xref: /netbsd-src/external/apache2/llvm/dist/llvm/docs/Passes.rst (revision 82d56013d7b633d116a93943de88e08335357a7c)
1..
2    If Passes.html is up to date, the following "one-liner" should print
3    an empty diff.
4
5    egrep -e '^<tr><td><a href="#.*">-.*</a></td><td>.*</td></tr>$' \
6          -e '^  <a name=".*">.*</a>$' < Passes.html >html; \
7    perl >help <<'EOT' && diff -u help html; rm -f help html
8    open HTML, "<Passes.html" or die "open: Passes.html: $!\n";
9    while (<HTML>) {
10      m:^<tr><td><a href="#(.*)">-.*</a></td><td>.*</td></tr>$: or next;
11      $order{$1} = sprintf("%03d", 1 + int %order);
12    }
13    open HELP, "../Release/bin/opt -help|" or die "open: opt -help: $!\n";
14    while (<HELP>) {
15      m:^    -([^ ]+) +- (.*)$: or next;
16      my $o = $order{$1};
17      $o = "000" unless defined $o;
18      push @x, "$o<tr><td><a href=\"#$1\">-$1</a></td><td>$2</td></tr>\n";
19      push @y, "$o  <a name=\"$1\">-$1: $2</a>\n";
20    }
21    @x = map { s/^\d\d\d//; $_ } sort @x;
22    @y = map { s/^\d\d\d//; $_ } sort @y;
23    print @x, @y;
24    EOT
25
26    This (real) one-liner can also be helpful when converting comments to HTML:
27
28    perl -e '$/ = undef; for (split(/\n/, <>)) { s:^ *///? ?::; print "  <p>\n" if !$on && $_ =~ /\S/; print "  </p>\n" if $on && $_ =~ /^\s*$/; print "  $_\n"; $on = ($_ =~ /\S/); } print "  </p>\n" if $on'
29
30====================================
31LLVM's Analysis and Transform Passes
32====================================
33
34.. contents::
35    :local:
36
37Introduction
38============
39
40This document serves as a high level summary of the optimization features that
41LLVM provides.  Optimizations are implemented as Passes that traverse some
42portion of a program to either collect information or transform the program.
43The table below divides the passes that LLVM provides into three categories.
44Analysis passes compute information that other passes can use or for debugging
45or program visualization purposes.  Transform passes can use (or invalidate)
46the analysis passes.  Transform passes all mutate the program in some way.
47Utility passes provides some utility but don't otherwise fit categorization.
48For example passes to extract functions to bitcode or write a module to bitcode
49are neither analysis nor transform passes.  The table of contents above
50provides a quick summary of each pass and links to the more complete pass
51description later in the document.
52
53Analysis Passes
54===============
55
56This section describes the LLVM Analysis Passes.
57
58``-aa-eval``: Exhaustive Alias Analysis Precision Evaluator
59-----------------------------------------------------------
60
61This is a simple N^2 alias analysis accuracy evaluator.  Basically, for each
62function in the program, it simply queries to see how the alias analysis
63implementation answers alias queries between each pair of pointers in the
64function.
65
66This is inspired and adapted from code by: Naveen Neelakantam, Francesco
67Spadini, and Wojciech Stryjewski.
68
69``-basic-aa``: Basic Alias Analysis (stateless AA impl)
70-------------------------------------------------------
71
72A basic alias analysis pass that implements identities (two different globals
73cannot alias, etc), but does no stateful analysis.
74
75``-basiccg``: Basic CallGraph Construction
76------------------------------------------
77
78Yet to be written.
79
80``-count-aa``: Count Alias Analysis Query Responses
81---------------------------------------------------
82
83A pass which can be used to count how many alias queries are being made and how
84the alias analysis implementation being used responds.
85
86.. _passes-da:
87
88``-da``: Dependence Analysis
89----------------------------
90
91Dependence analysis framework, which is used to detect dependences in memory
92accesses.
93
94``-debug-aa``: AA use debugger
95------------------------------
96
97This simple pass checks alias analysis users to ensure that if they create a
98new value, they do not query AA without informing it of the value.  It acts as
99a shim over any other AA pass you want.
100
101Yes keeping track of every value in the program is expensive, but this is a
102debugging pass.
103
104``-domfrontier``: Dominance Frontier Construction
105-------------------------------------------------
106
107This pass is a simple dominator construction algorithm for finding forward
108dominator frontiers.
109
110``-domtree``: Dominator Tree Construction
111-----------------------------------------
112
113This pass is a simple dominator construction algorithm for finding forward
114dominators.
115
116
117``-dot-callgraph``: Print Call Graph to "dot" file
118--------------------------------------------------
119
120This pass, only available in ``opt``, prints the call graph into a ``.dot``
121graph.  This graph can then be processed with the "dot" tool to convert it to
122postscript or some other suitable format.
123
124``-dot-cfg``: Print CFG of function to "dot" file
125-------------------------------------------------
126
127This pass, only available in ``opt``, prints the control flow graph into a
128``.dot`` graph.  This graph can then be processed with the :program:`dot` tool
129to convert it to postscript or some other suitable format.
130
131``-dot-cfg-only``: Print CFG of function to "dot" file (with no function bodies)
132--------------------------------------------------------------------------------
133
134This pass, only available in ``opt``, prints the control flow graph into a
135``.dot`` graph, omitting the function bodies.  This graph can then be processed
136with the :program:`dot` tool to convert it to postscript or some other suitable
137format.
138
139``-dot-dom``: Print dominance tree of function to "dot" file
140------------------------------------------------------------
141
142This pass, only available in ``opt``, prints the dominator tree into a ``.dot``
143graph.  This graph can then be processed with the :program:`dot` tool to
144convert it to postscript or some other suitable format.
