There's actually a room to improve this patch. Instead of not merging
sections that have different alignements, we can choose the section that
has the largest alignment requirement among all sections that are otherwise
considered the same. Then all section alignments are satisfied, so we can
merge them.
I don't know if that improvement could make any difference for real-world
input, so I'll leave it alone. Would be interesting to revisit later.
llvm-svn: 248581
This is an LLD extension to MSVC link.exe command line. MSVC linker
does not write symbol tables for executables. We do unless no /debug
option is given.
There's a situation that we want to enable debug info but don't want
to emit the symbol table. One example is when we are comparing output
file size. With this patch, you can tell the linker to not create
a symbol table by just specifying /nosymtab.
llvm-svn: 248225
std::distance(C->Relocs.end(), C->Relocs.begin()) is the same as NumRelocs
which is already added to the hash value. What we are missing here is the
section size.
llvm-svn: 248202
This patch fixes a regression introduced by r247964. Relocations that
are referring the same symbol should be considered equal, but they
were not if they were pointing to non-section chunks.
llvm-svn: 248132
Previously, InputFile::parse() was run in batch. We construct a list
of all input files and call parse() on each file using parallel_for_each.
That means we cannot start parsing files until we get a complete list
of input files, although InputFile::parse() is safe to call from anywhere.
This patch makes it asynchronous. As soon as we add a file to the symbol
table, we now start parsing the file using std::async().
This change shortens self-hosting time (650 ms) by 28 ms. It's about 4%
improvement.
llvm-svn: 248109
I made the field an atomic pointer in hope that we would be able to
parallelize the symbol resolver soon, but that's not going to happen
soon. This patch reverts that change for the sake of readability.
llvm-svn: 248104
InputFile::parse() can be called in parallel with other calls of
the same function. By doing that, time to self-link improves from
741 ms to 654 ms or 12% faster.
This is probably the last low hanging fruit in terms of parallelism.
Input file parsing and symbol table insertion takes 450 ms in total.
If we want to optimize further, we probably have to parallelize
symbol table insertion using concurrent hashmap or something.
That's doable, but that's not easy, especially if you want to keep
the exact same semantics and linking order. I'm not going to do that
at least soon.
Anyway, compared to r248019 (the change before the first attempt for
parallelism), we achieved 36% performance improvement from 1022 ms
to 654 ms. MSVC linker takes 3.3 seconds to link the same program.
MSVC's ICF feature is very slow for some reason, but even if we
disable the feature, it still takes about 1.2 seconds.
Our number is probably good enough.
llvm-svn: 248078
Self-hosting took 801 ms on my machine. Of which this function took
69 ms. Now it takes 37 ms. That is about 4% overall performance
improvement.
llvm-svn: 248052
The LLD's ICF algorithm is highly parallelizable. This patch does that
using parallel_for_each.
ICF accounted for about one third of total execution time. Previously,
it took 324 ms when self-hosting. Now it takes only 62 ms.
Of course your mileage may vary. My machine is a beefy 24-core Xeon machine,
so you may not see this much speedup. But this optimization should be
effective even for 2-core machine, since I saw speedup (324 ms -> 189 ms)
when setting parallelism parameter to 2.
llvm-svn: 248038
Previously, ICF created a vector for each SectionChunk. The vector
contained pointers to successors, which are namely associative sections
and COMDAT relocation targets. The reason I created vectors is because
I thought that that would make section comparison faster.
It did make the comparison faster. When self-linking, for example, it
saved about 10 ms on each iteration. The time we spent on constructing
the vectors was 124 ms. If we iterate more than 12 times, return from
the investment exceeds the initial cost.
In reality, it usually needs 5 iterations. So we shouldn't construct
the vectors.
llvm-svn: 247963
equalsConstants() is the heaviest function in ICF, and that consumes
more than half of total ICF execution time. Of which, section content
comparison accounts for roughly one third.
Previously, we compared section contents at the beginning of the
function after comparing their checksums. The comparison is very
likely to succeed because when the control reaches that comparison,
their checksums are always equal. And because checksums are 64-bit
CRC, they are unlikely to collide.
We compared relocations and associative sections after that.
If they are different, the time we spent on byte-by-byte comparison
of section contents were wasted.
