada가3*bon macOS before.