llvm.org GIT mirror llvm / release_90 docs / CommandGuide / llvm-mca.rst

Tree @release_90 (Download .tar.gz)

llvm-mca.rst @release_90view markup · raw · history · blame

llvm-mca - LLVM Machine Code Analyzer


:program:`llvm-mca` [options] [input]


:program:`llvm-mca` is a performance analysis tool that uses information 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, :program:`llvm-mca` estimates the Instructions Per Cycle (IPC), as well as hardware resource pressure. The analysis and reporting style were inspired by the IACA tool from Intel.

For example, you can compile code with clang, output assembly, and pipe it directly into :program:`llvm-mca` for analysis:

$ clang foo.c -O2 -target x86_64-unknown-unknown -S -o - | llvm-mca -mcpu=btver2

Or for Intel syntax:

$ clang foo.c -O2 -target x86_64-unknown-unknown -mllvm -x86-asm-syntax=intel -S -o - | llvm-mca -mcpu=btver2

Scheduling models are not just used to compute instruction latencies and throughput, but also to understand what processor resources are available and how to simulate them.

By design, the quality of the analysis conducted by :program:`llvm-mca` is inevitably affected by the quality of the scheduling models in LLVM.

If you see that the performance report is not accurate for a processor, please file a bug against the appropriate backend.


If input is "-" or omitted, :program:`llvm-mca` reads from standard input. Otherwise, it will read from the specified filename.

If the :option:`-o` option is omitted, then :program:`llvm-mca` will send its output to standard output if the input is from standard input. If the :option:`-o` option specifies "-", then the output will also be sent to standard output.


:program:`llvm-mca` returns 0 on success. Otherwise, an error message is printed to standard error, and the tool returns 1.


:program:`llvm-mca` allows for the optional usage of special code comments to mark regions of the assembly code to be analyzed. A comment starting with substring LLVM-MCA-BEGIN marks the beginning of a code region. A comment starting with substring LLVM-MCA-END marks the end of a code region. For example:

If no user-defined region is specified, then :program:`llvm-mca` assumes a default region which contains every instruction in the input file. Every region is analyzed in isolation, and the final performance report is the union of all the reports generated for every code region.

Code regions can have names. For example:

The code from the example above defines a region named "A simple example" with a single instruction in it. Note how the region name doesn't have to be repeated in the LLVM-MCA-END directive. In the absence of overlapping regions, an anonymous LLVM-MCA-END directive always ends the currently active user defined region.

Example of nesting regions:

Example of overlapping regions:

Note that multiple anonymous regions cannot overlap. Also, overlapping regions cannot have the same name.

There is no support for marking regions from high-level source code, like C or C++. As a workaround, inline assembly directives may be used:

int foo(int a, int b) {
  __asm volatile("# LLVM-MCA-BEGIN foo");
  a += 42;
  __asm volatile("# LLVM-MCA-END");
  a *= b;
  return a;

However, this interferes with optimizations like loop vectorization and may have an impact on the code generated. This is because the __asm statements are seen as real code having important side effects, which limits how the code around them can be transformed. If users want to make use of inline assembly to emit markers, then the recommendation is to always verify that the output assembly is equivalent to the assembly generated in the absence of markers. The Clang options to emit optimization reports can also help in detecting missed optimizations.


:program:`llvm-mca` takes assembly code as input. The 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 Pipeline module to generate a performance report.

The Pipeline module simulates the execution of the machine code sequence in a loop of iterations (default is 100). During this process, the pipeline collects a number of execution related statistics. At the end of this process, the pipeline generates and prints a report from the collected statistics.

Here is an example of a 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. The following result can be produced via the following command using the example located at test/tools/llvm-mca/X86/BtVer2/dot-product.s:

$ llvm-mca -mtriple=x86_64-unknown-unknown -mcpu=btver2 -iterations=300 dot-product.s

According to this report, the dot-product kernel has been executed 300 times, for a total of 900 simulated instructions. The total number of simulated micro opcodes (uOps) is also 900.

