Added some missing spaces between the module identifier and the start of
the debug message. Also added a ":" after the module identifier to make
this look a little nicer.
llvm-svn: 255259
- This simplifies the CallSite class, arg_begin / arg_end are now
simple wrapper getters.
- In several places, we were creating CallSite instances solely to call
arg_begin and arg_end. With this change, that's no longer required.
llvm-svn: 255226
loading the source Module, linking the function in the destination
module, and destroying the source Module before repeating with the
next function to import (potentially from the same Module).
Ideally we would keep the source Module alive and import the next
Function needed from this Module. Unfortunately this is not possible
because the linker does not leave it in a usable state.
However we can do better by first computing the list of all candidates
per Module, and only then load the source Module and import all the
function we need for it.
The trick to process callees is to materialize function in the source
module when building the list of function to import, and inspect them
in their source module, collecting the list of callees for each
callee.
When we move the the actual import, we will import from each source
module exactly once. Each source module is loaded exactly once.
The only drawback it that it requires to have all the lazy-loaded
source Module in memory at the same time.
Currently this patch already improves considerably the link time,
a multithreaded link of llvm-dis on my laptop was:
real 1m12.175s user 6m32.430s sys 0m10.529s
and is now:
real 0m40.697s user 2m10.237s sys 0m4.375s
Note: this is the full link time (linker+Import+Optimizer+CodeGen)
Differential Revision: http://reviews.llvm.org/D15178
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 255100
For an invoke with operand bundles, the [op_begin(), op_end()-3] range
can contain things other than invoke arguments. This change teaches
PruneEH to use arg_begin() and arg_end() explicitly.
llvm-svn: 255073
Summary:
Add a field on the PassManagerBuilder that clang or gold can use to pass
down a pointer to the function index in memory to use for importing when
the ThinLTO backend is triggered. Add support to supply this to the
function import pass.
Reviewers: joker.eph, dexonsmith
Subscribers: davidxl, llvm-commits, joker.eph
Differential Revision: http://reviews.llvm.org/D15024
llvm-svn: 254926
For efficiency reason, when importing multiple functions for the same Module,
we can avoid reparsing it every time.
Differential Revision: http://reviews.llvm.org/D15102
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 254486
When linking static archive, there is no individual module files to
load. Instead they can be mmap'ed and could be initialized from a
buffer directly. The callback provide flexibility to override the
scheme for loading module from the summary.
Differential Revision: http://reviews.llvm.org/D15101
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 254479
This fixes buildbots in systems that std::to_string is not present. It
also tidies the output of the diagnostic to render doubles a bit better
(thanks Ben Kramer for help with string streams and format).
llvm-svn: 254261
This adds two thresholds to the sample profiler to affect inlining
decisions: the concept of global hotness and coldness.
Functions that have accumulated more than a certain fraction of samples at
runtime, are annotated with the InlineHint attribute. Conversely,
functions that accumulate less than a certain fraction of samples, are
annotated with the Cold attribute.
This is very similar to the hints emitted by Clang when using
instrumentation profiles.
Notice that this is a very blunt instrument. A function may have
globally collected a significant fraction of samples, but that does not
necessarily mean that every callsite for that function is hot.
Ideally, we would annotate each callsite with the samples collected at
that callsite. This way, the inliner can incorporate all these weights
into its cost model.
Once the inliner offers this functionality, we can change the hints
emitted here to a more precise per-callsite annotation. For now, this is
providing some measure of speedups with our internal benchmarks. I've
observed speedups of up to 23% (though the geo mean is about 3%). I expect
these numbers to improve as the inliner gets better annotations.
llvm-svn: 254212
Based on testing of internal benchmarks, I'm lowering this threshold to
a value of 0.1%. This means that SamplePGO will respect 99.9% of the
original inline decisions when following a profile.
The performance difference is noticeable in some tests. With the
previous threshold, the speedups over baseline -O2 was about 0.63%. With
the new default, the speedups are around 3% on average.
The point of this threshold is not to do more aggressive inlining. When
an inlined callsite crosses this threshold, SamplePGO will redo the
inline decision so that it can better apply the input profile.
