Extract the buffered filer writer code used by value profile
writer and turn it into common/sharable buffered fileIO
interfaces. Added a test case for the buffered file writer and
rewrite the VP dumping using the new APIs.
llvm-svn: 256604
The profile reader no longer depends on this field to be updated and point
to owning func's vp data. The VP data also no longer needs to be allocated
in a contiguous memory space.
Differential Revision: http://reviews.llvm.org/D15258
llvm-svn: 256543
Patch by: Johan Engelen
On windows, opening in text mode will result in
line ending chars to be appended leading to
profile corruption.
llvm-svn: 255684
Value profile runtime depends on libc which breaks
buffer API implemenation with current file organization.
Test case is also updated to check more symbols.
llvm-svn: 255294
InstrProfiling.h file declares profile runtime public APIs.
It has become a dumping place for many different things, which
needs cleanups. In this change, core type declarations and
portability macros are moved to a new file InstrProfilingPort.h.
llvm-svn: 255270
This allows the profile runtime to pick the right impl
for cmp&swap for a given target.
Differential Revision: http://reviews.llvm.org/D15248
llvm-svn: 255173
cmp&swap is not well supported -- the new test
case triggers some assembler error.
This is a partial fix to the general problem (lack
of atomics operation support for certain targets).
llvm-svn: 254701
With the latest refactoring and code sharing patches landed,
it is possible to unify the value profile implementation between
raw and indexed profile. This is part in prfofile runtime.
Differential Revision: http://reviews.llvm.org/D15057
llvm-svn: 254678
(patch suggested by silvas)
With this patch, the IO information is wrapped in struct
ProfDataIOVec, and interface of writerCallback takes a vector
of IOVec and a pointer to writer context pointer.
Differential Revision: http://reviews.llvm.org/D14859
llvm-svn: 253764
There seems to be a problem in system header (stdint.h) of FreeBSD
where uint8_t nor uint16_t are defined. Explicitly define the key
types as done for FreeBSD i386.
llvm-svn: 253703
Value profile enumerator change to match LLVM code
ProfData new member field name change to match LLVM code
ProfData member type change to match LLVM code
Do not use lower case for types that are internal to implementation (not exposed to APIs)
There is no functional change. This is a preparation patch to enable more code sharing
in follow up patches
Differential Revision: http://reviews.llvm.org/D14841
llvm-svn: 253700
MIPS build bots failed due to lack of 64bit atomic operations.
The fix is to disable VP for MIPS target until a better solution
is found.
llvm-svn: 253687
1. Added missing public API decl in InstrProfiling.h
2. Clang formatting fix
3. Added more comments for new VP code
4. refactor the VP allocation code to make it more readable.
llvm-svn: 253508
This change adds extends the data structures and adds in the routines
for handling runtime calls for value profiling. The profile data format
is modified and the version number is incremented.
llvm-svn: 253483
- This is to handle a corner case where profile lib is linked
in but non of the modules are instrumented (On linux, since
we avoided the overhead to emit runtime hook use functions so
this is the side effect of that size optimization).
- Added a profile runtime test case to cover all scenarios of
shared library builds.
Differential Revision: http://reviews.llvm.org/D14468
llvm-svn: 253098
Runtime code implicitly depends on the defintions and const
values defined in LLVM proper, but currently such dependency
is made implicitly by duplicating code across two dirs. As
part of the PGO cleanup effort, there will be changes to share
common sources. This is a preparation patch to enable it (NFC).
Differential Revision: http://reviews.llvm.org/D14487
llvm-svn: 252570
To make them the same as the common def in InstrProfData.inc.
This is a preparation to make the runtime code to use the
template as well. NFC
llvm-svn: 252421
This patch introduces a well defined header struct
to represent raw profile header instead of using raw array.
Previously the raw array is used in two different files and
is very error prone when header structure is re-organized.
