v7.1.00

Jawk 7.1 makes Jawk easy to install — curl -fsSL https://jawk.io/get | sh and you have a
jawk command — and takes a deep pass over the numeric core: printf/sprintf are now
implemented natively with POSIX AWK / gawk semantics, integral arithmetic is computed exactly in
64 bits, and getline follows gawk down to which variables each form sets. Along the way, scripts
run measurably faster.

📦 One-command installation

• Install the jawk executable with one command — curl -fsSL https://jawk.io/get | sh on
  Linux/macOS, irm https://jawk.io/get.ps1 | iex on Windows. No sudo: the launcher goes into
  ~/.local/bin (or %LOCALAPPDATA%\Jawk), and it locates a Java runtime (Java 8 or later) at
  run time, so upgrading Java never requires a reinstall. JAWK_VERSION pins a release,
  JAWK_INSTALL_DIR changes where the launcher lands. Every release now also publishes a
  version-less jawk-standalone.jar alias with .sha256 checksums, so
  releases/latest/download/jawk-standalone.jar is a stable download URL
  by @bertysentry in #584
• JAWK_CLASSPATH in the launchers — set it and the launcher runs the CLI with the standalone
  jar plus your entries on the JVM class path, so custom
  extension jars load with -l <fully.qualified.ClassName>.
  Previously the launcher always used java -jar, which ignores every class path setting
  by @bertysentry in #588
• CI smoke tests for both installers — the real scripts are exercised end to end on every
  change, so a broken jawk.io/get cannot reach the site
  by @bertysentry in #587

🖨️ Native printf with AWK semantics

printf/sprintf no longer delegate to the Printf4J library, which emulated glibc's printf();
they are implemented in the new io.jawk.jrt.AwkPrintf with the semantics POSIX specifies and gawk
implements, verified against gawk 5.0/5.1
by @bertysentry in #543

• %s uses AWK's number-to-string rules — integral values print without a fractional part
  (printf "%s", i after i++ prints 1, not 1.0), non-integral values honor the script's
  current CONVFMT.
• %c prints the character for a numeric code point — numbers, numeric strings, and fields
  alike — or the first character of a string value.
• Dynamic precision %.*f joins dynamic width %*d, negative values included; gawk
  positional specifiers (%2$s) and the ' grouping flag (%'d) are supported.
• Out-of-range integers follow gawk — negative values wrap to unsigned 64-bit for
  %u/%o/%x/%X, and values beyond 64 bits print their full decimal expansion for %d/%i.
• NaN and infinities print as nan, inf, -inf instead of Java's NaN/Infinity — in
  print, printf, and number-to-string conversions.
• Correct rounding — %e/%f/%g round the exact binary value of the double, halfway cases
  to even, like the C library gawk uses (printf "%.0f", 2.5 prints 2), and the same rounding
  applies through CONVFMT and OFMT (#548).
• Too few arguments is a fatal error, as in gawk; unknown conversions — including %n, which
  Printf4J turned into a newline — print verbatim without consuming an argument.
• New JavaStringFormatAwkSink for Java embedders: a sink whose printf/sprintf use
  String.format(...), giving scripts Java-only conversions such as %,d and %tY.

🔢 Exact numbers

• Integral arithmetic is computed exactly in 64 bits — +, -, *, an evenly-dividing /,
  %, unary ±, ++/-- and the compound assignments, on variables, array elements, and fields
  alike, keep all their digits beyond 2^53: print 9007199254740992 + 1 prints 9007199254740993,
  where gawk computes in doubles and prints 9007199254740992 unless run with -M. Overflow falls
  back to floating point, and comparisons between two exact integers are exact as well. Values
  beyond the 64-bit range are no longer saturated either: print 2^100 prints the full
  1267650600228229401496703205376
  by @bertysentry in #552
• Fields keep their numeric value through += and ++, like = always did, and rebuilding
  $0 converts numeric fields with CONVFMT, as POSIX requires and gawk does:
  { $1 = 0.1 + 0.2; print } prints 0.3 ... instead of the raw 0.30000000000000004 ...
  by @bertysentry in #552
• Exponent literals are lexed as one number — print 2e3 prints 2000; previously the lexer
  stopped before the exponent, so 2e3 parsed as 2 concatenated with an uninitialized variable
  e3. Long numeric strings now keep all their digits instead of stopping after 26 characters, an
  integer literal too large for 64 bits is a floating-point constant instead of a
  NumberFormatException, and substr()'s length argument accepts every AWK number form
  by @bertysentry in #544
• Array subscripts convert with the CONVFMT in effect when the subscript is used, as POSIX
  requires and gawk does, so a later CONVFMT change no longer retroactively rewrites existing
  keys — and integral subscripts beyond the 64-bit range key as their exact digits
  by @bertysentry in #551
• Blank-padded numeric strings are recognized as strnums — the record " 12 " compares
  numerically ($0 == 12 is true), as in gawk; string constants such as x = " 12 " still compare
  as strings (#550)
  by @bertysentry in #573

