Merging two type streams is one of the most time consuming
parts of generating a PDB, and as such it needs to be as
fast as possible. The visitor abstractions used for interoperating
nicely with many different types of inputs and outputs have
been used widely and help greatly for testability and implementing
tools, but the abstractions build up and get in the way of
performance.
This patch removes all of the visitation stuff from the type
stream merger, essentially re-inventing the leaf / member switch
and loop, but at a very low level. This allows us many other
optimizations, such as not actually deserializing *any* records
(even member records which don't describe their own length), as
the operation of "figure out how long this record is" is somewhat
faster than "figure out how long this record *and* get all its
fields out". Furthermore, whereas before we had to deserialize,
re-write type indices, then re-serialize, now we don't have to
do any of those 3 steps. We just find out where the type indices
are and pull them directly out of the byte stream and re-write
them.
This is worth a 50-60% performance increase. On top of all other
optimizations that have been applied this week, I now get the
following numbers when linking lld.exe and lld.pdb
MSVC: 25.67s
Before This Patch: 18.59s
After This Patch: 8.92s
So this is a huge performance win.
Differential Revision: https://reviews.llvm.org/D33564
llvm-svn: 303935
Previously this code was defensive to the situation in which the debug
info scopes would lead to a different subprogram from the subprogram in
the CU's subprogram list (this could've happened with linkonce
functions, etc as per the comment being removed). Since the CU<>SP link
reversal this is no longer possible.
llvm-svn: 303933
I forgot to forward the chain, causing some missing instruction
dependencies. The test crashes the compiler without this patch.
Inspired by the test case, D33519 also tries to remove the extra sync.
Differential Revision: https://reviews.llvm.org/D33573
llvm-svn: 303931
We have wrappers for several other ValueTracking methods that take care of passing all of the analysis and assumption cache parameters. This extends it to isKnownToBeAPowerOfTwo.
llvm-svn: 303924
Right now scalarpre doesn't have phi-translate support, so it will miss some
simple pre opportunities. Like the following testcase, current scalarpre cannot
recognize the last "a * b" is fully redundent because a and b used by the last
"a * b" expr are both defined by phis.
long a[100], b[100], g1, g2, g3;
__attribute__((pure)) long goo();
void foo(long a, long b, long c, long d) {
g1 = a * b;
if (__builtin_expect(g2 > 3, 0)) {
a = c;
b = d;
g2 = a * b;
}
g3 = a * b; // fully redundant.
}
The patch adds phi-translate support in scalarpre. This is only a temporary
solution before the newpre based on newgvn is available.
Differential Revision: https://reviews.llvm.org/D32252
llvm-svn: 303923
Rename the DEBUG_TYPE to match the names of corresponding passes where
it makes sense. Also establish the pattern of simply referencing
DEBUG_TYPE instead of repeating the passname where possible.
llvm-svn: 303921
Originally this was intended to be set up so that when linking
a PDB which refers to a type server, it would only visit the
PDB once, and on subsequent visitations it would just skip it
since all the records had already been added.
Due to some C++ scoping issues, this was not occurring and it
was revisiting the type server every time, which caused every
record to end up being thrown away on all subsequent visitations.
This doesn't affect the performance of linking clang-cl generated
object files because we don't use type servers, but when linking
object files and libraries generated with /Zi via MSVC, this means
only 1 object file has to be linked instead of N object files, so
the speedup is quite large.
llvm-svn: 303920
Previously, every time we wanted to serialize a field list record, we
would create a new copy of FieldListRecordBuilder, which would in turn
create a temporary instance of TypeSerializer, which itself had a
std::vector<> that was about 128K in size. So this 128K allocation was
happening every time. We can re-use the same instance over and over, we
just have to clear its internal hash table and seen records list between
each run. This saves us from the constant re-allocations.
This is worth an ~18.5% speed increase (3.75s -> 3.05s) in my tests.