This is a small cleanup with NFC.
llvm-svn: 250561
On Linux, the profile runtime can use __start_SECTNAME and __stop_SECTNAME
symbols defined by the linker to locate the start and end location of
a named section (with C name). This eliminates the need for instrumented
binary to call __llvm_profile_register_function during start-up time.
llvm-svn: 250200
Summary: This refactoring moves much of the Apple-specific behavior into a function in AddCompilerRT. The next cleanup patch will remove more of the if(APPLE) checks in the outlying CMakeLists.
This patch adds a bunch of new functionality to add_compiler_rt_runtime so that the target names don't need to be reconstructed outside the call. It also updates some of the call sites to exercise the new functionality, but does not update all uses fully. Subsequent patches will further update call sites and move to using the new features.
Reviewers: filcab, bogner, kubabrecka, zaks.anna, glider, samsonov
Subscribers: beanz, rengolin, llvm-commits
Differential Revision: http://reviews.llvm.org/D12292
llvm-svn: 245970
Summary: This patch consolidates add_compiler_rt_osx_static_runtime and add_compiler_rt_darwin_dynamic_runtime into a single new function add_compiler_rt_darwin_runtime.
Reviewers: filcab, samsonov, bogner
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D12106
llvm-svn: 245317
When the file is initialized, this patch checks whether the path
specifies a directory. If so, it creates the directory tree before
truncating the file.
Use default.profdata instead of pgo-data for default indexed profile name.
llvm-svn: 241824
Summary:
Use CMake's cmake_parse_arguments() instead.
It's called in a slightly different way, but supports all our use cases.
It's in CMake 2.8.8, which is our minimum supported version.
CMake 3.0 doc (roughly the same. No direct link to 2.8.8 doc):
http://www.cmake.org/cmake/help/v3.0/module/CMakeParseArguments.html?highlight=cmake_parse_arguments
Since I was already changing these calls, I changed ARCH and LIB into
ARCHS and LIBS to make it more clear that they're lists of arguments.
Reviewers: eugenis, samsonov, beanz
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D10529
llvm-svn: 240120
the profile output from the command line via -fprofile-instr-generate=<path>,
where the specified output path/file will be overridden by the
LLVM_PROFILE_FILE environment variable.
Several changes are made to the runtime to support this:
Add a new interface __llvm_profile_override_default_filename that will
set the profile output filename, but allows LLVM_PROFILE_FILE to override.
This is the interface used by the new option.
Refactor the pid-expansion done for LLVM_PROFILE_FILE into a separate
routine that can be shared by the various filename setting routines
(so that the filename from the option can also use the "%p" syntax).
Move the truncation into setFilename, and only truncate if there is a
new filename specified (to maintain support for appending to the same
profile file in the case of multiple shared objects built with profiling).
Move the handling for a NULL filename passed to __llvm_profile_set_filename and
__llvm_profile_override_default_filename into the new setFilenamePossiblyWithPid
routine. This now correctly resets the output file to default.profraw
instead of NULL.
The handling for a null LLVM_PROFILE_FILE (which should not reset) is done
by caller setFilenameFromEnvironment.
Patch by Teresa Johnson.
llvm-svn: 236055
This avoids crashing or corrupting data if multiple concurrent
processes write to the same .gcda file. This is hard to test, since
the previous behaviour was a data race that often worked out, and it
ignores errors in flock to fall back to the old racy behaviour so that
it won't degrade the behaviour on filesystems that don't support
flock.
llvm-svn: 234036
Debugging a missing profile is a bit painful right now. We can make
people's lives a bit easier by adding a knob to enable printing a
helpful error message for such failures.
llvm-svn: 226312
On Darwin, compiler_rt uses magic linker symbols to find the profile
counters in the __DATA segment. This is a reasonable method for
normal, hosted, userspace programs. However programs with custom
memory layouts, such as the kernel, will need to tell compiler_rt
explicitly where to find these sections.
Patch by Lawrence D'Anna. Thanks!
llvm-svn: 223840
This change removes libclang_rt.profile-pic-<arch>.a version of
profile runtime. Instead, it's sufficient to always build
libclang_rt.profile-<arch>.a with -fPIC, as it can be linked into
both executables and shared objects.
llvm-svn: 221952
The naming scheme we're using for counters in profile data shares a
prefix with some fixed names we use for the runtime, notably
__llvm_profile_data_begin and _end. Embarrassingly, this means a
function called begin() can't be instrumented.