145
146``-dot-dom-only``: Print dominance tree of function to "dot" file (with no function bodies)
147-------------------------------------------------------------------------------------------
148
149This pass, only available in ``opt``, prints the dominator tree into a ``.dot``
150graph, omitting the function bodies.  This graph can then be processed with the
151:program:`dot` tool to convert it to postscript or some other suitable format.
152
153``-dot-postdom``: Print postdominance tree of function to "dot" file
154--------------------------------------------------------------------
155
156This pass, only available in ``opt``, prints the post dominator tree into a
157``.dot`` graph.  This graph can then be processed with the :program:`dot` tool
158to convert it to postscript or some other suitable format.
159
160``-dot-postdom-only``: Print postdominance tree of function to "dot" file (with no function bodies)
161---------------------------------------------------------------------------------------------------
162
163This pass, only available in ``opt``, prints the post dominator tree into a
164``.dot`` graph, omitting the function bodies.  This graph can then be processed
165with the :program:`dot` tool to convert it to postscript or some other suitable
166format.
167
168``-globalsmodref-aa``: Simple mod/ref analysis for globals
169----------------------------------------------------------
170
171This simple pass provides alias and mod/ref information for global values that
172do not have their address taken, and keeps track of whether functions read or
173write memory (are "pure").  For this simple (but very common) case, we can
174provide pretty accurate and useful information.
175
176``-instcount``: Counts the various types of ``Instruction``\ s
177--------------------------------------------------------------
178
179This pass collects the count of all instructions and reports them.
180
181``-intervals``: Interval Partition Construction
182-----------------------------------------------
183
184This analysis calculates and represents the interval partition of a function,
185or a preexisting interval partition.
186
187In this way, the interval partition may be used to reduce a flow graph down to
188its degenerate single node interval partition (unless it is irreducible).
189
190``-iv-users``: Induction Variable Users
191---------------------------------------
192
193Bookkeeping for "interesting" users of expressions computed from induction
194variables.
195
196``-lazy-value-info``: Lazy Value Information Analysis
197-----------------------------------------------------
198
199Interface for lazy computation of value constraint information.
200
201``-libcall-aa``: LibCall Alias Analysis
202---------------------------------------
203
204LibCall Alias Analysis.
205
206``-lint``: Statically lint-checks LLVM IR
207-----------------------------------------
208
209This pass statically checks for common and easily-identified constructs which
210produce undefined or likely unintended behavior in LLVM IR.
211
212It is not a guarantee of correctness, in two ways.  First, it isn't
213comprehensive.  There are checks which could be done statically which are not
214yet implemented.  Some of these are indicated by TODO comments, but those
215aren't comprehensive either.  Second, many conditions cannot be checked
216statically.  This pass does no dynamic instrumentation, so it can't check for
217all possible problems.
218
219Another limitation is that it assumes all code will be executed.  A store
220through a null pointer in a basic block which is never reached is harmless, but
221this pass will warn about it anyway.
222
223Optimization passes may make conditions that this pass checks for more or less
224obvious.  If an optimization pass appears to be introducing a warning, it may
225be that the optimization pass is merely exposing an existing condition in the
226code.
227
228This code may be run before :ref:`instcombine <passes-instcombine>`.  In many
229cases, instcombine checks for the same kinds of things and turns instructions
230with undefined behavior into unreachable (or equivalent).  Because of this,
231this pass makes some effort to look through bitcasts and so on.
232
233``-loops``: Natural Loop Information
234------------------------------------
235
236This analysis is used to identify natural loops and determine the loop depth of
237various nodes of the CFG.  Note that the loops identified may actually be
238several natural loops that share the same header node... not just a single
239natural loop.
240
241``-memdep``: Memory Dependence Analysis
242---------------------------------------
243
244An analysis that determines, for a given memory operation, what preceding
245memory operations it depends on.  It builds on alias analysis information, and
246tries to provide a lazy, caching interface to a common kind of alias
247information query.
248
249``-module-debuginfo``: Decodes module-level debug info
250------------------------------------------------------
251
252This pass decodes the debug info metadata in a module and prints in a
253(sufficiently-prepared-) human-readable form.
254
255For example, run this pass from ``opt`` along with the ``-analyze`` option, and
256it'll print to standard output.
257
258``-postdomfrontier``: Post-Dominance Frontier Construction
259----------------------------------------------------------
260
261This pass is a simple post-dominator construction algorithm for finding
262post-dominator frontiers.
263
264``-postdomtree``: Post-Dominator Tree Construction
265--------------------------------------------------
266
267This pass is a simple post-dominator construction algorithm for finding
268post-dominators.
269
270``-print-alias-sets``: Alias Set Printer
271----------------------------------------
272
273Yet to be written.
274
275``-print-callgraph``: Print a call graph
276----------------------------------------
277
278This pass, only available in ``opt``, prints the call graph to standard error
279in a human-readable form.
280
281``-print-callgraph-sccs``: Print SCCs of the Call Graph
282-------------------------------------------------------
283
284This pass, only available in ``opt``, prints the SCCs of the call graph to
285standard error in a human-readable form.
286
287``-print-cfg-sccs``: Print SCCs of each function CFG
288----------------------------------------------------
289
290This pass, only available in ``opt``, printsthe SCCs of each function CFG to
291standard error in a human-readable fom.
292
293``-print-dom-info``: Dominator Info Printer
294-------------------------------------------
295
296Dominator Info Printer.
297
298``-print-externalfnconstants``: Print external fn callsites passed constants
299----------------------------------------------------------------------------
300
301This pass, only available in ``opt``, prints out call sites to external
302functions that are called with constant arguments.  This can be useful when
303looking for standard library functions we should constant fold or handle in
304alias analyses.
