1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
|
//===- AMDGPULegalizerInfo.cpp -----------------------------------*- C++ -*-==//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
/// \file
/// This file implements the targeting of the Machinelegalizer class for
/// AMDGPU.
/// \todo This should be generated by TableGen.
//===----------------------------------------------------------------------===//
#include "AMDGPU.h"
#include "AMDGPULegalizerInfo.h"
#include "llvm/CodeGen/TargetOpcodes.h"
#include "llvm/CodeGen/ValueTypes.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/Type.h"
#include "llvm/Support/Debug.h"
using namespace llvm;
using namespace LegalizeActions;
AMDGPULegalizerInfo::AMDGPULegalizerInfo(const SISubtarget &ST,
const GCNTargetMachine &TM) {
using namespace TargetOpcode;
const LLT S1= LLT::scalar(1);
const LLT V2S16 = LLT::vector(2, 16);
const LLT S32 = LLT::scalar(32);
const LLT S64 = LLT::scalar(64);
const LLT P1 = LLT::pointer(AMDGPUAS::GLOBAL_ADDRESS, 64);
const LLT P2 = LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64);
setAction({G_ADD, S32}, Legal);
setAction({G_MUL, S32}, Legal);
setAction({G_AND, S32}, Legal);
setAction({G_OR, S32}, Legal);
setAction({G_XOR, S32}, Legal);
setAction({G_BITCAST, V2S16}, Legal);
setAction({G_BITCAST, 1, S32}, Legal);
setAction({G_BITCAST, S32}, Legal);
setAction({G_BITCAST, 1, V2S16}, Legal);
// FIXME: i1 operands to intrinsics should always be legal, but other i1
// values may not be legal. We need to figure out how to distinguish
// between these two scenarios.
setAction({G_CONSTANT, S1}, Legal);
setAction({G_CONSTANT, S32}, Legal);
setAction({G_CONSTANT, S64}, Legal);
setAction({G_FCONSTANT, S32}, Legal);
setAction({G_FCONSTANT, S64}, Legal);
setAction({G_IMPLICIT_DEF, S32}, Legal);
setAction({G_IMPLICIT_DEF, S64}, Legal);
setAction({G_FADD, S32}, Legal);
setAction({G_FCMP, S1}, Legal);
setAction({G_FCMP, 1, S32}, Legal);
setAction({G_FCMP, 1, S64}, Legal);
setAction({G_FMUL, S32}, Legal);
setAction({G_ZEXT, S64}, Legal);
setAction({G_ZEXT, 1, S32}, Legal);
setAction({G_FPTOSI, S32}, Legal);
setAction({G_FPTOSI, 1, S32}, Legal);
setAction({G_FPTOUI, S32}, Legal);
setAction({G_FPTOUI, 1, S32}, Legal);
setAction({G_GEP, P1}, Legal);
setAction({G_GEP, P2}, Legal);
setAction({G_GEP, 1, S64}, Legal);
setAction({G_ICMP, S1}, Legal);
setAction({G_ICMP, 1, S32}, Legal);
setAction({G_LOAD, P1}, Legal);
setAction({G_LOAD, P2}, Legal);
setAction({G_LOAD, S32}, Legal);
setAction({G_LOAD, 1, P1}, Legal);
setAction({G_LOAD, 1, P2}, Legal);
setAction({G_SELECT, S32}, Legal);
setAction({G_SELECT, 1, S1}, Legal);
setAction({G_SHL, S32}, Legal);
setAction({G_STORE, S32}, Legal);
setAction({G_STORE, 1, P1}, Legal);
// FIXME: When RegBankSelect inserts copies, it will only create new
// registers with scalar types. This means we can end up with
// G_LOAD/G_STORE/G_GEP instruction with scalar types for their pointer
// operands. In assert builds, the instruction selector will assert
// if it sees a generic instruction which isn't legal, so we need to
// tell it that scalar types are legal for pointer operands
setAction({G_GEP, S64}, Legal);
setAction({G_LOAD, 1, S64}, Legal);
setAction({G_STORE, 1, S64}, Legal);
for (unsigned Op : {G_EXTRACT_VECTOR_ELT, G_INSERT_VECTOR_ELT}) {
getActionDefinitionsBuilder(Op)
.legalIf([=](const LegalityQuery &Query) {
const LLT &VecTy = Query.Types[1];
const LLT &IdxTy = Query.Types[2];
return VecTy.getSizeInBits() % 32 == 0 &&
VecTy.getSizeInBits() <= 512 &&
IdxTy.getSizeInBits() == 32;
});
}
// FIXME: Doesn't handle extract of illegal sizes.
getActionDefinitionsBuilder(G_EXTRACT)
.unsupportedIf([=](const LegalityQuery &Query) {
return Query.Types[0].getSizeInBits() >= Query.Types[1].getSizeInBits();
})
.legalIf([=](const LegalityQuery &Query) {
const LLT &Ty0 = Query.Types[0];
const LLT &Ty1 = Query.Types[1];
return (Ty0.getSizeInBits() % 32 == 0) &&
(Ty1.getSizeInBits() % 32 == 0);
});
// Merge/Unmerge
for (unsigned Op : {G_MERGE_VALUES, G_UNMERGE_VALUES}) {
unsigned BigTyIdx = Op == G_MERGE_VALUES ? 0 : 1;
unsigned LitTyIdx = Op == G_MERGE_VALUES ? 1 : 0;
getActionDefinitionsBuilder(Op)
.legalIf([=](const LegalityQuery &Query) {
const LLT &BigTy = Query.Types[BigTyIdx];
const LLT &LitTy = Query.Types[LitTyIdx];
return BigTy.getSizeInBits() % 32 == 0 &&
LitTy.getSizeInBits() % 32 == 0 &&
BigTy.getSizeInBits() <= 512;
})
// Any vectors left are the wrong size. Scalarize them.
.fewerElementsIf([](const LegalityQuery &Query) { return true; },
[](const LegalityQuery &Query) {
return std::make_pair(
0, Query.Types[0].getElementType());
})
.fewerElementsIf([](const LegalityQuery &Query) { return true; },
[](const LegalityQuery &Query) {
return std::make_pair(
1, Query.Types[1].getElementType());
});
}
computeTables();
}
|