The report is structured in three main sections. The first section collects a few performance numbers; the goal of this section is to give a very quick overview of the performance throughput. Important performance indicators are IPC, uOps Per Cycle, and Block RThroughput (Block Reciprocal Throughput).

IPC is computed dividing the total number of simulated instructions by the total number of cycles. In the absence of loop-carried data dependencies, the observed IPC tends to a theoretical maximum which can be computed by dividing the number of instructions of a single iteration by the Block RThroughput.

Field 'uOps Per Cycle' is computed dividing the total number of simulated micro opcodes by the total number of cycles. A delta between Dispatch Width and this field is an indicator of a performance issue. In the absence of loop-carried data dependencies, the observed 'uOps Per Cycle' should tend to a theoretical maximum throughput which can be computed by dividing the number of uOps of a single iteration by the Block RThroughput.

Field uOps Per Cycle is bounded from above by the dispatch width. That is because the dispatch width limits the maximum size of a dispatch group. Both IPC and 'uOps Per Cycle' are limited by the amount of hardware parallelism. The availability of hardware resources affects the resource pressure distribution, and it limits the number of instructions that can be executed in parallel every cycle. A delta between Dispatch Width and the theoretical maximum uOps per Cycle (computed by dividing the number of uOps of a single iteration by the Block RTrhoughput) is an indicator of a performance bottleneck caused by the lack of hardware resources. In general, the lower the Block RThroughput, the better.

In this example, uOps per iteration/Block RThroughput is 1.50. Since there are no loop-carried dependencies, the observed uOps Per Cycle is expected to approach 1.50 when the number of iterations tends to infinity. The delta between the Dispatch Width (2.00), and the theoretical maximum uOp throughput (1.50) is an indicator of a performance bottleneck caused by the lack of hardware resources, and the Resource pressure view can help to identify the problematic resource usage.

The second 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 'HasSideEffects').

The third section is the 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 instruction vmulps always executes on resource unit [6] (JFPU1 - floating point pipeline #1), consuming an average of 1 resource cycle per iteration. Note that on AMD Jaguar, vector floating-point multiply can only be issued to pipeline JFPU1, while horizontal floating-point additions can only be issued to pipeline JFPU0.

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

The timeline view produces a detailed report of each instruction's state transitions through an instruction pipeline. This view is enabled by the command line option -timeline. As instructions transition through the various stages of the pipeline, their states are depicted in the view report. These states are represented by the following characters:

  • 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.

Below is the timeline view for a subset of the dot-product example located in test/tools/llvm-mca/X86/BtVer2/dot-product.s and processed by :program:`llvm-mca` using the following command:

$ llvm-mca -mtriple=x86_64-unknown-unknown -mcpu=btver2 -iterations=3 -timeline dot-product.s

The timeline view is interesting because it shows instruction state changes during execution. It also gives an idea of how the tool processes instructions executed on the target, and how their timing information might be calculated.

The timeline view is structured in two tables. The first table shows instructions changing 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 instruction index (i.e., where it appears in the code sequence). Since this example was generated using 3 iterations: -iterations=3, the iteration indices range from 0-2 inclusively.

Excluding the first and last column, the remaining columns are in cycles. Cycles are numbered sequentially starting from 0.

From the example output above, we know the following:

  • 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., vmulps from iteration #1) does not have to wait in the scheduler's queue for the operands to become available. By the time 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 until cycle 10).

In the example, all instructions are in a RAW (Read After Write) dependency chain. Register %xmm2 written by 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 impact 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 physical 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 :program:`llvm-mca`, 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. The difference between the two counters is a good indicator of how large of an impact data dependencies had on the execution of the 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 larger (i.e., more than 1-3cy), especially when compared to other low latency instructions.

Extra Statistics to Further Diagnose Performance Issues

The -all-stats command line option 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 -all-stats output generated by :program:`llvm-mca` for 300 iterations of the dot-product example discussed in the previous sections.