305
306``-print-function``: Print function to stderr
307---------------------------------------------
308
309The ``PrintFunctionPass`` class is designed to be pipelined with other
310``FunctionPasses``, and prints out the functions of the module as they are
311processed.
312
313``-print-module``: Print module to stderr
314-----------------------------------------
315
316This pass simply prints out the entire module when it is executed.
317
318.. _passes-print-used-types:
319
320``-print-used-types``: Find Used Types
321--------------------------------------
322
323This pass is used to seek out all of the types in use by the program.  Note
324that this analysis explicitly does not include types only used by the symbol
325table.
326
327``-regions``: Detect single entry single exit regions
328-----------------------------------------------------
329
330The ``RegionInfo`` pass detects single entry single exit regions in a function,
331where a region is defined as any subgraph that is connected to the remaining
332graph at only two spots.  Furthermore, a hierarchical region tree is built.
333
334.. _passes-scalar-evolution:
335
336``-scalar-evolution``: Scalar Evolution Analysis
337------------------------------------------------
338
339The ``ScalarEvolution`` analysis can be used to analyze and categorize scalar
340expressions in loops.  It specializes in recognizing general induction
341variables, representing them with the abstract and opaque ``SCEV`` class.
342Given this analysis, trip counts of loops and other important properties can be
343obtained.
344
345This analysis is primarily useful for induction variable substitution and
346strength reduction.
347
348``-scev-aa``: ScalarEvolution-based Alias Analysis
349--------------------------------------------------
350
351Simple alias analysis implemented in terms of ``ScalarEvolution`` queries.
352
353This differs from traditional loop dependence analysis in that it tests for
354dependencies within a single iteration of a loop, rather than dependencies
355between different iterations.
356
357``ScalarEvolution`` has a more complete understanding of pointer arithmetic
358than ``BasicAliasAnalysis``' collection of ad-hoc analyses.
359
360``-stack-safety``: Stack Safety Analysis
361------------------------------------------------
362
363The ``StackSafety`` analysis can be used to determine if stack allocated
364variables can be considered safe from memory access bugs.
365
366This analysis' primary purpose is to be used by sanitizers to avoid unnecessary
367instrumentation of safe variables.
368
369``-targetdata``: Target Data Layout
370-----------------------------------
371
372Provides other passes access to information on how the size and alignment
373required by the target ABI for various data types.
374
375Transform Passes
376================
377
378This section describes the LLVM Transform Passes.
379
380``-adce``: Aggressive Dead Code Elimination
381-------------------------------------------
382
383ADCE aggressively tries to eliminate code.  This pass is similar to :ref:`DCE
384<passes-dce>` but it assumes that values are dead until proven otherwise.  This
385is similar to :ref:`SCCP <passes-sccp>`, except applied to the liveness of
386values.
387
388``-always-inline``: Inliner for ``always_inline`` functions
389-----------------------------------------------------------
390
391A custom inliner that handles only functions that are marked as "always
392inline".
393
394``-argpromotion``: Promote 'by reference' arguments to scalars
395--------------------------------------------------------------
396
397This pass promotes "by reference" arguments to be "by value" arguments.  In
398practice, this means looking for internal functions that have pointer
399arguments.  If it can prove, through the use of alias analysis, that an
400argument is *only* loaded, then it can pass the value into the function instead
401of the address of the value.  This can cause recursive simplification of code
402and lead to the elimination of allocas (especially in C++ template code like
403the STL).
404
405This pass also handles aggregate arguments that are passed into a function,
406scalarizing them if the elements of the aggregate are only loaded.  Note that
407it refuses to scalarize aggregates which would require passing in more than
408three operands to the function, because passing thousands of operands for a
409large array or structure is unprofitable!
410
411Note that this transformation could also be done for arguments that are only
412stored to (returning the value instead), but does not currently.  This case
413would be best handled when and if LLVM starts supporting multiple return values
414from functions.
415
416``-bb-vectorize``: Basic-Block Vectorization
417--------------------------------------------
418
419This pass combines instructions inside basic blocks to form vector
420instructions.  It iterates over each basic block, attempting to pair compatible
421instructions, repeating this process until no additional pairs are selected for
422vectorization.  When the outputs of some pair of compatible instructions are
423used as inputs by some other pair of compatible instructions, those pairs are
424part of a potential vectorization chain.  Instruction pairs are only fused into
425vector instructions when they are part of a chain longer than some threshold
426length.  Moreover, the pass attempts to find the best possible chain for each
427pair of compatible instructions.  These heuristics are intended to prevent
428vectorization in cases where it would not yield a performance increase of the
429resulting code.
430
431``-block-placement``: Profile Guided Basic Block Placement
432----------------------------------------------------------
433
434This pass is a very simple profile guided basic block placement algorithm.  The
435idea is to put frequently executed blocks together at the start of the function
436and hopefully increase the number of fall-through conditional branches.  If
437there is no profile information for a particular function, this pass basically
438orders blocks in depth-first order.
439
440``-break-crit-edges``: Break critical edges in CFG
441--------------------------------------------------
442
443Break all of the critical edges in the CFG by inserting a dummy basic block.
444It may be "required" by passes that cannot deal with critical edges.  This
445transformation obviously invalidates the CFG, but can update forward dominator
446(set, immediate dominators, tree, and frontier) information.
447
448``-codegenprepare``: Optimize for code generation
449-------------------------------------------------
450
451This pass munges the code in the input function to better prepare it for
452SelectionDAG-based code generation.  This works around limitations in its
453basic-block-at-a-time approach.  It should eventually be removed.
