Previously applying 1 million relocations took about 2 seconds on my
Xeon 2.4GHz 8 core workstation. After this patch, it takes about 300
milliseconds. As a result, time to link chrome.dll becomes 23 seconds
to 21 seconds.
llvm-svn: 231454
We extend an underlying file before mmap'ing it, but it's not needed
on Windows. Extending file is slow on Windows, so we should avoid doing that.
The difference gets larger as the size of an output file gets larger.
It shove off 2 seconds out of 25 seconds when linking chrome.dll with LLD,
for example.
llvm-svn: 231452
Unlike GDB, we tackle the problem of representing vector types in different styles by having a synthetic child provider that recognizes the format you're trying to apply to the variable, and coming up with the right type and number of child values to match that format
This makes for a more compact representation and less visual noise
Fixes rdar://5429347
llvm-svn: 231449
Summary:
There are two packages in libgo which have
known failures when running the "make check" rule.
This change disables those packages in the tests so
that we can run libgo tests without them until the
root causes are identified and resolved.
Test Plan: ran check-libgo rule
Reviewers: pcc
Reviewed By: pcc
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D8003
llvm-svn: 231448
We would set the body of a struct type (therefore making it non-opaque)
but were forgetting to move it to the non-opaque set.
Fixes pr22807.
llvm-svn: 231442
These refactored computations check whether or not we are at a stage
of the sequence where we can perform a match. This patch moves the
computation out of the main dataflow and into
{BottomUp,TopDown}PtrState.
llvm-svn: 231439
This initialization occurs when we see a new retain or release. Before
we performed the actual initialization inline in the dataflow. That is
just messy.
llvm-svn: 231438
This will enable the main ObjCARCOpts dataflow to work with higher
level concepts such as "can this ptr state be modified by this ref
count" and not need to understand the nitty gritty details of how that
is determined. This makes the dataflow cleaner.
llvm-svn: 231437
In the resolver, we maintain a list of undefined symbols, and when we
visit an archive file, we check that file if undefined symbols can be
resolved using files in the archive. The archive file class provides
find() function to lookup a symbol.
Previously, we call find() for each undefined symbols. Archive files
may be visited multiple times if they are in a --start-group and
--end-group. If we visit a file M times and if we have N undefined
symbols, find() is called M*N times. I found that that is one of the
most significant bottlenecks in LLD when linking a large executable.
find() is not a very cheap operation because it looks up a hash table
for a given string. And a string, or a symbol name, can be pretty long
if you are dealing with C++ symbols.
We can eliminate the bottleneck.
Calling find() with the same symbol multiple times is a waste. If a
result of looking up a symbol is "not found", it stays "not found"
forever because the symbol simply doesn't exist in the archive.
Thus, we should call find() only for newly-added undefined symbols.
This optimization makes O(M*N) O(N).
In this patch, all undefined symbols are added to a vector. For each
archive/shared library file, we maintain a start position P. All
symbols [0, P) are already searched. [P, end of the vector) are not
searched yet. For each file, we scan the vector only once.
This patch changes the order in which undefined symbols are looked for.
Previously, we iterated over the result of _symbolTable.undefines().
Now we iterate over the new vector. This is a benign change but caused
differences in output if remaining undefines exist. This is why some
tests are updated.
The performance improvement of this patch seems sometimes significant.
Previously, linking chrome.dll on my workstation (Xeon 2.4GHz 8 cores)
took about 70 seconds. Now it takes (only?) 30 seconds!
http://reviews.llvm.org/D8091
llvm-svn: 231434
_reverseRef is a multimap from atoms to atoms. The map contains
reverse edges of "layout-before" and "group" edges for dead-stripping.
The type of the variable was DenseMap<Atom *, DenseSet<Atom *>>.
This patch changes that to std::unordered_multimap<Atom *, Atom *>.
A DenseMap with a value type of DenseSet was not fast. Inserting 900k
items to the map took about 1.6 seconds on my workstation.
unordered_multimap on the other hand took only 0.6 seconds.
