As far as I know, adding the support for restrict didn’t trigger any bugs in GCC
That’s very impressive for gcc. IIRC adding restrict to LLVM triiggered major bugs and miscompilations at least for the first two attempts. As they said they need to do a crater run to be sure, but even passing the initial smoke test is an achievement for gcc.
However, I’m surprised the code is “only” 3% faster using restric annotation. IIRC the speed-ups were about 5% for LLVM so maybe there is still some performance to gain on the gcc side?
To be fair, I heard that some bugs for restrict fixed in LLVM were also fixed in GCC as a result of finding them using the LLVM codegen of rustc.
For the 3%, this is in only one particular benchmark. Since I was curious, I compiled my test program with the LLVM codegen and disabled noalias: there’s also a 3% performance improvement.
That’s very impressive for gcc. IIRC adding restrict to LLVM triiggered major bugs and miscompilations at least for the first two attempts. As they said they need to do a crater run to be sure, but even passing the initial smoke test is an achievement for gcc.
However, I’m surprised the code is “only” 3% faster using restric annotation. IIRC the speed-ups were about 5% for LLVM so maybe there is still some performance to gain on the gcc side?
To be fair, I heard that some bugs for
restrictfixed in LLVM were also fixed in GCC as a result of finding them using the LLVM codegen of rustc.For the 3%, this is in only one particular benchmark. Since I was curious, I compiled my test program with the LLVM codegen and disabled
noalias: there’s also a 3% performance improvement.