If we look at the Dynamic Dispatch Stall Cycles table, we see the counter for SCHEDQ reports 272 cycles. This counter is incremented every time the dispatch logic is unable to dispatch a full group because the scheduler's queue is full.

Looking at the Dispatch Logic table, we see that the pipeline was only able to dispatch two micro opcodes 51.5% of the time. The dispatch group was limited to one micro opcode 44.6% of the cycles, which corresponds to 272 cycles. The dispatch statistics are displayed by either using the command option -all-stats or -dispatch-stats.

The next table, Schedulers, presents a histogram displaying a count, representing the number of micro opcodes issued on some number of cycles. In this case, of the 610 simulated cycles, single opcodes were issued 306 times (50.2%) and there were 7 cycles where no opcodes were issued.

The Scheduler's queue usage table shows that the average and maximum number of buffer entries (i.e., scheduler queue entries) used at runtime. Resource JFPU01 reached its maximum (18 of 18 queue entries). Note that AMD Jaguar implements three schedulers:

  • JALU01 - A scheduler for ALU instructions.
  • JFPU01 - A scheduler floating point operations.
  • JLSAGU - A scheduler for address generation.

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.

A full scheduler 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. The scheduler statistics are displayed by using the command option -all-stats or -scheduler-stats.

The next table, Retire Control Unit, presents a histogram displaying a count, representing the number of instructions retired on some number of cycles. In this case, of the 610 simulated cycles, two instructions were retired during the same cycle 399 times (65.4%) and there were 109 cycles where no instructions were retired. The retire statistics are displayed by using the command option -all-stats or -retire-stats.

The last table presented is Register File statistics. Each physical register file (PRF) used by the pipeline is presented in this table. In the case of AMD Jaguar, there are two register files, one for floating-point registers (JFpuPRF) and one for integer registers (JIntegerPRF). The table shows that of the 900 instructions processed, there were 900 mappings created. Since this dot-product example utilized only floating point registers, the JFPuPRF was responsible for creating the 900 mappings. However, we see that the pipeline only used a maximum of 35 of 72 available register slots at any given time. We can conclude that the floating point PRF was the only register file used for the example, and that it was never resource constrained. The register file statistics are displayed by using the command option -all-stats or -register-file-stats.

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

Instruction Flow

This section describes the instruction flow through the default pipeline of :program:`llvm-mca`, as well as the functional units involved in the process.

The default pipeline implements the following sequence of stages used to process instructions.

  • 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 default pipeline 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. :program:`llvm-mca` assumes that instructions have all been decoded and placed into a queue before the simulation start. Also, :program:`llvm-mca` does not model branch prediction.

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 simulated hardware schedulers.

The size of a dispatch group depends on the availability of the simulated hardware resources. The processor dispatch width defaults to the value of the IssueWidth in LLVM's scheduling model.

An instruction can be dispatched if:

  • The size of the dispatch group is smaller than processor's dispatch width.
  • There are enough entries in the reorder buffer.
  • There are enough physical registers to do register renaming.
  • The schedulers are not full.

Scheduling models can optionally specify which register files are available on the processor. :program:`llvm-mca` uses that information to initialize register file descriptors. Users can limit the number of physical registers that are globally available for register renaming by using the command option -register-file-size. A value of zero for this option means unbounded. By knowing how many registers are available for renaming, the tool can predict dispatch stalls caused by the lack of physical registers.

The number of reorder buffer entries consumed by an instruction depends on the number of micro-opcodes specified for that instruction by the target scheduling model. The reorder buffer is responsible for tracking the progress of instructions that are "in-flight", and retiring them in program order. The number of entries in the reorder buffer defaults to the value specified by field MicroOpBufferSize in the target scheduling model.

Instructions that are dispatched to the schedulers consume scheduler buffer entries. :program:`llvm-mca` queries the scheduling model to determine the set of buffered resources consumed by an instruction. Buffered resources are treated like scheduler resources.

