1 // Copyright 2015 The Go Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style
3 // license that can be found in the LICENSE file.
19 "cmd/compile/internal/abi"
20 "cmd/compile/internal/base"
21 "cmd/compile/internal/ir"
22 "cmd/compile/internal/liveness"
23 "cmd/compile/internal/objw"
24 "cmd/compile/internal/reflectdata"
25 "cmd/compile/internal/ssa"
26 "cmd/compile/internal/staticdata"
27 "cmd/compile/internal/typecheck"
28 "cmd/compile/internal/types"
36 var ssaConfig *ssa.Config
37 var ssaCaches []ssa.Cache
39 var ssaDump string // early copy of $GOSSAFUNC; the func name to dump output for
40 var ssaDir string // optional destination for ssa dump file
41 var ssaDumpStdout bool // whether to dump to stdout
42 var ssaDumpCFG string // generate CFGs for these phases
43 const ssaDumpFile = "ssa.html"
45 // ssaDumpInlined holds all inlined functions when ssaDump contains a function name.
46 var ssaDumpInlined []*ir.Func
48 func DumpInline(fn *ir.Func) {
49 if ssaDump != "" && ssaDump == ir.FuncName(fn) {
50 ssaDumpInlined = append(ssaDumpInlined, fn)
55 ssaDump = os.Getenv("GOSSAFUNC")
56 ssaDir = os.Getenv("GOSSADIR")
58 if strings.HasSuffix(ssaDump, "+") {
59 ssaDump = ssaDump[:len(ssaDump)-1]
62 spl := strings.Split(ssaDump, ":")
71 types_ := ssa.NewTypes()
77 // Generate a few pointer types that are uncommon in the frontend but common in the backend.
78 // Caching is disabled in the backend, so generating these here avoids allocations.
79 _ = types.NewPtr(types.Types[types.TINTER]) // *interface{}
80 _ = types.NewPtr(types.NewPtr(types.Types[types.TSTRING])) // **string
81 _ = types.NewPtr(types.NewSlice(types.Types[types.TINTER])) // *[]interface{}
82 _ = types.NewPtr(types.NewPtr(types.ByteType)) // **byte
83 _ = types.NewPtr(types.NewSlice(types.ByteType)) // *[]byte
84 _ = types.NewPtr(types.NewSlice(types.Types[types.TSTRING])) // *[]string
85 _ = types.NewPtr(types.NewPtr(types.NewPtr(types.Types[types.TUINT8]))) // ***uint8
86 _ = types.NewPtr(types.Types[types.TINT16]) // *int16
87 _ = types.NewPtr(types.Types[types.TINT64]) // *int64
88 _ = types.NewPtr(types.ErrorType) // *error
89 types.NewPtrCacheEnabled = false
90 ssaConfig = ssa.NewConfig(base.Ctxt.Arch.Name, *types_, base.Ctxt, base.Flag.N == 0, Arch.SoftFloat)
91 ssaConfig.Race = base.Flag.Race
92 ssaCaches = make([]ssa.Cache, base.Flag.LowerC)
94 // Set up some runtime functions we'll need to call.
95 ir.Syms.AssertE2I = typecheck.LookupRuntimeFunc("assertE2I")
96 ir.Syms.AssertE2I2 = typecheck.LookupRuntimeFunc("assertE2I2")
97 ir.Syms.AssertI2I = typecheck.LookupRuntimeFunc("assertI2I")
98 ir.Syms.AssertI2I2 = typecheck.LookupRuntimeFunc("assertI2I2")
99 ir.Syms.CgoCheckMemmove = typecheck.LookupRuntimeFunc("cgoCheckMemmove")
100 ir.Syms.CgoCheckPtrWrite = typecheck.LookupRuntimeFunc("cgoCheckPtrWrite")
101 ir.Syms.CheckPtrAlignment = typecheck.LookupRuntimeFunc("checkptrAlignment")
102 ir.Syms.Deferproc = typecheck.LookupRuntimeFunc("deferproc")
103 ir.Syms.DeferprocStack = typecheck.LookupRuntimeFunc("deferprocStack")
104 ir.Syms.Deferreturn = typecheck.LookupRuntimeFunc("deferreturn")
105 ir.Syms.Duffcopy = typecheck.LookupRuntimeFunc("duffcopy")
106 ir.Syms.Duffzero = typecheck.LookupRuntimeFunc("duffzero")
107 ir.Syms.GCWriteBarrier[0] = typecheck.LookupRuntimeFunc("gcWriteBarrier1")
108 ir.Syms.GCWriteBarrier[1] = typecheck.LookupRuntimeFunc("gcWriteBarrier2")
109 ir.Syms.GCWriteBarrier[2] = typecheck.LookupRuntimeFunc("gcWriteBarrier3")
110 ir.Syms.GCWriteBarrier[3] = typecheck.LookupRuntimeFunc("gcWriteBarrier4")
111 ir.Syms.GCWriteBarrier[4] = typecheck.LookupRuntimeFunc("gcWriteBarrier5")
112 ir.Syms.GCWriteBarrier[5] = typecheck.LookupRuntimeFunc("gcWriteBarrier6")
113 ir.Syms.GCWriteBarrier[6] = typecheck.LookupRuntimeFunc("gcWriteBarrier7")
114 ir.Syms.GCWriteBarrier[7] = typecheck.LookupRuntimeFunc("gcWriteBarrier8")
115 ir.Syms.Goschedguarded = typecheck.LookupRuntimeFunc("goschedguarded")
116 ir.Syms.Growslice = typecheck.LookupRuntimeFunc("growslice")
117 ir.Syms.Memmove = typecheck.LookupRuntimeFunc("memmove")
118 ir.Syms.Msanread = typecheck.LookupRuntimeFunc("msanread")
119 ir.Syms.Msanwrite = typecheck.LookupRuntimeFunc("msanwrite")
120 ir.Syms.Msanmove = typecheck.LookupRuntimeFunc("msanmove")
121 ir.Syms.Asanread = typecheck.LookupRuntimeFunc("asanread")
122 ir.Syms.Asanwrite = typecheck.LookupRuntimeFunc("asanwrite")
123 ir.Syms.Newobject = typecheck.LookupRuntimeFunc("newobject")
124 ir.Syms.Newproc = typecheck.LookupRuntimeFunc("newproc")
125 ir.Syms.Panicdivide = typecheck.LookupRuntimeFunc("panicdivide")
126 ir.Syms.PanicdottypeE = typecheck.LookupRuntimeFunc("panicdottypeE")
127 ir.Syms.PanicdottypeI = typecheck.LookupRuntimeFunc("panicdottypeI")
128 ir.Syms.Panicnildottype = typecheck.LookupRuntimeFunc("panicnildottype")
129 ir.Syms.Panicoverflow = typecheck.LookupRuntimeFunc("panicoverflow")
130 ir.Syms.Panicshift = typecheck.LookupRuntimeFunc("panicshift")
131 ir.Syms.Raceread = typecheck.LookupRuntimeFunc("raceread")
132 ir.Syms.Racereadrange = typecheck.LookupRuntimeFunc("racereadrange")
133 ir.Syms.Racewrite = typecheck.LookupRuntimeFunc("racewrite")
134 ir.Syms.Racewriterange = typecheck.LookupRuntimeFunc("racewriterange")
135 ir.Syms.WBZero = typecheck.LookupRuntimeFunc("wbZero")
136 ir.Syms.WBMove = typecheck.LookupRuntimeFunc("wbMove")
137 ir.Syms.X86HasPOPCNT = typecheck.LookupRuntimeVar("x86HasPOPCNT") // bool
138 ir.Syms.X86HasSSE41 = typecheck.LookupRuntimeVar("x86HasSSE41") // bool
139 ir.Syms.X86HasFMA = typecheck.LookupRuntimeVar("x86HasFMA") // bool
140 ir.Syms.ARMHasVFPv4 = typecheck.LookupRuntimeVar("armHasVFPv4") // bool
141 ir.Syms.ARM64HasATOMICS = typecheck.LookupRuntimeVar("arm64HasATOMICS") // bool
142 ir.Syms.Staticuint64s = typecheck.LookupRuntimeVar("staticuint64s")
143 ir.Syms.Typedmemmove = typecheck.LookupRuntimeFunc("typedmemmove")
144 ir.Syms.Udiv = typecheck.LookupRuntimeVar("udiv") // asm func with special ABI
145 ir.Syms.WriteBarrier = typecheck.LookupRuntimeVar("writeBarrier") // struct { bool; ... }
146 ir.Syms.Zerobase = typecheck.LookupRuntimeVar("zerobase")
148 if Arch.LinkArch.Family == sys.Wasm {
149 BoundsCheckFunc[ssa.BoundsIndex] = typecheck.LookupRuntimeFunc("goPanicIndex")
150 BoundsCheckFunc[ssa.BoundsIndexU] = typecheck.LookupRuntimeFunc("goPanicIndexU")
151 BoundsCheckFunc[ssa.BoundsSliceAlen] = typecheck.LookupRuntimeFunc("goPanicSliceAlen")
152 BoundsCheckFunc[ssa.BoundsSliceAlenU] = typecheck.LookupRuntimeFunc("goPanicSliceAlenU")
153 BoundsCheckFunc[ssa.BoundsSliceAcap] = typecheck.LookupRuntimeFunc("goPanicSliceAcap")
154 BoundsCheckFunc[ssa.BoundsSliceAcapU] = typecheck.LookupRuntimeFunc("goPanicSliceAcapU")
155 BoundsCheckFunc[ssa.BoundsSliceB] = typecheck.LookupRuntimeFunc("goPanicSliceB")
156 BoundsCheckFunc[ssa.BoundsSliceBU] = typecheck.LookupRuntimeFunc("goPanicSliceBU")
157 BoundsCheckFunc[ssa.BoundsSlice3Alen] = typecheck.LookupRuntimeFunc("goPanicSlice3Alen")
158 BoundsCheckFunc[ssa.BoundsSlice3AlenU] = typecheck.LookupRuntimeFunc("goPanicSlice3AlenU")
159 BoundsCheckFunc[ssa.BoundsSlice3Acap] = typecheck.LookupRuntimeFunc("goPanicSlice3Acap")
160 BoundsCheckFunc[ssa.BoundsSlice3AcapU] = typecheck.LookupRuntimeFunc("goPanicSlice3AcapU")
161 BoundsCheckFunc[ssa.BoundsSlice3B] = typecheck.LookupRuntimeFunc("goPanicSlice3B")
162 BoundsCheckFunc[ssa.BoundsSlice3BU] = typecheck.LookupRuntimeFunc("goPanicSlice3BU")
163 BoundsCheckFunc[ssa.BoundsSlice3C] = typecheck.LookupRuntimeFunc("goPanicSlice3C")
164 BoundsCheckFunc[ssa.BoundsSlice3CU] = typecheck.LookupRuntimeFunc("goPanicSlice3CU")
165 BoundsCheckFunc[ssa.BoundsConvert] = typecheck.LookupRuntimeFunc("goPanicSliceConvert")
167 BoundsCheckFunc[ssa.BoundsIndex] = typecheck.LookupRuntimeFunc("panicIndex")
168 BoundsCheckFunc[ssa.BoundsIndexU] = typecheck.LookupRuntimeFunc("panicIndexU")
169 BoundsCheckFunc[ssa.BoundsSliceAlen] = typecheck.LookupRuntimeFunc("panicSliceAlen")
170 BoundsCheckFunc[ssa.BoundsSliceAlenU] = typecheck.LookupRuntimeFunc("panicSliceAlenU")
171 BoundsCheckFunc[ssa.BoundsSliceAcap] = typecheck.LookupRuntimeFunc("panicSliceAcap")
172 BoundsCheckFunc[ssa.BoundsSliceAcapU] = typecheck.LookupRuntimeFunc("panicSliceAcapU")
173 BoundsCheckFunc[ssa.BoundsSliceB] = typecheck.LookupRuntimeFunc("panicSliceB")
174 BoundsCheckFunc[ssa.BoundsSliceBU] = typecheck.LookupRuntimeFunc("panicSliceBU")
175 BoundsCheckFunc[ssa.BoundsSlice3Alen] = typecheck.LookupRuntimeFunc("panicSlice3Alen")
176 BoundsCheckFunc[ssa.BoundsSlice3AlenU] = typecheck.LookupRuntimeFunc("panicSlice3AlenU")
177 BoundsCheckFunc[ssa.BoundsSlice3Acap] = typecheck.LookupRuntimeFunc("panicSlice3Acap")
178 BoundsCheckFunc[ssa.BoundsSlice3AcapU] = typecheck.LookupRuntimeFunc("panicSlice3AcapU")
179 BoundsCheckFunc[ssa.BoundsSlice3B] = typecheck.LookupRuntimeFunc("panicSlice3B")
180 BoundsCheckFunc[ssa.BoundsSlice3BU] = typecheck.LookupRuntimeFunc("panicSlice3BU")
181 BoundsCheckFunc[ssa.BoundsSlice3C] = typecheck.LookupRuntimeFunc("panicSlice3C")
182 BoundsCheckFunc[ssa.BoundsSlice3CU] = typecheck.LookupRuntimeFunc("panicSlice3CU")
183 BoundsCheckFunc[ssa.BoundsConvert] = typecheck.LookupRuntimeFunc("panicSliceConvert")
185 if Arch.LinkArch.PtrSize == 4 {
186 ExtendCheckFunc[ssa.BoundsIndex] = typecheck.LookupRuntimeVar("panicExtendIndex")
187 ExtendCheckFunc[ssa.BoundsIndexU] = typecheck.LookupRuntimeVar("panicExtendIndexU")
188 ExtendCheckFunc[ssa.BoundsSliceAlen] = typecheck.LookupRuntimeVar("panicExtendSliceAlen")
189 ExtendCheckFunc[ssa.BoundsSliceAlenU] = typecheck.LookupRuntimeVar("panicExtendSliceAlenU")
190 ExtendCheckFunc[ssa.BoundsSliceAcap] = typecheck.LookupRuntimeVar("panicExtendSliceAcap")
191 ExtendCheckFunc[ssa.BoundsSliceAcapU] = typecheck.LookupRuntimeVar("panicExtendSliceAcapU")
192 ExtendCheckFunc[ssa.BoundsSliceB] = typecheck.LookupRuntimeVar("panicExtendSliceB")
193 ExtendCheckFunc[ssa.BoundsSliceBU] = typecheck.LookupRuntimeVar("panicExtendSliceBU")
194 ExtendCheckFunc[ssa.BoundsSlice3Alen] = typecheck.LookupRuntimeVar("panicExtendSlice3Alen")
195 ExtendCheckFunc[ssa.BoundsSlice3AlenU] = typecheck.LookupRuntimeVar("panicExtendSlice3AlenU")
196 ExtendCheckFunc[ssa.BoundsSlice3Acap] = typecheck.LookupRuntimeVar("panicExtendSlice3Acap")
197 ExtendCheckFunc[ssa.BoundsSlice3AcapU] = typecheck.LookupRuntimeVar("panicExtendSlice3AcapU")
198 ExtendCheckFunc[ssa.BoundsSlice3B] = typecheck.LookupRuntimeVar("panicExtendSlice3B")
199 ExtendCheckFunc[ssa.BoundsSlice3BU] = typecheck.LookupRuntimeVar("panicExtendSlice3BU")
200 ExtendCheckFunc[ssa.BoundsSlice3C] = typecheck.LookupRuntimeVar("panicExtendSlice3C")
201 ExtendCheckFunc[ssa.BoundsSlice3CU] = typecheck.LookupRuntimeVar("panicExtendSlice3CU")
204 // Wasm (all asm funcs with special ABIs)
205 ir.Syms.WasmDiv = typecheck.LookupRuntimeVar("wasmDiv")
206 ir.Syms.WasmTruncS = typecheck.LookupRuntimeVar("wasmTruncS")
207 ir.Syms.WasmTruncU = typecheck.LookupRuntimeVar("wasmTruncU")
208 ir.Syms.SigPanic = typecheck.LookupRuntimeFunc("sigpanic")
211 // AbiForBodylessFuncStackMap returns the ABI for a bodyless function's stack map.
212 // This is not necessarily the ABI used to call it.
213 // Currently (1.17 dev) such a stack map is always ABI0;
214 // any ABI wrapper that is present is nosplit, hence a precise
215 // stack map is not needed there (the parameters survive only long
216 // enough to call the wrapped assembly function).
217 // This always returns a freshly copied ABI.
218 func AbiForBodylessFuncStackMap(fn *ir.Func) *abi.ABIConfig {
219 return ssaConfig.ABI0.Copy() // No idea what races will result, be safe
222 // abiForFunc implements ABI policy for a function, but does not return a copy of the ABI.
223 // Passing a nil function returns the default ABI based on experiment configuration.
224 func abiForFunc(fn *ir.Func, abi0, abi1 *abi.ABIConfig) *abi.ABIConfig {
225 if buildcfg.Experiment.RegabiArgs {
226 // Select the ABI based on the function's defining ABI.
233 case obj.ABIInternal:
234 // TODO(austin): Clean up the nomenclature here.
235 // It's not clear that "abi1" is ABIInternal.
238 base.Fatalf("function %v has unknown ABI %v", fn, fn.ABI)
239 panic("not reachable")
244 if fn.Pragma&ir.RegisterParams != 0 { // TODO(register args) remove after register abi is working
251 // dvarint writes a varint v to the funcdata in symbol x and returns the new offset.
252 func dvarint(x *obj.LSym, off int, v int64) int {
253 if v < 0 || v > 1e9 {
254 panic(fmt.Sprintf("dvarint: bad offset for funcdata - %v", v))
257 return objw.Uint8(x, off, uint8(v))
259 off = objw.Uint8(x, off, uint8((v&127)|128))
261 return objw.Uint8(x, off, uint8(v>>7))
263 off = objw.Uint8(x, off, uint8(((v>>7)&127)|128))
265 return objw.Uint8(x, off, uint8(v>>14))
267 off = objw.Uint8(x, off, uint8(((v>>14)&127)|128))
269 return objw.Uint8(x, off, uint8(v>>21))
271 off = objw.Uint8(x, off, uint8(((v>>21)&127)|128))
272 return objw.Uint8(x, off, uint8(v>>28))
275 // emitOpenDeferInfo emits FUNCDATA information about the defers in a function
276 // that is using open-coded defers. This funcdata is used to determine the active
277 // defers in a function and execute those defers during panic processing.
279 // The funcdata is all encoded in varints (since values will almost always be less than
280 // 128, but stack offsets could potentially be up to 2Gbyte). All "locations" (offsets)
281 // for stack variables are specified as the number of bytes below varp (pointer to the
282 // top of the local variables) for their starting address. The format is:
284 // - Offset of the deferBits variable
285 // - Number of defers in the function
286 // - Information about each defer call, in reverse order of appearance in the function:
287 // - Offset of the closure value to call
288 func (s *state) emitOpenDeferInfo() {
289 x := base.Ctxt.Lookup(s.curfn.LSym.Name + ".opendefer")
290 x.Set(obj.AttrContentAddressable, true)
291 s.curfn.LSym.Func().OpenCodedDeferInfo = x
293 off = dvarint(x, off, -s.deferBitsTemp.FrameOffset())
294 off = dvarint(x, off, int64(len(s.openDefers)))
296 // Write in reverse-order, for ease of running in that order at runtime
297 for i := len(s.openDefers) - 1; i >= 0; i-- {
299 off = dvarint(x, off, -r.closureNode.FrameOffset())
303 func okOffset(offset int64) int64 {
304 if offset == types.BOGUS_FUNARG_OFFSET {
305 panic(fmt.Errorf("Bogus offset %d", offset))
310 // buildssa builds an SSA function for fn.
311 // worker indicates which of the backend workers is doing the processing.
312 func buildssa(fn *ir.Func, worker int) *ssa.Func {
313 name := ir.FuncName(fn)
315 if ssaDump != "" { // match either a simple name e.g. "(*Reader).Reset", package.name e.g. "compress/gzip.(*Reader).Reset", or subpackage name "gzip.(*Reader).Reset"
316 pkgDotName := base.Ctxt.Pkgpath + "." + name
317 printssa = name == ssaDump ||
318 strings.HasSuffix(pkgDotName, ssaDump) && (pkgDotName == ssaDump || strings.HasSuffix(pkgDotName, "/"+ssaDump))
320 var astBuf *bytes.Buffer
322 astBuf = &bytes.Buffer{}
323 ir.FDumpList(astBuf, "buildssa-enter", fn.Enter)
324 ir.FDumpList(astBuf, "buildssa-body", fn.Body)
325 ir.FDumpList(astBuf, "buildssa-exit", fn.Exit)
327 fmt.Println("generating SSA for", name)
328 fmt.Print(astBuf.String())
336 s.hasdefer = fn.HasDefer()
337 if fn.Pragma&ir.CgoUnsafeArgs != 0 {
338 s.cgoUnsafeArgs = true
340 s.checkPtrEnabled = ir.ShouldCheckPtr(fn, 1)
344 log: printssa && ssaDumpStdout,
348 s.f = ssa.NewFunc(&fe)
351 s.f.Config = ssaConfig
352 s.f.Cache = &ssaCaches[worker]
355 s.f.PrintOrHtmlSSA = printssa
356 if fn.Pragma&ir.Nosplit != 0 {
359 s.f.ABI0 = ssaConfig.ABI0.Copy() // Make a copy to avoid racy map operations in type-register-width cache.
360 s.f.ABI1 = ssaConfig.ABI1.Copy()
361 s.f.ABIDefault = abiForFunc(nil, s.f.ABI0, s.f.ABI1)
362 s.f.ABISelf = abiForFunc(fn, s.f.ABI0, s.f.ABI1)
364 s.panics = map[funcLine]*ssa.Block{}
365 s.softFloat = s.config.SoftFloat
367 // Allocate starting block
368 s.f.Entry = s.f.NewBlock(ssa.BlockPlain)
369 s.f.Entry.Pos = fn.Pos()
374 ssaDF = filepath.Join(ssaDir, base.Ctxt.Pkgpath+"."+name+".html")
375 ssaD := filepath.Dir(ssaDF)
376 os.MkdirAll(ssaD, 0755)
378 s.f.HTMLWriter = ssa.NewHTMLWriter(ssaDF, s.f, ssaDumpCFG)
379 // TODO: generate and print a mapping from nodes to values and blocks
380 dumpSourcesColumn(s.f.HTMLWriter, fn)
381 s.f.HTMLWriter.WriteAST("AST", astBuf)
384 // Allocate starting values
385 s.labels = map[string]*ssaLabel{}
386 s.fwdVars = map[ir.Node]*ssa.Value{}
387 s.startmem = s.entryNewValue0(ssa.OpInitMem, types.TypeMem)
389 s.hasOpenDefers = base.Flag.N == 0 && s.hasdefer && !s.curfn.OpenCodedDeferDisallowed()
391 case base.Debug.NoOpenDefer != 0:
392 s.hasOpenDefers = false
393 case s.hasOpenDefers && (base.Ctxt.Flag_shared || base.Ctxt.Flag_dynlink) && base.Ctxt.Arch.Name == "386":
394 // Don't support open-coded defers for 386 ONLY when using shared
395 // libraries, because there is extra code (added by rewriteToUseGot())
396 // preceding the deferreturn/ret code that we don't track correctly.
397 s.hasOpenDefers = false
399 if s.hasOpenDefers && len(s.curfn.Exit) > 0 {
400 // Skip doing open defers if there is any extra exit code (likely
401 // race detection), since we will not generate that code in the
402 // case of the extra deferreturn/ret segment.
403 s.hasOpenDefers = false
406 // Similarly, skip if there are any heap-allocated result
407 // parameters that need to be copied back to their stack slots.
408 for _, f := range s.curfn.Type().Results().FieldSlice() {
409 if !f.Nname.(*ir.Name).OnStack() {
410 s.hasOpenDefers = false
415 if s.hasOpenDefers &&
416 s.curfn.NumReturns*s.curfn.NumDefers > 15 {
417 // Since we are generating defer calls at every exit for
418 // open-coded defers, skip doing open-coded defers if there are
419 // too many returns (especially if there are multiple defers).
420 // Open-coded defers are most important for improving performance
421 // for smaller functions (which don't have many returns).
422 s.hasOpenDefers = false
425 s.sp = s.entryNewValue0(ssa.OpSP, types.Types[types.TUINTPTR]) // TODO: use generic pointer type (unsafe.Pointer?) instead
426 s.sb = s.entryNewValue0(ssa.OpSB, types.Types[types.TUINTPTR])
428 s.startBlock(s.f.Entry)
429 s.vars[memVar] = s.startmem
431 // Create the deferBits variable and stack slot. deferBits is a
432 // bitmask showing which of the open-coded defers in this function
433 // have been activated.
434 deferBitsTemp := typecheck.TempAt(src.NoXPos, s.curfn, types.Types[types.TUINT8])
435 deferBitsTemp.SetAddrtaken(true)
436 s.deferBitsTemp = deferBitsTemp
437 // For this value, AuxInt is initialized to zero by default
438 startDeferBits := s.entryNewValue0(ssa.OpConst8, types.Types[types.TUINT8])
439 s.vars[deferBitsVar] = startDeferBits
440 s.deferBitsAddr = s.addr(deferBitsTemp)
441 s.store(types.Types[types.TUINT8], s.deferBitsAddr, startDeferBits)
442 // Make sure that the deferBits stack slot is kept alive (for use
443 // by panics) and stores to deferBits are not eliminated, even if
444 // all checking code on deferBits in the function exit can be
445 // eliminated, because the defer statements were all
447 s.vars[memVar] = s.newValue1Apos(ssa.OpVarLive, types.TypeMem, deferBitsTemp, s.mem(), false)
450 var params *abi.ABIParamResultInfo
451 params = s.f.ABISelf.ABIAnalyze(fn.Type(), true)
453 // The backend's stackframe pass prunes away entries from the fn's
454 // Dcl list, including PARAMOUT nodes that correspond to output
455 // params passed in registers. Walk the Dcl list and capture these
456 // nodes to a side list, so that we'll have them available during
457 // DWARF-gen later on. See issue 48573 for more details.
458 var debugInfo ssa.FuncDebug
459 for _, n := range fn.Dcl {
460 if n.Class == ir.PPARAMOUT && n.IsOutputParamInRegisters() {
461 debugInfo.RegOutputParams = append(debugInfo.RegOutputParams, n)
464 fn.DebugInfo = &debugInfo
466 // Generate addresses of local declarations
467 s.decladdrs = map[*ir.Name]*ssa.Value{}
468 for _, n := range fn.Dcl {
471 // Be aware that blank and unnamed input parameters will not appear here, but do appear in the type
472 s.decladdrs[n] = s.entryNewValue2A(ssa.OpLocalAddr, types.NewPtr(n.Type()), n, s.sp, s.startmem)
474 s.decladdrs[n] = s.entryNewValue2A(ssa.OpLocalAddr, types.NewPtr(n.Type()), n, s.sp, s.startmem)
476 // processed at each use, to prevent Addr coming
479 s.Fatalf("local variable with class %v unimplemented", n.Class)
483 s.f.OwnAux = ssa.OwnAuxCall(fn.LSym, params)
485 // Populate SSAable arguments.
486 for _, n := range fn.Dcl {
487 if n.Class == ir.PPARAM {
489 v := s.newValue0A(ssa.OpArg, n.Type(), n)
491 s.addNamedValue(n, v) // This helps with debugging information, not needed for compilation itself.
492 } else { // address was taken AND/OR too large for SSA
493 paramAssignment := ssa.ParamAssignmentForArgName(s.f, n)
494 if len(paramAssignment.Registers) > 0 {
495 if TypeOK(n.Type()) { // SSA-able type, so address was taken -- receive value in OpArg, DO NOT bind to var, store immediately to memory.
496 v := s.newValue0A(ssa.OpArg, n.Type(), n)
497 s.store(n.Type(), s.decladdrs[n], v)
498 } else { // Too big for SSA.
499 // Brute force, and early, do a bunch of stores from registers
500 // TODO fix the nasty storeArgOrLoad recursion in ssa/expand_calls.go so this Just Works with store of a big Arg.
501 s.storeParameterRegsToStack(s.f.ABISelf, paramAssignment, n, s.decladdrs[n], false)
508 // Populate closure variables.
510 clo := s.entryNewValue0(ssa.OpGetClosurePtr, s.f.Config.Types.BytePtr)
511 offset := int64(types.PtrSize) // PtrSize to skip past function entry PC field
512 for _, n := range fn.ClosureVars {
515 typ = types.NewPtr(typ)
518 offset = types.RoundUp(offset, typ.Alignment())
519 ptr := s.newValue1I(ssa.OpOffPtr, types.NewPtr(typ), offset, clo)
522 // If n is a small variable captured by value, promote
523 // it to PAUTO so it can be converted to SSA.
525 // Note: While we never capture a variable by value if
526 // the user took its address, we may have generated
527 // runtime calls that did (#43701). Since we don't
528 // convert Addrtaken variables to SSA anyway, no point
529 // in promoting them either.
530 if n.Byval() && !n.Addrtaken() && TypeOK(n.Type()) {
532 fn.Dcl = append(fn.Dcl, n)
533 s.assign(n, s.load(n.Type(), ptr), false, 0)
538 ptr = s.load(typ, ptr)
540 s.setHeapaddr(fn.Pos(), n, ptr)
544 // Convert the AST-based IR to the SSA-based IR
550 // fallthrough to exit
551 if s.curBlock != nil {
552 s.pushLine(fn.Endlineno)
557 for _, b := range s.f.Blocks {
558 if b.Pos != src.NoXPos {
559 s.updateUnsetPredPos(b)
563 s.f.HTMLWriter.WritePhase("before insert phis", "before insert phis")
567 // Main call to ssa package to compile function
571 s.emitOpenDeferInfo()
574 // Record incoming parameter spill information for morestack calls emitted in the assembler.
575 // This is done here, using all the parameters (used, partially used, and unused) because
576 // it mimics the behavior of the former ABI (everything stored) and because it's not 100%
577 // clear if naming conventions are respected in autogenerated code.
578 // TODO figure out exactly what's unused, don't spill it. Make liveness fine-grained, also.
579 for _, p := range params.InParams() {
580 typs, offs := p.RegisterTypesAndOffsets()
581 for i, t := range typs {
582 o := offs[i] // offset within parameter
583 fo := p.FrameOffset(params) // offset of parameter in frame
584 reg := ssa.ObjRegForAbiReg(p.Registers[i], s.f.Config)
585 s.f.RegArgs = append(s.f.RegArgs, ssa.Spill{Reg: reg, Offset: fo + o, Type: t})
592 func (s *state) storeParameterRegsToStack(abi *abi.ABIConfig, paramAssignment *abi.ABIParamAssignment, n *ir.Name, addr *ssa.Value, pointersOnly bool) {
593 typs, offs := paramAssignment.RegisterTypesAndOffsets()
594 for i, t := range typs {
595 if pointersOnly && !t.IsPtrShaped() {
598 r := paramAssignment.Registers[i]
600 op, reg := ssa.ArgOpAndRegisterFor(r, abi)
601 aux := &ssa.AuxNameOffset{Name: n, Offset: o}
602 v := s.newValue0I(op, t, reg)
604 p := s.newValue1I(ssa.OpOffPtr, types.NewPtr(t), o, addr)
609 // zeroResults zeros the return values at the start of the function.
610 // We need to do this very early in the function. Defer might stop a
611 // panic and show the return values as they exist at the time of
612 // panic. For precise stacks, the garbage collector assumes results
613 // are always live, so we need to zero them before any allocations,
614 // even allocations to move params/results to the heap.
615 func (s *state) zeroResults() {
616 for _, f := range s.curfn.Type().Results().FieldSlice() {
617 n := f.Nname.(*ir.Name)
619 // The local which points to the return value is the
620 // thing that needs zeroing. This is already handled
621 // by a Needzero annotation in plive.go:(*liveness).epilogue.
624 // Zero the stack location containing f.
625 if typ := n.Type(); TypeOK(typ) {
626 s.assign(n, s.zeroVal(typ), false, 0)
628 if typ.HasPointers() {
629 s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, n, s.mem())
631 s.zero(n.Type(), s.decladdrs[n])
636 // paramsToHeap produces code to allocate memory for heap-escaped parameters
637 // and to copy non-result parameters' values from the stack.
638 func (s *state) paramsToHeap() {
639 do := func(params *types.Type) {
640 for _, f := range params.FieldSlice() {
642 continue // anonymous or blank parameter
644 n := f.Nname.(*ir.Name)
645 if ir.IsBlank(n) || n.OnStack() {
649 if n.Class == ir.PPARAM {
650 s.move(n.Type(), s.expr(n.Heapaddr), s.decladdrs[n])
655 typ := s.curfn.Type()
661 // newHeapaddr allocates heap memory for n and sets its heap address.
662 func (s *state) newHeapaddr(n *ir.Name) {
663 s.setHeapaddr(n.Pos(), n, s.newObject(n.Type(), nil))
666 // setHeapaddr allocates a new PAUTO variable to store ptr (which must be non-nil)
667 // and then sets it as n's heap address.
668 func (s *state) setHeapaddr(pos src.XPos, n *ir.Name, ptr *ssa.Value) {
669 if !ptr.Type.IsPtr() || !types.Identical(n.Type(), ptr.Type.Elem()) {
670 base.FatalfAt(n.Pos(), "setHeapaddr %L with type %v", n, ptr.Type)
673 // Declare variable to hold address.
674 addr := ir.NewNameAt(pos, &types.Sym{Name: "&" + n.Sym().Name, Pkg: types.LocalPkg})
675 addr.SetType(types.NewPtr(n.Type()))
676 addr.Class = ir.PAUTO
679 s.curfn.Dcl = append(s.curfn.Dcl, addr)
680 types.CalcSize(addr.Type())
682 if n.Class == ir.PPARAMOUT {
683 addr.SetIsOutputParamHeapAddr(true)
687 s.assign(addr, ptr, false, 0)
690 // newObject returns an SSA value denoting new(typ).
691 func (s *state) newObject(typ *types.Type, rtype *ssa.Value) *ssa.Value {
693 return s.newValue1A(ssa.OpAddr, types.NewPtr(typ), ir.Syms.Zerobase, s.sb)
696 rtype = s.reflectType(typ)
698 return s.rtcall(ir.Syms.Newobject, true, []*types.Type{types.NewPtr(typ)}, rtype)[0]
701 func (s *state) checkPtrAlignment(n *ir.ConvExpr, v *ssa.Value, count *ssa.Value) {
702 if !n.Type().IsPtr() {
703 s.Fatalf("expected pointer type: %v", n.Type())
705 elem, rtypeExpr := n.Type().Elem(), n.ElemRType
708 s.Fatalf("expected array type: %v", elem)
710 elem, rtypeExpr = elem.Elem(), n.ElemElemRType
713 // Casting from larger type to smaller one is ok, so for smallest type, do nothing.
714 if elem.Alignment() == 1 && (size == 0 || size == 1 || count == nil) {
718 count = s.constInt(types.Types[types.TUINTPTR], 1)
720 if count.Type.Size() != s.config.PtrSize {
721 s.Fatalf("expected count fit to a uintptr size, have: %d, want: %d", count.Type.Size(), s.config.PtrSize)
724 if rtypeExpr != nil {
725 rtype = s.expr(rtypeExpr)
727 rtype = s.reflectType(elem)
729 s.rtcall(ir.Syms.CheckPtrAlignment, true, nil, v, rtype, count)
732 // reflectType returns an SSA value representing a pointer to typ's
733 // reflection type descriptor.
734 func (s *state) reflectType(typ *types.Type) *ssa.Value {
735 // TODO(mdempsky): Make this Fatalf under Unified IR; frontend needs
736 // to supply RType expressions.
737 lsym := reflectdata.TypeLinksym(typ)
738 return s.entryNewValue1A(ssa.OpAddr, types.NewPtr(types.Types[types.TUINT8]), lsym, s.sb)
741 func dumpSourcesColumn(writer *ssa.HTMLWriter, fn *ir.Func) {
742 // Read sources of target function fn.
743 fname := base.Ctxt.PosTable.Pos(fn.Pos()).Filename()
744 targetFn, err := readFuncLines(fname, fn.Pos().Line(), fn.Endlineno.Line())
746 writer.Logf("cannot read sources for function %v: %v", fn, err)
749 // Read sources of inlined functions.
750 var inlFns []*ssa.FuncLines
751 for _, fi := range ssaDumpInlined {
753 fname := base.Ctxt.PosTable.Pos(fi.Pos()).Filename()
754 fnLines, err := readFuncLines(fname, fi.Pos().Line(), elno.Line())
756 writer.Logf("cannot read sources for inlined function %v: %v", fi, err)
759 inlFns = append(inlFns, fnLines)
762 sort.Sort(ssa.ByTopo(inlFns))
764 inlFns = append([]*ssa.FuncLines{targetFn}, inlFns...)
767 writer.WriteSources("sources", inlFns)
770 func readFuncLines(file string, start, end uint) (*ssa.FuncLines, error) {
771 f, err := os.Open(os.ExpandEnv(file))
778 scanner := bufio.NewScanner(f)
779 for scanner.Scan() && ln <= end {
781 lines = append(lines, scanner.Text())
785 return &ssa.FuncLines{Filename: file, StartLineno: start, Lines: lines}, nil
788 // updateUnsetPredPos propagates the earliest-value position information for b
789 // towards all of b's predecessors that need a position, and recurs on that
790 // predecessor if its position is updated. B should have a non-empty position.
791 func (s *state) updateUnsetPredPos(b *ssa.Block) {
792 if b.Pos == src.NoXPos {
793 s.Fatalf("Block %s should have a position", b)
795 bestPos := src.NoXPos
796 for _, e := range b.Preds {
801 if bestPos == src.NoXPos {
803 for _, v := range b.Values {
807 if v.Pos != src.NoXPos {
808 // Assume values are still in roughly textual order;
809 // TODO: could also seek minimum position?
816 s.updateUnsetPredPos(p) // We do not expect long chains of these, thus recursion is okay.
820 // Information about each open-coded defer.
821 type openDeferInfo struct {
822 // The node representing the call of the defer
824 // If defer call is closure call, the address of the argtmp where the
825 // closure is stored.
827 // The node representing the argtmp where the closure is stored - used for
828 // function, method, or interface call, to store a closure that panic
829 // processing can use for this defer.
834 // configuration (arch) information
837 // function we're building
844 labels map[string]*ssaLabel
846 // unlabeled break and continue statement tracking
847 breakTo *ssa.Block // current target for plain break statement
848 continueTo *ssa.Block // current target for plain continue statement
850 // current location where we're interpreting the AST
853 // variable assignments in the current block (map from variable symbol to ssa value)
854 // *Node is the unique identifier (an ONAME Node) for the variable.
855 // TODO: keep a single varnum map, then make all of these maps slices instead?
856 vars map[ir.Node]*ssa.Value
858 // fwdVars are variables that are used before they are defined in the current block.
859 // This map exists just to coalesce multiple references into a single FwdRef op.
860 // *Node is the unique identifier (an ONAME Node) for the variable.
861 fwdVars map[ir.Node]*ssa.Value
863 // all defined variables at the end of each block. Indexed by block ID.
864 defvars []map[ir.Node]*ssa.Value
866 // addresses of PPARAM and PPARAMOUT variables on the stack.
867 decladdrs map[*ir.Name]*ssa.Value
869 // starting values. Memory, stack pointer, and globals pointer
873 // value representing address of where deferBits autotmp is stored
874 deferBitsAddr *ssa.Value
875 deferBitsTemp *ir.Name
877 // line number stack. The current line number is top of stack
879 // the last line number processed; it may have been popped
882 // list of panic calls by function name and line number.