By respecting most original inline decisions, we can apply more of the
input profile because the shape of the code follows the profile more
closely.
In the next series, I'll be looking at adding some inline hints for the
cold callsites and for toplevel functions that are hot/cold as well.
llvm-svn: 254211
They are as much trouble as aliases to declarations. They are requiring
the code generator to define a symbol with the same value as another
symbol, but the second symbol is undefined.
If representing this is important for some optimization, we could add
support for available_externally aliases. They would be *required* to
point to a declaration (or available_externally definition).
llvm-svn: 254170
Add a simple initial heuristic to control importing based on the number
of instructions recorded in the function's summary. Add option to
control the limit, and test using option.
llvm-svn: 254036
When the original binary is executed and sampled, the resulting profile
contains information on the original inline stack. We currently follow
the original inline plan if we notice that the inlined callsite has more
than 0 samples to it.
A better way is to determine whether the callsite is actually worth
inlining. If the callsite accumulates a small fraction of the samples
spent in the parent function, then we don't want to bother inlining it
(as it means that the callsite is actually cold).
This patch introduces a threshold expressed in percentage of samples
in relation to the parent function. If the callsite uses less than N%
of the total samples used by its parent, the original inline decision is
not re-applied.
I've set the threshold to the very arbitrary value of 5%. I'm yet to do
any actual experiments to see what's a good value. I wanted to separate
the basic mechanism from the tuning.
llvm-svn: 254034
This patch implements a minimum spanning tree (MST) based instrumentation for
PGO. The use of MST guarantees minimum number of CFG edges getting
instrumented. An addition optimization is to instrument the less executed
edges to further reduce the instrumentation overhead. The patch contains both the
instrumentation and the use of the profile to set the branch weights.
Differential Revision: http://reviews.llvm.org/D12781
llvm-svn: 254021
Analyze imported function bodies and add any new external calls to
the worklist for importing. Currently no controls on the importing
so this will end up importing everything possible in the call tree
below the importing module. Basic profitability checks coming next.
Update test to check for iteratively inlined functions.
llvm-svn: 254011
Skip imports for weak_any aliases as well. Fix the test to check
non-import of weak aliases and functions, and import of normal alias.
llvm-svn: 253991
Summary:
This is a helper to perform cross-module import for ThinLTO. Right now
it is importing naively every possible called functions.
Reviewers: tejohnson
Subscribers: dexonsmith, llvm-commits
Differential Revision: http://reviews.llvm.org/D14914
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 253954
The existing coverage tracker counts the number of records that were used
from the input profile. An alternative view of coverage is to check how
many available samples were applied.
This way, if the profile contains several records with few samples, it
doesn't really matter much that they were not applied. The more
interesting records to apply are the ones that contribute many samples.
llvm-svn: 253912
If a function was originally inlined but not actually hot at runtime,
its samples will not be counted inside the parent function. This throws
off the coverage calculation because it expects to find more used
records than it should.
Fixed by ignoring functions that will not be inlined into the parent.
Currently, this is inlined functions with 0 samples. In subsequent
patches, I'll change this to mean "cold" functions.
llvm-svn: 253716
This reverts r253661.
Turns out that the assignment is not redundant (despite the Clang static analyzer claiming the opposite).
The variable is being used by the lambda function AddUsersToWorklistIfCapturing().
llvm-svn: 253696
While debugging some sampling coverage problems, I found this useful:
When applying samples from a profile, it helps to also know what line
offset and discriminator the sample belongs to. This makes it easy to
correlate against the input profile.
llvm-svn: 253670
We currently bail out of global localization if the global has non-instruction users. However, often these can be simple bitcasts or constant-GEPs, which we can easily turn into instructions before localizing. Be a bit more aggressive.
llvm-svn: 253584
This provides a way to force a function to have certain attributes from the command line. This can be useful when debugging or doing workload exploration, where manually editing IR is tedious or not possible (due to build systems etc).