Use of the map also got faster -- originally markLive took 1.3 seconds
in the same test case, and it now took 1.0 seconds. In total we shove
off 1.3 seconds out of 27 seconds in that test case.
llvm-svn: 231432
We used to save out and eagerly load a (potentially huge) table of merged
formerly-canonical declarations when we loaded each module. This was extremely
inefficient in the presence of large amounts of merging, and didn't actually
save any merging lookup work, because we still needed to perform name lookup to
check that our merged declaration lists were complete. This also resulted in a
loss of laziness -- even if we only needed an early declaration of an entity, we
would eagerly pull in all declarations that had been merged into it regardless.
We now store the relevant fragments of the table within the declarations
themselves. In detail:
* The first declaration of each entity within a module stores a list of first
declarations from imported modules that are merged into it.
* Loading that declaration pre-loads those other entities, so that they appear
earlier within the redeclaration chain.
* The name lookup tables list the most recent local lookup result, if there
is one, or all directly-imported lookup results if not.
llvm-svn: 231424
This is a temporary fix until a more comprehensive fix can be made for finding functions that we call in expressions.
We find "NSLog" in ClangExpressionDeclMap::FindExternalVisibleDecls() in after a call to target->GetImages().FindFunctions(...). Note that there are two symbols: NSLog from CFNetwork which is not external, and NSLog from Foundation which _is_ external. We do something with the external symbol with:
if (extern_symbol)
{
AddOneFunction (context, NULL, extern_symbol, current_id);
context.m_found.function = true;
}
Then later we try to lookup the _Z5NSLogP8NSStringz name and we don't find it so we call ClangExpressionDeclMap::GetFunctionAddress() with "_Z5NSLogP8NSStringz" as the name and the sc_list_size is zero at the "if" statement at line 568 because we don't find the mangled name and we extract the basename "NSLog" and call:
FindCodeSymbolInContext(ConstString(basename), m_parser_vars->m_sym_ctx, sc_list);
sc_list_size = sc_list.GetSize();
and we get a list size of two again, and we proceed to search for the symbol again, this time ignoring the external vs non-external-ness of the symbols that we find. This fix ensures we prioritize the external symbol until we get a real fix from Sean Callanan when he gets back to make sure we don't do multiple lookups for the same symbol we already resolved.
<rdar://problem/19879282>
llvm-svn: 231420
arithmetic type to a vector so that the arithmatic type
matches the vector element type. Without which it crashes
in Code Gen. rdar://20000762
llvm-svn: 231419
Fixes llgo following Duncan's changes to debug info in r231082. llgo needs
to replace composite types, which are no longer represented using MDTuple.
llvm-svn: 231416
At this point, we should have decent coverage of the involved code. I've got a few more test cases to cleanup and submit, but what's here is already reasonable.
I've got a collection of liveness tests which will be posted for review along with a decent liveness algorithm in the next few days. Once those are in, the code in this file should be well tested and I can start renaming things without risk of serious breakage.
llvm-svn: 231414
It's not that easy. If we're only checking -fsanitize=shift-base we
still need to verify that exponent has sane value, otherwise
UBSan-inserted checks for base will contain undefined behavior
themselves.
llvm-svn: 231409
This patch reduces code size for all AVX targets and increases speed for some chips.
SSE 4.1 introduced the useless (see code comments) 2-register form of BLENDV and
only in the packed float/double flavors.
AVX subsequently made the instruction useful by adding a 4-register operand form.
So we just need to paper over the lack of scalar forms of this instruction, complicate
the code to choose float or double forms, and use blendv on scalars since all FP is in
xmm registers anyway.
This gives us an approximately 50% speed up for a blendv microbenchmark sequence
on SandyBridge and Haswell:
blendv : 29.73 cycles/iter
logic : 43.15 cycles/iter
No new test cases with this patch because:
1. fast-isel-select-sse.ll tests the positive side for regular X86 lowering and fast-isel
2. sse-minmax.ll and fp-select-cmp-and.ll confirm that we're not firing for scalar selects without AVX
3. fp-select-cmp-and.ll and logical-load-fold.ll confirm that we're not firing for scalar selects with constants.
http://llvm.org/bugs/show_bug.cgi?id=22483
Differential Revision: http://reviews.llvm.org/D8063
llvm-svn: 231408
Yaron Keren