Lifting from Bob Wilson's notes: The hash value that we compute and
store in PGO profile data to detect out-of-date profiles does not
include enough information. This means that many significant changes to
the source will not cause compiler warnings about the profile being out
of date, and worse, we may continue to use the outdated profile data to
make bad optimization decisions. There is some tension here because
some source changes won't affect PGO and we don't want to invalidate the
profile unnecessarily.
This patch adds a new hashing scheme which is more sensitive to loop
nesting, conditions, and out-of-order control flow. Here are examples
which show snippets which get the same hash under the current scheme,
and different hashes under the new scheme:
Loop Nesting Example
--------------------
// Snippet 1
while (foo()) {
while (bar()) {}
}
// Snippet 2
while (foo()) {}
while (bar()) {}
Condition Example
-----------------
// Snippet 1
if (foo())
bar();
baz();
// Snippet 2
if (foo())
bar();
else
baz();
Out-of-order Control Flow Example
---------------------------------
// Snippet 1
while (foo()) {
if (bar()) {}
baz();
}
// Snippet 2
while (foo()) {
if (bar())
continue;
baz();
}
In each of these cases, it's useful to differentiate between the
snippets because swapping their profiles gives bad optimization hints.
The new hashing scheme considers some logical operators in an effort to
detect more changes in conditions. This isn't a perfect scheme. E.g, it
does not produce the same hash for these equivalent snippets:
// Snippet 1
bool c = !a || b;
if (d && e) {}
// Snippet 2
bool f = d && e;
bool c = !a || b;
if (f) {}
This would require an expensive data flow analysis. Short of that, the
new hashing scheme looks reasonably complete, based on a scan over the
statements we place counters on.
Profiles which use the old version of the PGO hash remain valid and can
be used without issue (there are tests in tree which check this).
rdar://17068282
Differential Revision: https://reviews.llvm.org/D39446
llvm-svn: 318229
The root cause of the issues reported in D32406 and D34680 is that clang
instruments functions without bodies. Make it stop doing that, and also
teach it how to use old (incorrectly generated) profiles without
crashing.
llvm-svn: 306883
Fix the fact that we don't assign profile counters to constructors in
classes with virtual bases, or constructors with variadic parameters.
Differential Revision: https://reviews.llvm.org/D30131
llvm-svn: 296062
This sets the maximum entry count among all functions in the program to the module using module flags. This allows the optimizer to use this information.
Differential Revision: http://reviews.llvm.org/D15163
llvm-svn: 255918
(This is part-2 of the patch -- fixing test cases)
Before the patch, -fprofile-instr-generate compile will fail
if no integrated-as is specified when the file contains
any static functions (the -S output is also invalid).
This patch fixed the issue. With the change, the index format
version will be bumped up by 1. Backward compatibility is
preserved with this change.
Differential Revision: http://reviews.llvm.org/D15243
llvm-svn: 255366
This patch adds support for specifying where the profile is emitted in a
way similar to GCC. These flags are used to specify directories instead
of filenames. When -fprofile-generate=DIR is used, the compiler will
generate code to write to <DIR>/default.profraw.
The patch also adds a couple of extensions: LLVM_PROFILE_FILE can still be
used to override the directory and file name to use and -fprofile-use
accepts both directories and filenames.
To simplify the set of flags used in the backend, all the flags get
canonicalized to -fprofile-instr-{generate,use} when passed to the
backend. The decision to use a default name for the profile is done
in the driver.
llvm-svn: 241825
We were assigning the counter for the body of the loop to the loop
variable initialization for some reason here, but our tests completely
lacked coverage for range-for loops. This fixes that and makes the
logic generally more similar to the logic for a regular for.
llvm-svn: 236277
We've added support for a multiple functions with the same name in
LLVM's profile data, so the lookup returning the function hash it
found doesn't make sense anymore. Update to pass in the hash we
expect.
This also adds a test that the version 1 format is still readable,
since the new API is expected to handle that.
llvm-svn: 214586
Improve the warning when building with -fprofile-instr-use and a file
appears not to have been profiled at all. This keys on whether a
function is defined in the main file or not to avoid false negatives
when one includes a header with functions that have been profiled.
llvm-svn: 211760
In preparation for using a binary format for instrumentation based
profiling, explicitly treat the test inputs as text and transform them
before running. This will allow us to leave the checked in files in
human readable format once the instrumentation format is binary.
No functional change.
llvm-svn: 206509
The function hash should change when control flow changes. This patch
hashes the type of each AST node that affects counters, rather than just
counting how many there are. These types are combined into a small
enumerator that currently has 16 values.
The new hash algorithm packs the enums for consecutively visited types
into a `uint64_t`. In order to save space for new types, the types are
assumed to be 6-bit values (instead of 4-bit). In order to minimize
overhead for functions with little control flow, the `uint64_t` is used
directly as a hash if it never fills up; if it does, it's passed through
an MD5 context.
<rdar://problem/16435801>
llvm-svn: 206397
CapturedStmt was being ignored by instrumentation based profiling, and
its counters attributed to the containing function. Instead, we need
to treat this as a top level entity, like we do with blocks.
llvm-svn: 206231
Until now we were generating duplicate counters for lambdas: one set
in the function where the lambda was declared and another for the
lambda itself. Instead, we should skip over the bodies of lambdas in
their containing contexts.
llvm-svn: 206081
If all of our weights are zero when calculating branch weights, it
means we haven't profiled the code in question. Avoid creating a
metadata node that says all branches are equally likely in this case.
The test also checks constructs that hit the other createBranchWeights
overload. These were already working.
llvm-svn: 205606
The hash itself is still the number of counters, which isn't all that
useful, but this separates the API changes from the actual
implementation of the hash and will make it easier to transition to
the ProfileData library once it's implemented.
llvm-svn: 204186
This updates CodeGenPGO to use the ProfileDataReader introduced to
llvm in r203703 and the new API for writing out the profile introduced
to compiler-rt in r203710.
llvm-svn: 203711
PGO counters are 64-bit and branch weights are 32-bit. Scale them down
when necessary, instead of just taking the lower 32 bits.
<rdar://problem/16276448>
llvm-svn: 203592
Some of this data had gotten out of date, so we weren't quite testing
what we thought we were. This also moves the outdated data test to its
own file to simplify regenerating the test data.
llvm-svn: 203546
These tests are logically related, but they're spread about several
different CodeGen directories. Consolidate them in one place to make
them easier to manage.
llvm-svn: 203541
Vedant Kumar