883 // Used to deduplicate panic calls.
884 panics map[funcLine]*ssa.Block
887 hasdefer bool // whether the function contains a defer statement
889 hasOpenDefers bool // whether we are doing open-coded defers
890 checkPtrEnabled bool // whether to insert checkptr instrumentation
892 // If doing open-coded defers, list of info about the defer calls in
893 // scanning order. Hence, at exit we should run these defers in reverse
894 // order of this list
895 openDefers []*openDeferInfo
896 // For open-coded defers, this is the beginning and end blocks of the last
897 // defer exit code that we have generated so far. We use these to share
898 // code between exits if the shareDeferExits option (disabled by default)
900 lastDeferExit *ssa.Block // Entry block of last defer exit code we generated
901 lastDeferFinalBlock *ssa.Block // Final block of last defer exit code we generated
902 lastDeferCount int // Number of defers encountered at that point
904 prevCall *ssa.Value // the previous call; use this to tie results to the call op.
907 type funcLine struct {
913 type ssaLabel struct {
914 target *ssa.Block // block identified by this label
915 breakTarget *ssa.Block // block to break to in control flow node identified by this label
916 continueTarget *ssa.Block // block to continue to in control flow node identified by this label
919 // label returns the label associated with sym, creating it if necessary.
920 func (s *state) label(sym *types.Sym) *ssaLabel {
921 lab := s.labels[sym.Name]
924 s.labels[sym.Name] = lab
929 func (s *state) Logf(msg string, args ...interface{}) { s.f.Logf(msg, args...) }
930 func (s *state) Log() bool { return s.f.Log() }
931 func (s *state) Fatalf(msg string, args ...interface{}) {
932 s.f.Frontend().Fatalf(s.peekPos(), msg, args...)
934 func (s *state) Warnl(pos src.XPos, msg string, args ...interface{}) { s.f.Warnl(pos, msg, args...) }
935 func (s *state) Debug_checknil() bool { return s.f.Frontend().Debug_checknil() }
937 func ssaMarker(name string) *ir.Name {
938 return typecheck.NewName(&types.Sym{Name: name})
942 // marker node for the memory variable
943 memVar = ssaMarker("mem")
945 // marker nodes for temporary variables
946 ptrVar = ssaMarker("ptr")
947 lenVar = ssaMarker("len")
948 capVar = ssaMarker("cap")
949 typVar = ssaMarker("typ")
950 okVar = ssaMarker("ok")
951 deferBitsVar = ssaMarker("deferBits")
952 ternaryVar = ssaMarker("ternary")
955 // startBlock sets the current block we're generating code in to b.
956 func (s *state) startBlock(b *ssa.Block) {
957 if s.curBlock != nil {
958 s.Fatalf("starting block %v when block %v has not ended", b, s.curBlock)
961 s.vars = map[ir.Node]*ssa.Value{}
962 for n := range s.fwdVars {
967 // endBlock marks the end of generating code for the current block.
968 // Returns the (former) current block. Returns nil if there is no current
969 // block, i.e. if no code flows to the current execution point.
970 func (s *state) endBlock() *ssa.Block {
975 for len(s.defvars) <= int(b.ID) {
976 s.defvars = append(s.defvars, nil)
978 s.defvars[b.ID] = s.vars
982 // Empty plain blocks get the line of their successor (handled after all blocks created),
983 // except for increment blocks in For statements (handled in ssa conversion of OFOR),
984 // and for blocks ending in GOTO/BREAK/CONTINUE.
992 // pushLine pushes a line number on the line number stack.
993 func (s *state) pushLine(line src.XPos) {
995 // the frontend may emit node with line number missing,
996 // use the parent line number in this case.
998 if base.Flag.K != 0 {
999 base.Warn("buildssa: unknown position (line 0)")
1005 s.line = append(s.line, line)
1008 // popLine pops the top of the line number stack.
1009 func (s *state) popLine() {
1010 s.line = s.line[:len(s.line)-1]
1013 // peekPos peeks the top of the line number stack.
1014 func (s *state) peekPos() src.XPos {
1015 return s.line[len(s.line)-1]
1018 // newValue0 adds a new value with no arguments to the current block.
1019 func (s *state) newValue0(op ssa.Op, t *types.Type) *ssa.Value {
1020 return s.curBlock.NewValue0(s.peekPos(), op, t)
1023 // newValue0A adds a new value with no arguments and an aux value to the current block.
1024 func (s *state) newValue0A(op ssa.Op, t *types.Type, aux ssa.Aux) *ssa.Value {
1025 return s.curBlock.NewValue0A(s.peekPos(), op, t, aux)
1028 // newValue0I adds a new value with no arguments and an auxint value to the current block.
1029 func (s *state) newValue0I(op ssa.Op, t *types.Type, auxint int64) *ssa.Value {
1030 return s.curBlock.NewValue0I(s.peekPos(), op, t, auxint)
1033 // newValue1 adds a new value with one argument to the current block.
1034 func (s *state) newValue1(op ssa.Op, t *types.Type, arg *ssa.Value) *ssa.Value {
1035 return s.curBlock.NewValue1(s.peekPos(), op, t, arg)
1038 // newValue1A adds a new value with one argument and an aux value to the current block.
1039 func (s *state) newValue1A(op ssa.Op, t *types.Type, aux ssa.Aux, arg *ssa.Value) *ssa.Value {
1040 return s.curBlock.NewValue1A(s.peekPos(), op, t, aux, arg)
1043 // newValue1Apos adds a new value with one argument and an aux value to the current block.
1044 // isStmt determines whether the created values may be a statement or not
1045 // (i.e., false means never, yes means maybe).
1046 func (s *state) newValue1Apos(op ssa.Op, t *types.Type, aux ssa.Aux, arg *ssa.Value, isStmt bool) *ssa.Value {
1048 return s.curBlock.NewValue1A(s.peekPos(), op, t, aux, arg)
1050 return s.curBlock.NewValue1A(s.peekPos().WithNotStmt(), op, t, aux, arg)
1053 // newValue1I adds a new value with one argument and an auxint value to the current block.
1054 func (s *state) newValue1I(op ssa.Op, t *types.Type, aux int64, arg *ssa.Value) *ssa.Value {
1055 return s.curBlock.NewValue1I(s.peekPos(), op, t, aux, arg)
1058 // newValue2 adds a new value with two arguments to the current block.
1059 func (s *state) newValue2(op ssa.Op, t *types.Type, arg0, arg1 *ssa.Value) *ssa.Value {
1060 return s.curBlock.NewValue2(s.peekPos(), op, t, arg0, arg1)
1063 // newValue2A adds a new value with two arguments and an aux value to the current block.
1064 func (s *state) newValue2A(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1 *ssa.Value) *ssa.Value {
1065 return s.curBlock.NewValue2A(s.peekPos(), op, t, aux, arg0, arg1)
1068 // newValue2Apos adds a new value with two arguments and an aux value to the current block.
1069 // isStmt determines whether the created values may be a statement or not
1070 // (i.e., false means never, yes means maybe).
1071 func (s *state) newValue2Apos(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1 *ssa.Value, isStmt bool) *ssa.Value {
1073 return s.curBlock.NewValue2A(s.peekPos(), op, t, aux, arg0, arg1)
1075 return s.curBlock.NewValue2A(s.peekPos().WithNotStmt(), op, t, aux, arg0, arg1)
1078 // newValue2I adds a new value with two arguments and an auxint value to the current block.
1079 func (s *state) newValue2I(op ssa.Op, t *types.Type, aux int64, arg0, arg1 *ssa.Value) *ssa.Value {
1080 return s.curBlock.NewValue2I(s.peekPos(), op, t, aux, arg0, arg1)
1083 // newValue3 adds a new value with three arguments to the current block.
1084 func (s *state) newValue3(op ssa.Op, t *types.Type, arg0, arg1, arg2 *ssa.Value) *ssa.Value {
1085 return s.curBlock.NewValue3(s.peekPos(), op, t, arg0, arg1, arg2)
1088 // newValue3I adds a new value with three arguments and an auxint value to the current block.
1089 func (s *state) newValue3I(op ssa.Op, t *types.Type, aux int64, arg0, arg1, arg2 *ssa.Value) *ssa.Value {
1090 return s.curBlock.NewValue3I(s.peekPos(), op, t, aux, arg0, arg1, arg2)
1093 // newValue3A adds a new value with three arguments and an aux value to the current block.
1094 func (s *state) newValue3A(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1, arg2 *ssa.Value) *ssa.Value {
1095 return s.curBlock.NewValue3A(s.peekPos(), op, t, aux, arg0, arg1, arg2)
1098 // newValue3Apos adds a new value with three arguments and an aux value to the current block.
1099 // isStmt determines whether the created values may be a statement or not
1100 // (i.e., false means never, yes means maybe).
1101 func (s *state) newValue3Apos(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1, arg2 *ssa.Value, isStmt bool) *ssa.Value {
1103 return s.curBlock.NewValue3A(s.peekPos(), op, t, aux, arg0, arg1, arg2)
1105 return s.curBlock.NewValue3A(s.peekPos().WithNotStmt(), op, t, aux, arg0, arg1, arg2)
1108 // newValue4 adds a new value with four arguments to the current block.
1109 func (s *state) newValue4(op ssa.Op, t *types.Type, arg0, arg1, arg2, arg3 *ssa.Value) *ssa.Value {
1110 return s.curBlock.NewValue4(s.peekPos(), op, t, arg0, arg1, arg2, arg3)
1113 // newValue4I adds a new value with four arguments and an auxint value to the current block.
1114 func (s *state) newValue4I(op ssa.Op, t *types.Type, aux int64, arg0, arg1, arg2, arg3 *ssa.Value) *ssa.Value {
1115 return s.curBlock.NewValue4I(s.peekPos(), op, t, aux, arg0, arg1, arg2, arg3)
1118 func (s *state) entryBlock() *ssa.Block {
1120 if base.Flag.N > 0 && s.curBlock != nil {
1121 // If optimizations are off, allocate in current block instead. Since with -N
1122 // we're not doing the CSE or tighten passes, putting lots of stuff in the
1123 // entry block leads to O(n^2) entries in the live value map during regalloc.
1130 // entryNewValue0 adds a new value with no arguments to the entry block.
1131 func (s *state) entryNewValue0(op ssa.Op, t *types.Type) *ssa.Value {
1132 return s.entryBlock().NewValue0(src.NoXPos, op, t)
1135 // entryNewValue0A adds a new value with no arguments and an aux value to the entry block.
1136 func (s *state) entryNewValue0A(op ssa.Op, t *types.Type, aux ssa.Aux) *ssa.Value {
1137 return s.entryBlock().NewValue0A(src.NoXPos, op, t, aux)
1140 // entryNewValue1 adds a new value with one argument to the entry block.
1141 func (s *state) entryNewValue1(op ssa.Op, t *types.Type, arg *ssa.Value) *ssa.Value {
1142 return s.entryBlock().NewValue1(src.NoXPos, op, t, arg)
1145 // entryNewValue1I adds a new value with one argument and an auxint value to the entry block.
1146 func (s *state) entryNewValue1I(op ssa.Op, t *types.Type, auxint int64, arg *ssa.Value) *ssa.Value {
1147 return s.entryBlock().NewValue1I(src.NoXPos, op, t, auxint, arg)
1150 // entryNewValue1A adds a new value with one argument and an aux value to the entry block.
1151 func (s *state) entryNewValue1A(op ssa.Op, t *types.Type, aux ssa.Aux, arg *ssa.Value) *ssa.Value {
1152 return s.entryBlock().NewValue1A(src.NoXPos, op, t, aux, arg)
1155 // entryNewValue2 adds a new value with two arguments to the entry block.
1156 func (s *state) entryNewValue2(op ssa.Op, t *types.Type, arg0, arg1 *ssa.Value) *ssa.Value {
1157 return s.entryBlock().NewValue2(src.NoXPos, op, t, arg0, arg1)
1160 // entryNewValue2A adds a new value with two arguments and an aux value to the entry block.
1161 func (s *state) entryNewValue2A(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1 *ssa.Value) *ssa.Value {
1162 return s.entryBlock().NewValue2A(src.NoXPos, op, t, aux, arg0, arg1)
1165 // const* routines add a new const value to the entry block.
1166 func (s *state) constSlice(t *types.Type) *ssa.Value {
1167 return s.f.ConstSlice(t)
1169 func (s *state) constInterface(t *types.Type) *ssa.Value {
1170 return s.f.ConstInterface(t)
1172 func (s *state) constNil(t *types.Type) *ssa.Value { return s.f.ConstNil(t) }
1173 func (s *state) constEmptyString(t *types.Type) *ssa.Value {
1174 return s.f.ConstEmptyString(t)
1176 func (s *state) constBool(c bool) *ssa.Value {
1177 return s.f.ConstBool(types.Types[types.TBOOL], c)
1179 func (s *state) constInt8(t *types.Type, c int8) *ssa.Value {
1180 return s.f.ConstInt8(t, c)
1182 func (s *state) constInt16(t *types.Type, c int16) *ssa.Value {
1183 return s.f.ConstInt16(t, c)
1185 func (s *state) constInt32(t *types.Type, c int32) *ssa.Value {
1186 return s.f.ConstInt32(t, c)
1188 func (s *state) constInt64(t *types.Type, c int64) *ssa.Value {
1189 return s.f.ConstInt64(t, c)
1191 func (s *state) constFloat32(t *types.Type, c float64) *ssa.Value {
1192 return s.f.ConstFloat32(t, c)
1194 func (s *state) constFloat64(t *types.Type, c float64) *ssa.Value {
1195 return s.f.ConstFloat64(t, c)
1197 func (s *state) constInt(t *types.Type, c int64) *ssa.Value {
1198 if s.config.PtrSize == 8 {
1199 return s.constInt64(t, c)
1201 if int64(int32(c)) != c {
1202 s.Fatalf("integer constant too big %d", c)
1204 return s.constInt32(t, int32(c))
1206 func (s *state) constOffPtrSP(t *types.Type, c int64) *ssa.Value {
1207 return s.f.ConstOffPtrSP(t, c, s.sp)
1210 // newValueOrSfCall* are wrappers around newValue*, which may create a call to a
1211 // soft-float runtime function instead (when emitting soft-float code).
1212 func (s *state) newValueOrSfCall1(op ssa.Op, t *types.Type, arg *ssa.Value) *ssa.Value {
1214 if c, ok := s.sfcall(op, arg); ok {
1218 return s.newValue1(op, t, arg)
1220 func (s *state) newValueOrSfCall2(op ssa.Op, t *types.Type, arg0, arg1 *ssa.Value) *ssa.Value {
1222 if c, ok := s.sfcall(op, arg0, arg1); ok {
1226 return s.newValue2(op, t, arg0, arg1)
1229 type instrumentKind uint8
1232 instrumentRead = iota
1237 func (s *state) instrument(t *types.Type, addr *ssa.Value, kind instrumentKind) {
1238 s.instrument2(t, addr, nil, kind)
1241 // instrumentFields instruments a read/write operation on addr.
1242 // If it is instrumenting for MSAN or ASAN and t is a struct type, it instruments
1243 // operation for each field, instead of for the whole struct.
1244 func (s *state) instrumentFields(t *types.Type, addr *ssa.Value, kind instrumentKind) {
1245 if !(base.Flag.MSan || base.Flag.ASan) || !t.IsStruct() {
1246 s.instrument(t, addr, kind)
1249 for _, f := range t.Fields().Slice() {
1250 if f.Sym.IsBlank() {
1253 offptr := s.newValue1I(ssa.OpOffPtr, types.NewPtr(f.Type), f.Offset, addr)
1254 s.instrumentFields(f.Type, offptr, kind)
1258 func (s *state) instrumentMove(t *types.Type, dst, src *ssa.Value) {
1260 s.instrument2(t, dst, src, instrumentMove)
1262 s.instrument(t, src, instrumentRead)
1263 s.instrument(t, dst, instrumentWrite)
1267 func (s *state) instrument2(t *types.Type, addr, addr2 *ssa.Value, kind instrumentKind) {
1268 if !s.curfn.InstrumentBody() {
1274 return // can't race on zero-sized things
1277 if ssa.IsSanitizerSafeAddr(addr) {
1284 if addr2 != nil && kind != instrumentMove {
1285 panic("instrument2: non-nil addr2 for non-move instrumentation")
1290 case instrumentRead:
1291 fn = ir.Syms.Msanread
1292 case instrumentWrite:
1293 fn = ir.Syms.Msanwrite
1294 case instrumentMove:
1295 fn = ir.Syms.Msanmove
1297 panic("unreachable")
1300 } else if base.Flag.Race && t.NumComponents(types.CountBlankFields) > 1 {
1301 // for composite objects we have to write every address
1302 // because a write might happen to any subobject.
1303 // composites with only one element don't have subobjects, though.
1305 case instrumentRead:
1306 fn = ir.Syms.Racereadrange
1307 case instrumentWrite:
1308 fn = ir.Syms.Racewriterange
1310 panic("unreachable")
1313 } else if base.Flag.Race {
1314 // for non-composite objects we can write just the start
1315 // address, as any write must write the first byte.
1317 case instrumentRead:
1318 fn = ir.Syms.Raceread
1319 case instrumentWrite:
1320 fn = ir.Syms.Racewrite
1322 panic("unreachable")
1324 } else if base.Flag.ASan {
1326 case instrumentRead:
1327 fn = ir.Syms.Asanread
1328 case instrumentWrite:
1329 fn = ir.Syms.Asanwrite
1331 panic("unreachable")
1335 panic("unreachable")
1338 args := []*ssa.Value{addr}
1340 args = append(args, addr2)
1343 args = append(args, s.constInt(types.Types[types.TUINTPTR], w))
1345 s.rtcall(fn, true, nil, args...)
1348 func (s *state) load(t *types.Type, src *ssa.Value) *ssa.Value {
1349 s.instrumentFields(t, src, instrumentRead)
1350 return s.rawLoad(t, src)
1353 func (s *state) rawLoad(t *types.Type, src *ssa.Value) *ssa.Value {
1354 return s.newValue2(ssa.OpLoad, t, src, s.mem())
1357 func (s *state) store(t *types.Type, dst, val *ssa.Value) {
1358 s.vars[memVar] = s.newValue3A(ssa.OpStore, types.TypeMem, t, dst, val, s.mem())
1361 func (s *state) zero(t *types.Type, dst *ssa.Value) {
1362 s.instrument(t, dst, instrumentWrite)
1363 store := s.newValue2I(ssa.OpZero, types.TypeMem, t.Size(), dst, s.mem())
1365 s.vars[memVar] = store
1368 func (s *state) move(t *types.Type, dst, src *ssa.Value) {
1369 s.moveWhichMayOverlap(t, dst, src, false)
1371 func (s *state) moveWhichMayOverlap(t *types.Type, dst, src *ssa.Value, mayOverlap bool) {
1372 s.instrumentMove(t, dst, src)
1373 if mayOverlap && t.IsArray() && t.NumElem() > 1 && !ssa.IsInlinableMemmove(dst, src, t.Size(), s.f.Config) {
1374 // Normally, when moving Go values of type T from one location to another,
1375 // we don't need to worry about partial overlaps. The two Ts must either be
1376 // in disjoint (nonoverlapping) memory or in exactly the same location.
1377 // There are 2 cases where this isn't true:
1378 // 1) Using unsafe you can arrange partial overlaps.
1379 // 2) Since Go 1.17, you can use a cast from a slice to a ptr-to-array.
1380 // https://go.dev/ref/spec#Conversions_from_slice_to_array_pointer
1381 // This feature can be used to construct partial overlaps of array types.
1383 // p := (*[2]int)(a[:])
1384 // q := (*[2]int)(a[1:])
1386 // We don't care about solving 1. Or at least, we haven't historically
1387 // and no one has complained.
1388 // For 2, we need to ensure that if there might be partial overlap,
1389 // then we can't use OpMove; we must use memmove instead.
1390 // (memmove handles partial overlap by copying in the correct
1391 // direction. OpMove does not.)
1393 // Note that we have to be careful here not to introduce a call when
1394 // we're marshaling arguments to a call or unmarshaling results from a call.
1395 // Cases where this is happening must pass mayOverlap to false.
1396 // (Currently this only happens when unmarshaling results of a call.)
1397 if t.HasPointers() {
1398 s.rtcall(ir.Syms.Typedmemmove, true, nil, s.reflectType(t), dst, src)
1399 // We would have otherwise implemented this move with straightline code,
1400 // including a write barrier. Pretend we issue a write barrier here,
1401 // so that the write barrier tests work. (Otherwise they'd need to know
1402 // the details of IsInlineableMemmove.)
1403 s.curfn.SetWBPos(s.peekPos())
1405 s.rtcall(ir.Syms.Memmove, true, nil, dst, src, s.constInt(types.Types[types.TUINTPTR], t.Size()))
1407 ssa.LogLargeCopy(s.f.Name, s.peekPos(), t.Size())
1410 store := s.newValue3I(ssa.OpMove, types.TypeMem, t.Size(), dst, src, s.mem())
1412 s.vars[memVar] = store
1415 // stmtList converts the statement list n to SSA and adds it to s.
1416 func (s *state) stmtList(l ir.Nodes) {
1417 for _, n := range l {
1422 // stmt converts the statement n to SSA and adds it to s.
1423 func (s *state) stmt(n ir.Node) {
1427 // If s.curBlock is nil, and n isn't a label (which might have an associated goto somewhere),
1428 // then this code is dead. Stop here.
1429 if s.curBlock == nil && n.Op() != ir.OLABEL {
1433 s.stmtList(n.Init())
1437 n := n.(*ir.BlockStmt)
1441 case ir.ODCLCONST, ir.ODCLTYPE, ir.OFALL:
1443 // Expression statements
1445 n := n.(*ir.CallExpr)
1446 if ir.IsIntrinsicCall(n) {
1453 n := n.(*ir.CallExpr)
1454 s.callResult(n, callNormal)
1455 if n.Op() == ir.OCALLFUNC && n.X.Op() == ir.ONAME && n.X.(*ir.Name).Class == ir.PFUNC {
1456 if fn := n.X.Sym().Name; base.Flag.CompilingRuntime && fn == "throw" ||
1457 n.X.Sym().Pkg == ir.Pkgs.Runtime && (fn == "throwinit" || fn == "gopanic" || fn == "panicwrap" || fn == "block" || fn == "panicmakeslicelen" || fn == "panicmakeslicecap" || fn == "panicunsafeslicelen" || fn == "panicunsafeslicenilptr" || fn == "panicunsafestringlen" || fn == "panicunsafestringnilptr") {
1460 b.Kind = ssa.BlockExit
1462 // TODO: never rewrite OPANIC to OCALLFUNC in the
1463 // first place. Need to wait until all backends
1468 n := n.(*ir.GoDeferStmt)
1469 if base.Debug.Defer > 0 {
1470 var defertype string
1471 if s.hasOpenDefers {
1472 defertype = "open-coded"
1473 } else if n.Esc() == ir.EscNever {
1474 defertype = "stack-allocated"
1476 defertype = "heap-allocated"
1478 base.WarnfAt(n.Pos(), "%s defer", defertype)
1480 if s.hasOpenDefers {
1481 s.openDeferRecord(n.Call.(*ir.CallExpr))
1484 if n.Esc() == ir.EscNever {
1487 s.callResult(n.Call.(*ir.CallExpr), d)
1490 n := n.(*ir.GoDeferStmt)
1491 s.callResult(n.Call.(*ir.CallExpr), callGo)
1493 case ir.OAS2DOTTYPE:
1494 n := n.(*ir.AssignListStmt)
1495 var res, resok *ssa.Value
1496 if n.Rhs[0].Op() == ir.ODOTTYPE2 {
1497 res, resok = s.dottype(n.Rhs[0].(*ir.TypeAssertExpr), true)
1499 res, resok = s.dynamicDottype(n.Rhs[0].(*ir.DynamicTypeAssertExpr), true)
1502 if !TypeOK(n.Rhs[0].Type()) {
1503 if res.Op != ssa.OpLoad {
1504 s.Fatalf("dottype of non-load")
1507 if res.Args[1] != mem {
1508 s.Fatalf("memory no longer live from 2-result dottype load")
1513 s.assign(n.Lhs[0], res, deref, 0)
1514 s.assign(n.Lhs[1], resok, false, 0)
1518 // We come here only when it is an intrinsic call returning two values.
1519 n := n.(*ir.AssignListStmt)
1520 call := n.Rhs[0].(*ir.CallExpr)
1521 if !ir.IsIntrinsicCall(call) {
1522 s.Fatalf("non-intrinsic AS2FUNC not expanded %v", call)
1524 v := s.intrinsicCall(call)
1525 v1 := s.newValue1(ssa.OpSelect0, n.Lhs[0].Type(), v)
1526 v2 := s.newValue1(ssa.OpSelect1, n.Lhs[1].Type(), v)
1527 s.assign(n.Lhs[0], v1, false, 0)
1528 s.assign(n.Lhs[1], v2, false, 0)
1533 if v := n.X; v.Esc() == ir.EscHeap {
1538 n := n.(*ir.LabelStmt)
1541 // Nothing to do because the label isn't targetable. See issue 52278.
1546 // The label might already have a target block via a goto.
1547 if lab.target == nil {
1548 lab.target = s.f.NewBlock(ssa.BlockPlain)
1551 // Go to that label.
1552 // (We pretend "label:" is preceded by "goto label", unless the predecessor is unreachable.)
1553 if s.curBlock != nil {
1555 b.AddEdgeTo(lab.target)
1557 s.startBlock(lab.target)
1560 n := n.(*ir.BranchStmt)
1564 if lab.target == nil {
1565 lab.target = s.f.NewBlock(ssa.BlockPlain)
1569 b.Pos = s.lastPos.WithIsStmt() // Do this even if b is an empty block.
1570 b.AddEdgeTo(lab.target)
1573 n := n.(*ir.AssignStmt)
1574 if n.X == n.Y && n.X.Op() == ir.ONAME {
1575 // An x=x assignment. No point in doing anything
1576 // here. In addition, skipping this assignment
1577 // prevents generating:
1580 // which is bad because x is incorrectly considered
1581 // dead before the vardef. See issue #14904.
1585 // mayOverlap keeps track of whether the LHS and RHS might
1586 // refer to partially overlapping memory. Partial overlapping can
1587 // only happen for arrays, see the comment in moveWhichMayOverlap.
1589 // If both sides of the assignment are not dereferences, then partial
1590 // overlap can't happen. Partial overlap can only occur only when the
1591 // arrays referenced are strictly smaller parts of the same base array.
1592 // If one side of the assignment is a full array, then partial overlap
1593 // can't happen. (The arrays are either disjoint or identical.)
1594 mayOverlap := n.X.Op() == ir.ODEREF && (n.Y != nil && n.Y.Op() == ir.ODEREF)
1595 if n.Y != nil && n.Y.Op() == ir.ODEREF {
1596 p := n.Y.(*ir.StarExpr).X
1597 for p.Op() == ir.OCONVNOP {
1598 p = p.(*ir.ConvExpr).X
1600 if p.Op() == ir.OSPTR && p.(*ir.UnaryExpr).X.Type().IsString() {
1601 // Pointer fields of strings point to unmodifiable memory.
1602 // That memory can't overlap with the memory being written.
1611 case ir.OSTRUCTLIT, ir.OARRAYLIT, ir.OSLICELIT:
1612 // All literals with nonzero fields have already been
1613 // rewritten during walk. Any that remain are just T{}
1614 // or equivalents. Use the zero value.
1615 if !ir.IsZero(rhs) {
1616 s.Fatalf("literal with nonzero value in SSA: %v", rhs)
1620 rhs := rhs.(*ir.CallExpr)
1621 // Check whether we're writing the result of an append back to the same slice.
1622 // If so, we handle it specially to avoid write barriers on the fast
1623 // (non-growth) path.
1624 if !ir.SameSafeExpr(n.X, rhs.Args[0]) || base.Flag.N != 0 {
1627 // If the slice can be SSA'd, it'll be on the stack,
1628 // so there will be no write barriers,
1629 // so there's no need to attempt to prevent them.
1631 if base.Debug.Append > 0 { // replicating old diagnostic message
1632 base.WarnfAt(n.Pos(), "append: len-only update (in local slice)")
1636 if base.Debug.Append > 0 {
1637 base.WarnfAt(n.Pos(), "append: len-only update")
1644 if ir.IsBlank(n.X) {
1646 // Just evaluate rhs for side-effects.
1664 r = nil // Signal assign to use OpZero.
1677 if rhs != nil && (rhs.Op() == ir.OSLICE || rhs.Op() == ir.OSLICE3 || rhs.Op() == ir.OSLICESTR) && ir.SameSafeExpr(rhs.(*ir.SliceExpr).X, n.X) {
1678 // We're assigning a slicing operation back to its source.
1679 // Don't write back fields we aren't changing. See issue #14855.
1680 rhs := rhs.(*ir.SliceExpr)
1681 i, j, k := rhs.Low, rhs.High, rhs.Max
1682 if i != nil && (i.Op() == ir.OLITERAL && i.Val().Kind() == constant.Int && ir.Int64Val(i) == 0) {
1683 // [0:...] is the same as [:...]
1686 // TODO: detect defaults for len/cap also.
1687 // Currently doesn't really work because (*p)[:len(*p)] appears here as:
1690 // if j != nil && (j.Op == OLEN && SameSafeExpr(j.Left, n.Left)) {
1693 // if k != nil && (k.Op == OCAP && SameSafeExpr(k.Left, n.Left)) {
1707 s.assignWhichMayOverlap(n.X, r, deref, skip, mayOverlap)
1711 if ir.IsConst(n.Cond, constant.Bool) {
1712 s.stmtList(n.Cond.Init())
1713 if ir.BoolVal(n.Cond) {
1721 bEnd := s.f.NewBlock(ssa.BlockPlain)
1726 var bThen *ssa.Block
1727 if len(n.Body) != 0 {
1728 bThen = s.f.NewBlock(ssa.BlockPlain)
1732 var bElse *ssa.Block
1733 if len(n.Else) != 0 {
1734 bElse = s.f.NewBlock(ssa.BlockPlain)
1738 s.condBranch(n.Cond, bThen, bElse, likely)
1740 if len(n.Body) != 0 {
1743 if b := s.endBlock(); b != nil {
1747 if len(n.Else) != 0 {
1750 if b := s.endBlock(); b != nil {
1757 n := n.(*ir.ReturnStmt)
1758 s.stmtList(n.Results)
1760 b.Pos = s.lastPos.WithIsStmt()
1763 n := n.(*ir.TailCallStmt)
1764 s.callResult(n.Call, callTail)
1767 b.Kind = ssa.BlockRetJmp // could use BlockExit. BlockRetJmp is mostly for clarity.
1770 case ir.OCONTINUE, ir.OBREAK:
1771 n := n.(*ir.BranchStmt)
1774 // plain break/continue
1782 // labeled break/continue; look up the target
1787 to = lab.continueTarget
1789 to = lab.breakTarget
1794 b.Pos = s.lastPos.WithIsStmt() // Do this even if b is an empty block.
1798 // OFOR: for Ninit; Left; Right { Nbody }
1799 // cond (Left); body (Nbody); incr (Right)
1800 n := n.(*ir.ForStmt)
1801 base.Assert(!n.DistinctVars) // Should all be rewritten before escape analysis
1802 bCond := s.f.NewBlock(ssa.BlockPlain)
1803 bBody := s.f.NewBlock(ssa.BlockPlain)
1804 bIncr := s.f.NewBlock(ssa.BlockPlain)
1805 bEnd := s.f.NewBlock(ssa.BlockPlain)
1807 // ensure empty for loops have correct position; issue #30167
1810 // first, jump to condition test
1814 // generate code to test condition
1817 s.condBranch(n.Cond, bBody, bEnd, 1)
1820 b.Kind = ssa.BlockPlain
1824 // set up for continue/break in body
1825 prevContinue := s.continueTo
1826 prevBreak := s.breakTo
1827 s.continueTo = bIncr
1830 if sym := n.Label; sym != nil {
1833 lab.continueTarget = bIncr
1834 lab.breakTarget = bEnd
1841 // tear down continue/break
1842 s.continueTo = prevContinue
1843 s.breakTo = prevBreak
1845 lab.continueTarget = nil
1846 lab.breakTarget = nil
1849 // done with body, goto incr
1850 if b := s.endBlock(); b != nil {
1859 if b := s.endBlock(); b != nil {
1861 // It can happen that bIncr ends in a block containing only VARKILL,
1862 // and that muddles the debugging experience.
1863 if b.Pos == src.NoXPos {
1870 case ir.OSWITCH, ir.OSELECT:
1871 // These have been mostly rewritten by the front end into their Nbody fields.
1872 // Our main task is to correctly hook up any break statements.
1873 bEnd := s.f.NewBlock(ssa.BlockPlain)
1875 prevBreak := s.breakTo
1879 if n.Op() == ir.OSWITCH {
1880 n := n.(*ir.SwitchStmt)
1884 n := n.(*ir.SelectStmt)
1893 lab.breakTarget = bEnd
1896 // generate body code
1899 s.breakTo = prevBreak
1901 lab.breakTarget = nil
1904 // walk adds explicit OBREAK nodes to the end of all reachable code paths.
1905 // If we still have a current block here, then mark it unreachable.
1906 if s.curBlock != nil {
1909 b.Kind = ssa.BlockExit
1915 n := n.(*ir.JumpTableStmt)
1917 // Make blocks we'll need.
1918 jt := s.f.NewBlock(ssa.BlockJumpTable)
1919 bEnd := s.f.NewBlock(ssa.BlockPlain)
1921 // The only thing that needs evaluating is the index we're looking up.
1922 idx := s.expr(n.Idx)
1923 unsigned := idx.Type.IsUnsigned()
1925 // Extend so we can do everything in uintptr arithmetic.
1926 t := types.Types[types.TUINTPTR]
1927 idx = s.conv(nil, idx, idx.Type, t)
1929 // The ending condition for the current block decides whether we'll use
1930 // the jump table at all.
1931 // We check that min <= idx <= max and jump around the jump table
1932 // if that test fails.
1933 // We implement min <= idx <= max with 0 <= idx-min <= max-min, because
1934 // we'll need idx-min anyway as the control value for the jump table.
1937 min, _ = constant.Uint64Val(n.Cases[0])
1938 max, _ = constant.Uint64Val(n.Cases[len(n.Cases)-1])
1940 mn, _ := constant.Int64Val(n.Cases[0])
1941 mx, _ := constant.Int64Val(n.Cases[len(n.Cases)-1])
1945 // Compare idx-min with max-min, to see if we can use the jump table.
1946 idx = s.newValue2(s.ssaOp(ir.OSUB, t), t, idx, s.uintptrConstant(min))
1947 width := s.uintptrConstant(max - min)
1948 cmp := s.newValue2(s.ssaOp(ir.OLE, t), types.Types[types.TBOOL], idx, width)
1950 b.Kind = ssa.BlockIf
1952 b.AddEdgeTo(jt) // in range - use jump table
1953 b.AddEdgeTo(bEnd) // out of range - no case in the jump table will trigger
1954 b.Likely = ssa.BranchLikely // TODO: assumes missing the table entirely is unlikely. True?
1956 // Build jump table block.
1959 if base.Flag.Cfg.SpectreIndex {
1960 idx = s.newValue2(ssa.OpSpectreSliceIndex, t, idx, width)
1964 // Figure out where we should go for each index in the table.
1965 table := make([]*ssa.Block, max-min+1)
1966 for i := range table {
1967 table[i] = bEnd // default target
1969 for i := range n.Targets {
1971 lab := s.label(n.Targets[i])
1972 if lab.target == nil {
1973 lab.target = s.f.NewBlock(ssa.BlockPlain)
1977 val, _ = constant.Uint64Val(c)
1979 vl, _ := constant.Int64Val(c)
1982 // Overwrite the default target.
1983 table[val-min] = lab.target
1985 for _, t := range table {
1993 n := n.(*ir.UnaryExpr)
1998 n := n.(*ir.InlineMarkStmt)
1999 s.newValue1I(ssa.OpInlMark, types.TypeVoid, n.Index, s.mem())
2002 s.Fatalf("unhandled stmt %v", n.Op())
2006 // If true, share as many open-coded defer exits as possible (with the downside of
2007 // worse line-number information)
2008 const shareDeferExits = false
2010 // exit processes any code that needs to be generated just before returning.
2011 // It returns a BlockRet block that ends the control flow. Its control value
2012 // will be set to the final memory state.
2013 func (s *state) exit() *ssa.Block {
2015 if s.hasOpenDefers {
2016 if shareDeferExits && s.lastDeferExit != nil && len(s.openDefers) == s.lastDeferCount {
2017 if s.curBlock.Kind != ssa.BlockPlain {
2018 panic("Block for an exit should be BlockPlain")
2020 s.curBlock.AddEdgeTo(s.lastDeferExit)
2022 return s.lastDeferFinalBlock
2026 s.rtcall(ir.Syms.Deferreturn, true, nil)
2032 // Do actual return.
2033 // These currently turn into self-copies (in many cases).
2034 resultFields := s.curfn.Type().Results().FieldSlice()
2035 results := make([]*ssa.Value, len(resultFields)+1, len(resultFields)+1)
2036 m = s.newValue0(ssa.OpMakeResult, s.f.OwnAux.LateExpansionResultType())
2037 // Store SSAable and heap-escaped PPARAMOUT variables back to stack locations.
2038 for i, f := range resultFields {
2039 n := f.Nname.(*ir.Name)
2040 if s.canSSA(n) { // result is in some SSA variable
2041 if !n.IsOutputParamInRegisters() && n.Type().HasPointers() {
2042 // We are about to store to the result slot.
2043 s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, n, s.mem())
2045 results[i] = s.variable(n, n.Type())
2046 } else if !n.OnStack() { // result is actually heap allocated
2047 // We are about to copy the in-heap result to the result slot.
2048 if n.Type().HasPointers() {
2049 s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, n, s.mem())
2051 ha := s.expr(n.Heapaddr)
2052 s.instrumentFields(n.Type(), ha, instrumentRead)
2053 results[i] = s.newValue2(ssa.OpDereference, n.Type(), ha, s.mem())
2054 } else { // result is not SSA-able; not escaped, so not on heap, but too large for SSA.
2055 // Before register ABI this ought to be a self-move, home=dest,
2056 // With register ABI, it's still a self-move if parameter is on stack (i.e., too big or overflowed)
2057 // No VarDef, as the result slot is already holding live value.
2058 results[i] = s.newValue2(ssa.OpDereference, n.Type(), s.addr(n), s.mem())
2062 // Run exit code. Today, this is just racefuncexit, in -race mode.
2063 // TODO(register args) this seems risky here with a register-ABI, but not clear it is right to do it earlier either.
2064 // Spills in register allocation might just fix it.
2065 s.stmtList(s.curfn.Exit)
2067 results[len(results)-1] = s.mem()
2068 m.AddArgs(results...)
