1llvm-mca - LLVM Machine Code Analyzer
2=====================================
3
4.. program:: llvm-mca
5
6SYNOPSIS
7--------
8
9:program:`llvm-mca` [*options*] [input]
10
11DESCRIPTION
12-----------
13
14:program:`llvm-mca` is a performance analysis tool that uses information
15available in LLVM (e.g. scheduling models) to statically measure the performance
16of machine code in a specific CPU.
17
18Performance is measured in terms of throughput as well as processor resource
19consumption. The tool currently works for processors with an out-of-order
20backend, for which there is a scheduling model available in LLVM.
21
22The main goal of this tool is not just to predict the performance of the code
23when run on the target, but also help with diagnosing potential performance
24issues.
25
26Given an assembly code sequence, :program:`llvm-mca` estimates the Instructions
27Per Cycle (IPC), as well as hardware resource pressure. The analysis and
28reporting style were inspired by the IACA tool from Intel.
29
30For example, you can compile code with clang, output assembly, and pipe it
31directly into :program:`llvm-mca` for analysis:
32
33.. code-block:: bash
34
35  $ clang foo.c -O2 -target x86_64-unknown-unknown -S -o - | llvm-mca -mcpu=btver2
36
37Or for Intel syntax:
38
39.. code-block:: bash
40
41  $ clang foo.c -O2 -target x86_64-unknown-unknown -mllvm -x86-asm-syntax=intel -S -o - | llvm-mca -mcpu=btver2
42
43Scheduling models are not just used to compute instruction latencies and
44throughput, but also to understand what processor resources are available
45and how to simulate them.
46
47By design, the quality of the analysis conducted by :program:`llvm-mca` is
48inevitably affected by the quality of the scheduling models in LLVM.
49
50If you see that the performance report is not accurate for a processor,
51please `file a bug <https://bugs.llvm.org/enter_bug.cgi?product=libraries>`_
52against the appropriate backend.
53
54OPTIONS
55-------
56
57If ``input`` is "``-``" or omitted, :program:`llvm-mca` reads from standard
58input. Otherwise, it will read from the specified filename.
59
60If the :option:`-o` option is omitted, then :program:`llvm-mca` will send its output
61to standard output if the input is from standard input.  If the :option:`-o`
62option specifies "``-``", then the output will also be sent to standard output.
63
64
65.. option:: -help
66
67 Print a summary of command line options.
68
69.. option:: -o <filename>
70
71 Use ``<filename>`` as the output filename. See the summary above for more
72 details.
73
74.. option:: -mtriple=<target triple>
75
76 Specify a target triple string.
77
78.. option:: -march=<arch>
79
80 Specify the architecture for which to analyze the code. It defaults to the
81 host default target.
82
83.. option:: -mcpu=<cpuname>
84
85  Specify the processor for which to analyze the code.  By default, the cpu name
86  is autodetected from the host.
87
88.. option:: -output-asm-variant=<variant id>
89
90 Specify the output assembly variant for the report generated by the tool.
91 On x86, possible values are [0, 1]. A value of 0 (vic. 1) for this flag enables
92 the AT&T (vic. Intel) assembly format for the code printed out by the tool in
93 the analysis report.
94
95.. option:: -print-imm-hex
96
97 Prefer hex format for numeric literals in the output assembly printed as part
98 of the report.
99
100.. option:: -dispatch=<width>
101
102 Specify a different dispatch width for the processor. The dispatch width
103 defaults to field 'IssueWidth' in the processor scheduling model.  If width is
104 zero, then the default dispatch width is used.
105
106.. option:: -register-file-size=<size>
107
108 Specify the size of the register file. When specified, this flag limits how
109 many physical registers are available for register renaming purposes. A value
110 of zero for this flag means "unlimited number of physical registers".
111
112.. option:: -iterations=<number of iterations>
113
114 Specify the number of iterations to run. If this flag is set to 0, then the
115 tool sets the number of iterations to a default value (i.e. 100).
116
117.. option:: -noalias=<bool>
118
119  If set, the tool assumes that loads and stores don't alias. This is the
120  default behavior.
121
122.. option:: -lqueue=<load queue size>
123
124  Specify the size of the load queue in the load/store unit emulated by the tool.
125  By default, the tool assumes an unbound number of entries in the load queue.
126  A value of zero for this flag is ignored, and the default load queue size is
127  used instead.
128
129.. option:: -squeue=<store queue size>
130
131  Specify the size of the store queue in the load/store unit emulated by the
132  tool. By default, the tool assumes an unbound number of entries in the store
133  queue. A value of zero for this flag is ignored, and the default store queue
134  size is used instead.
135
136.. option:: -timeline
137
138  Enable the timeline view.
139
140.. option:: -timeline-max-iterations=<iterations>
141
142  Limit the number of iterations to print in the timeline view. By default, the
143  timeline view prints information for up to 10 iterations.
144
145.. option:: -timeline-max-cycles=<cycles>
146
147  Limit the number of cycles in the timeline view. By default, the number of
148  cycles is set to 80.