454
455``-constmerge``: Merge Duplicate Global Constants
456-------------------------------------------------
457
458Merges duplicate global constants together into a single constant that is
459shared.  This is useful because some passes (i.e., TraceValues) insert a lot of
460string constants into the program, regardless of whether or not an existing
461string is available.
462
463.. _passes-dce:
464
465``-dce``: Dead Code Elimination
466-------------------------------
467
468Dead code elimination is similar to :ref:`dead instruction elimination
469<passes-die>`, but it rechecks instructions that were used by removed
470instructions to see if they are newly dead.
471
472``-deadargelim``: Dead Argument Elimination
473-------------------------------------------
474
475This pass deletes dead arguments from internal functions.  Dead argument
476elimination removes arguments which are directly dead, as well as arguments
477only passed into function calls as dead arguments of other functions.  This
478pass also deletes dead arguments in a similar way.
479
480This pass is often useful as a cleanup pass to run after aggressive
481interprocedural passes, which add possibly-dead arguments.
482
483``-deadtypeelim``: Dead Type Elimination
484----------------------------------------
485
486This pass is used to cleanup the output of GCC.  It eliminate names for types
487that are unused in the entire translation unit, using the :ref:`find used types
488<passes-print-used-types>` pass.
489
490.. _passes-die:
491
492``-die``: Dead Instruction Elimination
493--------------------------------------
494
495Dead instruction elimination performs a single pass over the function, removing
496instructions that are obviously dead.
497
498``-dse``: Dead Store Elimination
499--------------------------------
500
501A trivial dead store elimination that only considers basic-block local
502redundant stores.
503
504.. _passes-function-attrs:
505
506``-function-attrs``: Deduce function attributes
507-----------------------------------------------
508
509A simple interprocedural pass which walks the call-graph, looking for functions
510which do not access or only read non-local memory, and marking them
511``readnone``/``readonly``.  In addition, it marks function arguments (of
512pointer type) "``nocapture``" if a call to the function does not create any
513copies of the pointer value that outlive the call.  This more or less means
514that the pointer is only dereferenced, and not returned from the function or
515stored in a global.  This pass is implemented as a bottom-up traversal of the
516call-graph.
517
518``-globaldce``: Dead Global Elimination
519---------------------------------------
520
521This transform is designed to eliminate unreachable internal globals from the
522program.  It uses an aggressive algorithm, searching out globals that are known
523to be alive.  After it finds all of the globals which are needed, it deletes
524whatever is left over.  This allows it to delete recursive chunks of the
525program which are unreachable.
526
527``-globalopt``: Global Variable Optimizer
528-----------------------------------------
529
530This pass transforms simple global variables that never have their address
531taken.  If obviously true, it marks read/write globals as constant, deletes
532variables only stored to, etc.
533
534``-gvn``: Global Value Numbering
535--------------------------------
536
537This pass performs global value numbering to eliminate fully and partially
538redundant instructions.  It also performs redundant load elimination.
539
540.. _passes-indvars:
541
542``-indvars``: Canonicalize Induction Variables
543----------------------------------------------
544
545This transformation analyzes and transforms the induction variables (and
546computations derived from them) into simpler forms suitable for subsequent
547analysis and transformation.
548
549This transformation makes the following changes to each loop with an
550identifiable induction variable:
551
552* All loops are transformed to have a *single* canonical induction variable
553  which starts at zero and steps by one.
554* The canonical induction variable is guaranteed to be the first PHI node in
555  the loop header block.
556* Any pointer arithmetic recurrences are raised to use array subscripts.
557
558If the trip count of a loop is computable, this pass also makes the following
559changes:
560
561* The exit condition for the loop is canonicalized to compare the induction
562  value against the exit value.  This turns loops like:
563
564  .. code-block:: c++
565
566    for (i = 7; i*i < 1000; ++i)
567
568    into
569
570  .. code-block:: c++
571
572    for (i = 0; i != 25; ++i)
573
574* Any use outside of the loop of an expression derived from the indvar is
575  changed to compute the derived value outside of the loop, eliminating the
576  dependence on the exit value of the induction variable.  If the only purpose
577  of the loop is to compute the exit value of some derived expression, this
578  transformation will make the loop dead.
579
580This transformation should be followed by strength reduction after all of the
581desired loop transformations have been performed.  Additionally, on targets
582where it is profitable, the loop could be transformed to count down to zero
583(the "do loop" optimization).
584
585``-inline``: Function Integration/Inlining
586------------------------------------------
587
588Bottom-up inlining of functions into callees.
589
590.. _passes-instcombine:
591
592``-instcombine``: Combine redundant instructions
593------------------------------------------------
594
595Combine instructions to form fewer, simple instructions.  This pass does not
596modify the CFG. This pass is where algebraic simplification happens.
597
598This pass combines things like:
599
600.. code-block:: llvm
601
602  %Y = add i32 %X, 1
603  %Z = add i32 %Y, 1
604
605into:
606
607.. code-block:: llvm
608
609  %Z = add i32 %X, 2
610
611This is a simple worklist driven algorithm.
612
613This pass guarantees that the following canonicalizations are performed on the
614program:
615
616#. If a binary operator has a constant operand, it is moved to the right-hand
617   side.
618#. Bitwise operators with constant operands are always grouped so that shifts
619   are performed first, then ``or``\ s, then ``and``\ s, then ``xor``\ s.
620#. Compare instructions are converted from ``<``, ``>``, ``≤``, or ``≥`` to
621   ``=`` or ``≠`` if possible.
622#. All ``cmp`` instructions on boolean values are replaced with logical
623   operations.
624#. ``add X, X`` is represented as ``mul X, 2`` ⇒ ``shl X, 1``
625#. Multiplies with a constant power-of-two argument are transformed into
626   shifts.