2071 b.Kind = ssa.BlockRet
2073 if s.hasdefer && s.hasOpenDefers {
2074 s.lastDeferFinalBlock = b
2079 type opAndType struct {
2084 var opToSSA = map[opAndType]ssa.Op{
2085 {ir.OADD, types.TINT8}: ssa.OpAdd8,
2086 {ir.OADD, types.TUINT8}: ssa.OpAdd8,
2087 {ir.OADD, types.TINT16}: ssa.OpAdd16,
2088 {ir.OADD, types.TUINT16}: ssa.OpAdd16,
2089 {ir.OADD, types.TINT32}: ssa.OpAdd32,
2090 {ir.OADD, types.TUINT32}: ssa.OpAdd32,
2091 {ir.OADD, types.TINT64}: ssa.OpAdd64,
2092 {ir.OADD, types.TUINT64}: ssa.OpAdd64,
2093 {ir.OADD, types.TFLOAT32}: ssa.OpAdd32F,
2094 {ir.OADD, types.TFLOAT64}: ssa.OpAdd64F,
2096 {ir.OSUB, types.TINT8}: ssa.OpSub8,
2097 {ir.OSUB, types.TUINT8}: ssa.OpSub8,
2098 {ir.OSUB, types.TINT16}: ssa.OpSub16,
2099 {ir.OSUB, types.TUINT16}: ssa.OpSub16,
2100 {ir.OSUB, types.TINT32}: ssa.OpSub32,
2101 {ir.OSUB, types.TUINT32}: ssa.OpSub32,
2102 {ir.OSUB, types.TINT64}: ssa.OpSub64,
2103 {ir.OSUB, types.TUINT64}: ssa.OpSub64,
2104 {ir.OSUB, types.TFLOAT32}: ssa.OpSub32F,
2105 {ir.OSUB, types.TFLOAT64}: ssa.OpSub64F,
2107 {ir.ONOT, types.TBOOL}: ssa.OpNot,
2109 {ir.ONEG, types.TINT8}: ssa.OpNeg8,
2110 {ir.ONEG, types.TUINT8}: ssa.OpNeg8,
2111 {ir.ONEG, types.TINT16}: ssa.OpNeg16,
2112 {ir.ONEG, types.TUINT16}: ssa.OpNeg16,
2113 {ir.ONEG, types.TINT32}: ssa.OpNeg32,
2114 {ir.ONEG, types.TUINT32}: ssa.OpNeg32,
2115 {ir.ONEG, types.TINT64}: ssa.OpNeg64,
2116 {ir.ONEG, types.TUINT64}: ssa.OpNeg64,
2117 {ir.ONEG, types.TFLOAT32}: ssa.OpNeg32F,
2118 {ir.ONEG, types.TFLOAT64}: ssa.OpNeg64F,
2120 {ir.OBITNOT, types.TINT8}: ssa.OpCom8,
2121 {ir.OBITNOT, types.TUINT8}: ssa.OpCom8,
2122 {ir.OBITNOT, types.TINT16}: ssa.OpCom16,
2123 {ir.OBITNOT, types.TUINT16}: ssa.OpCom16,
2124 {ir.OBITNOT, types.TINT32}: ssa.OpCom32,
2125 {ir.OBITNOT, types.TUINT32}: ssa.OpCom32,
2126 {ir.OBITNOT, types.TINT64}: ssa.OpCom64,
2127 {ir.OBITNOT, types.TUINT64}: ssa.OpCom64,
2129 {ir.OIMAG, types.TCOMPLEX64}: ssa.OpComplexImag,
2130 {ir.OIMAG, types.TCOMPLEX128}: ssa.OpComplexImag,
2131 {ir.OREAL, types.TCOMPLEX64}: ssa.OpComplexReal,
2132 {ir.OREAL, types.TCOMPLEX128}: ssa.OpComplexReal,
2134 {ir.OMUL, types.TINT8}: ssa.OpMul8,
2135 {ir.OMUL, types.TUINT8}: ssa.OpMul8,
2136 {ir.OMUL, types.TINT16}: ssa.OpMul16,
2137 {ir.OMUL, types.TUINT16}: ssa.OpMul16,
2138 {ir.OMUL, types.TINT32}: ssa.OpMul32,
2139 {ir.OMUL, types.TUINT32}: ssa.OpMul32,
2140 {ir.OMUL, types.TINT64}: ssa.OpMul64,
2141 {ir.OMUL, types.TUINT64}: ssa.OpMul64,
2142 {ir.OMUL, types.TFLOAT32}: ssa.OpMul32F,
2143 {ir.OMUL, types.TFLOAT64}: ssa.OpMul64F,
2145 {ir.ODIV, types.TFLOAT32}: ssa.OpDiv32F,
2146 {ir.ODIV, types.TFLOAT64}: ssa.OpDiv64F,
2148 {ir.ODIV, types.TINT8}: ssa.OpDiv8,
2149 {ir.ODIV, types.TUINT8}: ssa.OpDiv8u,
2150 {ir.ODIV, types.TINT16}: ssa.OpDiv16,
2151 {ir.ODIV, types.TUINT16}: ssa.OpDiv16u,
2152 {ir.ODIV, types.TINT32}: ssa.OpDiv32,
2153 {ir.ODIV, types.TUINT32}: ssa.OpDiv32u,
2154 {ir.ODIV, types.TINT64}: ssa.OpDiv64,
2155 {ir.ODIV, types.TUINT64}: ssa.OpDiv64u,
2157 {ir.OMOD, types.TINT8}: ssa.OpMod8,
2158 {ir.OMOD, types.TUINT8}: ssa.OpMod8u,
2159 {ir.OMOD, types.TINT16}: ssa.OpMod16,
2160 {ir.OMOD, types.TUINT16}: ssa.OpMod16u,
2161 {ir.OMOD, types.TINT32}: ssa.OpMod32,
2162 {ir.OMOD, types.TUINT32}: ssa.OpMod32u,
2163 {ir.OMOD, types.TINT64}: ssa.OpMod64,
2164 {ir.OMOD, types.TUINT64}: ssa.OpMod64u,
2166 {ir.OAND, types.TINT8}: ssa.OpAnd8,
2167 {ir.OAND, types.TUINT8}: ssa.OpAnd8,
2168 {ir.OAND, types.TINT16}: ssa.OpAnd16,
2169 {ir.OAND, types.TUINT16}: ssa.OpAnd16,
2170 {ir.OAND, types.TINT32}: ssa.OpAnd32,
2171 {ir.OAND, types.TUINT32}: ssa.OpAnd32,
2172 {ir.OAND, types.TINT64}: ssa.OpAnd64,
2173 {ir.OAND, types.TUINT64}: ssa.OpAnd64,
2175 {ir.OOR, types.TINT8}: ssa.OpOr8,
2176 {ir.OOR, types.TUINT8}: ssa.OpOr8,
2177 {ir.OOR, types.TINT16}: ssa.OpOr16,
2178 {ir.OOR, types.TUINT16}: ssa.OpOr16,
2179 {ir.OOR, types.TINT32}: ssa.OpOr32,
2180 {ir.OOR, types.TUINT32}: ssa.OpOr32,
2181 {ir.OOR, types.TINT64}: ssa.OpOr64,
2182 {ir.OOR, types.TUINT64}: ssa.OpOr64,
2184 {ir.OXOR, types.TINT8}: ssa.OpXor8,
2185 {ir.OXOR, types.TUINT8}: ssa.OpXor8,
2186 {ir.OXOR, types.TINT16}: ssa.OpXor16,
2187 {ir.OXOR, types.TUINT16}: ssa.OpXor16,
2188 {ir.OXOR, types.TINT32}: ssa.OpXor32,
2189 {ir.OXOR, types.TUINT32}: ssa.OpXor32,
2190 {ir.OXOR, types.TINT64}: ssa.OpXor64,
2191 {ir.OXOR, types.TUINT64}: ssa.OpXor64,
2193 {ir.OEQ, types.TBOOL}: ssa.OpEqB,
2194 {ir.OEQ, types.TINT8}: ssa.OpEq8,
2195 {ir.OEQ, types.TUINT8}: ssa.OpEq8,
2196 {ir.OEQ, types.TINT16}: ssa.OpEq16,
2197 {ir.OEQ, types.TUINT16}: ssa.OpEq16,
2198 {ir.OEQ, types.TINT32}: ssa.OpEq32,
2199 {ir.OEQ, types.TUINT32}: ssa.OpEq32,
2200 {ir.OEQ, types.TINT64}: ssa.OpEq64,
2201 {ir.OEQ, types.TUINT64}: ssa.OpEq64,
2202 {ir.OEQ, types.TINTER}: ssa.OpEqInter,
2203 {ir.OEQ, types.TSLICE}: ssa.OpEqSlice,
2204 {ir.OEQ, types.TFUNC}: ssa.OpEqPtr,
2205 {ir.OEQ, types.TMAP}: ssa.OpEqPtr,
2206 {ir.OEQ, types.TCHAN}: ssa.OpEqPtr,
2207 {ir.OEQ, types.TPTR}: ssa.OpEqPtr,
2208 {ir.OEQ, types.TUINTPTR}: ssa.OpEqPtr,
2209 {ir.OEQ, types.TUNSAFEPTR}: ssa.OpEqPtr,
2210 {ir.OEQ, types.TFLOAT64}: ssa.OpEq64F,
2211 {ir.OEQ, types.TFLOAT32}: ssa.OpEq32F,
2213 {ir.ONE, types.TBOOL}: ssa.OpNeqB,
2214 {ir.ONE, types.TINT8}: ssa.OpNeq8,
2215 {ir.ONE, types.TUINT8}: ssa.OpNeq8,
2216 {ir.ONE, types.TINT16}: ssa.OpNeq16,
2217 {ir.ONE, types.TUINT16}: ssa.OpNeq16,
2218 {ir.ONE, types.TINT32}: ssa.OpNeq32,
2219 {ir.ONE, types.TUINT32}: ssa.OpNeq32,
2220 {ir.ONE, types.TINT64}: ssa.OpNeq64,
2221 {ir.ONE, types.TUINT64}: ssa.OpNeq64,
2222 {ir.ONE, types.TINTER}: ssa.OpNeqInter,
2223 {ir.ONE, types.TSLICE}: ssa.OpNeqSlice,
2224 {ir.ONE, types.TFUNC}: ssa.OpNeqPtr,
2225 {ir.ONE, types.TMAP}: ssa.OpNeqPtr,
2226 {ir.ONE, types.TCHAN}: ssa.OpNeqPtr,
2227 {ir.ONE, types.TPTR}: ssa.OpNeqPtr,
2228 {ir.ONE, types.TUINTPTR}: ssa.OpNeqPtr,
2229 {ir.ONE, types.TUNSAFEPTR}: ssa.OpNeqPtr,
2230 {ir.ONE, types.TFLOAT64}: ssa.OpNeq64F,
2231 {ir.ONE, types.TFLOAT32}: ssa.OpNeq32F,
2233 {ir.OLT, types.TINT8}: ssa.OpLess8,
2234 {ir.OLT, types.TUINT8}: ssa.OpLess8U,
2235 {ir.OLT, types.TINT16}: ssa.OpLess16,
2236 {ir.OLT, types.TUINT16}: ssa.OpLess16U,
2237 {ir.OLT, types.TINT32}: ssa.OpLess32,
2238 {ir.OLT, types.TUINT32}: ssa.OpLess32U,
2239 {ir.OLT, types.TINT64}: ssa.OpLess64,
2240 {ir.OLT, types.TUINT64}: ssa.OpLess64U,
2241 {ir.OLT, types.TFLOAT64}: ssa.OpLess64F,
2242 {ir.OLT, types.TFLOAT32}: ssa.OpLess32F,
2244 {ir.OLE, types.TINT8}: ssa.OpLeq8,
2245 {ir.OLE, types.TUINT8}: ssa.OpLeq8U,
2246 {ir.OLE, types.TINT16}: ssa.OpLeq16,
2247 {ir.OLE, types.TUINT16}: ssa.OpLeq16U,
2248 {ir.OLE, types.TINT32}: ssa.OpLeq32,
2249 {ir.OLE, types.TUINT32}: ssa.OpLeq32U,
2250 {ir.OLE, types.TINT64}: ssa.OpLeq64,
2251 {ir.OLE, types.TUINT64}: ssa.OpLeq64U,
2252 {ir.OLE, types.TFLOAT64}: ssa.OpLeq64F,
2253 {ir.OLE, types.TFLOAT32}: ssa.OpLeq32F,
2256 func (s *state) concreteEtype(t *types.Type) types.Kind {
2262 if s.config.PtrSize == 8 {
2267 if s.config.PtrSize == 8 {
2268 return types.TUINT64
2270 return types.TUINT32
2271 case types.TUINTPTR:
2272 if s.config.PtrSize == 8 {
2273 return types.TUINT64
2275 return types.TUINT32
2279 func (s *state) ssaOp(op ir.Op, t *types.Type) ssa.Op {
2280 etype := s.concreteEtype(t)
2281 x, ok := opToSSA[opAndType{op, etype}]
2283 s.Fatalf("unhandled binary op %v %s", op, etype)
2288 type opAndTwoTypes struct {
2294 type twoTypes struct {
2299 type twoOpsAndType struct {
2302 intermediateType types.Kind
2305 var fpConvOpToSSA = map[twoTypes]twoOpsAndType{
2307 {types.TINT8, types.TFLOAT32}: {ssa.OpSignExt8to32, ssa.OpCvt32to32F, types.TINT32},
2308 {types.TINT16, types.TFLOAT32}: {ssa.OpSignExt16to32, ssa.OpCvt32to32F, types.TINT32},
2309 {types.TINT32, types.TFLOAT32}: {ssa.OpCopy, ssa.OpCvt32to32F, types.TINT32},
2310 {types.TINT64, types.TFLOAT32}: {ssa.OpCopy, ssa.OpCvt64to32F, types.TINT64},
2312 {types.TINT8, types.TFLOAT64}: {ssa.OpSignExt8to32, ssa.OpCvt32to64F, types.TINT32},
2313 {types.TINT16, types.TFLOAT64}: {ssa.OpSignExt16to32, ssa.OpCvt32to64F, types.TINT32},
2314 {types.TINT32, types.TFLOAT64}: {ssa.OpCopy, ssa.OpCvt32to64F, types.TINT32},
2315 {types.TINT64, types.TFLOAT64}: {ssa.OpCopy, ssa.OpCvt64to64F, types.TINT64},
2317 {types.TFLOAT32, types.TINT8}: {ssa.OpCvt32Fto32, ssa.OpTrunc32to8, types.TINT32},
2318 {types.TFLOAT32, types.TINT16}: {ssa.OpCvt32Fto32, ssa.OpTrunc32to16, types.TINT32},
2319 {types.TFLOAT32, types.TINT32}: {ssa.OpCvt32Fto32, ssa.OpCopy, types.TINT32},
2320 {types.TFLOAT32, types.TINT64}: {ssa.OpCvt32Fto64, ssa.OpCopy, types.TINT64},
2322 {types.TFLOAT64, types.TINT8}: {ssa.OpCvt64Fto32, ssa.OpTrunc32to8, types.TINT32},
2323 {types.TFLOAT64, types.TINT16}: {ssa.OpCvt64Fto32, ssa.OpTrunc32to16, types.TINT32},
2324 {types.TFLOAT64, types.TINT32}: {ssa.OpCvt64Fto32, ssa.OpCopy, types.TINT32},
2325 {types.TFLOAT64, types.TINT64}: {ssa.OpCvt64Fto64, ssa.OpCopy, types.TINT64},
2327 {types.TUINT8, types.TFLOAT32}: {ssa.OpZeroExt8to32, ssa.OpCvt32to32F, types.TINT32},
2328 {types.TUINT16, types.TFLOAT32}: {ssa.OpZeroExt16to32, ssa.OpCvt32to32F, types.TINT32},
2329 {types.TUINT32, types.TFLOAT32}: {ssa.OpZeroExt32to64, ssa.OpCvt64to32F, types.TINT64}, // go wide to dodge unsigned
2330 {types.TUINT64, types.TFLOAT32}: {ssa.OpCopy, ssa.OpInvalid, types.TUINT64}, // Cvt64Uto32F, branchy code expansion instead
2332 {types.TUINT8, types.TFLOAT64}: {ssa.OpZeroExt8to32, ssa.OpCvt32to64F, types.TINT32},
2333 {types.TUINT16, types.TFLOAT64}: {ssa.OpZeroExt16to32, ssa.OpCvt32to64F, types.TINT32},
2334 {types.TUINT32, types.TFLOAT64}: {ssa.OpZeroExt32to64, ssa.OpCvt64to64F, types.TINT64}, // go wide to dodge unsigned
2335 {types.TUINT64, types.TFLOAT64}: {ssa.OpCopy, ssa.OpInvalid, types.TUINT64}, // Cvt64Uto64F, branchy code expansion instead
2337 {types.TFLOAT32, types.TUINT8}: {ssa.OpCvt32Fto32, ssa.OpTrunc32to8, types.TINT32},
2338 {types.TFLOAT32, types.TUINT16}: {ssa.OpCvt32Fto32, ssa.OpTrunc32to16, types.TINT32},
2339 {types.TFLOAT32, types.TUINT32}: {ssa.OpCvt32Fto64, ssa.OpTrunc64to32, types.TINT64}, // go wide to dodge unsigned
2340 {types.TFLOAT32, types.TUINT64}: {ssa.OpInvalid, ssa.OpCopy, types.TUINT64}, // Cvt32Fto64U, branchy code expansion instead
2342 {types.TFLOAT64, types.TUINT8}: {ssa.OpCvt64Fto32, ssa.OpTrunc32to8, types.TINT32},
2343 {types.TFLOAT64, types.TUINT16}: {ssa.OpCvt64Fto32, ssa.OpTrunc32to16, types.TINT32},
2344 {types.TFLOAT64, types.TUINT32}: {ssa.OpCvt64Fto64, ssa.OpTrunc64to32, types.TINT64}, // go wide to dodge unsigned
2345 {types.TFLOAT64, types.TUINT64}: {ssa.OpInvalid, ssa.OpCopy, types.TUINT64}, // Cvt64Fto64U, branchy code expansion instead
2348 {types.TFLOAT64, types.TFLOAT32}: {ssa.OpCvt64Fto32F, ssa.OpCopy, types.TFLOAT32},
2349 {types.TFLOAT64, types.TFLOAT64}: {ssa.OpRound64F, ssa.OpCopy, types.TFLOAT64},
2350 {types.TFLOAT32, types.TFLOAT32}: {ssa.OpRound32F, ssa.OpCopy, types.TFLOAT32},
2351 {types.TFLOAT32, types.TFLOAT64}: {ssa.OpCvt32Fto64F, ssa.OpCopy, types.TFLOAT64},
2354 // this map is used only for 32-bit arch, and only includes the difference
2355 // on 32-bit arch, don't use int64<->float conversion for uint32
2356 var fpConvOpToSSA32 = map[twoTypes]twoOpsAndType{
2357 {types.TUINT32, types.TFLOAT32}: {ssa.OpCopy, ssa.OpCvt32Uto32F, types.TUINT32},
2358 {types.TUINT32, types.TFLOAT64}: {ssa.OpCopy, ssa.OpCvt32Uto64F, types.TUINT32},
2359 {types.TFLOAT32, types.TUINT32}: {ssa.OpCvt32Fto32U, ssa.OpCopy, types.TUINT32},
2360 {types.TFLOAT64, types.TUINT32}: {ssa.OpCvt64Fto32U, ssa.OpCopy, types.TUINT32},
2363 // uint64<->float conversions, only on machines that have instructions for that
2364 var uint64fpConvOpToSSA = map[twoTypes]twoOpsAndType{
2365 {types.TUINT64, types.TFLOAT32}: {ssa.OpCopy, ssa.OpCvt64Uto32F, types.TUINT64},
2366 {types.TUINT64, types.TFLOAT64}: {ssa.OpCopy, ssa.OpCvt64Uto64F, types.TUINT64},
2367 {types.TFLOAT32, types.TUINT64}: {ssa.OpCvt32Fto64U, ssa.OpCopy, types.TUINT64},
2368 {types.TFLOAT64, types.TUINT64}: {ssa.OpCvt64Fto64U, ssa.OpCopy, types.TUINT64},
2371 var shiftOpToSSA = map[opAndTwoTypes]ssa.Op{
2372 {ir.OLSH, types.TINT8, types.TUINT8}: ssa.OpLsh8x8,
2373 {ir.OLSH, types.TUINT8, types.TUINT8}: ssa.OpLsh8x8,
2374 {ir.OLSH, types.TINT8, types.TUINT16}: ssa.OpLsh8x16,
2375 {ir.OLSH, types.TUINT8, types.TUINT16}: ssa.OpLsh8x16,
2376 {ir.OLSH, types.TINT8, types.TUINT32}: ssa.OpLsh8x32,
2377 {ir.OLSH, types.TUINT8, types.TUINT32}: ssa.OpLsh8x32,
2378 {ir.OLSH, types.TINT8, types.TUINT64}: ssa.OpLsh8x64,
2379 {ir.OLSH, types.TUINT8, types.TUINT64}: ssa.OpLsh8x64,
2381 {ir.OLSH, types.TINT16, types.TUINT8}: ssa.OpLsh16x8,
2382 {ir.OLSH, types.TUINT16, types.TUINT8}: ssa.OpLsh16x8,
2383 {ir.OLSH, types.TINT16, types.TUINT16}: ssa.OpLsh16x16,
2384 {ir.OLSH, types.TUINT16, types.TUINT16}: ssa.OpLsh16x16,
2385 {ir.OLSH, types.TINT16, types.TUINT32}: ssa.OpLsh16x32,
2386 {ir.OLSH, types.TUINT16, types.TUINT32}: ssa.OpLsh16x32,
2387 {ir.OLSH, types.TINT16, types.TUINT64}: ssa.OpLsh16x64,
2388 {ir.OLSH, types.TUINT16, types.TUINT64}: ssa.OpLsh16x64,
2390 {ir.OLSH, types.TINT32, types.TUINT8}: ssa.OpLsh32x8,
2391 {ir.OLSH, types.TUINT32, types.TUINT8}: ssa.OpLsh32x8,
2392 {ir.OLSH, types.TINT32, types.TUINT16}: ssa.OpLsh32x16,
2393 {ir.OLSH, types.TUINT32, types.TUINT16}: ssa.OpLsh32x16,
2394 {ir.OLSH, types.TINT32, types.TUINT32}: ssa.OpLsh32x32,
2395 {ir.OLSH, types.TUINT32, types.TUINT32}: ssa.OpLsh32x32,
2396 {ir.OLSH, types.TINT32, types.TUINT64}: ssa.OpLsh32x64,
2397 {ir.OLSH, types.TUINT32, types.TUINT64}: ssa.OpLsh32x64,
2399 {ir.OLSH, types.TINT64, types.TUINT8}: ssa.OpLsh64x8,
2400 {ir.OLSH, types.TUINT64, types.TUINT8}: ssa.OpLsh64x8,
2401 {ir.OLSH, types.TINT64, types.TUINT16}: ssa.OpLsh64x16,
2402 {ir.OLSH, types.TUINT64, types.TUINT16}: ssa.OpLsh64x16,
2403 {ir.OLSH, types.TINT64, types.TUINT32}: ssa.OpLsh64x32,
2404 {ir.OLSH, types.TUINT64, types.TUINT32}: ssa.OpLsh64x32,
2405 {ir.OLSH, types.TINT64, types.TUINT64}: ssa.OpLsh64x64,
2406 {ir.OLSH, types.TUINT64, types.TUINT64}: ssa.OpLsh64x64,
2408 {ir.ORSH, types.TINT8, types.TUINT8}: ssa.OpRsh8x8,
2409 {ir.ORSH, types.TUINT8, types.TUINT8}: ssa.OpRsh8Ux8,
2410 {ir.ORSH, types.TINT8, types.TUINT16}: ssa.OpRsh8x16,
2411 {ir.ORSH, types.TUINT8, types.TUINT16}: ssa.OpRsh8Ux16,
2412 {ir.ORSH, types.TINT8, types.TUINT32}: ssa.OpRsh8x32,
2413 {ir.ORSH, types.TUINT8, types.TUINT32}: ssa.OpRsh8Ux32,
2414 {ir.ORSH, types.TINT8, types.TUINT64}: ssa.OpRsh8x64,
2415 {ir.ORSH, types.TUINT8, types.TUINT64}: ssa.OpRsh8Ux64,
2417 {ir.ORSH, types.TINT16, types.TUINT8}: ssa.OpRsh16x8,
2418 {ir.ORSH, types.TUINT16, types.TUINT8}: ssa.OpRsh16Ux8,
2419 {ir.ORSH, types.TINT16, types.TUINT16}: ssa.OpRsh16x16,
2420 {ir.ORSH, types.TUINT16, types.TUINT16}: ssa.OpRsh16Ux16,
2421 {ir.ORSH, types.TINT16, types.TUINT32}: ssa.OpRsh16x32,
2422 {ir.ORSH, types.TUINT16, types.TUINT32}: ssa.OpRsh16Ux32,
2423 {ir.ORSH, types.TINT16, types.TUINT64}: ssa.OpRsh16x64,
2424 {ir.ORSH, types.TUINT16, types.TUINT64}: ssa.OpRsh16Ux64,
2426 {ir.ORSH, types.TINT32, types.TUINT8}: ssa.OpRsh32x8,
2427 {ir.ORSH, types.TUINT32, types.TUINT8}: ssa.OpRsh32Ux8,
2428 {ir.ORSH, types.TINT32, types.TUINT16}: ssa.OpRsh32x16,
2429 {ir.ORSH, types.TUINT32, types.TUINT16}: ssa.OpRsh32Ux16,
2430 {ir.ORSH, types.TINT32, types.TUINT32}: ssa.OpRsh32x32,
2431 {ir.ORSH, types.TUINT32, types.TUINT32}: ssa.OpRsh32Ux32,
2432 {ir.ORSH, types.TINT32, types.TUINT64}: ssa.OpRsh32x64,
2433 {ir.ORSH, types.TUINT32, types.TUINT64}: ssa.OpRsh32Ux64,
2435 {ir.ORSH, types.TINT64, types.TUINT8}: ssa.OpRsh64x8,
2436 {ir.ORSH, types.TUINT64, types.TUINT8}: ssa.OpRsh64Ux8,
2437 {ir.ORSH, types.TINT64, types.TUINT16}: ssa.OpRsh64x16,
2438 {ir.ORSH, types.TUINT64, types.TUINT16}: ssa.OpRsh64Ux16,
2439 {ir.ORSH, types.TINT64, types.TUINT32}: ssa.OpRsh64x32,
2440 {ir.ORSH, types.TUINT64, types.TUINT32}: ssa.OpRsh64Ux32,
2441 {ir.ORSH, types.TINT64, types.TUINT64}: ssa.OpRsh64x64,
2442 {ir.ORSH, types.TUINT64, types.TUINT64}: ssa.OpRsh64Ux64,
2445 func (s *state) ssaShiftOp(op ir.Op, t *types.Type, u *types.Type) ssa.Op {
2446 etype1 := s.concreteEtype(t)
2447 etype2 := s.concreteEtype(u)
2448 x, ok := shiftOpToSSA[opAndTwoTypes{op, etype1, etype2}]
2450 s.Fatalf("unhandled shift op %v etype=%s/%s", op, etype1, etype2)
2455 func (s *state) uintptrConstant(v uint64) *ssa.Value {
2456 if s.config.PtrSize == 4 {
2457 return s.newValue0I(ssa.OpConst32, types.Types[types.TUINTPTR], int64(v))
2459 return s.newValue0I(ssa.OpConst64, types.Types[types.TUINTPTR], int64(v))
2462 func (s *state) conv(n ir.Node, v *ssa.Value, ft, tt *types.Type) *ssa.Value {
2463 if ft.IsBoolean() && tt.IsKind(types.TUINT8) {
2464 // Bool -> uint8 is generated internally when indexing into runtime.staticbyte.
2465 return s.newValue1(ssa.OpCvtBoolToUint8, tt, v)
2467 if ft.IsInteger() && tt.IsInteger() {
2469 if tt.Size() == ft.Size() {
2471 } else if tt.Size() < ft.Size() {
2473 switch 10*ft.Size() + tt.Size() {
2475 op = ssa.OpTrunc16to8
2477 op = ssa.OpTrunc32to8
2479 op = ssa.OpTrunc32to16
2481 op = ssa.OpTrunc64to8
2483 op = ssa.OpTrunc64to16
2485 op = ssa.OpTrunc64to32
2487 s.Fatalf("weird integer truncation %v -> %v", ft, tt)
2489 } else if ft.IsSigned() {
2491 switch 10*ft.Size() + tt.Size() {
2493 op = ssa.OpSignExt8to16
2495 op = ssa.OpSignExt8to32
2497 op = ssa.OpSignExt8to64
2499 op = ssa.OpSignExt16to32
2501 op = ssa.OpSignExt16to64
2503 op = ssa.OpSignExt32to64
2505 s.Fatalf("bad integer sign extension %v -> %v", ft, tt)
2509 switch 10*ft.Size() + tt.Size() {
2511 op = ssa.OpZeroExt8to16
2513 op = ssa.OpZeroExt8to32
2515 op = ssa.OpZeroExt8to64
2517 op = ssa.OpZeroExt16to32
2519 op = ssa.OpZeroExt16to64
2521 op = ssa.OpZeroExt32to64
2523 s.Fatalf("weird integer sign extension %v -> %v", ft, tt)
2526 return s.newValue1(op, tt, v)
2529 if ft.IsComplex() && tt.IsComplex() {
2531 if ft.Size() == tt.Size() {
2538 s.Fatalf("weird complex conversion %v -> %v", ft, tt)
2540 } else if ft.Size() == 8 && tt.Size() == 16 {
2541 op = ssa.OpCvt32Fto64F
2542 } else if ft.Size() == 16 && tt.Size() == 8 {
2543 op = ssa.OpCvt64Fto32F
2545 s.Fatalf("weird complex conversion %v -> %v", ft, tt)
2547 ftp := types.FloatForComplex(ft)
2548 ttp := types.FloatForComplex(tt)
2549 return s.newValue2(ssa.OpComplexMake, tt,
2550 s.newValueOrSfCall1(op, ttp, s.newValue1(ssa.OpComplexReal, ftp, v)),
2551 s.newValueOrSfCall1(op, ttp, s.newValue1(ssa.OpComplexImag, ftp, v)))
2554 if tt.IsComplex() { // and ft is not complex
2555 // Needed for generics support - can't happen in normal Go code.
2556 et := types.FloatForComplex(tt)
2557 v = s.conv(n, v, ft, et)
2558 return s.newValue2(ssa.OpComplexMake, tt, v, s.zeroVal(et))
2561 if ft.IsFloat() || tt.IsFloat() {
2562 conv, ok := fpConvOpToSSA[twoTypes{s.concreteEtype(ft), s.concreteEtype(tt)}]
2563 if s.config.RegSize == 4 && Arch.LinkArch.Family != sys.MIPS && !s.softFloat {
2564 if conv1, ok1 := fpConvOpToSSA32[twoTypes{s.concreteEtype(ft), s.concreteEtype(tt)}]; ok1 {
2568 if Arch.LinkArch.Family == sys.ARM64 || Arch.LinkArch.Family == sys.Wasm || Arch.LinkArch.Family == sys.S390X || s.softFloat {
2569 if conv1, ok1 := uint64fpConvOpToSSA[twoTypes{s.concreteEtype(ft), s.concreteEtype(tt)}]; ok1 {
2574 if Arch.LinkArch.Family == sys.MIPS && !s.softFloat {
2575 if ft.Size() == 4 && ft.IsInteger() && !ft.IsSigned() {
2576 // tt is float32 or float64, and ft is also unsigned
2578 return s.uint32Tofloat32(n, v, ft, tt)
2581 return s.uint32Tofloat64(n, v, ft, tt)
2583 } else if tt.Size() == 4 && tt.IsInteger() && !tt.IsSigned() {
2584 // ft is float32 or float64, and tt is unsigned integer
2586 return s.float32ToUint32(n, v, ft, tt)
2589 return s.float64ToUint32(n, v, ft, tt)
2595 s.Fatalf("weird float conversion %v -> %v", ft, tt)
2597 op1, op2, it := conv.op1, conv.op2, conv.intermediateType
2599 if op1 != ssa.OpInvalid && op2 != ssa.OpInvalid {
2600 // normal case, not tripping over unsigned 64
2601 if op1 == ssa.OpCopy {
2602 if op2 == ssa.OpCopy {
2605 return s.newValueOrSfCall1(op2, tt, v)
2607 if op2 == ssa.OpCopy {
2608 return s.newValueOrSfCall1(op1, tt, v)
2610 return s.newValueOrSfCall1(op2, tt, s.newValueOrSfCall1(op1, types.Types[it], v))
2612 // Tricky 64-bit unsigned cases.
2614 // tt is float32 or float64, and ft is also unsigned
2616 return s.uint64Tofloat32(n, v, ft, tt)
2619 return s.uint64Tofloat64(n, v, ft, tt)
2621 s.Fatalf("weird unsigned integer to float conversion %v -> %v", ft, tt)
2623 // ft is float32 or float64, and tt is unsigned integer
2625 return s.float32ToUint64(n, v, ft, tt)
2628 return s.float64ToUint64(n, v, ft, tt)
2630 s.Fatalf("weird float to unsigned integer conversion %v -> %v", ft, tt)
2634 s.Fatalf("unhandled OCONV %s -> %s", ft.Kind(), tt.Kind())
2638 // expr converts the expression n to ssa, adds it to s and returns the ssa result.
2639 func (s *state) expr(n ir.Node) *ssa.Value {
2640 return s.exprCheckPtr(n, true)
2643 func (s *state) exprCheckPtr(n ir.Node, checkPtrOK bool) *ssa.Value {
2644 if ir.HasUniquePos(n) {
2645 // ONAMEs and named OLITERALs have the line number
2646 // of the decl, not the use. See issue 14742.
2651 s.stmtList(n.Init())
2653 case ir.OBYTES2STRTMP:
2654 n := n.(*ir.ConvExpr)
2655 slice := s.expr(n.X)
2656 ptr := s.newValue1(ssa.OpSlicePtr, s.f.Config.Types.BytePtr, slice)
2657 len := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], slice)
2658 return s.newValue2(ssa.OpStringMake, n.Type(), ptr, len)
2659 case ir.OSTR2BYTESTMP:
2660 n := n.(*ir.ConvExpr)
2662 ptr := s.newValue1(ssa.OpStringPtr, s.f.Config.Types.BytePtr, str)
2663 len := s.newValue1(ssa.OpStringLen, types.Types[types.TINT], str)
2664 return s.newValue3(ssa.OpSliceMake, n.Type(), ptr, len, len)
2666 n := n.(*ir.UnaryExpr)
2667 aux := n.X.(*ir.Name).Linksym()
2668 // OCFUNC is used to build function values, which must
2669 // always reference ABIInternal entry points.
2670 if aux.ABI() != obj.ABIInternal {
2671 s.Fatalf("expected ABIInternal: %v", aux.ABI())
2673 return s.entryNewValue1A(ssa.OpAddr, n.Type(), aux, s.sb)
2676 if n.Class == ir.PFUNC {
2677 // "value" of a function is the address of the function's closure
2678 sym := staticdata.FuncLinksym(n)
2679 return s.entryNewValue1A(ssa.OpAddr, types.NewPtr(n.Type()), sym, s.sb)
2682 return s.variable(n, n.Type())
2684 return s.load(n.Type(), s.addr(n))
2685 case ir.OLINKSYMOFFSET:
2686 n := n.(*ir.LinksymOffsetExpr)
2687 return s.load(n.Type(), s.addr(n))
2689 n := n.(*ir.NilExpr)
2693 return s.constSlice(t)
2694 case t.IsInterface():
2695 return s.constInterface(t)
2697 return s.constNil(t)
2700 switch u := n.Val(); u.Kind() {
2702 i := ir.IntVal(n.Type(), u)
2703 switch n.Type().Size() {
2705 return s.constInt8(n.Type(), int8(i))
2707 return s.constInt16(n.Type(), int16(i))
2709 return s.constInt32(n.Type(), int32(i))
2711 return s.constInt64(n.Type(), i)
2713 s.Fatalf("bad integer size %d", n.Type().Size())
2716 case constant.String:
2717 i := constant.StringVal(u)
2719 return s.constEmptyString(n.Type())
2721 return s.entryNewValue0A(ssa.OpConstString, n.Type(), ssa.StringToAux(i))
2723 return s.constBool(constant.BoolVal(u))
2724 case constant.Float:
2725 f, _ := constant.Float64Val(u)
2726 switch n.Type().Size() {
2728 return s.constFloat32(n.Type(), f)
2730 return s.constFloat64(n.Type(), f)
2732 s.Fatalf("bad float size %d", n.Type().Size())
2735 case constant.Complex:
2736 re, _ := constant.Float64Val(constant.Real(u))
2737 im, _ := constant.Float64Val(constant.Imag(u))
2738 switch n.Type().Size() {
2740 pt := types.Types[types.TFLOAT32]
2741 return s.newValue2(ssa.OpComplexMake, n.Type(),
2742 s.constFloat32(pt, re),
2743 s.constFloat32(pt, im))
2745 pt := types.Types[types.TFLOAT64]
2746 return s.newValue2(ssa.OpComplexMake, n.Type(),
2747 s.constFloat64(pt, re),
2748 s.constFloat64(pt, im))
2750 s.Fatalf("bad complex size %d", n.Type().Size())
2754 s.Fatalf("unhandled OLITERAL %v", u.Kind())
2758 n := n.(*ir.ConvExpr)
2762 // Assume everything will work out, so set up our return value.
2763 // Anything interesting that happens from here is a fatal.
2769 // Special case for not confusing GC and liveness.
2770 // We don't want pointers accidentally classified
2771 // as not-pointers or vice-versa because of copy
2773 if to.IsPtrShaped() != from.IsPtrShaped() {
2774 return s.newValue2(ssa.OpConvert, to, x, s.mem())
2777 v := s.newValue1(ssa.OpCopy, to, x) // ensure that v has the right type
2780 if to.Kind() == types.TFUNC && from.IsPtrShaped() {
2784 // named <--> unnamed type or typed <--> untyped const
2785 if from.Kind() == to.Kind() {
2789 // unsafe.Pointer <--> *T
2790 if to.IsUnsafePtr() && from.IsPtrShaped() || from.IsUnsafePtr() && to.IsPtrShaped() {
2791 if s.checkPtrEnabled && checkPtrOK && to.IsPtr() && from.IsUnsafePtr() {
2792 s.checkPtrAlignment(n, v, nil)
2798 if to.Kind() == types.TMAP && from.IsPtr() &&
2799 to.MapType().Hmap == from.Elem() {
2803 types.CalcSize(from)
2805 if from.Size() != to.Size() {
2806 s.Fatalf("CONVNOP width mismatch %v (%d) -> %v (%d)\n", from, from.Size(), to, to.Size())
2809 if etypesign(from.Kind()) != etypesign(to.Kind()) {
2810 s.Fatalf("CONVNOP sign mismatch %v (%s) -> %v (%s)\n", from, from.Kind(), to, to.Kind())
2814 if base.Flag.Cfg.Instrumenting {
2815 // These appear to be fine, but they fail the
2816 // integer constraint below, so okay them here.