The syntax is -force-attribute=function_name:attribute_name
All function attributes are parsed except alignstack as it requires an argument.
llvm-svn: 253550
While setting function attributes we check all instructions that may access memory. For a call instruction we check all arguments. The special check is required for pointers.
I added vector-of-pointers to the call arguments types that should be checked.
Differential Revision: http://reviews.llvm.org/D14693
llvm-svn: 253363
Address Duncan Exon Smith's comments on D14148, which was added after the patch had been LGTM'd and committed:
* clang-format one area where whitespace diffs occurred.
* Add a threshold to limit the store/load dominance checks as they are quadratic.
llvm-svn: 253192
Global to local demotion can speed up programs that use globals a lot. It is particularly useful with LTO, when the entire call graph is known and most functions have been internalized.
For a global to be demoted, it must only be accessed by one function and that function:
1. Must never recurse directly or indirectly, else the GV would be clobbered.
2. Must never rely on the value in GV at the start of the function (apart from the initializer).
GlobalOpt can already do this, but it is hamstrung and only ever tries to demote globals inside "main", because C++ gives extra guarantees about how main is called - once and only once.
In LTO mode, we can often prove the first property (if the function is internal by this point, we know enough about the callgraph to determine if it could possibly recurse). FunctionAttrs now infers the "norecurse" attribute for this reason.
The second property can be proven for a subset of functions by proving that all loads from GV are dominated by a store to GV. This is conservative in the name of compile time - this only requires a DominatorTree which is fairly cheap in the grand scheme of things. We could do more fancy stuff with MemoryDependenceAnalysis too to catch more cases but this appears to catch most of the useful ones in my testing.
llvm-svn: 253168
This reapplies r252949. I've changed the type of FuncName to be
std::string instead of StringRef in emitFnAttrCompatCheck.
Original commit message for r252949:
Provide a way to specify inliner's attribute compatibility and merging
rules using table-gen. NFC.
This commit adds new classes CompatRule and MergeRule to Attributes.td,
which are used to generate code to check attribute compatibility and
merge attributes of the caller and callee.
rdar://problem/19836465
llvm-svn: 252990
rules using table-gen. NFC.
This commit adds new classes CompatRule and MergeRule to Attributes.td,
which are used to generate code to check attribute compatibility and
merge attributes of the caller and callee.
rdar://problem/19836465
llvm-svn: 252949
A function can be marked as norecurse if:
* The SCC to which it belongs has cardinality 1; and either
a) It does not call any non-norecurse function. This includes self-recursion; or
b) It only has one callsite and the function that callsite is within is marked norecurse.
a) is best propagated bottom-up and b) is best propagated top-down.
We build up the norecurse attributes bottom-up using the existing SCC pass, and mark functions with no obvious recursion (but not provably norecurse) to sweep later, top-down.
llvm-svn: 252862
When working with tokens, it is often the case that one has instructions
which consume a token and produce a new token. Currently, we have no
mechanism to represent an initial token state.
Instead, we can create a notional "empty token" by inventing a new
constant which captures the semantics we would like. This new constant
is called ConstantTokenNone and is written textually as "token none".
Differential Revision: http://reviews.llvm.org/D14581
llvm-svn: 252811
When GlobalOpt splits an internal, global variable with an aggregate type, it
should propagate the externally_initialized flag to the newly created globals.
This makes the pass safe for our downstream use of this flag, while still
allowing some useful optimisations (such as removing dead parts of the split
aggregate) to be performed.
Differential Revision: http://reviews.llvm.org/D13382
llvm-svn: 252490
Summary:
Teach the FunctionAttrs to do the right thing for IR with operand
bundles.
Reviewers: reames, chandlerc
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D14408
llvm-svn: 252387
Summary:
This change fixes an iterator wraparound bug in
`determinePointerReadAttrs`.
Ideally, ++'ing off the `end()` of an iplist should result in a failed
assert, but currently iplist seems to silently wrap to the head of the
list on `end()++`. This is why the bad behavior is difficult to
demonstrate.
Reviewers: chandlerc, reames
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D14350
llvm-svn: 252386
Some implicit ilist iterator conversions have crept back into Analysis,
Transforms, Hexagon, and llvm-stress. This removes them.
I'll commit a patch immediately after this to disallow them (in a
separate patch so that it's easy to revert if necessary).
llvm-svn: 252371
Previously, subprograms contained a metadata reference to the function they
described. Because most clients need to get or set a subprogram for a given
function rather than the other way around, this created unneeded inefficiency.
For example, many passes needed to call the function llvm::makeSubprogramMap()
to build a mapping from functions to subprograms, and the IR linker needed to
fix up function references in a way that caused quadratic complexity in the IR
linking phase of LTO.
This change reverses the direction of the edge by storing the subprogram as
function-level metadata and removing DISubprogram's function field.
Since this is an IR change, a bitcode upgrade has been provided.
Fixes PR23367. An upgrade script for textual IR for out-of-tree clients is
attached to the PR.
Differential Revision: http://reviews.llvm.org/D14265
llvm-svn: 252219
Summary:
Remove the loop over the uses of the CallSite in ArgumentUsesTracker.
Since we have the `Use *` for actual argument operand, we can just use
pointer subtraction.
The time complexity remains the same though (except for a vararg
argument) -- `std::advance` is O(UseIndex) for the ArgumentList
iterator.
The real motivation is to make a later change adding support for operand
bundles simpler.
Reviewers: reames, chandlerc, nlewycky
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D14363
llvm-svn: 252141
Summary:
The goal of this pass is to perform store-to-load forwarding across the
backedge of a loop. E.g.:
for (i)
A[i + 1] = A[i] + B[i]
=>
T = A[0]
for (i)
T = T + B[i]
A[i + 1] = T
The pass relies on loop dependence analysis via LoopAccessAnalisys to
find opportunities of loop-carried dependences with a distance of one
between a store and a load. Since it's using LoopAccessAnalysis, it was
easy to also add support for versioning away may-aliasing intervening
stores that would otherwise prevent this transformation.
This optimization is also performed by Load-PRE in GVN without the
option of multi-versioning. As was discussed with Daniel Berlin in
http://reviews.llvm.org/D9548, this is inferior to a more loop-aware
solution applied here. Hopefully, we will be able to remove some
complexity from GVN/MemorySSA as a consequence.
In the long run, we may want to extend this pass (or create a new one if
there is little overlap) to also eliminate loop-indepedent redundant
loads and store that *require* versioning due to may-aliasing
intervening stores/loads. I have some motivating cases for store
elimination. My plan right now is to wait for MemorySSA to come online
first rather than using memdep for this.
The main motiviation for this pass is the 456.hmmer loop in SPECint2006
where after distributing the original loop and vectorizing the top part,
we are left with the critical path exposed in the bottom loop. Being
able to promote the memory dependence into a register depedence (even
though the HW does perform store-to-load fowarding as well) results in a
major gain (~20%). This gain also transfers over to x86: it's
around 8-10%.
Right now the pass is off by default and can be enabled
with -enable-loop-load-elim. On the LNT testsuite, there are two
performance changes (negative number -> improvement):
1. -28% in Polybench/linear-algebra/solvers/dynprog: the length of the
critical paths is reduced
2. +2% in Polybench/stencils/adi: Unfortunately, I couldn't reproduce this
outside of LNT
The pass is scheduled after the loop vectorizer (which is after loop
distribution). The rational is to try to reuse LAA state, rather than
recomputing it. The order between LV and LLE is not critical because
normally LV does not touch scalar st->ld forwarding cases where
vectorizing would inhibit the CPU's st->ld forwarding to kick in.
LoopLoadElimination requires LAA to provide the full set of dependences
(including forward dependences). LAA is known to omit loop-independent
dependences in certain situations. The big comment before
removeDependencesFromMultipleStores explains why this should not occur
for the cases that we're interested in.
Reviewers: dberlin, hfinkel
Subscribers: junbuml, dberlin, mssimpso, rengolin, sanjoy, llvm-commits
Differential Revision: http://reviews.llvm.org/D13259
llvm-svn: 252017
This reverts commit r251837, due to a number of bot failures of the form:
/home/grosser/buildslave/perf-x86_64-penryn-O3-polly-fast/llvm.obj/tools/llvm-link/Release+Asserts/llvm-link.o:llvm-link.cpp:function
loadIndex(llvm::LLVMContext&, llvm::Module const*): error: undefined
reference to
'llvm::object::FunctionIndexObjectFile::create(llvm::MemoryBufferRef,
llvm::LLVMContext&, llvm::Module const*, bool)'
/home/grosser/buildslave/perf-x86_64-penryn-O3-polly-fast/llvm.obj/tools/llvm-link/Release+Asserts/llvm-link.o:llvm-link.cpp:function
loadIndex(llvm::LLVMContext&, llvm::Module const*): error: undefined
reference to 'llvm::object::FunctionIndexObjectFile::takeIndex()'
I'm not sure why these are happening - I added Object to the requred
libraries in tools/llvm-link/LLVMBuild.txt and the LLVM_LINK_COMPONENTS
in tools/llvm-link/CMakeLists.txt. Confirmed for my build that these
symbols come out of libLLVMObject.a. What am I missing?
llvm-svn: 251841
Summary:
Support for necessary linkage changes and symbol renaming during
ThinLTO function importing.
Also includes llvm-link support for manually importing functions
and associated llvm-link based tests.
Note that this does not include support for intelligently importing
metadata, which is currently imported duplicate times. That support will
be in the follow-on patch, and currently is ignored by the tests.
Reviewers: dexonsmith, joker.eph, davidxl
Subscribers: tobiasvk, tejohnson, llvm-commits
Differential Revision: http://reviews.llvm.org/D13515
llvm-svn: 251837
The initial coverage checking code for sample records failed to count
records inside inlined profiles. This change fixes the oversight.
llvm-svn: 251752
from its pass harness by providing a lambda to query for AA results.
This allows the legacy pass to easily provide a lambda that uses the
special helpers to construct function AA results from a legacy CGSCC
pass. With the new pass manager (the next patch) the lambda just
directly wraps the intuitive query API.
llvm-svn: 251715
transformations in FunctionAttrs rather than building a new one each
time.
This isn't trivial because there are different heuristics from different
passes for exactly what set they want. The primary difference is whether
an *overridable* function completely disables the synthesis of
attributes. I've modeled this by directly testing for overridable, and
using the common set that excludes external and opt-none functions.
This does cause some changes by disabling more optimizations in the face
of opt-none. Specifically, we were still optimizing *calls* to opt-none
functions based on their attributes, just not the bodies. It seems
better to be conservative on both fronts given the intended semanticas
here (best effort to not assume or disturb anything). I've not tried to
test this change as it seems complex, brittle, and not important to the
implicit contract of opt-none. Instead, it seems more like a choice that
should be dictated by the simplified implementation and the change to be
acceptable differences within the space of opt-none.
A big benefit here is that these transformations no longer rely on the
legacy pass manager's SCC types, they just work on generic sets of
function pointers. This will make it easy to re-use their logic in the
new pass manager.
I've also made the transforms static functions instead of members where
trivial while I was touching the signatures.
llvm-svn: 251640
This adds the flag -mllvm -sample-profile-check-coverage=N to the
SampleProfile pass. N is the percent of input sample records that the
user expects to apply. If the pass does not use N% (or more) of the
sample records in the input, it emits a warning.
This is useful to detect some forms of stale profiles. If the code has
drifted enough from the original profile, there will be records that do
not match the IR anymore.
This will not detect cases where a sample profile record for line L is
referring to some other instructions that also used to be at line L.
llvm-svn: 251568
The pass was keeping around a lot of per-function data (visited blocks,
edges, dominance, etc) that is just taking up memory for no reason. In
fact, from function to function it could potentially confuse the
propagator since some maps are indexed by line offsets which can be
common between functions.
llvm-svn: 251531
I think these were affected by a change way back when to stop printing newlines in Value::dump() by default. This change simply allows the debug output to be readable.
NFC.
llvm-svn: 251517
When emitting a remark for a conditional branch annotation, the remark
uses the line location information of the conditional branch in the
message. In some cases, that information is unavailable and the
optimization would segfaul. I'm still not sure whether this is a bug or
WAI, but the optimizer should not die because of this.
llvm-svn: 251420
No functionality changed here, but the indentation is substantially
reduced and IMO the code is much easier to read. I've also added some
helpful comments.
This is just a clean-up I wrote while studying the code, and that has
been in my backlog for a while.
llvm-svn: 251381
This adds a couple of optimization remarks to the SamplePGO
transformation. When it decides to inline a hot function (to mimic the
inline stack and repeat useful inline decisions in the original build).
It will also report branch destinations. For instance, given the code
fragment:
6 if (i < 1000)
7 sum -= i;
8 else
9 sum += -i * rand();
If the 'else' branch is taken most of the time, building this code with
-Rpass=sample-profile will produce:
a.cc:9:14: remark: most popular destination for conditional branches at small.cc:6:9 [-Rpass=sample-profile]
sum += -i * rand();
^
llvm-svn: 251330
In some cases (as illustrated in the unittest), lineno can be less than the heade_lineno because the function body are included from some other files. In this case, offset will be negative. This patch makes clang still able to match the profile to IR in this situation.
http://reviews.llvm.org/D13914
llvm-svn: 250873
This adjusts all integers in the reader/writer to reflect the types
stored on profile files. They should all be unsigned 32-bit or 64-bit
values. Changed all associated internal types to be uint32_t or
uint64_t.
The only place that needed some adjustments is in the sample profile
transformation. Altough the weight read from the profile are 64-bit
values, the internal API for branch weights only accepts 32-bit values.
The pass now saturates weights that overflow uint32_t.
llvm-svn: 250427
Now that all the known faults with GlobalsAA have been fixed, flip the big switch on -enable-non-lto-gmr again.
Feel free to pester me with any more bugs found, and don't hesitate to flip the switch back off.
llvm-svn: 250157
GlobalOpt currently merges stores into the initialisers of internal,
externally_initialized globals, but should not do so as the value of the global
may change between the initialiser and any code in the module being run.
llvm-svn: 250035
http://reviews.llvm.org/D13576
As we are using hierarchical profile, there is no need to keep HeaderLineno a member variable. This is because each level of the inline stack will have its own header lineno. One should use the head lineno of its own inline stack level instead of the actual symbol.
llvm-svn: 249848
Otherwise, the map will observe changes as long as MergeFunctions is alive. This
is bad because follow-up passes could replace-all-uses-with on the key of an
entry in the map. The value handle callback of ValueMap however asserts that the
key type matches.
rdar://22971893
llvm-svn: 249327
Place new and update dbg.declare calls immediately after the
corresponding alloca.
Current code in replaceDbgDeclareForAlloca puts the new dbg.declare
at the end of the basic block. LLVM codegen has problems emitting
debug info in a situation when dbg.declare appears after all uses of
the variable. This usually kinda works for inlining and ASan (two
users of this function) but not for SafeStack (see the pending change
in http://reviews.llvm.org/D13178).
llvm-svn: 248769
Patch by Jake VanAdrighem!
Summary:
Fix the way we sort the llvm.used and llvm.compiler.used members.
This bug seems to have been introduced in rL183756 through a set of improper casts to GlobalValue*. In subsequent patches this problem was missed and transformed into a getName call on a ConstantExpr.
Reviewers: silvas
Subscribers: silvas, llvm-commits
Differential Revision: http://reviews.llvm.org/D12851
llvm-svn: 248728
Loop unswitching produces conditional branches with constant condition,
and it's beneficial for later passes to clean this up with simplify-cfg.
We do this after the second invocation of loop-unswitch, but not after
the first one. Not doing so might cause problem for passes like
LoopUnroll, whose estimate of loop body size would be less accurate.
Reviewers: hfinkel
Differential Revision: http://reviews.llvm.org/D13064
llvm-svn: 248460