627#. … etc.
628
629This pass can also simplify calls to specific well-known function calls (e.g.
630runtime library functions).  For example, a call ``exit(3)`` that occurs within
631the ``main()`` function can be transformed into simply ``return 3``. Whether or
632not library calls are simplified is controlled by the
633:ref:`-function-attrs <passes-function-attrs>` pass and LLVM's knowledge of
634library calls on different targets.
635
636.. _passes-aggressive-instcombine:
637
638``-aggressive-instcombine``: Combine expression patterns
639--------------------------------------------------------
640
641Combine expression patterns to form expressions with fewer, simple instructions.
642This pass does not modify the CFG.
643
644For example, this pass reduce width of expressions post-dominated by TruncInst
645into smaller width when applicable.
646
647It differs from instcombine pass in that it contains pattern optimization that
648requires higher complexity than the O(1), thus, it should run fewer times than
649instcombine pass.
650
651``-internalize``: Internalize Global Symbols
652--------------------------------------------
653
654This pass loops over all of the functions in the input module, looking for a
655main function.  If a main function is found, all other functions and all global
656variables with initializers are marked as internal.
657
658``-ipsccp``: Interprocedural Sparse Conditional Constant Propagation
659--------------------------------------------------------------------
660
661An interprocedural variant of :ref:`Sparse Conditional Constant Propagation
662<passes-sccp>`.
663
664``-jump-threading``: Jump Threading
665-----------------------------------
666
667Jump threading tries to find distinct threads of control flow running through a
668basic block.  This pass looks at blocks that have multiple predecessors and
669multiple successors.  If one or more of the predecessors of the block can be
670proven to always cause a jump to one of the successors, we forward the edge
671from the predecessor to the successor by duplicating the contents of this
672block.
673
674An example of when this can occur is code like this:
675
676.. code-block:: c++
677
678  if () { ...
679    X = 4;
680  }
681  if (X < 3) {
682
683In this case, the unconditional branch at the end of the first if can be
684revectored to the false side of the second if.
685
686.. _passes-lcssa:
687
688``-lcssa``: Loop-Closed SSA Form Pass
689-------------------------------------
690
691This pass transforms loops by placing phi nodes at the end of the loops for all
692values that are live across the loop boundary.  For example, it turns the left
693into the right code:
694
695.. code-block:: c++
696
697  for (...)                for (...)
698      if (c)                   if (c)
699          X1 = ...                 X1 = ...
700      else                     else
701          X2 = ...                 X2 = ...
702      X3 = phi(X1, X2)         X3 = phi(X1, X2)
703  ... = X3 + 4              X4 = phi(X3)
704                              ... = X4 + 4
705
706This is still valid LLVM; the extra phi nodes are purely redundant, and will be
707trivially eliminated by ``InstCombine``.  The major benefit of this
708transformation is that it makes many other loop optimizations, such as
709``LoopUnswitch``\ ing, simpler. You can read more in the
710:ref:`loop terminology section for the LCSSA form <loop-terminology-lcssa>`.
711
712.. _passes-licm:
713
714``-licm``: Loop Invariant Code Motion
715-------------------------------------
716
717This pass performs loop invariant code motion, attempting to remove as much
718code from the body of a loop as possible.  It does this by either hoisting code
719into the preheader block, or by sinking code to the exit blocks if it is safe.
720This pass also promotes must-aliased memory locations in the loop to live in
721registers, thus hoisting and sinking "invariant" loads and stores.
722
723Hoisting operations out of loops is a canonicalization transform. It enables
724and simplifies subsequent optimizations in the middle-end. Rematerialization
725of hoisted instructions to reduce register pressure is the responsibility of
726the back-end, which has more accurate information about register pressure and
727also handles other optimizations than LICM that increase live-ranges.
728
729This pass uses alias analysis for two purposes:
730
731#. Moving loop invariant loads and calls out of loops.  If we can determine
732   that a load or call inside of a loop never aliases anything stored to, we
733   can hoist it or sink it like any other instruction.
734
735#. Scalar Promotion of Memory.  If there is a store instruction inside of the
736   loop, we try to move the store to happen AFTER the loop instead of inside of
737   the loop.  This can only happen if a few conditions are true:
738
739   #. The pointer stored through is loop invariant.
740   #. There are no stores or loads in the loop which *may* alias the pointer.
741      There are no calls in the loop which mod/ref the pointer.
742
743   If these conditions are true, we can promote the loads and stores in the
744   loop of the pointer to use a temporary alloca'd variable.  We then use the
745   :ref:`mem2reg <passes-mem2reg>` functionality to construct the appropriate
746   SSA form for the variable.
747
748``-loop-deletion``: Delete dead loops
749-------------------------------------
750
751This file implements the Dead Loop Deletion Pass.  This pass is responsible for
752eliminating loops with non-infinite computable trip counts that have no side
753effects or volatile instructions, and do not contribute to the computation of
754the function's return value.
755
756.. _passes-loop-extract:
757
758``-loop-extract``: Extract loops into new functions
759---------------------------------------------------
760
761A pass wrapper around the ``ExtractLoop()`` scalar transformation to extract
762each top-level loop into its own new function.  If the loop is the *only* loop
763in a given function, it is not touched.  This is a pass most useful for
764debugging via bugpoint.
765
766``-loop-extract-single``: Extract at most one loop into a new function
767----------------------------------------------------------------------
768
769Similar to :ref:`Extract loops into new functions <passes-loop-extract>`, this
770pass extracts one natural loop from the program into a function if it can.
771This is used by :program:`bugpoint`.
