brintos

brintos / llvm-project-archived public Read only

0
0
Text · 24.1 KiB · 6978a9b Raw
719 lines · plain
1.. _loop-terminology:2 3===========================================4LLVM Loop Terminology (and Canonical Forms)5===========================================6 7.. contents::8   :local:9 10Loop Definition11===============12 13Loops are an important concept for a code optimizer. In LLVM, detection14of loops in a control-flow graph is done by :ref:`loopinfo`. It is based15on the following definition.16 17A loop is a subset of nodes from the control-flow graph (CFG; where18nodes represent basic blocks) with the following properties:19 201. The induced subgraph (which is the subgraph that contains all the21   edges from the CFG within the loop) is strongly connected22   (every node is reachable from all others).23 242. All edges from outside the subset into the subset point to the same25   node, called the **header**. As a consequence, the header dominates26   all nodes in the loop (i.e. every execution path to any of the loop's27   node will have to pass through the header).28 293. The loop is the maximum subset with these properties. That is, no30   additional nodes from the CFG can be added such that the induced31   subgraph would still be strongly connected and the header would32   remain the same.33 34In computer science literature, this is often called a *natural loop*.35In LLVM, a more generalized definition is called a36:ref:`cycle <cycle-terminology>`.37 38 39Terminology40-----------41 42The definition of a loop comes with some additional terminology:43 44* An **entering block** (or **loop predecessor**) is a non-loop node45  that has an edge into the loop (necessarily the header). If there is46  only one entering block, and its only edge is to the47  header, it is also called the loop's **preheader**. The preheader48  dominates the loop without itself being part of the loop.49 50* A **latch** is a loop node that has an edge to the header.51 52* A **backedge** is an edge from a latch to the header.53 54* An **exiting edge** is an edge from inside the loop to a node outside55  of the loop. The source of such an edge is called an **exiting block**, its56  target is an **exit block**.57 58.. image:: ./loop-terminology.svg59   :width: 400 px60 61 62Important Notes63---------------64 65This loop definition has some noteworthy consequences:66 67* A node can be the header of at most one loop. As such, a loop can be68  identified by its header. Due to the header being the only entry into69  a loop, it can be called a Single-Entry-Multiple-Exits (SEME) region.70 71 72* For basic blocks that are not reachable from the function's entry, the73  concept of loops is undefined. This follows from the concept of74  dominance being undefined as well.75 76 77* The smallest loop consists of a single basic block that branches to78  itself. In this case that block is the header, latch (and exiting79  block if it has another edge to a different block) at the same time.80  A single block that has no branch to itself is not considered a loop,81  even though it is trivially strongly connected.82 83.. image:: ./loop-single.svg84   :width: 300 px85 86In this case, the role of header, exiting block and latch fall to the87same node. :ref:`loopinfo` reports this as:88 89.. code-block:: console90 91  $ opt input.ll -passes='print<loops>'92  Loop at depth 1 containing: %for.body<header><latch><exiting>93 94 95* Loops can be nested inside each other. That is, a loop's node set can96  be a subset of another loop with a different loop header. The loop97  hierarchy in a function forms a forest: Each top-level loop is the98  root of the tree of the loops nested inside it.99 100.. image:: ./loop-nested.svg101   :width: 350 px102 103 104* It is not possible that two loops share only a few of their nodes.105  Two loops are either disjoint or one is nested inside the other. In106  the example below the left and right subsets both violate the107  maximality condition. Only the merge of both sets is considered a loop.108 109.. image:: ./loop-nonmaximal.svg110   :width: 250 px111 112 113* It is also possible that two logical loops share a header, but are114  considered a single loop by LLVM:115 116.. code-block:: C117 118  for (int i = 0; i < 128; ++i)119    for (int j = 0; j < 128; ++j)120      body(i,j);121 122which might be