• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- SpeculateAroundPHIs.h - Speculate around PHIs ------------*- 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 #ifndef LLVM_TRANSFORMS_SCALAR_SPECULATEAROUNDPHIS_H
10 #define LLVM_TRANSFORMS_SCALAR_SPECULATEAROUNDPHIS_H
11 
12 #include "llvm/ADT/SetVector.h"
13 #include "llvm/Analysis/AssumptionCache.h"
14 #include "llvm/IR/Dominators.h"
15 #include "llvm/IR/Function.h"
16 #include "llvm/IR/PassManager.h"
17 #include "llvm/Support/Compiler.h"
18 #include <vector>
19 
20 namespace llvm {
21 
22 /// This pass handles simple speculating of  instructions around PHIs when
23 /// doing so is profitable for a particular target despite duplicated
24 /// instructions.
25 ///
26 /// The motivating example are PHIs of constants which will require
27 /// materializing the constants along each edge. If the PHI is used by an
28 /// instruction where the target can materialize the constant as part of the
29 /// instruction, it is profitable to speculate those instructions around the
30 /// PHI node. This can reduce dynamic instruction count as well as decrease
31 /// register pressure.
32 ///
33 /// Consider this IR for example:
34 ///   ```
35 ///   entry:
36 ///     br i1 %flag, label %a, label %b
37 ///
38 ///   a:
39 ///     br label %exit
40 ///
41 ///   b:
42 ///     br label %exit
43 ///
44 ///   exit:
45 ///     %p = phi i32 [ 7, %a ], [ 11, %b ]
46 ///     %sum = add i32 %arg, %p
47 ///     ret i32 %sum
48 ///   ```
49 /// To materialize the inputs to this PHI node may require an explicit
50 /// instruction. For example, on x86 this would turn into something like
51 ///   ```
52 ///     testq %eax, %eax
53 ///     movl $7, %rNN
54 ///     jne .L
55 ///     movl $11, %rNN
56 ///   .L:
57 ///     addl %edi, %rNN
58 ///     movl %rNN, %eax
59 ///     retq
60 ///   ```
61 /// When these constants can be folded directly into another instruction, it
62 /// would be preferable to avoid the potential for register pressure (above we
63 /// can easily avoid it, but that isn't always true) and simply duplicate the
64 /// instruction using the PHI:
65 ///   ```
66 ///   entry:
67 ///     br i1 %flag, label %a, label %b
68 ///
69 ///   a:
70 ///     %sum.1 = add i32 %arg, 7
71 ///     br label %exit
72 ///
73 ///   b:
74 ///     %sum.2 = add i32 %arg, 11
75 ///     br label %exit
76 ///
77 ///   exit:
78 ///     %p = phi i32 [ %sum.1, %a ], [ %sum.2, %b ]
79 ///     ret i32 %p
80 ///   ```
81 /// Which will generate something like the following on x86:
82 ///   ```
83 ///     testq %eax, %eax
84 ///     addl $7, %edi
85 ///     jne .L
86 ///     addl $11, %edi
87 ///   .L:
88 ///     movl %edi, %eax
89 ///     retq
90 ///   ```
91 ///
92 /// It is important to note that this pass is never intended to handle more
93 /// complex cases where speculating around PHIs allows simplifications of the
94 /// IR itself or other subsequent optimizations. Those can and should already
95 /// be handled before this pass is ever run by a more powerful analysis that
96 /// can reason about equivalences and common subexpressions. Classically, those
97 /// cases would be handled by a GVN-powered PRE or similar transform. This
98 /// pass, in contrast, is *only* interested in cases where despite no
99 /// simplifications to the IR itself, speculation is *faster* to execute. The
100 /// result of this is that the cost models which are appropriate to consider
101 /// here are relatively simple ones around execution and codesize cost, without
102 /// any need to consider simplifications or other transformations.
103 struct SpeculateAroundPHIsPass : PassInfoMixin<SpeculateAroundPHIsPass> {
104   /// Run the pass over the function.
105   PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM);
106 };
107 
108 } // end namespace llvm
109 
110 #endif // LLVM_TRANSFORMS_SCALAR_SPECULATEAROUNDPHIS_H
111