This modifies the runtime names so as not to collide with the
instrumentation.
llvm-svn: 217166
Shared objects are hard. After this commit, we do the right thing when
profiling two separate shared objects that have been dlopen'd with
`RTLD_LOCAL`, when the main executable is *not* being profiled.
This mainly simplifies the writer logic.
- At initialization, determine the output filename and truncate the
file. Depending on whether shared objects can see each other, this
may happen multiple times.
- At exit, each executable writes its own profile in append mode.
<rdar://problem/16918688>
llvm-svn: 209053
These tests were XPASS-ing on Linux bots creating Mach-O, which makes
sense, since the real difference is the object format.
I'm hoping a short-term fix to get these tests passing on ELF is to
create two copies of the runtime -- one built with -fPIC, and one
without. A follow-up patch will change clang's driver to pick between
them depending on whether `-shared` is specified.
llvm-svn: 208947
Change the API of the instrumented profiling library to work with shared
objects.
- Most things are now declared hidden, so that each executable gets
its own copy.
- Initialization hooks up a linked list of writers.
- The raw format with shared objects that are profiled consists of a
concatenated series of profiles. llvm-profdata knows how to deal
with that since r208938.
<rdar://problem/16918688>
llvm-svn: 208940
GCC -pedantic warns that the initialization of Header is not constant:
InstrProfilingFile.c:31:5: error: initializer element is not computable at load time [-Werror]
LLVM defaults to enabling -pedantic. If this warning is unhelpful, we
can consider revisiting that decision.
llvm-svn: 207784
Add the test infrastructure for testing lib/profile and a single test.
This initial commit only enables the tests on Darwin, but they'll be
enabled on Linux soon after.
<rdar://problem/16458307>
llvm-svn: 205256
Soon there will be an option to build compiler-rt parts as shared libraries
on Linux. Extracted from http://llvm-reviews.chandlerc.com/D3042
by Yuri Gribov.
llvm-svn: 205183
Change the name of the default profile dumped by compiler-rt to
default.profraw. This distinguishes it more clearly from the
(incompatible) format output by llvm-profdata that is read by clang
-fprofile-instr-use.
llvm-svn: 204676
Since the profile can come from 32-bit machines, the reader needs to
check the pointer size. Change the magic number to facilitate this.
<rdar://problem/16400648>
llvm-svn: 204556
This is a bit of a stab in the dark as I'm not sure I've got these
source files compiling correctly locally. (and the warning only
reproduces on a 32bit build anyway)
llvm-svn: 204521
Apparently, MSVC has stdint.h now? Let's see if the buildbots complain.
I'm not convinced that the build system is even set up for MSVC to build
this file, but...
llvm-svn: 204515
Write __llvm_profile_write_buffer(), which uses the same logic as
__llvm_profile_write_file(), but writes directly to a provided `char*`
buffer instead.
<rdar://problem/15943240>
llvm-svn: 204499
It was misguided to plan to rely on __llvm_profile_write_buffer() in
__llvm_profile_write_file(). It's less complex to duplicate the writing
logic than to mmap the file.
Since it's here to stay, move `FILE*`-based writing logic into
InstrProfilingFile.c.
<rdar://problem/15943240>
llvm-svn: 204498
Move functions around to prepare for some other changes.
- Merge InstrProfilingExtras.h with InstrProfiling.h. There's no
benefit to having these split.
- Rename InstrProfilingExtras.c to InstrProfilingFile.c.
- Split actual buffer writing code out of InstrProfiling.c into
InstrProfilingBuffer.c.
- Drive-by corrections of a couple of header comments.
<rdar://problem/15943240>
llvm-svn: 204497
Add logic to do a printf-style substitution of %p for the process pid in
the filename.
It's getting increasingly awkward to work on lib/profile without test
infrastructure. This needs to be fixed!