represented in LLVM-IR as follows. Note that there is123only a single header and hence just a single loop.124 125.. image:: ./loop-merge.svg126   :width: 400 px127 128The :ref:`LoopSimplify <loop-terminology-loop-simplify>` pass will129detect the loop and ensure separate headers for the outer and inner loop.130 131.. image:: ./loop-separate.svg132   :width: 400 px133 134* A cycle in the CFG does not imply there is a loop. The example below135  shows such a CFG, where there is no header node that dominates all136  other nodes in the cycle. This is called **irreducible control-flow**.137 138.. image:: ./loop-irreducible.svg139   :width: 150 px140 141The term reducible results from the ability to collapse the CFG into a142single node by successively replacing one of three base structures with143a single node: A sequential execution of basic blocks, acyclic conditional144branches (or switches), and a basic block looping on itself.145`Wikipedia <https://en.wikipedia.org/wiki/Control-flow_graph#Reducibility>`_146has a more formal definition, which basically says that every cycle has147a dominating header.148 149 150* Irreducible control-flow can occur at any level of the loop nesting.151  That is, a loop that itself does not contain any loops can still have152  cyclic control flow in its body; a loop that is not nested inside153  another loop can still be part of an outer cycle; and there can be154  additional cycles between any two loops where one is contained in the other.155  However, an LLVM :ref:`cycle<cycle-terminology>` covers both, loops and156  irreducible control flow.157 158 159* The `FixIrreducible <https://llvm.org/doxygen/FixIrreducible_8h.html>`_160  pass can transform irreducible control flow into loops by inserting161  new loop headers. It is not included in any default optimization pass162  pipeline, but is required for some back-end targets.163 164 165* Exiting edges are not the only way to break out of a loop. Other166  possibilities are unreachable terminators, [[noreturn]] functions,167  exceptions, signals, and your computer's power button.168 169 170* A basic block "inside" the loop that does not have a path back to the171  loop (i.e. to a latch or header) is not considered part of the loop.172  This is illustrated by the following code.173 174.. code-block:: C175 176  for (unsigned i = 0; i <= n; ++i) {177    if (c1) {178      // When reaching this block, we will have exited the loop.179      do_something();180      break;181    }182    if (c2) {183      // abort(), never returns, so we have exited the loop.184      abort();185    }186    if (c3) {187      // The unreachable allows the compiler to assume that this will not rejoin the loop.188      do_something();189      __builtin_unreachable();190    }191    if (c4) {192      // This statically infinite loop is not nested because control-flow will not continue with the for-loop.193      while(true) {194        do_something();195      }196    }197  }198 199 200* There is no requirement for the control flow to eventually leave the201  loop, i.e. a loop can be infinite. A **statically infinite loop** is a202  loop that has no exiting edges. A **dynamically infinite loop** has203  exiting edges, but it is possible to be never taken. This may happen204  only under some circumstances, such as when n == UINT_MAX in the code205  below.206 207.. code-block:: C208 209  for (unsigned i = 0; i <= n; ++i)210    body(i);211 212It is possible for the optimizer to turn a dynamically infinite loop213into a statically infinite loop, for instance when it can prove that the214exiting condition is always false. Because the exiting edge is never215taken, the optimizer can change the conditional branch into an216unconditional one.217 218If a is loop is annotated with219:ref:`llvm.loop.mustprogress <langref_llvm_loop_mustprogress>` metadata,220the compiler is allowed to assume that it will eventually terminate, even221if it cannot prove it. For instance, it may remove a mustprogress-loop222that does not have any side-effect in its body even though the program223could be stuck in that loop forever. Languages such as C and224`C++ <https://eel.is/c++draft/intro.progress#1>`_ have such225forward-progress guarantees for some loops. Also see the226:ref:`mustprogress <langref_mustprogress>` and227:ref:`willreturn <langref_willreturn>` function attributes, as