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