Intel Sandy Bridge Shapes Up On GCC 4.7 Compiler
Phoronix: Intel Sandy Bridge Shapes Up On GCC 4.7 Compiler
Back in January I wrote about how open-source compilers are quickly maturing for Intel Sandy Bridge CPUs and offering early support for Intel Ivy Bridge and Intel Haswell processors. Both GCC and LLVM have been quick to take advantage of the new instruction set extensions and other capabilities of these latest -- and very impressive -- Intel processors. With the release of GCC 4.7 quickly approaching, here is an updated set of GNU Compiler Collection Fortran/C/C++ benchmarks from the Intel Core i7 3960X Sandy Bridge Extreme Edition test-bed.
Looks like you're printing the LDFLAGS rather than the C(XX)FLAGS
A slow time to compile is fine with me. I would rather the Linux kernel take all day to compile, but produce an incredibly fast binary. As long as there is an alternative mode that compiles a bit faster, for build farms and automated tests and things where performance of the built binary isn't critical, but you want to produce a "correct" binary that you can either debug or run tests on...
Kind of impressive that a compiler release makes this much of a difference. I wonder how much more performance you can squeeze out with LTO?
I suppose you know, but both PNG issues are still there (tiny unreadable text, and long lines being clipped).
It's amazing to find out that almost every new GCC release is worse than one of previous ones.
Please, always test how generic code behaves (-O3 -march=i686 -mtune=generic on x86 platform, or just -O3 on x64) - such a test shows universal architectural improvements, not just some fixes tailored for specific CPU architectures.
It can be worse in one or a couple particular tests, this is always the case with compilers given the complexity of optimization these days. However looking at these tests the only ones where it shows a worse results are in synthetic benchmarks of which particularly Himeno Pressure Cooker has shown a wide amount of weird results over different compiler toolchains for a long time (it used to be compiled in these tests with NO optimization options which in turn meant it favoured compilers which defaulted to doing optimization).
Originally Posted by birdie
I have very little interest in synthetic benchmarks as they generally do not reflect real-world performance, looking at the real-world benchmarks in this testsuite (C-Ray, Ogg, Flac, LZMA) then GCC 4.7 improves or remains the same (as it does on most synthetic aswell).
As for the timed compilation tests I wonder if Micheal really built the RC as 'release' as I seriously doubt 4.7 has gotten slower in compiling.
Tags for this Thread