llvm-project/llvm/tools/llvm-mca

llvm-mca - LLVM Machine Code Analyzer
-------------------------------------

llvm-mca is a performance analysis tool that uses information which is already
available in LLVM (e.g. scheduling models) to statically measure the performance
of machine code in a specific cpu.

Performance is measured in terms of throughput as well as processor resource
consumption.  The tool currently works for processors with an out-of-order
backend, for which there is a scheduling model available in LLVM.

The main goal of this tool is not just to predict the performance of the code
when run on the target, but also help with diagnosing potential performance
issues.

Given an assembly code sequence, llvm-mca estimates the IPC (instructions Per
cycle), as well as hardware resources pressure. The analysis and reporting style
were inspired by the IACA tool from Intel.

The presence of long data dependency chains, as well as poor usage of hardware
resources may lead to bottlenecks in the back-end.  The tool is able to generate
a detailed report which should help with identifying and analyzing sources of
bottlenecks.

Scheduling models are mostly used to compute instruction latencies, to obtain
read-advance information, and understand how processor resources are used by
instructions.  By design, the quality of the performance analysis conducted by
the tool is inevitably affected by the quality of the target scheduling models. 

However, scheduling models intentionally do not describe all processors details,
since the goal is just to enable the scheduling of machine instructions during
compilation. That means, there are processor details which are not important for
the purpose of scheduling instructions (and therefore not described by the
scheduling model), but are very important for this tool.

A few examples of details that are missing in scheduling models are:
 - Actual dispatch width (it often differs from the issue width).
 - Number of read/write ports in the register file(s).
 - Length of the load/store queue in the LSUnit.

It is also very difficult to find a "good" abstract model to describe the
behavior of out-of-order processors. So, we have to keep in mind that all of
these aspects are going to affect the quality of the static analysis performed
by the tool.

An extensive list of known limitations is reported in one of the last sections
of this document. There is also a section related to design problems which must
be addressed (hopefully with the help of the community).  At the moment, the
tool has been mostly tested for x86 targets, but there are still several
limitations, some of which could be overcome by integrating extra information
into the scheduling models.

How the tool works
------------------

The tool takes assembly code as input. Assembly code is parsed into a sequence
of MCInst with the help of the existing LLVM target assembly parsers. The parsed
sequence of MCInst is then analyzed by a 'Backend' module to generate a
performance report.

The Backend module internally emulates the execution of the machine code
sequence in a loop of iterations (which by default is 100). At the end of this
process, the backend collects a number of statistics which are then printed out
in the form of a report.

Here is an example of performance report generated by the tool for a dot-product
of two packed float vectors of four elements. The analysis is conducted for
target x86, cpu btver2:

///////////////////

Iterations:     300
Instructions:   900
Total Cycles:   610
Dispatch Width: 2
IPC:            1.48


Resources:
[0] - JALU0
[1] - JALU1
[2] - JDiv
[3] - JFPM
[4] - JFPU0
[5] - JFPU1
[6] - JLAGU
[7] - JSAGU
[8] - JSTC
[9] - JVIMUL


Resource pressure per iteration:
[0]    [1]    [2]    [3]    [4]    [5]    [6]    [7]    [8]    [9]    
 -      -      -      -     2.00   1.00    -      -      -      -     

Resource pressure by instruction:
[0]    [1]    [2]    [3]    [4]    [5]    [6]    [7]    [8]    [9]    	Instructions:
 -      -      -      -      -     1.00    -      -      -      -     	vmulps	%xmm0, %xmm1, %xmm2
 -      -      -      -     1.00    -      -      -      -      -     	vhaddps	%xmm2, %xmm2, %xmm3
 -      -      -      -     1.00    -      -      -      -      -     	vhaddps	%xmm3, %xmm3, %xmm4


Instruction Info:
[1]: #uOps
[2]: Latency
[3]: RThroughput
[4]: MayLoad
[5]: MayStore
[6]: HasSideEffects

[1]    [2]    [3]    [4]    [5]    [6]	Instructions:
 1      2     1.00                    	vmulps	%xmm0, %xmm1, %xmm2
 1      3     1.00                    	vhaddps	%xmm2, %xmm2, %xmm3
 1      3     1.00                    	vhaddps	%xmm3, %xmm3, %xmm4

///////////////////

According to this report, the dot-product kernel has been executed 300 times,
for a total of 900 instructions dynamically executed.

