SPEC2000: Comparison of LLVM-3.2 and GCC4.8 on x86

The goal of these runs is to compare the performance of the latest versions of LLVM and GCC on x86 in 32-bit mode. I used the following similar options:

-O1 -mtune=corei7 -march=i686, -O2 -mtune=corei7 -march=i686, -O3 -mtune=corei7 -march=i686, and -O4 -mtune=corei7 -march=i686 for LLVM
-O1 -mtune=corei7 -march=i686, -O2 -mtune=corei7 -march=i686, -O3 -mtune=corei7 -march=i686, and -Ofast -fno-fast-math -flto -fwhole-program -mtune=corei7 -march=i686 for GCC

GCC -Ofast means -ffast-math by default whereas LLVM -O4 still uses normal floating point math, therefore GCC in -Ofast ran with -fno-fast-math to make a correct comparison. For the same reason LLVM run with -march=i686 as by default it uses -march=x86-64 when it generates 32-bit code on x86-64 target.

LLVM can not produce a correct code for 254.gap on any optimization level. Therefore -O0 was used for both LLVM and GCC for 254.gap.

The used machine is 3.4Ghz Intel Core i7-2600 with 8GB memory under Fedora Core17.

GCC-4.8 was configured with --enable-checking=release. LLVM-3.2 with CLANG and Dragonegg for x86 were built in release mode. The sources are taken from LLVM download site. All changes are given relative to LLVM-3.2 run with -O1.

Compilation Speed
Code Size

Here the code size is the corresponding text and data segment size.
- SPECInt2000
- SPECFP2000
Generated Code Performace
- SPECInt2000
- SPECFP2000
  
  I removed data for 179.art as the results are varying very much depending on optimization modes.

Last modified: 02/04/2013 - vmakarov at redhat dot com

Return to index page.

The used machine is 3.4Ghz Intel Core i7-2600 with 8GB memory under Fedora Core17.

Compilation Speed

Code Size

SPECInt2000

SPECFP2000

Generated Code Performace

SPECInt2000

SPECFP2000