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
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
|
package compiler
// This file contains helper functions to create calls to LLVM intrinsics.
import (
"go/token"
"strconv"
"strings"
"tinygo.org/x/go-llvm"
)
// Define unimplemented intrinsic functions.
//
// Some functions are either normally implemented in Go assembly (like
// sync/atomic functions) or intentionally left undefined to be implemented
// directly in the compiler (like runtime/volatile functions). Either way, look
// for these and implement them if this is the case.
func (b *builder) defineIntrinsicFunction() {
name := b.fn.RelString(nil)
switch {
case name == "runtime.memcpy" || name == "runtime.memmove":
b.createMemoryCopyImpl()
case name == "runtime.memzero":
b.createMemoryZeroImpl()
case name == "runtime.KeepAlive":
b.createKeepAliveImpl()
case strings.HasPrefix(name, "runtime/volatile.Load"):
b.createVolatileLoad()
case strings.HasPrefix(name, "runtime/volatile.Store"):
b.createVolatileStore()
case strings.HasPrefix(name, "sync/atomic.") && token.IsExported(b.fn.Name()):
b.createFunctionStart(true)
returnValue := b.createAtomicOp(b.fn.Name())
if !returnValue.IsNil() {
b.CreateRet(returnValue)
} else {
b.CreateRetVoid()
}
}
}
// createMemoryCopyImpl creates a call to a builtin LLVM memcpy or memmove
// function, declaring this function if needed. These calls are treated
// specially by optimization passes possibly resulting in better generated code,
// and will otherwise be lowered to regular libc memcpy/memmove calls.
func (b *builder) createMemoryCopyImpl() {
b.createFunctionStart(true)
fnName := "llvm." + b.fn.Name() + ".p0.p0.i" + strconv.Itoa(b.uintptrType.IntTypeWidth())
llvmFn := b.mod.NamedFunction(fnName)
if llvmFn.IsNil() {
fnType := llvm.FunctionType(b.ctx.VoidType(), []llvm.Type{b.dataPtrType, b.dataPtrType, b.uintptrType, b.ctx.Int1Type()}, false)
llvmFn = llvm.AddFunction(b.mod, fnName, fnType)
}
var params []llvm.Value
for _, param := range b.fn.Params {
params = append(params, b.getValue(param, getPos(b.fn)))
}
params = append(params, llvm.ConstInt(b.ctx.Int1Type(), 0, false))
b.CreateCall(llvmFn.GlobalValueType(), llvmFn, params, "")
b.CreateRetVoid()
}
// createMemoryZeroImpl creates calls to llvm.memset.* to zero a block of
// memory, declaring the function if needed. These calls will be lowered to
// regular libc memset calls if they aren't optimized out in a different way.
func (b *builder) createMemoryZeroImpl() {
b.createFunctionStart(true)
llvmFn := b.getMemsetFunc()
params := []llvm.Value{
b.getValue(b.fn.Params[0], getPos(b.fn)),
llvm.ConstInt(b.ctx.Int8Type(), 0, false),
b.getValue(b.fn.Params[1], getPos(b.fn)),
llvm.ConstInt(b.ctx.Int1Type(), 0, false),
}
b.CreateCall(llvmFn.GlobalValueType(), llvmFn, params, "")
b.CreateRetVoid()
}
// Return the llvm.memset.p0.i8 function declaration.
func (c *compilerContext) getMemsetFunc() llvm.Value {
fnName := "llvm.memset.p0.i" + strconv.Itoa(c.uintptrType.IntTypeWidth())
llvmFn := c.mod.NamedFunction(fnName)
if llvmFn.IsNil() {
fnType := llvm.FunctionType(c.ctx.VoidType(), []llvm.Type{c.dataPtrType, c.ctx.Int8Type(), c.uintptrType, c.ctx.Int1Type()}, false)
llvmFn = llvm.AddFunction(c.mod, fnName, fnType)
}
return llvmFn
}
// createKeepAlive creates the runtime.KeepAlive function. It is implemented
// using inline assembly.
func (b *builder) createKeepAliveImpl() {
b.createFunctionStart(true)
// Get the underlying value of the interface value.