Differential Revision: https://reviews.llvm.org/D33506
llvm-svn: 303919
Previously it would do a character by character search for a null
terminator, to account for the fact that an arbitrary stream need not
store its data contiguously so you couldn't just do a memchr. However, the
stream API has a function which will return the longest contiguous chunk
without doing a copy, and by using this function we can do a memchr on the
individual chunks. For certain types of streams like data from object
files etc, this is guaranteed to find the null terminator with only a
single memchr, but even with discontiguous streams such as
MappedBlockStream, it's rare that any given string will cross a block
boundary, so even those will almost always be satisfied with a single
memchr.
This optimization is worth a 10-12% reduction in link time (4.2 seconds ->
3.75 seconds)
Differential Revision: https://reviews.llvm.org/D33503
llvm-svn: 303918
Summary:
DbiStreamBuilder calculated the offset of the source file names inside
the file info substream as the size of the file info substream minus
the size of the file names. Since the file info substream is padded to
a multiple of 4 bytes, this caused the first file name to be aligned
on a 4-byte boundary. By contrast, DbiModuleList would read the file
names immediately after the file name offset table, without skipping
to the next 4-byte boundary. This change makes it so that the file
names are written to the location where DbiModuleList expects them,
and puts any necessary padding for the file info substream after the
file names instead of before it.
Reviewers: amccarth, rnk, zturner
Reviewed By: amccarth, zturner
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D33475
llvm-svn: 303917
It was using the number of blocks of the entire PDB file as the number
of blocks of each stream that was created. This was only an issue in
the readLongestContiguousChunk function, which was never called prior.
This bug surfaced when I updated an algorithm to use this function and
the algorithm broke.
llvm-svn: 303916
Also, include global entries for all data symbols, not
just external ones, since these are referenced by the
relocation records.
Add a test case that includes unnamed data.
Differential Revision: https://reviews.llvm.org/D33079
llvm-svn: 303915
A profile shows the majority of time doing type merging is spent
deserializing records from sequences of bytes into friendly C++ structures
that we can easily access members of in order to find the type indices to
re-write.
Records are prefixed with their length, however, and most records have
type indices that appear at fixed offsets in the record. For these
records, we can save some cycles by just looking at the right place in the
byte sequence and re-writing the value, then skipping the record in the
type stream. This saves us from the costly deserialization of examining
every field, including potentially null terminated strings which are the
slowest, even though it was unnecessary to begin with.
In addition, we apply another optimization. Previously, after
deserializing a record and re-writing its type indices, we would
unconditionally re-serialize it in order to compute the hash of the
re-written record. This would result in an alloc and memcpy for every
record. If no type indices were re-written, however, this was an
unnecessary allocation. In this patch re-writing is made two phase. The
first phase discovers the indices that need to be rewritten and their new
values. This information is passed through to the de-duplication code,
which only copies and re-writes type indices in the serialized byte
sequence if at least one type index is different.
Some records have type indices which only appear after variable length
strings, or which have lists of type indices, or various other situations
that can make it tricky to make this optimization. While I'm not giving up
on optimizing these cases as well, for now we can get the easy cases out
of the way and lay the groundwork for more complicated cases later.
This patch yields another 50% speedup on top of the already large speedups
submitted over the past 2 days. In two tests I have run, I went from 9
seconds to 3 seconds, and from 16 seconds to 8 seconds.
Differential Revision: https://reviews.llvm.org/D33480
llvm-svn: 303914
By default, CMake uses a 32-bit toolchain, even when on a 64-bit platform targeting a 64-bit build. However, due to the size of the binaries involved, this can cause linker instabilities (such as the linker running out of memory). Guide people to the correct solution to get CMake to use the native toolchain.
llvm-svn: 303912
PPC::GETtlsADDR is lowered to a branch and a nop, by the assembly
printer. Its size was incorrectly marked as 4, correct it to 8. The
incorrect size can cause incorrect branch relaxation in
PPCBranchSelector under the right conditions.
llvm-svn: 303904
Summary:
This is used in the Linux kernel, and effectively just means "print an
address". This brings back r193593.
