This is an archive of the discontinued LLVM Phabricator instance.

Add a compact dumping mode to `llvm-pdbdump`.
ClosedPublic

Authored by zturner on Jan 10 2017, 5:50 PM.

Details

Summary

This adds a command line option -compact-records. Currently this will only affect the behavior of tpi-records and -ipi-records but can eventually be updated to affect the behavior of symbol records as well.

Instead of dumping full record details, it outputs a format like this:

TPI Version: 20040203
Record count: 75
Records [
  Index: 0x1000 (8 bytes, offset 0) LF_ARGLIST "()"
  Index: 0x1001 (16 bytes, offset 8) LF_PROCEDURE "int ()"
  Index: 0x1002 (76 bytes, offset 24) LF_FIELDLIST "<field list>"
  Index: 0x1003 (120 bytes, offset 100) LF_ENUM "__vc_attributes::threadingAttribute::threading_e"
  Index: 0x1004 (100 bytes, offset 220) LF_STRUCTURE "__vc_attributes::threadingAttribute"
  Index: 0x1005 (12 bytes, offset 320) LF_POINTER "const __vc_attributes::threadingAttribute*"
  Index: 0x1006 (12 bytes, offset 332) LF_ARGLIST "(__vc_attributes::threadingAttribute::threading_e)"
  Index: 0x1007 (28 bytes, offset 344) LF_MFUNCTION "void __vc_attributes::threadingAttribute::(__vc_attributes::threadingAttribute::threading_e)"
  Index: 0x1008 (28 bytes, offset 372) LF_MFUNCTION "void __vc_attributes::threadingAttribute::()"
  Index: 0x1009 (20 bytes, offset 400) LF_METHODLIST ""
  Index: 0x100a (68 bytes, offset 420) LF_FIELDLIST "<field list>"
  Index: 0x100b (100 bytes, offset 488) LF_STRUCTURE "__vc_attributes::threadingAttribute"
  Index: 0x100c (48 bytes, offset 588) LF_FIELDLIST "<field list>"
  Index: 0x100d (120 bytes, offset 636) LF_ENUM "__vc_attributes::event_receiverAttribute::type_e"
  Index: 0x100e (112 bytes, offset 756) LF_STRUCTURE "__vc_attributes::event_receiverAttribute"
  Index: 0x100f (12 bytes, offset 868) LF_POINTER "const __vc_attributes::event_receiverAttribute*"
  Index: 0x1010 (16 bytes, offset 880) LF_ARGLIST "(__vc_attributes::event_receiverAttribute::type_e, bool)"
  Index: 0x1011 (28 bytes, offset 896) LF_MFUNCTION "void __vc_attributes::event_receiverAttribute::(__vc_attributes::event_receiverAttribute::type_e, bool)"
  Index: 0x1012 (12 bytes, offset 924) LF_ARGLIST "(__vc_attributes::event_receiverAttribute::type_e)"
  Index: 0x1013 (28 bytes, offset 936) LF_MFUNCTION "void __vc_attributes::event_receiverAttribute::(__vc_attributes::event_receiverAttribute::type_e)"
  Index: 0x1014 (28 bytes, offset 964) LF_MFUNCTION "void __vc_attributes::event_receiverAttribute::()"
  Index: 0x1015 (28 bytes, offset 992) LF_METHODLIST ""
  Index: 0x1016 (96 bytes, offset 1,020) LF_FIELDLIST "<field list>"
  Index: 0x1017 (112 bytes, offset 1,116) LF_STRUCTURE "__vc_attributes::event_receiverAttribute"

In addition, this patch prints out the type index offset array in the TPI / IPI streams. After printing all type records, you will get some output that looks like this:

  Index: 0x1691 (16 bytes, offset 153,924) LF_ARRAY ""
  Index: 0x1692 (720 bytes, offset 153,940) LF_FIELDLIST "<field list>"
  Index: 0x1693 (92 bytes, offset 154,660) LF_STRUCTURE "_IMAGE_LOAD_CONFIG_DIRECTORY32"
  Index: 0x1694 (7,484 bytes, offset 154,752) LF_FIELDLIST "<field list>"
  Index: 0x1695 (52 bytes, offset 162,236) LF_ENUM "CXCursorKind"
  Index: 0x1696 (7,660 bytes, offset 162,288) LF_FIELDLIST "<field list>"
  Index: 0x1697 (52 bytes, offset 169,948) LF_ENUM "CXCursorKind"
  Index: 0x1698 (7,716 bytes, offset 170,000) LF_FIELDLIST "<field list>"
  Index: 0x1699 (52 bytes, offset 177,716) LF_ENUM "CXCursorKind"
  TypeIndexOffsets [
    Index: 0x1000, Offset: 0
    Index: 0x103e, Offset: 9,408
    Index: 0x1093, Offset: 16,392
    Index: 0x1104, Offset: 24,624
    Index: 0x11b8, Offset: 32,808
    Index: 0x126a, Offset: 40,964
    Index: 0x133f, Offset: 52,388
    Index: 0x1343, Offset: 115,832
    Index: 0x1355, Offset: 123,100
    Index: 0x1431, Offset: 131,080
    Index: 0x1523, Offset: 141,164
    Index: 0x15c0, Offset: 147,472
    Index: 0x1695, Offset: 162,236
    Index: 0x1697, Offset: 169,948
    Index: 0x1699, Offset: 177,716
  ]
]

With these two options combined, I was able to decipher the way the Index / Offsets array is laid out. It is an array of pairs, sorted by type index, where each pair represents the Type Index of the first item in a "chunk" of records, and the offset at which that record begins. The number of records in chunk N is IndexOffsets[N+1].First - IndexOffsets[N].First, and the spacing is chosen such that each chunk contains the minimum number of consecutive records such that a chunk is >= 8KB.

There are some exceptions to this which I'm investigating. See for example Index 0x1695 and 0x1697 above, the distance between them is < 8KB. But it seems to hold true almost everywhere else, so it shouldn't be too hard to figure out the rest of this edge case.

Diff Detail

Event Timeline

zturner updated this revision to Diff 83906.Jan 10 2017, 5:50 PM
zturner retitled this revision from to Add a compact dumping mode to `llvm-pdbdump`..
zturner updated this object.
zturner added reviewers: amccarth, inglorion, ruiu, rnk.
zturner added a subscriber: llvm-commits.
ruiu added inline comments.Jan 10 2017, 5:56 PM
llvm/tools/llvm-pdbdump/LLVMOutputStyle.cpp
577

You have using namespace llvm.

llvm/tools/llvm-pdbdump/llvm-pdbdump.cpp
589

I think errs() is unbuffered, so you don't need this. Even if it is buffered, exit() should flush it.

This revision was automatically updated to reflect the committed changes.