• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- SpeculateAroundPHIs.h - Speculate around PHIs ------------*- C++ -*-===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 
10 #ifndef LLVM_TRANSFORMS_SCALAR_SPECULATEAROUNDPHIS_H
11 #define LLVM_TRANSFORMS_SCALAR_SPECULATEAROUNDPHIS_H
12 
13 #include "llvm/ADT/SetVector.h"
14 #include "llvm/Analysis/AssumptionCache.h"
15 #include "llvm/IR/Dominators.h"
16 #include "llvm/IR/Function.h"
17 #include "llvm/IR/PassManager.h"
18 #include "llvm/Support/Compiler.h"
19 #include <vector>
20 
21 namespace llvm {
22 
23 /// This pass handles simple speculating of  instructions around PHIs when
24 /// doing so is profitable for a particular target despite duplicated
25 /// instructions.
26 ///
27 /// The motivating example are PHIs of constants which will require
28 /// materializing the constants along each edge. If the PHI is used by an
29 /// instruction where the target can materialize the constant as part of the
30 /// instruction, it is profitable to speculate those instructions around the
31 /// PHI node. This can reduce dynamic instruction count as well as decrease
32 /// register pressure.
33 ///
34 /// Consider this IR for example:
35 ///   ```
36 ///   entry:
37 ///     br i1 %flag, label %a, label %b
38 ///
39 ///   a:
40 ///     br label %exit
41 ///
42 ///   b:
43 ///     br label %exit
44 ///
45 ///   exit:
46 ///     %p = phi i32 [ 7, %a ], [ 11, %b ]
47 ///     %sum = add i32 %arg, %p
48 ///     ret i32 %sum
49 ///   ```
50 /// To materialize the inputs to this PHI node may require an explicit
51 /// instruction. For example, on x86 this would turn into something like
52 ///   ```
53 ///     testq %eax, %eax
54 ///     movl $7, %rNN
55 ///     jne .L
56 ///     movl $11, %rNN
57 ///   .L:
58 ///     addl %edi, %rNN
59 ///     movl %rNN, %eax
60 ///     retq
61 ///   ```
62 /// When these constants can be folded directly into another instruction, it
63 /// would be preferable to avoid the potential for register pressure (above we
64 /// can easily avoid it, but that isn't always true) and simply duplicate the
65 /// instruction using the PHI:
66 ///   ```
67 ///   entry:
68 ///     br i1 %flag, label %a, label %b
69 ///
70 ///   a:
71 ///     %sum.1 = add i32 %arg, 7
72 ///     br label %exit
73 ///
74 ///   b:
75 ///     %sum.2 = add i32 %arg, 11
76 ///     br label %exit
77 ///
78 ///   exit:
79 ///     %p = phi i32 [ %sum.1, %a ], [ %sum.2, %b ]
80 ///     ret i32 %p
81 ///   ```
82 /// Which will generate something like the following on x86:
83 ///   ```
84 ///     testq %eax, %eax
85 ///     addl $7, %edi
86 ///     jne .L
87 ///     addl $11, %edi
88 ///   .L:
89 ///     movl %edi, %eax
90 ///     retq
91 ///   ```
92 ///
93 /// It is important to note that this pass is never intended to handle more
94 /// complex cases where speculating around PHIs allows simplifications of the
95 /// IR itself or other subsequent optimizations. Those can and should already
96 /// be handled before this pass is ever run by a more powerful analysis that
97 /// can reason about equivalences and common subexpressions. Classically, those
98 /// cases would be handled by a GVN-powered PRE or similar transform. This
99 /// pass, in contrast, is *only* interested in cases where despite no
100 /// simplifications to the IR itself, speculation is *faster* to execute. The
101 /// result of this is that the cost models which are appropriate to consider
102 /// here are relatively simple ones around execution and codesize cost, without
103 /// any need to consider simplifications or other transformations.
104 struct SpeculateAroundPHIsPass : PassInfoMixin<SpeculateAroundPHIsPass> {
105   /// Run the pass over the function.
106   PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM);
107 };
108 
109 } // end namespace llvm
110 
111 #endif // LLVM_TRANSFORMS_SCALAR_SPECULATEAROUNDPHIS_H
112