Reviewed by: Renato Golin
Reviewers: t.p.northover, rengolin, richard.barton.arm, kristof.beyls
Subscribers: aemerson, javed.absar, llvm-commits, eraman
Differential Revision: https://reviews.llvm.org/D33558
llvm-svn: 303901
This fixes an oversight in r300522, which changed alloca
dbg.values to no longer emit a DW_OP_deref.
The array.ll testcase was regenerated from source.
Fixes PR33166:
https://bugs.llvm.org/show_bug.cgi?id=33166
llvm-svn: 303897
Turns out gold doesn't use the DW_AT_GNU_pubnames to decide whether to
parse the rest of the DIEs when building gdb-index. This causes gold to
trip over LLVM's output when there are DW_FORM_ref_addr present.
Gold does use the presence of a debug_gnu_pub{names,types} entry for the
CU to skip parsing the debug_info portion, so make sure that's included
even when empty (technically, when empty there couldn't be any ref_addr
anyway - it only came up when gmlt didn't produce any (even non-empty)
pubnames - but given what that reveals about gold's implementation, this
seems like a good thing to do for consistency).
llvm-svn: 303894
Summary:
Previously, the yaml2pdb subcommand of llvm-pdbdump only
included object file names in module info if a module info stream was
present. This change makes it so that we include the object file name
even if there is no module info stream for the module. As a result,
running
llvm-pdbdump pdb2yaml -dbi-module-info original.pdb > original.yaml &&
llvm-pdbdump yaml2pdb -pdb=new.pdb original.yaml && llvm-pdbdump
pdb2yaml -dbi-module-info new.pdb > new.yaml now produces identical
original.yaml and new.yaml files.
Reviewers: amccarth, zturner
Reviewed By: zturner
Subscribers: fhahn, llvm-commits
Differential Revision: https://reviews.llvm.org/D33463
llvm-svn: 303891
There's probably a lot more like this (see also comments in D33338 about responsibility),
but I suspect we don't usually get a visible manifestation.
Given the recent interest in improving InstCombine efficiency, another potential micro-opt
that could be repeated several times in this function: morph the existing icmp pred/operands
instead of creating a new instruction.
llvm-svn: 303860
AVX512_VPOPCNTDQ is a new feature set that was published by Intel.
The patch represents the LLVM side of the addition of two new intrinsic based instructions (vpopcntd and vpopcntq).
Differential Revision: https://reviews.llvm.org/D33169
llvm-svn: 303858
This patch provides an initial prototype for a pass that sinks instructions based on GVN information, similar to GVNHoist. It is not yet ready for commiting but I've uploaded it to gather some initial thoughts.
This pass attempts to sink instructions into successors, reducing static
instruction count and enabling if-conversion.
We use a variant of global value numbering to decide what can be sunk.
Consider:
[ %a1 = add i32 %b, 1 ] [ %c1 = add i32 %d, 1 ]
[ %a2 = xor i32 %a1, 1 ] [ %c2 = xor i32 %c1, 1 ]
\ /
[ %e = phi i32 %a2, %c2 ]
[ add i32 %e, 4 ]
GVN would number %a1 and %c1 differently because they compute different
results - the VN of an instruction is a function of its opcode and the
transitive closure of its operands. This is the key property for hoisting
and CSE.
What we want when sinking however is for a numbering that is a function of
the *uses* of an instruction, which allows us to answer the question "if I
replace %a1 with %c1, will it contribute in an equivalent way to all
successive instructions?". The (new) PostValueTable class in GVN provides this
mapping.
This pass has some shown really impressive improvements especially for codesize already on internal benchmarks, so I have high hopes it can replace all the sinking logic in SimplifyCFG.
Differential revision: https://reviews.llvm.org/D24805
llvm-svn: 303850
instrumenting code.
