[project @ 2002-06-18 12:43:42 by simonpj]

Notes about par/seq

diff --git a/ghc/docs/comm/rts-libs/prelude.html b/ghc/docs/comm/rts-libs/prelude.html
index 72147ab..4ad6c20 100644 (file)
--- a/ghc/docs/comm/rts-libs/prelude.html
+++ b/ghc/docs/comm/rts-libs/prelude.html
@@ -12,6 +12,30 @@
       values, RULES pragmas, and so on) to make the heavily used Prelude code
       as fast as possible.
 
       values, RULES pragmas, and so on) to make the heavily used Prelude code
       as fast as possible.
 

+    <hr>
+    <h4>Par, seq, and lazy</h4>
+
+    In GHC.Conc you will dinf
+<blockquote><pre>
+  pseq a b = a `seq` lazy b
+</pre></blockquote>
+   What's this "lazy" thing.  Well, <tt>pseq</tt> is a <tt>seq</tt> for a parallel setting.
+   We really mean "evaluate a, then b".  But if the strictness analyser sees that pseq is strict
+   in b, then b might be evaluated <em>before</em> a, which is all wrong. 
+<p>
+Solution: wrap the 'b' in a call to <tt>GHC.Base.lazy</tt>.  This function is just the identity function,
+except that it's put into the built-in environment in MkId.lhs.  That is, the MkId.lhs defn over-rides the
+inlining and strictness information that comes in from GHC.Base.hi.  And that makes <tt>lazy</tt> look
+lazy, and have no inlining.  So the strictness analyser gets no traction.
+<p>
+In the worker/wrapper phase, after strictness analysis, <tt>lazy</tt> is "manually" inlined (see WorkWrap.lhs),
+so we get all the efficiency back.
+<p>
+This supersedes an earlier scheme involving an even grosser hack in which par# and seq# returned an
+Int#.  Now there is no seq# operator at all.
+
+
+    <hr>

     <h4>fold/build</h4>
     <p>
       There is a lot of magic in <a
     <h4>fold/build</h4>
     <p>
       There is a lot of magic in <a