well as228the older :ref:`llvm.sideeffect <llvm_sideeffect>` intrinsic.229 230* The number of executions of the loop header before leaving the loop is231  the **loop trip count** (or **iteration count**). If the loop should232  not be executed at all, a **loop guard** must skip the entire loop:233 234.. image:: ./loop-guard.svg235   :width: 500 px236 237Since the first thing a loop header might do is to check whether there238is another execution and if not, immediately exit without doing any work239(also see :ref:`loop-terminology-loop-rotate`), loop trip count is not240the best measure of a loop's number of iterations. For instance, the241number of header executions of the code below for a non-positive n242(before loop rotation) is 1, even though the loop body is not executed243at all.244 245.. code-block:: C246 247  for (int i = 0; i < n; ++i)248    body(i);249 250A better measure is the **backedge-taken count**, which is the number of251times any of the backedges is taken before the loop. It is one less than252the trip count for executions that enter the header.253 254 255.. _loopinfo:256 257LoopInfo258========259 260LoopInfo is the core analysis for obtaining information about loops.261There are few key implications of the definitions given above which262are important for working successfully with this interface.263 264* LoopInfo does not contain information about non-loop cycles.  As a265  result, it is not suitable for any algorithm which requires complete266  cycle detection for correctness.267 268* LoopInfo provides an interface for enumerating all top level loops269  (e.g. those not contained in any other loop).  From there, you may270  walk the tree of sub-loops rooted in that top level loop.271 272* Loops which become statically unreachable during optimization *must*273  be removed from LoopInfo. If this can not be done for some reason,274  then the optimization is *required* to preserve the static275  reachability of the loop.276 277 278.. _loop-terminology-loop-simplify:279 280Loop Simplify Form281==================282 283The Loop Simplify Form is a canonical form that makes284several analyses and transformations simpler and more effective.285It is ensured by the LoopSimplify286(:ref:`-loop-simplify <passes-loop-simplify>`) pass and is automatically287added by the pass managers when scheduling a LoopPass.288This pass is implemented in289`LoopSimplify.h <https://llvm.org/doxygen/LoopSimplify_8h_source.html>`_.290When it is successful, the loop has:291 292* A preheader.293* A single backedge (which implies that there is a single latch).294* Dedicated exits. That is, no exit block for the loop295  has a predecessor that is outside the loop. This implies296  that all exit blocks are dominated by the loop header.297 298.. _loop-terminology-lcssa:299 300Loop Closed SSA (LCSSA)301=======================302 303A program is in Loop Closed SSA Form if it is in SSA form304and all values that are defined in a loop are used only inside305this loop.306 307Programs written in LLVM IR are always in SSA form but not necessarily308in LCSSA. To achieve the latter, for each value that is live across the309loop boundary, single entry PHI nodes are inserted to each of the exit blocks310[#lcssa-construction]_ in order to "close" these values inside the loop.311In particular, consider the following loop:312 313.. code-block:: C314 315    c = ...;316    for (...) {317      if (c)318        X1 = ...319      else320        X2 = ...321      X3 = phi(X1, X2);  // X3 defined322    }323 324    ... = X3 + 4;  // X3 used, i.e. live325                   // outside the loop326 327In the inner loop, the X3 is defined inside the loop, but used328outside of it. In Loop Closed SSA form, this would be represented as follows:329 330.. code-block:: C331 332    c = ...;333    for (...) {334      if (c)335        X1 = ...336      else337        X2 = ...338      X3 = phi(X1, X2);339    }340    X4 = phi(X3);341 342    ... = X4 + 4;343 344This is still valid LLVM; the extra phi nodes are purely redundant,345but all LoopPass'es are required to preserve them.346This form is ensured by the LCSSA (:ref:`-lcssa <passes-lcssa>`)347pass and is added automatically by the LoopPassManager when348scheduling a LoopPass.349After the loop optimizations are done, these extra phi nodes350will be deleted by :ref:`-instcombine <passes-instcombine>`.351 