The report is structured in three main sections.  A first section collects a few
performance numbers; the goal of this section is to give a very quick overview
of the performance throughput. In this example, the two important perforamce
indicators are a) the predicted total number of cycles, and b) the IPC. 
IPC is probably the most important throughput indicator. A big delta between the
Dispatch Width and the computed IPC is an indicator of potential performance
issues.

The second section is the so-called "resource pressure view".  This view reports
the average number of resource cycles consumed every iteration by instructions
for every processor resource unit available on the target.  Information is
structured in two tables. The first table reports the number of resource cycles
spent on average every iteration. The second table correlates the resource
cycles to the machine instruction in the sequence. For example, every iteration
of the dot-product, instruction 'vmulps' always executes on resource unit [5]
(JFPU1 - floating point pipeline #1), consuming an average of 1 resource cycle
per iteration.  Note that on Jaguar, vector FP multiply can only be issued to
pipeline JFPU1, while horizontal FP adds can only be issued to pipeline JFPU0.

The third (and last) section of the report shows the latency and reciprocal
throughput of every instruction in the sequence. That section also reports extra
information related to the number of micro opcodes, and opcode properties (i.e.
'MayLoad', 'MayStore' and 'UnmodeledSideEffects').

The resource pressure view helps with identifying bottlenecks caused by high
usage of specific hardware resources.  Situations with resource pressure mainly
concentrated on a few resources should, in general, be avoided.  Ideally,
pressure should be uniformly distributed between multiple resources.

Timeline View
-------------

A detailed report of each instruction's state transitions over time can be
enabled using the command line flag '-timeline'.  This prints an extra section
in the report which contains the so-called "timeline view".  Below is the
timeline view for the dot-product example from the previous section.

///////////////
Timeline view:
     	          012345
Index	0123456789      

[0,0]	DeeER.    .    .	vmulps	%xmm0, %xmm1, %xmm2
[0,1]	D==eeeER  .    .	vhaddps	%xmm2, %xmm2, %xmm3
[0,2]	.D====eeeER    .	vhaddps	%xmm3, %xmm3, %xmm4

[1,0]	.DeeE-----R    .	vmulps	%xmm0, %xmm1, %xmm2
[1,1]	. D=eeeE---R   .	vhaddps	%xmm2, %xmm2, %xmm3
[1,2]	. D====eeeER   .	vhaddps	%xmm3, %xmm3, %xmm4

[2,0]	.  DeeE-----R  .	vmulps	%xmm0, %xmm1, %xmm2
[2,1]	.  D====eeeER  .	vhaddps	%xmm2, %xmm2, %xmm3
[2,2]	.   D======eeeER	vhaddps	%xmm3, %xmm3, %xmm4


Average Wait times (based on the timeline view):
[0]: Executions
[1]: Average time spent waiting in a scheduler's queue
[2]: Average time spent waiting in a scheduler's queue while ready
[3]: Average time elapsed from WB until retire stage

      [0]    [1]    [2]    [3]
0.     3     1.0    1.0    3.3  	vmulps	%xmm0, %xmm1, %xmm2
1.     3     3.3    0.7    1.0  	vhaddps	%xmm2, %xmm2, %xmm3
2.     3     5.7    0.0    0.0  	vhaddps	%xmm3, %xmm3, %xmm4
///////////////

The timeline view is very interesting because it shows how instructions changed
in state during execution.  It also gives an idea of how the tool "sees"
instructions executed on the target.

The timeline view is structured in two tables.  The first table shows how
instructions change in state over time (measured in cycles); the second table
(named "Average Wait times") reports useful timing statistics which should help
diagnose performance bottlenecks caused by long data dependencies and
sub-optimal usage of hardware resources.

An instruction in the timeline view is identified by a pair of indices, where
the 'first' index identifies an iteration, and the 'second' index is the actual
instruction index (i.e. where it appears in the code sequence).

Excluding the first and last column, the remaining columns are in cycles.  Cycles
are numbered sequentially starting from 0.  The following characters are used to
describe the state of an instruction:

 D : Instruction dispatched.
 e : Instruction executing.
 E : Instruction executed.
 R : Instruction retired.
 = : Instruction already dispatched, waiting to be executed.
 - : Instruction executed, waiting to be retired.

Based on the timeline view from the example, we know that:
  - Instruction [1, 0] was dispatched at cycle 1.
  - Instruction [1, 0] started executing at cycle 2.
  - Instruction [1, 0] reached the write back stage at cycle 4.
  - Instruction [1, 0] was retired at cycle 10.

