cmark

diff --git a/Makefile b/Makefile
@@ -127,6 +127,7 @@ progit/progit.md: progit
 bench: $(BENCHFILE)
 	{ sudo renice 99 $$$$; \
 	  for x in `seq 1 $(NUMRUNS)` ; do \
+	  /usr/bin/env time -p ${PROG} </dev/null >/dev/null ; \
 	  /usr/bin/env time -p ${PROG} $< >/dev/null ; \
 		  done \
 	} 2>&1  | grep 'real' | awk '{print $$2}' | python3 'bench/stats.py'

diff --git a/bench/stats.py b/bench/stats.py
@@ -3,7 +3,15 @@
 import sys
 import statistics
 
-values = [ float(x) for x in sys.stdin.readlines()]
+def pairs(l, n):
+        return zip(*[l[i::n] for i in range(n)])
+
+# data comes in pairs:
+#    n - time for running the program with no input
+#    m - time for running it with the benchmark input
+# we measure (m - n)
+
+values = [ float(y) - float(x) for (x,y) in pairs(sys.stdin.readlines(),2)]
 
 print("mean = %.4f, median = %.4f, stdev = %.4f" %
     (statistics.mean(values), statistics.median(values),

diff --git a/benchmarks.md b/benchmarks.md
@@ -4,23 +4,30 @@ Some benchmarks, run on an ancient Thinkpad running Intel Core 2 Duo at 2GHz.
 
 |Implementation     |  Time (sec)| Factor  |
 |-------------------|-----------:|--------:|
-| Markdown.pl       | 2921.30    | 14606.5 |
-| kramdown          | 20.76      |   103.8 |
-| PHP markdown      | 20.475     |   102.4 |
-| lunamark          | 6.195      |    40.0 |
-| cheapskate        | 5.645      |    28.2 |
-| peg-markdown      | 5.330      |    26.7 |
-| **commonmark.js** | 2.730      |    13.6 |
-| marked            | 1.870      |     9.4 |
-| discount          | 1.660      |     8.3 |
-| **cmark**         | 0.320      |     1.6 |
+| Markdown.pl       | 2921.24    | 14606.2 |
+| PHP markdown      | 20.85      |   104.3 |
+| kramdown          | 20.83      |   104.1 |
+| lunamark          | 6.295      |    31.5 |
+| cheapskate        | 5.760      |    28.8 |
+| peg-markdown      | 5.450      |    27.3 |
+| **commonmark.js** | 2.675      |    13.4 |
+| marked            | 1.855      |     9.3 |
+| discount          | 1.705      |     8.5 |
+| **cmark**         | 0.315      |     1.6 |
 | sundown           | 0.200      |     1.0 |
 
+
 To run these benchmarks, use `make bench PROG=/path/to/program`.
 
 The input text is a 10MB Markdown file built by concatenating 20 copies
 of the Markdown source of the first edition of [*Pro
 Git*](https://github.com/progit/progit/tree/master/en) by Scott Chacon.
 
-`time` is used to measure execution speed.
+`time` is used to measure execution speed.  The reported
+time is the *difference* between the time to run the program
+with the benchmark input and the time to run it with no input.
+(This procedure ensures that implementation dynamic languages are
+not prenalized by startup time.) A median of ten runs is taken.  The
+process is reniced to a high priority so that the system doesn't
+interrupt runs.