xref: /freebsd/contrib/llvm-project/llvm/include/llvm/Transforms/Scalar/TailRecursionElimination.h (revision 5ca8e32633c4ffbbcd6762e5888b6a4ba0708c6c)
1 //===---- TailRecursionElimination.h ----------------------------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file transforms calls of the current function (self recursion) followed
10 // by a return instruction with a branch to the entry of the function, creating
11 // a loop.  This pass also implements the following extensions to the basic
12 // algorithm:
13 //
14 //  1. Trivial instructions between the call and return do not prevent the
15 //     transformation from taking place, though currently the analysis cannot
16 //     support moving any really useful instructions (only dead ones).
17 //  2. This pass transforms functions that are prevented from being tail
18 //     recursive by an associative and commutative expression to use an
19 //     accumulator variable, thus compiling the typical naive factorial or
20 //     'fib' implementation into efficient code.
21 //  3. TRE is performed if the function returns void, if the return
22 //     returns the result returned by the call, or if the function returns a
23 //     run-time constant on all exits from the function.  It is possible, though
24 //     unlikely, that the return returns something else (like constant 0), and
25 //     can still be TRE'd.  It can be TRE'd if ALL OTHER return instructions in
26 //     the function return the exact same value.
27 //  4. If it can prove that callees do not access their caller stack frame,
28 //     they are marked as eligible for tail call elimination (by the code
29 //     generator).
30 //
31 // There are several improvements that could be made:
32 //
33 //  1. If the function has any alloca instructions, these instructions will be
34 //     moved out of the entry block of the function, causing them to be
35 //     evaluated each time through the tail recursion.  Safely keeping allocas
36 //     in the entry block requires analysis to proves that the tail-called
37 //     function does not read or write the stack object.
38 //  2. Tail recursion is only performed if the call immediately precedes the
39 //     return instruction.  It's possible that there could be a jump between
40 //     the call and the return.
41 //  3. There can be intervening operations between the call and the return that
42 //     prevent the TRE from occurring.  For example, there could be GEP's and
43 //     stores to memory that will not be read or written by the call.  This
44 //     requires some substantial analysis (such as with DSA) to prove safe to
45 //     move ahead of the call, but doing so could allow many more TREs to be
46 //     performed, for example in TreeAdd/TreeAlloc from the treeadd benchmark.
47 //  4. The algorithm we use to detect if callees access their caller stack
48 //     frames is very primitive.
49 //
50 //===----------------------------------------------------------------------===//
51 
52 #ifndef LLVM_TRANSFORMS_SCALAR_TAILRECURSIONELIMINATION_H
53 #define LLVM_TRANSFORMS_SCALAR_TAILRECURSIONELIMINATION_H
54 
55 #include "llvm/IR/PassManager.h"
56 
57 namespace llvm {
58 
59 class Function;
60 
61 struct TailCallElimPass : PassInfoMixin<TailCallElimPass> {
62   PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM);
63 };
64 }
65 
66 #endif // LLVM_TRANSFORMS_SCALAR_TAILRECURSIONELIMINATION_H
67