interfaceValue := b.getValue(b.fn.Params[0], getPos(b.fn))
pointerValue := b.CreateExtractValue(interfaceValue, 1, "")
// Create an equivalent of the following C code, which is basically just a
// nop but ensures the pointerValue is kept alive:
//
// __asm__ __volatile__("" : : "r"(pointerValue))
//
// It should be portable to basically everything as the "r" register type
// exists basically everywhere.
asmType := llvm.FunctionType(b.ctx.VoidType(), []llvm.Type{b.dataPtrType}, false)
asmFn := llvm.InlineAsm(asmType, "", "r", true, false, 0, false)
b.createCall(asmType, asmFn, []llvm.Value{pointerValue}, "")
b.CreateRetVoid()
}
var mathToLLVMMapping = map[string]string{
"math.Ceil": "llvm.ceil.f64",
"math.Exp": "llvm.exp.f64",
"math.Exp2": "llvm.exp2.f64",
"math.Floor": "llvm.floor.f64",
"math.Log": "llvm.log.f64",
"math.Sqrt": "llvm.sqrt.f64",
"math.Trunc": "llvm.trunc.f64",
}
// defineMathOp defines a math function body as a call to a LLVM intrinsic,
// instead of the regular Go implementation. This allows LLVM to reason about
// the math operation and (depending on the architecture) allows it to lower the
// operation to very fast floating point instructions. If this is not possible,
// LLVM will emit a call to a libm function that implements the same operation.
//
// One example of an optimization that LLVM can do is to convert
// float32(math.Sqrt(float64(v))) to a 32-bit floating point operation, which is
// beneficial on architectures where 64-bit floating point operations are (much)
// more expensive than 32-bit ones.
func (b *builder) defineMathOp() {
b.createFunctionStart(true)
llvmName := mathToLLVMMapping[b.fn.RelString(nil)]
if llvmName == "" {
panic("unreachable: unknown math operation") // sanity check
}
llvmFn := b.mod.NamedFunction(llvmName)
if llvmFn.IsNil() {
// The intrinsic doesn't exist yet, so declare it.
// At the moment, all supported intrinsics have the form "double
// foo(double %x)" so we can hardcode the signature here.
llvmType := llvm.FunctionType(b.ctx.DoubleType(), []llvm.Type{b.ctx.DoubleType()}, false)
llvmFn = llvm.AddFunction(b.mod, llvmName, llvmType)
}
// Create a call to the intrinsic.
args := make([]llvm.Value, len(b.fn.Params))
for i, param := range b.fn.Params {
args[i] = b.getValue(param, getPos(b.fn))
}
result := b.CreateCall(llvmFn.GlobalValueType(), llvmFn, args, "")
b.CreateRet(result)
}
// Implement most math/bits functions.
//
// This implements all the functions that operate on bits. It does not yet
// implement the arithmetic functions (like bits.Add), which also have LLVM
// intrinsics.
func (b *builder) defineMathBitsIntrinsic() bool {
if b.fn.Pkg.Pkg.Path() != "math/bits" {
return false
}
name := b.fn.Name()
switch name {
case "LeadingZeros", "LeadingZeros8", "LeadingZeros16", "LeadingZeros32", "LeadingZeros64",
"TrailingZeros", "TrailingZeros8", "TrailingZeros16", "TrailingZeros32", "TrailingZeros64":
b.createFunctionStart(true)
param := b.getValue(b.fn.Params[0], b.fn.Pos())
valueType := param.Type()
var intrinsicName string
if strings.HasPrefix(name, "Leading") { // LeadingZeros
intrinsicName = "llvm.ctlz.i" + strconv.Itoa(valueType.IntTypeWidth())
} else { // TrailingZeros
intrinsicName = "llvm.cttz.i" + strconv.Itoa(valueType.IntTypeWidth())
}
llvmFn := b.mod.NamedFunction(intrinsicName)
llvmFnType := llvm.FunctionType(valueType, []llvm.Type{valueType, b.ctx.Int1Type()}, false)
if llvmFn.IsNil() {
llvmFn = llvm.AddFunction(b.mod, intrinsicName, llvmFnType)
}
result := b.createCall(llvmFnType, llvmFn, []llvm.Value{
param,
llvm.ConstInt(b.ctx.Int1Type(), 0, false),
}, "")
result = b.createZExtOrTrunc(result, b.intType)
b.CreateRet(result)
return true
case "Len", "Len8", "Len16", "Len32", "Len64":
// bits.Len can be implemented as:
// (unsafe.Sizeof(v) * 8) - bits.LeadingZeros(n)
// Not sure why this isn't already done in the standard library, as it
// is much simpler than a lookup table.