This patch moves the comparison at the end of function. If the
comparison fails, the time we spent on relocation comparison are
wasted, but as I wrote it's very unlikely to happen.
LLD took 1198 ms to link itself to produce a 27.11 MB executable.
Of which, ICF accounted for 536 ms. This patch cuts it by 90 ms,
which is 17% speedup of ICF and 7.5% speedup overall. All numbers
are median of ten runs.
llvm-svn: 247961
Basically the concept of "liveness" is for sections (or chunks in LLD
terminology) and not for symbols. Symbols are always available or live,
or otherwise it indicates a link failure.
Previously, we had isLive() and markLive() methods for DefinedSymbol.
They are confusing methods. What they actually did is to act as a proxy
to backing section chunks. We can simplify eliminate these methods
and call section chunk's methods directly.
llvm-svn: 247869
Only live symbols are written to the symbol table. Because isLive()
returned false if dead-stripping was disabled entirely, only
non-COMDAT sections were written to the symbol table. This patch fixes
the issue.
llvm-svn: 247856
This patch defines ICF class and defines ICF-related functions as
members of the class. By doing this we can move code that are
related only to ICF from SectionChunk to the newly-defined class.
This also eliminates a global variable "NextID".
llvm-svn: 247802
This is a patch to make LLD to be on par with MSVC in terms of ICF
effectiveness. MSVC produces a 27.14MB executable when linking LLD.
LLD previously produced a 27.61MB when self-linking. Now the size
is reduced to 27.11MB. Note that without ICF the size is 29.63MB.
In r247387, I implemented an algorithm that handles section graphs
as cyclic graphs and merge them using SCC. The algorithm did not
always work as intended as I demonstrated in r247721. The new
algortihm implemented in this patch is different from the previous
one. If you are interested the details, you want to read the file
comment of ICF.cpp.
llvm-svn: 247770
Previously, LLD's ICF couldn't merge cyclic graphs. That was unfortunate
because, in COFF, cyclic graphs are not exceptional at all. That is
pretty common.
In this patch, sections are grouped by Tarjan's strongly connected
component algorithm to get acyclic graphs. And then we try to merge
SCCs whose outdegree is zero, and remove them from the graph. This
makes other SCCs to have outdegree zero, so we can repeat the
process until all SCCs are removed. When comparing two SCCs, we handle
cycles properly.
This algorithm works better than previous one. Previously, self-linking
produced a 29.0MB executable. It now produces a 27.7MB. There's still some
gap compared to MSVC linker which produces a 27.1MB executable for the
same input. So the gap is narrowed, but still LLD is not on par with MSVC.
I'll investigate that later.
llvm-svn: 247387
Identical COMDAT Folding is a feature to merge COMDAT sections
by contents. Two sections are considered the same if their contents,
relocations, attributes, etc, are all the same.
An interesting fact is that MSVC linker takes "iterations" parameter
for ICF because the algorithm they are using is iterative. Merging
two sections could make more sections to be mergeable because
different relocations could now point to the same section. ICF is
repeated until we get a convergence (until no section can be merged).
This algorithm is not fast. Usually it needs three iterations until a
convergence is obtained.
In the new algorithm implemented in this patch, we consider sections
and relocations as a directed acyclic graph, and we try to merge
sections whose outdegree is zero. Sections with outdegree zero are then
removed from the graph, which makes other sections to have outdegree
zero. We repeat that until all sections are processed. In this
algorithm, we don't iterate over the same sections many times.
There's an apparent issue in the algorithm -- the section graph is
not guaranteed to be acyclic. It's actually pretty often cyclic.
So this algorithm cannot eliminate all possible duplicates.
That's OK for now because the previous algorithm was not able to
eliminate cycles too. I'll address the issue in a follow-up patch.
llvm-svn: 246878
Previously, we calculated our own hash values for section contents.
Of coruse that's slow because we had to access all bytes in sections.
Fortunately, COFF objects usually contain hash values for COMDAT
sections. We can use that to speed up Identical COMDAT Folding.
llvm-svn: 246869
The option is added in MSVC 2015, and there's no documentation about
what the option is. This patch is to ignore the option for now, so that
at least LLD is usable with MSVC 2015.
llvm-svn: 246780
Rui Ueyama