Recommend "icc -fp-model precise" if doing "make check".

Without this we get 24 test failures, as the default -fp-model is a
bit fast and loose.  (Ideally our tests wouldn't object so much about
minor fluctuations in the least significant digits, as that's not
really the thing we're trying to validate.)

Author: jkbonfield

INSTALL

@@ -197,6 +197,16 @@ compiler or SSL implementation.  It is assumed you already have the core set
 of packages for the given distribution - the lists may be incomplete if
 this is not the case.
 
+Intel Compiler
+--------------
+
+At some optimisation levels the Intel compiler may use approximate
+floating point mathematics.  We do not believe this to produce
+substantially erroneous results, but it can cause the test harness to
+fail due to minor fluctuations in the least significiant digits.  If
+you wish to use icc with the test harness, it is recommended you use
+the "-fp-model precise" icc option.
+
 Debian / Ubuntu
 ---------------