<rdar://problem/16383358>
llvm-svn: 204414
These functions are in the profile runtime. PGO comes later.
Unfortunately, there's only room for 16 characters in a Darwin section,
so use __llvm_prf_ instead of __llvm_profile_ for section names.
<rdar://problem/15943240>
llvm-svn: 204391
__llvm_pgo_write_default_file() was a bad name, since it checked the
environment (it wasn't just a default file).
- Change __llvm_pgo_write_file() to __llvm_pgo_write_file_with_name()
and make it static.
- Rename __llvm_pgo_write_default_file() to __llvm_pgo_write_file().
- Add __llvm_pgo_set_filename(), which sets the filename for
subsequent calls to __llvm_pgo_write_file().
<rdar://problem/15943240>
llvm-svn: 204381
Instead of relying on explicit static initialization from translation
units, create a new file, InstrProfilingRuntime.cc, with an
__llvm_pgo_runtime variable. After this commit (and its pair in clang),
the driver will create a use of this variable. Unless the user defines
their own version, the new object file will get pulled in, including
that C++ static initialization that calls
__llvm_pgo_register_write_atexit.
The result is that, at least on Darwin, static initialization typically
consists of a single function call, which registers a writeout functino
atexit. Furthermore, users can skip even this behaviour by defining
their own __llvm_pgo_runtime.
<rdar://problem/15943240>
llvm-svn: 204380
Currently we register instrumentation data at runtime to determine the
bounds of the sections where the data lives. Soon we'll implement
platform-specific linker magic to determine this at link time.
Move this logic to a separate file, so that our build system can choose
the correct platform-specific code.
No functionality change intended.
<rdar://problem/15943240>
llvm-svn: 204299
Split implementation files along a uses-libc/shouldn't-use-libc
boundary.
- InstrProfiling.h is a shared header.
- InstrProfiling.c provides an API to extract profiling data from the
runtime, but avoids the use of libc. Currently this is a lie:
__llvm_pgo_write_buffer() uses `FILE*` and related functions. It
will be updated soon to write to a `char*` buffer instead.
- InstrProfilingExtras.c provides a more convenient API for
interfacing with the profiling runtime, but has logic that does (and
will continue to) use libc.
<rdar://problem/15943240>
llvm-svn: 204268
In instrumentation-based profiling, we need a set of data structures to
represent the counters. Previously, these were built up during static
initialization. Now, they're shoved into a specially-named section so
that they show up as an array.
As a consequence of the reorganizing symbols, instrumentation data
structures for linkonce functions are now correctly coalesced.
This is the first step in a larger project to minimize runtime overhead
and dependencies in instrumentation-based profilng. The larger picture
includes removing all initialization overhead and making the dependency
on libc optional.
<rdar://problem/15943240>
llvm-svn: 204079
This will break without the corresponding change in clang, which I've
reverted until I figure out how to get it to link properly.
This reverts commit r203710.
llvm-svn: 203713
Also rename the default output file from "pgo-data" to "default.profdata".
The ".profdata" suffix is consistent with the name of the new llvm-profdata
tool.
llvm-svn: 201808
Adding the ARM RT sources to the CMake files, and enabling some
sanitizers to also build on ARM. This is far from supported or
production quality, but enabling it to build will get us errors
that we can actually fix.
Having said that, the Compiler-RT and the Asan libraries are
know to work on some variations of ARM.
llvm-svn: 200546
Otherwise on (some) 64-bit systems, -Wformat will trigger a warning
because uint64_t is an 'unsigned long' not an 'unsigned long long'.
Consequently, PGOProfiling.c would fail to build if -Werror and
-Wformat are both enabled.
llvm-svn: 198644
This is fairly minimal support for instrumentation based PGO. The data
format is inefficient, and the output file name is hardcoded to
pgo-data.
llvm-svn: 198638
(thanks!) by deferring the free of the filename until we finish writing
the coverage data to that file.
