The main changes are in:
include/lld/Core/Reference.h
include/lld/ReaderWriter/Reader.h
Everything else is details to support the main change.
1) Registration based Readers
Previously, lld had a tangled interdependency with all the Readers. It would
have been impossible to make a streamlined linker (say for a JIT) which
just supported one file format and one architecture (no yaml, no archives, etc).
The old model also required a LinkingContext to read an object file, which
would have made .o inspection tools awkward.
The new model is that there is a global Registry object. You programmatically
register the Readers you want with the registry object. Whenever you need to
read/parse a file, you ask the registry to do it, and the registry tries each
registered reader.
For ease of use with the existing lld code base, there is one Registry
object inside the LinkingContext object.
2) Changing kind value to be a tuple
Beside Readers, the registry also keeps track of the mapping for Reference
Kind values to and from strings. Along with that, this patch also fixes
an ambiguity with the previous Reference::Kind values. The problem was that
we wanted to reuse existing relocation type values as Reference::Kind values.
But then how can the YAML write know how to convert a value to a string? The
fix is to change the 32-bit Reference::Kind into a tuple with an 8-bit namespace
(e.g. ELF, COFFF, etc), an 8-bit architecture (e.g. x86_64, PowerPC, etc), and
a 16-bit value. This tuple system allows conversion to and from strings with
no ambiguities.
llvm-svn: 197727
Executable files do not use a string table, so section names longer than 8
characters are not permitted. Long section names should just be truncated.
llvm-svn: 197470
If NONAME option is given for an export, that symbol will be exported only by
its ordinal. LLD will not emit the symbol name to the export table.
llvm-svn: 197371
OrdinalBase is an addend to the ordinals. We used to always set 1 to the field.
Although it produced a valid a DLL export table, it'd be a waste if the first
ordinal does not start with 1 -- we had to have NULL fields at the beginning of
the export address table. By setting the ordinal base, we can eliminate the
NULL fields.
llvm-svn: 197367
You can specify exported function's ordinal by /export:func,@<number> command
line option, but LLD ignored the option until now. This patch implements the
feature.
Ordinal is basically the index into the exported function address table. So,
for example, if /export:foo,@42 is specified, the linker writes foo's address
to 42th entry in the address table. Windows supports import-by-ordinal; you
can not only import a function by name, but by its ordinal. If you want to
allow your DLL users to import your functions by their ordinals, you need to
make sure that your functions are always exported with the same ordinals.
This is the feature for that situation.
llvm-svn: 197364
The following are the most significant peculiarities of MIPS target:
- MIPS ABI requires some special tags in the dynamic table.
- GOT consists of two parts local and global. The local part contains
entries refer locally visible symbols. The global part contains entries
refer global symbols.
- Entries in the .dynsym section which have corresponded entries in the
GOT should be:
* Emitted at the end of .dynsym section
* Sorted accordingly to theirs GOT counterparts
- There are "paired" relocations. One or more R_MIPS_HI16 and R_MIPS_GOT16
relocations should be followed by R_MIPS_LO16 relocation. To calculate
result of R_MIPS_HI16 and R_MIPS_GOT16 relocations we need to combine
addends from these relocations and paired R_MIPS_LO16 relocation.
The patch reviewed by Michael Spencer, Shankar Easwaran, Rui Ueyama.
http://llvm-reviews.chandlerc.com/D2156
llvm-svn: 197342
The only data in .edata whose length varies is the string. This patch moves
all the strings to the end of the section, so that 16-bit or 32-bit integers
are aligned on correct boundaries.
llvm-svn: 197213
This is the first patch to emit data for the DLL export table. The DLL export
table is the data used by the Windows loader to find the address of exported
function from DLL. With this patch, LLD is able to emit a valid DLL export
table which the Windows loader can interpret and load.
The data structure of the DLL export table is described in the Microsoft
PE/COFF Specification, section 5.3.
DLL support is not complete yet; the linker needs to emit an import library
for a DLL, otherwise the linker cannot link against the DLL. We also do not
support export-only-by-ordinal yet.
llvm-svn: 197212
DLLNameAtom is an atom whose content is a string. IdataAtom is not going to
be the only place we need such atom, so I want to generalize it.
llvm-svn: 197137
I'm planning to create a new pass for the DLL export table, and I want to use
the class both from IdataPass and the new pass, EdataPass. So move the class to
a common place.
llvm-svn: 197132
/DLLEXPORT is a command line option to export a symbol. __declspec(dllexport)
uses that to make the linker to export DLLExport'ed functions, by adding the
option to .drectve section.
This patch implements the parser of the command line option.
llvm-svn: 197122
Before this patch, we had the following class hierarchy.
Chunk -> AtomChunk -> SectionChunk -> GenericSectionChunk
-> BaseRelocChunk
-> HeaderChunk
Chunk represented the generic concept of contiguous range in an output
file. AtomChunk represented a chunk consists of atoms.
That class hierarchy had many issues: 1) BaseRelocChunk does not really
consist of atoms, so inheriting from AtomChunk was plainly wrong, and 2)
the hierarchy is unecessarily too deep.
This patch correct them. The new hierachy is shown below.
Chunk -> SectionChunk -> AtomChunk
-> BaseRelocChunk
-> HeaderChunk
In the new hierarchy, AtomChunk represents a chunk consists of atoms. Other
types of sections (currently only BaseRelocChunk) should inherit directly
from SectionChunk.
llvm-svn: 197038
This patch is to basically move the functionality to construct Data Directory
from IdataPass to WriterPECOFF.
Data Directory is a part of the PE/COFF header and contains the addresses of
the import tables.
