[PATCH] D99487: [CodeGen] Port basic block sections from ELF to COFF

Reid Kleckner via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Fri May 7 12:49:01 PDT 2021


rnk added inline comments.


================
Comment at: llvm/lib/CodeGen/TargetLoweringObjectFileImpl.cpp:1712
+          COFF::IMAGE_SCN_MEM_READ | COFF::IMAGE_SCN_LNK_COMDAT,
+      SectionKind::getText(), COMDATSymName,
+      COFF::IMAGE_COMDAT_SELECT_NODUPLICATES, UniqueID);
----------------
mstorsjo wrote:
> TaoPan wrote:
> > rnk wrote:
> > > tmsriram wrote:
> > > > MaskRay wrote:
> > > > > TaoPan wrote:
> > > > > > tmsriram wrote:
> > > > > > > COMDATSymName can be folded to be equal to MBB.getSymbol()->getName() here?  Plus, you are not preserving the .text.hot prefix that the original function might get here.  Is this future work?  If the original function is named .text.hot.foo, the basic block will still be named .text.foo.__part.1 which is not right.
> > > > > > > 
> > > > > > > Plus, what about exception handling sections like ".eh.*"?
> > > > > > Thanks! I'll redesign section name and comdat symbol name.
> > > > > > The text section with prefix "hot" and "unlikely" won't be constructed here, I added COFF text section prefix "hot" and "unlikely" in D92073. In ELF override function, also not handling text section with prefix "hot" and "unlikely".
> > > > > > The text section with prefix "split" will be constructed here, I plan to add related code in MFS COFF patch.
> > > > > > Also, exception handling section is future work that support basic block sections Windows COFF exception handling.
> > > > > This is complex. PE-COFF has multiple COMDAT seletion kinds. I want to see a holistic plan how every component is going to be implemented.
> > > > The basic block should just mimic the COMDAT type of its containing function, is there a reason to do anything more with it here?
> > > After thinking about it a bit, I think the entry block should use the regular selection kind, and all of the auxilliary MBB sections should use IMAGE_COMDAT_SELECT_ASSOCIATIVE. They should be associated with the main function symbol, unless the function itself is associated with something else, in which case the BBs use that symbol.
> > > 
> > > This will ensure that if the main function section prevails, they are included, and if it does not prevail, they are discarded. Does that make sense?
> > Thanks! I think set entry block sections as regular IMAGE_COMDAT_SELECT_NODUPLICATES and set the auxilliary MBB sections as IMAGE_COMDAT_SELECT_ASSOCIATIVE is fine. I changed.
> @rnk - I'm not quite familiar with the concrete implications of this work here, but would this need to be mapped to the GNU style comdats, for compatibility with ld.bfd for mingw targets? (LLD itself works fine with proper associative comdats though.)
I think basic block sections are kind of in the same category as LTO: it's a very sophisticated optimization feature, and it's probably OK if it's LLD-only. I think it also requires special linker support that might make it LLD-only, but I've forgotten the details.

It also has potentially very high object file size overhead, and it will be important to do everything possible to minimize that object file size overhead. Long gnu-style section names for every basic block section are likely to make the feature unusably slow, so maybe it's worth removing these "if mingw" conditionals. We can leave behind a comment by the use of IMAGE_COMDAT_SELECT_ASSOCIATIVE indicating that gnu-style section names are not needed because the feature is only designed to work with LLD.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D99487/new/

https://reviews.llvm.org/D99487



More information about the llvm-commits mailing list