352Note that an exit block is outside of a loop, so how can such a phi "close"353the value inside the loop since it uses it outside of it ? First of all,354for phi nodes, as355`mentioned in the LangRef <https://llvm.org/docs/LangRef.html#phi-instruction>`_:356"the use of each incoming value is deemed to occur on the edge from the357corresponding predecessor block to the current block". Now, an358edge to an exit block is considered outside of the loop because359if we take that edge, it leads us clearly out of the loop.360 361However, an edge doesn't actually contain any IR, so in source code,362we have to choose a convention of whether the use happens in363the current block or in the respective predecessor. For LCSSA's purpose,364we consider the use happens in the latter (so as to consider the365use inside) [#point-of-use-phis]_.366 367The major benefit of LCSSA is that it makes many other loop optimizations368simpler.369 370First of all, a simple observation is that if one needs to see all371the outside users, they can just iterate over all the (loop closing)372PHI nodes in the exit blocks (the alternative would be to373scan the def-use chain [#def-use-chain]_ of all instructions in the loop).374 375Then, consider for example376:ref:`simple-loop-unswitch <passes-simple-loop-unswitch>` ing the loop above.377Because it is in LCSSA form, we know that any value defined inside of378the loop will be used either only inside the loop or in a loop closing379PHI node. In this case, the only loop closing PHI node is X4.380This means that we can just copy the loop and change the X4381accordingly, like so:382 383.. code-block:: C384 385    c = ...;386    if (c) {387      for (...) {388        if (true)389          X1 = ...390        else391          X2 = ...392        X3 = phi(X1, X2);393      }394    } else {395      for (...) {396        if (false)397          X1' = ...398        else399          X2' = ...400        X3' = phi(X1', X2');401      }402    }403    X4 = phi(X3, X3')404 405Now, all uses of X4 will get the updated value (in general,406if a loop is in LCSSA form, in any loop transformation,407we only need to update the loop closing PHI nodes for the changes408to take effect).  If we did not have Loop Closed SSA form, it means that X3 could409possibly be used outside the loop. So, we would have to introduce the410X4 (which is the new X3) and replace all uses of X3 with that.411However, we should note that because LLVM keeps a def-use chain412[#def-use-chain]_ for each Value, we wouldn't need413to perform data-flow analysis to find and replace all the uses414(there is even a utility function, replaceAllUsesWith(),415that performs this transformation by iterating the def-use chain).416 417Another important advantage is that the behavior of all uses418of an induction variable is the same.  Without this, you need to419distinguish the case when the variable is used outside of420the loop it is defined in, for example:421 422.. code-block:: C423 424  for (i = 0; i < 100; i++) {425    for (j = 0; j < 100; j++) {426      k = i + j;427      use(k);    // use 1428    }429    use(k);      // use 2430  }431 432Looking from the outer loop with the normal SSA form, the first use of k433is not well-behaved, while the second one is an induction variable with434base 100 and step 1.  Although, in practice, and in the LLVM context,435such cases can be handled effectively by SCEV. Scalar Evolution436(:ref:`scalar-evolution <passes-scalar-evolution>`) or SCEV, is a437(analysis) pass that analyzes and categorizes the evolution of scalar438expressions in loops.439 440In general, it's easier to use SCEV in loops that are in LCSSA form.441The evolution of a scalar (loop-variant) expression that442SCEV can analyze is, by definition, relative to a loop.443An expression is represented in LLVM by an444`llvm::Instruction <https://llvm.org/doxygen/classllvm_1_1Instruction.html>`_.445If the expression is inside two (or more) loops (which can only446happen if the loops are nested, like in the example above) and you want447to get an analysis of its evolution (from SCEV),448you have to also specify relative to what Loop you want it.449Specifically, you have to use450`getSCEVAtScope() <https://llvm.org/doxygen/classllvm_1_1ScalarEvolution.html#a21d6ee82eed29080d911dbb548a8bb68>`_.451 452However, if all loops are in LCSSA form, each expression is actually453represented by two different llvm::Instructions.  One inside the loop454and one outside, which is the loop-closing PHI node and represents455the value of the expression after the last iteration (effectively,456we break each loop-variant expression into two expressions and so, every457expression is at most in one loop).  You can now just use458`getSCEV() <https://llvm.org/doxygen/classllvm_1_1ScalarEvolution.html#a30bd18ac905eacf3601bc6a553a9ff49>`_.459and which of these two llvm::Instructions you pass to it disambiguates460the context / scope / relative loop.461 462.. rubric:: Footnotes463 464.. [#lcssa-construction] To insert these loop-closing PHI nodes, one has to465  (re-)compute dominance frontiers (if the loop has multiple exits).466 467.. [#point-of-use-phis] Considering the point of use of a PHI entry value468  to be in the respective predecessor is a convention across the whole LLVM.469  The reason is mostly practical; for example it preserves the dominance470  property of SSA. It is also just an overapproximation of the actual471  number of uses; the incoming block could branch to another block in which472  case the value is not actually used but there are no side-effects (it might473  increase its live range which is not relevant in LCSSA though).474  Furthermore, we can gain some intuition if we consider liveness:475  A PHI is *usually* inserted in the current block because the value can't476  be used from this point and onwards (i.e. the current block is a dominance477  frontier). It doesn't make sense to consider that the value is used in478  the current block (because of the PHI) since the value stops being live479  before the PHI. In some sense the PHI definition just "replaces" the original480  value definition and doesn't actually use it. It should be stressed that481  this analogy is only used as an example and does not pose any strict482  requirements. For example, the value might dominate the current block483  but we can still insert a PHI (as we do with LCSSA PHI nodes) *and*484  use the original value afterwards (in which case the two live ranges overlap,485  although in LCSSA (the whole point is that) we never do that).486 487 488.. [#def-use-chain] A property of SSA is that there exists a def-use chain489  for each definition, which is a list of all the uses of this definition.490  LLVM implements this property by keeping a list of all the uses of a Value491  in an internal data structure.492 493"More Canonical" Loops494======================495 496.. _loop-terminology-loop-rotate:497 498Rotated Loops499-------------500 501Loops are rotated by the LoopRotate (:ref:`loop-rotate <passes-loop-rotate>`)502pass, which converts loops into do/while style loops and is503implemented in504`LoopRotation.h <https://llvm.org/doxygen/LoopRotation_8h_source.html>`_.  Example:505 506.. code-block:: C507 508  void test(int n) {509    for (int i = 0; i < n; i += 1)510      // Loop body511  }512 513is transformed to:514 515.. code-block:: C516 517  void test(int n) {518    int i = 0;519    do {520      // Loop body521      i += 1;522    } while (i < n);523  }524 525**Warning**: This transformation is valid only if the compiler526can prove that the loop body will be executed at least once. Otherwise,527it has to insert a guard which will test it at runtime. In the example528above, that would be:529 530.. code-block:: C531 532  void test(int n) {533    int i = 0;534    if (n > 0) {535      do {536        // Loop body537        i += 1;538      } while (i < n);539    }540  }541 542It's important to understand the effect of loop rotation543at the LLVM IR level. We follow with the previous examples544in LLVM IR while also providing a graphical representation545of the control-flow graphs (CFG). You can get the same graphical546results by utilizing the :ref:`view-cfg <passes-view-cfg>` pass.547 548The initial **for** loop could be translated to:549 550.. code-block:: none551 552  define void @test(i32 %n) {553  entry:554    br label %for.header555 556  for.header:557    %i = phi i32 [ 0, %entry ], [ %i.next, %latch ]558    %cond = icmp slt i32 %i, %n559    br i1 %cond, label %body, label %exit560 561  body:562    ; Loop body563    br label %latch564 565  latch:566    %i.next = add nsw i32 %i, 1567    br label %for.header568 569  exit:570    ret void571  }572 573.. image:: ./loop-terminology-initial-loop.png574  :width: 400 px575 576Before we explain how LoopRotate will actually577transform this loop, here's how we could convert578it (by hand) to a do-while style loop.579 580.. code-block:: none581 582  define void @test(i32 %n) {583  entry:584    br label %body585 586  body:587    %i = phi i32 [ 0, %entry ], [ %i.next, %latch ]588    ; Loop body589    br label %latch590 591  latch:592    %i.next = add nsw i32 %i, 1593    %cond = icmp slt i32 %i.next, %n594    br i1 %cond, label %body, label %exit595 596  exit:597    ret void598  }599 600.. image:: ./loop-terminology-rotated-loop.png601  :width: 400 px602 603Note two things:604 605* The condition check was moved to the "bottom" of the loop, i.e.606  the latch. This is something that LoopRotate does by copying the header607  of the loop to the latch.608* The compiler in this case can't deduce that the loop will609  definitely execute at least once so the above transformation610  is not valid. As mentioned above, a guard has to be inserted,611  which is something that LoopRotate will do.612 613This is how LoopRotate transforms this loop:614 615.. code-block:: none616 617  define void @test(i32 %n) {618  entry:619    %guard_cond = icmp slt i32 0, %n620    br i1 %guard_cond, label %loop.preheader, label %exit621 622  loop.preheader:623    br label %body624 625  body:626    %i2 = phi i32 [ 0, %loop.preheader ], [ %i.next, %latch ]627    br label %latch628 629  latch:630    %i.next = add nsw i32 %i2, 1631    %cond = icmp slt i32 %i.next, %n632    br i1 %cond, label %body, label %loop.exit633 634  loop.exit:635    br label %exit636 637  exit:638    ret void639  }640 641.. image:: ./loop-terminology-guarded-loop.png642  :width: 500 px643 644The result is a little bit more complicated than we may expect645because LoopRotate ensures that the loop is in646:ref:`Loop Simplify Form <loop-terminology-loop-simplify>`647after rotation.648In this case, it inserted the %loop.preheader basic block so649that the loop has a preheader and it introduced the %loop.exit650basic block so that the loop has dedicated exits651(otherwise, %exit would be jumped from both %latch and %entry,652but %entry is not contained in the loop).653Note that a loop has to be in Loop Simplify Form beforehand654too for LoopRotate to be applied successfully.655 656The main advantage of this form is that it allows hoisting657invariant instructions, especially loads, into the preheader.658That could be done in non-rotated loops as well but with659some disadvantages.  Let's illustrate them with an example:660 661.. code-block:: C662 663  for (int i = 0; i < n; ++i) {664    auto v = *p;665    use(v);666  }667 668We assume that loading from p is invariant and use(v) is some669statement that uses v.670If we wanted to execute the load only once we could move it671"out" of the loop body, resulting in this:672 673.. code-block:: C674 675  auto v = *p;676  for (int i = 0; i < n; ++i) {677    use(v);678  }679 680However, now, in the case that n <= 0, in the initial form,681the loop body would never execute, and so, the load would682never execute.  This is a problem mainly for semantic reasons.683Consider the case in which n <= 0 and loading from p is invalid.684In the initial program there would be no error.  However, with this685transformation we would introduce one, effectively breaking686the initial semantics.687 688To avoid both of these problems, we can insert a guard:689 690.. code-block:: C691 692  if (n > 0) {  // loop guard693    auto v = *p;694    for (int i = 0; i < n; ++i) {695      use(v);696    }697  }698 699This is certainly better but it could be improved slightly. Notice700that the check for whether n is bigger than 0 is executed twice (and701n does not change in between).  Once when we check the guard condition702and once in the first execution of the loop.  To avoid that, we could703do an unconditional first execution and insert the loop condition704in the end. This effectively means transforming the loop into a do-while loop:705 706.. code-block:: C707 708  if (0 < n) {709    auto v = *p;710    do {711      use(v);712      ++i;713    } while (i < n);714  }715 716Note that LoopRotate does not generally do such717hoisting.  Rather, it is an enabling transformation for other718passes like Loop-Invariant Code Motion (:ref:`-licm <passes-licm>`).719