Instruction [1, 0] (i.e. the vmulps from iteration #1) doesn't have to wait in
the scheduler's queue for the operands to become available. By the time the
vmulps is dispatched, operands are already available, and pipeline JFPU1 is
ready to serve another instruction.  So the instruction can be immediately
issued on the JFPU1 pipeline. That is demonstrated by the fact that the
instruction only spent 1cy in the scheduler's queue.

There is a gap of 5 cycles between the write-back stage and the retire event.
That is because instructions must retire in program order, so [1,0] has to wait
for [0, 2] to be retired first (i.e it has to wait unti cycle 10).

In the dot-product example, all instructions are in a RAW (Read After Write)
dependency chain.  Register %xmm2 written by the vmulps is immediately used by
the first vhaddps, and register %xmm3 written by the first vhaddps is used by
the second vhaddps.  Long data dependencies negatively affect the ILP
(Instruction Level Parallelism).

In the dot-product example, there are anti-dependencies introduced by
instructions from different iterations.  However, those dependencies can be
removed at register renaming stage (at the cost of allocating register aliases,
and therefore consuming temporary registers).

Table "Average Wait times" helps diagnose performance issues that are caused by
the presence of long latency instructions and potentially long data dependencies
which may limit the ILP.  Note that the tool by default assumes at least 1cy
between the dispatch event and the issue event.

When the performance is limited by data dependencies and/or long latency
instructions, the number of cycles spent while in the "ready" state is expected
to be very small when compared with the total number of cycles spent in the
scheduler's queue.  So the difference between the two counters is a good
indicator of how big of an impact data dependencies had on the execution of
instructions.  When performance is mostly limited by the lack of hardware
resources, the delta between the two counters is small.  However, the number of
cycles spent in the queue tends to be bigger (i.e. more than 1-3cy) especially
when compared with other low latency instructions.

Extra statistics to further diagnose performance issues.
--------------------------------------------------------

Flag '-verbose' enables extra statistics and performance counters for the
dispatch logic, the reorder buffer, the retire control unit and the register
file.

Below is an example of verbose output generated by the tool for the dot-product
example discussed in the previous sections.

///////////////////
Iterations:     300
Instructions:   900
Total Cycles:   610
Dispatch Width: 2
IPC:            1.48


Dynamic Dispatch Stall Cycles:
RAT     - Register unavailable:                      0
RCU     - Retire tokens unavailable:                 0
SCHEDQ  - Scheduler full:                            272
LQ      - Load queue full:                           0
SQ      - Store queue full:                          0
GROUP   - Static restrictions on the dispatch group: 0


Register Alias Table:
Total number of mappings created: 900
Max number of mappings used:      35


Dispatch Logic - number of cycles where we saw N instructions dispatched:
[# dispatched], [# cycles]
 0,              24  (3.9%)
 1,              272  (44.6%)
 2,              314  (51.5%)


Schedulers - number of cycles where we saw N instructions issued:
[# issued], [# cycles]
 0,          7  (1.1%)
 1,          306  (50.2%)
 2,          297  (48.7%)


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%)


Scheduler's queue usage:
JALU01,  0/20
JFPU01,  18/18
JLSAGU,  0/12
///////////////////

Based on the verbose report, the backend was only able to dispatch two
instructions 51.5% of the time.  The dispatch group was limited to one
instruction 44.6% of the cycles, which corresponds to 272 cycles.

If we look at section "Dynamic Dispatch Stall Cycles", we can see how counter
SCHEDQ reports 272 cycles.  Counter SCHEDQ is incremented every time the dispatch
logic is unable to dispatch a full group of two instructions because the
scheduler's queue is full.

Section "Scheduler's queue usage" shows how the maximum number of buffer entries
(i.e. scheduler's queue entries) used at runtime for resource JFPU01 reached its
maximum. Note that AMD Jaguar implements three schedulers:
  * JALU01 - A scheduler for ALU instructions
  * JLSAGU - A scheduler for address generation
  * JFPU01 - A scheduler floating point operations.

The dot-product is a kernel of three floating point instructions (a vector
multiply followed by two horizontal adds).  That explains why only the floating
point scheduler appears to be used according to section "Scheduler's queue
usage".

A full scheduler's queue is either caused by data dependency chains, or by a
sub-optimal usage of hardware resources.  Sometimes, resource pressure can be
mitigated by rewriting the kernel using different instructions that consume
different scheduler resources.  Schedulers with a small queue are less resilient
to bottlenecks caused by the presence of long data dependencies.

In this example, we can conclude that the IPC is mostly limited by data
dependencies, and not by resource pressure.

LLVM-MCA instruction flow
-------------------------

This section describes the instruction flow through the out-of-order backend, as
well as the functional units involved in the process. 

