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
