== Progress == * EEMBC - Ran on multiple boards, collected statistical data, wrote some scripts to do it again - Found a few regressions on GCC 4.9, reported internally - Found 4 benchmarks in which LLVM is a lot worse - Looking at RGBCMY which GCC is able to vectorise - LLVM transforms the pointer access into a PHI cycle, which is wrongly interpreted as a reduction variable by the loop vectorizer.
* Background - Patch reviews and discussions as usual - TCWG rack has arrived! - Got another Chromebook and an ODroid XU to play with next week == Plan == * Update script to rank the compilers according to EEMBC results on the same machine * Try to teach the loop vectorizer to not treat load/store PHIs as a reduction variable * Set up some temporary buildbots with new hardware
_______________________________________________ linaro-toolchain mailing list linaro-toolchain@lists.linaro.org http://lists.linaro.org/mailman/listinfo/linaro-toolchain