An instruction goes through a default sequence of stages:
    - Dispatch (Instruction is dispatched to the schedulers).
    - Issue (Instruction is issued to the processor pipelines).
    - Write Back (Instruction is executed, and results are written back).
    - Retire (Instruction is retired; writes are architecturally committed).

The tool only models the out-of-order portion of a processor. Therefore, the
instruction fetch and decode stages are not modeled. Performance bottlenecks in
the frontend are not diagnosed by this tool.  The tool assumes that instructions
have all been decoded and placed in a queue.  Also, the tool doesn't know
anything about branch prediction.

The long term plan is to make the process customizable, so that processors can
define their own. This is a future work.

Instruction Dispatch
--------------------

During the Dispatch stage, instructions are picked in program order from a queue
of already decoded instructions, and dispatched in groups to the hardware
schedulers.  The dispatch logic is implemented by class DispatchUnit in file
Dispatch.h.

The size of a dispatch group depends on the availability of hardware resources,
and it cannot exceed the value of field 'DispatchWidth' in class DispatchUnit.
Note that field DispatchWidth defaults to the value of field 'IssueWidth' from
the scheduling model.

Users can override the DispatchWidth value with flag "-dispatch=<N>" (where 'N'
is an unsigned quantity).

An instruction can be dispatched if:
 - The size of the dispatch group is smaller than DispatchWidth
 - There are enough entries in the reorder buffer
 - There are enough temporary registers to do register renaming
 - Schedulers are not full.

Since r329067, scheduling models can now optionally specify which register files
are available on the processor. Class DispatchUnit(see Dispatch.h) would use
that information to initialize register file descriptors.

By default, if the model doesn't describe register files, the tool
(optimistically) assumes a single register file with an unbounded number of
temporary registers.  Users can limit the number of temporary registers that are
globally available for register renaming using flag `-register-file-size=<N>`,
where N is the number of temporaries.  A value of zero for N means 'unbounded'.
Knowing how many temporaries are available for register renaming, the tool can
predict dispatch stalls caused by the lack of temporaries.

The number of reorder buffer entries consumed by an instruction depends on the
number of micro-opcodes it specifies in the target scheduling model (see field
'NumMicroOpcodes' of tablegen class ProcWriteResources and its derived classes;
TargetSchedule.td).

The reorder buffer is implemented by class RetireControlUnit (see Dispatch.h).
Its goal is to track the progress of instructions that are "in-flight", and
retire instructions in program order.  The number of entries in the reorder
buffer defaults to the value of field 'MicroOpBufferSize' from the target
scheduling model.

Instructions that are dispatched to the schedulers consume scheduler buffer
entries.  The tool queries the scheduling model to figure out the set of
buffered resources consumed by an instruction.  Buffered resources are treated
like "scheduler" resources, and the field 'BufferSize' (from the processor
resource tablegen definition) defines the size of the scheduler's queue.

Zero latency instructions (for example NOP instructions) don't consume scheduler
resources.  However, those instructions still reserve a number of slots in the
reorder buffer.

Instruction Issue
-----------------

As mentioned in the previous section, each scheduler resource implements a queue
of instructions.  An instruction has to wait in the scheduler's queue until
input register operands become available.  Only at that point, does the
instruction becomes eligible for execution and may be issued (potentially
out-of-order) to a pipeline for execution.

Instruction latencies can be computed by the tool with the help of the
scheduling model; latency values are defined by the scheduling model through
ProcWriteResources objects.

Class Scheduler (see file Scheduler.h) knows how to emulate multiple processor
schedulers.  A Scheduler is responsible for tracking data dependencies, and
dynamically select which processor resources are consumed/used by instructions.

Internally, the Scheduler class delegates the management of processor resource
units and resource groups to the ResourceManager class.  ResourceManager is also
responsible for selecting resource units that are effectively consumed by
instructions.  For example, if an instruction consumes 1cy of a resource group,
the ResourceManager object selects one of the available units from the group; by
default, it uses a round-robin selector to guarantee that resource usage is
uniformly distributed between all units of a group.

Internally, class Scheduler implements three instruction queues:
  - WaitQueue: a queue of instructions whose operands are not ready yet.
  - ReadyQueue: a queue of instructions ready to execute.
  - IssuedQueue: a queue of instructions executing.

Depending on the operands availability, instructions that are dispatched to the
Scheduler are either placed into the WaitQueue or into the ReadyQueue.

Every cycle, class Scheduler checks if instructions can be moved from the
WaitQueue to the ReadyQueue, and if instructions from the ReadyQueue can be
issued to the underlying pipelines.  The algorithm prioritizes older
instructions over younger instructions.

