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")
954 // startBlock sets the current block we're generating code in to b.
955 func (s *state) startBlock(b *ssa.Block) {
956 if s.curBlock != nil {
957 s.Fatalf("starting block %v when block %v has not ended", b, s.curBlock)
960 s.vars = map[ir.Node]*ssa.Value{}
961 for n := range s.fwdVars {
966 // endBlock marks the end of generating code for the current block.
967 // Returns the (former) current block. Returns nil if there is no current
968 // block, i.e. if no code flows to the current execution point.
969 func (s *state) endBlock() *ssa.Block {
974 for len(s.defvars) <= int(b.ID) {
975 s.defvars = append(s.defvars, nil)
977 s.defvars[b.ID] = s.vars
981 // Empty plain blocks get the line of their successor (handled after all blocks created),
982 // except for increment blocks in For statements (handled in ssa conversion of OFOR),
983 // and for blocks ending in GOTO/BREAK/CONTINUE.
991 // pushLine pushes a line number on the line number stack.
992 func (s *state) pushLine(line src.XPos) {
994 // the frontend may emit node with line number missing,
995 // use the parent line number in this case.
997 if base.Flag.K != 0 {
998 base.Warn("buildssa: unknown position (line 0)")
1004 s.line = append(s.line, line)
1007 // popLine pops the top of the line number stack.
1008 func (s *state) popLine() {
1009 s.line = s.line[:len(s.line)-1]
1012 // peekPos peeks the top of the line number stack.
1013 func (s *state) peekPos() src.XPos {
1014 return s.line[len(s.line)-1]
1017 // newValue0 adds a new value with no arguments to the current block.
1018 func (s *state) newValue0(op ssa.Op, t *types.Type) *ssa.Value {
1019 return s.curBlock.NewValue0(s.peekPos(), op, t)
1022 // newValue0A adds a new value with no arguments and an aux value to the current block.
1023 func (s *state) newValue0A(op ssa.Op, t *types.Type, aux ssa.Aux) *ssa.Value {
1024 return s.curBlock.NewValue0A(s.peekPos(), op, t, aux)
1027 // newValue0I adds a new value with no arguments and an auxint value to the current block.
1028 func (s *state) newValue0I(op ssa.Op, t *types.Type, auxint int64) *ssa.Value {
1029 return s.curBlock.NewValue0I(s.peekPos(), op, t, auxint)
1032 // newValue1 adds a new value with one argument to the current block.
1033 func (s *state) newValue1(op ssa.Op, t *types.Type, arg *ssa.Value) *ssa.Value {
1034 return s.curBlock.NewValue1(s.peekPos(), op, t, arg)
1037 // newValue1A adds a new value with one argument and an aux value to the current block.
1038 func (s *state) newValue1A(op ssa.Op, t *types.Type, aux ssa.Aux, arg *ssa.Value) *ssa.Value {
1039 return s.curBlock.NewValue1A(s.peekPos(), op, t, aux, arg)
1042 // newValue1Apos adds a new value with one argument and an aux value to the current block.
1043 // isStmt determines whether the created values may be a statement or not
1044 // (i.e., false means never, yes means maybe).
1045 func (s *state) newValue1Apos(op ssa.Op, t *types.Type, aux ssa.Aux, arg *ssa.Value, isStmt bool) *ssa.Value {
1047 return s.curBlock.NewValue1A(s.peekPos(), op, t, aux, arg)
1049 return s.curBlock.NewValue1A(s.peekPos().WithNotStmt(), op, t, aux, arg)
1052 // newValue1I adds a new value with one argument and an auxint value to the current block.
1053 func (s *state) newValue1I(op ssa.Op, t *types.Type, aux int64, arg *ssa.Value) *ssa.Value {
1054 return s.curBlock.NewValue1I(s.peekPos(), op, t, aux, arg)
1057 // newValue2 adds a new value with two arguments to the current block.
1058 func (s *state) newValue2(op ssa.Op, t *types.Type, arg0, arg1 *ssa.Value) *ssa.Value {
1059 return s.curBlock.NewValue2(s.peekPos(), op, t, arg0, arg1)
1062 // newValue2A adds a new value with two arguments and an aux value to the current block.
1063 func (s *state) newValue2A(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1 *ssa.Value) *ssa.Value {
1064 return s.curBlock.NewValue2A(s.peekPos(), op, t, aux, arg0, arg1)
1067 // newValue2Apos adds a new value with two arguments and an aux value to the current block.
