> This patch seems to make rather too many assumptions about host and > target compilers. Certainly code like this can't go into > target-independent code like lto-wrapper. That's true. The point of this patch was to show what is needed to support x86->MIC OpenMP offloading, as we currently see it. We are ready to extend existing code making it more versatile, but keeping this needed functionality.
> Also, I'm not sure you can > assume you'll get ELF files out of the OpenACC target compiler; I'd very > prefer a solution that doesn't rely on objcopy. Yep, that's an issue I suspected but wasn't sure of. The idea was to prepare an image in the following steps: 1. Compile with target compiler. 2. Post-process it (in our case, call objcopy and perform partial linking). 3. Pass it to the host linker as a usual object file with the image and its size placed in the known symbols with defined names (e.g. _omp_<target-name>_image). I would like to keep step 3 unchanged, while the steps 1 and 2 could be easily combined into a single one. In that case we could say that output of a target compiler should be an object file for host linker with several symbols defined: _omp_<target-name>_image, _omp_<target-name>_size, etc. For x86->MIC one would use gcc+objcopy for this, and for OpenACC offloading one could use gcc+some of other target-specific utils (if any of them needed). Will this scheme work for OpenACC? Michael > Bernd