Objects of class ResourceState (see Scheduler.h) describe processor resources.
There is an instance of class ResourceState for each single processor resource
specified by the scheduling model.  A ResourceState object for a processor
resource with multiple units dynamically tracks the availability of every single
unit.  For example, the ResourceState of a resource group tracks the
availability of every resource in that group.  Internally, ResourceState
implements a round-robin selector to dynamically pick the next unit to use from
the group.

Write-Back and Retire Stage
---------------------------

Issued instructions are moved from the ReadyQueue to the IssuedQueue.  There,
instructions wait until they reach the write-back stage.  At that point, they
get removed from the queue and the retire control unit is notified.

On the event of "instruction executed", the retire control unit flags the
instruction as "ready to retire".

Instruction are retired in program order; an "instruction retired" event is sent
to the register file which frees the temporary registers allocated for the
instruction at register renaming stage.

Load/Store Unit and Memory Consistency Model
--------------------------------------------

The tool attempts to emulate out-of-order execution of memory operations.  Class
LSUnit (see file LSUnit.h) emulates a load/store unit implementing queues for
speculative execution of loads and stores.
 
Each load (or store) consumes an entry in the load (or store) queue.  The number
of slots in the load/store queues is unknown by the tool, since there is no
mention of it in the scheduling model.  In practice, users can specify flag
`-lqueue=N` (vic. `-squeue=N`) to limit the number of entries in the queue to be
equal to exactly N (an unsigned value). If N is zero, then the tool assumes an
unbounded queue (this is the default).

LSUnit implements a relaxed consistency model for memory loads and stores. The
rules are:
1) A younger load is allowed to pass an older load only if there is no
   intervening store in between the two loads.
2) An younger store is not allowed to pass an older store.
3) A younger store is not allowed to pass an older load.
4) A younger load is allowed to pass an older store provided that the load does
   not alias with the store.

By default, this class conservatively (i.e. pessimistically) assumes that loads
always may-alias store operations.  Essentially, this LSUnit doesn't perform any
sort of alias analysis to rule out cases where loads and stores don't overlap
with each other.  The downside of this approach however is that younger loads are
never allowed to pass older stores.  To make it possible for a younger load to
pass an older store, users can use the command line flag -noalias.  Under
'noalias', a younger load is always allowed to pass an older store.

Note that, in the case of write-combining memory, rule 2. could be relaxed a bit
to allow reordering of non-aliasing store operations.  That being said, at the
moment, there is no way to further relax the memory model (flag -noalias is the
only option).  Essentially, there is no option to specify a different memory
type (for example: write-back, write-combining, write-through; etc.) and
consequently to weaken or strengthen the memory model.

Other limitations are:
 * LSUnit doesn't know when store-to-load forwarding may occur.
 * LSUnit doesn't know anything about the cache hierarchy and memory types.
 * LSUnit doesn't know how to identify serializing operations and memory fences.

No assumption is made on the store buffer size.  As mentioned before, LSUnit
conservatively assumes a may-alias relation between loads and stores, and it
doesn't attempt to identify cases where store-to-load forwarding would occur in
practice.

LSUnit doesn't attempt to predict whether a load or store hits or misses the L1
cache.  It only knows if an instruction "MayLoad" and/or "MayStore".  For loads,
the scheduling model provides an "optimistic" load-to-use latency (which usually
matches the load-to-use latency for when there is a hit in the L1D).

Class MCInstrDesc in LLVM doesn't know about serializing operations, nor
memory-barrier like instructions.  LSUnit conservatively assumes that an
instruction which has both 'MayLoad' and 'UnmodeledSideEffects' behaves like a
"soft" load-barrier.  That means, it serializes loads without forcing a flush of
the load queue.  Similarly, instructions flagged with both 'MayStore' and
'UnmodeledSideEffects' are treated like store barriers.  A full memory barrier
is a 'MayLoad' and 'MayStore' instruction with 'UnmodeledSideEffects'.  This is
inaccurate, but it is the best that we can do at the moment with the current
information available in LLVM.

A load/store barrier consumes one entry of the load/store queue.  A load/store
barrier enforces ordering of loads/stores.  A younger load cannot pass a load
barrier.  Also, a younger store cannot pass a store barrier.  A younger load has
to wait for the memory/load barrier to execute.  A load/store barrier is
"executed" when it becomes the oldest entry in the load/store queue(s). That
also means, by construction, all the older loads/stores have been executed.