This is important in the new pass manager. The old pass manager's
inliner has a small DCE routine embedded within it. The new pass manager
relies on the actual GlobalDCE pass for this.
Without this patch, instrumentation profiling with the new PM results in
massive code bloat in the object files because the instrumentation
itself ends up preventing DCE from working to remove the code.
We should probably change the instrumentation (and/or DCE) so that we
can eliminate dead code even if instrumented, but we shouldn't even
spend the time generating instrumentation for that code so this still
seems like a good patch.
Differential Revision: https://reviews.llvm.org/D33535
llvm-svn: 303845
pass.
The original logic only considered direct successors of the hoisted
domtree nodes, but that isn't really enough. If there are other basic
blocks that are completely within the subtree, their successors could
just as easily be impacted by the hoisting.
The more I think about it, the more I think the correct update here is
to hoist every block on the dominance frontier which has an idom in the
chain we hoist across. However, this is subtle enough that I'd
definitely appreciate some more eyes on it.
Sadly, if this is the correct algorithm, it requires computing a (highly
localized) dominance frontier. I've done this in the simplest (IE, least
code) way I could come up with, but that may be too naive. Suggestions
welcome here, dominance update algorithms are not an area I've studied
much, so I don't have strong opinions.
In good news, with this patch, turning on simple unswitch passes the
LLVM test suite for me with asserts enabled.
Differential Revision: https://reviews.llvm.org/D32740
llvm-svn: 303843
If Op is equal to array_lengthof, the lookup would be out of bounds, but we were only checking for greater than. I suspect nothing ever passes in the equal value because its a sentinel to mark the end of the builtin opcodes and not a real opcode.
So really this fix is just so that the code looks right and makes sense.
llvm-svn: 303840
having it internally allocate the loop.
This is a much more flexible API and necessary in the new loop unswitch
to reasonably support both new and old PMs in common code. It also just
seems like a cleaner separation of concerns.
NFC, this should just be a pure refactoring.
Differential Revision: https://reviews.llvm.org/D33528
llvm-svn: 303834
Summary:
This allows to keep handlers installed by sanitizers.
In other cases third-party code can replace handlers after libFuzzer
initialization anyway.
Reviewers: kcc
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D33522
llvm-svn: 303828
Coverage instrumentation which does not instrument full post-dominators
and full-dominators may skip valid paths, as the reasoning for skipping
blocks may become circular.
This patch fixes that, by only skipping
full post-dominators with multiple predecessors, as such predecessors by
definition can not be full-dominators.
llvm-svn: 303827
Summary:
Frontend generates store instructions after allocas, for example:
```
define i8* @f(i64 %this) "coroutine.presplit"="1" personality i32 0 {
entry:
%this.addr = alloca i64
store i64 %this, i64* %this.addr
..
%hdl = call i8* @llvm.coro.begin(token %id, i8* %alloc)
```
Such instructions may require spilling into coro.frame, but, coro-frame address is only available after coro.begin and thus needs to be moved after coro.begin.
The only instructions that should not be moved are the arguments of coro.begin and all of their operands.
Reviewers: GorNishanov, majnemer
Reviewed By: GorNishanov
Subscribers: llvm-commits, EricWF
Differential Revision: https://reviews.llvm.org/D33527
llvm-svn: 303825
There are some VectorShuffle Nodes in SDAG which can be selected to XXSLDWI
instruction, this patch recognizes them and does the selection to improve the
PPC performance.
llvm-svn: 303822
This change allows llvm-nm to print symbols found in import libraries,
in part by allowing COFFImportFiles to be casted to SymbolicFiles.
Patch by Dave Lee!
llvm-svn: 303821
The swapped operands in the first test is a manifestation of an
inefficiency for vectors that doesn't exist for scalars because
the IRBuilder checks for an all-ones mask for scalars, but not
vectors.
llvm-svn: 303818