772
773``-loop-reduce``: Loop Strength Reduction
774-----------------------------------------
775
776This pass performs a strength reduction on array references inside loops that
777have as one or more of their components the loop induction variable.  This is
778accomplished by creating a new value to hold the initial value of the array
779access for the first iteration, and then creating a new GEP instruction in the
780loop to increment the value by the appropriate amount.
781
782.. _passes-loop-rotate:
783
784``-loop-rotate``: Rotate Loops
785------------------------------
786
787A simple loop rotation transformation.  A summary of it can be found in
788:ref:`Loop Terminology for Rotated Loops <loop-terminology-loop-rotate>`.
789
790
791.. _passes-loop-simplify:
792
793``-loop-simplify``: Canonicalize natural loops
794----------------------------------------------
795
796This pass performs several transformations to transform natural loops into a
797simpler form, which makes subsequent analyses and transformations simpler and
798more effective. A summary of it can be found in
799:ref:`Loop Terminology, Loop Simplify Form <loop-terminology-loop-simplify>`.
800
801Loop pre-header insertion guarantees that there is a single, non-critical entry
802edge from outside of the loop to the loop header.  This simplifies a number of
803analyses and transformations, such as :ref:`LICM <passes-licm>`.
804
805Loop exit-block insertion guarantees that all exit blocks from the loop (blocks
806which are outside of the loop that have predecessors inside of the loop) only
807have predecessors from inside of the loop (and are thus dominated by the loop
808header).  This simplifies transformations such as store-sinking that are built
809into LICM.
810
811This pass also guarantees that loops will have exactly one backedge.
812
813Note that the :ref:`simplifycfg <passes-simplifycfg>` pass will clean up blocks
814which are split out but end up being unnecessary, so usage of this pass should
815not pessimize generated code.
816
817This pass obviously modifies the CFG, but updates loop information and
818dominator information.
819
820``-loop-unroll``: Unroll loops
821------------------------------
822
823This pass implements a simple loop unroller.  It works best when loops have
824been canonicalized by the :ref:`indvars <passes-indvars>` pass, allowing it to
825determine the trip counts of loops easily.
826
827``-loop-unroll-and-jam``: Unroll and Jam loops
828----------------------------------------------
829
830This pass implements a simple unroll and jam classical loop optimisation pass.
831It transforms loop from:
832
833.. code-block:: c++
834
835  for i.. i+= 1              for i.. i+= 4
836    for j..                    for j..
837      code(i, j)                 code(i, j)
838                                 code(i+1, j)
839                                 code(i+2, j)
840                                 code(i+3, j)
841                             remainder loop
842
843Which can be seen as unrolling the outer loop and "jamming" (fusing) the inner
844loops into one. When variables or loads can be shared in the new inner loop, this
845can lead to significant performance improvements. It uses
846:ref:`Dependence Analysis <passes-da>` for proving the transformations are safe.
847
848.. _passes-loop-unswitch:
849
850``-loop-unswitch``: Unswitch loops
851----------------------------------
852
853This pass transforms loops that contain branches on loop-invariant conditions
854to have multiple loops.  For example, it turns the left into the right code:
855
856.. code-block:: c++
857
858  for (...)                  if (lic)
859      A                          for (...)
860      if (lic)                       A; B; C
861          B                  else
862      C                          for (...)
863                                     A; C
864
865This can increase the size of the code exponentially (doubling it every time a
866loop is unswitched) so we only unswitch if the resultant code will be smaller
867than a threshold.
868
869This pass expects :ref:`LICM <passes-licm>` to be run before it to hoist
870invariant conditions out of the loop, to make the unswitching opportunity
871obvious.
872
873``-loweratomic``: Lower atomic intrinsics to non-atomic form
874------------------------------------------------------------
875
876This pass lowers atomic intrinsics to non-atomic form for use in a known
877non-preemptible environment.
878
879The pass does not verify that the environment is non-preemptible (in general
880this would require knowledge of the entire call graph of the program including
881any libraries which may not be available in bitcode form); it simply lowers
882every atomic intrinsic.
883
884``-lowerinvoke``: Lower invokes to calls, for unwindless code generators
885------------------------------------------------------------------------
886
887This transformation is designed for use by code generators which do not yet
888support stack unwinding.  This pass converts ``invoke`` instructions to
889``call`` instructions, so that any exception-handling ``landingpad`` blocks
890become dead code (which can be removed by running the ``-simplifycfg`` pass
891afterwards).
892
893``-lowerswitch``: Lower ``SwitchInst``\ s to branches
894-----------------------------------------------------
895
896Rewrites switch instructions with a sequence of branches, which allows targets
897to get away with not implementing the switch instruction until it is
898convenient.
899
900.. _passes-mem2reg:
901
902``-mem2reg``: Promote Memory to Register
903----------------------------------------
904
905This file promotes memory references to be register references.  It promotes
906alloca instructions which only have loads and stores as uses.  An ``alloca`` is
907transformed by using dominator frontiers to place phi nodes, then traversing
908the function in depth-first order to rewrite loads and stores as appropriate.
909This is just the standard SSA construction algorithm to construct "pruned" SSA
910form.
911
912``-memcpyopt``: MemCpy Optimization
913-----------------------------------
914
915This pass performs various transformations related to eliminating ``memcpy``
916calls, or transforming sets of stores into ``memset``\ s.
917
918``-mergefunc``: Merge Functions
919-------------------------------
920
921This pass looks for equivalent functions that are mergable and folds them.
922
923Total-ordering is introduced among the functions set: we define comparison
924that answers for every two functions which of them is greater. It allows to
925arrange functions into the binary tree.
926
927For every new function we check for equivalent in tree.
928
929If equivalent exists we fold such functions. If both functions are overridable,
930we move the functionality into a new internal function and leave two
931overridable thunks to it.