In conclusion the full set of rules is:
  1. A store may not pass a previous store.
  2. A load may not pass a previous store unless flag 'NoAlias' is set.
  3. A load may pass a previous load.
  4. A store may not pass a previous load (regardless of flag 'NoAlias').
  5. A load has to wait until an older load barrier is fully executed.
  6. A store has to wait until an older store barrier is fully executed.

Known limitations
-----------------
Previous sections described cases where the tool is missing information to give
an accurate report.  For example, the first sections of this document explained
how the lack of knowledge about the processor negatively affects the performance
analysis.  The lack of knowledge is often a consequence of how scheduling models
are defined; as mentioned before, scheduling models intentionally don't describe
processors in fine details. That being said, the LLVM machine model can be
extended to expose more details, as long as they are opt-in for targets.

The accuracy of the performance analysis is also affected by assumptions made by
the processor model used by the tool.

Most recent Intel and AMD processors implement dedicated LoopBuffer/OpCache in
the hardware frontend to speedup the throughput in the presence of tight loops.
The presence of these buffers complicates the decoding logic, and requires
knowledge on the branch predictor too.  Class 'SchedMachineModel' in tablegen
provides a field named 'LoopMicroOpBufferSize' which is used to describe loop
buffers.  However, the purpose of that field is to enable loop unrolling of
tight loops; essentially, it affects the cost model used by pass loop-unroll.

At the current state, the tool only describes the out-of-order portion of a
processor, and consequently doesn't try to predict the frontend throughput. That
being said, this tool could be definitely extended in future to also account for
the hardware frontend when doing performance analysis.  This would inevitably
require extra (extensive) processor knowledge related to all the available
decoding paths in the hardware frontend, as well as branch prediction.

Currently, the tool assumes a zero-latency "perfect" fetch&decode
stage; the full sequence of decoded instructions is immediately visible to the
dispatch logic from the start.

The tool doesn't know about simultaneous mutithreading.  According to the tool,
processor resources are not statically/dynamically partitioned.  Processor
resources are fully available to the hardware thread executing the
microbenchmark.

The execution model implemented by this tool assumes that instructions are
firstly dispatched in groups to hardware schedulers, and then issued to
pipelines for execution.  The model assumes dynamic scheduling of instructions.
Instructions are placed in a queue and potentially executed out-of-order (based
on the operand availability). The dispatch stage is definitely distinct from the
issue stage. This will change in future; as mentioned in the first section, the
end goal is to let processors customize the process.

This model doesn't correctly describe processors where the dispatch/issue is a
single stage. This is what happens for example in VLIW processors, where
instructions are packaged and statically scheduled at compile time; it is up to
the compiler to predict the latency of instructions and package issue groups
accordingly. For such targets, there is no dynamic scheduling done by the
hardware.

Existing classes (DispatchUnit, Scheduler, etc.) could be extended/adapted to
support processors with a single dispatch/issue stage. The execution flow would
require some changes in the way how existing components (i.e.  DispatchUnit,
Scheduler, etc.) interact. This can be a future development.

The following sections describes other known limitations.  The goal is not to
provide an extensive list of limitations; we want to report what we believe are
the most important limitations, and suggest possible methods to overcome them.

Load/Store barrier instructions and serializing operations
----------------------------------------------------------
Section "Load/Store Unit and Memory Consistency Model" already mentioned how
LLVM doesn't know about serializing operations and memory barriers.  Most of it
boils down to the fact that class MCInstrDesc (intentionally) doesn't expose
those properties.  Instead, both serializing operations and memory barriers
"have side-effects" according to MCInstrDesc.  That is because, at least for
scheduling purposes, knowing that an instruction has unmodeled side effects is
often enough to treat the instruction like a compiler scheduling barrier.

A performance analysis tool could use the extra knowledge on barriers and
serializing operations to generate a more accurate performance report.  One way
to improve this is by reserving a couple of bits in field 'Flags' from class
MCInstrDesc: one bit for barrier operations, and another bit to mark
instructions as serializing operations.

Lack of support for instruction itineraries
-------------------------------------------
The current version of the tool doesn't know how to process instruction
itineraries.  This is probably one of the most important limitations, since it
affects a few out-of-order processors in LLVM.

As mentioned in section 'Instruction Issue', class Scheduler delegates to an
instance of class ResourceManager the handling of processor resources.
ResourceManager is where most of the scheduling logic is implemented.

Adding support for instruction itineraries requires that we teach
ResourceManager how to handle functional units and instruction stages.  This
development can be a future extension, and it would probably require a few
changes to the ResourceManager interface.

