[PATCH] D28423: Fix assertions on lazy-loading of Metadata TBAA attachments

[Mehdi AMINI via Phabricator via llvm-commits]

mehdi_amini added inline comments.


================
Comment at: llvm/lib/Bitcode/Reader/MetadataLoader.cpp:1613
           lazyLoadOneMetadata(Idx, Placeholders);
+          resolveForwardRefsAndPlaceholders(Placeholders);
+        }
----------------
tejohnson wrote:
> I suppose there is no loss in doing this more frequently since we know we need to resolve these anyway? Is there any disadvantage to doing the resolveForwardRefsAndPlaceholders more frequently? If so, should it only be done for TBAA attachments? If not, then should we be doing it more frequently elsewhere?
I'm looking into this, especially because since committing the lazy-loading I don't reproduce the improvements I had on my early version. I likely fixed a bug between my prototype and the final version.
I have >10% of the *total* link time spent in RAUW MDNodes.


https://reviews.llvm.org/D28423