149
150.. option:: -resource-pressure
151
152  Enable the resource pressure view. This is enabled by default.
153
154.. option:: -register-file-stats
155
156  Enable register file usage statistics.
157
158.. option:: -dispatch-stats
159
160  Enable extra dispatch statistics. This view collects and analyzes instruction
161  dispatch events, as well as static/dynamic dispatch stall events. This view
162  is disabled by default.
163
164.. option:: -scheduler-stats
165
166  Enable extra scheduler statistics. This view collects and analyzes instruction
167  issue events. This view is disabled by default.
168
169.. option:: -retire-stats
170
171  Enable extra retire control unit statistics. This view is disabled by default.
172
173.. option:: -instruction-info
174
175  Enable the instruction info view. This is enabled by default.
176
177.. option:: -show-encoding
178
179  Enable the printing of instruction encodings within the instruction info view.
180
181.. option:: -all-stats
182
183  Print all hardware statistics. This enables extra statistics related to the
184  dispatch logic, the hardware schedulers, the register file(s), and the retire
185  control unit. This option is disabled by default.
186
187.. option:: -all-views
188
189  Enable all the view.
190
191.. option:: -instruction-tables
192
193  Prints resource pressure information based on the static information
194  available from the processor model. This differs from the resource pressure
195  view because it doesn't require that the code is simulated. It instead prints
196  the theoretical uniform distribution of resource pressure for every
197  instruction in sequence.
198
199.. option:: -bottleneck-analysis
200
201  Print information about bottlenecks that affect the throughput. This analysis
202  can be expensive, and it is disabled by default.  Bottlenecks are highlighted
203  in the summary view.
204
205
206EXIT STATUS
207-----------
208
209:program:`llvm-mca` returns 0 on success. Otherwise, an error message is printed
210to standard error, and the tool returns 1.
211
212USING MARKERS TO ANALYZE SPECIFIC CODE BLOCKS
213---------------------------------------------
214:program:`llvm-mca` allows for the optional usage of special code comments to
215mark regions of the assembly code to be analyzed.  A comment starting with
216substring ``LLVM-MCA-BEGIN`` marks the beginning of a code region. A comment
217starting with substring ``LLVM-MCA-END`` marks the end of a code region.  For
218example:
219
220.. code-block:: none
221
222  # LLVM-MCA-BEGIN
223    ...
224  # LLVM-MCA-END
225
226If no user-defined region is specified, then :program:`llvm-mca` assumes a
227default region which contains every instruction in the input file.  Every region
228is analyzed in isolation, and the final performance report is the union of all
229the reports generated for every code region.
230
231Code regions can have names. For example:
232
233.. code-block:: none
234
235  # LLVM-MCA-BEGIN A simple example
236    add %eax, %eax
237  # LLVM-MCA-END
238
239The code from the example above defines a region named "A simple example" with a
240single instruction in it. Note how the region name doesn't have to be repeated
241in the ``LLVM-MCA-END`` directive. In the absence of overlapping regions,
242an anonymous ``LLVM-MCA-END`` directive always ends the currently active user
243defined region.
244
245Example of nesting regions:
246
247.. code-block:: none
248
249  # LLVM-MCA-BEGIN foo
250    add %eax, %edx
251  # LLVM-MCA-BEGIN bar
252    sub %eax, %edx
253  # LLVM-MCA-END bar
254  # LLVM-MCA-END foo
255
256Example of overlapping regions:
257
258.. code-block:: none
259
260  # LLVM-MCA-BEGIN foo
261    add %eax, %edx
262  # LLVM-MCA-BEGIN bar
263    sub %eax, %edx
264  # LLVM-MCA-END foo
265    add %eax, %edx
266  # LLVM-MCA-END bar
267
268Note that multiple anonymous regions cannot overlap. Also, overlapping regions
269cannot have the same name.
270
271There is no support for marking regions from high-level source code, like C or
272C++. As a workaround, inline assembly directives may be used:
273
274.. code-block:: c++
275
276  int foo(int a, int b) {
277    __asm volatile("# LLVM-MCA-BEGIN foo");
278    a += 42;
279    __asm volatile("# LLVM-MCA-END");
280    a *= b;
281    return a;
282  }
283
284However, this interferes with optimizations like loop vectorization and may have
285an impact on the code generated. This is because the ``__asm`` statements are
286seen as real code having important side effects, which limits how the code
287around them can be transformed. If users want to make use of inline assembly
288to emit markers, then the recommendation is to always verify that the output
289assembly is equivalent to the assembly generated in the absence of markers.
290The `Clang options to emit optimization reports <https://clang.llvm.org/docs/UsersManual.html#options-to-emit-optimization-reports>`_
291can also help in detecting missed optimizations.
292
293HOW LLVM-MCA WORKS
294------------------
295
296:program:`llvm-mca` takes assembly code as input. The assembly code is parsed
297into a sequence of MCInst with the help of the existing LLVM target assembly
298parsers. The parsed sequence of MCInst is then analyzed by a ``Pipeline`` module
299to generate a performance report.