We used to represent the link from Data Directory to the import tables as
relocation references. The idea behind it is that, because relocation
references are processed by the Writer, we wouldn't have to do anything special
to fill the addresses of the import tables. I thought that the addresses would
be set "automatically".
But it turned out that that design made the pass and the writer rather
complicated. In order to make relocation references between Data Directory to
the import tables, these data structures needed to be represented as Atom.
However, because Data Directory is not a section content but a part of the
PE/COFF header, it did not fit well as an Atom. So we ended up having
complicated code both in IdataPass and the writer.
This patch simplifies it.
One side effect of this patch is that we now have ".idata.a", ".idata.d" and
"idata.t" sections for the import address table, the import directory table,
and the import lookup table. The writer looks for the sections by name to find
the start addresses of the sections. We probably should have a better way to
find a specific atom from the core linking result, but currently using the
section name seems to be the easiest way to do that. The Windows loader do not
care about the import table's section layout.
llvm-svn: 197016
If section size is not multiple of 512, the writer added NULL bytes at the end
of it to make it so. That is not required by the PE/COFF spec, and the MSVC's
linker does not do that too. So we don't need to do that, too.
llvm-svn: 197002
Code to create COFF section header was scattered across many member functions
of SectionChunk. Consolidate it to a member function of SectionHeaderTableChunk.
llvm-svn: 196895
/ALTERNATENAME is a rarely-used, undocumented command line option that is
needed to link LLD for release build. It seems that the option is for defining
an weak alias; /alternatename:foo=bar defines weak symbol "foo" for "bar".
If "foo" is defined in an input file, it'll be linked normally and the command
line option will have no effect. If it's not defined, "foo" will be handled
as an alias for "bar".
This patch implements the parser for the option. The actual weak alias handling
will be implemented in a separate patch.
llvm-svn: 196743
Because compare() and its heper functions no longer have to be members of
LayoutPass class, we can remove it from the class. No functionality change.
llvm-svn: 196715
The comparator used in the layout pass has many calls of map::find(). Because
std::sort runs the comparator N*log2(N) times, the maps are looked up with the
same key again and again. The map lookup is not a very fast operation. It made
the pass slow.
This patch eliminates the duplicate map lookups using decorate-sort-undecorate
idiom. The pass used to take 1.1 seconds when linking LLD with LLD on Windows,
but it now takes only 0.3 seconds. Overall performance gain in that case is from
6.1 seconds to 5.2 seconds.
Differential Revision: http://llvm-reviews.chandlerc.com/D2358
llvm-svn: 196714
GroupedSectionsPass was a complicated pass. That pass's job was to reorder
atoms by section name, so that the atoms with the same section prefix will be
emitted consecutively to the executable. The pass added layout edges to atoms,
and let the layout pass to actually reorder them.
This patch simplifies the design by making GroupedSectionPass to directly
reorder atoms, rather than adding layout edges. This resembles ELF's
ArrayOrderPass.
This patch improves the performance of LLD; it used to take 7.1 seconds to
link LLD with LLD on my Macbook Pro, but it now takes 6.1 seconds.
llvm-svn: 196628
Currently we do not de-duplicate library files specified by /defaultlib option.
As a result, the same files are added multiple times to the input graph. In
particular, some popular files, such as kernel32.lib or oldnames.lib, are added
more than 10 times during linking of LLD. That makes the linker slower, as it
needs to parse the same file again and again.
This patch solves the issue by de-duplicating. The same file will be added only
once to the input graph. This patch improved the LLD linking time from 10.5
seconds to 7.7 seconds on my 4-core Core i7 Macbook Pro.
llvm-svn: 196504
Emitting idata atoms to their own section would make debugging easier.
The Windows loader do not really care about whether the DLL import table is
in .rdata or its own .idata section, so there is no change in functionality.
llvm-svn: 196458
If /functionpadmin is specified, the linker is supposed to make room at the
beginning of each function, so that self-modifying program would easily
hotpatch existing functions. Since I'm not sure if this feature is really used,
I'll make LLD to ignore the option for now.
llvm-svn: 196363
/DEBUG option is to make the linker to emit debug information to the resulting
executable. It's not for enable debugging of the linker itself.
llvm-svn: 196040
This is a patch to let the PECOFF writer to use the information passed
by the parser for /section option. The implementation of /section should
now be complete.
llvm-svn: 195893
/MERGE option is a bit complicated for many reasons. Firstly, it takes both
positive and negative arguments. That means we have to have one of three
distinctive values (set, clear or unchange) for each permission bit. In this
patch we represent the three values using two bitmasks.
Secondly, the permissions specified by the parameter is bitwise or-ed with the
default permissions of a section. There is an exception for that rule; if one
of READ, WRITE or EXECUTE bit is specified, unspecified bits need to be
cleared. (So if you specify only WRITE for example, the resulting section will
not have WRITE nor EXECUTE bits.)
Lastly, multiple /merge options are allowed.
llvm-svn: 195882
Atom ordinals are the indeces in a file. Currently the PECOFF reader assigns
ordinals for each section, so it's (incorrectly) assigning duplicate ordinals.
llvm-svn: 195852
Instead of having multiple SectionChunks for each section (.text, .data,
.rdata and .bss), we could have one chunk writer that can emit any sections.
This patch does that -- removing all section-sepcific chunk writers and
replace them with one "generic" writer.
This change should simplify the code because it eliminates similar-but-
slightly-different classes.
It also fixes an issue in the previous design. Before this patch, we could
emit only limited set of sections (i.e. .text, .data, .rdata and .bss). With
this patch, we can emit any sections.
llvm-svn: 195797
Nick Kledzik