GCC 3.3 compile speed regression - AN ANSWER

[Matt Austern]

On Tuesday, February 11, 2003, at 11:52 PM, Joseph S. Myers wrote:

> On Tue, 11 Feb 2003, Matt Austern wrote:
>
>> All of my measurements suggest that improving the speed of 
>> optimization
>> phases is irrelevant.  What's slow is the front end: parsing and
>> semantic analysis.  -O0 compilations are slow, not just -O3.
>
> The -fsyntax-only figures posted
> <http://gcc.gnu.org/ml/gcc/2003-02/msg00488.html> show it is not the 
> front
> end; -O0 -fsyntax-only has become faster, while the time for 
> unoptimized
> code generation (the time for -O0 minus that for -O0 -fsyntax-only) has
> doubled between 2.95 branch and mainline and the times to optimize have
> become worse as well.

We're obviously testing different things.  For the kinds of compilations
where Apple is most concerned about speed, the back end is so far down
in the noise that using -fsyntax-only makes no important difference.  
Here's
an example:

bash-2.05$ ~/root33/bin/g++ -S -ftime-report  -O0  -DQT_NO_DEBUG 
-DQT_SHARED -I/home/austern/qt-x11-free-3.1.1/mkspecs/linux-g++ -I. 
-I/usr/include/freetype2 -I../../../include -I/usr/X11R6/include 
-I/usr/X11R6/include -I.moc/release-shared/  structureparser.cpp

Execution times (seconds)
  garbage collection    :   1.30 (24%) usr   0.00 ( 0%) sys   1.38 (21%) 
wall
  cfg cleanup           :   0.02 ( 0%) usr   0.00 ( 0%) sys   0.00 ( 0%) 
wall
  trivially dead code   :   0.01 ( 0%) usr   0.00 ( 0%) sys   0.00 ( 0%) 
wall
  preprocessing         :   0.35 ( 6%) usr   0.21 (18%) sys   0.38 ( 6%) 
wall
  lexical analysis      :   0.49 ( 9%) usr   0.20 (17%) sys   0.75 (11%) 
wall
  parser                :   2.30 (42%) usr   0.31 (26%) sys   2.50 (38%) 
wall
  name lookup           :   0.91 (17%) usr   0.43 (37%) sys   1.38 (21%) 
wall
  varconst              :   0.02 ( 0%) usr   0.00 ( 0%) sys   0.00 ( 0%) 
wall
  branch prediction     :   0.00 ( 0%) usr   0.01 ( 1%) sys   0.00 ( 0%) 
wall
  local alloc           :   0.02 ( 0%) usr   0.00 ( 0%) sys   0.25 ( 4%) 
wall
  global alloc          :   0.03 ( 1%) usr   0.00 ( 0%) sys   0.00 ( 0%) 
wall
  symout                :   0.00 ( 0%) usr   0.01 ( 1%) sys   0.00 ( 0%) 
wall
  TOTAL                 :   5.47             1.17             6.62
bash-2.05$ ~/root33/bin/g++ -S -ftime-report -fsyntax-only  -O0  
-DQT_NO_DEBUG -DQT_SHARED 
-I/home/austern/qt-x11-free-3.1.1/mkspecs/linux-g++ -I. 
-I/usr/include/freetype2 -I../../../include -I/usr/X11R6/include 
-I/usr/X11R6/include -I.moc/release-shared/  structureparser.cpp

Execution times (seconds)
  garbage collection    :   1.17 (21%) usr   0.01 ( 1%) sys   1.12 (16%) 
wall
  preprocessing         :   0.37 ( 7%) usr   0.13 (10%) sys   0.75 (11%) 
wall
  lexical analysis      :   0.33 ( 6%) usr   0.28 (21%) sys   0.75 (11%) 
wall
  parser                :   2.40 (42%) usr   0.27 (20%) sys   2.88 (40%) 
wall
  name lookup           :   1.04 (18%) usr   0.65 (48%) sys   1.25 (18%) 
wall
  varconst              :   0.37 ( 7%) usr   0.02 ( 1%) sys   0.38 ( 5%) 
wall
  TOTAL                 :   5.69             1.36             7.12

In this example the -fsyntax-only run happens to be slower.  That's not
always the case, but it illustrates my point: for this -O0 compilation, 
the
time taken by the back end is smaller than the measurement error.

			--Matt