On 10/13/2017 12:18 PM, Geoff Wozniak wrote: > My team and I are working on a procedural abstraction (PA) optimization > for GCC based on the ARM Embedded Toolchain distribution that uses GCC > 6.2.1. Our initial attempt to get something functional has been > successful, but it is a very messy solution because we don't see a > reasonable way to do an analysis of a compilation unit or whole program > at the level or RTL. From what we have been able to gleen (note: we're > all relative newcomers to the GCC code base), all RTL transformation and > analysis is done at the level of functions only. Once pass_final runs, > the RTL has been output to assembler, meaning no more changes to RTL is > possible. We have not been able to find a reasonable way to collect > all the RTL for a compilation unit to do an analysis across functions. > Any solution we have come up with involves some serious changes to > cgraphunit.c and final.c, among others. While we're willing to put in > the effort to make these changes, we're wondering if anyone knows of > something that would allow global RTL analysis without making > large-scale changes. It seems as though the structure of the pass > manager does not allow for this. The code structure suggests that we > should either do an IPA pass on GIMPLE or we work as an assembler pass, > thus forgoing work on RTL at all. Is this is what we should be doing? > If anyone has some guidance here, it would be appreciated. -- Geoff There are no real mechanisms for doing IPA optimizations or analysis on RTL.
If you could reformulate to work as an IPA pass on gimple that would be my recommendation. You could probably use pieces of the IPA-ICF infrastructure to help identify semantically equivalent blocks to stitch together and outline. Jeff
