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
YAML files tend to be very large compared to binary formats because of ASCII
format inefficiency. And the YAML reader consumes an excessively large amount
of memory when parsing a large file. It's very slow too.
For example, I observed that 6MB executable became 120MB YAML file, and the
YAML reader consumed more than 1.5GB memory to load it. The YAML reader even
caused OOM error on 32 bit, causing the entire process to fail.
This patch sets the limit on the YAML file size the linker will try to load in
the RoundTripYAML test as a safeguard.
llvm-svn: 194666
Enable this for the following flavors
a) core
b) gnu
c) darwin
Its disabled for the flavor PECOFF. Convenient markers are added with FIXME
comments in the Driver that would be removed and code removed from each flavor.
llvm-svn: 193585
Disable tests to be run with REQUIRES: disable. Note disable is not added to the
config by the test runner Mkaefiles, so essentially disables the test.
Code changes would be required to fix these tests :-
test/darwin/hello-world.objtxt
test/elf/check.test
test/elf/phdr.test
test/elf/ppc.test
test/elf/undef-from-main-dso.test
test/elf/X86_64/note-sections-ro_plus_rw.test
test/pecoff/alignment.test
test/pecoff/base-reloc.test
test/pecoff/bss-section.test
test/pecoff/drectve.test
test/pecoff/dynamic.test
test/pecoff/dynamicbase.test
test/pecoff/entry.test
test/pecoff/hello.test
test/pecoff/imagebase.test
test/pecoff/importlib.test
test/pecoff/lib.test
test/pecoff/multi.test
test/pecoff/reloc.test
test/pecoff/weak-external.test
llvm-svn: 193300
This patch fixes a bug in r190608. The results of a comparison function
passed to std::sort must be transitive, which is, if a < b and b < c, and if
a != b, a < c must be also true. CompareAtoms::compare did not actually
guarantee the transitivity. As a result the sort results were sometimes just
wrong.
Consider there are three atoms, X, Y, and Z, whose file ordinals are 1, 2, 3,
respectively. Z has a property "layout-after X". In this case, all the
following conditionals become true:
X < Y because X's ordinal is less than Y's
Y < Z because Y's ordinal is less than Z's
Z < X because of the layout-after relationship
This is not of course transitive. The reason why this happened is because
we used follow-on relationships for comparison if two atoms falls in the same
follow-on chain, but we used each atom's properties if they did not. This patch
fixes the issue by using follow-on root atoms for comparison to get consistent
results.
Differential Revision: http://llvm-reviews.chandlerc.com/D1980
llvm-svn: 193029
Instead of showing multiple lines of debug messages, show only one message
by CompareAtoms::operator(). Here is an example.
Before:
Sorting _main .text
Sorting by sectionPos(2,2)
Sorting by override
Sorting _main .text
Sorting by sectionPos(2,2)
Sorting by override
After:
Layout: '_main' > '.text' (override (1, 0))
Layout: '_main' > '.text' (override (1, 0))
Differential Revision: http://llvm-reviews.chandlerc.com/D1964
llvm-svn: 192941
Found this with asan. Code assumes that find doesn't return end, thus if
both atoms didn't have followon roots it would still compare their positions.
llvm-svn: 191865
We need to order atoms that exist in the same chain. This is to make sure that
the command line order is preserved when we emit the atoms to the output file.
Credits: BigCheese for finding the bug.
Adds a test which otherwise would fail.
llvm-svn: 190608
Found that having a method to check the strucutre of the followon graph makes
it easy to debug file readers. The method checks if there's no wrong edge in
followOnNexts and followOnRoots. It is called only when debuggging is enabled
for LayoutPass.
Reviewers: shankarke
CC: llvm-commits
Differential Revision: http://llvm-reviews.chandlerc.com/D922
llvm-svn: 183553
DEBUG_TYPE is already defined at the beginning of this file. We don't
want to have two different debug flags for a single pass.
Reviewers: shankarke
CC: llvm-commits
Differential Revision: http://llvm-reviews.chandlerc.com/D845
llvm-svn: 182543
be laid out by their ordinal overrides first, there was a bug that two atoms may
get the same override index due to which atoms were not ordered properly. This
commit fixes the problem.
Now the atoms are ordered by
- Section Position hints
- Atom override (Using layout-after/layout-before/in-group)
- Content Permissions
- Content Type
- File Ordinal
This also fixes the problem of running c++ static executables that
was broken by an earlier patch.
llvm-svn: 182494
Summary:
Split buildFollowOnTable to small functions to improve
code readability and remove code duplication. No change
in functionality.
CC: llvm-commits
Differential Revision: http://llvm-reviews.chandlerc.com/D790
llvm-svn: 181749
Layoutpass by ordering atoms if they appear in the override
list first and then looking at the way of ordering atoms in
the default way.
The fix also fixes issues with the sizes of the sections,
that appear in the output properly too.
The commit also adds a testcase(orderatoms-by-override.test)
to test it and fixes all the other relevant testcases.
llvm-svn: 181605
The major changes are:
1) LinkerOptions has been merged into TargetInfo
2) LinkerInvocation has been merged into Driver
3) Drivers no longer convert arguments into an intermediate (core) argument
list, but instead create a TargetInfo object and call setter methods on
it. This is only how in-process linking would work. That is, you can
programmatically set up a TargetInfo object which controls the linking.
4) Lots of tweaks to test suite to work with driver changes
5) Add the DarwinDriver
6) I heavily doxygen commented TargetInfo.h
Things to do after this patch is committed:
a) Consider renaming TargetInfo, given its new roll.
b) Consider pulling the list of input files out of TargetInfo. This will
enable in-process clients to create one TargetInfo the re-use it with
different input file lists.
c) Work out a way for Drivers to format the warnings and error done in
core linking.
llvm-svn: 178776
Rui Ueyama