2817 // Sample non-integer conversion: map[string]string -> *uint8
2821 if etypesign(from.Kind()) == 0 {
2822 s.Fatalf("CONVNOP unrecognized non-integer %v -> %v\n", from, to)
2826 // integer, same width, same sign
2830 n := n.(*ir.ConvExpr)
2832 return s.conv(n, x, n.X.Type(), n.Type())
2835 n := n.(*ir.TypeAssertExpr)
2836 res, _ := s.dottype(n, false)
2839 case ir.ODYNAMICDOTTYPE:
2840 n := n.(*ir.DynamicTypeAssertExpr)
2841 res, _ := s.dynamicDottype(n, false)
2845 case ir.OLT, ir.OEQ, ir.ONE, ir.OLE, ir.OGE, ir.OGT:
2846 n := n.(*ir.BinaryExpr)
2849 if n.X.Type().IsComplex() {
2850 pt := types.FloatForComplex(n.X.Type())
2851 op := s.ssaOp(ir.OEQ, pt)
2852 r := s.newValueOrSfCall2(op, types.Types[types.TBOOL], s.newValue1(ssa.OpComplexReal, pt, a), s.newValue1(ssa.OpComplexReal, pt, b))
2853 i := s.newValueOrSfCall2(op, types.Types[types.TBOOL], s.newValue1(ssa.OpComplexImag, pt, a), s.newValue1(ssa.OpComplexImag, pt, b))
2854 c := s.newValue2(ssa.OpAndB, types.Types[types.TBOOL], r, i)
2859 return s.newValue1(ssa.OpNot, types.Types[types.TBOOL], c)
2861 s.Fatalf("ordered complex compare %v", n.Op())
2865 // Convert OGE and OGT into OLE and OLT.
2869 op, a, b = ir.OLE, b, a
2871 op, a, b = ir.OLT, b, a
2873 if n.X.Type().IsFloat() {
2875 return s.newValueOrSfCall2(s.ssaOp(op, n.X.Type()), types.Types[types.TBOOL], a, b)
2877 // integer comparison
2878 return s.newValue2(s.ssaOp(op, n.X.Type()), types.Types[types.TBOOL], a, b)
2880 n := n.(*ir.BinaryExpr)
2883 if n.Type().IsComplex() {
2884 mulop := ssa.OpMul64F
2885 addop := ssa.OpAdd64F
2886 subop := ssa.OpSub64F
2887 pt := types.FloatForComplex(n.Type()) // Could be Float32 or Float64
2888 wt := types.Types[types.TFLOAT64] // Compute in Float64 to minimize cancellation error
2890 areal := s.newValue1(ssa.OpComplexReal, pt, a)
2891 breal := s.newValue1(ssa.OpComplexReal, pt, b)
2892 aimag := s.newValue1(ssa.OpComplexImag, pt, a)
2893 bimag := s.newValue1(ssa.OpComplexImag, pt, b)
2895 if pt != wt { // Widen for calculation
2896 areal = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, areal)
2897 breal = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, breal)
2898 aimag = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, aimag)
2899 bimag = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, bimag)
2902 xreal := s.newValueOrSfCall2(subop, wt, s.newValueOrSfCall2(mulop, wt, areal, breal), s.newValueOrSfCall2(mulop, wt, aimag, bimag))
2903 ximag := s.newValueOrSfCall2(addop, wt, s.newValueOrSfCall2(mulop, wt, areal, bimag), s.newValueOrSfCall2(mulop, wt, aimag, breal))
2905 if pt != wt { // Narrow to store back
2906 xreal = s.newValueOrSfCall1(ssa.OpCvt64Fto32F, pt, xreal)
2907 ximag = s.newValueOrSfCall1(ssa.OpCvt64Fto32F, pt, ximag)
2910 return s.newValue2(ssa.OpComplexMake, n.Type(), xreal, ximag)
2913 if n.Type().IsFloat() {
2914 return s.newValueOrSfCall2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2917 return s.newValue2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2920 n := n.(*ir.BinaryExpr)
2923 if n.Type().IsComplex() {
2924 // TODO this is not executed because the front-end substitutes a runtime call.
2925 // That probably ought to change; with modest optimization the widen/narrow
2926 // conversions could all be elided in larger expression trees.
2927 mulop := ssa.OpMul64F
2928 addop := ssa.OpAdd64F
2929 subop := ssa.OpSub64F
2930 divop := ssa.OpDiv64F
2931 pt := types.FloatForComplex(n.Type()) // Could be Float32 or Float64
2932 wt := types.Types[types.TFLOAT64] // Compute in Float64 to minimize cancellation error
2934 areal := s.newValue1(ssa.OpComplexReal, pt, a)
2935 breal := s.newValue1(ssa.OpComplexReal, pt, b)
2936 aimag := s.newValue1(ssa.OpComplexImag, pt, a)
2937 bimag := s.newValue1(ssa.OpComplexImag, pt, b)
2939 if pt != wt { // Widen for calculation
2940 areal = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, areal)
2941 breal = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, breal)
2942 aimag = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, aimag)
2943 bimag = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, bimag)
2946 denom := s.newValueOrSfCall2(addop, wt, s.newValueOrSfCall2(mulop, wt, breal, breal), s.newValueOrSfCall2(mulop, wt, bimag, bimag))
2947 xreal := s.newValueOrSfCall2(addop, wt, s.newValueOrSfCall2(mulop, wt, areal, breal), s.newValueOrSfCall2(mulop, wt, aimag, bimag))
2948 ximag := s.newValueOrSfCall2(subop, wt, s.newValueOrSfCall2(mulop, wt, aimag, breal), s.newValueOrSfCall2(mulop, wt, areal, bimag))
2950 // TODO not sure if this is best done in wide precision or narrow
2951 // Double-rounding might be an issue.
2952 // Note that the pre-SSA implementation does the entire calculation
2953 // in wide format, so wide is compatible.
2954 xreal = s.newValueOrSfCall2(divop, wt, xreal, denom)
2955 ximag = s.newValueOrSfCall2(divop, wt, ximag, denom)
2957 if pt != wt { // Narrow to store back
2958 xreal = s.newValueOrSfCall1(ssa.OpCvt64Fto32F, pt, xreal)
2959 ximag = s.newValueOrSfCall1(ssa.OpCvt64Fto32F, pt, ximag)
2961 return s.newValue2(ssa.OpComplexMake, n.Type(), xreal, ximag)
2963 if n.Type().IsFloat() {
2964 return s.newValueOrSfCall2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2966 return s.intDivide(n, a, b)
2968 n := n.(*ir.BinaryExpr)
2971 return s.intDivide(n, a, b)
2972 case ir.OADD, ir.OSUB:
2973 n := n.(*ir.BinaryExpr)
2976 if n.Type().IsComplex() {
2977 pt := types.FloatForComplex(n.Type())
2978 op := s.ssaOp(n.Op(), pt)
2979 return s.newValue2(ssa.OpComplexMake, n.Type(),
2980 s.newValueOrSfCall2(op, pt, s.newValue1(ssa.OpComplexReal, pt, a), s.newValue1(ssa.OpComplexReal, pt, b)),
2981 s.newValueOrSfCall2(op, pt, s.newValue1(ssa.OpComplexImag, pt, a), s.newValue1(ssa.OpComplexImag, pt, b)))
2983 if n.Type().IsFloat() {
2984 return s.newValueOrSfCall2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2986 return s.newValue2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2987 case ir.OAND, ir.OOR, ir.OXOR:
2988 n := n.(*ir.BinaryExpr)
2991 return s.newValue2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2993 n := n.(*ir.BinaryExpr)
2996 b = s.newValue1(s.ssaOp(ir.OBITNOT, b.Type), b.Type, b)
2997 return s.newValue2(s.ssaOp(ir.OAND, n.Type()), a.Type, a, b)
2998 case ir.OLSH, ir.ORSH:
2999 n := n.(*ir.BinaryExpr)
3004 cmp := s.newValue2(s.ssaOp(ir.OLE, bt), types.Types[types.TBOOL], s.zeroVal(bt), b)
3005 s.check(cmp, ir.Syms.Panicshift)
3006 bt = bt.ToUnsigned()
3008 return s.newValue2(s.ssaShiftOp(n.Op(), n.Type(), bt), a.Type, a, b)
3009 case ir.OANDAND, ir.OOROR:
3010 // To implement OANDAND (and OOROR), we introduce a
3011 // new temporary variable to hold the result. The
3012 // variable is associated with the OANDAND node in the
3013 // s.vars table (normally variables are only
3014 // associated with ONAME nodes). We convert
3021 // Using var in the subsequent block introduces the
3022 // necessary phi variable.
3023 n := n.(*ir.LogicalExpr)
3028 b.Kind = ssa.BlockIf
3030 // In theory, we should set b.Likely here based on context.
3031 // However, gc only gives us likeliness hints
3032 // in a single place, for plain OIF statements,
3033 // and passing around context is finnicky, so don't bother for now.
3035 bRight := s.f.NewBlock(ssa.BlockPlain)
3036 bResult := s.f.NewBlock(ssa.BlockPlain)
3037 if n.Op() == ir.OANDAND {
3039 b.AddEdgeTo(bResult)
3040 } else if n.Op() == ir.OOROR {
3041 b.AddEdgeTo(bResult)
3045 s.startBlock(bRight)
3050 b.AddEdgeTo(bResult)
3052 s.startBlock(bResult)
3053 return s.variable(n, types.Types[types.TBOOL])
3055 n := n.(*ir.BinaryExpr)
3058 return s.newValue2(ssa.OpComplexMake, n.Type(), r, i)
3062 n := n.(*ir.UnaryExpr)
3064 if n.Type().IsComplex() {
3065 tp := types.FloatForComplex(n.Type())
3066 negop := s.ssaOp(n.Op(), tp)
3067 return s.newValue2(ssa.OpComplexMake, n.Type(),
3068 s.newValue1(negop, tp, s.newValue1(ssa.OpComplexReal, tp, a)),
3069 s.newValue1(negop, tp, s.newValue1(ssa.OpComplexImag, tp, a)))
3071 return s.newValue1(s.ssaOp(n.Op(), n.Type()), a.Type, a)
3072 case ir.ONOT, ir.OBITNOT:
3073 n := n.(*ir.UnaryExpr)
3075 return s.newValue1(s.ssaOp(n.Op(), n.Type()), a.Type, a)
3076 case ir.OIMAG, ir.OREAL:
3077 n := n.(*ir.UnaryExpr)
3079 return s.newValue1(s.ssaOp(n.Op(), n.X.Type()), n.Type(), a)
3081 n := n.(*ir.UnaryExpr)
3085 n := n.(*ir.AddrExpr)
3089 n := n.(*ir.ResultExpr)
3090 if s.prevCall == nil || s.prevCall.Op != ssa.OpStaticLECall && s.prevCall.Op != ssa.OpInterLECall && s.prevCall.Op != ssa.OpClosureLECall {
3091 panic("Expected to see a previous call")
3095 panic(fmt.Errorf("ORESULT %v does not match call %s", n, s.prevCall))
3097 return s.resultOfCall(s.prevCall, which, n.Type())
3100 n := n.(*ir.StarExpr)
3101 p := s.exprPtr(n.X, n.Bounded(), n.Pos())
3102 return s.load(n.Type(), p)
3105 n := n.(*ir.SelectorExpr)
3106 if n.X.Op() == ir.OSTRUCTLIT {
3107 // All literals with nonzero fields have already been
3108 // rewritten during walk. Any that remain are just T{}
3109 // or equivalents. Use the zero value.
3110 if !ir.IsZero(n.X) {
3111 s.Fatalf("literal with nonzero value in SSA: %v", n.X)
3113 return s.zeroVal(n.Type())
3115 // If n is addressable and can't be represented in
3116 // SSA, then load just the selected field. This
3117 // prevents false memory dependencies in race/msan/asan
3119 if ir.IsAddressable(n) && !s.canSSA(n) {
3121 return s.load(n.Type(), p)
3124 return s.newValue1I(ssa.OpStructSelect, n.Type(), int64(fieldIdx(n)), v)
3127 n := n.(*ir.SelectorExpr)
3128 p := s.exprPtr(n.X, n.Bounded(), n.Pos())
3129 p = s.newValue1I(ssa.OpOffPtr, types.NewPtr(n.Type()), n.Offset(), p)
3130 return s.load(n.Type(), p)
3133 n := n.(*ir.IndexExpr)
3135 case n.X.Type().IsString():
3136 if n.Bounded() && ir.IsConst(n.X, constant.String) && ir.IsConst(n.Index, constant.Int) {
3137 // Replace "abc"[1] with 'b'.
3138 // Delayed until now because "abc"[1] is not an ideal constant.
3139 // See test/fixedbugs/issue11370.go.
3140 return s.newValue0I(ssa.OpConst8, types.Types[types.TUINT8], int64(int8(ir.StringVal(n.X)[ir.Int64Val(n.Index)])))
3143 i := s.expr(n.Index)
3144 len := s.newValue1(ssa.OpStringLen, types.Types[types.TINT], a)
3145 i = s.boundsCheck(i, len, ssa.BoundsIndex, n.Bounded())
3146 ptrtyp := s.f.Config.Types.BytePtr
3147 ptr := s.newValue1(ssa.OpStringPtr, ptrtyp, a)
3148 if ir.IsConst(n.Index, constant.Int) {
3149 ptr = s.newValue1I(ssa.OpOffPtr, ptrtyp, ir.Int64Val(n.Index), ptr)
3151 ptr = s.newValue2(ssa.OpAddPtr, ptrtyp, ptr, i)
3153 return s.load(types.Types[types.TUINT8], ptr)
3154 case n.X.Type().IsSlice():
3156 return s.load(n.X.Type().Elem(), p)
3157 case n.X.Type().IsArray():
3158 if TypeOK(n.X.Type()) {
3159 // SSA can handle arrays of length at most 1.
3160 bound := n.X.Type().NumElem()
3162 i := s.expr(n.Index)
3164 // Bounds check will never succeed. Might as well
3165 // use constants for the bounds check.
3166 z := s.constInt(types.Types[types.TINT], 0)
3167 s.boundsCheck(z, z, ssa.BoundsIndex, false)
3168 // The return value won't be live, return junk.
3169 // But not quite junk, in case bounds checks are turned off. See issue 48092.
3170 return s.zeroVal(n.Type())
3172 len := s.constInt(types.Types[types.TINT], bound)
3173 s.boundsCheck(i, len, ssa.BoundsIndex, n.Bounded()) // checks i == 0
3174 return s.newValue1I(ssa.OpArraySelect, n.Type(), 0, a)
3177 return s.load(n.X.Type().Elem(), p)
3179 s.Fatalf("bad type for index %v", n.X.Type())
3183 case ir.OLEN, ir.OCAP:
3184 n := n.(*ir.UnaryExpr)
3186 case n.X.Type().IsSlice():
3187 op := ssa.OpSliceLen
3188 if n.Op() == ir.OCAP {
3191 return s.newValue1(op, types.Types[types.TINT], s.expr(n.X))
3192 case n.X.Type().IsString(): // string; not reachable for OCAP
3193 return s.newValue1(ssa.OpStringLen, types.Types[types.TINT], s.expr(n.X))
3194 case n.X.Type().IsMap(), n.X.Type().IsChan():
3195 return s.referenceTypeBuiltin(n, s.expr(n.X))
3197 return s.constInt(types.Types[types.TINT], n.X.Type().NumElem())
3201 n := n.(*ir.UnaryExpr)
3203 if n.X.Type().IsSlice() {
3205 return s.newValue1(ssa.OpSlicePtr, n.Type(), a)
3207 return s.newValue1(ssa.OpSlicePtrUnchecked, n.Type(), a)
3209 return s.newValue1(ssa.OpStringPtr, n.Type(), a)
3213 n := n.(*ir.UnaryExpr)
3215 return s.newValue1(ssa.OpITab, n.Type(), a)
3218 n := n.(*ir.UnaryExpr)
3220 return s.newValue1(ssa.OpIData, n.Type(), a)
3223 n := n.(*ir.BinaryExpr)
3226 return s.newValue2(ssa.OpIMake, n.Type(), tab, data)
3228 case ir.OSLICEHEADER:
3229 n := n.(*ir.SliceHeaderExpr)
3233 return s.newValue3(ssa.OpSliceMake, n.Type(), p, l, c)
3235 case ir.OSTRINGHEADER:
3236 n := n.(*ir.StringHeaderExpr)
3239 return s.newValue2(ssa.OpStringMake, n.Type(), p, l)
3241 case ir.OSLICE, ir.OSLICEARR, ir.OSLICE3, ir.OSLICE3ARR:
3242 n := n.(*ir.SliceExpr)
3243 check := s.checkPtrEnabled && n.Op() == ir.OSLICE3ARR && n.X.Op() == ir.OCONVNOP && n.X.(*ir.ConvExpr).X.Type().IsUnsafePtr()
3244 v := s.exprCheckPtr(n.X, !check)
3245 var i, j, k *ssa.Value
3255 p, l, c := s.slice(v, i, j, k, n.Bounded())
3257 // Emit checkptr instrumentation after bound check to prevent false positive, see #46938.
3258 s.checkPtrAlignment(n.X.(*ir.ConvExpr), v, s.conv(n.Max, k, k.Type, types.Types[types.TUINTPTR]))
3260 return s.newValue3(ssa.OpSliceMake, n.Type(), p, l, c)
3263 n := n.(*ir.SliceExpr)
3272 p, l, _ := s.slice(v, i, j, nil, n.Bounded())
3273 return s.newValue2(ssa.OpStringMake, n.Type(), p, l)
3275 case ir.OSLICE2ARRPTR:
3276 // if arrlen > slice.len {
3280 n := n.(*ir.ConvExpr)
3282 nelem := n.Type().Elem().NumElem()
3283 arrlen := s.constInt(types.Types[types.TINT], nelem)
3284 cap := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], v)
3285 s.boundsCheck(arrlen, cap, ssa.BoundsConvert, false)
3286 op := ssa.OpSlicePtr
3288 op = ssa.OpSlicePtrUnchecked
3290 return s.newValue1(op, n.Type(), v)
3293 n := n.(*ir.CallExpr)
3294 if ir.IsIntrinsicCall(n) {
3295 return s.intrinsicCall(n)
3300 n := n.(*ir.CallExpr)
3301 return s.callResult(n, callNormal)
3304 n := n.(*ir.CallExpr)
3305 return s.newValue1(ssa.OpGetG, n.Type(), s.mem())
3307 case ir.OGETCALLERPC:
3308 n := n.(*ir.CallExpr)
3309 return s.newValue0(ssa.OpGetCallerPC, n.Type())
3311 case ir.OGETCALLERSP:
3312 n := n.(*ir.CallExpr)
3313 return s.newValue1(ssa.OpGetCallerSP, n.Type(), s.mem())
3316 return s.append(n.(*ir.CallExpr), false)
3318 case ir.OMIN, ir.OMAX:
3319 return s.minMax(n.(*ir.CallExpr))
3321 case ir.OSTRUCTLIT, ir.OARRAYLIT:
3322 // All literals with nonzero fields have already been
3323 // rewritten during walk. Any that remain are just T{}
3324 // or equivalents. Use the zero value.
3325 n := n.(*ir.CompLitExpr)
3327 s.Fatalf("literal with nonzero value in SSA: %v", n)
3329 return s.zeroVal(n.Type())
3332 n := n.(*ir.UnaryExpr)
3333 var rtype *ssa.Value
3334 if x, ok := n.X.(*ir.DynamicType); ok && x.Op() == ir.ODYNAMICTYPE {
3335 rtype = s.expr(x.RType)
3337 return s.newObject(n.Type().Elem(), rtype)
3340 n := n.(*ir.BinaryExpr)
3344 // Force len to uintptr to prevent misuse of garbage bits in the
3345 // upper part of the register (#48536).
3346 len = s.conv(n, len, len.Type, types.Types[types.TUINTPTR])
3348 return s.newValue2(ssa.OpAddPtr, n.Type(), ptr, len)
3351 s.Fatalf("unhandled expr %v", n.Op())
3356 func (s *state) resultOfCall(c *ssa.Value, which int64, t *types.Type) *ssa.Value {
3357 aux := c.Aux.(*ssa.AuxCall)
3358 pa := aux.ParamAssignmentForResult(which)
3359 // TODO(register args) determine if in-memory TypeOK is better loaded early from SelectNAddr or later when SelectN is expanded.
3360 // SelectN is better for pattern-matching and possible call-aware analysis we might want to do in the future.
3361 if len(pa.Registers) == 0 && !TypeOK(t) {
3362 addr := s.newValue1I(ssa.OpSelectNAddr, types.NewPtr(t), which, c)
3363 return s.rawLoad(t, addr)
3365 return s.newValue1I(ssa.OpSelectN, t, which, c)
3368 func (s *state) resultAddrOfCall(c *ssa.Value, which int64, t *types.Type) *ssa.Value {
3369 aux := c.Aux.(*ssa.AuxCall)
3370 pa := aux.ParamAssignmentForResult(which)
3371 if len(pa.Registers) == 0 {
3372 return s.newValue1I(ssa.OpSelectNAddr, types.NewPtr(t), which, c)
3374 _, addr := s.temp(c.Pos, t)
3375 rval := s.newValue1I(ssa.OpSelectN, t, which, c)
3376 s.vars[memVar] = s.newValue3Apos(ssa.OpStore, types.TypeMem, t, addr, rval, s.mem(), false)
3380 // append converts an OAPPEND node to SSA.
3381 // If inplace is false, it converts the OAPPEND expression n to an ssa.Value,
3382 // adds it to s, and returns the Value.
3383 // If inplace is true, it writes the result of the OAPPEND expression n
3384 // back to the slice being appended to, and returns nil.
3385 // inplace MUST be set to false if the slice can be SSA'd.
3386 // Note: this code only handles fixed-count appends. Dotdotdot appends
3387 // have already been rewritten at this point (by walk).
3388 func (s *state) append(n *ir.CallExpr, inplace bool) *ssa.Value {
3389 // If inplace is false, process as expression "append(s, e1, e2, e3)":
3391 // ptr, len, cap := s
3393 // if uint(len) > uint(cap) {
3394 // ptr, len, cap = growslice(ptr, len, cap, 3, typ)
3395 // Note that len is unmodified by growslice.
3397 // // with write barriers, if needed:
3398 // *(ptr+(len-3)) = e1
3399 // *(ptr+(len-2)) = e2
3400 // *(ptr+(len-1)) = e3
3401 // return makeslice(ptr, len, cap)
3404 // If inplace is true, process as statement "s = append(s, e1, e2, e3)":
3407 // ptr, len, cap := s
3409 // if uint(len) > uint(cap) {
3410 // ptr, len, cap = growslice(ptr, len, cap, 3, typ)
3411 // vardef(a) // if necessary, advise liveness we are writing a new a
3412 // *a.cap = cap // write before ptr to avoid a spill
3413 // *a.ptr = ptr // with write barrier
3416 // // with write barriers, if needed:
3417 // *(ptr+(len-3)) = e1
3418 // *(ptr+(len-2)) = e2
3419 // *(ptr+(len-1)) = e3
3421 et := n.Type().Elem()
3422 pt := types.NewPtr(et)
3425 sn := n.Args[0] // the slice node is the first in the list
3426 var slice, addr *ssa.Value
3429 slice = s.load(n.Type(), addr)
3434 // Allocate new blocks
3435 grow := s.f.NewBlock(ssa.BlockPlain)
3436 assign := s.f.NewBlock(ssa.BlockPlain)
3438 // Decomposse input slice.
3439 p := s.newValue1(ssa.OpSlicePtr, pt, slice)
3440 l := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], slice)
3441 c := s.newValue1(ssa.OpSliceCap, types.Types[types.TINT], slice)
3443 // Add number of new elements to length.
3444 nargs := s.constInt(types.Types[types.TINT], int64(len(n.Args)-1))
3445 l = s.newValue2(s.ssaOp(ir.OADD, types.Types[types.TINT]), types.Types[types.TINT], l, nargs)
3447 // Decide if we need to grow
3448 cmp := s.newValue2(s.ssaOp(ir.OLT, types.Types[types.TUINT]), types.Types[types.TBOOL], c, l)
3450 // Record values of ptr/len/cap before branch.
3458 b.Kind = ssa.BlockIf
3459 b.Likely = ssa.BranchUnlikely
3466 taddr := s.expr(n.X)
3467 r := s.rtcall(ir.Syms.Growslice, true, []*types.Type{n.Type()}, p, l, c, nargs, taddr)
3469 // Decompose output slice
3470 p = s.newValue1(ssa.OpSlicePtr, pt, r[0])
3471 l = s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], r[0])
3472 c = s.newValue1(ssa.OpSliceCap, types.Types[types.TINT], r[0])
3478 if sn.Op() == ir.ONAME {
3480 if sn.Class != ir.PEXTERN {
3481 // Tell liveness we're about to build a new slice
3482 s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, sn, s.mem())
3485 capaddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, types.SliceCapOffset, addr)
3486 s.store(types.Types[types.TINT], capaddr, c)
3487 s.store(pt, addr, p)
3493 // assign new elements to slots
3494 s.startBlock(assign)
3495 p = s.variable(ptrVar, pt) // generates phi for ptr
3496 l = s.variable(lenVar, types.Types[types.TINT]) // generates phi for len
3498 c = s.variable(capVar, types.Types[types.TINT]) // generates phi for cap
3502 // Update length in place.
3503 // We have to wait until here to make sure growslice succeeded.
3504 lenaddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, types.SliceLenOffset, addr)
3505 s.store(types.Types[types.TINT], lenaddr, l)
3509 type argRec struct {
3510 // if store is true, we're appending the value v. If false, we're appending the
3515 args := make([]argRec, 0, len(n.Args[1:]))
3516 for _, n := range n.Args[1:] {
3517 if TypeOK(n.Type()) {
3518 args = append(args, argRec{v: s.expr(n), store: true})
3521 args = append(args, argRec{v: v})
3525 // Write args into slice.
3526 oldLen := s.newValue2(s.ssaOp(ir.OSUB, types.Types[types.TINT]), types.Types[types.TINT], l, nargs)
3527 p2 := s.newValue2(ssa.OpPtrIndex, pt, p, oldLen)
3528 for i, arg := range args {
3529 addr := s.newValue2(ssa.OpPtrIndex, pt, p2, s.constInt(types.Types[types.TINT], int64(i)))
3531 s.storeType(et, addr, arg.v, 0, true)
3533 s.move(et, addr, arg.v)
3537 // The following deletions have no practical effect at this time
3538 // because state.vars has been reset by the preceding state.startBlock.
3539 // They only enforce the fact that these variables are no longer need in
3540 // the current scope.
3541 delete(s.vars, ptrVar)
3542 delete(s.vars, lenVar)
3544 delete(s.vars, capVar)
3551 return s.newValue3(ssa.OpSliceMake, n.Type(), p, l, c)
3554 // minMax converts an OMIN/OMAX builtin call into SSA.
3555 func (s *state) minMax(n *ir.CallExpr) *ssa.Value {
3556 // The OMIN/OMAX builtin is variadic, but its semantics are
3557 // equivalent to left-folding a binary min/max operation across the
3559 fold := func(op func(x, a *ssa.Value) *ssa.Value) *ssa.Value {
3560 x := s.expr(n.Args[0])
3561 for _, arg := range n.Args[1:] {
3562 x = op(x, s.expr(arg))
3569 if typ.IsFloat() || typ.IsString() {
3570 // min/max semantics for floats are tricky because of NaNs and
3571 // negative zero, so we let the runtime handle this instead.
3573 // Strings are conceptually simpler, but we currently desugar
3574 // string comparisons during walk, not ssagen.
3578 case types.TFLOAT32:
3585 case types.TFLOAT64:
3600 fn := typecheck.LookupRuntimeFunc(name)
3602 return fold(func(x, a *ssa.Value) *ssa.Value {
3603 return s.rtcall(fn, true, []*types.Type{typ}, x, a)[0]
3607 lt := s.ssaOp(ir.OLT, typ)
3609 return fold(func(x, a *ssa.Value) *ssa.Value {
3613 return s.ternary(s.newValue2(lt, types.Types[types.TBOOL], a, x), a, x)
3616 return s.ternary(s.newValue2(lt, types.Types[types.TBOOL], x, a), a, x)
3618 panic("unreachable")
3622 // ternary emits code to evaluate cond ? x : y.
3623 func (s *state) ternary(cond, x, y *ssa.Value) *ssa.Value {
3624 bThen := s.f.NewBlock(ssa.BlockPlain)
3625 bElse := s.f.NewBlock(ssa.BlockPlain)
3626 bEnd := s.f.NewBlock(ssa.BlockPlain)
3629 b.Kind = ssa.BlockIf
3635 s.vars[ternaryVar] = x
3636 s.endBlock().AddEdgeTo(bEnd)
3639 s.vars[ternaryVar] = y
3640 s.endBlock().AddEdgeTo(bEnd)
3643 r := s.variable(ternaryVar, x.Type)
3644 delete(s.vars, ternaryVar)
3648 // condBranch evaluates the boolean expression cond and branches to yes
3649 // if cond is true and no if cond is false.
3650 // This function is intended to handle && and || better than just calling
3651 // s.expr(cond) and branching on the result.
3652 func (s *state) condBranch(cond ir.Node, yes, no *ssa.Block, likely int8) {
3655 cond := cond.(*ir.LogicalExpr)
3656 mid := s.f.NewBlock(ssa.BlockPlain)
3657 s.stmtList(cond.Init())
3658 s.condBranch(cond.X, mid, no, max8(likely, 0))
3660 s.condBranch(cond.Y, yes, no, likely)
3662 // Note: if likely==1, then both recursive calls pass 1.
3663 // If likely==-1, then we don't have enough information to decide
3664 // whether the first branch is likely or not. So we pass 0 for
3665 // the likeliness of the first branch.
3666 // TODO: have the frontend give us branch prediction hints for
3667 // OANDAND and OOROR nodes (if it ever has such info).
3669 cond := cond.(*ir.LogicalExpr)
3670 mid := s.f.NewBlock(ssa.BlockPlain)
3671 s.stmtList(cond.Init())
3672 s.condBranch(cond.X, yes, mid, min8(likely, 0))
3674 s.condBranch(cond.Y, yes, no, likely)
3676 // Note: if likely==-1, then both recursive calls pass -1.
3677 // If likely==1, then we don't have enough info to decide
3678 // the likelihood of the first branch.
3680 cond := cond.(*ir.UnaryExpr)
3681 s.stmtList(cond.Init())
3682 s.condBranch(cond.X, no, yes, -likely)
3685 cond := cond.(*ir.ConvExpr)
3686 s.stmtList(cond.Init())
3687 s.condBranch(cond.X, yes, no, likely)
3692 b.Kind = ssa.BlockIf
3694 b.Likely = ssa.BranchPrediction(likely) // gc and ssa both use -1/0/+1 for likeliness
3702 skipPtr skipMask = 1 << iota
3707 // assign does left = right.
3708 // Right has already been evaluated to ssa, left has not.
3709 // If deref is true, then we do left = *right instead (and right has already been nil-checked).
3710 // If deref is true and right == nil, just do left = 0.
3711 // skip indicates assignments (at the top level) that can be avoided.
3712 // mayOverlap indicates whether left&right might partially overlap in memory. Default is false.
3713 func (s *state) assign(left ir.Node, right *ssa.Value, deref bool, skip skipMask) {
3714 s.assignWhichMayOverlap(left, right, deref, skip, false)
3716 func (s *state) assignWhichMayOverlap(left ir.Node, right *ssa.Value, deref bool, skip skipMask, mayOverlap bool) {
3717 if left.Op() == ir.ONAME && ir.IsBlank(left) {
3724 s.Fatalf("can SSA LHS %v but not RHS %s", left, right)
3726 if left.Op() == ir.ODOT {
3727 // We're assigning to a field of an ssa-able value.
3728 // We need to build a new structure with the new value for the
3729 // field we're assigning and the old values for the other fields.
3731 // type T struct {a, b, c int}
3734 // For the x.b = 5 assignment we want to generate x = T{x.a, 5, x.c}
3736 // Grab information about the structure type.
3737 left := left.(*ir.SelectorExpr)
3740 idx := fieldIdx(left)
3742 // Grab old value of structure.
3743 old := s.expr(left.X)
3745 // Make new structure.
3746 new := s.newValue0(ssa.StructMakeOp(t.NumFields()), t)
3748 // Add fields as args.
3749 for i := 0; i < nf; i++ {
3753 new.AddArg(s.newValue1I(ssa.OpStructSelect, t.FieldType(i), int64(i), old))
3757 // Recursively assign the new value we've made to the base of the dot op.
3758 s.assign(left.X, new, false, 0)
3759 // TODO: do we need to update named values here?
3762 if left.Op() == ir.OINDEX && left.(*ir.IndexExpr).X.Type().IsArray() {
3763 left := left.(*ir.IndexExpr)
3764 s.pushLine(left.Pos())
3766 // We're assigning to an element of an ssa-able array.
3771 i := s.expr(left.Index) // index
3773 // The bounds check must fail. Might as well
3774 // ignore the actual index and just use zeros.
3775 z := s.constInt(types.Types[types.TINT], 0)
3776 s.boundsCheck(z, z, ssa.BoundsIndex, false)
3780 s.Fatalf("assigning to non-1-length array")
3782 // Rewrite to a = [1]{v}
3783 len := s.constInt(types.Types[types.TINT], 1)
3784 s.boundsCheck(i, len, ssa.BoundsIndex, false) // checks i == 0
3785 v := s.newValue1(ssa.OpArrayMake1, t, right)
3786 s.assign(left.X, v, false, 0)
3789 left := left.(*ir.Name)
3790 // Update variable assignment.
3791 s.vars[left] = right
3792 s.addNamedValue(left, right)
3796 // If this assignment clobbers an entire local variable, then emit
3797 // OpVarDef so liveness analysis knows the variable is redefined.
3798 if base, ok := clobberBase(left).(*ir.Name); ok && base.OnStack() && skip == 0 && t.HasPointers() {
3799 s.vars[memVar] = s.newValue1Apos(ssa.OpVarDef, types.TypeMem, base, s.mem(), !ir.IsAutoTmp(base))
3802 // Left is not ssa-able. Compute its address.
3803 addr := s.addr(left)
3804 if ir.IsReflectHeaderDataField(left) {
3805 // Package unsafe's documentation says storing pointers into
3806 // reflect.SliceHeader and reflect.StringHeader's Data fields
3807 // is valid, even though they have type uintptr (#19168).
3808 // Mark it pointer type to signal the writebarrier pass to
3809 // insert a write barrier.
3810 t = types.Types[types.TUNSAFEPTR]
3813 // Treat as a mem->mem move.
3817 s.moveWhichMayOverlap(t, addr, right, mayOverlap)
3821 // Treat as a store.
3822 s.storeType(t, addr, right, skip, !ir.IsAutoTmp(left))
3825 // zeroVal returns the zero value for type t.
3826 func (s *state) zeroVal(t *types.Type) *ssa.Value {
3831 return s.constInt8(t, 0)
3833 return s.constInt16(t, 0)
3835 return s.constInt32(t, 0)
3837 return s.constInt64(t, 0)
3839 s.Fatalf("bad sized integer type %v", t)
3844 return s.constFloat32(t, 0)
3846 return s.constFloat64(t, 0)
3848 s.Fatalf("bad sized float type %v", t)
3853 z := s.constFloat32(types.Types[types.TFLOAT32], 0)
3854 return s.entryNewValue2(ssa.OpComplexMake, t, z, z)
3856 z := s.constFloat64(types.Types[types.TFLOAT64], 0)
3857 return s.entryNewValue2(ssa.OpComplexMake, t, z, z)
3859 s.Fatalf("bad sized complex type %v", t)
3863 return s.constEmptyString(t)
3864 case t.IsPtrShaped():
3865 return s.constNil(t)
3867 return s.constBool(false)
3868 case t.IsInterface():
3869 return s.constInterface(t)
3871 return s.constSlice(t)
3874 v := s.entryNewValue0(ssa.StructMakeOp(t.NumFields()), t)
3875 for i := 0; i < n; i++ {
3876 v.AddArg(s.zeroVal(t.FieldType(i)))
3880 switch t.NumElem() {
3882 return s.entryNewValue0(ssa.OpArrayMake0, t)
3884 return s.entryNewValue1(ssa.OpArrayMake1, t, s.zeroVal(t.Elem()))
3887 s.Fatalf("zero for type %v not implemented", t)
3894 callNormal callKind = iota
3901 type sfRtCallDef struct {
3906 var softFloatOps map[ssa.Op]sfRtCallDef
3908 func softfloatInit() {
3909 // Some of these operations get transformed by sfcall.
3910 softFloatOps = map[ssa.Op]sfRtCallDef{
3911 ssa.OpAdd32F: {typecheck.LookupRuntimeFunc("fadd32"), types.TFLOAT32},
3912 ssa.OpAdd64F: {typecheck.LookupRuntimeFunc("fadd64"), types.TFLOAT64},
3913 ssa.OpSub32F: {typecheck.LookupRuntimeFunc("fadd32"), types.TFLOAT32},
3914 ssa.OpSub64F: {typecheck.LookupRuntimeFunc("fadd64"), types.TFLOAT64},
3915 ssa.OpMul32F: {typecheck.LookupRuntimeFunc("fmul32"), types.TFLOAT32},
3916 ssa.OpMul64F: {typecheck.LookupRuntimeFunc("fmul64"), types.TFLOAT64},
3917 ssa.OpDiv32F: {typecheck.LookupRuntimeFunc("fdiv32"), types.TFLOAT32},
3918 ssa.OpDiv64F: {typecheck.LookupRuntimeFunc("fdiv64"), types.TFLOAT64},
3920 ssa.OpEq64F: {typecheck.LookupRuntimeFunc("feq64"), types.TBOOL},
3921 ssa.OpEq32F: {typecheck.LookupRuntimeFunc("feq32"), types.TBOOL},
3922 ssa.OpNeq64F: {typecheck.LookupRuntimeFunc("feq64"), types.TBOOL},
3923 ssa.OpNeq32F: {typecheck.LookupRuntimeFunc("feq32"), types.TBOOL},
3924 ssa.OpLess64F: {typecheck.LookupRuntimeFunc("fgt64"), types.TBOOL},
3925 ssa.OpLess32F: {typecheck.LookupRuntimeFunc("fgt32"), types.TBOOL},
3926 ssa.OpLeq64F: {typecheck.LookupRuntimeFunc("fge64"), types.TBOOL},
3927 ssa.OpLeq32F: {typecheck.LookupRuntimeFunc("fge32"), types.TBOOL},
3929 ssa.OpCvt32to32F: {typecheck.LookupRuntimeFunc("fint32to32"), types.TFLOAT32},
3930 ssa.OpCvt32Fto32: {typecheck.LookupRuntimeFunc("f32toint32"), types.TINT32},
3931 ssa.OpCvt64to32F: {typecheck.LookupRuntimeFunc("fint64to32"), types.TFLOAT32},
3932 ssa.OpCvt32Fto64: {typecheck.LookupRuntimeFunc("f32toint64"), types.TINT64},
3933 ssa.OpCvt64Uto32F: {typecheck.LookupRuntimeFunc("fuint64to32"), types.TFLOAT32},
3934 ssa.OpCvt32Fto64U: {typecheck.LookupRuntimeFunc("f32touint64"), types.TUINT64},
3935 ssa.OpCvt32to64F: {typecheck.LookupRuntimeFunc("fint32to64"), types.TFLOAT64},
3936 ssa.OpCvt64Fto32: {typecheck.LookupRuntimeFunc("f64toint32"), types.TINT32},
3937 ssa.OpCvt64to64F: {typecheck.LookupRuntimeFunc("fint64to64"), types.TFLOAT64},
3938 ssa.OpCvt64Fto64: {typecheck.LookupRuntimeFunc("f64toint64"), types.TINT64},
3939 ssa.OpCvt64Uto64F: {typecheck.LookupRuntimeFunc("fuint64to64"), types.TFLOAT64},
3940 ssa.OpCvt64Fto64U: {typecheck.LookupRuntimeFunc("f64touint64"), types.TUINT64},
3941 ssa.OpCvt32Fto64F: {typecheck.LookupRuntimeFunc("f32to64"), types.TFLOAT64},
3942 ssa.OpCvt64Fto32F: {typecheck.LookupRuntimeFunc("f64to32"), types.TFLOAT32},
3946 // TODO: do not emit sfcall if operation can be optimized to constant in later
3948 func (s *state) sfcall(op ssa.Op, args ...*ssa.Value) (*ssa.Value, bool) {
3949 f2i := func(t *types.Type) *types.Type {
3951 case types.TFLOAT32:
3952 return types.Types[types.TUINT32]
3953 case types.TFLOAT64:
3954 return types.Types[types.TUINT64]
3959 if callDef, ok := softFloatOps[op]; ok {
3965 args[0], args[1] = args[1], args[0]
3968 args[1] = s.newValue1(s.ssaOp(ir.ONEG, types.Types[callDef.rtype]), args[1].Type, args[1])
3971 // runtime functions take uints for floats and returns uints.
