[MLton-devel] cvs commit: self-compile improvements using the profiler

[Stephen Weeks]

sweeks      03/01/28 21:33:59

  Modified:    mlton/codegen/x86-codegen x86-allocate-registers.fun
  Log:
  This checkin fixes one recently introduced performance problem that
  has been especially noticeable on self compiles.  The problem was
  introduced on December 16, which was the same time as I was doing a
  large checkin of profiling code.  So, I had thought that the profiling
  stuff was responsible.  Turns out that had nothing to do with it.  The
  problem was due to a couple of innocuous-looking asserts that had been
  added -- the RegisterAllocation.toString call was costing quite a bit,
  and happened even though the asserts didn't.
  
  Before this checkin a self-compile had the following performance.
  	MLton finished in 445.61 + 374.75 (46% GC)
  	total allocated: 69,855,314,616 bytes
  
  And afterwards, here was the performance.
  	MLton finished in 318.05 + 199.32 (39% GC)
  	total allocated: 37,234,766,632 bytes
  
  Not yet down where we used to be (under 240s total), but a lot
  better.
  
  The new profiling tools were incredibly helpful in finding this
  problem, especially the new -graph option, which let me zero in on the
  relevant parts of the call graph.  The mlprof call that finally did
  the trick was:
  
  	mlprof -graph '(pred "x86-allocate-registers.fun: 794")' \
  		-show-line true \
  		mlton-compile mlmon.out
  
  This led me to the following graph, which once I looked at the code a
  little, helped me to see the problem.  Anyways, this has been a nice
  demonstration to me of the ease of use and utility of the profiling
  tools, even on a very large program.  It also makes me think there is
  a lot more that could be done, especially with the call graphs and
  program understanding and debugging.
  
  digraph "call-stack graph" {
  label = "call-stack graph"
  n0 [shape = "box", label = "chooseRegister\ncodegen/x86-codegen/x86-allocate-registers.fun: 1308\n0.0% 1.5% 0.0%\n"]
  n0 -> n1 []
  n2 [shape = "box", label = "spillAndReissue\ncodegen/x86-codegen/x86-allocate-registers.fun: 1257\n"]
  n2 -> n1 []
  n3 [shape = "box", label = "pre\ncodegen/x86-codegen/x86-allocate-registers.fun: 3830\n0.3% 24.4% 0.0%\n"]
  n3 -> n1 []
  n4 [shape = "box", label = "post\ncodegen/x86-codegen/x86-allocate-registers.fun: 4046\n0.3% 27.6% 0.0%\n"]
  n4 -> n1 []
  n1 [shape = "box", label = "toString\ncodegen/x86-codegen/x86-allocate-registers.fun: 794\n0.0% 44.0% 0.0%\n"]
  n5 [shape = "box", label = "codegen/x86-codegen/x86-allocate-registers.fun: 10747\n0.2% 60.6% 0.0%\n"]
  n5 -> n4 []
  n5 -> n1 []
  }

Revision  Changes    Path
1.27      +14 -6     mlton/mlton/codegen/x86-codegen/x86-allocate-registers.fun

Index: x86-allocate-registers.fun
===================================================================
RCS file: /cvsroot/mlton/mlton/mlton/codegen/x86-codegen/x86-allocate-registers.fun,v
retrieving revision 1.26
retrieving revision 1.27
diff -u -r1.26 -r1.27
--- x86-allocate-registers.fun	21 Jan 2003 22:56:03 -0000	1.26
+++ x86-allocate-registers.fun	29 Jan 2003 05:33:58 -0000	1.27
@@ -3839,9 +3839,13 @@
 	       registerAllocation: t}
 	= let
 	    val ra = registerAllocation 
-	    val _ = Assert.assert
-	            ("pre: " ^ (toString ra),
-		     fn () => unique ra)
+	    val _ =
+	       if true
+		  then ()
+	       else
+		  Assert.assert
+		  ("pre: " ^ (toString ra),
+		   fn () => unique ra)
 
 	    val dead_memlocs = dead
 	    val commit_memlocs = commit
@@ -4057,9 +4061,13 @@
 		registerAllocation: t}
 	= let 
 	    val ra = registerAllocation
-	    val _ = Assert.assert
-	            ("post: " ^ (toString ra),
-		     fn () => unique ra)
+	    val _ =
+	       if true
+		  then ()
+	       else
+		  Assert.assert
+		  ("post: " ^ (toString ra),
+		   fn () => unique ra)
 
 	    val (final_uses_registers,
 		 final_defs_registers,





-------------------------------------------------------
This SF.NET email is sponsored by:
SourceForge Enterprise Edition + IBM + LinuxWorld = Something 2 See!
http://www.vasoftware.com
_______________________________________________
MLton-devel mailing list
MLton-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/mlton-devel