[LLVMbugs] [Bug 507] NEW: [archive reading] Dramatically speed up reading archives without a symtab

[bugzilla-daemon at cs.uiuc.edu]

http://llvm.cs.uiuc.edu/bugs/show_bug.cgi?id=507

           Summary: [archive reading] Dramatically speed up reading archives
                    without a symtab
           Product: libraries
           Version: 1.4
          Platform: All
        OS/Version: All
            Status: NEW
          Severity: enhancement
          Priority: P2
         Component: Archive library
        AssignedTo: unassignedbugs at nondot.org
        ReportedBy: sabre at nondot.org


The code in Archive::findModulesDefiningSymbols that handles building a symbol
table for an archive without one is incredibly inefficient.  There are two
reasons for this:

1. The GetBytecodeSymbols (exported by the Bytecode Reader) function used by
   this method does not take a ModuleProvider, it takes either a filename or
   a byte buffer.

   The problem is that this causes findModulesDefiningSymbols to parse the
   module from the archive file just to find its symbols.... then
   immediately throw it away.  It would be much better to use the
   ModuleProvider that the archive already has, load into it, which will
   avoid the reload of the data when needed later.

2. We should be able to entirely avoid reading the function bodies to build
   the symbol table (which would speed up both llvm-ar/ranlib and
   gccld/llvm-ld).  To do this, we have to rearrange the bytecode slightly, so
   that we know what linkage the functions will have without having to read
   the function body (like globals, the linkage should be stored in the
   ModuleGlobalInfo table).

Both of these changes will speed up linking a lot by reducing IO and by not
building large data structures that are never used.

-Chris



------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.