From 58d90767821271f9546b8f34b7db85d896784941 Mon Sep 17 00:00:00 2001 From: Miguel Sofer Date: Thu, 16 Aug 2007 20:39:34 +0000 Subject: * generic/tclExecute.c: check the two most frequent instructions before the switch. Reduces both runtime and obj size a tiny bit. --- ChangeLog | 5 +++++ generic/tclExecute.c | 18 +++++++++++++++++- 2 files changed, 22 insertions(+), 1 deletion(-) diff --git a/ChangeLog b/ChangeLog index 47b9063..9aae28e 100644 --- a/ChangeLog +++ b/ChangeLog @@ -1,3 +1,8 @@ +2007-08-16 Miguel Sofer + + * generic/tclExecute.c: check the two most frequent instructions + before the switch. Reduces both runtime and obj size a tiny bit. + 2007-08-16 Don Porter * generic/tclCompExpr.c: Added a "constant" field to the OpNode diff --git a/generic/tclExecute.c b/generic/tclExecute.c index 43bd58a..5b3185b 100644 --- a/generic/tclExecute.c +++ b/generic/tclExecute.c @@ -12,7 +12,7 @@ * See the file "license.terms" for information on usage and redistribution of * this file, and for a DISCLAIMER OF ALL WARRANTIES. * - * RCS: @(#) $Id: tclExecute.c,v 1.314 2007/08/14 21:04:28 msofer Exp $ + * RCS: @(#) $Id: tclExecute.c,v 1.315 2007/08/16 20:39:34 msofer Exp $ */ #include "tclInt.h" @@ -1640,6 +1640,21 @@ TclExecuteByteCode( } } + /* + * These two instructions account for 26% of all instructions (according + * to measurements on tclbench by Ben Vitale + * [http://www.cs.toronto.edu/syslab/pubs/tcl2005-vitale-zaleski.pdf] + * Resolving them before the switch reduces the cost of branch + * mispredictions, seems to improve runtime by 5% to 15%, and (amazingly!) + * reduces total obj size. + */ + + if (*pc == INST_LOAD_SCALAR1) { + goto instLoadScalar1; + } else if (*pc == INST_PUSH1) { + goto instPush1Peephole; + } + switch (*pc) { case INST_RETURN_IMM: { int code = TclGetInt4AtPtr(pc+1); @@ -2257,6 +2272,7 @@ TclExecuteByteCode( Tcl_Obj *objPtr; case INST_LOAD_SCALAR1: + instLoadScalar1: opnd = TclGetUInt1AtPtr(pc+1); varPtr = &(compiledLocals[opnd]); while (TclIsVarLink(varPtr)) { -- cgit v0.12