📥 getline, redirections, and special filenames

• Redirected getline sets exactly the variables gawk documents — getline < file and
  cmd | getline set $0 and NF only; getline var < file and cmd | getline var set var
  only. Previously every redirected form also advanced NR, so a side read shifted the record
  numbers of the whole main input, and took over FILENAME
  by @bertysentry in #572
• A getline that cannot open its file returns -1 with ERRNO carrying the gawk-style
  description, so the idiomatic while ((getline line < f) > 0) guard and probing for optional
  files work; previously the underlying Java IOException aborted the whole script
  by @bertysentry in #572
• A getline that reads nothing leaves its target untouched, as POSIX requires — no more
  clearing $0/NF or assigning the empty string at end of input or on error
  by @bertysentry in #572
• getline into a field — getline $n < file, cmd | getline $n, and plain getline $n —
  stores the record into that field and rebuilds $0; previously any of these forms corrupted the
  interpreter's operand stack and crashed the script on the first record
  by @bertysentry in #572
• The gawk special filenames /dev/stdout, /dev/stderr, /dev/stdin and their /dev/fd/N
  spellings route to the streams the process already holds open, instead of being opened — and
  truncated — as regular files. close() flushes the redirection and reports success without ever
  closing the underlying stream
  by @bertysentry in #564
• /dev/null designates the null device on Windows (NUL) in redirections, in getline, and
  as an input operand, as gawk's Windows port does. Previously it was opened as a relative path,
  creating and truncating dev\null — and keeping the output that was meant to be discarded
  by @bertysentry in #568

🖥️ CLI and process integration

• POSIX attached option-arguments — -fprog.awk, -vx=1, -F: and the Jawk-specific -L,
  -K, -l now work like their spaced forms, as in gawk, mawk, BWK awk, and goawk; previously the
  glued form was rejected with Unknown parameter
  by @bertysentry in #576
• -V / --version prints the Jawk version and the Java runtime that executes it, and the
  usage printed by -h now names the command you actually typed — the installed launchers pass it
  through the new JAWK_PROGRAM_NAME variable — instead of always describing a java -jar
  invocation the user never made
  by @bertysentry in #593
• system() and command-pipe children inherit the standard input, as POSIX requires and gawk,
  mawk, and BWK awk do: echo hi | jawk 'BEGIN { "sort" | getline l; print l }' prints hi, and
  terminal-aware commands such as "stty size" | getline can reach the controlling terminal.
  Previously the child's standard input was always closed, so stdin filters read nothing
  by @bertysentry in #577

🧩 Language and correctness

• next is accepted inside user-defined functions, as in gawk, mawk, and BWK awk: called from
  an input rule it abandons the current record and unwinds the active calls; called from BEGIN,
  END, BEGINFILE, or ENDFILE it is a runtime fatal error, matching gawk. Previously any next
  inside a function was rejected at compile time
  by @bertysentry in #582
• Conditional expressions feeding a further operation now optimize correctly — shapes like
  (v ? v : 24) * 2, -(v ? 5 : 24), or $(v ? 1 : 2). The peephole literal fold used to merge
  the false branch's literal with the operation at the branches' join point, so (60 ? 60 : 24) * 2
  yielded 48 and leaked a value onto the operand stack that displaced later print operands
  by @bertysentry in #579
• --profile no longer leaks function timings on nextfile — a call abandoned by nextfile is
  recorded as exited with its actual duration, instead of leaking a stack entry (unbounded memory
  growth on long profiled runs) and being reported with wildly wrong durations at program exit
  by @bertysentry in #581

⚡ Performance

• Flat Object[] operand stack replaces the boxed ArrayDeque in the AVM: push/pop are
  single array accesses the JIT can fully inline, and argument slices are bulk-copied. −13.5% on a
  tight arithmetic loop, −5% to −14% across the expression benchmarks
  by @bertysentry in #555
• AwkPrintf hot paths — per-locale symbol cache, parsed-format cache, and a guarded rounding
  fast path: 1.3×–9.8× faster formatting (%f of 0.1 3.0×, CONVFMT of 3.14159 3.2×), with
  output verified bit-identical — 19.6M rounding comparisons against the exact-expansion reference
  and 600k differential-fuzz cases, zero mismatches
  by @bertysentry in #560
• Non-throwing associative-array key normalization — every non-integer key (all SUBSEP-joined
  arr[x,y] keys, all string keys) used to allocate a NumberFormatException and fill in its stack
  trace, ~600 ns on every get, put, and in test; it is now a hand-rolled scan with parity to
  Long.parseLong locked in by tests
  by @bertysentry in #561
• AVM.executeTuples kept under HotSpot's huge-method limit — the dispatch loop sat at 7994
  bytecodes (8013 under ECJ), right at the 8000-bytecode cliff past which HotSpot never JIT-compiles
  a method at all: crossing it made every AWK script ~4× slower, silently. It is now 6728 bytecodes,
  with a test that fails the build above 7500
  by @bertysentry in #563

⚠️ Breaking changes

• AwkSink.printf(...) and AwkSink.sprintf(...) now take CONVFMT — printf receives it
  between ofmt and format, and sprintf(convfmt, format, values...) is the single sprintf
  method and the customization point. Custom sinks must be updated. The org.metricshub:printf4j
  dependency is gone (#543).
• Programs precompiled with an earlier version are rejected and must be recompiled: they may
  carry constants folded in floating point (#552).
• The broken stdin extension is removed, along with the BlockObject/BlockManager API whose
  only user it was. Its StdinHasInput() and StdinBlock() never reported a pending end of input,
  hanging every read loop that guarded its StdinGetline() calls, and no other AWK implementation
  provides these functions. -l stdin is now rejected with Unknown extension 'stdin'; read
  standard input through the regular input loop, getline, or /dev/stdin
  by @bertysentry in #594

Every user-visible change above is described with its before/after on the
behavior changes page.

📚 Documentation

• Pre-release documentation audit — the site and README now lead with the jawk command, the
  installation guide is simplified, and over-broad statements were tightened
  by @bertysentry in #589
• Release plumbing for v7.1.00-SNAPSHOT by @github-actions[bot] in #542

Full Changelog: v7.0.01...v7.1.00

Dependabot

• build(deps): bump org.apache.maven:apache-maven from 3.8.6 to 3.9.16 by @dependabot[bot] in #554