Instructions that affect control flow are not correctly modeled
---------------------------------------------------------------
Examples of instructions that affect the control flow are: return, indirect
branches, calls, etc.  The tool doesn't try to predict/evaluate branch targets.
In particular, the tool doesn't model any sort of branch prediction, nor does it
attempt to track changes to the program counter.  The tool always assumes that
the input assembly sequence is the body of a microbenchmark (a simple loop
executed for a number of iterations). The "next" instruction in sequence is
always the next instruction to dispatch.

Call instructions default to an arbitrary high latency of 100cy. A warning is
generated if the tool encounters a call instruction in the sequence.  Return
instructions are not evaluated, and therefore control flow is not affected.
However, the tool still queries the processor scheduling model to obtain latency
information for instructions that affect the control flow.

Known limitations on X86 processors
-----------------------------------

1) Partial register updates versus full register updates.

On x86-64, a 32-bit GPR write fully updates the super-register. Example:
      add %edi %eax    ## eax += edi

Here, register %eax aliases the lower half of 64-bit register %rax. On x86-64,
register %rax is fully updated by the 'add' (the upper half of %rax is zeroed).
Essentially, it "kills" any previous definition of (the upper half of) register
%rax.

On the other hand, 8/16 bit register writes only perform a so-called "partial
register update". Example:
      add %di, %ax     ## ax += di

Here, register %eax is only partially updated. To be more specific, the lower
half of %eax is set, and the upper half is left unchanged. There is also no
change in the upper 48 bits of register %rax.

To get accurate performance analysis, the tool has to know which instructions
perform a partial register update, and which instructions fully update the
destination's super-register.

One way to expose this information is (again) via tablegen.  For example, we
could add a flag in the tablegen instruction class to tag instructions that
perform partial register updates. Something like this: 'bit
hasPartialRegisterUpdate = 1'. However, this would force a `let
hasPartialRegisterUpdate = 0` on several instruction definitions.

Another approach is to have a MCSubtargetInfo hook similar to this:
    virtual bool updatesSuperRegisters(unsigned short opcode) { return false; }

Targets will be able to override this method if needed.  Again, this is just an
idea. But the plan is to have this fixed as a future development.

2) Macro Op fusion.

The tool doesn't know about macro-op fusion. On modern x86 processors, a
'cmp/test' followed by a 'jmp' is fused into a single macro operation.  The
advantage is that the fused pair only consumes a single slot in the dispatch
group. 

As a future development, the tool should be extended to address macro-fusion.
Ideally, we could have LLVM generate a table enumerating all the opcode pairs
that can be fused together. That table could be exposed to the tool via the
MCSubtargetInfo interface.  This is just an idea; there may be better ways to
implement this.

3) Intel processors: mixing legacy SSE with AVX instructions.

On modern Intel processors with AVX, mixing legacy SSE code with AVX code
negatively impacts the performance.  The tool is not aware of this issue, and
the performance penalty is not accounted when doing the analysis. This is
something that we would like to improve in future.

4) Zero-latency register moves and Zero-idioms.

Most modern AMD/Intel processors know how to optimize out register-register
moves and zero idioms at register renaming stage. The tool doesn't know
about these patterns, and this may negatively impact the performance analysis.

Known design problems
---------------------
This section describes two design issues that are currently affecting the tool.
The long term plan is to "fix" these issues.
Both limitations would be easily fixed if we teach the tool how to directly
manipulate MachineInstr objects (instead of MCInst objects).

1) Variant instructions not correctly modeled.

The tool doesn't know how to analyze instructions with a "variant" scheduling
class descriptor. A variant scheduling class needs to be resolved dynamically.
The "actual" scheduling class often depends on the subtarget, as well as
properties of the specific MachineInstr object.

Unfortunately, the tool manipulates MCInst, and it doesn't know anything about
MachineInstr. As a consequence, the tool cannot use the existing machine
subtarget hooks that are normally used to resolve the variant scheduling class.
This is a major design issue which mostly affects ARM/AArch64 targets.  It
mostly boils down to the fact that the existing scheduling framework was meant
to work for MachineInstr.

When the tool encounters a "variant" instruction, it assumes a generic 1cy
latency. However, the tool would not be able to tell which processor resources
are effectively consumed by the variant instruction.

2) MCInst and MCInstrDesc.

Performance analysis tools require data dependency information to correctly
predict the runtime performance of the code. This tool must always be able to
obtain the set of implicit/explicit register defs/uses for every instruction of
the input assembly sequence.

