The tramp3d-v4.cpp testcase (attached to PR23928) shows regressions in compile
time. For a first comparison consider compilation with leafify disabled and
-fno-exceptions -O2 -march=pentium4 -ffast-math -fpeel-loops
compile flags. All results with checking-disable and profile-bootstrapped
compilers from the HEAD of their branches. Run-time for 10 iterations is
included as a reference, too.
compile run
gcc-3.4 1m02s 25.9s
gcc-4.0 1m43s 15.2s
gcc-4.1 2m50s 8.8s
The following measurements are with leafification enabled (patched
3.4 and 4.0, -Dleafify=flatten for 4.1):
gcc-3.4 1m20s 10.5s
gcc-4.0 2m18s 8.6s
gcc-4.1 3m2s 8.5s
gcc-4.1+ 3m8s 6.5s
gcc-4.1+ is with the tree-forwprop and array-aliasing patches.
We see that even the fair comparison (same inlining via the
leafify patch) regresses considerably from 3.4 to 4.0 and 4.1.
Given the non-existant run-time improvement from 4.0 to 4.1
the regression there is even more disappointing.
Just for the complete report, here are some -O0 -fno-exception
numbers (w/o leafify):
3.4 22.3s
4.0 15.8s
4.1 16.8s
--
Summary: Compile time regressions with tramp3d
Product: gcc
Version: 4.1.0
Status: UNCONFIRMED
Keywords: compile-time-hog
Severity: normal
Priority: P2
Component: tree-optimization
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: rguenth at gcc dot gnu dot org
CC: gcc-bugs at gcc dot gnu dot org
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23955