3972 // Convert to uints so we use the right calling convention.
3973 for i, a := range args {
3974 if a.Type.IsFloat() {
3975 args[i] = s.newValue1(ssa.OpCopy, f2i(a.Type), a)
3979 rt := types.Types[callDef.rtype]
3980 result := s.rtcall(callDef.rtfn, true, []*types.Type{f2i(rt)}, args...)[0]
3982 result = s.newValue1(ssa.OpCopy, rt, result)
3984 if op == ssa.OpNeq32F || op == ssa.OpNeq64F {
3985 result = s.newValue1(ssa.OpNot, result.Type, result)
3992 var intrinsics map[intrinsicKey]intrinsicBuilder
3994 // An intrinsicBuilder converts a call node n into an ssa value that
3995 // implements that call as an intrinsic. args is a list of arguments to the func.
3996 type intrinsicBuilder func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value
3998 type intrinsicKey struct {
4005 intrinsics = map[intrinsicKey]intrinsicBuilder{}
4010 var lwatomics []*sys.Arch
4011 for _, a := range &sys.Archs {
4012 all = append(all, a)
4018 if a.Family != sys.PPC64 {
4019 lwatomics = append(lwatomics, a)
4023 // add adds the intrinsic b for pkg.fn for the given list of architectures.
4024 add := func(pkg, fn string, b intrinsicBuilder, archs ...*sys.Arch) {
4025 for _, a := range archs {
4026 intrinsics[intrinsicKey{a, pkg, fn}] = b
4029 // addF does the same as add but operates on architecture families.
4030 addF := func(pkg, fn string, b intrinsicBuilder, archFamilies ...sys.ArchFamily) {
4032 for _, f := range archFamilies {
4034 panic("too many architecture families")
4038 for _, a := range all {
4039 if m>>uint(a.Family)&1 != 0 {
4040 intrinsics[intrinsicKey{a, pkg, fn}] = b
4044 // alias defines pkg.fn = pkg2.fn2 for all architectures in archs for which pkg2.fn2 exists.
4045 alias := func(pkg, fn, pkg2, fn2 string, archs ...*sys.Arch) {
4047 for _, a := range archs {
4048 if b, ok := intrinsics[intrinsicKey{a, pkg2, fn2}]; ok {
4049 intrinsics[intrinsicKey{a, pkg, fn}] = b
4054 panic(fmt.Sprintf("attempted to alias undefined intrinsic: %s.%s", pkg, fn))
4058 /******** runtime ********/
4059 if !base.Flag.Cfg.Instrumenting {
4060 add("runtime", "slicebytetostringtmp",
4061 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4062 // Compiler frontend optimizations emit OBYTES2STRTMP nodes
4063 // for the backend instead of slicebytetostringtmp calls
4064 // when not instrumenting.
4065 return s.newValue2(ssa.OpStringMake, n.Type(), args[0], args[1])
4069 addF("runtime/internal/math", "MulUintptr",
4070 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4071 if s.config.PtrSize == 4 {
4072 return s.newValue2(ssa.OpMul32uover, types.NewTuple(types.Types[types.TUINT], types.Types[types.TUINT]), args[0], args[1])
4074 return s.newValue2(ssa.OpMul64uover, types.NewTuple(types.Types[types.TUINT], types.Types[types.TUINT]), args[0], args[1])
4076 sys.AMD64, sys.I386, sys.Loong64, sys.MIPS64, sys.RISCV64, sys.ARM64)
4077 alias("runtime", "mulUintptr", "runtime/internal/math", "MulUintptr", all...)
4078 add("runtime", "KeepAlive",
4079 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4080 data := s.newValue1(ssa.OpIData, s.f.Config.Types.BytePtr, args[0])
4081 s.vars[memVar] = s.newValue2(ssa.OpKeepAlive, types.TypeMem, data, s.mem())
4085 add("runtime", "getclosureptr",
4086 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4087 return s.newValue0(ssa.OpGetClosurePtr, s.f.Config.Types.Uintptr)
4091 add("runtime", "getcallerpc",
4092 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4093 return s.newValue0(ssa.OpGetCallerPC, s.f.Config.Types.Uintptr)
4097 add("runtime", "getcallersp",
4098 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4099 return s.newValue1(ssa.OpGetCallerSP, s.f.Config.Types.Uintptr, s.mem())
4103 addF("runtime", "publicationBarrier",
4104 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4105 s.vars[memVar] = s.newValue1(ssa.OpPubBarrier, types.TypeMem, s.mem())
4108 sys.ARM64, sys.PPC64)
4110 brev_arch := []sys.ArchFamily{sys.AMD64, sys.I386, sys.ARM64, sys.ARM, sys.S390X}
4111 if buildcfg.GOPPC64 >= 10 {
4112 // Use only on Power10 as the new byte reverse instructions that Power10 provide
4113 // make it worthwhile as an intrinsic
4114 brev_arch = append(brev_arch, sys.PPC64)
4116 /******** runtime/internal/sys ********/
4117 addF("runtime/internal/sys", "Bswap32",
4118 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4119 return s.newValue1(ssa.OpBswap32, types.Types[types.TUINT32], args[0])
4122 addF("runtime/internal/sys", "Bswap64",
4123 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4124 return s.newValue1(ssa.OpBswap64, types.Types[types.TUINT64], args[0])
4128 /****** Prefetch ******/
4129 makePrefetchFunc := func(op ssa.Op) func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4130 return func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4131 s.vars[memVar] = s.newValue2(op, types.TypeMem, args[0], s.mem())
4136 // Make Prefetch intrinsics for supported platforms
4137 // On the unsupported platforms stub function will be eliminated
4138 addF("runtime/internal/sys", "Prefetch", makePrefetchFunc(ssa.OpPrefetchCache),
4139 sys.AMD64, sys.ARM64, sys.PPC64)
4140 addF("runtime/internal/sys", "PrefetchStreamed", makePrefetchFunc(ssa.OpPrefetchCacheStreamed),
4141 sys.AMD64, sys.ARM64, sys.PPC64)
4143 /******** runtime/internal/atomic ********/
4144 addF("runtime/internal/atomic", "Load",
4145 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4146 v := s.newValue2(ssa.OpAtomicLoad32, types.NewTuple(types.Types[types.TUINT32], types.TypeMem), args[0], s.mem())
4147 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4148 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT32], v)
4150 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4151 addF("runtime/internal/atomic", "Load8",
4152 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4153 v := s.newValue2(ssa.OpAtomicLoad8, types.NewTuple(types.Types[types.TUINT8], types.TypeMem), args[0], s.mem())
4154 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4155 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT8], v)
4157 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4158 addF("runtime/internal/atomic", "Load64",
4159 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4160 v := s.newValue2(ssa.OpAtomicLoad64, types.NewTuple(types.Types[types.TUINT64], types.TypeMem), args[0], s.mem())
4161 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4162 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT64], v)
4164 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4165 addF("runtime/internal/atomic", "LoadAcq",
4166 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4167 v := s.newValue2(ssa.OpAtomicLoadAcq32, types.NewTuple(types.Types[types.TUINT32], types.TypeMem), args[0], s.mem())
4168 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4169 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT32], v)
4171 sys.PPC64, sys.S390X)
4172 addF("runtime/internal/atomic", "LoadAcq64",
4173 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4174 v := s.newValue2(ssa.OpAtomicLoadAcq64, types.NewTuple(types.Types[types.TUINT64], types.TypeMem), args[0], s.mem())
4175 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4176 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT64], v)
4179 addF("runtime/internal/atomic", "Loadp",
4180 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4181 v := s.newValue2(ssa.OpAtomicLoadPtr, types.NewTuple(s.f.Config.Types.BytePtr, types.TypeMem), args[0], s.mem())
4182 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4183 return s.newValue1(ssa.OpSelect0, s.f.Config.Types.BytePtr, v)
4185 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4187 addF("runtime/internal/atomic", "Store",
4188 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4189 s.vars[memVar] = s.newValue3(ssa.OpAtomicStore32, types.TypeMem, args[0], args[1], s.mem())
4192 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4193 addF("runtime/internal/atomic", "Store8",
4194 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4195 s.vars[memVar] = s.newValue3(ssa.OpAtomicStore8, types.TypeMem, args[0], args[1], s.mem())
4198 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4199 addF("runtime/internal/atomic", "Store64",
4200 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4201 s.vars[memVar] = s.newValue3(ssa.OpAtomicStore64, types.TypeMem, args[0], args[1], s.mem())
4204 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4205 addF("runtime/internal/atomic", "StorepNoWB",
4206 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4207 s.vars[memVar] = s.newValue3(ssa.OpAtomicStorePtrNoWB, types.TypeMem, args[0], args[1], s.mem())
4210 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.RISCV64, sys.S390X)
4211 addF("runtime/internal/atomic", "StoreRel",
4212 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4213 s.vars[memVar] = s.newValue3(ssa.OpAtomicStoreRel32, types.TypeMem, args[0], args[1], s.mem())
4216 sys.PPC64, sys.S390X)
4217 addF("runtime/internal/atomic", "StoreRel64",
4218 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4219 s.vars[memVar] = s.newValue3(ssa.OpAtomicStoreRel64, types.TypeMem, args[0], args[1], s.mem())
4224 addF("runtime/internal/atomic", "Xchg",
4225 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4226 v := s.newValue3(ssa.OpAtomicExchange32, types.NewTuple(types.Types[types.TUINT32], types.TypeMem), args[0], args[1], s.mem())
4227 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4228 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT32], v)
4230 sys.AMD64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4231 addF("runtime/internal/atomic", "Xchg64",
4232 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4233 v := s.newValue3(ssa.OpAtomicExchange64, types.NewTuple(types.Types[types.TUINT64], types.TypeMem), args[0], args[1], s.mem())
4234 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4235 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT64], v)
4237 sys.AMD64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4239 type atomicOpEmitter func(s *state, n *ir.CallExpr, args []*ssa.Value, op ssa.Op, typ types.Kind)
4241 makeAtomicGuardedIntrinsicARM64 := func(op0, op1 ssa.Op, typ, rtyp types.Kind, emit atomicOpEmitter) intrinsicBuilder {
4243 return func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4244 // Target Atomic feature is identified by dynamic detection
4245 addr := s.entryNewValue1A(ssa.OpAddr, types.Types[types.TBOOL].PtrTo(), ir.Syms.ARM64HasATOMICS, s.sb)
4246 v := s.load(types.Types[types.TBOOL], addr)
4248 b.Kind = ssa.BlockIf
4250 bTrue := s.f.NewBlock(ssa.BlockPlain)
4251 bFalse := s.f.NewBlock(ssa.BlockPlain)
4252 bEnd := s.f.NewBlock(ssa.BlockPlain)
4255 b.Likely = ssa.BranchLikely
4257 // We have atomic instructions - use it directly.
4259 emit(s, n, args, op1, typ)
4260 s.endBlock().AddEdgeTo(bEnd)
4262 // Use original instruction sequence.
4263 s.startBlock(bFalse)
4264 emit(s, n, args, op0, typ)
4265 s.endBlock().AddEdgeTo(bEnd)
4269 if rtyp == types.TNIL {
4272 return s.variable(n, types.Types[rtyp])
4277 atomicXchgXaddEmitterARM64 := func(s *state, n *ir.CallExpr, args []*ssa.Value, op ssa.Op, typ types.Kind) {
4278 v := s.newValue3(op, types.NewTuple(types.Types[typ], types.TypeMem), args[0], args[1], s.mem())
4279 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4280 s.vars[n] = s.newValue1(ssa.OpSelect0, types.Types[typ], v)
4282 addF("runtime/internal/atomic", "Xchg",
4283 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicExchange32, ssa.OpAtomicExchange32Variant, types.TUINT32, types.TUINT32, atomicXchgXaddEmitterARM64),
4285 addF("runtime/internal/atomic", "Xchg64",
4286 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicExchange64, ssa.OpAtomicExchange64Variant, types.TUINT64, types.TUINT64, atomicXchgXaddEmitterARM64),
4289 addF("runtime/internal/atomic", "Xadd",
4290 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4291 v := s.newValue3(ssa.OpAtomicAdd32, types.NewTuple(types.Types[types.TUINT32], types.TypeMem), args[0], args[1], s.mem())
4292 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4293 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT32], v)
4295 sys.AMD64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4296 addF("runtime/internal/atomic", "Xadd64",
4297 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4298 v := s.newValue3(ssa.OpAtomicAdd64, types.NewTuple(types.Types[types.TUINT64], types.TypeMem), args[0], args[1], s.mem())
4299 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4300 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT64], v)
4302 sys.AMD64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4304 addF("runtime/internal/atomic", "Xadd",
4305 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAdd32, ssa.OpAtomicAdd32Variant, types.TUINT32, types.TUINT32, atomicXchgXaddEmitterARM64),
4307 addF("runtime/internal/atomic", "Xadd64",
4308 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAdd64, ssa.OpAtomicAdd64Variant, types.TUINT64, types.TUINT64, atomicXchgXaddEmitterARM64),
4311 addF("runtime/internal/atomic", "Cas",
4312 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4313 v := s.newValue4(ssa.OpAtomicCompareAndSwap32, types.NewTuple(types.Types[types.TBOOL], types.TypeMem), args[0], args[1], args[2], s.mem())
4314 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4315 return s.newValue1(ssa.OpSelect0, types.Types[types.TBOOL], v)
4317 sys.AMD64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4318 addF("runtime/internal/atomic", "Cas64",
4319 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4320 v := s.newValue4(ssa.OpAtomicCompareAndSwap64, types.NewTuple(types.Types[types.TBOOL], types.TypeMem), args[0], args[1], args[2], s.mem())
4321 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4322 return s.newValue1(ssa.OpSelect0, types.Types[types.TBOOL], v)
4324 sys.AMD64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4325 addF("runtime/internal/atomic", "CasRel",
4326 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4327 v := s.newValue4(ssa.OpAtomicCompareAndSwap32, types.NewTuple(types.Types[types.TBOOL], types.TypeMem), args[0], args[1], args[2], s.mem())
4328 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4329 return s.newValue1(ssa.OpSelect0, types.Types[types.TBOOL], v)
4333 atomicCasEmitterARM64 := func(s *state, n *ir.CallExpr, args []*ssa.Value, op ssa.Op, typ types.Kind) {
4334 v := s.newValue4(op, types.NewTuple(types.Types[types.TBOOL], types.TypeMem), args[0], args[1], args[2], s.mem())
4335 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4336 s.vars[n] = s.newValue1(ssa.OpSelect0, types.Types[typ], v)
4339 addF("runtime/internal/atomic", "Cas",
4340 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicCompareAndSwap32, ssa.OpAtomicCompareAndSwap32Variant, types.TUINT32, types.TBOOL, atomicCasEmitterARM64),
4342 addF("runtime/internal/atomic", "Cas64",
4343 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicCompareAndSwap64, ssa.OpAtomicCompareAndSwap64Variant, types.TUINT64, types.TBOOL, atomicCasEmitterARM64),
4346 addF("runtime/internal/atomic", "And8",
4347 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4348 s.vars[memVar] = s.newValue3(ssa.OpAtomicAnd8, types.TypeMem, args[0], args[1], s.mem())
4351 sys.AMD64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4352 addF("runtime/internal/atomic", "And",
4353 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4354 s.vars[memVar] = s.newValue3(ssa.OpAtomicAnd32, types.TypeMem, args[0], args[1], s.mem())
4357 sys.AMD64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4358 addF("runtime/internal/atomic", "Or8",
4359 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4360 s.vars[memVar] = s.newValue3(ssa.OpAtomicOr8, types.TypeMem, args[0], args[1], s.mem())
4363 sys.AMD64, sys.ARM64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4364 addF("runtime/internal/atomic", "Or",
4365 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4366 s.vars[memVar] = s.newValue3(ssa.OpAtomicOr32, types.TypeMem, args[0], args[1], s.mem())
4369 sys.AMD64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4371 atomicAndOrEmitterARM64 := func(s *state, n *ir.CallExpr, args []*ssa.Value, op ssa.Op, typ types.Kind) {
4372 s.vars[memVar] = s.newValue3(op, types.TypeMem, args[0], args[1], s.mem())
4375 addF("runtime/internal/atomic", "And8",
4376 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAnd8, ssa.OpAtomicAnd8Variant, types.TNIL, types.TNIL, atomicAndOrEmitterARM64),
4378 addF("runtime/internal/atomic", "And",
4379 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAnd32, ssa.OpAtomicAnd32Variant, types.TNIL, types.TNIL, atomicAndOrEmitterARM64),
4381 addF("runtime/internal/atomic", "Or8",
4382 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicOr8, ssa.OpAtomicOr8Variant, types.TNIL, types.TNIL, atomicAndOrEmitterARM64),
4384 addF("runtime/internal/atomic", "Or",
4385 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicOr32, ssa.OpAtomicOr32Variant, types.TNIL, types.TNIL, atomicAndOrEmitterARM64),
4388 // Aliases for atomic load operations
4389 alias("runtime/internal/atomic", "Loadint32", "runtime/internal/atomic", "Load", all...)
4390 alias("runtime/internal/atomic", "Loadint64", "runtime/internal/atomic", "Load64", all...)
4391 alias("runtime/internal/atomic", "Loaduintptr", "runtime/internal/atomic", "Load", p4...)
4392 alias("runtime/internal/atomic", "Loaduintptr", "runtime/internal/atomic", "Load64", p8...)
4393 alias("runtime/internal/atomic", "Loaduint", "runtime/internal/atomic", "Load", p4...)
4394 alias("runtime/internal/atomic", "Loaduint", "runtime/internal/atomic", "Load64", p8...)
4395 alias("runtime/internal/atomic", "LoadAcq", "runtime/internal/atomic", "Load", lwatomics...)
4396 alias("runtime/internal/atomic", "LoadAcq64", "runtime/internal/atomic", "Load64", lwatomics...)
4397 alias("runtime/internal/atomic", "LoadAcquintptr", "runtime/internal/atomic", "LoadAcq", p4...)
4398 alias("sync", "runtime_LoadAcquintptr", "runtime/internal/atomic", "LoadAcq", p4...) // linknamed
4399 alias("runtime/internal/atomic", "LoadAcquintptr", "runtime/internal/atomic", "LoadAcq64", p8...)
4400 alias("sync", "runtime_LoadAcquintptr", "runtime/internal/atomic", "LoadAcq64", p8...) // linknamed
4402 // Aliases for atomic store operations
4403 alias("runtime/internal/atomic", "Storeint32", "runtime/internal/atomic", "Store", all...)
4404 alias("runtime/internal/atomic", "Storeint64", "runtime/internal/atomic", "Store64", all...)
4405 alias("runtime/internal/atomic", "Storeuintptr", "runtime/internal/atomic", "Store", p4...)
4406 alias("runtime/internal/atomic", "Storeuintptr", "runtime/internal/atomic", "Store64", p8...)
4407 alias("runtime/internal/atomic", "StoreRel", "runtime/internal/atomic", "Store", lwatomics...)
4408 alias("runtime/internal/atomic", "StoreRel64", "runtime/internal/atomic", "Store64", lwatomics...)
4409 alias("runtime/internal/atomic", "StoreReluintptr", "runtime/internal/atomic", "StoreRel", p4...)
4410 alias("sync", "runtime_StoreReluintptr", "runtime/internal/atomic", "StoreRel", p4...) // linknamed
4411 alias("runtime/internal/atomic", "StoreReluintptr", "runtime/internal/atomic", "StoreRel64", p8...)
4412 alias("sync", "runtime_StoreReluintptr", "runtime/internal/atomic", "StoreRel64", p8...) // linknamed
4414 // Aliases for atomic swap operations
4415 alias("runtime/internal/atomic", "Xchgint32", "runtime/internal/atomic", "Xchg", all...)
4416 alias("runtime/internal/atomic", "Xchgint64", "runtime/internal/atomic", "Xchg64", all...)
4417 alias("runtime/internal/atomic", "Xchguintptr", "runtime/internal/atomic", "Xchg", p4...)
4418 alias("runtime/internal/atomic", "Xchguintptr", "runtime/internal/atomic", "Xchg64", p8...)
4420 // Aliases for atomic add operations
4421 alias("runtime/internal/atomic", "Xaddint32", "runtime/internal/atomic", "Xadd", all...)
4422 alias("runtime/internal/atomic", "Xaddint64", "runtime/internal/atomic", "Xadd64", all...)
4423 alias("runtime/internal/atomic", "Xadduintptr", "runtime/internal/atomic", "Xadd", p4...)
4424 alias("runtime/internal/atomic", "Xadduintptr", "runtime/internal/atomic", "Xadd64", p8...)
4426 // Aliases for atomic CAS operations
4427 alias("runtime/internal/atomic", "Casint32", "runtime/internal/atomic", "Cas", all...)
4428 alias("runtime/internal/atomic", "Casint64", "runtime/internal/atomic", "Cas64", all...)
4429 alias("runtime/internal/atomic", "Casuintptr", "runtime/internal/atomic", "Cas", p4...)
4430 alias("runtime/internal/atomic", "Casuintptr", "runtime/internal/atomic", "Cas64", p8...)
4431 alias("runtime/internal/atomic", "Casp1", "runtime/internal/atomic", "Cas", p4...)
4432 alias("runtime/internal/atomic", "Casp1", "runtime/internal/atomic", "Cas64", p8...)
4433 alias("runtime/internal/atomic", "CasRel", "runtime/internal/atomic", "Cas", lwatomics...)
4435 /******** math ********/
4436 addF("math", "sqrt",
4437 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4438 return s.newValue1(ssa.OpSqrt, types.Types[types.TFLOAT64], args[0])
4440 sys.I386, sys.AMD64, sys.ARM, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X, sys.Wasm)
4441 addF("math", "Trunc",
4442 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4443 return s.newValue1(ssa.OpTrunc, types.Types[types.TFLOAT64], args[0])
4445 sys.ARM64, sys.PPC64, sys.S390X, sys.Wasm)
4446 addF("math", "Ceil",
4447 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4448 return s.newValue1(ssa.OpCeil, types.Types[types.TFLOAT64], args[0])
4450 sys.ARM64, sys.PPC64, sys.S390X, sys.Wasm)
4451 addF("math", "Floor",
4452 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4453 return s.newValue1(ssa.OpFloor, types.Types[types.TFLOAT64], args[0])
4455 sys.ARM64, sys.PPC64, sys.S390X, sys.Wasm)
4456 addF("math", "Round",
4457 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4458 return s.newValue1(ssa.OpRound, types.Types[types.TFLOAT64], args[0])
4460 sys.ARM64, sys.PPC64, sys.S390X)
4461 addF("math", "RoundToEven",
4462 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4463 return s.newValue1(ssa.OpRoundToEven, types.Types[types.TFLOAT64], args[0])
4465 sys.ARM64, sys.S390X, sys.Wasm)
4467 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4468 return s.newValue1(ssa.OpAbs, types.Types[types.TFLOAT64], args[0])
4470 sys.ARM64, sys.ARM, sys.PPC64, sys.RISCV64, sys.Wasm, sys.MIPS, sys.MIPS64)
4471 addF("math", "Copysign",
4472 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4473 return s.newValue2(ssa.OpCopysign, types.Types[types.TFLOAT64], args[0], args[1])
4475 sys.PPC64, sys.RISCV64, sys.Wasm)
4477 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4478 return s.newValue3(ssa.OpFMA, types.Types[types.TFLOAT64], args[0], args[1], args[2])
4480 sys.ARM64, sys.PPC64, sys.RISCV64, sys.S390X)
4482 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4483 if !s.config.UseFMA {
4484 s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
4485 return s.variable(n, types.Types[types.TFLOAT64])
4488 if buildcfg.GOAMD64 >= 3 {
4489 return s.newValue3(ssa.OpFMA, types.Types[types.TFLOAT64], args[0], args[1], args[2])
4492 v := s.entryNewValue0A(ssa.OpHasCPUFeature, types.Types[types.TBOOL], ir.Syms.X86HasFMA)
4494 b.Kind = ssa.BlockIf
4496 bTrue := s.f.NewBlock(ssa.BlockPlain)
4497 bFalse := s.f.NewBlock(ssa.BlockPlain)
4498 bEnd := s.f.NewBlock(ssa.BlockPlain)
4501 b.Likely = ssa.BranchLikely // >= haswell cpus are common
4503 // We have the intrinsic - use it directly.
4505 s.vars[n] = s.newValue3(ssa.OpFMA, types.Types[types.TFLOAT64], args[0], args[1], args[2])
4506 s.endBlock().AddEdgeTo(bEnd)
4508 // Call the pure Go version.
4509 s.startBlock(bFalse)
4510 s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
4511 s.endBlock().AddEdgeTo(bEnd)
4515 return s.variable(n, types.Types[types.TFLOAT64])
4519 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4520 if !s.config.UseFMA {
4521 s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
4522 return s.variable(n, types.Types[types.TFLOAT64])
4524 addr := s.entryNewValue1A(ssa.OpAddr, types.Types[types.TBOOL].PtrTo(), ir.Syms.ARMHasVFPv4, s.sb)
4525 v := s.load(types.Types[types.TBOOL], addr)
4527 b.Kind = ssa.BlockIf
4529 bTrue := s.f.NewBlock(ssa.BlockPlain)
4530 bFalse := s.f.NewBlock(ssa.BlockPlain)
4531 bEnd := s.f.NewBlock(ssa.BlockPlain)
4534 b.Likely = ssa.BranchLikely
4536 // We have the intrinsic - use it directly.
4538 s.vars[n] = s.newValue3(ssa.OpFMA, types.Types[types.TFLOAT64], args[0], args[1], args[2])
4539 s.endBlock().AddEdgeTo(bEnd)
4541 // Call the pure Go version.
4542 s.startBlock(bFalse)
4543 s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
4544 s.endBlock().AddEdgeTo(bEnd)
4548 return s.variable(n, types.Types[types.TFLOAT64])
4552 makeRoundAMD64 := func(op ssa.Op) func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4553 return func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4554 if buildcfg.GOAMD64 >= 2 {
4555 return s.newValue1(op, types.Types[types.TFLOAT64], args[0])
4558 v := s.entryNewValue0A(ssa.OpHasCPUFeature, types.Types[types.TBOOL], ir.Syms.X86HasSSE41)
4560 b.Kind = ssa.BlockIf
4562 bTrue := s.f.NewBlock(ssa.BlockPlain)
4563 bFalse := s.f.NewBlock(ssa.BlockPlain)
4564 bEnd := s.f.NewBlock(ssa.BlockPlain)
4567 b.Likely = ssa.BranchLikely // most machines have sse4.1 nowadays
4569 // We have the intrinsic - use it directly.
4571 s.vars[n] = s.newValue1(op, types.Types[types.TFLOAT64], args[0])
4572 s.endBlock().AddEdgeTo(bEnd)
4574 // Call the pure Go version.
4575 s.startBlock(bFalse)
4576 s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
4577 s.endBlock().AddEdgeTo(bEnd)
4581 return s.variable(n, types.Types[types.TFLOAT64])
4584 addF("math", "RoundToEven",
4585 makeRoundAMD64(ssa.OpRoundToEven),
4587 addF("math", "Floor",
4588 makeRoundAMD64(ssa.OpFloor),
4590 addF("math", "Ceil",
4591 makeRoundAMD64(ssa.OpCeil),
4593 addF("math", "Trunc",
4594 makeRoundAMD64(ssa.OpTrunc),
4597 /******** math/bits ********/
4598 addF("math/bits", "TrailingZeros64",
4599 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4600 return s.newValue1(ssa.OpCtz64, types.Types[types.TINT], args[0])
4602 sys.AMD64, sys.I386, sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4603 addF("math/bits", "TrailingZeros32",
4604 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4605 return s.newValue1(ssa.OpCtz32, types.Types[types.TINT], args[0])
4607 sys.AMD64, sys.I386, sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4608 addF("math/bits", "TrailingZeros16",
4609 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4610 x := s.newValue1(ssa.OpZeroExt16to32, types.Types[types.TUINT32], args[0])
4611 c := s.constInt32(types.Types[types.TUINT32], 1<<16)
4612 y := s.newValue2(ssa.OpOr32, types.Types[types.TUINT32], x, c)
4613 return s.newValue1(ssa.OpCtz32, types.Types[types.TINT], y)
4616 addF("math/bits", "TrailingZeros16",
4617 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4618 return s.newValue1(ssa.OpCtz16, types.Types[types.TINT], args[0])
4620 sys.AMD64, sys.I386, sys.ARM, sys.ARM64, sys.Wasm)
4621 addF("math/bits", "TrailingZeros16",
4622 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4623 x := s.newValue1(ssa.OpZeroExt16to64, types.Types[types.TUINT64], args[0])
4624 c := s.constInt64(types.Types[types.TUINT64], 1<<16)
4625 y := s.newValue2(ssa.OpOr64, types.Types[types.TUINT64], x, c)
4626 return s.newValue1(ssa.OpCtz64, types.Types[types.TINT], y)
4628 sys.S390X, sys.PPC64)
4629 addF("math/bits", "TrailingZeros8",
4630 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4631 x := s.newValue1(ssa.OpZeroExt8to32, types.Types[types.TUINT32], args[0])
4632 c := s.constInt32(types.Types[types.TUINT32], 1<<8)
4633 y := s.newValue2(ssa.OpOr32, types.Types[types.TUINT32], x, c)
4634 return s.newValue1(ssa.OpCtz32, types.Types[types.TINT], y)
4637 addF("math/bits", "TrailingZeros8",
4638 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4639 return s.newValue1(ssa.OpCtz8, types.Types[types.TINT], args[0])
4641 sys.AMD64, sys.I386, sys.ARM, sys.ARM64, sys.Wasm)
4642 addF("math/bits", "TrailingZeros8",
4643 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4644 x := s.newValue1(ssa.OpZeroExt8to64, types.Types[types.TUINT64], args[0])
4645 c := s.constInt64(types.Types[types.TUINT64], 1<<8)
4646 y := s.newValue2(ssa.OpOr64, types.Types[types.TUINT64], x, c)
4647 return s.newValue1(ssa.OpCtz64, types.Types[types.TINT], y)
4650 alias("math/bits", "ReverseBytes64", "runtime/internal/sys", "Bswap64", all...)
4651 alias("math/bits", "ReverseBytes32", "runtime/internal/sys", "Bswap32", all...)
4652 // ReverseBytes inlines correctly, no need to intrinsify it.
4653 // Nothing special is needed for targets where ReverseBytes16 lowers to a rotate
4654 // On Power10, 16-bit rotate is not available so use BRH instruction
4655 if buildcfg.GOPPC64 >= 10 {
4656 addF("math/bits", "ReverseBytes16",
4657 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4658 return s.newValue1(ssa.OpBswap16, types.Types[types.TUINT], args[0])
4663 addF("math/bits", "Len64",
4664 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4665 return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], args[0])
4667 sys.AMD64, sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4668 addF("math/bits", "Len32",
4669 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4670 return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], args[0])
4672 sys.AMD64, sys.ARM64, sys.PPC64)
4673 addF("math/bits", "Len32",
4674 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4675 if s.config.PtrSize == 4 {
4676 return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], args[0])
4678 x := s.newValue1(ssa.OpZeroExt32to64, types.Types[types.TUINT64], args[0])
4679 return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], x)
4681 sys.ARM, sys.S390X, sys.MIPS, sys.Wasm)
4682 addF("math/bits", "Len16",
4683 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4684 if s.config.PtrSize == 4 {
4685 x := s.newValue1(ssa.OpZeroExt16to32, types.Types[types.TUINT32], args[0])
4686 return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], x)
4688 x := s.newValue1(ssa.OpZeroExt16to64, types.Types[types.TUINT64], args[0])
4689 return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], x)
4691 sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4692 addF("math/bits", "Len16",
4693 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4694 return s.newValue1(ssa.OpBitLen16, types.Types[types.TINT], args[0])
4697 addF("math/bits", "Len8",
4698 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4699 if s.config.PtrSize == 4 {
4700 x := s.newValue1(ssa.OpZeroExt8to32, types.Types[types.TUINT32], args[0])
4701 return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], x)
4703 x := s.newValue1(ssa.OpZeroExt8to64, types.Types[types.TUINT64], args[0])
4704 return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], x)
4706 sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4707 addF("math/bits", "Len8",
4708 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4709 return s.newValue1(ssa.OpBitLen8, types.Types[types.TINT], args[0])
4712 addF("math/bits", "Len",
4713 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4714 if s.config.PtrSize == 4 {
4715 return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], args[0])
4717 return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], args[0])
4719 sys.AMD64, sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4720 // LeadingZeros is handled because it trivially calls Len.
4721 addF("math/bits", "Reverse64",
4722 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4723 return s.newValue1(ssa.OpBitRev64, types.Types[types.TINT], args[0])
4726 addF("math/bits", "Reverse32",
4727 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4728 return s.newValue1(ssa.OpBitRev32, types.Types[types.TINT], args[0])
4731 addF("math/bits", "Reverse16",
4732 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4733 return s.newValue1(ssa.OpBitRev16, types.Types[types.TINT], args[0])
4736 addF("math/bits", "Reverse8",
4737 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4738 return s.newValue1(ssa.OpBitRev8, types.Types[types.TINT], args[0])
4741 addF("math/bits", "Reverse",
4742 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4743 return s.newValue1(ssa.OpBitRev64, types.Types[types.TINT], args[0])
4746 addF("math/bits", "RotateLeft8",
4747 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4748 return s.newValue2(ssa.OpRotateLeft8, types.Types[types.TUINT8], args[0], args[1])
4751 addF("math/bits", "RotateLeft16",
4752 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4753 return s.newValue2(ssa.OpRotateLeft16, types.Types[types.TUINT16], args[0], args[1])
4756 addF("math/bits", "RotateLeft32",
4757 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4758 return s.newValue2(ssa.OpRotateLeft32, types.Types[types.TUINT32], args[0], args[1])
4760 sys.AMD64, sys.ARM, sys.ARM64, sys.S390X, sys.PPC64, sys.Wasm, sys.Loong64)
4761 addF("math/bits", "RotateLeft64",
4762 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4763 return s.newValue2(ssa.OpRotateLeft64, types.Types[types.TUINT64], args[0], args[1])
4765 sys.AMD64, sys.ARM64, sys.S390X, sys.PPC64, sys.Wasm, sys.Loong64)
4766 alias("math/bits", "RotateLeft", "math/bits", "RotateLeft64", p8...)
4768 makeOnesCountAMD64 := func(op ssa.Op) func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4769 return func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4770 if buildcfg.GOAMD64 >= 2 {
4771 return s.newValue1(op, types.Types[types.TINT], args[0])
4774 v := s.entryNewValue0A(ssa.OpHasCPUFeature, types.Types[types.TBOOL], ir.Syms.X86HasPOPCNT)
4776 b.Kind = ssa.BlockIf
4778 bTrue := s.f.NewBlock(ssa.BlockPlain)
4779 bFalse := s.f.NewBlock(ssa.BlockPlain)
4780 bEnd := s.f.NewBlock(ssa.BlockPlain)
4783 b.Likely = ssa.BranchLikely // most machines have popcnt nowadays
4785 // We have the intrinsic - use it directly.
4787 s.vars[n] = s.newValue1(op, types.Types[types.TINT], args[0])
4788 s.endBlock().AddEdgeTo(bEnd)
4790 // Call the pure Go version.
4791 s.startBlock(bFalse)
4792 s.vars[n] = s.callResult(n, callNormal) // types.Types[TINT]
4793 s.endBlock().AddEdgeTo(bEnd)
4797 return s.variable(n, types.Types[types.TINT])
4800 addF("math/bits", "OnesCount64",
4801 makeOnesCountAMD64(ssa.OpPopCount64),
4803 addF("math/bits", "OnesCount64",
4804 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4805 return s.newValue1(ssa.OpPopCount64, types.Types[types.TINT], args[0])
4807 sys.PPC64, sys.ARM64, sys.S390X, sys.Wasm)
4808 addF("math/bits", "OnesCount32",
4809 makeOnesCountAMD64(ssa.OpPopCount32),
4811 addF("math/bits", "OnesCount32",
4812 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4813 return s.newValue1(ssa.OpPopCount32, types.Types[types.TINT], args[0])
4815 sys.PPC64, sys.ARM64, sys.S390X, sys.Wasm)
4816 addF("math/bits", "OnesCount16",
4817 makeOnesCountAMD64(ssa.OpPopCount16),
4819 addF("math/bits", "OnesCount16",
4820 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4821 return s.newValue1(ssa.OpPopCount16, types.Types[types.TINT], args[0])
4823 sys.ARM64, sys.S390X, sys.PPC64, sys.Wasm)
4824 addF("math/bits", "OnesCount8",
4825 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4826 return s.newValue1(ssa.OpPopCount8, types.Types[types.TINT], args[0])
4828 sys.S390X, sys.PPC64, sys.Wasm)
4829 addF("math/bits", "OnesCount",
4830 makeOnesCountAMD64(ssa.OpPopCount64),
4832 addF("math/bits", "Mul64",
4833 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4834 return s.newValue2(ssa.OpMul64uhilo, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1])
4836 sys.AMD64, sys.ARM64, sys.PPC64, sys.S390X, sys.MIPS64, sys.RISCV64, sys.Loong64)
4837 alias("math/bits", "Mul", "math/bits", "Mul64", p8...)
4838 alias("runtime/internal/math", "Mul64", "math/bits", "Mul64", p8...)
4839 addF("math/bits", "Add64",
4840 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4841 return s.newValue3(ssa.OpAdd64carry, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1], args[2])
4843 sys.AMD64, sys.ARM64, sys.PPC64, sys.S390X, sys.RISCV64, sys.Loong64)
4844 alias("math/bits", "Add", "math/bits", "Add64", p8...)
4845 addF("math/bits", "Sub64",
4846 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4847 return s.newValue3(ssa.OpSub64borrow, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1], args[2])
4849 sys.AMD64, sys.ARM64, sys.PPC64, sys.S390X, sys.RISCV64, sys.Loong64)
4850 alias("math/bits", "Sub", "math/bits", "Sub64", p8...)
4851 addF("math/bits", "Div64",
4852 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4853 // check for divide-by-zero/overflow and panic with appropriate message
4854 cmpZero := s.newValue2(s.ssaOp(ir.ONE, types.Types[types.TUINT64]), types.Types[types.TBOOL], args[2], s.zeroVal(types.Types[types.TUINT64]))
4855 s.check(cmpZero, ir.Syms.Panicdivide)
4856 cmpOverflow := s.newValue2(s.ssaOp(ir.OLT, types.Types[types.TUINT64]), types.Types[types.TBOOL], args[0], args[2])
4857 s.check(cmpOverflow, ir.Syms.Panicoverflow)
4858 return s.newValue3(ssa.OpDiv128u, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1], args[2])
4861 alias("math/bits", "Div", "math/bits", "Div64", sys.ArchAMD64)
4863 alias("runtime/internal/sys", "TrailingZeros8", "math/bits", "TrailingZeros8", all...)