300
301The Pipeline module simulates the execution of the machine code sequence in a
302loop of iterations (default is 100). During this process, the pipeline collects
303a number of execution related statistics. At the end of this process, the
304pipeline generates and prints a report from the collected statistics.
305
306Here is an example of a performance report generated by the tool for a
307dot-product of two packed float vectors of four elements. The analysis is
308conducted for target x86, cpu btver2.  The following result can be produced via
309the following command using the example located at
310``test/tools/llvm-mca/X86/BtVer2/dot-product.s``:
311
312.. code-block:: bash
313
314  $ llvm-mca -mtriple=x86_64-unknown-unknown -mcpu=btver2 -iterations=300 dot-product.s
315
316.. code-block:: none
317
318  Iterations:        300
319  Instructions:      900
320  Total Cycles:      610
321  Total uOps:        900
322
323  Dispatch Width:    2
324  uOps Per Cycle:    1.48
325  IPC:               1.48
326  Block RThroughput: 2.0
327
328
329  Instruction Info:
330  [1]: #uOps
331  [2]: Latency
332  [3]: RThroughput
333  [4]: MayLoad
334  [5]: MayStore
335  [6]: HasSideEffects (U)
336
337  [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
338   1      2     1.00                        vmulps	%xmm0, %xmm1, %xmm2
339   1      3     1.00                        vhaddps	%xmm2, %xmm2, %xmm3
340   1      3     1.00                        vhaddps	%xmm3, %xmm3, %xmm4
341
342
343  Resources:
344  [0]   - JALU0
345  [1]   - JALU1
346  [2]   - JDiv
347  [3]   - JFPA
348  [4]   - JFPM
349  [5]   - JFPU0
350  [6]   - JFPU1
351  [7]   - JLAGU
352  [8]   - JMul
353  [9]   - JSAGU
354  [10]  - JSTC
355  [11]  - JVALU0
356  [12]  - JVALU1
357  [13]  - JVIMUL
358
359
360  Resource pressure per iteration:
361  [0]    [1]    [2]    [3]    [4]    [5]    [6]    [7]    [8]    [9]    [10]   [11]   [12]   [13]
362   -      -      -     2.00   1.00   2.00   1.00    -      -      -      -      -      -      -
363
364  Resource pressure by instruction:
365  [0]    [1]    [2]    [3]    [4]    [5]    [6]    [7]    [8]    [9]    [10]   [11]   [12]   [13]   Instructions:
366   -      -      -      -     1.00    -     1.00    -      -      -      -      -      -      -     vmulps	%xmm0, %xmm1, %xmm2
367   -      -      -     1.00    -     1.00    -      -      -      -      -      -      -      -     vhaddps	%xmm2, %xmm2, %xmm3
368   -      -      -     1.00    -     1.00    -      -      -      -      -      -      -      -     vhaddps	%xmm3, %xmm3, %xmm4
369
370According to this report, the dot-product kernel has been executed 300 times,
371for a total of 900 simulated instructions. The total number of simulated micro
372opcodes (uOps) is also 900.
373
374The report is structured in three main sections.  The first section collects a
375few performance numbers; the goal of this section is to give a very quick
376overview of the performance throughput. Important performance indicators are
377**IPC**, **uOps Per Cycle**, and  **Block RThroughput** (Block Reciprocal
378Throughput).
379
380Field *DispatchWidth* is the maximum number of micro opcodes that are dispatched
381to the out-of-order backend every simulated cycle.
382
383IPC is computed dividing the total number of simulated instructions by the total
384number of cycles.
385
386Field *Block RThroughput* is the reciprocal of the block throughput. Block
387throuhgput is a theoretical quantity computed as the maximum number of blocks
388(i.e. iterations) that can be executed per simulated clock cycle in the absence
389of loop carried dependencies. Block throughput is is superiorly
390limited by the dispatch rate, and the availability of hardware resources.
391
392In the absence of loop-carried data dependencies, the observed IPC tends to a
393theoretical maximum which can be computed by dividing the number of instructions
394of a single iteration by the `Block RThroughput`.
395
396Field 'uOps Per Cycle' is computed dividing the total number of simulated micro
397opcodes by the total number of cycles. A delta between Dispatch Width and this
398field is an indicator of a performance issue. In the absence of loop-carried
399data dependencies, the observed 'uOps Per Cycle' should tend to a theoretical
400maximum throughput which can be computed by dividing the number of uOps of a
401single iteration by the `Block RThroughput`.
402
403Field *uOps Per Cycle* is bounded from above by the dispatch width. That is
404because the dispatch width limits the maximum size of a dispatch group. Both IPC
405and 'uOps Per Cycle' are limited by the amount of hardware parallelism. The
406availability of hardware resources affects the resource pressure distribution,
407and it limits the number of instructions that can be executed in parallel every
408cycle.  A delta between Dispatch Width and the theoretical maximum uOps per
409Cycle (computed by dividing the number of uOps of a single iteration by the
410`Block RThroughput`) is an indicator of a performance bottleneck caused by the
411lack of hardware resources.
412In general, the lower the Block RThroughput, the better.
413
414In this example, ``uOps per iteration/Block RThroughput`` is 1.50. Since there
415are no loop-carried dependencies, the observed `uOps Per Cycle` is expected to
416approach 1.50 when the number of iterations tends to infinity. The delta between
417the Dispatch Width (2.00), and the theoretical maximum uOp throughput (1.50) is
418an indicator of a performance bottleneck caused by the lack of hardware
419resources, and the *Resource pressure view* can help to identify the problematic
420resource usage.
421
422The second section of the report is the `instruction info view`. It shows the
423latency and reciprocal throughput of every instruction in the sequence. It also
424reports extra information related to the number of micro opcodes, and opcode
425properties (i.e., 'MayLoad', 'MayStore', and 'HasSideEffects').
426
427Field *RThroughput* is the reciprocal of the instruction throughput. Throughput
428is computed as the maximum number of instructions of a same type that can be
429executed per clock cycle in the absence of operand dependencies. In this
430example, the reciprocal throughput of a vector float multiply is 1
431cycles/instruction.  That is because the FP multiplier JFPM is only available
432from pipeline JFPU1.
433
434Instruction encodings are displayed within the instruction info view when flag
435`-show-encoding` is specified.
436
437Below is an example of `-show-encoding` output for the dot-product kernel:
438
439.. code-block:: none
440
441  Instruction Info:
442  [1]: #uOps
443  [2]: Latency
444  [3]: RThroughput
445  [4]: MayLoad
446  [5]: MayStore
447  [6]: HasSideEffects (U)
448  [7]: Encoding Size
449
450  [1]    [2]    [3]    [4]    [5]    [6]    [7]    Encodings:                    Instructions:
451   1      2     1.00                         4     c5 f0 59 d0                   vmulps	%xmm0, %xmm1, %xmm2
452   1      4     1.00                         4     c5 eb 7c da                   vhaddps	%xmm2, %xmm2, %xmm3
453   1      4     1.00                         4     c5 e3 7c e3                   vhaddps	%xmm3, %xmm3, %xmm4
454
455The `Encoding Size` column shows the size in bytes of instructions.  The
456`Encodings` column shows the actual instruction encodings (byte sequences in
457hex).
458
459The third section is the *Resource pressure view*.  This view reports
460the average number of resource cycles consumed every iteration by instructions
461for every processor resource unit available on the target.  Information is
462structured in two tables. The first table reports the number of resource cycles
463spent on average every iteration. The second table correlates the resource
464cycles to the machine instruction in the sequence. For example, every iteration
465of the instruction vmulps always executes on resource unit [6]
466(JFPU1 - floating point pipeline #1), consuming an average of 1 resource cycle
467per iteration.  Note that on AMD Jaguar, vector floating-point multiply can
468only be issued to pipeline JFPU1, while horizontal floating-point additions can
469only be issued to pipeline JFPU0.
470
471The resource pressure view helps with identifying bottlenecks caused by high
472usage of specific hardware resources.  Situations with resource pressure mainly
473concentrated on a few resources should, in general, be avoided.  Ideally,
474pressure should be uniformly distributed between multiple resources.
475
476Timeline View
477^^^^^^^^^^^^^
478The timeline view produces a detailed report of each instruction's state
479transitions through an instruction pipeline.  This view is enabled by the
480command line option ``-timeline``.  As instructions transition through the
481various stages of the pipeline, their states are depicted in the view report.
482These states are represented by the following characters:
483
484* D : Instruction dispatched.
485* e : Instruction executing.
486* E : Instruction executed.
487* R : Instruction retired.
488* = : Instruction already dispatched, waiting to be executed.
489* \- : Instruction executed, waiting to be retired.
490
491Below is the timeline view for a subset of the dot-product example located in
492``test/tools/llvm-mca/X86/BtVer2/dot-product.s`` and processed by
493:program:`llvm-mca` using the following command:
494
495.. code-block:: bash
496
497  $ llvm-mca -mtriple=x86_64-unknown-unknown -mcpu=btver2 -iterations=3 -timeline dot-product.s
498
499.. code-block:: none
500
501  Timeline view:
502                      012345
503  Index     0123456789
504
505  [0,0]     DeeER.    .    .   vmulps	%xmm0, %xmm1, %xmm2
506  [0,1]     D==eeeER  .    .   vhaddps	%xmm2, %xmm2, %xmm3
507  [0,2]     .D====eeeER    .   vhaddps	%xmm3, %xmm3, %xmm4
508  [1,0]     .DeeE-----R    .   vmulps	%xmm0, %xmm1, %xmm2
509  [1,1]     . D=eeeE---R   .   vhaddps	%xmm2, %xmm2, %xmm3
510  [1,2]     . D====eeeER   .   vhaddps	%xmm3, %xmm3, %xmm4
511  [2,0]     .  DeeE-----R  .   vmulps	%xmm0, %xmm1, %xmm2
512  [2,1]     .  D====eeeER  .   vhaddps	%xmm2, %xmm2, %xmm3
513  [2,2]     .   D======eeeER   vhaddps	%xmm3, %xmm3, %xmm4
514
515
516  Average Wait times (based on the timeline view):
517  [0]: Executions
518  [1]: Average time spent waiting in a scheduler's queue
519  [2]: Average time spent waiting in a scheduler's queue while ready
520  [3]: Average time elapsed from WB until retire stage
521
522        [0]    [1]    [2]    [3]
523  0.     3     1.0    1.0    3.3       vmulps	%xmm0, %xmm1, %xmm2
524  1.     3     3.3    0.7    1.0       vhaddps	%xmm2, %xmm2, %xmm3
525  2.     3     5.7    0.0    0.0       vhaddps	%xmm3, %xmm3, %xmm4
526
527The timeline view is interesting because it shows instruction state changes
528during execution.  It also gives an idea of how the tool processes instructions
529executed on the target, and how their timing information might be calculated.
530
531The timeline view is structured in two tables.  The first table shows
532instructions changing state over time (measured in cycles); the second table
533(named *Average Wait times*) reports useful timing statistics, which should
534help diagnose performance bottlenecks caused by long data dependencies and
535sub-optimal usage of hardware resources.
536
537An instruction in the timeline view is identified by a pair of indices, where
538the first index identifies an iteration, and the second index is the
539instruction index (i.e., where it appears in the code sequence).  Since this
540example was generated using 3 iterations: ``-iterations=3``, the iteration
541indices range from 0-2 inclusively.
542
543Excluding the first and last column, the remaining columns are in cycles.
544Cycles are numbered sequentially starting from 0.
545
546From the example output above, we know the following:
547
548* Instruction [1,0] was dispatched at cycle 1.
549* Instruction [1,0] started executing at cycle 2.
550* Instruction [1,0] reached the write back stage at cycle 4.
551* Instruction [1,0] was retired at cycle 10.
552
553Instruction [1,0] (i.e., vmulps from iteration #1) does not have to wait in the
554scheduler's queue for the operands to become available. By the time vmulps is
555dispatched, operands are already available, and pipeline JFPU1 is ready to
556serve another instruction.  So the instruction can be immediately issued on the
557JFPU1 pipeline. That is demonstrated by the fact that the instruction only
558spent 1cy in the scheduler's queue.
559
560There is a gap of 5 cycles between the write-back stage and the retire event.
561That is because instructions must retire in program order, so [1,0] has to wait
562for [0,2] to be retired first (i.e., it has to wait until cycle 10).
563
564In the example, all instructions are in a RAW (Read After Write) dependency
565chain.  Register %xmm2 written by vmulps is immediately used by the first
566vhaddps, and register %xmm3 written by the first vhaddps is used by the second
567vhaddps.  Long data dependencies negatively impact the ILP (Instruction Level
568Parallelism).
569
570In the dot-product example, there are anti-dependencies introduced by
571instructions from different iterations.  However, those dependencies can be
572removed at register renaming stage (at the cost of allocating register aliases,
573and therefore consuming physical registers).
574
575Table *Average Wait times* helps diagnose performance issues that are caused by
576the presence of long latency instructions and potentially long data dependencies
577which may limit the ILP.  Note that :program:`llvm-mca`, by default, assumes at
578least 1cy between the dispatch event and the issue event.
579
580When the performance is limited by data dependencies and/or long latency
581instructions, the number of cycles spent while in the *ready* state is expected
582to be very small when compared with the total number of cycles spent in the
583scheduler's queue.  The difference between the two counters is a good indicator
584of how large of an impact data dependencies had on the execution of the
585instructions.  When performance is mostly limited by the lack of hardware
586resources, the delta between the two counters is small.  However, the number of
587cycles spent in the queue tends to be larger (i.e., more than 1-3cy),
588especially when compared to other low latency instructions.
589
590Bottleneck Analysis
591^^^^^^^^^^^^^^^^^^^
592The ``-bottleneck-analysis`` command line option enables the analysis of
593performance bottlenecks.
594
595This analysis is potentially expensive. It attempts to correlate increases in
596backend pressure (caused by pipeline resource pressure and data dependencies) to
597dynamic dispatch stalls.
598
599Below is an example of ``-bottleneck-analysis`` output generated by
600:program:`llvm-mca` for 500 iterations of the dot-product example on btver2.
601
602.. code-block:: none
603
604
605  Cycles with backend pressure increase [ 48.07% ]
606  Throughput Bottlenecks:
607    Resource Pressure       [ 47.77% ]
608    - JFPA  [ 47.77% ]
609    - JFPU0  [ 47.77% ]
610    Data Dependencies:      [ 0.30% ]
611    - Register Dependencies [ 0.30% ]
612    - Memory Dependencies   [ 0.00% ]
613
614  Critical sequence based on the simulation:
615
616                Instruction                         Dependency Information
617   +----< 2.    vhaddps %xmm3, %xmm3, %xmm4
618   |
619   |    < loop carried >
620   |
621   |      0.    vmulps  %xmm0, %xmm1, %xmm2
622   +----> 1.    vhaddps %xmm2, %xmm2, %xmm3         ## RESOURCE interference:  JFPA [ probability: 74% ]
623   +----> 2.    vhaddps %xmm3, %xmm3, %xmm4         ## REGISTER dependency:  %xmm3
624   |
625   |    < loop carried >
626   |
627   +----> 1.    vhaddps %xmm2, %xmm2, %xmm3         ## RESOURCE interference:  JFPA [ probability: 74% ]
628
629
630According to the analysis, throughput is limited by resource pressure and not by
631data dependencies.  The analysis observed increases in backend pressure during
63248.07% of the simulated run. Almost all those pressure increase events were
633caused by contention on processor resources JFPA/JFPU0.
634
635The `critical sequence` is the most expensive sequence of instructions according
636to the simulation. It is annotated to provide extra information about critical
637register dependencies and resource interferences between instructions.
638
639Instructions from the critical sequence are expected to significantly impact
640performance. By construction, the accuracy of this analysis is strongly
641dependent on the simulation and (as always) by the quality of the processor
642model in llvm.
643
644
645Extra Statistics to Further Diagnose Performance Issues
646^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
647The ``-all-stats`` command line option enables extra statistics and performance
648counters for the dispatch logic, the reorder buffer, the retire control unit,
649and the register file.
650
651Below is an example of ``-all-stats`` output generated by  :program:`llvm-mca`
652for 300 iterations of the dot-product example discussed in the previous
653sections.
654
655.. code-block:: none
656
657  Dynamic Dispatch Stall Cycles:
658  RAT     - Register unavailable:                      0
659  RCU     - Retire tokens unavailable:                 0
660  SCHEDQ  - Scheduler full:                            272  (44.6%)
661  LQ      - Load queue full:                           0
662  SQ      - Store queue full:                          0
663  GROUP   - Static restrictions on the dispatch group: 0
664
665
666  Dispatch Logic - number of cycles where we saw N micro opcodes dispatched:
667  [# dispatched], [# cycles]
668   0,              24  (3.9%)
669   1,              272  (44.6%)
670   2,              314  (51.5%)
671
672
673  Schedulers - number of cycles where we saw N micro opcodes issued:
674  [# issued], [# cycles]
675   0,          7  (1.1%)
676   1,          306  (50.2%)
677   2,          297  (48.7%)
678
679  Scheduler's queue usage:
680  [1] Resource name.
681  [2] Average number of used buffer entries.
682  [3] Maximum number of used buffer entries.
683  [4] Total number of buffer entries.
684
685   [1]            [2]        [3]        [4]
686  JALU01           0          0          20
687  JFPU01           17         18         18
688  JLSAGU           0          0          12
689
690
691  Retire Control Unit - number of cycles where we saw N instructions retired:
692  [# retired], [# cycles]
693   0,           109  (17.9%)
694   1,           102  (16.7%)
695   2,           399  (65.4%)
696
697  Total ROB Entries:                64
698  Max Used ROB Entries:             35  ( 54.7% )
699  Average Used ROB Entries per cy:  32  ( 50.0% )
700
701
702  Register File statistics:
703  Total number of mappings created:    900
704  Max number of mappings used:         35
705
706  *  Register File #1 -- JFpuPRF:
707     Number of physical registers:     72
708     Total number of mappings created: 900
709     Max number of mappings used:      35
710
711  *  Register File #2 -- JIntegerPRF:
712     Number of physical registers:     64
713     Total number of mappings created: 0
714     Max number of mappings used:      0
715
716If we look at the *Dynamic Dispatch Stall Cycles* table, we see the counter for
717SCHEDQ reports 272 cycles.  This counter is incremented every time the dispatch
718logic is unable to dispatch a full group because the scheduler's queue is full.
719
720Looking at the *Dispatch Logic* table, we see that the pipeline was only able to
721dispatch two micro opcodes 51.5% of the time.  The dispatch group was limited to
722one micro opcode 44.6% of the cycles, which corresponds to 272 cycles.  The
723dispatch statistics are displayed by either using the command option
724``-all-stats`` or ``-dispatch-stats``.
725
726The next table, *Schedulers*, presents a histogram displaying a count,
727representing the number of micro opcodes issued on some number of cycles. In
728this case, of the 610 simulated cycles, single opcodes were issued 306 times
729(50.2%) and there were 7 cycles where no opcodes were issued.
730
731The *Scheduler's queue usage* table shows that the average and maximum number of
732buffer entries (i.e., scheduler queue entries) used at runtime.  Resource JFPU01
733reached its maximum (18 of 18 queue entries). Note that AMD Jaguar implements
734three schedulers:
735
736* JALU01 - A scheduler for ALU instructions.
737* JFPU01 - A scheduler floating point operations.
738* JLSAGU - A scheduler for address generation.
739
740The dot-product is a kernel of three floating point instructions (a vector
741multiply followed by two horizontal adds).  That explains why only the floating
742point scheduler appears to be used.
743
744A full scheduler queue is either caused by data dependency chains or by a
745sub-optimal usage of hardware resources.  Sometimes, resource pressure can be
746mitigated by rewriting the kernel using different instructions that consume
747different scheduler resources.  Schedulers with a small queue are less resilient
748to bottlenecks caused by the presence of long data dependencies.  The scheduler
749statistics are displayed by using the command option ``-all-stats`` or
750``-scheduler-stats``.
751
752The next table, *Retire Control Unit*, presents a histogram displaying a count,
753representing the number of instructions retired on some number of cycles.  In
754this case, of the 610 simulated cycles, two instructions were retired during the
755same cycle 399 times (65.4%) and there were 109 cycles where no instructions
756were retired.  The retire statistics are displayed by using the command option
757``-all-stats`` or ``-retire-stats``.
758
759The last table presented is *Register File statistics*.  Each physical register
760file (PRF) used by the pipeline is presented in this table.  In the case of AMD
761Jaguar, there are two register files, one for floating-point registers (JFpuPRF)
762and one for integer registers (JIntegerPRF).  The table shows that of the 900
763instructions processed, there were 900 mappings created.  Since this dot-product
764example utilized only floating point registers, the JFPuPRF was responsible for
765creating the 900 mappings.  However, we see that the pipeline only used a
766maximum of 35 of 72 available register slots at any given time. We can conclude
767that the floating point PRF was the only register file used for the example, and
768that it was never resource constrained.  The register file statistics are
769displayed by using the command option ``-all-stats`` or
770``-register-file-stats``.
771
772In this example, we can conclude that the IPC is mostly limited by data
773dependencies, and not by resource pressure.
774
775Instruction Flow
776^^^^^^^^^^^^^^^^
777This section describes the instruction flow through the default pipeline of
778:program:`llvm-mca`, as well as the functional units involved in the process.
779
780The default pipeline implements the following sequence of stages used to
781process instructions.
782
783* Dispatch (Instruction is dispatched to the schedulers).
784* Issue (Instruction is issued to the processor pipelines).
785* Write Back (Instruction is executed, and results are written back).
786* Retire (Instruction is retired; writes are architecturally committed).
787
788The default pipeline only models the out-of-order portion of a processor.
789Therefore, the instruction fetch and decode stages are not modeled. Performance
790bottlenecks in the frontend are not diagnosed. :program:`llvm-mca` assumes that
791instructions have all been decoded and placed into a queue before the simulation
792start.  Also, :program:`llvm-mca` does not model branch prediction.
793
794Instruction Dispatch
795""""""""""""""""""""
796During the dispatch stage, instructions are picked in program order from a
797queue of already decoded instructions, and dispatched in groups to the
798simulated hardware schedulers.
799
800The size of a dispatch group depends on the availability of the simulated
801hardware resources.  The processor dispatch width defaults to the value
802of the ``IssueWidth`` in LLVM's scheduling model.
803
804An instruction can be dispatched if:
805
806* The size of the dispatch group is smaller than processor's dispatch width.
807* There are enough entries in the reorder buffer.
808* There are enough physical registers to do register renaming.
809* The schedulers are not full.
810
811Scheduling models can optionally specify which register files are available on
812the processor. :program:`llvm-mca` uses that information to initialize register
813file descriptors.  Users can limit the number of physical registers that are
814globally available for register renaming by using the command option
815``-register-file-size``.  A value of zero for this option means *unbounded*. By
816knowing how many registers are available for renaming, the tool can predict
817dispatch stalls caused by the lack of physical registers.
818
819The number of reorder buffer entries consumed by an instruction depends on the
820number of micro-opcodes specified for that instruction by the target scheduling
821model.  The reorder buffer is responsible for tracking the progress of
822instructions that are "in-flight", and retiring them in program order.  The
823number of entries in the reorder buffer defaults to the value specified by field
824`MicroOpBufferSize` in the target scheduling model.
825
826Instructions that are dispatched to the schedulers consume scheduler buffer
827entries. :program:`llvm-mca` queries the scheduling model to determine the set
828of buffered resources consumed by an instruction.  Buffered resources are
829treated like scheduler resources.
830
831Instruction Issue
832"""""""""""""""""
833Each processor scheduler implements a buffer of instructions.  An instruction
834has to wait in the scheduler's buffer until input register operands become
835available.  Only at that point, does the instruction becomes eligible for
836execution and may be issued (potentially out-of-order) for execution.
837Instruction latencies are computed by :program:`llvm-mca` with the help of the
838scheduling model.
839
840:program:`llvm-mca`'s scheduler is designed to simulate multiple processor
841schedulers.  The scheduler is responsible for tracking data dependencies, and
842dynamically selecting which processor resources are consumed by instructions.
843It delegates the management of processor resource units and resource groups to a
844resource manager.  The resource manager is responsible for selecting resource
845units that are consumed by instructions.  For example, if an instruction
846consumes 1cy of a resource group, the resource manager selects one of the
847available units from the group; by default, the resource manager uses a
848round-robin selector to guarantee that resource usage is uniformly distributed
849between all units of a group.
850
851:program:`llvm-mca`'s scheduler internally groups instructions into three sets:
852
853* WaitSet: a set of instructions whose operands are not ready.
854* ReadySet: a set of instructions ready to execute.
855* IssuedSet: a set of instructions executing.
856
857Depending on the operands availability, instructions that are dispatched to the
858scheduler are either placed into the WaitSet or into the ReadySet.
859
860Every cycle, the scheduler checks if instructions can be moved from the WaitSet
861to the ReadySet, and if instructions from the ReadySet can be issued to the
862underlying pipelines. The algorithm prioritizes older instructions over younger
863instructions.
864
865Write-Back and Retire Stage
866"""""""""""""""""""""""""""
867Issued instructions are moved from the ReadySet to the IssuedSet.  There,
868instructions wait until they reach the write-back stage.  At that point, they
869get removed from the queue and the retire control unit is notified.
870
871When instructions are executed, the retire control unit flags the instruction as
872"ready to retire."
873
874Instructions are retired in program order.  The register file is notified of the
875retirement so that it can free the physical registers that were allocated for
876the instruction during the register renaming stage.
877
878Load/Store Unit and Memory Consistency Model
879""""""""""""""""""""""""""""""""""""""""""""
880To simulate an out-of-order execution of memory operations, :program:`llvm-mca`
881utilizes a simulated load/store unit (LSUnit) to simulate the speculative
882execution of loads and stores.
883
884Each load (or store) consumes an entry in the load (or store) queue. Users can
885specify flags ``-lqueue`` and ``-squeue`` to limit the number of entries in the
886load and store queues respectively. The queues are unbounded by default.
887
888The LSUnit implements a relaxed consistency model for memory loads and stores.
889The rules are:
890
8911. A younger load is allowed to pass an older load only if there are no
892   intervening stores or barriers between the two loads.
8932. A younger load is allowed to pass an older store provided that the load does
894   not alias with the store.
8953. A younger store is not allowed to pass an older store.
8964. A younger store is not allowed to pass an older load.
897
898By default, the LSUnit optimistically assumes that loads do not alias
899(`-noalias=true`) store operations.  Under this assumption, younger loads are
900always allowed to pass older stores.  Essentially, the LSUnit does not attempt
901to run any alias analysis to predict when loads and stores do not alias with
902each other.
903
904Note that, in the case of write-combining memory, rule 3 could be relaxed to
905allow reordering of non-aliasing store operations.  That being said, at the
906moment, there is no way to further relax the memory model (``-noalias`` is the
907only option).  Essentially, there is no option to specify a different memory
908type (e.g., write-back, write-combining, write-through; etc.) and consequently
909to weaken, or strengthen, the memory model.
910
911Other limitations are:
912
913* The LSUnit does not know when store-to-load forwarding may occur.
914* The LSUnit does not know anything about cache hierarchy and memory types.
915* The LSUnit does not know how to identify serializing operations and memory
916  fences.
917
918The LSUnit does not attempt to predict if a load or store hits or misses the L1
919cache.  It only knows if an instruction "MayLoad" and/or "MayStore."  For
920loads, the scheduling model provides an "optimistic" load-to-use latency (which
921usually matches the load-to-use latency for when there is a hit in the L1D).
922
923:program:`llvm-mca` does not know about serializing operations or memory-barrier
924like instructions.  The LSUnit conservatively assumes that an instruction which
925has both "MayLoad" and unmodeled side effects behaves like a "soft"
926load-barrier.  That means, it serializes loads without forcing a flush of the
927load queue.  Similarly, instructions that "MayStore" and have unmodeled side
928effects are treated like store barriers.  A full memory barrier is a "MayLoad"
929and "MayStore" instruction with unmodeled side effects.  This is inaccurate, but
930it is the best that we can do at the moment with the current information
931available in LLVM.
932
933A load/store barrier consumes one entry of the load/store queue.  A load/store
934barrier enforces ordering of loads/stores.  A younger load cannot pass a load
935barrier.  Also, a younger store cannot pass a store barrier.  A younger load
936has to wait for the memory/load barrier to execute.  A load/store barrier is
937"executed" when it becomes the oldest entry in the load/store queue(s). That
938also means, by construction, all of the older loads/stores have been executed.
939
940In conclusion, the full set of load/store consistency rules are:
941
942#. A store may not pass a previous store.
943#. A store may not pass a previous load (regardless of ``-noalias``).
944#. A store has to wait until an older store barrier is fully executed.
945#. A load may pass a previous load.
946#. A load may not pass a previous store unless ``-noalias`` is set.
947#. A load has to wait until an older load barrier is fully executed.
948