1 | /* |
2 | * Copyright (C) 2016-2017 Apple Inc. All rights reserved. |
3 | * |
4 | * Redistribution and use in source and binary forms, with or without |
5 | * modification, are permitted provided that the following conditions |
6 | * are met: |
7 | * 1. Redistributions of source code must retain the above copyright |
8 | * notice, this list of conditions and the following disclaimer. |
9 | * 2. Redistributions in binary form must reproduce the above copyright |
10 | * notice, this list of conditions and the following disclaimer in the |
11 | * documentation and/or other materials provided with the distribution. |
12 | * |
13 | * THIS SOFTWARE IS PROVIDED BY APPLE INC. ``AS IS'' AND ANY |
14 | * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
15 | * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR |
16 | * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL APPLE INC. OR |
17 | * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, |
18 | * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, |
19 | * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR |
20 | * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY |
21 | * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT |
22 | * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE |
23 | * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
24 | */ |
25 | |
26 | #pragma once |
27 | |
28 | #if ENABLE(B3_JIT) |
29 | |
30 | #include "AirArg.h" |
31 | #include "AirInst.h" |
32 | #include <wtf/Vector.h> |
33 | |
34 | namespace JSC { namespace B3 { |
35 | |
36 | class Value; |
37 | |
38 | namespace Air { |
39 | |
40 | class Code; |
41 | |
42 | inline Opcode moveFor(Bank bank, Width width) |
43 | { |
44 | switch (width) { |
45 | case Width32: |
46 | return bank == GP ? Move32 : MoveFloat; |
47 | case Width64: |
48 | return bank == GP ? Move : MoveDouble; |
49 | default: |
50 | RELEASE_ASSERT_NOT_REACHED(); |
51 | return Oops; |
52 | } |
53 | } |
54 | |
55 | class ShufflePair { |
56 | public: |
57 | ShufflePair() |
58 | { |
59 | } |
60 | |
61 | ShufflePair(const Arg& src, const Arg& dst, Width width) |
62 | : m_src(src) |
63 | , m_dst(dst) |
64 | , m_width(width) |
65 | { |
66 | } |
67 | |
68 | const Arg& src() const { return m_src; } |
69 | const Arg& dst() const { return m_dst; } |
70 | |
71 | // The width determines the kind of move we do. You can only choose Width32 or Width64 right now. |
72 | // For GP, it picks between Move32 and Move. For FP, it picks between MoveFloat and MoveDouble. |
73 | Width width() const { return m_width; } |
74 | |
75 | Bank bank() const; |
76 | |
77 | // Creates an instruction sequence for the move represented by this shuffle pair. |
78 | // You need to pass Code because we may need to create a tmp. |
79 | Vector<Inst, 2> insts(Code&, Value* origin) const; |
80 | |
81 | void dump(PrintStream&) const; |
82 | |
83 | private: |
84 | Arg m_src; |
85 | Arg m_dst; |
86 | Width m_width { Width8 }; |
87 | }; |
88 | |
89 | // Create a Shuffle instruction. |
90 | Inst createShuffle(Value* origin, const Vector<ShufflePair>&); |
91 | |
92 | // Perform a shuffle of a given type. The scratch argument is mandatory. You should pass it as |
93 | // follows: If you know that you have scratch registers or temporaries available - that is, they're |
94 | // registers that are not mentioned in the shuffle, have the same type as the shuffle, and are not |
95 | // live at the shuffle - then you can pass them. If you don't have scratch registers available or if |
96 | // you don't feel like looking for them, you can pass memory locations. It's always safe to pass a |
97 | // pair of memory locations, and replacing either memory location with a register can be viewed as an |
98 | // optimization. It's a pretty important optimization. Some more notes: |
99 | // |
100 | // - We define scratch registers as things that are not live before the shuffle and are not one of |
101 | // the destinations of the shuffle. Not being live before the shuffle also means that they cannot |
102 | // be used for any of the sources of the shuffle. |
103 | // |
104 | // - A second scratch location is only needed when you have shuffle pairs where memory is used both |
105 | // as source and destination. |
106 | // |
107 | // - You're guaranteed not to need any scratch locations if there is a Swap instruction available for |
108 | // the type and you don't have any memory locations that are both the source and the destination of |
109 | // some pairs. GP supports Swap on x86 while FP never supports Swap. |
110 | // |
111 | // - Passing memory locations as scratch if are running emitShuffle() before register allocation is |
112 | // silly, since that will cause emitShuffle() to pick some specific registers when it does need |
113 | // scratch. One easy way to avoid that predicament is to ensure that you call emitShuffle() after |
114 | // register allocation. For this reason we could add a Shuffle instruction so that we can defer |
115 | // shufflings until after regalloc. |
116 | // |
117 | // - Shuffles with memory=>memory pairs are not very well tuned. You should avoid them if you want |
118 | // performance. If you need to do them, then making sure that you reserve a temporary is one way to |
119 | // get acceptable performance. |
120 | // |
121 | // NOTE: Use this method (and its friend below) to emit shuffles after register allocation. Before |
122 | // register allocation it is much better to simply use the Shuffle instruction. |
123 | Vector<Inst> emitShuffle( |
124 | Code& code, Vector<ShufflePair>, std::array<Arg, 2> scratch, Bank, Value* origin); |
125 | |
126 | // Perform a shuffle that involves any number of types. Pass scratch registers or memory locations |
127 | // for each type according to the rules above. |
128 | Vector<Inst> emitShuffle( |
129 | Code& code, const Vector<ShufflePair>&, |
130 | const std::array<Arg, 2>& gpScratch, const std::array<Arg, 2>& fpScratch, |
131 | Value* origin); |
132 | |
133 | } } } // namespace JSC::B3::Air |
134 | |
135 | #endif // ENABLE(B3_JIT) |
136 | |