4864 alias("runtime/internal/sys", "TrailingZeros32", "math/bits", "TrailingZeros32", all...)
4865 alias("runtime/internal/sys", "TrailingZeros64", "math/bits", "TrailingZeros64", all...)
4866 alias("runtime/internal/sys", "Len8", "math/bits", "Len8", all...)
4867 alias("runtime/internal/sys", "Len64", "math/bits", "Len64", all...)
4868 alias("runtime/internal/sys", "OnesCount64", "math/bits", "OnesCount64", all...)
4870 /******** sync/atomic ********/
4872 // Note: these are disabled by flag_race in findIntrinsic below.
4873 alias("sync/atomic", "LoadInt32", "runtime/internal/atomic", "Load", all...)
4874 alias("sync/atomic", "LoadInt64", "runtime/internal/atomic", "Load64", all...)
4875 alias("sync/atomic", "LoadPointer", "runtime/internal/atomic", "Loadp", all...)
4876 alias("sync/atomic", "LoadUint32", "runtime/internal/atomic", "Load", all...)
4877 alias("sync/atomic", "LoadUint64", "runtime/internal/atomic", "Load64", all...)
4878 alias("sync/atomic", "LoadUintptr", "runtime/internal/atomic", "Load", p4...)
4879 alias("sync/atomic", "LoadUintptr", "runtime/internal/atomic", "Load64", p8...)
4881 alias("sync/atomic", "StoreInt32", "runtime/internal/atomic", "Store", all...)
4882 alias("sync/atomic", "StoreInt64", "runtime/internal/atomic", "Store64", all...)
4883 // Note: not StorePointer, that needs a write barrier. Same below for {CompareAnd}Swap.
4884 alias("sync/atomic", "StoreUint32", "runtime/internal/atomic", "Store", all...)
4885 alias("sync/atomic", "StoreUint64", "runtime/internal/atomic", "Store64", all...)
4886 alias("sync/atomic", "StoreUintptr", "runtime/internal/atomic", "Store", p4...)
4887 alias("sync/atomic", "StoreUintptr", "runtime/internal/atomic", "Store64", p8...)
4889 alias("sync/atomic", "SwapInt32", "runtime/internal/atomic", "Xchg", all...)
4890 alias("sync/atomic", "SwapInt64", "runtime/internal/atomic", "Xchg64", all...)
4891 alias("sync/atomic", "SwapUint32", "runtime/internal/atomic", "Xchg", all...)
4892 alias("sync/atomic", "SwapUint64", "runtime/internal/atomic", "Xchg64", all...)
4893 alias("sync/atomic", "SwapUintptr", "runtime/internal/atomic", "Xchg", p4...)
4894 alias("sync/atomic", "SwapUintptr", "runtime/internal/atomic", "Xchg64", p8...)
4896 alias("sync/atomic", "CompareAndSwapInt32", "runtime/internal/atomic", "Cas", all...)
4897 alias("sync/atomic", "CompareAndSwapInt64", "runtime/internal/atomic", "Cas64", all...)
4898 alias("sync/atomic", "CompareAndSwapUint32", "runtime/internal/atomic", "Cas", all...)
4899 alias("sync/atomic", "CompareAndSwapUint64", "runtime/internal/atomic", "Cas64", all...)
4900 alias("sync/atomic", "CompareAndSwapUintptr", "runtime/internal/atomic", "Cas", p4...)
4901 alias("sync/atomic", "CompareAndSwapUintptr", "runtime/internal/atomic", "Cas64", p8...)
4903 alias("sync/atomic", "AddInt32", "runtime/internal/atomic", "Xadd", all...)
4904 alias("sync/atomic", "AddInt64", "runtime/internal/atomic", "Xadd64", all...)
4905 alias("sync/atomic", "AddUint32", "runtime/internal/atomic", "Xadd", all...)
4906 alias("sync/atomic", "AddUint64", "runtime/internal/atomic", "Xadd64", all...)
4907 alias("sync/atomic", "AddUintptr", "runtime/internal/atomic", "Xadd", p4...)
4908 alias("sync/atomic", "AddUintptr", "runtime/internal/atomic", "Xadd64", p8...)
4910 /******** math/big ********/
4911 alias("math/big", "mulWW", "math/bits", "Mul64", p8...)
4914 // findIntrinsic returns a function which builds the SSA equivalent of the
4915 // function identified by the symbol sym. If sym is not an intrinsic call, returns nil.
4916 func findIntrinsic(sym *types.Sym) intrinsicBuilder {
4917 if sym == nil || sym.Pkg == nil {
4921 if sym.Pkg == ir.Pkgs.Runtime {
4924 if base.Flag.Race && pkg == "sync/atomic" {
4925 // The race detector needs to be able to intercept these calls.
4926 // We can't intrinsify them.
4929 // Skip intrinsifying math functions (which may contain hard-float
4930 // instructions) when soft-float
4931 if Arch.SoftFloat && pkg == "math" {
4936 if ssa.IntrinsicsDisable {
4937 if pkg == "runtime" && (fn == "getcallerpc" || fn == "getcallersp" || fn == "getclosureptr") {
4938 // These runtime functions don't have definitions, must be intrinsics.
4943 return intrinsics[intrinsicKey{Arch.LinkArch.Arch, pkg, fn}]
4946 func IsIntrinsicCall(n *ir.CallExpr) bool {
4950 name, ok := n.X.(*ir.Name)
4954 return findIntrinsic(name.Sym()) != nil
4957 // intrinsicCall converts a call to a recognized intrinsic function into the intrinsic SSA operation.
4958 func (s *state) intrinsicCall(n *ir.CallExpr) *ssa.Value {
4959 v := findIntrinsic(n.X.Sym())(s, n, s.intrinsicArgs(n))
4960 if ssa.IntrinsicsDebug > 0 {
4965 if x.Op == ssa.OpSelect0 || x.Op == ssa.OpSelect1 {
4968 base.WarnfAt(n.Pos(), "intrinsic substitution for %v with %s", n.X.Sym().Name, x.LongString())
4973 // intrinsicArgs extracts args from n, evaluates them to SSA values, and returns them.
4974 func (s *state) intrinsicArgs(n *ir.CallExpr) []*ssa.Value {
4975 args := make([]*ssa.Value, len(n.Args))
4976 for i, n := range n.Args {
4982 // openDeferRecord adds code to evaluate and store the function for an open-code defer
4983 // call, and records info about the defer, so we can generate proper code on the
4984 // exit paths. n is the sub-node of the defer node that is the actual function
4985 // call. We will also record funcdata information on where the function is stored
4986 // (as well as the deferBits variable), and this will enable us to run the proper
4987 // defer calls during panics.
4988 func (s *state) openDeferRecord(n *ir.CallExpr) {
4989 if len(n.Args) != 0 || n.Op() != ir.OCALLFUNC || n.X.Type().NumResults() != 0 {
4990 s.Fatalf("defer call with arguments or results: %v", n)
4993 opendefer := &openDeferInfo{
4997 // We must always store the function value in a stack slot for the
4998 // runtime panic code to use. But in the defer exit code, we will
4999 // call the function directly if it is a static function.
5000 closureVal := s.expr(fn)
5001 closure := s.openDeferSave(fn.Type(), closureVal)
5002 opendefer.closureNode = closure.Aux.(*ir.Name)
5003 if !(fn.Op() == ir.ONAME && fn.(*ir.Name).Class == ir.PFUNC) {
5004 opendefer.closure = closure
5006 index := len(s.openDefers)
5007 s.openDefers = append(s.openDefers, opendefer)
5009 // Update deferBits only after evaluation and storage to stack of
5010 // the function is successful.
5011 bitvalue := s.constInt8(types.Types[types.TUINT8], 1<<uint(index))
5012 newDeferBits := s.newValue2(ssa.OpOr8, types.Types[types.TUINT8], s.variable(deferBitsVar, types.Types[types.TUINT8]), bitvalue)
5013 s.vars[deferBitsVar] = newDeferBits
5014 s.store(types.Types[types.TUINT8], s.deferBitsAddr, newDeferBits)
5017 // openDeferSave generates SSA nodes to store a value (with type t) for an
5018 // open-coded defer at an explicit autotmp location on the stack, so it can be
5019 // reloaded and used for the appropriate call on exit. Type t must be a function type
5020 // (therefore SSAable). val is the value to be stored. The function returns an SSA
5021 // value representing a pointer to the autotmp location.
5022 func (s *state) openDeferSave(t *types.Type, val *ssa.Value) *ssa.Value {
5024 s.Fatalf("openDeferSave of non-SSA-able type %v val=%v", t, val)
5026 if !t.HasPointers() {
5027 s.Fatalf("openDeferSave of pointerless type %v val=%v", t, val)
5030 temp := typecheck.TempAt(pos.WithNotStmt(), s.curfn, t)
5031 temp.SetOpenDeferSlot(true)
5032 var addrTemp *ssa.Value
5033 // Use OpVarLive to make sure stack slot for the closure is not removed by
5034 // dead-store elimination
5035 if s.curBlock.ID != s.f.Entry.ID {
5036 // Force the tmp storing this defer function to be declared in the entry
5037 // block, so that it will be live for the defer exit code (which will
5038 // actually access it only if the associated defer call has been activated).
5039 if t.HasPointers() {
5040 s.defvars[s.f.Entry.ID][memVar] = s.f.Entry.NewValue1A(src.NoXPos, ssa.OpVarDef, types.TypeMem, temp, s.defvars[s.f.Entry.ID][memVar])
5042 s.defvars[s.f.Entry.ID][memVar] = s.f.Entry.NewValue1A(src.NoXPos, ssa.OpVarLive, types.TypeMem, temp, s.defvars[s.f.Entry.ID][memVar])
5043 addrTemp = s.f.Entry.NewValue2A(src.NoXPos, ssa.OpLocalAddr, types.NewPtr(temp.Type()), temp, s.sp, s.defvars[s.f.Entry.ID][memVar])
5045 // Special case if we're still in the entry block. We can't use
5046 // the above code, since s.defvars[s.f.Entry.ID] isn't defined
5047 // until we end the entry block with s.endBlock().
5048 if t.HasPointers() {
5049 s.vars[memVar] = s.newValue1Apos(ssa.OpVarDef, types.TypeMem, temp, s.mem(), false)
5051 s.vars[memVar] = s.newValue1Apos(ssa.OpVarLive, types.TypeMem, temp, s.mem(), false)
5052 addrTemp = s.newValue2Apos(ssa.OpLocalAddr, types.NewPtr(temp.Type()), temp, s.sp, s.mem(), false)
5054 // Since we may use this temp during exit depending on the
5055 // deferBits, we must define it unconditionally on entry.
5056 // Therefore, we must make sure it is zeroed out in the entry
5057 // block if it contains pointers, else GC may wrongly follow an
5058 // uninitialized pointer value.
5059 temp.SetNeedzero(true)
5060 // We are storing to the stack, hence we can avoid the full checks in
5061 // storeType() (no write barrier) and do a simple store().
5062 s.store(t, addrTemp, val)
5066 // openDeferExit generates SSA for processing all the open coded defers at exit.
5067 // The code involves loading deferBits, and checking each of the bits to see if
5068 // the corresponding defer statement was executed. For each bit that is turned
5069 // on, the associated defer call is made.
5070 func (s *state) openDeferExit() {
5071 deferExit := s.f.NewBlock(ssa.BlockPlain)
5072 s.endBlock().AddEdgeTo(deferExit)
5073 s.startBlock(deferExit)
5074 s.lastDeferExit = deferExit
5075 s.lastDeferCount = len(s.openDefers)
5076 zeroval := s.constInt8(types.Types[types.TUINT8], 0)
5077 // Test for and run defers in reverse order
5078 for i := len(s.openDefers) - 1; i >= 0; i-- {
5079 r := s.openDefers[i]
5080 bCond := s.f.NewBlock(ssa.BlockPlain)
5081 bEnd := s.f.NewBlock(ssa.BlockPlain)
5083 deferBits := s.variable(deferBitsVar, types.Types[types.TUINT8])
5084 // Generate code to check if the bit associated with the current
5086 bitval := s.constInt8(types.Types[types.TUINT8], 1<<uint(i))
5087 andval := s.newValue2(ssa.OpAnd8, types.Types[types.TUINT8], deferBits, bitval)
5088 eqVal := s.newValue2(ssa.OpEq8, types.Types[types.TBOOL], andval, zeroval)
5090 b.Kind = ssa.BlockIf
5094 bCond.AddEdgeTo(bEnd)
5097 // Clear this bit in deferBits and force store back to stack, so
5098 // we will not try to re-run this defer call if this defer call panics.
5099 nbitval := s.newValue1(ssa.OpCom8, types.Types[types.TUINT8], bitval)
5100 maskedval := s.newValue2(ssa.OpAnd8, types.Types[types.TUINT8], deferBits, nbitval)
5101 s.store(types.Types[types.TUINT8], s.deferBitsAddr, maskedval)
5102 // Use this value for following tests, so we keep previous
5104 s.vars[deferBitsVar] = maskedval
5106 // Generate code to call the function call of the defer, using the
5107 // closure that were stored in argtmps at the point of the defer
5110 stksize := fn.Type().ArgWidth()
5111 var callArgs []*ssa.Value
5113 if r.closure != nil {
5114 v := s.load(r.closure.Type.Elem(), r.closure)
5115 s.maybeNilCheckClosure(v, callDefer)
5116 codeptr := s.rawLoad(types.Types[types.TUINTPTR], v)
5117 aux := ssa.ClosureAuxCall(s.f.ABIDefault.ABIAnalyzeTypes(nil, nil, nil))
5118 call = s.newValue2A(ssa.OpClosureLECall, aux.LateExpansionResultType(), aux, codeptr, v)
5120 aux := ssa.StaticAuxCall(fn.(*ir.Name).Linksym(), s.f.ABIDefault.ABIAnalyzeTypes(nil, nil, nil))
5121 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
5123 callArgs = append(callArgs, s.mem())
5124 call.AddArgs(callArgs...)
5125 call.AuxInt = stksize
5126 s.vars[memVar] = s.newValue1I(ssa.OpSelectN, types.TypeMem, 0, call)
5127 // Make sure that the stack slots with pointers are kept live
5128 // through the call (which is a pre-emption point). Also, we will
5129 // use the first call of the last defer exit to compute liveness
5130 // for the deferreturn, so we want all stack slots to be live.
5131 if r.closureNode != nil {
5132 s.vars[memVar] = s.newValue1Apos(ssa.OpVarLive, types.TypeMem, r.closureNode, s.mem(), false)
5140 func (s *state) callResult(n *ir.CallExpr, k callKind) *ssa.Value {
5141 return s.call(n, k, false)
5144 func (s *state) callAddr(n *ir.CallExpr, k callKind) *ssa.Value {
5145 return s.call(n, k, true)
5148 // Calls the function n using the specified call type.
5149 // Returns the address of the return value (or nil if none).
5150 func (s *state) call(n *ir.CallExpr, k callKind, returnResultAddr bool) *ssa.Value {
5152 var callee *ir.Name // target function (if static)
5153 var closure *ssa.Value // ptr to closure to run (if dynamic)
5154 var codeptr *ssa.Value // ptr to target code (if dynamic)
5155 var rcvr *ssa.Value // receiver to set
5157 var ACArgs []*types.Type // AuxCall args
5158 var ACResults []*types.Type // AuxCall results
5159 var callArgs []*ssa.Value // For late-expansion, the args themselves (not stored, args to the call instead).
5161 callABI := s.f.ABIDefault
5163 if k != callNormal && k != callTail && (len(n.Args) != 0 || n.Op() == ir.OCALLINTER || n.X.Type().NumResults() != 0) {
5164 s.Fatalf("go/defer call with arguments: %v", n)
5169 if (k == callNormal || k == callTail) && fn.Op() == ir.ONAME && fn.(*ir.Name).Class == ir.PFUNC {
5172 if buildcfg.Experiment.RegabiArgs {
5173 // This is a static call, so it may be
5174 // a direct call to a non-ABIInternal
5175 // function. fn.Func may be nil for
5176 // some compiler-generated functions,
5177 // but those are all ABIInternal.
5179 callABI = abiForFunc(fn.Func, s.f.ABI0, s.f.ABI1)
5182 // TODO(register args) remove after register abi is working
5183 inRegistersImported := fn.Pragma()&ir.RegisterParams != 0
5184 inRegistersSamePackage := fn.Func != nil && fn.Func.Pragma&ir.RegisterParams != 0
5185 if inRegistersImported || inRegistersSamePackage {
5191 closure = s.expr(fn)
5192 if k != callDefer && k != callDeferStack {
5193 // Deferred nil function needs to panic when the function is invoked,
5194 // not the point of defer statement.
5195 s.maybeNilCheckClosure(closure, k)
5198 if fn.Op() != ir.ODOTINTER {
5199 s.Fatalf("OCALLINTER: n.Left not an ODOTINTER: %v", fn.Op())
5201 fn := fn.(*ir.SelectorExpr)
5202 var iclosure *ssa.Value
5203 iclosure, rcvr = s.getClosureAndRcvr(fn)
5204 if k == callNormal {
5205 codeptr = s.load(types.Types[types.TUINTPTR], iclosure)
5211 params := callABI.ABIAnalyze(n.X.Type(), false /* Do not set (register) nNames from caller side -- can cause races. */)
5212 types.CalcSize(fn.Type())
5213 stksize := params.ArgWidth() // includes receiver, args, and results
5215 res := n.X.Type().Results()
5216 if k == callNormal || k == callTail {
5217 for _, p := range params.OutParams() {
5218 ACResults = append(ACResults, p.Type)
5223 if k == callDeferStack {
5224 // Make a defer struct d on the stack.
5226 s.Fatalf("deferprocStack with non-zero stack size %d: %v", stksize, n)
5230 d := typecheck.TempAt(n.Pos(), s.curfn, t)
5232 if t.HasPointers() {
5233 s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, d, s.mem())
5237 // Must match deferstruct() below and src/runtime/runtime2.go:_defer.
5238 // 0: started, set in deferprocStack
5239 // 1: heap, set in deferprocStack
5241 // 3: sp, set in deferprocStack
5242 // 4: pc, set in deferprocStack
5244 s.store(closure.Type,
5245 s.newValue1I(ssa.OpOffPtr, closure.Type.PtrTo(), t.FieldOff(5), addr),
5247 // 6: panic, set in deferprocStack
5248 // 7: link, set in deferprocStack
5253 // Call runtime.deferprocStack with pointer to _defer record.
5254 ACArgs = append(ACArgs, types.Types[types.TUINTPTR])
5255 aux := ssa.StaticAuxCall(ir.Syms.DeferprocStack, s.f.ABIDefault.ABIAnalyzeTypes(nil, ACArgs, ACResults))
5256 callArgs = append(callArgs, addr, s.mem())
5257 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
5258 call.AddArgs(callArgs...)
5259 call.AuxInt = int64(types.PtrSize) // deferprocStack takes a *_defer arg
5261 // Store arguments to stack, including defer/go arguments and receiver for method calls.
5262 // These are written in SP-offset order.
5263 argStart := base.Ctxt.Arch.FixedFrameSize
5265 if k != callNormal && k != callTail {
5266 // Write closure (arg to newproc/deferproc).
5267 ACArgs = append(ACArgs, types.Types[types.TUINTPTR]) // not argExtra
5268 callArgs = append(callArgs, closure)
5269 stksize += int64(types.PtrSize)
5270 argStart += int64(types.PtrSize)
5273 // Set receiver (for interface calls).
5275 callArgs = append(callArgs, rcvr)
5282 for _, p := range params.InParams() { // includes receiver for interface calls
5283 ACArgs = append(ACArgs, p.Type)
5286 // Split the entry block if there are open defers, because later calls to
5287 // openDeferSave may cause a mismatch between the mem for an OpDereference
5288 // and the call site which uses it. See #49282.
5289 if s.curBlock.ID == s.f.Entry.ID && s.hasOpenDefers {
5291 b.Kind = ssa.BlockPlain
5292 curb := s.f.NewBlock(ssa.BlockPlain)
5297 for i, n := range args {
5298 callArgs = append(callArgs, s.putArg(n, t.Params().Field(i).Type))
5301 callArgs = append(callArgs, s.mem())
5305 case k == callDefer:
5306 aux := ssa.StaticAuxCall(ir.Syms.Deferproc, s.f.ABIDefault.ABIAnalyzeTypes(nil, ACArgs, ACResults)) // TODO paramResultInfo for DeferProc
5307 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
5309 aux := ssa.StaticAuxCall(ir.Syms.Newproc, s.f.ABIDefault.ABIAnalyzeTypes(nil, ACArgs, ACResults))
5310 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux) // TODO paramResultInfo for NewProc
5311 case closure != nil:
5312 // rawLoad because loading the code pointer from a
5313 // closure is always safe, but IsSanitizerSafeAddr
5314 // can't always figure that out currently, and it's
5315 // critical that we not clobber any arguments already
5316 // stored onto the stack.
5317 codeptr = s.rawLoad(types.Types[types.TUINTPTR], closure)
5318 aux := ssa.ClosureAuxCall(callABI.ABIAnalyzeTypes(nil, ACArgs, ACResults))
5319 call = s.newValue2A(ssa.OpClosureLECall, aux.LateExpansionResultType(), aux, codeptr, closure)
5320 case codeptr != nil:
5321 // Note that the "receiver" parameter is nil because the actual receiver is the first input parameter.
5322 aux := ssa.InterfaceAuxCall(params)
5323 call = s.newValue1A(ssa.OpInterLECall, aux.LateExpansionResultType(), aux, codeptr)
5325 aux := ssa.StaticAuxCall(callTargetLSym(callee), params)
5326 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
5328 call.Op = ssa.OpTailLECall
5329 stksize = 0 // Tail call does not use stack. We reuse caller's frame.
5332 s.Fatalf("bad call type %v %v", n.Op(), n)
5334 call.AddArgs(callArgs...)
5335 call.AuxInt = stksize // Call operations carry the argsize of the callee along with them
5338 s.vars[memVar] = s.newValue1I(ssa.OpSelectN, types.TypeMem, int64(len(ACResults)), call)
5339 // Insert VarLive opcodes.
5340 for _, v := range n.KeepAlive {
5342 s.Fatalf("KeepAlive variable %v must have Addrtaken set", v)
5345 case ir.PAUTO, ir.PPARAM, ir.PPARAMOUT:
5347 s.Fatalf("KeepAlive variable %v must be Auto or Arg", v)
5349 s.vars[memVar] = s.newValue1A(ssa.OpVarLive, types.TypeMem, v, s.mem())
5352 // Finish block for defers
5353 if k == callDefer || k == callDeferStack {
5355 b.Kind = ssa.BlockDefer
5357 bNext := s.f.NewBlock(ssa.BlockPlain)
5359 // Add recover edge to exit code.
5360 r := s.f.NewBlock(ssa.BlockPlain)
5364 b.Likely = ssa.BranchLikely
5368 if res.NumFields() == 0 || k != callNormal {
5369 // call has no return value. Continue with the next statement.
5373 if returnResultAddr {
5374 return s.resultAddrOfCall(call, 0, fp.Type)
5376 return s.newValue1I(ssa.OpSelectN, fp.Type, 0, call)
5379 // maybeNilCheckClosure checks if a nil check of a closure is needed in some
5380 // architecture-dependent situations and, if so, emits the nil check.
5381 func (s *state) maybeNilCheckClosure(closure *ssa.Value, k callKind) {
5382 if Arch.LinkArch.Family == sys.Wasm || buildcfg.GOOS == "aix" && k != callGo {
5383 // On AIX, the closure needs to be verified as fn can be nil, except if it's a call go. This needs to be handled by the runtime to have the "go of nil func value" error.
5384 // TODO(neelance): On other architectures this should be eliminated by the optimization steps
5389 // getClosureAndRcvr returns values for the appropriate closure and receiver of an
5391 func (s *state) getClosureAndRcvr(fn *ir.SelectorExpr) (*ssa.Value, *ssa.Value) {
5393 itab := s.newValue1(ssa.OpITab, types.Types[types.TUINTPTR], i)
5395 itabidx := fn.Offset() + 2*int64(types.PtrSize) + 8 // offset of fun field in runtime.itab
5396 closure := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.UintptrPtr, itabidx, itab)
5397 rcvr := s.newValue1(ssa.OpIData, s.f.Config.Types.BytePtr, i)
5398 return closure, rcvr
5401 // etypesign returns the signed-ness of e, for integer/pointer etypes.
5402 // -1 means signed, +1 means unsigned, 0 means non-integer/non-pointer.
5403 func etypesign(e types.Kind) int8 {
5405 case types.TINT8, types.TINT16, types.TINT32, types.TINT64, types.TINT:
5407 case types.TUINT8, types.TUINT16, types.TUINT32, types.TUINT64, types.TUINT, types.TUINTPTR, types.TUNSAFEPTR:
5413 // addr converts the address of the expression n to SSA, adds it to s and returns the SSA result.
5414 // The value that the returned Value represents is guaranteed to be non-nil.
5415 func (s *state) addr(n ir.Node) *ssa.Value {
5416 if n.Op() != ir.ONAME {
5422 s.Fatalf("addr of canSSA expression: %+v", n)
5425 t := types.NewPtr(n.Type())
5426 linksymOffset := func(lsym *obj.LSym, offset int64) *ssa.Value {
5427 v := s.entryNewValue1A(ssa.OpAddr, t, lsym, s.sb)
5428 // TODO: Make OpAddr use AuxInt as well as Aux.
5430 v = s.entryNewValue1I(ssa.OpOffPtr, v.Type, offset, v)
5435 case ir.OLINKSYMOFFSET:
5436 no := n.(*ir.LinksymOffsetExpr)
5437 return linksymOffset(no.Linksym, no.Offset_)
5440 if n.Heapaddr != nil {
5441 return s.expr(n.Heapaddr)
5446 return linksymOffset(n.Linksym(), 0)
5453 s.Fatalf("addr of undeclared ONAME %v. declared: %v", n, s.decladdrs)
5456 return s.newValue2Apos(ssa.OpLocalAddr, t, n, s.sp, s.mem(), !ir.IsAutoTmp(n))
5458 case ir.PPARAMOUT: // Same as PAUTO -- cannot generate LEA early.
5459 // ensure that we reuse symbols for out parameters so
5460 // that cse works on their addresses
5461 return s.newValue2Apos(ssa.OpLocalAddr, t, n, s.sp, s.mem(), true)
5463 s.Fatalf("variable address class %v not implemented", n.Class)
5467 // load return from callee
5468 n := n.(*ir.ResultExpr)
5469 return s.resultAddrOfCall(s.prevCall, n.Index, n.Type())
5471 n := n.(*ir.IndexExpr)
5472 if n.X.Type().IsSlice() {
5474 i := s.expr(n.Index)
5475 len := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], a)
5476 i = s.boundsCheck(i, len, ssa.BoundsIndex, n.Bounded())
5477 p := s.newValue1(ssa.OpSlicePtr, t, a)
5478 return s.newValue2(ssa.OpPtrIndex, t, p, i)
5481 i := s.expr(n.Index)
5482 len := s.constInt(types.Types[types.TINT], n.X.Type().NumElem())
5483 i = s.boundsCheck(i, len, ssa.BoundsIndex, n.Bounded())
5484 return s.newValue2(ssa.OpPtrIndex, types.NewPtr(n.X.Type().Elem()), a, i)
5487 n := n.(*ir.StarExpr)
5488 return s.exprPtr(n.X, n.Bounded(), n.Pos())
5490 n := n.(*ir.SelectorExpr)
5492 return s.newValue1I(ssa.OpOffPtr, t, n.Offset(), p)
5494 n := n.(*ir.SelectorExpr)
5495 p := s.exprPtr(n.X, n.Bounded(), n.Pos())
5496 return s.newValue1I(ssa.OpOffPtr, t, n.Offset(), p)
5498 n := n.(*ir.ConvExpr)
5499 if n.Type() == n.X.Type() {
5503 return s.newValue1(ssa.OpCopy, t, addr) // ensure that addr has the right type
5504 case ir.OCALLFUNC, ir.OCALLINTER:
5505 n := n.(*ir.CallExpr)
5506 return s.callAddr(n, callNormal)
5507 case ir.ODOTTYPE, ir.ODYNAMICDOTTYPE:
5509 if n.Op() == ir.ODOTTYPE {
5510 v, _ = s.dottype(n.(*ir.TypeAssertExpr), false)
5512 v, _ = s.dynamicDottype(n.(*ir.DynamicTypeAssertExpr), false)
5514 if v.Op != ssa.OpLoad {
5515 s.Fatalf("dottype of non-load")
5517 if v.Args[1] != s.mem() {
5518 s.Fatalf("memory no longer live from dottype load")
5522 s.Fatalf("unhandled addr %v", n.Op())
5527 // canSSA reports whether n is SSA-able.
5528 // n must be an ONAME (or an ODOT sequence with an ONAME base).
5529 func (s *state) canSSA(n ir.Node) bool {
5530 if base.Flag.N != 0 {
5535 if nn.Op() == ir.ODOT {
5536 nn := nn.(*ir.SelectorExpr)
5540 if nn.Op() == ir.OINDEX {
5541 nn := nn.(*ir.IndexExpr)
5542 if nn.X.Type().IsArray() {
5549 if n.Op() != ir.ONAME {
5552 return s.canSSAName(n.(*ir.Name)) && TypeOK(n.Type())
5555 func (s *state) canSSAName(name *ir.Name) bool {
5556 if name.Addrtaken() || !name.OnStack() {
5562 // TODO: handle this case? Named return values must be
5563 // in memory so that the deferred function can see them.
5564 // Maybe do: if !strings.HasPrefix(n.String(), "~") { return false }
5565 // Or maybe not, see issue 18860. Even unnamed return values
5566 // must be written back so if a defer recovers, the caller can see them.
5569 if s.cgoUnsafeArgs {
5570 // Cgo effectively takes the address of all result args,
5571 // but the compiler can't see that.
5576 // TODO: try to make more variables SSAable?
5579 // TypeOK reports whether variables of type t are SSA-able.
5580 func TypeOK(t *types.Type) bool {
5582 if t.Size() > int64(4*types.PtrSize) {
5583 // 4*Widthptr is an arbitrary constant. We want it
5584 // to be at least 3*Widthptr so slices can be registerized.
5585 // Too big and we'll introduce too much register pressure.
5590 // We can't do larger arrays because dynamic indexing is
5591 // not supported on SSA variables.
5592 // TODO: allow if all indexes are constant.
5593 if t.NumElem() <= 1 {
5594 return TypeOK(t.Elem())
5598 if t.NumFields() > ssa.MaxStruct {
5601 for _, t1 := range t.Fields().Slice() {
5602 if !TypeOK(t1.Type) {
5612 // exprPtr evaluates n to a pointer and nil-checks it.
5613 func (s *state) exprPtr(n ir.Node, bounded bool, lineno src.XPos) *ssa.Value {
5615 if bounded || n.NonNil() {
5616 if s.f.Frontend().Debug_checknil() && lineno.Line() > 1 {
5617 s.f.Warnl(lineno, "removed nil check")
5625 // nilCheck generates nil pointer checking code.
5626 // Used only for automatically inserted nil checks,
5627 // not for user code like 'x != nil'.
5628 func (s *state) nilCheck(ptr *ssa.Value) {
5629 if base.Debug.DisableNil != 0 || s.curfn.NilCheckDisabled() {
5632 s.newValue2(ssa.OpNilCheck, types.TypeVoid, ptr, s.mem())
5635 // boundsCheck generates bounds checking code. Checks if 0 <= idx <[=] len, branches to exit if not.
5636 // Starts a new block on return.
5637 // On input, len must be converted to full int width and be nonnegative.
5638 // Returns idx converted to full int width.
5639 // If bounded is true then caller guarantees the index is not out of bounds
5640 // (but boundsCheck will still extend the index to full int width).
5641 func (s *state) boundsCheck(idx, len *ssa.Value, kind ssa.BoundsKind, bounded bool) *ssa.Value {
5642 idx = s.extendIndex(idx, len, kind, bounded)
5644 if bounded || base.Flag.B != 0 {
5645 // If bounded or bounds checking is flag-disabled, then no check necessary,
5646 // just return the extended index.
5648 // Here, bounded == true if the compiler generated the index itself,
5649 // such as in the expansion of a slice initializer. These indexes are
5650 // compiler-generated, not Go program variables, so they cannot be
5651 // attacker-controlled, so we can omit Spectre masking as well.
5653 // Note that we do not want to omit Spectre masking in code like:
5655 // if 0 <= i && i < len(x) {
5659 // Lucky for us, bounded==false for that code.
5660 // In that case (handled below), we emit a bound check (and Spectre mask)
5661 // and then the prove pass will remove the bounds check.
5662 // In theory the prove pass could potentially remove certain
5663 // Spectre masks, but it's very delicate and probably better
5664 // to be conservative and leave them all in.
5668 bNext := s.f.NewBlock(ssa.BlockPlain)
5669 bPanic := s.f.NewBlock(ssa.BlockExit)
5671 if !idx.Type.IsSigned() {
5673 case ssa.BoundsIndex:
5674 kind = ssa.BoundsIndexU
5675 case ssa.BoundsSliceAlen:
5676 kind = ssa.BoundsSliceAlenU
5677 case ssa.BoundsSliceAcap:
5678 kind = ssa.BoundsSliceAcapU
5679 case ssa.BoundsSliceB:
5680 kind = ssa.BoundsSliceBU
5681 case ssa.BoundsSlice3Alen:
5682 kind = ssa.BoundsSlice3AlenU
5683 case ssa.BoundsSlice3Acap:
5684 kind = ssa.BoundsSlice3AcapU
5685 case ssa.BoundsSlice3B:
5686 kind = ssa.BoundsSlice3BU
5687 case ssa.BoundsSlice3C:
5688 kind = ssa.BoundsSlice3CU
5693 if kind == ssa.BoundsIndex || kind == ssa.BoundsIndexU {
5694 cmp = s.newValue2(ssa.OpIsInBounds, types.Types[types.TBOOL], idx, len)
5696 cmp = s.newValue2(ssa.OpIsSliceInBounds, types.Types[types.TBOOL], idx, len)
5699 b.Kind = ssa.BlockIf
5701 b.Likely = ssa.BranchLikely
5705 s.startBlock(bPanic)
5706 if Arch.LinkArch.Family == sys.Wasm {
5707 // TODO(khr): figure out how to do "register" based calling convention for bounds checks.
5708 // Should be similar to gcWriteBarrier, but I can't make it work.
5709 s.rtcall(BoundsCheckFunc[kind], false, nil, idx, len)
5711 mem := s.newValue3I(ssa.OpPanicBounds, types.TypeMem, int64(kind), idx, len, s.mem())
5712 s.endBlock().SetControl(mem)
5716 // In Spectre index mode, apply an appropriate mask to avoid speculative out-of-bounds accesses.
5717 if base.Flag.Cfg.SpectreIndex {
5718 op := ssa.OpSpectreIndex
5719 if kind != ssa.BoundsIndex && kind != ssa.BoundsIndexU {
5720 op = ssa.OpSpectreSliceIndex
5722 idx = s.newValue2(op, types.Types[types.TINT], idx, len)
5728 // If cmp (a bool) is false, panic using the given function.
5729 func (s *state) check(cmp *ssa.Value, fn *obj.LSym) {
5731 b.Kind = ssa.BlockIf
5733 b.Likely = ssa.BranchLikely
5734 bNext := s.f.NewBlock(ssa.BlockPlain)
5736 pos := base.Ctxt.PosTable.Pos(line)
5737 fl := funcLine{f: fn, base: pos.Base(), line: pos.Line()}
5738 bPanic := s.panics[fl]
5740 bPanic = s.f.NewBlock(ssa.BlockPlain)
5741 s.panics[fl] = bPanic
5742 s.startBlock(bPanic)
5743 // The panic call takes/returns memory to ensure that the right
5744 // memory state is observed if the panic happens.
5745 s.rtcall(fn, false, nil)
5752 func (s *state) intDivide(n ir.Node, a, b *ssa.Value) *ssa.Value {
5755 case ssa.OpConst8, ssa.OpConst16, ssa.OpConst32, ssa.OpConst64:
5761 // do a size-appropriate check for zero
5762 cmp := s.newValue2(s.ssaOp(ir.ONE, n.Type()), types.Types[types.TBOOL], b, s.zeroVal(n.Type()))
5763 s.check(cmp, ir.Syms.Panicdivide)
5765 return s.newValue2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
5768 // rtcall issues a call to the given runtime function fn with the listed args.
5769 // Returns a slice of results of the given result types.
5770 // The call is added to the end of the current block.
5771 // If returns is false, the block is marked as an exit block.
5772 func (s *state) rtcall(fn *obj.LSym, returns bool, results []*types.Type, args ...*ssa.Value) []*ssa.Value {
5774 // Write args to the stack
5775 off := base.Ctxt.Arch.FixedFrameSize
5776 var callArgs []*ssa.Value
5777 var callArgTypes []*types.Type
5779 for _, arg := range args {
5781 off = types.RoundUp(off, t.Alignment())
5783 callArgs = append(callArgs, arg)
5784 callArgTypes = append(callArgTypes, t)
5787 off = types.RoundUp(off, int64(types.RegSize))
5791 aux := ssa.StaticAuxCall(fn, s.f.ABIDefault.ABIAnalyzeTypes(nil, callArgTypes, results))
5792 callArgs = append(callArgs, s.mem())
5793 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
5794 call.AddArgs(callArgs...)
5795 s.vars[memVar] = s.newValue1I(ssa.OpSelectN, types.TypeMem, int64(len(results)), call)
5800 b.Kind = ssa.BlockExit
5802 call.AuxInt = off - base.Ctxt.Arch.FixedFrameSize
5803 if len(results) > 0 {
5804 s.Fatalf("panic call can't have results")
5810 res := make([]*ssa.Value, len(results))
5811 for i, t := range results {
5812 off = types.RoundUp(off, t.Alignment())
5813 res[i] = s.resultOfCall(call, int64(i), t)
5816 off = types.RoundUp(off, int64(types.PtrSize))
5818 // Remember how much callee stack space we needed.
5824 // do *left = right for type t.
5825 func (s *state) storeType(t *types.Type, left, right *ssa.Value, skip skipMask, leftIsStmt bool) {
5826 s.instrument(t, left, instrumentWrite)
5828 if skip == 0 && (!t.HasPointers() || ssa.IsStackAddr(left)) {
5829 // Known to not have write barrier. Store the whole type.
5830 s.vars[memVar] = s.newValue3Apos(ssa.OpStore, types.TypeMem, t, left, right, s.mem(), leftIsStmt)
5834 // store scalar fields first, so write barrier stores for
5835 // pointer fields can be grouped together, and scalar values
5836 // don't need to be live across the write barrier call.
5837 // TODO: if the writebarrier pass knows how to reorder stores,
5838 // we can do a single store here as long as skip==0.
5839 s.storeTypeScalars(t, left, right, skip)
5840 if skip&skipPtr == 0 && t.HasPointers() {
5841 s.storeTypePtrs(t, left, right)
5845 // do *left = right for all scalar (non-pointer) parts of t.
5846 func (s *state) storeTypeScalars(t *types.Type, left, right *ssa.Value, skip skipMask) {
5848 case t.IsBoolean() || t.IsInteger() || t.IsFloat() || t.IsComplex():
5849 s.store(t, left, right)
5850 case t.IsPtrShaped():
5851 if t.IsPtr() && t.Elem().NotInHeap() {
5852 s.store(t, left, right) // see issue 42032
5854 // otherwise, no scalar fields.
5856 if skip&skipLen != 0 {
5859 len := s.newValue1(ssa.OpStringLen, types.Types[types.TINT], right)
5860 lenAddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, s.config.PtrSize, left)
5861 s.store(types.Types[types.TINT], lenAddr, len)
5863 if skip&skipLen == 0 {
5864 len := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], right)
5865 lenAddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, s.config.PtrSize, left)
5866 s.store(types.Types[types.TINT], lenAddr, len)
5868 if skip&skipCap == 0 {
5869 cap := s.newValue1(ssa.OpSliceCap, types.Types[types.TINT], right)
5870 capAddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, 2*s.config.PtrSize, left)
5871 s.store(types.Types[types.TINT], capAddr, cap)
5873 case t.IsInterface():
5874 // itab field doesn't need a write barrier (even though it is a pointer).
5875 itab := s.newValue1(ssa.OpITab, s.f.Config.Types.BytePtr, right)
5876 s.store(types.Types[types.TUINTPTR], left, itab)
5879 for i := 0; i < n; i++ {
5880 ft := t.FieldType(i)
5881 addr := s.newValue1I(ssa.OpOffPtr, ft.PtrTo(), t.FieldOff(i), left)
5882 val := s.newValue1I(ssa.OpStructSelect, ft, int64(i), right)
5883 s.storeTypeScalars(ft, addr, val, 0)
5885 case t.IsArray() && t.NumElem() == 0:
5887 case t.IsArray() && t.NumElem() == 1:
5888 s.storeTypeScalars(t.Elem(), left, s.newValue1I(ssa.OpArraySelect, t.Elem(), 0, right), 0)
5890 s.Fatalf("bad write barrier type %v", t)
5894 // do *left = right for all pointer parts of t.
5895 func (s *state) storeTypePtrs(t *types.Type, left, right *ssa.Value) {
5897 case t.IsPtrShaped():
5898 if t.IsPtr() && t.Elem().NotInHeap() {
5899 break // see issue 42032
5901 s.store(t, left, right)
5903 ptr := s.newValue1(ssa.OpStringPtr, s.f.Config.Types.BytePtr, right)
5904 s.store(s.f.Config.Types.BytePtr, left, ptr)
5906 elType := types.NewPtr(t.Elem())
5907 ptr := s.newValue1(ssa.OpSlicePtr, elType, right)
5908 s.store(elType, left, ptr)
5909 case t.IsInterface():
5910 // itab field is treated as a scalar.
5911 idata := s.newValue1(ssa.OpIData, s.f.Config.Types.BytePtr, right)
5912 idataAddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.BytePtrPtr, s.config.PtrSize, left)
5913 s.store(s.f.Config.Types.BytePtr, idataAddr, idata)
5916 for i := 0; i < n; i++ {
5917 ft := t.FieldType(i)
5918 if !ft.HasPointers() {
5921 addr := s.newValue1I(ssa.OpOffPtr, ft.PtrTo(), t.FieldOff(i), left)
5922 val := s.newValue1I(ssa.OpStructSelect, ft, int64(i), right)
5923 s.storeTypePtrs(ft, addr, val)
5925 case t.IsArray() && t.NumElem() == 0:
5927 case t.IsArray() && t.NumElem() == 1:
5928 s.storeTypePtrs(t.Elem(), left, s.newValue1I(ssa.OpArraySelect, t.Elem(), 0, right))
5930 s.Fatalf("bad write barrier type %v", t)
5934 // putArg evaluates n for the purpose of passing it as an argument to a function and returns the value for the call.
5935 func (s *state) putArg(n ir.Node, t *types.Type) *ssa.Value {
5938 a = s.newValue2(ssa.OpDereference, t, s.addr(n), s.mem())
5945 func (s *state) storeArgWithBase(n ir.Node, t *types.Type, base *ssa.Value, off int64) {
5946 pt := types.NewPtr(t)
5949 // Use special routine that avoids allocation on duplicate offsets.
5950 addr = s.constOffPtrSP(pt, off)
5952 addr = s.newValue1I(ssa.OpOffPtr, pt, off, base)
5962 s.storeType(t, addr, a, 0, false)
5965 // slice computes the slice v[i:j:k] and returns ptr, len, and cap of result.
5966 // i,j,k may be nil, in which case they are set to their default value.
5967 // v may be a slice, string or pointer to an array.
5968 func (s *state) slice(v, i, j, k *ssa.Value, bounded bool) (p, l, c *ssa.Value) {
5970 var ptr, len, cap *ssa.Value
5973 ptr = s.newValue1(ssa.OpSlicePtr, types.NewPtr(t.Elem()), v)
5974 len = s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], v)
5975 cap = s.newValue1(ssa.OpSliceCap, types.Types[types.TINT], v)
5977 ptr = s.newValue1(ssa.OpStringPtr, types.NewPtr(types.Types[types.TUINT8]), v)
5978 len = s.newValue1(ssa.OpStringLen, types.Types[types.TINT], v)
5981 if !t.Elem().IsArray() {
5982 s.Fatalf("bad ptr to array in slice %v\n", t)
5985 ptr = s.newValue1(ssa.OpCopy, types.NewPtr(t.Elem().Elem()), v)
5986 len = s.constInt(types.Types[types.TINT], t.Elem().NumElem())
5989 s.Fatalf("bad type in slice %v\n", t)
5992 // Set default values
5994 i = s.constInt(types.Types[types.TINT], 0)
6005 // Panic if slice indices are not in bounds.
6006 // Make sure we check these in reverse order so that we're always
6007 // comparing against a value known to be nonnegative. See issue 28797.
6010 kind := ssa.BoundsSlice3Alen
6012 kind = ssa.BoundsSlice3Acap
6014 k = s.boundsCheck(k, cap, kind, bounded)
6017 j = s.boundsCheck(j, k, ssa.BoundsSlice3B, bounded)
6019 i = s.boundsCheck(i, j, ssa.BoundsSlice3C, bounded)
6022 kind := ssa.BoundsSliceAlen
6024 kind = ssa.BoundsSliceAcap
6026 j = s.boundsCheck(j, k, kind, bounded)
6028 i = s.boundsCheck(i, j, ssa.BoundsSliceB, bounded)
6031 // Word-sized integer operations.
6032 subOp := s.ssaOp(ir.OSUB, types.Types[types.TINT])
6033 mulOp := s.ssaOp(ir.OMUL, types.Types[types.TINT])
6034 andOp := s.ssaOp(ir.OAND, types.Types[types.TINT])
6036 // Calculate the length (rlen) and capacity (rcap) of the new slice.
6037 // For strings the capacity of the result is unimportant. However,
6038 // we use rcap to test if we've generated a zero-length slice.
6039 // Use length of strings for that.
6040 rlen := s.newValue2(subOp, types.Types[types.TINT], j, i)
6042 if j != k && !t.IsString() {
6043 rcap = s.newValue2(subOp, types.Types[types.TINT], k, i)
6046 if (i.Op == ssa.OpConst64 || i.Op == ssa.OpConst32) && i.AuxInt == 0 {
6047 // No pointer arithmetic necessary.
6048 return ptr, rlen, rcap
6051 // Calculate the base pointer (rptr) for the new slice.
6053 // Generate the following code assuming that indexes are in bounds.
6054 // The masking is to make sure that we don't generate a slice
6055 // that points to the next object in memory. We cannot just set
6056 // the pointer to nil because then we would create a nil slice or
6061 // rptr = ptr + (mask(rcap) & (i * stride))
6063 // Where mask(x) is 0 if x==0 and -1 if x>0 and stride is the width
6064 // of the element type.
6065 stride := s.constInt(types.Types[types.TINT], ptr.Type.Elem().Size())
6067 // The delta is the number of bytes to offset ptr by.
6068 delta := s.newValue2(mulOp, types.Types[types.TINT], i, stride)
6070 // If we're slicing to the point where the capacity is zero,
6071 // zero out the delta.
6072 mask := s.newValue1(ssa.OpSlicemask, types.Types[types.TINT], rcap)
6073 delta = s.newValue2(andOp, types.Types[types.TINT], delta, mask)
6075 // Compute rptr = ptr + delta.
6076 rptr := s.newValue2(ssa.OpAddPtr, ptr.Type, ptr, delta)
6078 return rptr, rlen, rcap
6081 type u642fcvtTab struct {
6082 leq, cvt2F, and, rsh, or, add ssa.Op
6083 one func(*state, *types.Type, int64) *ssa.Value
6086 var u64_f64 = u642fcvtTab{
6088 cvt2F: ssa.OpCvt64to64F,
6090 rsh: ssa.OpRsh64Ux64,
6093 one: (*state).constInt64,
6096 var u64_f32 = u642fcvtTab{
6098 cvt2F: ssa.OpCvt64to32F,
6100 rsh: ssa.OpRsh64Ux64,
6103 one: (*state).constInt64,
6106 func (s *state) uint64Tofloat64(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6107 return s.uint64Tofloat(&u64_f64, n, x, ft, tt)
6110 func (s *state) uint64Tofloat32(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6111 return s.uint64Tofloat(&u64_f32, n, x, ft, tt)
6114 func (s *state) uint64Tofloat(cvttab *u642fcvtTab, n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6116 // result = (floatY) x
6118 // y = uintX(x) ; y = x & 1
6119 // z = uintX(x) ; z = z >> 1
6121 // result = floatY(z)
6122 // result = result + result
6125 // Code borrowed from old code generator.
6126 // What's going on: large 64-bit "unsigned" looks like
6127 // negative number to hardware's integer-to-float
6128 // conversion. However, because the mantissa is only
6129 // 63 bits, we don't need the LSB, so instead we do an
6130 // unsigned right shift (divide by two), convert, and
6131 // double. However, before we do that, we need to be
6132 // sure that we do not lose a "1" if that made the
6133 // difference in the resulting rounding. Therefore, we
6134 // preserve it, and OR (not ADD) it back in. The case
6135 // that matters is when the eleven discarded bits are
6136 // equal to 10000000001; that rounds up, and the 1 cannot
6137 // be lost else it would round down if the LSB of the
6138 // candidate mantissa is 0.
6139 cmp := s.newValue2(cvttab.leq, types.Types[types.TBOOL], s.zeroVal(ft), x)
6141 b.Kind = ssa.BlockIf
6143 b.Likely = ssa.BranchLikely
6145 bThen := s.f.NewBlock(ssa.BlockPlain)
6146 bElse := s.f.NewBlock(ssa.BlockPlain)
6147 bAfter := s.f.NewBlock(ssa.BlockPlain)
6151 a0 := s.newValue1(cvttab.cvt2F, tt, x)
6154 bThen.AddEdgeTo(bAfter)
6158 one := cvttab.one(s, ft, 1)
6159 y := s.newValue2(cvttab.and, ft, x, one)
6160 z := s.newValue2(cvttab.rsh, ft, x, one)
6161 z = s.newValue2(cvttab.or, ft, z, y)
6162 a := s.newValue1(cvttab.cvt2F, tt, z)
6163 a1 := s.newValue2(cvttab.add, tt, a, a)
6166 bElse.AddEdgeTo(bAfter)
6168 s.startBlock(bAfter)
6169 return s.variable(n, n.Type())
6172 type u322fcvtTab struct {
6173 cvtI2F, cvtF2F ssa.Op
6176 var u32_f64 = u322fcvtTab{
6177 cvtI2F: ssa.OpCvt32to64F,
6181 var u32_f32 = u322fcvtTab{
6182 cvtI2F: ssa.OpCvt32to32F,
6183 cvtF2F: ssa.OpCvt64Fto32F,
6186 func (s *state) uint32Tofloat64(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6187 return s.uint32Tofloat(&u32_f64, n, x, ft, tt)
6190 func (s *state) uint32Tofloat32(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6191 return s.uint32Tofloat(&u32_f32, n, x, ft, tt)
6194 func (s *state) uint32Tofloat(cvttab *u322fcvtTab, n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6196 // result = floatY(x)
6198 // result = floatY(float64(x) + (1<<32))
6200 cmp := s.newValue2(ssa.OpLeq32, types.Types[types.TBOOL], s.zeroVal(ft), x)
6202 b.Kind = ssa.BlockIf
6204 b.Likely = ssa.BranchLikely
6206 bThen := s.f.NewBlock(ssa.BlockPlain)
6207 bElse := s.f.NewBlock(ssa.BlockPlain)
6208 bAfter := s.f.NewBlock(ssa.BlockPlain)
6212 a0 := s.newValue1(cvttab.cvtI2F, tt, x)
6215 bThen.AddEdgeTo(bAfter)
6219 a1 := s.newValue1(ssa.OpCvt32to64F, types.Types[types.TFLOAT64], x)
6220 twoToThe32 := s.constFloat64(types.Types[types.TFLOAT64], float64(1<<32))
6221 a2 := s.newValue2(ssa.OpAdd64F, types.Types[types.TFLOAT64], a1, twoToThe32)
6222 a3 := s.newValue1(cvttab.cvtF2F, tt, a2)
6226 bElse.AddEdgeTo(bAfter)
6228 s.startBlock(bAfter)
6229 return s.variable(n, n.Type())
6232 // referenceTypeBuiltin generates code for the len/cap builtins for maps and channels.
6233 func (s *state) referenceTypeBuiltin(n *ir.UnaryExpr, x *ssa.Value) *ssa.Value {
6234 if !n.X.Type().IsMap() && !n.X.Type().IsChan() {
6235 s.Fatalf("node must be a map or a channel")
6241 // return *((*int)n)
6243 // return *(((*int)n)+1)
6246 nilValue := s.constNil(types.Types[types.TUINTPTR])
6247 cmp := s.newValue2(ssa.OpEqPtr, types.Types[types.TBOOL], x, nilValue)
6249 b.Kind = ssa.BlockIf
6251 b.Likely = ssa.BranchUnlikely
6253 bThen := s.f.NewBlock(ssa.BlockPlain)
6254 bElse := s.f.NewBlock(ssa.BlockPlain)
6255 bAfter := s.f.NewBlock(ssa.BlockPlain)
6257 // length/capacity of a nil map/chan is zero
6260 s.vars[n] = s.zeroVal(lenType)
6262 bThen.AddEdgeTo(bAfter)
6268 // length is stored in the first word for map/chan
6269 s.vars[n] = s.load(lenType, x)
6271 // capacity is stored in the second word for chan
6272 sw := s.newValue1I(ssa.OpOffPtr, lenType.PtrTo(), lenType.Size(), x)
6273 s.vars[n] = s.load(lenType, sw)
6275 s.Fatalf("op must be OLEN or OCAP")
6278 bElse.AddEdgeTo(bAfter)
6280 s.startBlock(bAfter)
6281 return s.variable(n, lenType)
6284 type f2uCvtTab struct {
6285 ltf, cvt2U, subf, or ssa.Op
6286 floatValue func(*state, *types.Type, float64) *ssa.Value
6287 intValue func(*state, *types.Type, int64) *ssa.Value
6291 var f32_u64 = f2uCvtTab{
6293 cvt2U: ssa.OpCvt32Fto64,
6296 floatValue: (*state).constFloat32,
6297 intValue: (*state).constInt64,
6301 var f64_u64 = f2uCvtTab{
6303 cvt2U: ssa.OpCvt64Fto64,
6306 floatValue: (*state).constFloat64,
6307 intValue: (*state).constInt64,
6311 var f32_u32 = f2uCvtTab{
6313 cvt2U: ssa.OpCvt32Fto32,
6316 floatValue: (*state).constFloat32,
6317 intValue: func(s *state, t *types.Type, v int64) *ssa.Value { return s.constInt32(t, int32(v)) },
6321 var f64_u32 = f2uCvtTab{
6323 cvt2U: ssa.OpCvt64Fto32,
6326 floatValue: (*state).constFloat64,
6327 intValue: func(s *state, t *types.Type, v int64) *ssa.Value { return s.constInt32(t, int32(v)) },
6331 func (s *state) float32ToUint64(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6332 return s.floatToUint(&f32_u64, n, x, ft, tt)
6334 func (s *state) float64ToUint64(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6335 return s.floatToUint(&f64_u64, n, x, ft, tt)
6338 func (s *state) float32ToUint32(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6339 return s.floatToUint(&f32_u32, n, x, ft, tt)
6342 func (s *state) float64ToUint32(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6343 return s.floatToUint(&f64_u32, n, x, ft, tt)
6346 func (s *state) floatToUint(cvttab *f2uCvtTab, n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6347 // cutoff:=1<<(intY_Size-1)
6348 // if x < floatX(cutoff) {
6349 // result = uintY(x)
6351 // y = x - floatX(cutoff)
6353 // result = z | -(cutoff)
6355 cutoff := cvttab.floatValue(s, ft, float64(cvttab.cutoff))
6356 cmp := s.newValue2(cvttab.ltf, types.Types[types.TBOOL], x, cutoff)
6358 b.Kind = ssa.BlockIf
6360 b.Likely = ssa.BranchLikely
6362 bThen := s.f.NewBlock(ssa.BlockPlain)
6363 bElse := s.f.NewBlock(ssa.BlockPlain)
6364 bAfter := s.f.NewBlock(ssa.BlockPlain)
6368 a0 := s.newValue1(cvttab.cvt2U, tt, x)
6371 bThen.AddEdgeTo(bAfter)
6375 y := s.newValue2(cvttab.subf, ft, x, cutoff)
6376 y = s.newValue1(cvttab.cvt2U, tt, y)
6377 z := cvttab.intValue(s, tt, int64(-cvttab.cutoff))
6378 a1 := s.newValue2(cvttab.or, tt, y, z)
6381 bElse.AddEdgeTo(bAfter)
6383 s.startBlock(bAfter)
6384 return s.variable(n, n.Type())
6387 // dottype generates SSA for a type assertion node.
6388 // commaok indicates whether to panic or return a bool.
6389 // If commaok is false, resok will be nil.
6390 func (s *state) dottype(n *ir.TypeAssertExpr, commaok bool) (res, resok *ssa.Value) {
6391 iface := s.expr(n.X) // input interface
6392 target := s.reflectType(n.Type()) // target type
6393 var targetItab *ssa.Value
6395 targetItab = s.expr(n.ITab)
6397 return s.dottype1(n.Pos(), n.X.Type(), n.Type(), iface, nil, target, targetItab, commaok)
6400 func (s *state) dynamicDottype(n *ir.DynamicTypeAssertExpr, commaok bool) (res, resok *ssa.Value) {
6401 iface := s.expr(n.X)
6402 var source, target, targetItab *ssa.Value
6403 if n.SrcRType != nil {
6404 source = s.expr(n.SrcRType)
6406 if !n.X.Type().IsEmptyInterface() && !n.Type().IsInterface() {
6407 byteptr := s.f.Config.Types.BytePtr
6408 targetItab = s.expr(n.ITab)
6409 // TODO(mdempsky): Investigate whether compiling n.RType could be
6410 // better than loading itab.typ.
6411 target = s.load(byteptr, s.newValue1I(ssa.OpOffPtr, byteptr, int64(types.PtrSize), targetItab)) // itab.typ
6413 target = s.expr(n.RType)
6415 return s.dottype1(n.Pos(), n.X.Type(), n.Type(), iface, source, target, targetItab, commaok)
6418 // dottype1 implements a x.(T) operation. iface is the argument (x), dst is the type we're asserting to (T)
6419 // and src is the type we're asserting from.
6420 // source is the *runtime._type of src
6421 // target is the *runtime._type of dst.
6422 // If src is a nonempty interface and dst is not an interface, targetItab is an itab representing (dst, src). Otherwise it is nil.
6423 // commaok is true if the caller wants a boolean success value. Otherwise, the generated code panics if the conversion fails.
6424 func (s *state) dottype1(pos src.XPos, src, dst *types.Type, iface, source, target, targetItab *ssa.Value, commaok bool) (res, resok *ssa.Value) {
6425 byteptr := s.f.Config.Types.BytePtr
6426 if dst.IsInterface() {
6427 if dst.IsEmptyInterface() {
6428 // Converting to an empty interface.
6429 // Input could be an empty or nonempty interface.
6430 if base.Debug.TypeAssert > 0 {
6431 base.WarnfAt(pos, "type assertion inlined")
6434 // Get itab/type field from input.
6435 itab := s.newValue1(ssa.OpITab, byteptr, iface)
6436 // Conversion succeeds iff that field is not nil.
6437 cond := s.newValue2(ssa.OpNeqPtr, types.Types[types.TBOOL], itab, s.constNil(byteptr))
6439 if src.IsEmptyInterface() && commaok {
6440 // Converting empty interface to empty interface with ,ok is just a nil check.
6444 // Branch on nilness.
6446 b.Kind = ssa.BlockIf
6448 b.Likely = ssa.BranchLikely
6449 bOk := s.f.NewBlock(ssa.BlockPlain)
6450 bFail := s.f.NewBlock(ssa.BlockPlain)
6455 // On failure, panic by calling panicnildottype.
6457 s.rtcall(ir.Syms.Panicnildottype, false, nil, target)
6459 // On success, return (perhaps modified) input interface.
6461 if src.IsEmptyInterface() {
6462 res = iface // Use input interface unchanged.
6465 // Load type out of itab, build interface with existing idata.
6466 off := s.newValue1I(ssa.OpOffPtr, byteptr, int64(types.PtrSize), itab)
6467 typ := s.load(byteptr, off)
6468 idata := s.newValue1(ssa.OpIData, byteptr, iface)
6469 res = s.newValue2(ssa.OpIMake, dst, typ, idata)
6474 // nonempty -> empty
6475 // Need to load type from itab
6476 off := s.newValue1I(ssa.OpOffPtr, byteptr, int64(types.PtrSize), itab)
6477 s.vars[typVar] = s.load(byteptr, off)
6480 // itab is nil, might as well use that as the nil result.
6482 s.vars[typVar] = itab
6486 bEnd := s.f.NewBlock(ssa.BlockPlain)
6488 bFail.AddEdgeTo(bEnd)
6490 idata := s.newValue1(ssa.OpIData, byteptr, iface)
6491 res = s.newValue2(ssa.OpIMake, dst, s.variable(typVar, byteptr), idata)
6493 delete(s.vars, typVar) // no practical effect, just to indicate typVar is no longer live.
6496 // converting to a nonempty interface needs a runtime call.
6497 if base.Debug.TypeAssert > 0 {
6498 base.WarnfAt(pos, "type assertion not inlined")
6501 fn := ir.Syms.AssertI2I
6502 if src.IsEmptyInterface() {
6503 fn = ir.Syms.AssertE2I
6505 data := s.newValue1(ssa.OpIData, types.Types[types.TUNSAFEPTR], iface)
6506 tab := s.newValue1(ssa.OpITab, byteptr, iface)
6507 tab = s.rtcall(fn, true, []*types.Type{byteptr}, target, tab)[0]
6508 return s.newValue2(ssa.OpIMake, dst, tab, data), nil
6510 fn := ir.Syms.AssertI2I2
6511 if src.IsEmptyInterface() {
6512 fn = ir.Syms.AssertE2I2
6514 res = s.rtcall(fn, true, []*types.Type{dst}, target, iface)[0]
6515 resok = s.newValue2(ssa.OpNeqInter, types.Types[types.TBOOL], res, s.constInterface(dst))
6519 if base.Debug.TypeAssert > 0 {
6520 base.WarnfAt(pos, "type assertion inlined")
6523 // Converting to a concrete type.
6524 direct := types.IsDirectIface(dst)
6525 itab := s.newValue1(ssa.OpITab, byteptr, iface) // type word of interface
6526 if base.Debug.TypeAssert > 0 {
6527 base.WarnfAt(pos, "type assertion inlined")
6529 var wantedFirstWord *ssa.Value
6530 if src.IsEmptyInterface() {
6531 // Looking for pointer to target type.
6532 wantedFirstWord = target
6534 // Looking for pointer to itab for target type and source interface.
6535 wantedFirstWord = targetItab
6538 var tmp ir.Node // temporary for use with large types
6539 var addr *ssa.Value // address of tmp
6540 if commaok && !TypeOK(dst) {
6541 // unSSAable type, use temporary.
6542 // TODO: get rid of some of these temporaries.
6543 tmp, addr = s.temp(pos, dst)
6546 cond := s.newValue2(ssa.OpEqPtr, types.Types[types.TBOOL], itab, wantedFirstWord)
6548 b.Kind = ssa.BlockIf
6550 b.Likely = ssa.BranchLikely
6552 bOk := s.f.NewBlock(ssa.BlockPlain)
6553 bFail := s.f.NewBlock(ssa.BlockPlain)
6558 // on failure, panic by calling panicdottype
6562 taddr = s.reflectType(src)
6564 if src.IsEmptyInterface() {
6565 s.rtcall(ir.Syms.PanicdottypeE, false, nil, itab, target, taddr)
6567 s.rtcall(ir.Syms.PanicdottypeI, false, nil, itab, target, taddr)
6570 // on success, return data from interface
6573 return s.newValue1(ssa.OpIData, dst, iface), nil
6575 p := s.newValue1(ssa.OpIData, types.NewPtr(dst), iface)
6576 return s.load(dst, p), nil
6579 // commaok is the more complicated case because we have
6580 // a control flow merge point.
6581 bEnd := s.f.NewBlock(ssa.BlockPlain)
6582 // Note that we need a new valVar each time (unlike okVar where we can
6583 // reuse the variable) because it might have a different type every time.
6584 valVar := ssaMarker("val")
6586 // type assertion succeeded
6590 s.vars[valVar] = s.newValue1(ssa.OpIData, dst, iface)
6592 p := s.newValue1(ssa.OpIData, types.NewPtr(dst), iface)
6593 s.vars[valVar] = s.load(dst, p)
6596 p := s.newValue1(ssa.OpIData, types.NewPtr(dst), iface)
6597 s.move(dst, addr, p)
6599 s.vars[okVar] = s.constBool(true)
6603 // type assertion failed
6606 s.vars[valVar] = s.zeroVal(dst)
6610 s.vars[okVar] = s.constBool(false)
6612 bFail.AddEdgeTo(bEnd)
6617 res = s.variable(valVar, dst)
6618 delete(s.vars, valVar) // no practical effect, just to indicate typVar is no longer live.
6620 res = s.load(dst, addr)
6622 resok = s.variable(okVar, types.Types[types.TBOOL])
6623 delete(s.vars, okVar) // ditto
6627 // temp allocates a temp of type t at position pos
6628 func (s *state) temp(pos src.XPos, t *types.Type) (*ir.Name, *ssa.Value) {
6629 tmp := typecheck.TempAt(pos, s.curfn, t)
6630 if t.HasPointers() {
6631 s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, tmp, s.mem())
6637 // variable returns the value of a variable at the current location.
6638 func (s *state) variable(n ir.Node, t *types.Type) *ssa.Value {
6648 if s.curBlock == s.f.Entry {
6649 // No variable should be live at entry.
6650 s.f.Fatalf("value %v (%v) incorrectly live at entry", n, v)
6652 // Make a FwdRef, which records a value that's live on block input.
6653 // We'll find the matching definition as part of insertPhis.
6654 v = s.newValue0A(ssa.OpFwdRef, t, fwdRefAux{N: n})
6656 if n.Op() == ir.ONAME {
6657 s.addNamedValue(n.(*ir.Name), v)
6662 func (s *state) mem() *ssa.Value {
6663 return s.variable(memVar, types.TypeMem)
6666 func (s *state) addNamedValue(n *ir.Name, v *ssa.Value) {
6667 if n.Class == ir.Pxxx {
6668 // Don't track our marker nodes (memVar etc.).
6671 if ir.IsAutoTmp(n) {
6672 // Don't track temporary variables.
6675 if n.Class == ir.PPARAMOUT {
6676 // Don't track named output values. This prevents return values
6677 // from being assigned too early. See #14591 and #14762. TODO: allow this.
6680 loc := ssa.LocalSlot{N: n, Type: n.Type(), Off: 0}
6681 values, ok := s.f.NamedValues[loc]
6683 s.f.Names = append(s.f.Names, &loc)
6684 s.f.CanonicalLocalSlots[loc] = &loc
6686 s.f.NamedValues[loc] = append(values, v)
6689 // Branch is an unresolved branch.
6690 type Branch struct {
6691 P *obj.Prog // branch instruction
6692 B *ssa.Block // target
6695 // State contains state needed during Prog generation.
6701 // Branches remembers all the branch instructions we've seen
6702 // and where they would like to go.
6705 // JumpTables remembers all the jump tables we've seen.
6706 JumpTables []*ssa.Block
6708 // bstart remembers where each block starts (indexed by block ID)
6711 maxarg int64 // largest frame size for arguments to calls made by the function
6713 // Map from GC safe points to liveness index, generated by
6714 // liveness analysis.
6715 livenessMap liveness.Map
6717 // partLiveArgs includes arguments that may be partially live, for which we
6718 // need to generate instructions that spill the argument registers.
6719 partLiveArgs map[*ir.Name]bool
6721 // lineRunStart records the beginning of the current run of instructions
6722 // within a single block sharing the same line number
6723 // Used to move statement marks to the beginning of such runs.
6724 lineRunStart *obj.Prog
6726 // wasm: The number of values on the WebAssembly stack. This is only used as a safeguard.
6727 OnWasmStackSkipped int
6730 func (s *State) FuncInfo() *obj.FuncInfo {
6731 return s.pp.CurFunc.LSym.Func()
6734 // Prog appends a new Prog.
6735 func (s *State) Prog(as obj.As) *obj.Prog {
6737 if objw.LosesStmtMark(as) {
6740 // Float a statement start to the beginning of any same-line run.
6741 // lineRunStart is reset at block boundaries, which appears to work well.
6742 if s.lineRunStart == nil || s.lineRunStart.Pos.Line() != p.Pos.Line() {
6744 } else if p.Pos.IsStmt() == src.PosIsStmt {
6745 s.lineRunStart.Pos = s.lineRunStart.Pos.WithIsStmt()
6746 p.Pos = p.Pos.WithNotStmt()
6751 // Pc returns the current Prog.
6752 func (s *State) Pc() *obj.Prog {
6756 // SetPos sets the current source position.
6757 func (s *State) SetPos(pos src.XPos) {
6761 // Br emits a single branch instruction and returns the instruction.
6762 // Not all architectures need the returned instruction, but otherwise
6763 // the boilerplate is common to all.
6764 func (s *State) Br(op obj.As, target *ssa.Block) *obj.Prog {
6766 p.To.Type = obj.TYPE_BRANCH
6767 s.Branches = append(s.Branches, Branch{P: p, B: target})
6771 // DebugFriendlySetPosFrom adjusts Pos.IsStmt subject to heuristics
6772 // that reduce "jumpy" line number churn when debugging.
6773 // Spill/fill/copy instructions from the register allocator,
6774 // phi functions, and instructions with a no-pos position
6775 // are examples of instructions that can cause churn.
6776 func (s *State) DebugFriendlySetPosFrom(v *ssa.Value) {
6778 case ssa.OpPhi, ssa.OpCopy, ssa.OpLoadReg, ssa.OpStoreReg:
6779 // These are not statements
6780 s.SetPos(v.Pos.WithNotStmt())
6783 if p != src.NoXPos {
6784 // If the position is defined, update the position.
6785 // Also convert default IsStmt to NotStmt; only
6786 // explicit statement boundaries should appear
6787 // in the generated code.
6788 if p.IsStmt() != src.PosIsStmt {
6789 if s.pp.Pos.IsStmt() == src.PosIsStmt && s.pp.Pos.SameFileAndLine(p) {
6790 // If s.pp.Pos already has a statement mark, then it was set here (below) for
6791 // the previous value. If an actual instruction had been emitted for that
6792 // value, then the statement mark would have been reset. Since the statement
6793 // mark of s.pp.Pos was not reset, this position (file/line) still needs a
6794 // statement mark on an instruction. If file and line for this value are
6795 // the same as the previous value, then the first instruction for this
6796 // value will work to take the statement mark. Return early to avoid
6797 // resetting the statement mark.
6799 // The reset of s.pp.Pos occurs in (*Progs).Prog() -- if it emits
6800 // an instruction, and the instruction's statement mark was set,
6801 // and it is not one of the LosesStmtMark instructions,
6802 // then Prog() resets the statement mark on the (*Progs).Pos.
6806 // Calls use the pos attached to v, but copy the statement mark from State
6810 s.SetPos(s.pp.Pos.WithNotStmt())
6815 // emit argument info (locations on stack) for traceback.
6816 func emitArgInfo(e *ssafn, f *ssa.Func, pp *objw.Progs) {
6817 ft := e.curfn.Type()
6818 if ft.NumRecvs() == 0 && ft.NumParams() == 0 {
6822 x := EmitArgInfo(e.curfn, f.OwnAux.ABIInfo())
6823 x.Set(obj.AttrContentAddressable, true)
6824 e.curfn.LSym.Func().ArgInfo = x
6826 // Emit a funcdata pointing at the arg info data.
6827 p := pp.Prog(obj.AFUNCDATA)
6828 p.From.SetConst(rtabi.FUNCDATA_ArgInfo)
6829 p.To.Type = obj.TYPE_MEM
6830 p.To.Name = obj.NAME_EXTERN
6834 // emit argument info (locations on stack) of f for traceback.
6835 func EmitArgInfo(f *ir.Func, abiInfo *abi.ABIParamResultInfo) *obj.LSym {
6836 x := base.Ctxt.Lookup(fmt.Sprintf("%s.arginfo%d", f.LSym.Name, f.ABI))
6837 // NOTE: do not set ContentAddressable here. This may be referenced from
6838 // assembly code by name (in this case f is a declaration).
6839 // Instead, set it in emitArgInfo above.
6841 PtrSize := int64(types.PtrSize)
6842 uintptrTyp := types.Types[types.TUINTPTR]
6844 isAggregate := func(t *types.Type) bool {
6845 return t.IsStruct() || t.IsArray() || t.IsComplex() || t.IsInterface() || t.IsString() || t.IsSlice()
6848 // Populate the data.
6849 // The data is a stream of bytes, which contains the offsets and sizes of the
6850 // non-aggregate arguments or non-aggregate fields/elements of aggregate-typed
6851 // arguments, along with special "operators". Specifically,
6852 // - for each non-aggrgate arg/field/element, its offset from FP (1 byte) and
6854 // - special operators:
6855 // - 0xff - end of sequence
6856 // - 0xfe - print { (at the start of an aggregate-typed argument)
6857 // - 0xfd - print } (at the end of an aggregate-typed argument)
6858 // - 0xfc - print ... (more args/fields/elements)
6859 // - 0xfb - print _ (offset too large)
6860 // These constants need to be in sync with runtime.traceback.go:printArgs.
6866 _offsetTooLarge = 0xfb
6867 _special = 0xf0 // above this are operators, below this are ordinary offsets
6871 limit = 10 // print no more than 10 args/components
6872 maxDepth = 5 // no more than 5 layers of nesting
6874 // maxLen is a (conservative) upper bound of the byte stream length. For
6875 // each arg/component, it has no more than 2 bytes of data (size, offset),
6876 // and no more than one {, }, ... at each level (it cannot have both the
6877 // data and ... unless it is the last one, just be conservative). Plus 1
6879 maxLen = (maxDepth*3+2)*limit + 1
6884 writebyte := func(o uint8) { wOff = objw.Uint8(x, wOff, o) }
6886 // Write one non-aggrgate arg/field/element.
6887 write1 := func(sz, offset int64) {
6888 if offset >= _special {
6889 writebyte(_offsetTooLarge)
6891 writebyte(uint8(offset))
6892 writebyte(uint8(sz))
6897 // Visit t recursively and write it out.
6898 // Returns whether to continue visiting.
6899 var visitType func(baseOffset int64, t *types.Type, depth int) bool
6900 visitType = func(baseOffset int64, t *types.Type, depth int) bool {
6902 writebyte(_dotdotdot)
6905 if !isAggregate(t) {
6906 write1(t.Size(), baseOffset)
6909 writebyte(_startAgg)
6911 if depth >= maxDepth {
6912 writebyte(_dotdotdot)
6918 case t.IsInterface(), t.IsString():
6919 _ = visitType(baseOffset, uintptrTyp, depth) &&
6920 visitType(baseOffset+PtrSize, uintptrTyp, depth)
6922 _ = visitType(baseOffset, uintptrTyp, depth) &&
6923 visitType(baseOffset+PtrSize, uintptrTyp, depth) &&
6924 visitType(baseOffset+PtrSize*2, uintptrTyp, depth)
6926 _ = visitType(baseOffset, types.FloatForComplex(t), depth) &&
6927 visitType(baseOffset+t.Size()/2, types.FloatForComplex(t), depth)
6929 if t.NumElem() == 0 {
6930 n++ // {} counts as a component
6933 for i := int64(0); i < t.NumElem(); i++ {
6934 if !visitType(baseOffset, t.Elem(), depth) {
6937 baseOffset += t.Elem().Size()
6940 if t.NumFields() == 0 {
6941 n++ // {} counts as a component
6944 for _, field := range t.Fields().Slice() {
6945 if !visitType(baseOffset+field.Offset, field.Type, depth) {
6955 if strings.Contains(f.LSym.Name, "[") {
6956 // Skip the dictionary argument - it is implicit and the user doesn't need to see it.
6960 for _, a := range abiInfo.InParams()[start:] {
6961 if !visitType(a.FrameOffset(abiInfo), a.Type, 0) {
6967 base.Fatalf("ArgInfo too large")
6973 // for wrapper, emit info of wrapped function.
6974 func emitWrappedFuncInfo(e *ssafn, pp *objw.Progs) {
6975 if base.Ctxt.Flag_linkshared {
6976 // Relative reference (SymPtrOff) to another shared object doesn't work.
6981 wfn := e.curfn.WrappedFunc
6986 wsym := wfn.Linksym()
6987 x := base.Ctxt.LookupInit(fmt.Sprintf("%s.wrapinfo", wsym.Name), func(x *obj.LSym) {
6988 objw.SymPtrOff(x, 0, wsym)
6989 x.Set(obj.AttrContentAddressable, true)
6991 e.curfn.LSym.Func().WrapInfo = x
6993 // Emit a funcdata pointing at the wrap info data.
6994 p := pp.Prog(obj.AFUNCDATA)
6995 p.From.SetConst(rtabi.FUNCDATA_WrapInfo)
6996 p.To.Type = obj.TYPE_MEM
6997 p.To.Name = obj.NAME_EXTERN
7001 // genssa appends entries to pp for each instruction in f.
7002 func genssa(f *ssa.Func, pp *objw.Progs) {
7004 s.ABI = f.OwnAux.Fn.ABI()
7006 e := f.Frontend().(*ssafn)
7008 s.livenessMap, s.partLiveArgs = liveness.Compute(e.curfn, f, e.stkptrsize, pp)
7009 emitArgInfo(e, f, pp)
7010 argLiveBlockMap, argLiveValueMap := liveness.ArgLiveness(e.curfn, f, pp)
7012 openDeferInfo := e.curfn.LSym.Func().OpenCodedDeferInfo
7013 if openDeferInfo != nil {
7014 // This function uses open-coded defers -- write out the funcdata
7015 // info that we computed at the end of genssa.
7016 p := pp.Prog(obj.AFUNCDATA)
7017 p.From.SetConst(rtabi.FUNCDATA_OpenCodedDeferInfo)
7018 p.To.Type = obj.TYPE_MEM
7019 p.To.Name = obj.NAME_EXTERN
7020 p.To.Sym = openDeferInfo
7023 emitWrappedFuncInfo(e, pp)
7025 // Remember where each block starts.
