LLVM gcc

[Guetti]

So, I tested it with my engine (bitboard based, 64-bit compiles, perft 6).

g++-4.2: NPS = 8079kNPS (-DNDEBUG -Wall -mcpu=G5 -mtune=G5 -O3 -fast)
g++-4.2: NPS = 7335kNPS (-DNDEBUG -Wall -arch ppc64 -O3)
llvm-g++: NPS = 6717kNPS (-DNDEBUG -Wall -arch ppc64 -O3)

So, llvm-g++ is a tad slower than g++-4.2 on 64-bit, and a good bit slower that g++4.2 using CPU specific optimizations. Using 32-bit compiles they were pretty much the same:

g++-4.2: NPS = 5214kNPS (-DNDEBUG -Wall -arch ppc -O3)
llvm-g++: NPS = 4987kNPS (-DNDEBUG -Wall -arch ppc -O3)
llvm-g++: NPS = 6032kNPS (-DNDEBUG -Wall -arch ppc -O4)

With Link-Time Optimization (use -O4 command-line option) llvm-g++ even smoked g++ on 32-bit. Unfortunately, I couldn't get Link-Time Optimization to work on 64-bit, I always got the error (invalid .machine argument: ppc970).

[krazyken]

From the Release notes

These are GCC 4.2 features that are not available in LLVM-GCC 4.2:

Profile-guided optimization
Auto vectorization
Stack canaries
OpenMP
Inline assembly that uses x87 registers
Generating debug information for optimized code

[Ron Murawski]

From the Release notes

These are GCC 4.2 features that are not available in LLVM-GCC 4.2:

Profile-guided optimization
Auto vectorization
Stack canaries
OpenMP
Inline assembly that uses x87 registers
Generating debug information for optimized code

The lack of PGO breaks it for serious chess engine use. I like to play around with "little languages" so LLVM still seems good.

Thanks for the information, Kenny.

Ron