2nd attempt: the first was in r296231, but it had a use after lifetime
bug.
Clang has logic to lower certain conditional expressions directly into llvm
select instructions. However, it does not emit the correct profile counter
increment as it does this: it emits an unconditional increment of the counter
for the 'then branch', even if the value selected is from the 'else branch'
(this is PR32019).
That means, given the following snippet, we would report that "0" is selected
twice, and that "1" is never selected:
int f1(int x) {
return x ? 0 : 1;
^2 ^0
}
f1(0);
f1(1);
Fix the problem by using the instrprof_increment_step intrinsic to do the
proper increment.
llvm-svn: 296245
Clang has logic to lower certain conditional expressions directly into
llvm select instructions. However, it does not emit the correct profile
counter increment as it does this: it emits an unconditional increment
of the counter for the 'then branch', even if the value selected is from
the 'else branch' (this is PR32019).
That means, given the following snippet, we would report that "0" is
selected twice, and that "1" is never selected:
int f1(int x) {
return x ? 0 : 1;
^2 ^0
}
f1(0);
f1(1);
Fix the problem by using the instrprof_increment_step intrinsic to do
the proper increment.
llvm-svn: 296231
Revert the two changes to thread CodeGenOptions into the TargetInfo allocation
and to fix the layering violation by moving CodeGenOptions into Basic.
Code Generation is arguably not particularly "basic". This addresses Richard's
post-commit review comments. This change purely does the mechanical revert and
will be followed up with an alternate approach to thread the desired information
into TargetInfo.
llvm-svn: 265806
This is a mechanical move of CodeGenOptions from libFrontend to libBasic. This
fixes the layering violation introduced earlier by threading CodeGenOptions into
TargetInfo. It should also fix the modules based self-hosting builds. NFC.
llvm-svn: 265702
Constructors and destructors may be represented by several functions
in IR. Only base structors correspond to source code, others are
small pieces of code and eventually call the base variant. In this
case instrumentation of non-base structors has little sense, this
fix remove it. Now profile data of a declaration corresponds to
exactly one function in IR, it agrees with the current logic of the
profile data loading.
This change fixes PR24996.
Differential Revision: http://reviews.llvm.org/D15158
llvm-svn: 254876
The fact that PGO has a say in how these branch weights are determined
isn't interesting to most of CodeGen, so it makes more sense for this
API to be accessible via CodeGenFunction rather than CodeGenPGO.
llvm-svn: 236380
No functional change. This just makes it more obvious that the logic
in ComputeRegionCounts only depends on the counter map and local
state.
llvm-svn: 236370
This removes the RegionCounter class, which is only used as a helper
in teh ComputeRegionCounts stmt visitor. This class is just an extra
layer of abstraction that makes the code harder to follow at this
point, and removing it makes the logic quite a bit more direct.
llvm-svn: 236364
The RegionCounter type does a lot of legwork, but most of it is only
meaningful within the implementation of CodeGenPGO. The uses elsewhere
in CodeGen generally just want to increment or read counters, so do
that directly.
llvm-svn: 235664
The logic for lowering profiling counters has been moved to an LLVM
pass. Emit the intrinsics rather than duplicating the whole pass in
clang.
llvm-svn: 223683
There's no need to use different names for the local variables than we
use in the profile itself, and it's a bit simpler and easier to debug
if we're consistent.
llvm-svn: 223173
It doesn't make much sense to have std::unique_ptrs of std::string and
std::vector. Avoid some useless indirection by using these types
directly.
llvm-svn: 223166
This patch adds the '-fcoverage-mapping' option which
allows clang to generate the coverage mapping information
that can be used to provide code coverage analysis using
the execution counts obtained from the instrumentation
based profiling (-fprofile-instr-generate).
llvm-svn: 214752
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
Update clang to use the InstrProfReader from LLVM to read
instrumentation based profile data. This also switches us from the
naive text format to the binary format, since that's what's
implemented in the reader.
llvm-svn: 206658
The variable is used to set the linkage for variables, and will become
different from function linkage in a follow-up commit.
<rdar://problem/15943240>
llvm-svn: 204407
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: 204390
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
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: 204080
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
In addition, for all functions, use the name from the llvm::Function to
identify the function in the profile data. Compute that "function name",
including the file name for local functions, once when assigning the PGO
counters and store it in the CodeGenPGO class.
Move the code to add InlineHint and Cold attributes out of StartFunction(),
because the "function name" string isn't available at that point.
llvm-svn: 203075
For C++ functions, we will continue to use the mangled name to identify
functions in the PGO profile data, but this name is confusing for things like
Objective-C methods. For functions with local linkage, we're also going to
include the file name to help distinguish those functions, so this changes to
use more generic variable names.
No functional changes.
llvm-svn: 203074
Move the PGO.assignRegionCounters() call out of StartFunction, because that
function is called from many places where it does not make sense to do PGO
instrumentation (e.g., compiler-generated helper functions). Change several
functions to take a StringRef argument for the unique name associated with
a function, so that the name can be set differently for things like Objective-C
methods and block literals.
llvm-svn: 203073
In r201528, I changed the PGO instrumentation counter for a "do" loop to not
include the fall-through count. That fall-through count is included later, b
it means that this assertion may fail for "do" loops.
llvm-svn: 202437
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
We collect a maximal function count among all functions in the pgo data file.
For functions that are hot, we set its InlineHint attribute. For functions that
are cold, we set its Cold attribute.
We currently treat functions with >= 30% of the maximal function count as hot
and functions with <= 1% of the maximal function count are treated as cold.
These two numbers are from preliminary tuning on SPEC.
This commit should not affect non-PGO builds and should boost performance on
instrumentation based PGO.
llvm-svn: 200874
adjustFallThroughCount isn't a good name, and the documentation was
even worse. This commit attempts to clarify what it's for and when to
use it.
llvm-svn: 199139
There are a number of places where we do PGO.setCurrentRegionCount(0)
directly after an unconditional branch. Give this operation a name so
that it's clearer why we're doing this.
llvm-svn: 199138
Vedant Kumar