7026 s.bstart = make([]*obj.Prog, f.NumBlocks())
7028 var progToValue map[*obj.Prog]*ssa.Value
7029 var progToBlock map[*obj.Prog]*ssa.Block
7030 var valueToProgAfter []*obj.Prog // The first Prog following computation of a value v; v is visible at this point.
7031 gatherPrintInfo := f.PrintOrHtmlSSA || ssa.GenssaDump[f.Name]
7032 if gatherPrintInfo {
7033 progToValue = make(map[*obj.Prog]*ssa.Value, f.NumValues())
7034 progToBlock = make(map[*obj.Prog]*ssa.Block, f.NumBlocks())
7035 f.Logf("genssa %s\n", f.Name)
7036 progToBlock[s.pp.Next] = f.Blocks[0]
7039 if base.Ctxt.Flag_locationlists {
7040 if cap(f.Cache.ValueToProgAfter) < f.NumValues() {
7041 f.Cache.ValueToProgAfter = make([]*obj.Prog, f.NumValues())
7043 valueToProgAfter = f.Cache.ValueToProgAfter[:f.NumValues()]
7044 for i := range valueToProgAfter {
7045 valueToProgAfter[i] = nil
7049 // If the very first instruction is not tagged as a statement,
7050 // debuggers may attribute it to previous function in program.
7051 firstPos := src.NoXPos
7052 for _, v := range f.Entry.Values {
7053 if v.Pos.IsStmt() == src.PosIsStmt && v.Op != ssa.OpArg && v.Op != ssa.OpArgIntReg && v.Op != ssa.OpArgFloatReg && v.Op != ssa.OpLoadReg && v.Op != ssa.OpStoreReg {
7055 v.Pos = firstPos.WithDefaultStmt()
7060 // inlMarks has an entry for each Prog that implements an inline mark.
7061 // It maps from that Prog to the global inlining id of the inlined body
7062 // which should unwind to this Prog's location.
7063 var inlMarks map[*obj.Prog]int32
7064 var inlMarkList []*obj.Prog
7066 // inlMarksByPos maps from a (column 1) source position to the set of
7067 // Progs that are in the set above and have that source position.
7068 var inlMarksByPos map[src.XPos][]*obj.Prog
7070 var argLiveIdx int = -1 // argument liveness info index
7072 // Emit basic blocks
7073 for i, b := range f.Blocks {
7074 s.bstart[b.ID] = s.pp.Next
7075 s.lineRunStart = nil
7076 s.SetPos(s.pp.Pos.WithNotStmt()) // It needs a non-empty Pos, but cannot be a statement boundary (yet).
7078 // Attach a "default" liveness info. Normally this will be
7079 // overwritten in the Values loop below for each Value. But
7080 // for an empty block this will be used for its control
7081 // instruction. We won't use the actual liveness map on a
7082 // control instruction. Just mark it something that is
7083 // preemptible, unless this function is "all unsafe".
7084 s.pp.NextLive = objw.LivenessIndex{StackMapIndex: -1, IsUnsafePoint: liveness.IsUnsafe(f)}
7086 if idx, ok := argLiveBlockMap[b.ID]; ok && idx != argLiveIdx {
7088 p := s.pp.Prog(obj.APCDATA)
7089 p.From.SetConst(rtabi.PCDATA_ArgLiveIndex)
7090 p.To.SetConst(int64(idx))
7093 // Emit values in block
7094 Arch.SSAMarkMoves(&s, b)
7095 for _, v := range b.Values {
7097 s.DebugFriendlySetPosFrom(v)
7099 if v.Op.ResultInArg0() && v.ResultReg() != v.Args[0].Reg() {
7100 v.Fatalf("input[0] and output not in same register %s", v.LongString())
7105 // memory arg needs no code
7107 // input args need no code
7108 case ssa.OpSP, ssa.OpSB:
7110 case ssa.OpSelect0, ssa.OpSelect1, ssa.OpSelectN, ssa.OpMakeResult:
7113 // nothing to do when there's a g register,
7114 // and checkLower complains if there's not
7115 case ssa.OpVarDef, ssa.OpVarLive, ssa.OpKeepAlive, ssa.OpWBend:
7116 // nothing to do; already used by liveness
7120 // nothing to do; no-op conversion for liveness
7121 if v.Args[0].Reg() != v.Reg() {
7122 v.Fatalf("OpConvert should be a no-op: %s; %s", v.Args[0].LongString(), v.LongString())
7125 p := Arch.Ginsnop(s.pp)
7126 if inlMarks == nil {
7127 inlMarks = map[*obj.Prog]int32{}
7128 inlMarksByPos = map[src.XPos][]*obj.Prog{}
7130 inlMarks[p] = v.AuxInt32()
7131 inlMarkList = append(inlMarkList, p)
7132 pos := v.Pos.AtColumn1()
7133 inlMarksByPos[pos] = append(inlMarksByPos[pos], p)
7134 firstPos = src.NoXPos
7137 // Special case for first line in function; move it to the start (which cannot be a register-valued instruction)
7138 if firstPos != src.NoXPos && v.Op != ssa.OpArgIntReg && v.Op != ssa.OpArgFloatReg && v.Op != ssa.OpLoadReg && v.Op != ssa.OpStoreReg {
7140 firstPos = src.NoXPos
7142 // Attach this safe point to the next
7144 s.pp.NextLive = s.livenessMap.Get(v)
7146 // let the backend handle it
7147 Arch.SSAGenValue(&s, v)
7150 if idx, ok := argLiveValueMap[v.ID]; ok && idx != argLiveIdx {
7152 p := s.pp.Prog(obj.APCDATA)
7153 p.From.SetConst(rtabi.PCDATA_ArgLiveIndex)
7154 p.To.SetConst(int64(idx))
7157 if base.Ctxt.Flag_locationlists {
7158 valueToProgAfter[v.ID] = s.pp.Next
7161 if gatherPrintInfo {
7162 for ; x != s.pp.Next; x = x.Link {
7167 // If this is an empty infinite loop, stick a hardware NOP in there so that debuggers are less confused.
7168 if s.bstart[b.ID] == s.pp.Next && len(b.Succs) == 1 && b.Succs[0].Block() == b {
7169 p := Arch.Ginsnop(s.pp)
7170 p.Pos = p.Pos.WithIsStmt()
7171 if b.Pos == src.NoXPos {
7172 b.Pos = p.Pos // It needs a file, otherwise a no-file non-zero line causes confusion. See #35652.
7173 if b.Pos == src.NoXPos {
7174 b.Pos = pp.Text.Pos // Sometimes p.Pos is empty. See #35695.
7177 b.Pos = b.Pos.WithBogusLine() // Debuggers are not good about infinite loops, force a change in line number
7179 // Emit control flow instructions for block
7181 if i < len(f.Blocks)-1 && base.Flag.N == 0 {
7182 // If -N, leave next==nil so every block with successors
7183 // ends in a JMP (except call blocks - plive doesn't like
7184 // select{send,recv} followed by a JMP call). Helps keep
7185 // line numbers for otherwise empty blocks.
7186 next = f.Blocks[i+1]
7190 Arch.SSAGenBlock(&s, b, next)
7191 if gatherPrintInfo {
7192 for ; x != s.pp.Next; x = x.Link {
7197 if f.Blocks[len(f.Blocks)-1].Kind == ssa.BlockExit {
7198 // We need the return address of a panic call to
7199 // still be inside the function in question. So if
7200 // it ends in a call which doesn't return, add a
7201 // nop (which will never execute) after the call.
7204 if openDeferInfo != nil {
7205 // When doing open-coded defers, generate a disconnected call to
7206 // deferreturn and a return. This will be used to during panic
7207 // recovery to unwind the stack and return back to the runtime.
7208 s.pp.NextLive = s.livenessMap.DeferReturn
7209 p := pp.Prog(obj.ACALL)
7210 p.To.Type = obj.TYPE_MEM
7211 p.To.Name = obj.NAME_EXTERN
7212 p.To.Sym = ir.Syms.Deferreturn
7214 // Load results into registers. So when a deferred function
7215 // recovers a panic, it will return to caller with right results.
7216 // The results are already in memory, because they are not SSA'd
7217 // when the function has defers (see canSSAName).
7218 for _, o := range f.OwnAux.ABIInfo().OutParams() {
7219 n := o.Name.(*ir.Name)
7220 rts, offs := o.RegisterTypesAndOffsets()
7221 for i := range o.Registers {
7222 Arch.LoadRegResult(&s, f, rts[i], ssa.ObjRegForAbiReg(o.Registers[i], f.Config), n, offs[i])
7229 if inlMarks != nil {
7232 // We have some inline marks. Try to find other instructions we're
7233 // going to emit anyway, and use those instructions instead of the
7235 for p := pp.Text; p != nil; p = p.Link {
7236 if p.As == obj.ANOP || p.As == obj.AFUNCDATA || p.As == obj.APCDATA || p.As == obj.ATEXT || p.As == obj.APCALIGN || Arch.LinkArch.Family == sys.Wasm {
7237 // Don't use 0-sized instructions as inline marks, because we need
7238 // to identify inline mark instructions by pc offset.
7239 // (Some of these instructions are sometimes zero-sized, sometimes not.
7240 // We must not use anything that even might be zero-sized.)
7241 // TODO: are there others?
7244 if _, ok := inlMarks[p]; ok {
7245 // Don't use inline marks themselves. We don't know
7246 // whether they will be zero-sized or not yet.
7249 if p.As == obj.ACALL || p.As == obj.ADUFFCOPY || p.As == obj.ADUFFZERO {
7252 pos := p.Pos.AtColumn1()
7253 s := inlMarksByPos[pos]
7257 for _, m := range s {
7258 // We found an instruction with the same source position as
7259 // some of the inline marks.
7260 // Use this instruction instead.
7261 p.Pos = p.Pos.WithIsStmt() // promote position to a statement
7262 pp.CurFunc.LSym.Func().AddInlMark(p, inlMarks[m])
7263 // Make the inline mark a real nop, so it doesn't generate any code.
7269 delete(inlMarksByPos, pos)
7271 // Any unmatched inline marks now need to be added to the inlining tree (and will generate a nop instruction).
7272 for _, p := range inlMarkList {
7273 if p.As != obj.ANOP {
7274 pp.CurFunc.LSym.Func().AddInlMark(p, inlMarks[p])
7278 if e.stksize == 0 && !hasCall {
7279 // Frameless leaf function. It doesn't need any preamble,
7280 // so make sure its first instruction isn't from an inlined callee.
7281 // If it is, add a nop at the start of the function with a position
7282 // equal to the start of the function.
7283 // This ensures that runtime.FuncForPC(uintptr(reflect.ValueOf(fn).Pointer())).Name()
7284 // returns the right answer. See issue 58300.
7285 for p := pp.Text; p != nil; p = p.Link {
7286 if p.As == obj.AFUNCDATA || p.As == obj.APCDATA || p.As == obj.ATEXT || p.As == obj.ANOP {
7289 if base.Ctxt.PosTable.Pos(p.Pos).Base().InliningIndex() >= 0 {
7290 // Make a real (not 0-sized) nop.
7291 nop := Arch.Ginsnop(pp)
7292 nop.Pos = e.curfn.Pos().WithIsStmt()
7294 // Unfortunately, Ginsnop puts the instruction at the
7295 // end of the list. Move it up to just before p.
7297 // Unlink from the current list.
7298 for x := pp.Text; x != nil; x = x.Link {
7304 // Splice in right before p.
7305 for x := pp.Text; x != nil; x = x.Link {
7318 if base.Ctxt.Flag_locationlists {
7319 var debugInfo *ssa.FuncDebug
7320 debugInfo = e.curfn.DebugInfo.(*ssa.FuncDebug)
7321 if e.curfn.ABI == obj.ABIInternal && base.Flag.N != 0 {
7322 ssa.BuildFuncDebugNoOptimized(base.Ctxt, f, base.Debug.LocationLists > 1, StackOffset, debugInfo)
7324 ssa.BuildFuncDebug(base.Ctxt, f, base.Debug.LocationLists, StackOffset, debugInfo)
7327 idToIdx := make([]int, f.NumBlocks())
7328 for i, b := range f.Blocks {
7331 // Note that at this moment, Prog.Pc is a sequence number; it's
7332 // not a real PC until after assembly, so this mapping has to
7334 debugInfo.GetPC = func(b, v ssa.ID) int64 {
7336 case ssa.BlockStart.ID:
7337 if b == f.Entry.ID {
7338 return 0 // Start at the very beginning, at the assembler-generated prologue.
7339 // this should only happen for function args (ssa.OpArg)
7342 case ssa.BlockEnd.ID:
7343 blk := f.Blocks[idToIdx[b]]
7344 nv := len(blk.Values)
7345 return valueToProgAfter[blk.Values[nv-1].ID].Pc
7346 case ssa.FuncEnd.ID:
7347 return e.curfn.LSym.Size
7349 return valueToProgAfter[v].Pc
7354 // Resolve branches, and relax DefaultStmt into NotStmt
7355 for _, br := range s.Branches {
7356 br.P.To.SetTarget(s.bstart[br.B.ID])
7357 if br.P.Pos.IsStmt() != src.PosIsStmt {
7358 br.P.Pos = br.P.Pos.WithNotStmt()
7359 } else if v0 := br.B.FirstPossibleStmtValue(); v0 != nil && v0.Pos.Line() == br.P.Pos.Line() && v0.Pos.IsStmt() == src.PosIsStmt {
7360 br.P.Pos = br.P.Pos.WithNotStmt()
7365 // Resolve jump table destinations.
7366 for _, jt := range s.JumpTables {
7367 // Convert from *Block targets to *Prog targets.
7368 targets := make([]*obj.Prog, len(jt.Succs))
7369 for i, e := range jt.Succs {
7370 targets[i] = s.bstart[e.Block().ID]
7372 // Add to list of jump tables to be resolved at assembly time.
7373 // The assembler converts from *Prog entries to absolute addresses
7374 // once it knows instruction byte offsets.
7375 fi := pp.CurFunc.LSym.Func()
7376 fi.JumpTables = append(fi.JumpTables, obj.JumpTable{Sym: jt.Aux.(*obj.LSym), Targets: targets})
7379 if e.log { // spew to stdout
7381 for p := pp.Text; p != nil; p = p.Link {
7382 if p.Pos.IsKnown() && p.InnermostFilename() != filename {
7383 filename = p.InnermostFilename()
7384 f.Logf("# %s\n", filename)
7388 if v, ok := progToValue[p]; ok {
7390 } else if b, ok := progToBlock[p]; ok {
7393 s = " " // most value and branch strings are 2-3 characters long
7395 f.Logf(" %-6s\t%.5d (%s)\t%s\n", s, p.Pc, p.InnermostLineNumber(), p.InstructionString())
7398 if f.HTMLWriter != nil { // spew to ssa.html
7399 var buf strings.Builder
7400 buf.WriteString("<code>")
7401 buf.WriteString("<dl class=\"ssa-gen\">")
7403 for p := pp.Text; p != nil; p = p.Link {
7404 // Don't spam every line with the file name, which is often huge.
7405 // Only print changes, and "unknown" is not a change.
7406 if p.Pos.IsKnown() && p.InnermostFilename() != filename {
7407 filename = p.InnermostFilename()
7408 buf.WriteString("<dt class=\"ssa-prog-src\"></dt><dd class=\"ssa-prog\">")
7409 buf.WriteString(html.EscapeString("# " + filename))
7410 buf.WriteString("</dd>")
7413 buf.WriteString("<dt class=\"ssa-prog-src\">")
7414 if v, ok := progToValue[p]; ok {
7415 buf.WriteString(v.HTML())
7416 } else if b, ok := progToBlock[p]; ok {
7417 buf.WriteString("<b>" + b.HTML() + "</b>")
7419 buf.WriteString("</dt>")
7420 buf.WriteString("<dd class=\"ssa-prog\">")
7421 fmt.Fprintf(&buf, "%.5d <span class=\"l%v line-number\">(%s)</span> %s", p.Pc, p.InnermostLineNumber(), p.InnermostLineNumberHTML(), html.EscapeString(p.InstructionString()))
7422 buf.WriteString("</dd>")
7424 buf.WriteString("</dl>")
7425 buf.WriteString("</code>")
7426 f.HTMLWriter.WriteColumn("genssa", "genssa", "ssa-prog", buf.String())
7428 if ssa.GenssaDump[f.Name] {
7429 fi := f.DumpFileForPhase("genssa")
7432 // inliningDiffers if any filename changes or if any line number except the innermost (last index) changes.
7433 inliningDiffers := func(a, b []src.Pos) bool {
7434 if len(a) != len(b) {
7438 if a[i].Filename() != b[i].Filename() {
7441 if i != len(a)-1 && a[i].Line() != b[i].Line() {
7448 var allPosOld []src.Pos
7449 var allPos []src.Pos
7451 for p := pp.Text; p != nil; p = p.Link {
7452 if p.Pos.IsKnown() {
7454 p.Ctxt.AllPos(p.Pos, func(pos src.Pos) { allPos = append(allPos, pos) })
7455 if inliningDiffers(allPos, allPosOld) {
7456 for _, pos := range allPos {
7457 fmt.Fprintf(fi, "# %s:%d\n", pos.Filename(), pos.Line())
7459 allPos, allPosOld = allPosOld, allPos // swap, not copy, so that they do not share slice storage.
7464 if v, ok := progToValue[p]; ok {
7466 } else if b, ok := progToBlock[p]; ok {
7469 s = " " // most value and branch strings are 2-3 characters long
7471 fmt.Fprintf(fi, " %-6s\t%.5d %s\t%s\n", s, p.Pc, ssa.StmtString(p.Pos), p.InstructionString())
7479 f.HTMLWriter.Close()
7483 func defframe(s *State, e *ssafn, f *ssa.Func) {
7486 s.maxarg = types.RoundUp(s.maxarg, e.stkalign)
7487 frame := s.maxarg + e.stksize
7488 if Arch.PadFrame != nil {
7489 frame = Arch.PadFrame(frame)
7492 // Fill in argument and frame size.
7493 pp.Text.To.Type = obj.TYPE_TEXTSIZE
7494 pp.Text.To.Val = int32(types.RoundUp(f.OwnAux.ArgWidth(), int64(types.RegSize)))
7495 pp.Text.To.Offset = frame
7499 // Insert code to spill argument registers if the named slot may be partially
7500 // live. That is, the named slot is considered live by liveness analysis,
7501 // (because a part of it is live), but we may not spill all parts into the
7502 // slot. This can only happen with aggregate-typed arguments that are SSA-able
7503 // and not address-taken (for non-SSA-able or address-taken arguments we always
7505 // Note: spilling is unnecessary in the -N/no-optimize case, since all values
7506 // will be considered non-SSAable and spilled up front.
7507 // TODO(register args) Make liveness more fine-grained to that partial spilling is okay.
7508 if f.OwnAux.ABIInfo().InRegistersUsed() != 0 && base.Flag.N == 0 {
7509 // First, see if it is already spilled before it may be live. Look for a spill
7510 // in the entry block up to the first safepoint.
7511 type nameOff struct {
7515 partLiveArgsSpilled := make(map[nameOff]bool)
7516 for _, v := range f.Entry.Values {
7520 if v.Op != ssa.OpStoreReg || v.Args[0].Op != ssa.OpArgIntReg {
7523 n, off := ssa.AutoVar(v)
7524 if n.Class != ir.PPARAM || n.Addrtaken() || !TypeOK(n.Type()) || !s.partLiveArgs[n] {
7527 partLiveArgsSpilled[nameOff{n, off}] = true
7530 // Then, insert code to spill registers if not already.
7531 for _, a := range f.OwnAux.ABIInfo().InParams() {
7532 n, ok := a.Name.(*ir.Name)
7533 if !ok || n.Addrtaken() || !TypeOK(n.Type()) || !s.partLiveArgs[n] || len(a.Registers) <= 1 {
7536 rts, offs := a.RegisterTypesAndOffsets()
7537 for i := range a.Registers {
7538 if !rts[i].HasPointers() {
7541 if partLiveArgsSpilled[nameOff{n, offs[i]}] {
7542 continue // already spilled
7544 reg := ssa.ObjRegForAbiReg(a.Registers[i], f.Config)
7545 p = Arch.SpillArgReg(pp, p, f, rts[i], reg, n, offs[i])
7550 // Insert code to zero ambiguously live variables so that the
7551 // garbage collector only sees initialized values when it
7552 // looks for pointers.
7555 // Opaque state for backend to use. Current backends use it to
7556 // keep track of which helper registers have been zeroed.
7559 // Iterate through declarations. Autos are sorted in decreasing
7560 // frame offset order.
7561 for _, n := range e.curfn.Dcl {
7565 if n.Class != ir.PAUTO {
7566 e.Fatalf(n.Pos(), "needzero class %d", n.Class)
7568 if n.Type().Size()%int64(types.PtrSize) != 0 || n.FrameOffset()%int64(types.PtrSize) != 0 || n.Type().Size() == 0 {
7569 e.Fatalf(n.Pos(), "var %L has size %d offset %d", n, n.Type().Size(), n.Offset_)
7572 if lo != hi && n.FrameOffset()+n.Type().Size() >= lo-int64(2*types.RegSize) {
7573 // Merge with range we already have.
7574 lo = n.FrameOffset()
7579 p = Arch.ZeroRange(pp, p, frame+lo, hi-lo, &state)
7582 lo = n.FrameOffset()
7583 hi = lo + n.Type().Size()
7586 // Zero final range.
7587 Arch.ZeroRange(pp, p, frame+lo, hi-lo, &state)
7590 // For generating consecutive jump instructions to model a specific branching
7591 type IndexJump struct {
7596 func (s *State) oneJump(b *ssa.Block, jump *IndexJump) {
7597 p := s.Br(jump.Jump, b.Succs[jump.Index].Block())
7601 // CombJump generates combinational instructions (2 at present) for a block jump,
7602 // thereby the behaviour of non-standard condition codes could be simulated
7603 func (s *State) CombJump(b, next *ssa.Block, jumps *[2][2]IndexJump) {
7605 case b.Succs[0].Block():
7606 s.oneJump(b, &jumps[0][0])
7607 s.oneJump(b, &jumps[0][1])
7608 case b.Succs[1].Block():
7609 s.oneJump(b, &jumps[1][0])
7610 s.oneJump(b, &jumps[1][1])
7613 if b.Likely != ssa.BranchUnlikely {
7614 s.oneJump(b, &jumps[1][0])
7615 s.oneJump(b, &jumps[1][1])
7616 q = s.Br(obj.AJMP, b.Succs[1].Block())
7618 s.oneJump(b, &jumps[0][0])
7619 s.oneJump(b, &jumps[0][1])
7620 q = s.Br(obj.AJMP, b.Succs[0].Block())
7626 // AddAux adds the offset in the aux fields (AuxInt and Aux) of v to a.
7627 func AddAux(a *obj.Addr, v *ssa.Value) {
7628 AddAux2(a, v, v.AuxInt)
7630 func AddAux2(a *obj.Addr, v *ssa.Value, offset int64) {
7631 if a.Type != obj.TYPE_MEM && a.Type != obj.TYPE_ADDR {
7632 v.Fatalf("bad AddAux addr %v", a)
7634 // add integer offset
7637 // If no additional symbol offset, we're done.
7641 // Add symbol's offset from its base register.
7642 switch n := v.Aux.(type) {
7644 a.Name = obj.NAME_EXTERN
7647 a.Name = obj.NAME_EXTERN
7650 if n.Class == ir.PPARAM || (n.Class == ir.PPARAMOUT && !n.IsOutputParamInRegisters()) {
7651 a.Name = obj.NAME_PARAM
7652 a.Sym = ir.Orig(n).(*ir.Name).Linksym()
7653 a.Offset += n.FrameOffset()
7656 a.Name = obj.NAME_AUTO
7657 if n.Class == ir.PPARAMOUT {
7658 a.Sym = ir.Orig(n).(*ir.Name).Linksym()
7662 a.Offset += n.FrameOffset()
7664 v.Fatalf("aux in %s not implemented %#v", v, v.Aux)
7668 // extendIndex extends v to a full int width.
7669 // panic with the given kind if v does not fit in an int (only on 32-bit archs).
7670 func (s *state) extendIndex(idx, len *ssa.Value, kind ssa.BoundsKind, bounded bool) *ssa.Value {
7671 size := idx.Type.Size()
7672 if size == s.config.PtrSize {
7675 if size > s.config.PtrSize {
7676 // truncate 64-bit indexes on 32-bit pointer archs. Test the
7677 // high word and branch to out-of-bounds failure if it is not 0.
7679 if idx.Type.IsSigned() {
7680 lo = s.newValue1(ssa.OpInt64Lo, types.Types[types.TINT], idx)
7682 lo = s.newValue1(ssa.OpInt64Lo, types.Types[types.TUINT], idx)
7684 if bounded || base.Flag.B != 0 {
7687 bNext := s.f.NewBlock(ssa.BlockPlain)
7688 bPanic := s.f.NewBlock(ssa.BlockExit)
7689 hi := s.newValue1(ssa.OpInt64Hi, types.Types[types.TUINT32], idx)
7690 cmp := s.newValue2(ssa.OpEq32, types.Types[types.TBOOL], hi, s.constInt32(types.Types[types.TUINT32], 0))
7691 if !idx.Type.IsSigned() {
7693 case ssa.BoundsIndex:
7694 kind = ssa.BoundsIndexU
7695 case ssa.BoundsSliceAlen:
7696 kind = ssa.BoundsSliceAlenU
7697 case ssa.BoundsSliceAcap:
7698 kind = ssa.BoundsSliceAcapU
7699 case ssa.BoundsSliceB:
7700 kind = ssa.BoundsSliceBU
7701 case ssa.BoundsSlice3Alen:
7702 kind = ssa.BoundsSlice3AlenU
7703 case ssa.BoundsSlice3Acap:
7704 kind = ssa.BoundsSlice3AcapU
7705 case ssa.BoundsSlice3B:
7706 kind = ssa.BoundsSlice3BU
7707 case ssa.BoundsSlice3C:
7708 kind = ssa.BoundsSlice3CU
7712 b.Kind = ssa.BlockIf
7714 b.Likely = ssa.BranchLikely
7718 s.startBlock(bPanic)
7719 mem := s.newValue4I(ssa.OpPanicExtend, types.TypeMem, int64(kind), hi, lo, len, s.mem())
7720 s.endBlock().SetControl(mem)
7726 // Extend value to the required size
7728 if idx.Type.IsSigned() {
7729 switch 10*size + s.config.PtrSize {
7731 op = ssa.OpSignExt8to32
7733 op = ssa.OpSignExt8to64
7735 op = ssa.OpSignExt16to32
7737 op = ssa.OpSignExt16to64
7739 op = ssa.OpSignExt32to64
7741 s.Fatalf("bad signed index extension %s", idx.Type)
7744 switch 10*size + s.config.PtrSize {
7746 op = ssa.OpZeroExt8to32
7748 op = ssa.OpZeroExt8to64
7750 op = ssa.OpZeroExt16to32
7752 op = ssa.OpZeroExt16to64
7754 op = ssa.OpZeroExt32to64
7756 s.Fatalf("bad unsigned index extension %s", idx.Type)
7759 return s.newValue1(op, types.Types[types.TINT], idx)
7762 // CheckLoweredPhi checks that regalloc and stackalloc correctly handled phi values.
7763 // Called during ssaGenValue.
7764 func CheckLoweredPhi(v *ssa.Value) {
7765 if v.Op != ssa.OpPhi {
7766 v.Fatalf("CheckLoweredPhi called with non-phi value: %v", v.LongString())
7768 if v.Type.IsMemory() {
7772 loc := f.RegAlloc[v.ID]
7773 for _, a := range v.Args {
7774 if aloc := f.RegAlloc[a.ID]; aloc != loc { // TODO: .Equal() instead?
7775 v.Fatalf("phi arg at different location than phi: %v @ %s, but arg %v @ %s\n%s\n", v, loc, a, aloc, v.Block.Func)
7780 // CheckLoweredGetClosurePtr checks that v is the first instruction in the function's entry block,
7781 // except for incoming in-register arguments.
7782 // The output of LoweredGetClosurePtr is generally hardwired to the correct register.
7783 // That register contains the closure pointer on closure entry.
7784 func CheckLoweredGetClosurePtr(v *ssa.Value) {
7785 entry := v.Block.Func.Entry
7786 if entry != v.Block {
7787 base.Fatalf("in %s, badly placed LoweredGetClosurePtr: %v %v", v.Block.Func.Name, v.Block, v)
7789 for _, w := range entry.Values {
7794 case ssa.OpArgIntReg, ssa.OpArgFloatReg:
7797 base.Fatalf("in %s, badly placed LoweredGetClosurePtr: %v %v", v.Block.Func.Name, v.Block, v)
7802 // CheckArgReg ensures that v is in the function's entry block.
7803 func CheckArgReg(v *ssa.Value) {
7804 entry := v.Block.Func.Entry
7805 if entry != v.Block {
7806 base.Fatalf("in %s, badly placed ArgIReg or ArgFReg: %v %v", v.Block.Func.Name, v.Block, v)
7810 func AddrAuto(a *obj.Addr, v *ssa.Value) {
7811 n, off := ssa.AutoVar(v)
7812 a.Type = obj.TYPE_MEM
7814 a.Reg = int16(Arch.REGSP)
7815 a.Offset = n.FrameOffset() + off
7816 if n.Class == ir.PPARAM || (n.Class == ir.PPARAMOUT && !n.IsOutputParamInRegisters()) {
7817 a.Name = obj.NAME_PARAM
7819 a.Name = obj.NAME_AUTO
7823 // Call returns a new CALL instruction for the SSA value v.
7824 // It uses PrepareCall to prepare the call.
7825 func (s *State) Call(v *ssa.Value) *obj.Prog {
7826 pPosIsStmt := s.pp.Pos.IsStmt() // The statement-ness fo the call comes from ssaGenState
7829 p := s.Prog(obj.ACALL)
7830 if pPosIsStmt == src.PosIsStmt {
7831 p.Pos = v.Pos.WithIsStmt()
7833 p.Pos = v.Pos.WithNotStmt()
7835 if sym, ok := v.Aux.(*ssa.AuxCall); ok && sym.Fn != nil {
7836 p.To.Type = obj.TYPE_MEM
7837 p.To.Name = obj.NAME_EXTERN
7840 // TODO(mdempsky): Can these differences be eliminated?
7841 switch Arch.LinkArch.Family {
7842 case sys.AMD64, sys.I386, sys.PPC64, sys.RISCV64, sys.S390X, sys.Wasm:
7843 p.To.Type = obj.TYPE_REG
7844 case sys.ARM, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64:
7845 p.To.Type = obj.TYPE_MEM
7847 base.Fatalf("unknown indirect call family")
7849 p.To.Reg = v.Args[0].Reg()
7854 // TailCall returns a new tail call instruction for the SSA value v.
7855 // It is like Call, but for a tail call.
7856 func (s *State) TailCall(v *ssa.Value) *obj.Prog {
7862 // PrepareCall prepares to emit a CALL instruction for v and does call-related bookkeeping.
7863 // It must be called immediately before emitting the actual CALL instruction,
7864 // since it emits PCDATA for the stack map at the call (calls are safe points).
7865 func (s *State) PrepareCall(v *ssa.Value) {
7866 idx := s.livenessMap.Get(v)
7867 if !idx.StackMapValid() {
7868 // See Liveness.hasStackMap.
7869 if sym, ok := v.Aux.(*ssa.AuxCall); !ok || !(sym.Fn == ir.Syms.WBZero || sym.Fn == ir.Syms.WBMove) {
7870 base.Fatalf("missing stack map index for %v", v.LongString())
7874 call, ok := v.Aux.(*ssa.AuxCall)
7877 // Record call graph information for nowritebarrierrec
7879 if nowritebarrierrecCheck != nil {
7880 nowritebarrierrecCheck.recordCall(s.pp.CurFunc, call.Fn, v.Pos)
7884 if s.maxarg < v.AuxInt {
7889 // UseArgs records the fact that an instruction needs a certain amount of
7890 // callee args space for its use.
7891 func (s *State) UseArgs(n int64) {
7897 // fieldIdx finds the index of the field referred to by the ODOT node n.
7898 func fieldIdx(n *ir.SelectorExpr) int {
7901 panic("ODOT's LHS is not a struct")
7904 for i, f := range t.Fields().Slice() {
7906 if f.Offset != n.Offset() {
7907 panic("field offset doesn't match")
7912 panic(fmt.Sprintf("can't find field in expr %v\n", n))
7914 // TODO: keep the result of this function somewhere in the ODOT Node
7915 // so we don't have to recompute it each time we need it.
7918 // ssafn holds frontend information about a function that the backend is processing.
7919 // It also exports a bunch of compiler services for the ssa backend.
7922 strings map[string]*obj.LSym // map from constant string to data symbols
7923 stksize int64 // stack size for current frame
7924 stkptrsize int64 // prefix of stack containing pointers
7926 // alignment for current frame.
7927 // NOTE: when stkalign > PtrSize, currently this only ensures the offsets of
7928 // objects in the stack frame are aligned. The stack pointer is still aligned
7932 log bool // print ssa debug to the stdout
7935 // StringData returns a symbol which
7936 // is the data component of a global string constant containing s.
7937 func (e *ssafn) StringData(s string) *obj.LSym {
7938 if aux, ok := e.strings[s]; ok {
7941 if e.strings == nil {
7942 e.strings = make(map[string]*obj.LSym)
7944 data := staticdata.StringSym(e.curfn.Pos(), s)
7949 func (e *ssafn) Auto(pos src.XPos, t *types.Type) *ir.Name {
7950 return typecheck.TempAt(pos, e.curfn, t) // Note: adds new auto to e.curfn.Func.Dcl list
7953 // SplitSlot returns a slot representing the data of parent starting at offset.
7954 func (e *ssafn) SplitSlot(parent *ssa.LocalSlot, suffix string, offset int64, t *types.Type) ssa.LocalSlot {
7957 if node.Class != ir.PAUTO || node.Addrtaken() {
7958 // addressed things and non-autos retain their parents (i.e., cannot truly be split)
7959 return ssa.LocalSlot{N: node, Type: t, Off: parent.Off + offset}
7962 s := &types.Sym{Name: node.Sym().Name + suffix, Pkg: types.LocalPkg}
7963 n := ir.NewNameAt(parent.N.Pos(), s)
7965 ir.AsNode(s.Def).Name().SetUsed(true)
7968 n.SetEsc(ir.EscNever)
7970 e.curfn.Dcl = append(e.curfn.Dcl, n)
7972 return ssa.LocalSlot{N: n, Type: t, Off: 0, SplitOf: parent, SplitOffset: offset}
7975 func (e *ssafn) CanSSA(t *types.Type) bool {
7979 // Logf logs a message from the compiler.
7980 func (e *ssafn) Logf(msg string, args ...interface{}) {
7982 fmt.Printf(msg, args...)
7986 func (e *ssafn) Log() bool {
7990 // Fatalf reports a compiler error and exits.
7991 func (e *ssafn) Fatalf(pos src.XPos, msg string, args ...interface{}) {
7993 nargs := append([]interface{}{ir.FuncName(e.curfn)}, args...)
7994 base.Fatalf("'%s': "+msg, nargs...)
7997 // Warnl reports a "warning", which is usually flag-triggered
7998 // logging output for the benefit of tests.
7999 func (e *ssafn) Warnl(pos src.XPos, fmt_ string, args ...interface{}) {
8000 base.WarnfAt(pos, fmt_, args...)
8003 func (e *ssafn) Debug_checknil() bool {
8004 return base.Debug.Nil != 0
8007 func (e *ssafn) UseWriteBarrier() bool {
8011 func (e *ssafn) Syslook(name string) *obj.LSym {
8013 case "goschedguarded":
8014 return ir.Syms.Goschedguarded
8015 case "writeBarrier":
8016 return ir.Syms.WriteBarrier
8018 return ir.Syms.WBZero
8020 return ir.Syms.WBMove
8021 case "cgoCheckMemmove":
8022 return ir.Syms.CgoCheckMemmove
8023 case "cgoCheckPtrWrite":
8024 return ir.Syms.CgoCheckPtrWrite
8026 e.Fatalf(src.NoXPos, "unknown Syslook func %v", name)
8030 func (e *ssafn) MyImportPath() string {
8031 return base.Ctxt.Pkgpath
8034 func (e *ssafn) Func() *ir.Func {
8038 func clobberBase(n ir.Node) ir.Node {
8039 if n.Op() == ir.ODOT {
8040 n := n.(*ir.SelectorExpr)
8041 if n.X.Type().NumFields() == 1 {
8042 return clobberBase(n.X)
8045 if n.Op() == ir.OINDEX {
8046 n := n.(*ir.IndexExpr)
8047 if n.X.Type().IsArray() && n.X.Type().NumElem() == 1 {
8048 return clobberBase(n.X)
8054 // callTargetLSym returns the correct LSym to call 'callee' using its ABI.
8055 func callTargetLSym(callee *ir.Name) *obj.LSym {
8056 if callee.Func == nil {
8057 // TODO(austin): This happens in case of interface method I.M from imported package.
8058 // It's ABIInternal, and would be better if callee.Func was never nil and we didn't
8060 return callee.Linksym()
8063 return callee.LinksymABI(callee.Func.ABI)
8066 func min8(a, b int8) int8 {
8073 func max8(a, b int8) int8 {
8080 // deferstruct makes a runtime._defer structure.
8081 func deferstruct() *types.Type {
8082 makefield := func(name string, typ *types.Type) *types.Field {
8083 // Unlike the global makefield function, this one needs to set Pkg
8084 // because these types might be compared (in SSA CSE sorting).
8085 // TODO: unify this makefield and the global one above.
8086 sym := &types.Sym{Name: name, Pkg: types.LocalPkg}
8087 return types.NewField(src.NoXPos, sym, typ)
8089 // These fields must match the ones in runtime/runtime2.go:_defer and
8090 // (*state).call above.
8091 fields := []*types.Field{
8092 makefield("started", types.Types[types.TBOOL]),
8093 makefield("heap", types.Types[types.TBOOL]),
8094 makefield("openDefer", types.Types[types.TBOOL]),
8095 makefield("sp", types.Types[types.TUINTPTR]),
8096 makefield("pc", types.Types[types.TUINTPTR]),
8097 // Note: the types here don't really matter. Defer structures
8098 // are always scanned explicitly during stack copying and GC,
8099 // so we make them uintptr type even though they are real pointers.
8100 makefield("fn", types.Types[types.TUINTPTR]),
8101 makefield("_panic", types.Types[types.TUINTPTR]),
8102 makefield("link", types.Types[types.TUINTPTR]),
8103 makefield("fd", types.Types[types.TUINTPTR]),
8104 makefield("varp", types.Types[types.TUINTPTR]),
8105 makefield("framepc", types.Types[types.TUINTPTR]),
8108 // build struct holding the above fields
8109 s := types.NewStruct(fields)
8111 types.CalcStructSize(s)
8115 // SpillSlotAddr uses LocalSlot information to initialize an obj.Addr
8116 // The resulting addr is used in a non-standard context -- in the prologue
8117 // of a function, before the frame has been constructed, so the standard
8118 // addressing for the parameters will be wrong.
8119 func SpillSlotAddr(spill ssa.Spill, baseReg int16, extraOffset int64) obj.Addr {
8121 Name: obj.NAME_NONE,
8124 Offset: spill.Offset + extraOffset,
8129 BoundsCheckFunc [ssa.BoundsKindCount]*obj.LSym
8130 ExtendCheckFunc [ssa.BoundsKindCount]*obj.LSym