932
933If there is no equivalent, then we add this function to tree.
934
935Lookup routine has O(log(n)) complexity, while whole merging process has
936complexity of O(n*log(n)).
937
938Read
939:doc:`this <MergeFunctions>`
940article for more details.
941
942``-mergereturn``: Unify function exit nodes
943-------------------------------------------
944
945Ensure that functions have at most one ``ret`` instruction in them.
946Additionally, it keeps track of which node is the new exit node of the CFG.
947
948``-partial-inliner``: Partial Inliner
949-------------------------------------
950
951This pass performs partial inlining, typically by inlining an ``if`` statement
952that surrounds the body of the function.
953
954``-prune-eh``: Remove unused exception handling info
955----------------------------------------------------
956
957This file implements a simple interprocedural pass which walks the call-graph,
958turning invoke instructions into call instructions if and only if the callee
959cannot throw an exception.  It implements this as a bottom-up traversal of the
960call-graph.
961
962``-reassociate``: Reassociate expressions
963-----------------------------------------
964
965This pass reassociates commutative expressions in an order that is designed to
966promote better constant propagation, GCSE, :ref:`LICM <passes-licm>`, PRE, etc.
967
968For example: 4 + (x + 5) ⇒ x + (4 + 5)
969
970In the implementation of this algorithm, constants are assigned rank = 0,
971function arguments are rank = 1, and other values are assigned ranks
972corresponding to the reverse post order traversal of current function (starting
973at 2), which effectively gives values in deep loops higher rank than values not
974in loops.
975
976``-rel-lookup-table-converter``: Relative lookup table converter
977----------------------------------------------------------------
978
979This pass converts lookup tables to PIC-friendly relative lookup tables.
980
981``-reg2mem``: Demote all values to stack slots
982----------------------------------------------
983
984This file demotes all registers to memory references.  It is intended to be the
985inverse of :ref:`mem2reg <passes-mem2reg>`.  By converting to ``load``
986instructions, the only values live across basic blocks are ``alloca``
987instructions and ``load`` instructions before ``phi`` nodes.  It is intended
988that this should make CFG hacking much easier.  To make later hacking easier,
989the entry block is split into two, such that all introduced ``alloca``
990instructions (and nothing else) are in the entry block.
991
992``-sroa``: Scalar Replacement of Aggregates
993------------------------------------------------------
994
995The well-known scalar replacement of aggregates transformation.  This transform
996breaks up ``alloca`` instructions of aggregate type (structure or array) into
997individual ``alloca`` instructions for each member if possible.  Then, if
998possible, it transforms the individual ``alloca`` instructions into nice clean
999scalar SSA form.
1000
1001.. _passes-sccp:
1002
1003``-sccp``: Sparse Conditional Constant Propagation
1004--------------------------------------------------
1005
1006Sparse conditional constant propagation and merging, which can be summarized
1007as:
1008
1009* Assumes values are constant unless proven otherwise
1010* Assumes BasicBlocks are dead unless proven otherwise
1011* Proves values to be constant, and replaces them with constants
1012* Proves conditional branches to be unconditional
1013
1014Note that this pass has a habit of making definitions be dead.  It is a good
1015idea to run a :ref:`DCE <passes-dce>` pass sometime after running this pass.
1016
1017.. _passes-simplifycfg:
1018
1019``-simplifycfg``: Simplify the CFG
1020----------------------------------
1021
1022Performs dead code elimination and basic block merging.  Specifically:
1023
1024* Removes basic blocks with no predecessors.
1025* Merges a basic block into its predecessor if there is only one and the
1026  predecessor only has one successor.
1027* Eliminates PHI nodes for basic blocks with a single predecessor.
1028* Eliminates a basic block that only contains an unconditional branch.
1029
1030``-sink``: Code sinking
1031-----------------------
1032
1033This pass moves instructions into successor blocks, when possible, so that they
1034aren't executed on paths where their results aren't needed.
1035
1036``-strip``: Strip all symbols from a module
1037-------------------------------------------
1038
1039Performs code stripping.  This transformation can delete:
1040
1041* names for virtual registers
1042* symbols for internal globals and functions
1043* debug information
1044
1045Note that this transformation makes code much less readable, so it should only
1046be used in situations where the strip utility would be used, such as reducing
1047code size or making it harder to reverse engineer code.
1048
1049``-strip-dead-debug-info``: Strip debug info for unused symbols
1050---------------------------------------------------------------
1051
1052.. FIXME: this description is the same as for -strip
1053
1054performs code stripping. this transformation can delete:
1055
1056* names for virtual registers
1057* symbols for internal globals and functions
1058* debug information
1059
1060note that this transformation makes code much less readable, so it should only
1061be used in situations where the strip utility would be used, such as reducing
1062code size or making it harder to reverse engineer code.
1063
1064``-strip-dead-prototypes``: Strip Unused Function Prototypes
1065------------------------------------------------------------
1066
1067This pass loops over all of the functions in the input module, looking for dead
1068declarations and removes them.  Dead declarations are declarations of functions
1069for which no implementation is available (i.e., declarations for unused library
1070functions).
1071
1072``-strip-debug-declare``: Strip all ``llvm.dbg.declare`` intrinsics
1073-------------------------------------------------------------------
1074
1075.. FIXME: this description is the same as for -strip
1076
1077This pass implements code stripping.  Specifically, it can delete:
1078
1079#. names for virtual registers
1080#. symbols for internal globals and functions
1081#. debug information
1082
1083Note that this transformation makes code much less readable, so it should only
1084be used in situations where the 'strip' utility would be used, such as reducing
1085code size or making it harder to reverse engineer code.
