gtbercea added a comment. In D60907#1473406 <https://reviews.llvm.org/D60907#1473406>, @Hahnfeld wrote:
> So the scheme is: `pow` is defined in `__clang_openmp_math.h` to call > `__kmpc_pow`. This lives in `libomptarget-nvptx` (both bc and static lib) and > just calls `pow` which works because `nvcc` and Clang in CUDA mode make sure > that the call gets routed into `libdevice`? > > Did you test that something like `pow(d, 2)` is optimized by LLVM to `d * d`? > There's a pass doing so (can't recall the name) and from my previous attempts > it didn't work well if you hid the function name instead of the known `pow` > one. The transformation was blocked because of a check in optimizePow() this was preventing pow(x,2) from becoming x*x. By adding the pow functions to the TLI the transformation now applies. This has now been fixed. SQRT is eliminated as per usual, no change for that. Repository: rC Clang CHANGES SINCE LAST ACTION https://reviews.llvm.org/D60907/new/ https://reviews.llvm.org/D60907 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits