This patch teaches the bottleneck analysis how to identify and print the most
expensive sequence of instructions according to the simulation. Fixes PR37494.
The goal is to help users identify the sequence of instruction which is most
critical for performance.
A dependency graph is internally used by the bottleneck analysis to describe
data dependencies and processor resource interferences between instructions.
There is one node in the graph for every instruction in the input assembly
sequence. The number of nodes in the graph is independent from the number of
iterations simulated by the tool. It means that a single node of the graph
represents all the possible instances of a same instruction contributed by the
simulated iterations.
Edges are dynamically "discovered" by the bottleneck analysis by observing
instruction state transitions and "backend pressure increase" events generated
by the Execute stage. Information from the events is used to identify critical
dependencies, and materialize edges in the graph. A dependency edge is uniquely
identified by a pair of node identifiers plus an instance of struct
DependencyEdge::Dependency (which provides more details about the actual
dependency kind).
The bottleneck analysis internally ranks dependency edges based on their impact
on the runtime (see field DependencyEdge::Dependency::Cost). To this end, each
edge of the graph has an associated cost. By default, the cost of an edge is a
function of its latency (in cycles). In practice, the cost of an edge is also a
function of the number of cycles where the dependency has been seen as
'contributing to backend pressure increases'. The idea is that the higher the
cost of an edge, the higher is the impact of the dependency on performance. To
put it in another way, the cost of an edge is a measure of criticality for
performance.
Note how a same edge may be found in multiple iteration of the simulated loop.
The logic that adds new edges to the graph checks if an equivalent dependency
already exists (duplicate edges are not allowed). If an equivalent dependency
edge is found, field DependencyEdge::Frequency of that edge is incremented by
one, and the new cost is cumulatively added to the existing edge cost.
At the end of simulation, costs are propagated to nodes through the edges of the
graph. The goal is to identify a critical sequence from a node of the root-set
(composed by node of the graph with no predecessors) to a 'sink node' with no
successors. Note that the graph is intentionally kept acyclic to minimize the
complexity of the critical sequence computation algorithm (complexity is
currently linear in the number of nodes in the graph).
The critical path is finally computed as a sequence of dependency edges. For
edges describing processor resource interferences, the view also prints a
so-called "interference probability" value (by dividing field
DependencyEdge::Frequency by the total number of iterations).
Examples of critical sequence computations can be found in tests added/modified
by this patch.
On output streams that support colored output, instructions from the critical
sequence are rendered with a different color.
Strictly speaking the analysis conducted by the bottleneck analysis view is not
a critical path analysis. The cost of an edge doesn't only depend on the
dependency latency. More importantly, the cost of a same edge may be computed
differently by different iterations.
The number of dependencies is discovered dynamically based on the events
generated by the simulator. However, their number is not fixed. This is
especially true for edges that model processor resource interferences; an
interference may not occur in every iteration. For that reason, it makes sense
to also print out a "probability of interference".
By construction, the accuracy of this analysis (as always) is strongly dependent
on the simulation (and therefore the quality of the information available in the
scheduling model).
That being said, the critical sequence effectively identifies a performance
criticality. Instructions from that sequence are expected to have a very big
impact on performance. So, users can take advantage of this information to focus
their attention on specific interactions between instructions.
In my experience, it works quite well in practice, and produces useful
output (in a reasonable amount time).
Differential Revision: https://reviews.llvm.org/D63543
llvm-svn: 364045
This patch slightly refactors data structures internally used by the bottleneck
analysis to track data and resource dependencies.
This patch also updates methods used to print out information about dependency
edges when in debug mode.
This is the last of a sequence of commits done in preparation for an upcoming
patch that fixes PR37494. No functional change intended.
llvm-svn: 363677
Bottleneck Analysis is one of the many views available in llvm-mca. Therefore,
it should be enabled when flag -all-views is passed in input to the tool.
llvm-svn: 362964
The resource pressure distribution computation is now delegated by class
BottleneckAnalysis to an instance of class PressureTracker.
Class PressureTracker is also responsible for:
- tracking users of processor resource units.
- tracking the number of delay cycles caused by increases in backpressure.
BottleneckAnalysis internally initializes a dependency graph. Each nodes
represents an instruction in the input code sequence. Edges of the dependency
graph are critical register/memory/resource dependencies. Dependencies are only
added to the graph if they are seen as critical by backend pressure events.
The DependencyGraph is currently unused. It is possible to print the dependency
graph (see method DependencyGraph::dump()) for debugging purposes.
The long term goal is to use the information stored by the dependency graph in
order to do critical path computation.
llvm-svn: 362246
This patch fixes PR41523
https://bugs.llvm.org/show_bug.cgi?id=41523
Regions can now nest/overlap provided that they have different names.
Anonymous regions cannot overlap.
Region end markers must specify the region name. The only exception is for when
there is only one user-defined region; in that particular case, the region end
marker doesn't need to specify a name.
Incorrect region end markers are no longer ignored. Instead, the tool reports an
error and we exit with an error code.
Added test cases to verify the new diagnostic error messages.
Updated the llvm-mca docs to reflect this feature change.
Differential Revision: https://reviews.llvm.org/D61676
llvm-svn: 360351
It makes more sense to print out the number of micro opcodes that are issued
every cycle rather than the number of instructions issued per cycle.
This behavior is also consistent with the dispatch-stats: numbers from the two
views can now be easily compared.
llvm-svn: 357919
This patch adds an experimental stage named MicroOpQueueStage.
MicroOpQueueStage can be used to simulate a hardware micro-op queue (basically,
a decoupling queue between 'decode' and 'dispatch'). Users can specify a queue
size, as well as a optional MaxIPC (which - in the absence of a "Decoders" stage
- can be used to simulate a different throughput from the decoders).
This stage is added to the default pipeline between the EntryStage and the
DispatchStage only if PipelineOption::MicroOpQueue is different than zero. By
default, llvm-mca sets PipelineOption::MicroOpQueue to the value of hidden flag
-micro-op-queue-size.
Throughput from the decoder can be simulated via another hidden flag named
-decoder-throughput. That flag allows us to quickly experiment with different
frontend throughputs. For targets that declare a loop buffer, flag
-decoder-throughput allows users to do multiple runs, each time simulating a
different throughput from the decoders.
This stage can/will be extended in future. For example, we could add a "buffer
full" event to notify bottlenecks caused by backpressure. flag
-decoder-throughput would probably go away if in future we delegate to another
stage (DecoderStage?) the simulation of a (potentially variable) throughput from
the decoders. For now, flag -decoder-throughput is "good enough" to run some
simple experiments.
Differential Revision: https://reviews.llvm.org/D59928
llvm-svn: 357248
Found by inspection when looking at the debug output of MCA.
This problem was latent, and none of the upstream models were affected by it.
No functional change intended.
llvm-svn: 357000
Summary:
Since bottleneck hints are enabled via user request, it can be
confusing if no bottleneck information is presented. Such is the
case when no bottlenecks are identified. This patch emits a message
in that case.
Reviewers: andreadb
Reviewed By: andreadb
Subscribers: tschuett, gbedwell, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D59098
llvm-svn: 355628
This patch adds a new flag named -bottleneck-analysis to print out information
about throughput bottlenecks.
MCA knows how to identify and classify dynamic dispatch stalls. However, it
doesn't know how to analyze and highlight kernel bottlenecks. The goal of this
patch is to teach MCA how to correlate increases in backend pressure to backend
stalls (and therefore, the loss of throughput).
From a Scheduler point of view, backend pressure is a function of the scheduler
buffer usage (i.e. how the number of uOps in the scheduler buffers changes over
time). Backend pressure increases (or decreases) when there is a mismatch
between the number of opcodes dispatched, and the number of opcodes issued in
the same cycle. Since buffer resources are limited, continuous increases in
backend pressure would eventually leads to dispatch stalls. So, there is a
strong correlation between dispatch stalls, and how backpressure changed over
time.
This patch teaches how to identify situations where backend pressure increases
due to:
- unavailable pipeline resources.
- data dependencies.
Data dependencies may delay execution of instructions and therefore increase the
time that uOps have to spend in the scheduler buffers. That often translates to
an increase in backend pressure which may eventually lead to a bottleneck.
Contention on pipeline resources may also delay execution of instructions, and
lead to a temporary increase in backend pressure.
Internally, the Scheduler classifies instructions based on whether register /
memory operands are available or not.
An instruction is marked as "ready to execute" only if data dependencies are
fully resolved.
Every cycle, the Scheduler attempts to execute all instructions that are ready
to execute. If an instruction cannot execute because of unavailable pipeline
resources, then the Scheduler internally updates a BusyResourceUnits mask with
the ID of each unavailable resource.
ExecuteStage is responsible for tracking changes in backend pressure. If backend
pressure increases during a cycle because of contention on pipeline resources,
then ExecuteStage sends a "backend pressure" event to the listeners.
That event would contain information about instructions delayed by resource
pressure, as well as the BusyResourceUnits mask.
Note that ExecuteStage also knows how to identify situations where backpressure
increased because of delays introduced by data dependencies.
The SummaryView observes "backend pressure" events and prints out a "bottleneck
report".
Example of bottleneck report:
```
Cycles with backend pressure increase [ 99.89% ]
Throughput Bottlenecks:
Resource Pressure [ 0.00% ]
Data Dependencies: [ 99.89% ]
- Register Dependencies [ 0.00% ]
- Memory Dependencies [ 99.89% ]
```
A bottleneck report is printed out only if increases in backend pressure
eventually caused backend stalls.
About the time complexity:
Time complexity is linear in the number of instructions in the
Scheduler::PendingSet.
The average slowdown tends to be in the range of ~5-6%.
For memory intensive kernels, the slowdown can be significant if flag
-noalias=false is specified. In the worst case scenario I have observed a
slowdown of ~30% when flag -noalias=false was specified.
We can definitely recover part of that slowdown if we optimize class LSUnit (by
doing extra bookkeeping to speedup queries). For now, this new analysis is
disabled by default, and it can be enabled via flag -bottleneck-analysis. Users
of MCA as a library can enable the generation of pressure events through the
constructor of ExecuteStage.
This patch partially addresses https://bugs.llvm.org/show_bug.cgi?id=37494
Differential Revision: https://reviews.llvm.org/D58728
llvm-svn: 355308
This patch adds a lookup table to speed up resource queries in the ResourceManager.
This patch also moves helper function 'getResourceStateIndex()' from
ResourceManager.cpp to Support.h, so that we can reuse that logic in the
SummaryView (and potentially other views in llvm-mca).
No functional change intended.
llvm-svn: 354470
This patch removes hidden codegen flag -print-schedule effectively reverting the
logic originally committed as r300311
(https://llvm.org/viewvc/llvm-project?view=revision&revision=300311).
Flag -print-schedule was originally introduced by r300311 to address PR32216
(https://bugs.llvm.org/show_bug.cgi?id=32216). That bug was about adding "Better
testing of schedule model instruction latencies/throughputs".
These days, we can use llvm-mca to test scheduling models. So there is no longer
a need for flag -print-schedule in LLVM. The main use case for PR32216 is
now addressed by llvm-mca.
Flag -print-schedule is mainly used for debugging purposes, and it is only
actually used by x86 specific tests. We already have extensive (latency and
throughput) tests under "test/tools/llvm-mca" for X86 processor models. That
means, most (if not all) existing -print-schedule tests for X86 are redundant.
When flag -print-schedule was first added to LLVM, several files had to be
modified; a few APIs gained new arguments (see for example method
MCAsmStreamer::EmitInstruction), and MCSubtargetInfo/TargetSubtargetInfo gained
a couple of getSchedInfoStr() methods.
Method getSchedInfoStr() had to originally work for both MCInst and
MachineInstr. The original implmentation of getSchedInfoStr() introduced a
subtle layering violation (reported as PR37160 and then fixed/worked-around by
r330615).
In retrospect, that new API could have been designed more optimally. We can
always query MCSchedModel to get the latency and throughput. More importantly,
the "sched-info" string should not have been generated by the subtarget.
Note, r317782 fixed an issue where "print-schedule" didn't work very well in the
presence of inline assembly. That commit is also reverted by this change.
Differential Revision: https://reviews.llvm.org/D57244
llvm-svn: 353043
This patch adds a new ReadAdvance definition named ReadInt2Fpu.
ReadInt2Fpu allows x86 scheduling models to accurately describe delays caused by
data transfers from the integer unit to the floating point unit.
ReadInt2Fpu currently defaults to a delay of zero cycles (i.e. no delay) for all
x86 models excluding BtVer2. That means, this patch is only a functional change
for the Jaguar cpu model only.
Tablegen definitions for instructions (V)PINSR* have been updated to account for
the new ReadInt2Fpu. That read is mapped to the the GPR input operand.
On Jaguar, int-to-fpu transfers are modeled as a +6cy delay. Before this patch,
that extra delay was added to the opcode latency. In practice, the insert opcode
only executes for 1cy. Most of the actual latency is actually contributed by the
so-called operand-latency. According to the AMD SOG for family 16h, (V)PINSR*
latency is defined by expression f+1, where f is defined as a forwarding delay
from the integer unit to the fpu.
When printing instruction latency from MCA (see InstructionInfoView.cpp) and LLC
(only when flag -print-schedule is speified), we now need to account for any
extra forwarding delays. We do this by checking if scheduling classes declare
any negative ReadAdvance entries. Quoting a code comment in TargetSchedule.td:
"A negative advance effectively increases latency, which may be used for
cross-domain stalls". When computing the instruction latency for the purpose of
our scheduling tests, we now add any extra delay to the formula. This avoids
regressing existing codegen and mca schedule tests. It comes with the cost of an
extra (but very simple) hook in MCSchedModel.
Differential Revision: https://reviews.llvm.org/D57056
llvm-svn: 351965
to reflect the new license.
We understand that people may be surprised that we're moving the header
entirely to discuss the new license. We checked this carefully with the
Foundation's lawyer and we believe this is the correct approach.
Essentially, all code in the project is now made available by the LLVM
project under our new license, so you will see that the license headers
include that license only. Some of our contributors have contributed
code under our old license, and accordingly, we have retained a copy of
our old license notice in the top-level files in each project and
repository.
llvm-svn: 351636
Field ResourceUnitMask was incorrectly defined as a 'const unsigned' mask. It
should have been a 64 bit quantity instead. That means, ResourceUnitMask was
always implicitly truncated to a 32 bit quantity.
This issue has been found by inspection. Surprisingly, that bug was latent, and
it never negatively affected any existing upstream targets.
This patch fixes the wrong definition of ResourceUnitMask, and adds a bunch of
extra debug prints to help debugging potential issues related to invalid
processor resource masks.
llvm-svn: 350820
Summary:
It's a bit tricky to add a test for the failing path right now, binary support will have an easier path to exercise the path here.
* Ran clang-format.
Reviewers: andreadb
Reviewed By: andreadb
Subscribers: tschuett, gbedwell, llvm-commits
Differential Revision: https://reviews.llvm.org/D55803
llvm-svn: 349659
Class InstrBuilder wrongly assumed that llvm targets were always able to return
a non-null pointer when createMCInstrAnalysis() was called on them.
This was causing crashes when simulating executions for targets that don't
provide an MCInstrAnalysis object.
This patch fixes the issue by making MCInstrAnalysis optional.
llvm-svn: 349352
This patch removes a (potentially) slow while loop in
DefaultResourceStrategy::select(). A better (and faster) approach is to do some
bit manipulation in order to shrink the range of candidate resources.
On a release build, this change gives an average speedup of ~10%.
llvm-svn: 348007
This patch adds the ability to specify via tablegen which processor resources
are load/store queue resources.
A new tablegen class named MemoryQueue can be optionally used to mark resources
that model load/store queues. Information about the load/store queue is
collected at 'CodeGenSchedule' stage, and analyzed by the 'SubtargetEmitter' to
initialize two new fields in struct MCExtraProcessorInfo named `LoadQueueID` and
`StoreQueueID`. Those two fields are identifiers for buffered resources used to
describe the load queue and the store queue.
Field `BufferSize` is interpreted as the number of entries in the queue, while
the number of units is a throughput indicator (i.e. number of available pickers
for loads/stores).
At construction time, LSUnit in llvm-mca checks for the presence of extra
processor information (i.e. MCExtraProcessorInfo) in the scheduling model. If
that information is available, and fields LoadQueueID and StoreQueueID are set
to a value different than zero (i.e. the invalid processor resource index), then
LSUnit initializes its LoadQueue/StoreQueue based on the BufferSize value
declared by the two processor resources.
With this patch, we more accurately track dynamic dispatch stalls caused by the
lack of LS tokens (i.e. load/store queue full). This is also shown by the
differences in two BdVer2 tests. Stalls that were previously classified as
generic SCHEDULER FULL stalls, are not correctly classified either as "load
queue full" or "store queue full".
About the differences in the -scheduler-stats view: those differences are
expected, because entries in the load/store queue are not released at
instruction issue stage. Instead, those are released at instruction executed
stage. This is the main reason why for the modified tests, the load/store
queues gets full before PdEx is full.
Differential Revision: https://reviews.llvm.org/D54957
llvm-svn: 347857
This reapplies r347767 (originally reviewed at: https://reviews.llvm.org/D55000)
with a fix for the missing std::move of the Error returned by the call to
Pipeline::runCycle().
Below is the original commit message from r347767.
If a user only cares about the overall latency, then the best/quickest way is to
change method Pipeline::run() so that it returns the total number of cycles to
the caller.
When the simulation pipeline is run, the number of cycles (or an error) is
returned from method Pipeline::run().
The advantage is that no hardware event listener is needed for computing that
latency. So, the whole process should be faster (and simpler - at least for that
particular use case).
llvm-svn: 347795
If a user only cares about the overall latency, then the best/quickest way is to
change method Pipeline::run() so that it returns the total number of cycles to
the caller.
When the simulation pipeline is run, the number of cycles (or an error) is
returned from method Pipeline::run().
The advantage is that no hardware event listener is needed for computing that
latency. So, the whole process should be faster (and simpler - at least for that
particular use case).
llvm-svn: 347767
By default, llvm-mca conservatively assumes that a register operand from the
variadic sequence is both a register read and a register write. That is because
MCInstrDesc doesn't describe extra variadic operands; we don't have enough
dataflow information to tell which register operands from the variadic sequence
is a definition, and which is a use instead.
However, if a variadic instruction is flagged 'mayStore' (but not 'mayLoad'),
and it has no 'unmodeledSideEffects', then llvm-mca (very) optimistically
assumes that any register operand in the variadic sequence is a register read
only. Conversely, if a variadic instruction is marked as 'mayLoad' (but not
'mayStore'), and it has no 'unmodeledSideEffects', then llvm-mca optimistically
assumes that any extra register operand is a register definition only.
These assumptions work quite well for variadic load/store multiple instructions
defined by the ARM backend.
llvm-svn: 347522
With this change, InstrBuilder emits an error if the MCInst sequence contains an
instruction with a variadic opcode, and a non-zero number of variadic operands.
Currently we don't know how to correctly analyze variadic opcodes. The problem
with variadic operands is that there is no information for them in the opcode
descriptor (i.e. MCInstrDesc). That means, we don't know which variadic operands
are defs, and which are uses.
In future, we could try to conservatively assume that any extra register
operands is both a register use and a register definition.
This patch fixes a subtle bug in the evaluation of read/write operands for ARM
VLD1 with implicit index update. Added test vld1-index-update.s
llvm-svn: 347503
RetireControlUnitStatistics now reports extra information about the ROB and the
avg/maximum number of entries consumed over the entire simulation.
Example:
Retire Control Unit - number of cycles where we saw N instructions retired:
[# retired], [# cycles]
0, 109 (17.9%)
1, 102 (16.7%)
2, 399 (65.4%)
Total ROB Entries: 64
Max Used ROB Entries: 35 ( 54.7% )
Average Used ROB Entries per cy: 32 ( 50.0% )
Documentation in llvm/docs/CommandGuide/llvmn-mca.rst has been updated to
reflect this change.
llvm-svn: 347493
Also, try to minimize the number of queries to the memory queues to speedup the
analysis.
On average, this change gives a small 2% speedup. For memcpy-like kernels, the
speedup is up to 5.5%.
llvm-svn: 347469
This avoids a heap allocation most of the times.
This patch gives a small but consistent 3% speedup on a release build (up to ~5%
on a debug build).
llvm-svn: 347464
This patch fixes an invalid memory read introduced by r346487.
Before this patch, partial register write had to query the latency of the
dependent full register write by calling a method on the full write descriptor.
However, if the full write is from an already retired instruction, chances are
that the EntryStage already reclaimed its memory.
In some parial register write tests, valgrind was reporting an invalid
memory read.
This change fixes the invalid memory access problem. Writes are now responsible
for tracking dependent partial register writes, and notify them in the event of
instruction issued.
That means, partial register writes no longer need to query their associated
full write to check when they are ready to execute.
Added test X86/BtVer2/partial-reg-update-7.s
llvm-svn: 347459
When looking at the tests committed by Roman at r346587, I noticed that numbers
reported by the resource pressure for PdAGU01 were wrong.
In particular, according to the aut-generated CHECK lines in tests
memcpy-like-test.s and store-throughput.s, resource pressure for PdAGU01
was not uniformly distributed among the two AGEN pipes.
It turns out that the reason why pressure was not correctly distributed, was
because the "resource selection strategy" object associated with PdAGU01 was not
correctly updated on the event of AGEN pipe used.
As a result, llvm-mca was not simulating a round-robin pipeline allocation for
PdAGU01. Instead, PdAGU1 was always prioritized over PdAGU0.
This patch fixes the issue; now processor resource strategy objects for
resources declaring multiple units, are correctly notified in the event of
"resource used".
llvm-svn: 346650
This was noticed when working on PR3946.
By construction, a group cannot be used as a "Super" resource. That constraint
is enforced by method `SubtargetEmitter::ExpandProcResource()`.
A Super resource S can be part of a group G. However, method
`SubtargetEmitter::ExpandProcResource()` would not update the number of
consumed resource cycles in G based on S.
In practice, this is perfectly fine because the resource usage is correctly
computed for processor resource units. However, llvm-mca should still check if G
is a buffered resource.
Before this patch, llvm-mca didn't correctly check if S was part of a group that
defines a buffer. So, the instruction descriptor was not correctly set.
For now, the semantic change introduced by this patch doesn't affect any of the
upstream scheduling models. However, it will allow to make some progress on PR3946.
llvm-svn: 346545
Use a simple SmallVector to track the lifetime of simulated instructions.
An ordered map was not needed because instructions are already picked in program
order. It is also much faster if we avoid searching for already retired
instructions at the end of every cycle.
The new policy only triggers a "garbage collection" when the number of retired
instructions becomes significantly big when compared with the total size of the
vector.
While working on this, I noticed that instructions were correctly retired, but
their internal state was not updated (i.e. there was no transition from the
EXECUTED state, to the RETIRED state). While this was not a problem for the
views, it prevented the EntryStage from correctly garbage collecting already
retired instructions. That was a bad oversight, and this patch fixes it.
The observed speedup on a debug build of llvm-mca after this patch is ~6%.
On a release build of llvm-mca, the observed speedup is ~%15%.
llvm-svn: 346487
This fixes PR39261.
FetchStage is a misnomer. It causes confusion with the frontend fetch stage,
which we don't currently simulate. I decided to rename it into EntryStage
mainly because this is meant to be a "source" stage for all pipelines.
Differential Revision: https://reviews.llvm.org/D54268
llvm-svn: 346419
Summary:
This patch introduces a CodeRegionGenerator class which is responsible for parsing some type of input and creating a 'CodeRegions' instance for use by llvm-mca. In the future, we will also have a CodeRegionGenerator subclass for converting an input object file into CodeRegions. For now, we only have the subclass for converting input assembly into CodeRegions.
This is mostly a NFC patch, as the logic remains close to the original, but now encapsulated in its own class and moved outside of llvm-mca.cpp.
Reviewers: andreadb, courbet, RKSimon
Reviewed By: andreadb
Subscribers: mgorny, tschuett, gbedwell, llvm-commits
Differential Revision: https://reviews.llvm.org/D54179
llvm-svn: 346344
This patch teaches view RegisterFileStatistics how to report events for
optimizable register moves.
For each processor register file, view RegisterFileStatistics reports the
following extra information:
- Number of optimizable register moves
- Number of register moves eliminated
- Number of zero moves (i.e. register moves that propagate a zero)
- Max Number of moves eliminated per cycle.
Differential Revision: https://reviews.llvm.org/D53976
llvm-svn: 345865