1086
1087``-strip-nondebug``: Strip all symbols, except dbg symbols, from a module
1088-------------------------------------------------------------------------
1089
1090.. FIXME: this description is the same as for -strip
1091
1092This pass implements code stripping.  Specifically, it can delete:
1093
1094#. names for virtual registers
1095#. symbols for internal globals and functions
1096#. debug information
1097
1098Note that this transformation makes code much less readable, so it should only
1099be used in situations where the 'strip' utility would be used, such as reducing
1100code size or making it harder to reverse engineer code.
1101
1102``-tailcallelim``: Tail Call Elimination
1103----------------------------------------
1104
1105This file transforms calls of the current function (self recursion) followed by
1106a return instruction with a branch to the entry of the function, creating a
1107loop.  This pass also implements the following extensions to the basic
1108algorithm:
1109
1110#. Trivial instructions between the call and return do not prevent the
1111   transformation from taking place, though currently the analysis cannot
1112   support moving any really useful instructions (only dead ones).
1113#. This pass transforms functions that are prevented from being tail recursive
1114   by an associative expression to use an accumulator variable, thus compiling
1115   the typical naive factorial or fib implementation into efficient code.
1116#. TRE is performed if the function returns void, if the return returns the
1117   result returned by the call, or if the function returns a run-time constant
1118   on all exits from the function.  It is possible, though unlikely, that the
1119   return returns something else (like constant 0), and can still be TRE'd.  It
1120   can be TRE'd if *all other* return instructions in the function return the
1121   exact same value.
1122#. If it can prove that callees do not access their caller stack frame, they
1123   are marked as eligible for tail call elimination (by the code generator).
1124
1125Utility Passes
1126==============
1127
1128This section describes the LLVM Utility Passes.
1129
1130``-deadarghaX0r``: Dead Argument Hacking (BUGPOINT USE ONLY; DO NOT USE)
1131------------------------------------------------------------------------
1132
1133Same as dead argument elimination, but deletes arguments to functions which are
1134external.  This is only for use by :doc:`bugpoint <Bugpoint>`.
1135
1136``-extract-blocks``: Extract Basic Blocks From Module (for bugpoint use)
1137------------------------------------------------------------------------
1138
1139This pass is used by bugpoint to extract all blocks from the module into their
1140own functions.
1141
1142``-instnamer``: Assign names to anonymous instructions
1143------------------------------------------------------
1144
1145This is a little utility pass that gives instructions names, this is mostly
1146useful when diffing the effect of an optimization because deleting an unnamed
1147instruction can change all other instruction numbering, making the diff very
1148noisy.
1149
1150.. _passes-verify:
1151
1152``-verify``: Module Verifier
1153----------------------------
1154
1155Verifies an LLVM IR code.  This is useful to run after an optimization which is
1156undergoing testing.  Note that llvm-as verifies its input before emitting
1157bitcode, and also that malformed bitcode is likely to make LLVM crash.  All
1158language front-ends are therefore encouraged to verify their output before
1159performing optimizing transformations.
1160
1161#. Both of a binary operator's parameters are of the same type.
1162#. Verify that the indices of mem access instructions match other operands.
1163#. Verify that arithmetic and other things are only performed on first-class
1164   types.  Verify that shifts and logicals only happen on integrals f.e.
1165#. All of the constants in a switch statement are of the correct type.
1166#. The code is in valid SSA form.
1167#. It is illegal to put a label into any other type (like a structure) or to
1168   return one.
1169#. Only phi nodes can be self referential: ``%x = add i32 %x``, ``%x`` is
1170   invalid.
1171#. PHI nodes must have an entry for each predecessor, with no extras.
1172#. PHI nodes must be the first thing in a basic block, all grouped together.
1173#. PHI nodes must have at least one entry.
1174#. All basic blocks should only end with terminator insts, not contain them.
1175#. The entry node to a function must not have predecessors.
1176#. All Instructions must be embedded into a basic block.
1177#. Functions cannot take a void-typed parameter.
1178#. Verify that a function's argument list agrees with its declared type.
1179#. It is illegal to specify a name for a void value.
1180#. It is illegal to have an internal global value with no initializer.
1181#. It is illegal to have a ``ret`` instruction that returns a value that does
1182   not agree with the function return value type.
1183#. Function call argument types match the function prototype.
1184#. All other things that are tested by asserts spread about the code.
1185
1186Note that this does not provide full security verification (like Java), but
1187instead just tries to ensure that code is well-formed.
1188
1189.. _passes-view-cfg:
1190
1191``-view-cfg``: View CFG of function
1192-----------------------------------
1193
1194Displays the control flow graph using the GraphViz tool.
1195
1196``-view-cfg-only``: View CFG of function (with no function bodies)
1197------------------------------------------------------------------
1198
1199Displays the control flow graph using the GraphViz tool, but omitting function
1200bodies.
1201
1202``-view-dom``: View dominance tree of function
1203----------------------------------------------
1204
1205Displays the dominator tree using the GraphViz tool.
1206
1207``-view-dom-only``: View dominance tree of function (with no function bodies)
1208-----------------------------------------------------------------------------
1209
1210Displays the dominator tree using the GraphViz tool, but omitting function
1211bodies.
1212
1213``-view-postdom``: View postdominance tree of function
1214------------------------------------------------------
1215
1216Displays the post dominator tree using the GraphViz tool.
1217
1218``-view-postdom-only``: View postdominance tree of function (with no function bodies)
1219-------------------------------------------------------------------------------------
1220
1221Displays the post dominator tree using the GraphViz tool, but omitting function
1222bodies.
1223
1224``-transform-warning``: Report missed forced transformations
1225------------------------------------------------------------
1226
1227Emits warnings about not yet applied forced transformations (e.g. from
1228``#pragma omp simd``).
1229