Bill, let me know if you'd prefer a different approach!
llvm-svn: 184895
through Bill's patch:
1) Correctly test the file descriptor after the sceond attempt at
creating the file.
2) Make the filename a global so that we can issue error messages from
other routines.
3) Check errno in several places and print it out so that errors are
easier to track down.
I don't really expect any of these to fix the current failures I'm
seeing, but I'm hopeful they'll at least let me debug them.
llvm-svn: 184799
We really need the C standard library to be available to implement the
profiling runtime library reasonably, and replicating everything in the
SDKs tree really isn't addressing any problems we have. Notably, all of
the sanitizer runtimes take the same approach, and this isn't a library
which could end up in a bootstrapping problem where the system headers
aren't even available.
This will hopefully prevent subsequent changes which start using various
other bits of C standard library to make things more debuggable.
llvm-svn: 184798
Using fwrite and fread was very *very* slow. The resulting code was multiple
times slower than GCC's implementation of gcov. Replace the fwrite/fread system
with an mmap() version.
If the `.gcda' file doesn't exist, we (re)allocate a buffer that we write
into. That gets written to the `.gcda' file in one chunk. If the `.gcda' file
already exists, we simply mmap() the file, modify the mapped data, and use
msync() to write the contents out to disk. It's much easier than implementing
our own buffering scheme, and we don't have to use fwrite's and fread's
buffering.
For those who are numbers-oriented, here are some timings:
GCC Verison
-----------
`.gcda' files don't exist: 23s
`.gcda' files do exist: 14s
LLVM Version (before this change)
---------------------------------
`.gcda' files don't exist: 28s
`.gcda' files do exist: 28s
LLVM Version (with this change)
-------------------------------
`.gcda' files don't exist: 18s
`.gcda' files do exist: 4s
It's a win-win-win-win-lose-win-win scenario!
<rdar://problem/13466086>
llvm-svn: 182563
The calls to fwrite/fread can be very expensive. GCC avoids this by using a
buffer to read and write from the file, thus limiting the number of fwrite/fread
calls.
<rdar://problem/13466086>
llvm-svn: 181924
The fread / fwrite calls were happening for each timer. However, that could be
pretty expensive for a large number of timers. Instead, read and write the
timers in one call.
This gives ~10% speedup in compilation time.
llvm-svn: 179990
This function replaces the call of `atexit' from being generated in the compile
units. Basically, it registers the "writeout" and "flush" functions (if
present). It will generate calls to the `atexit' function for cleanups and final
writeout functions, but only once. This is better than checking for `main',
because a library may not have a `main' function in it.
<rdar://problem/12439551>
llvm-svn: 177578
It may be prohibitively expensive to write out >1000 files at the same time. So
we would rather emit them serially. These functions allow the GCOV
implementation to register the functions that writeout the GCOV information per
compile unit. At exit, they are written.
<rdar://problem/12439551>
llvm-svn: 177436
The __llvm_gcov_flush() functions only work for the local compile unit. However,
when __gcov_flush() is called, the user expects all of the counters to be
flushed, not just the ones in the current compile unit.
This adds some library functions that register the flush functions. It also
defined __gcov_flush() so that loops through that list and calls the functions.
PR15191 & <rdar://problem/13167507>
llvm-svn: 177337
it. Fortunately, versions of gcov that predate the extra checksum also ignore
any extra data, so this isn't a problem. This matches the API change made in
r176745.
llvm-svn: 176746
didn't delete the original, and now they've diverged. I have no idea what's
going on. Apply my patch in r176173 to this one too, this one looks newer?
llvm-svn: 176236
With the advent of the __llvm_gcov_flush function, we need to be able to merge
counts into the .gcda files in an intelligent manner. This involves reading the
file if it exists, adding the counts together, and then writing the results.
<rdar://problem/12185886>
llvm-svn: 163923
sets of functions/objects into high level groups. Currently we have "builtins"
(the main compiler-rt code), "profile", and "asan".
- Use this to define CommonFunctions and ArchFunctions to only reference the
builtins functions.
llvm-svn: 145674
Xinliang David Li