parallel: --pipe --block with negative size works like --pipepart.

doc/haikus

@@ -4,11 +4,12 @@
  
 Quote of the month:
 
-ainda não inventaram palavras capazes de expressar minha gratidão aos desenvolvedores do GNU Parallel
--- nueidris ‪@nueidris.kawaii.social‬
 
-Und die Tage jetzt hab ich GNU parallel für mich entdeckt, auch ne nette Geschichte, gerade wenn's irgendwelche remote APIs sind.
--- Vince ‪@dd1des.bsky.social‬
+  ainda não inventaram palavras capazes de expressar minha gratidão aos desenvolvedores do GNU Parallel
+    -- nueidris ‪@nueidris.kawaii.social‬
+
+  Und die Tage jetzt hab ich GNU parallel für mich entdeckt, auch ne nette Geschichte, gerade wenn's irgendwelche remote APIs sind.
+    -- Vince ‪@dd1des.bsky.social‬
 
   GNU parallel is so satisfying
     -- ‪James Coman‬ ‪@jcoman.bsky.social‬

src/parallel

@@ -1165,6 +1165,8 @@ sub spreadstdin() {
     my $blocksize = int($Global::blocksize);
     my $in = *STDIN;
     my $timeout = $Global::blocktimeout;
+    my @parts;
+    my $everything_read;
 
     if($opt::skip_first_line) {
 	my $newline;
@@ -1178,12 +1180,14 @@ sub spreadstdin() {
     my $eof;
     my $garbage_read;
 
-    sub read_block() {
-	# Read a --blocksize from STDIN
+    sub read_with_alarm($) {
+	my ($readsize) = @_;
+	my ($nread,$alarm,$read_everything);
+	if($readsize < 0) {
+	    $readsize = -$readsize;
+	    $read_everything = 1;
+	}
 	# possibly interrupted by --blocktimeout
-	# Add up to the next full block
-	my $readsize = $blocksize - (length $buf) % $blocksize;
-	my ($nread,$alarm);
 	eval {
 	    local $SIG{ALRM} = sub { die "alarm\n" }; # NB: \n required
 	    # --blocktimeout (or 0 if not set)
@@ -1191,13 +1195,13 @@ sub spreadstdin() {
 	    if($] >= 5.026) {
 		do {
 		    $nread = sysread $in, $buf, $readsize, length $buf;
-		    $readsize -= $nread;
+		    if(not $read_everything) { $readsize -= $nread; }
 		} while($readsize and $nread);
 	    } else {
 		# Less efficient reading, but 32-bit sysread compatible
 		do {
 		    $nread = sysread($in,substr($buf,length $buf,0),$readsize,0);
-		    $readsize -= $nread;
+		    if(not $read_everything) { $readsize -= $nread; }
 		} while($readsize and $nread);
 	    }
 	    alarm 0;
@@ -1208,7 +1212,60 @@ sub spreadstdin() {
 	} else {
 	    $alarm = 0;
 	}
-	$eof = not ($nread or $alarm);
+	# Is this EOF?
+	return not ($nread or $alarm);
+    }
+    
+    sub read_block_per_jobslot() {
+	if(not $everything_read) {
+	    # Read everything in readsize of 2^17
+	    # 21 = 20s
+	    # 20 = 16s
+	    # 19 = 16s,15s
+	    # 18 = 15s
+	    # 17 = 14s
+	    # 16 = 14s
+	    # 15 = 14s
+	    $eof = read_with_alarm(-2**17);
+	    if($eof) {
+		$everything_read = 1;
+		# Trick the rest of the code to think we are not done reading blocks yet.
+		$eof = 0;
+		# Chop into parts
+		my $total_size = length($buf);
+		my $jobslots = 	$Global::max_jobs_running;
+		my $parts = -$Global::blocksize * $jobslots;
+		my $part_size = int($total_size / $parts);
+		for my $i (0 .. $parts - 1) {
+		    my $start = $i * $part_size;
+		    my $end = ($i == $parts - 1) ? $total_size : $start + $part_size;
+		    # Extract the chunk from buffer
+		    push @parts, substr($buf, $start, $end - $start);
+		}
+		$buf="";
+	    } else {
+		# What do we do here? Caused by alarm.
+		::die_bug("Read block from pipe failed");
+	    }
+	}
+	if(@parts) {
+	    # pop part and return that
+	    $buf .= shift @parts;
+	} else {
+	    # All parts are done: Let the rest of the code know we are EOF
+	    $eof = 1;
+	}
+    }
+
+    sub read_block() {
+	# Read a --blocksize from STDIN
+	if($Global::blocksize < 0) {
+	    read_block_per_jobslot();
+	} else {
+	    # Add up to the next full block
+	    my $readsize = $blocksize - (length $buf) % $blocksize;
+	    $eof = read_with_alarm($readsize);
+	}
     }
 
     sub pass_n_line_records() {

src/parallel.pod

@@ -740,19 +740,25 @@ length of one record. For performance reasons I<size> should be bigger
 than a two records. GNU B<parallel> will warn you and automatically
 increase the size if you choose a I<size> that is too small.
 
-If you use B<-N>, B<--block> should be bigger than N+1 records.
-
 I<size> defaults to 1M.
 
-When using B<--pipe-part> a negative block size is not interpreted as a
-blocksize but as the number of blocks each jobslot should have. So
-this will run 10*5 = 50 jobs in total:
+A negative block size is not interpreted as a blocksize but as the
+number of blocks each jobslot should have. So B<--block -3> will make
+3 jobs for each jobslot. In other words: this will run 3*5 = 15 jobs
+in total:
 
-  parallel --pipe-part -a myfile --block -10 -j5 wc
+  parallel --pipe-part -a myfile --block -3 -j5 wc
+  cat myfile | parallel --pipe --block -3 -j5 wc
 
-This is an efficient alternative to B<--round-robin> because data is
-never read by GNU B<parallel>, but you can still have very few
-jobslots process large amounts of data.
+B<--pipe-part --block> is an efficient alternative to B<--round-robin>
+because data is never read by GNU B<parallel>, but you can still have
+very few jobslots process huge amounts of data.
+
+On the other hand, B<--pipe --block> is quite I<inefficient>: It reads
+the whole file into memory before splitting it. Thus input must be
+able to fit in memory.
+
+If you use B<--block> -I<size>, input should be bigger than I<size>+1 records.
 
 See also: UNIT PREFIX B<-N> B<--pipe> B<--pipe-part> B<--round-robin>
 B<--block-timeout>

src/parallel_alternatives.pod

@@ -173,11 +173,13 @@ The following features are in some of the comparable tools:
 
 =back
 
-As every new version of the programs are not tested the table may be
+Since each new version of the programs is not tested, the table may be
 outdated.  Please file a bug report if you find errors (See REPORTING
 BUGS).
 
-parallel:
+=head2 GNU Parallel
+
+Summary (see legend above):
 
 =over