In the first section of this document, it was mentioned how the tool takes as
input an assembly sequence. That sequence is parsed into a MCInst sequence with
the help of assembly parsers available from the targets.

A MCInst is a very low-level instruction representation. The tool can inspect
the MCOperand sequence of an MCInst to identify register operands. However,
there is no way to tell register operands that are definitions from register
operands that are uses.

In LLVM, class MCInstrDesc is used to fully describe target instructions and
their operands. The opcode of a machine instruction (a MachineInstr object) can
be used to query the instruction set through method `MCInstrInfo::get' to obtain
the associated MCInstrDesc object.

However class MCInstrDesc describes properties and operands of MachineInstr
objects. Essentially, MCInstrDesc is not meant to be used to describe MCInst
objects.  To be more specific, MCInstrDesc objects are automatically generated
via tablegen from the instruction set description in the target .td files.  For
example, field `MCInstrDesc::NumDefs' is always equal to the cardinality of the
`(outs)` set from the tablegen instruction definition.

By construction, register definitions always appear at the beginning of the
MachineOperands list in MachineInstr. Basically, the (outs) are the first
operands of a MachineInstr, and the (ins) will come after in the machine operand
list. Knowing the number of register definitions is enough to identify
all the register operands that are definitions.

In a normal compilation process, MCInst objects are generated from MachineInstr
objects through a lowering step. By default the lowering logic simply iterates
over the machine operands of a MachineInstr, and converts/expands them into
equivalent MCOperand objects.

The default lowering strategy has the advantage of preserving all of the
above mentioned assumptions on the machine operand sequence. That means, register
definitions would still be at the beginning of the MCOperand sequence, and
register uses would come after.

Targets may still define custom lowering routines for specific opcodes. Some of
these routines may lower operands in a way that potentially breaks (some of) the
assumptions on the machine operand sequence which were valid for MachineInstr.
Luckily, this is not the most common form of lowering done by the targets, and
the vast majority of the MachineInstr are lowered based on the default strategy
which preserves the original machine operand sequence.  This is especially true
for x86, where the custom lowering logic always preserves the original (i.e.
from the MachineInstr) operand sequence.

This tool currently works under the strong (and potentially incorrect)
assumption that register def/uses in a MCInst can always be identified by
querying the machine instruction descriptor for the opcode. This assumption made
it possible to develop this tool and get good numbers at least for the
processors available in the x86 backend.

That being said, the analysis is still potentially incorrect for other targets.
So we plan (with the help of the community) to find a proper mechanism to map
when possible MCOperand indices back to MachineOperand indices of the equivalent
MachineInstr.  This would be equivalent to describing changes made by the
lowering step which affected the operand sequence. For example, we could have an
index for every register MCOperand (or -1, if the operand didn't exist in the
original MachineInstr). The mapping could look like this <0,1,3,2>.  Here,
MCOperand #2 was obtained from the lowering of MachineOperand #3. etc.

This information could be automatically generated via tablegen for all the
instructions whose custom lowering step breaks assumptions made by the tool on
the register operand sequence (In general, these instructions should be the
minority of a target's instruction set). Unfortunately, we don't have that
information now.  As a consequence, we assume that the number of explicit
register definitions is the same number specified in MCInstrDesc.  We also
assume that register definitions always come first in the operand sequence.

In conclusion: these are for now the strong assumptions made by the tool:
  * The number of explicit and implicit register definitions in a MCInst
    matches the number of explicit and implicit definitions specified by the
    MCInstrDesc object.
  * Register uses always come after register definitions.
  * If an opcode specifies an optional definition, then the optional
    definition is always the last register operand in the sequence.

Note that some of the information accessible from the MCInstrDesc is always
valid for MCInst. For example: implicit register defs, implicit register uses
and 'MayLoad/MayStore/HasUnmodeledSideEffects' opcode properties still apply to
MCInst. The tool knows about this, and uses that information during its
analysis.

Future work
-----------
 * Address limitations (described in section "Known limitations").
 * Let processors specify the selection strategy for processor resource groups
   and resources with multiple units. The tool currently uses a round-robin
   selector to pick the next resource to use.
 * Address limitations specifically described in section "Known limitations on
   X86 processors".
 * Address design issues identified in section "Known design problems".
 * Define a standard interface for "Views". This would let users customize the
   performance report generated by the tool.

When interfaces are mature/stable:
 * Move the logic into a library. This will enable a number of other
   interesting use cases.

Work is currently tracked on https://bugs.llvm.org. llvm-mca bugs are tagged
with prefix [llvm-mca]. You can easily find the full list of open bugs if you
search for that tag.