Thunks are now implemented by redirecting the relocation to the
symbol S, to a symbol TS in a Thunk. The Thunk will transfer control
to S. This has the following implications:
- All the side-effects of Thunks happen within createThunks()
- Thunks are no longer stored in InputSections and Symbols no longer
need to hold a pointer to a Thunk
- The synthetic Thunk sections need to be merged into OutputSections
This implementation is almost a direct conversion of the existing
Thunks with the following exceptions:
- Mips LA25 Thunks are placed before the InputSection that defines
the symbol that needs a Thunk.
- All ARM Thunks are placed at the end of the OutputSection of the
first caller to the Thunk.
Range extension Thunks are not supported yet so it is optimistically
assumed that all Thunks can be reused.
Differential Revision: https://reviews.llvm.org/D29129
llvm-svn: 293283
In a shared library an undefined symbol is implicitly imported. If the
symbol is called as a function a PLT entry is generated for it. When the
caller is a Thumb b.w a thunk to the PLT entry is needed as all PLT
entries are in ARM state.
This change allows undefined symbols to have thunks in the same way that
shared symbols may have thunks.
llvm-svn: 290951
StringRefZ is a class to represent a null-terminated string. String
length is computed lazily, so it's more efficient than StringRef to
represent strings in string table.
The motivation of defining this new class is to merge functions
that only differ in string types; we have many constructors that takes
`const char *` or `StringRef`. With StringRefZ, we can merge them.
Differential Revision: https://reviews.llvm.org/D27037
llvm-svn: 288172
Now that lld switched to lib/LTO, which always calls setDataLayout(),
we don't need this check anymore.
Thanks to Peter for pointing out!
llvm-svn: 287699
Previously, we do this piece of code to iterate over all input sections.
for (elf::ObjectFile<ELFT> *F : Symtab.getObjectFiles())
for (InputSectionBase<ELFT> *S : F->getSections())
It turned out that this mechanisms doesn't work well with synthetic
input sections because synthetic input sections don't belong to any
input file.
This patch defines a vector that contains all input sections including
synthetic ones.
llvm-svn: 286051
Previously, we have a lot of BumpPtrAllocators, but all these
allocators virtually have the same lifetime because they are
not freed until the linker finishes its job. This patch aggregates
them into a single allocator.
Differential revision: https://reviews.llvm.org/D26042
llvm-svn: 285452
We used to have one allocator per file, which reduces the advantage of
using an allocator in the first place.
This is a small speed up is most cases. The largest speedup was in
1.014X in chromium no-gc. The largest slowdown was scylla at 1.003X.
llvm-svn: 285205
Before the default was whatever number hardware_concurrency() returned.
Users can specify the number of threads via --lto-jobs=X option.
llvm-svn: 283787
We have a few "check" functions in Error.h. All of them are to
check for an error and strip an error object if there was no error,
except "void check(Error E)", which doesn't return anything.
This patch renames it and moves it to the .cpp file where it is used.
llvm-svn: 282764
I found out this wasn't tested when looking at Vedant's coverage bot
numbers, so, thanks to him. While I'm here, switch the error message
to be lld-compliant (first letter lowercase).
llvm-svn: 282335
Previously, all input files were owned by the symbol table.
Files were created at various places, such as the Driver, the lazy
symbols, or the bitcode compiler, and the ownership of new files
was transferred to the symbol table using std::unique_ptr.
All input files were then free'd when the symbol table is freed
which is on program exit.
I think we don't have to transfer ownership just to free all
instance at once on exit.
In this patch, all instances are automatically collected to a
vector and freed on exit. In this way, we no longer have to
use std::unique_ptr.
Differential Revision: https://reviews.llvm.org/D24493
llvm-svn: 281425
Before this lld was always creating common symbols itself. It worked,
but prevented them from being internalized when possible.
Now it preserves common symbols is the bitcode and they are internalized.
Fixes pr30184.
llvm-svn: 280242
Previously, each subclass of SymbolBody had a pointer to a source
file from which it was created. So, there was no single way to get
a source file for a symbol. We had getSourceFile<ELFT>(), but the
function was a bit inconvenient as it's a template.
This patch makes SymbolBody have a pointer to a source file.
If a symbol is not created from a file, the pointer has a nullptr.
llvm-svn: 275701
This fixes PR28218. Thanks to Rafael for spotting a failure in
the SHARED_LIBS build!
Differential Revision: http://reviews.llvm.org/D21577
llvm-svn: 273451
When we undefine, we also preserve type of symbol so that we get
it right in the combined LTO object.
Differential Revision: http://reviews.llvm.org/D20851
llvm-svn: 271403
This allows the combined LTO object to provide a definition with the same
name as a symbol that was internalized without causing a duplicate symbol
error. This normally happens during parallel codegen which externalizes
originally-internal symbols, for example.
In order to make this work, I needed to relax the undefined symbol error to
only report an error for symbols that are used in regular objects.
Differential Revision: http://reviews.llvm.org/D19954
llvm-svn: 268649
This patch implements a new design for the symbol table that stores
SymbolBodies within a memory region of the Symbol object. Symbols are mutated
by constructing SymbolBodies in place over existing SymbolBodies, rather
than by mutating pointers. As mentioned in the initial proposal [1], this
memory layout helps reduce the cache miss rate by improving memory locality.
Performance numbers:
old(s) new(s)
Without debug info:
chrome 7.178 6.432 (-11.5%)
LLVMgold.so 0.505 0.502 (-0.5%)
clang 0.954 0.827 (-15.4%)
llvm-as 0.052 0.045 (-15.5%)
With debug info:
scylla 5.695 5.613 (-1.5%)
clang 14.396 14.143 (-1.8%)
Performance counter results show that the fewer required indirections is
indeed the cause of the improved performance. For example, when linking
chrome, stalled cycles decreases from 14,556,444,002 to 12,959,238,310, and
instructions per cycle increases from 0.78 to 0.83. We are also executing
many fewer instructions (15,516,401,933 down to 15,002,434,310), probably
because we spend less time allocating SymbolBodies.
The new mechanism by which symbols are added to the symbol table is by calling
add* functions on the SymbolTable.
In this patch, I handle local symbols by storing them inside "unparented"
SymbolBodies. This is suboptimal, but if we do want to try to avoid allocating
these SymbolBodies, we can probably do that separately.
I also removed a few members from the SymbolBody class that were only being
used to pass information from the input file to the symbol table.
This patch implements the new design for the ELF linker only. I intend to
prepare a similar patch for the COFF linker.
[1] http://lists.llvm.org/pipermail/llvm-dev/2016-April/098832.html
Differential Revision: http://reviews.llvm.org/D19752
llvm-svn: 268178
Using multiple context used to be a really big memory saving because we
could free memory from each file while the linker proceeded with the
symbol resolution. We are getting lazier about reading data from the
bitcode, so I was curious if this was still a good tradeoff.
One thing that is a bit annoying is that we still have to copy the
symbol names. The problem is that the names are stored in the Module and
get freed when we move the module bits during linking.
Long term I think the solution is to add a symbol table to the bitcode.
That way IRObject file will not need to use a Module or a Context and we
can drop it while still keeping a StringRef to the names.
This patch is still be an interesting medium term improvement.
When linking llvm-as without debug info this patch is a small speedup:
master: 29.861877513 seconds
patch: 29.814533787 seconds
With debug info the numbers are
master: 34.765181469 seconds
patch: 34.563351584 seconds
The peak memory usage when linking llvm-as with debug info was
master: 599.10MB
patch: 600.13MB
llvm-svn: 267921
Rui Ueyama