b.createFunctionStart(true)
param := b.getValue(b.fn.Params[0], b.fn.Pos())
valueType := param.Type()
valueBits := valueType.IntTypeWidth()
intrinsicName := "llvm.ctlz.i" + strconv.Itoa(valueBits)
llvmFn := b.mod.NamedFunction(intrinsicName)
llvmFnType := llvm.FunctionType(valueType, []llvm.Type{valueType, b.ctx.Int1Type()}, false)
if llvmFn.IsNil() {
llvmFn = llvm.AddFunction(b.mod, intrinsicName, llvmFnType)
}
result := b.createCall(llvmFnType, llvmFn, []llvm.Value{
param,
llvm.ConstInt(b.ctx.Int1Type(), 0, false),
}, "")
result = b.createZExtOrTrunc(result, b.intType)
maxLen := llvm.ConstInt(b.intType, uint64(valueBits), false) // number of bits in the value
result = b.CreateSub(maxLen, result, "")
b.CreateRet(result)
return true
case "OnesCount", "OnesCount8", "OnesCount16", "OnesCount32", "OnesCount64":
b.createFunctionStart(true)
param := b.getValue(b.fn.Params[0], b.fn.Pos())
valueType := param.Type()
intrinsicName := "llvm.ctpop.i" + strconv.Itoa(valueType.IntTypeWidth())
llvmFn := b.mod.NamedFunction(intrinsicName)
llvmFnType := llvm.FunctionType(valueType, []llvm.Type{valueType}, false)
if llvmFn.IsNil() {
llvmFn = llvm.AddFunction(b.mod, intrinsicName, llvmFnType)
}
result := b.createCall(llvmFnType, llvmFn, []llvm.Value{param}, "")
result = b.createZExtOrTrunc(result, b.intType)
b.CreateRet(result)
return true
case "Reverse", "Reverse8", "Reverse16", "Reverse32", "Reverse64",
"ReverseBytes", "ReverseBytes16", "ReverseBytes32", "ReverseBytes64":
b.createFunctionStart(true)
param := b.getValue(b.fn.Params[0], b.fn.Pos())
valueType := param.Type()
var intrinsicName string
if strings.HasPrefix(name, "ReverseBytes") {
intrinsicName = "llvm.bswap.i" + strconv.Itoa(valueType.IntTypeWidth())
} else { // Reverse
intrinsicName = "llvm.bitreverse.i" + strconv.Itoa(valueType.IntTypeWidth())
}
llvmFn := b.mod.NamedFunction(intrinsicName)
llvmFnType := llvm.FunctionType(valueType, []llvm.Type{valueType}, false)
if llvmFn.IsNil() {
llvmFn = llvm.AddFunction(b.mod, intrinsicName, llvmFnType)
}
result := b.createCall(llvmFnType, llvmFn, []llvm.Value{param}, "")
b.CreateRet(result)
return true
case "RotateLeft", "RotateLeft8", "RotateLeft16", "RotateLeft32", "RotateLeft64":
// Warning: the documentation says these functions must be constant time.
// I do not think LLVM guarantees this, but there's a good chance LLVM
// already recognized the rotate instruction so it probably won't get
// any _worse_ by implementing these rotate functions.
b.createFunctionStart(true)
x := b.getValue(b.fn.Params[0], b.fn.Pos())
k := b.getValue(b.fn.Params[1], b.fn.Pos())
valueType := x.Type()
intrinsicName := "llvm.fshl.i" + strconv.Itoa(valueType.IntTypeWidth())
llvmFn := b.mod.NamedFunction(intrinsicName)
llvmFnType := llvm.FunctionType(valueType, []llvm.Type{valueType, valueType, valueType}, false)
if llvmFn.IsNil() {
llvmFn = llvm.AddFunction(b.mod, intrinsicName, llvmFnType)
}
k = b.createZExtOrTrunc(k, valueType)
result := b.createCall(llvmFnType, llvmFn, []llvm.Value{x, x, k}, "")
b.CreateRet(result)
return true
default:
return false
}
}
|