Index: htdocs/projects/ast-optimizer.html
===================================================================
RCS file: ast-optimizer.html
diff -N ast-optimizer.html
*** /dev/null	Tue May  5 13:32:27 1998
--- ast-optimizer.html	Sun Jul 22 07:10:43 2001
***************
*** 0 ****
--- 1,124 ----
+ <html>
+ <head>
+ <title>Abstract Syntax Tree Optimizations</title>
+ </head>
+ 
+ <body>
+ <h1 align=center>Abstract Syntax Tree Optimizations</h1>
+ 
+ <p>This page describes ongoing work to improve GCC's tree based
+ optimizations. There is a branch in CVS called,
+ <code>ast-optimizer-branch</code>, which is available to experiment with
+ these optimizers. As these stabilize, they can be submitted for review
+ onto the mainline development tree. Please contact <a
+ href="mailto:nathan@codesourcery.com">Nathan Sidwell,
+ nathan@codesourcery.com</a>, if you want to contribute.
+ 
+ <h2>Background &amp; Rationale</h2>
+ 
+ <p>GCC, in common with many other compilers, has more than one internal
+ representation of a program. The main ones are trees and RTL. The
+ trees, (or formally abstract syntax trees - ASTs) are generated during
+ parsing, and are close to the source language semantics. The RTL is
+ generated in the back end, and is close to the generated assembly
+ code. Ideally, the AST would contain all the semantic information of
+ the source program.
+ 
+ <p>Historically, GCC generated RTL one statement at a time, so
+ the AST did not stay around very long. This has changed with 'function
+ at a time' compilation (<a href="/news/inlining.html">Inliner</a>),
+ which both C and C++ frontends now implement. With the AST for complete
+ functions, and the additional semantic information they contain, the
+ opportunity for new optimizations presents itself.
+ 
+ <h2>Goals</h2>
+ 
+ <p>In addition to providing completely new optimization passes, there
+ are a number of sub goals.
+ 
+ <ul>
+ 
+ <li><h3>Provide a testing framework</h3>
+ 
+ <p>Although GCC's testsuite can verify an optimization's correctness,
+ there is no framework to verify the efficiency. Efficiency in both
+ compiler resources (memory and compile time), and produced executable
+ (object size and execution speed) both need to be measured. Without a
+ test framework, we will have no information about the effectiveness
+ and stability of optimizers.
+ <p>
+ 
+ <li><h3>Optimization tuning parameters</h3>
+ 
+ <p>A number of optimizations can be tuned by various parameters. Giving
+ the user a knob to twiddle is good, provided (a) it does something (b)
+ there is a measureable way of determining its effectiveness.
+ <p>
+ 
+ <li><h3>Move RTL optimizations to the AST level</h3>
+ 
+ <p>GCC has many optimizations that work at the RTL level. At the AST
+ level some of these can do a better job.
+ <p>
+ 
+ <li><h3>Move AST optimizations into one place</h3>
+ 
+ <p>GCC has a number of AST optimizations that attempt to optimize trees
+ during parsing. These have limited effectiveness, and complicate the
+ parser. It would be better to put these all in one place, where they can
+ be more effective (by being repeated, or using common data).
+ <p>
+ 
+ <li><h3>Move AST optimizers into the common middle end</h3>
+ 
+ <p>Although C and C++ both have function at a time mode, the AST new
+ inliner is only in the C++ frontend.
+ <p>
+ 
+ <li><h3>Provable optimizer performance</h3>
+ 
+ <p>It is often easy to implement a mis-optimization. Without evidence
+ that an optimizer actually does optimize, it will be hard to have it
+ accepted into the mainline. Obviously in the
+ <code>ast-optimizer-branch</code> branch, things are different.
+ <p>
+ 
+ </ul>
+ 
+ <h2>Status</h2>
+ 
+ The branch was created from the development mainline on 21 July 2001.
+ Its version string is 'ast-optimizer-branch'.
+ 
+ <ul>
+ 
+ <li><h3>New inlining algorithm</h3>
+ 
+ <p>The first AST inliner was top down, and exhibits problematic
+ compiler time &amp; memory usage at -O3. This is particularly nasty for
+ heavily templated C++ code. A new bottom up inliner is available for
+ testing, which it is hoped will provide better -O3 performance.
+ See <a href="/ml/gcc-patches/2001-07/msg00859.html">this thread</a>.
+ <p>
+ 
+ </ul>
+ 
+ <h2>Contributing</h2>
+ 
+ <p>Checkout the <code>ast-optimizer-branch</code> branch
+ 
+ <pre>
+ cvs co -r ast-optimizer-branch gcc
+ </pre>
+ 
+ then configure and build in the normal way.
+ 
+ <p>Please post patches in the usual way to the gcc-patches list, marked
+ <code>[ast-optimizer-branch]</code>. As this is a branch, and not the
+ mainline, the usual maintainer rules do not apply. This branch is
+ maintained by <a href="mailto:nathan@codesourcery.com">Nathan Sidwell,
+ nathan@codesourcery.com</a>. Approval from the usual maintainer will be
+ needed when submitting patches from the branch onto the mainline.
+ 
+ </body>
+ </html>
Index: htdocs/projects/index.html
===================================================================
RCS file: /cvs/gcc/wwwdocs/htdocs/projects/index.html,v
retrieving revision 1.15
diff -c -3 -p -r1.15 index.html
*** index.html	2001/07/19 05:13:12	1.15
--- index.html	2001/07/22 14:11:05
***************
*** 19,24 ****
--- 19,25 ----
  <li><a href="#ssa_conditional_constant_propagation_pass">SSA Conditional Constant Propagation Pass</a>
  <li><a href="#value_range_propagation_pass">Value range propagation pass</a>
  <li><a href="#automaton_based_pipeline_hazard_recognizer">Automaton based pipeline hazard recognizer</a>
+ <li><a href="#ast_optimizer">Tree optimization passes</a>
  </ul>
  <li><a href="#optimizer_inadequacies">Optimizer inadequacies</a>
  <li><a href="#ia64 projects">Projects to improve performance on IA-64</a>
*************** automaton based pipeline hazard recogniz
*** 136,142 ****
  instruction scheduling and it provides the base for a future sofware
  pipelining implementation.</p>
  
! 
  <h2><a name="optimizer_inadequacies">Optimizer inadequacies</h2>
  <p>We also have a page detailing <a href="optimize.html">optimizer
  inadequacies</a>, if you'd prefer to think about it in terms of problems
--- 137,146 ----
  instruction scheduling and it provides the base for a future sofware
  pipelining implementation.</p>
  
! <h3><a name="ast_optimizer">Tree optimization passes</a></h3>
! A separate page describes <a href="ast-optimizer.html">abstract
! syntax tree optimization passes</a> that are being worked on.
!   
  <h2><a name="optimizer_inadequacies">Optimizer inadequacies</h2>
  <p>We also have a page detailing <a href="optimize.html">optimizer
  inadequacies</a>, if you'd prefer to think about it in terms of problems
