frasercrmck wrote: > > > Seems reasonable, which architectures require this? I know that NVIDIA > > > deprecated the 32-bit `nvptx` target in CUDA 12 or something. > > > > > > I'm not an expert on CUDA but, AFAICT, even in 64-bit CUDA, certain > > pointers such as those pointing to shared memory are 32 bit, because the > > size of shared memory is somewhere in the kB range. This generates better > > code, fewer registers, etc. I'm not sure why the option isn't enabled by > > default, personally - it seems like `nvcc` is doing this by default. > > I was just playing with the option downstream and noticed this issue. > > I figured it was something like that, since it saves a register per address. > I don't know the history for why this isn't the default, it's pretty much > just a data layout modifier to state that certain address spaces are 32-bit. > Maybe @Artem-B or @jlebar can comment.
Just threw together a nonsensical example for godbolt: https://godbolt.org/z/bhdEhrxd7. Notice the `mov.u32 %r7, As`, etc. https://github.com/llvm/llvm-project/pull/111682 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