Instruction Issue

Each processor scheduler implements a buffer of instructions. An instruction has to wait in the scheduler's buffer 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) for execution. Instruction latencies are computed by :program:`llvm-mca` with the help of the scheduling model.

:program:`llvm-mca`'s scheduler is designed to simulate multiple processor schedulers. The scheduler is responsible for tracking data dependencies, and dynamically selecting which processor resources are consumed by instructions. It delegates the management of processor resource units and resource groups to a resource manager. The resource manager is responsible for selecting resource units that are consumed by instructions. For example, if an instruction consumes 1cy of a resource group, the resource manager selects one of the available units from the group; by default, the resource manager uses a round-robin selector to guarantee that resource usage is uniformly distributed between all units of a group.

:program:`llvm-mca`'s scheduler internally groups instructions into three sets:

  • WaitSet: a set of instructions whose operands are not ready.
  • ReadySet: a set of instructions ready to execute.
  • IssuedSet: a set of instructions executing.

Depending on the operands availability, instructions that are dispatched to the scheduler are either placed into the WaitSet or into the ReadySet.

Every cycle, the scheduler checks if instructions can be moved from the WaitSet to the ReadySet, and if instructions from the ReadySet can be issued to the underlying pipelines. The algorithm prioritizes older instructions over younger instructions.

Write-Back and Retire Stage

Issued instructions are moved from the ReadySet to the IssuedSet. 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.

When instructions are executed, the retire control unit flags the instruction as "ready to retire."

Instructions are retired in program order. The register file is notified of the retirement so that it can free the physical registers that were allocated for the instruction during the register renaming stage.

Load/Store Unit and Memory Consistency Model

To simulate an out-of-order execution of memory operations, :program:`llvm-mca` utilizes a simulated load/store unit (LSUnit) to simulate the speculative execution of loads and stores.

Each load (or store) consumes an entry in the load (or store) queue. Users can specify flags -lqueue and -squeue to limit the number of entries in the load and store queues respectively. The queues are unbounded by default.

The 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 are no intervening stores or barriers between the two loads.
  2. A younger load is allowed to pass an older store provided that the load does not alias with the store.
  3. A younger store is not allowed to pass an older store.
  4. A younger store is not allowed to pass an older load.

By default, the LSUnit optimistically assumes that loads do not alias (-noalias=true) store operations. Under this assumption, younger loads are always allowed to pass older stores. Essentially, the LSUnit does not attempt to run any alias analysis to predict when loads and stores do not alias with each other.

Note that, in the case of write-combining memory, rule 3 could be relaxed to allow reordering of non-aliasing store operations. That being said, at the moment, there is no way to further relax the memory model (-noalias is the only option). Essentially, there is no option to specify a different memory type (e.g., write-back, write-combining, write-through; etc.) and consequently to weaken, or strengthen, the memory model.

Other limitations are:

  • The LSUnit does not know when store-to-load forwarding may occur.
  • The LSUnit does not know anything about cache hierarchy and memory types.
  • The LSUnit does not know how to identify serializing operations and memory fences.

The LSUnit does not attempt to predict if 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).

:program:`llvm-mca` does not know about serializing operations or memory-barrier like instructions. The LSUnit conservatively assumes that an instruction which has both "MayLoad" and unmodeled side effects behaves like a "soft" load-barrier. That means, it serializes loads without forcing a flush of the load queue. Similarly, instructions that "MayStore" and have unmodeled side effects are treated like store barriers. A full memory barrier is a "MayLoad" and "MayStore" instruction with unmodeled side effects. 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 of the older loads/stores have been executed.

In conclusion, the full set of load/store consistency rules are:

  1. A store may not pass a previous store.
  2. A store may not pass a previous load (regardless of -noalias).
  3. A store has to wait until an older store barrier is fully executed.
  4. A load may pass a previous load.
  5. A load may not pass a previous store unless -noalias is set.
  6. A load has to wait until an older load barrier is fully executed.