[PATCH] D49700: [LLD] [COFF] Treat .xdata/.pdata$<sym> as implicitly associative to <sym> for MinGW

[Reid Kleckner via Phabricator via llvm-commits]

rnk added inline comments.


================
Comment at: COFF/InputFiles.cpp:263
 
+  std::map<StringRef, uint32_t> SymbolSectionIndexes;
   std::vector<const coff_aux_section_definition *> ComdatDefs(
----------------
Is there some way we can hash fewer symbols? It's unfortunate that we need a string hash table at all to parse object files. I wonder if there's a way we can defer this work, since for C++, a lot of comdat things end up getting discarded.

If we do need a string hash table, it should probably be a `DenseMap<StringRef, uint32_t>`, since this is performance critical.

Oh, here's an idea: maybe `createDefined` should return `Prevailing` somehow. Only add the symbol to the table if it's a prevailing comdat symbol in an executable section.


================
Comment at: test/COFF/associative-comdat-mingw.s:40-43
+        .section        .xdata$foo,"dr",associative,foo
+#        .linkonce       discard
+        .p2align        3
+        .long           42
----------------
I wonder if we could keep using associative comdats for .xdata and .pdata, but use the GCC-style section names. I wonder if ld.bfd would do the right thing with them. That would allow us to avoid the string hash table and save us a fair amount of link speed.


Repository:
  rLLD LLVM Linker

https://reviews.llvm.org/D49700