1068 // isStmt determines whether the created values may be a statement or not
1069 // (i.e., false means never, yes means maybe).
1070 func (s *state) newValue2Apos(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1 *ssa.Value, isStmt bool) *ssa.Value {
1072 return s.curBlock.NewValue2A(s.peekPos(), op, t, aux, arg0, arg1)
1074 return s.curBlock.NewValue2A(s.peekPos().WithNotStmt(), op, t, aux, arg0, arg1)
1077 // newValue2I adds a new value with two arguments and an auxint value to the current block.
1078 func (s *state) newValue2I(op ssa.Op, t *types.Type, aux int64, arg0, arg1 *ssa.Value) *ssa.Value {
1079 return s.curBlock.NewValue2I(s.peekPos(), op, t, aux, arg0, arg1)
1082 // newValue3 adds a new value with three arguments to the current block.
1083 func (s *state) newValue3(op ssa.Op, t *types.Type, arg0, arg1, arg2 *ssa.Value) *ssa.Value {
1084 return s.curBlock.NewValue3(s.peekPos(), op, t, arg0, arg1, arg2)
1087 // newValue3I adds a new value with three arguments and an auxint value to the current block.
1088 func (s *state) newValue3I(op ssa.Op, t *types.Type, aux int64, arg0, arg1, arg2 *ssa.Value) *ssa.Value {
1089 return s.curBlock.NewValue3I(s.peekPos(), op, t, aux, arg0, arg1, arg2)
1092 // newValue3A adds a new value with three arguments and an aux value to the current block.
1093 func (s *state) newValue3A(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1, arg2 *ssa.Value) *ssa.Value {
1094 return s.curBlock.NewValue3A(s.peekPos(), op, t, aux, arg0, arg1, arg2)
1097 // newValue3Apos adds a new value with three arguments and an aux value to the current block.
1098 // isStmt determines whether the created values may be a statement or not
1099 // (i.e., false means never, yes means maybe).
1100 func (s *state) newValue3Apos(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1, arg2 *ssa.Value, isStmt bool) *ssa.Value {
1102 return s.curBlock.NewValue3A(s.peekPos(), op, t, aux, arg0, arg1, arg2)
1104 return s.curBlock.NewValue3A(s.peekPos().WithNotStmt(), op, t, aux, arg0, arg1, arg2)
1107 // newValue4 adds a new value with four arguments to the current block.
1108 func (s *state) newValue4(op ssa.Op, t *types.Type, arg0, arg1, arg2, arg3 *ssa.Value) *ssa.Value {
1109 return s.curBlock.NewValue4(s.peekPos(), op, t, arg0, arg1, arg2, arg3)
1112 // newValue4I adds a new value with four arguments and an auxint value to the current block.
1113 func (s *state) newValue4I(op ssa.Op, t *types.Type, aux int64, arg0, arg1, arg2, arg3 *ssa.Value) *ssa.Value {
1114 return s.curBlock.NewValue4I(s.peekPos(), op, t, aux, arg0, arg1, arg2, arg3)
1117 func (s *state) entryBlock() *ssa.Block {
1119 if base.Flag.N > 0 && s.curBlock != nil {
1120 // If optimizations are off, allocate in current block instead. Since with -N
1121 // we're not doing the CSE or tighten passes, putting lots of stuff in the
1122 // entry block leads to O(n^2) entries in the live value map during regalloc.
1129 // entryNewValue0 adds a new value with no arguments to the entry block.
1130 func (s *state) entryNewValue0(op ssa.Op, t *types.Type) *ssa.Value {
1131 return s.entryBlock().NewValue0(src.NoXPos, op, t)
1134 // entryNewValue0A adds a new value with no arguments and an aux value to the entry block.
1135 func (s *state) entryNewValue0A(op ssa.Op, t *types.Type, aux ssa.Aux) *ssa.Value {
1136 return s.entryBlock().NewValue0A(src.NoXPos, op, t, aux)
1139 // entryNewValue1 adds a new value with one argument to the entry block.
1140 func (s *state) entryNewValue1(op ssa.Op, t *types.Type, arg *ssa.Value) *ssa.Value {
1141 return s.entryBlock().NewValue1(src.NoXPos, op, t, arg)
1144 // entryNewValue1I adds a new value with one argument and an auxint value to the entry block.
1145 func (s *state) entryNewValue1I(op ssa.Op, t *types.Type, auxint int64, arg *ssa.Value) *ssa.Value {
1146 return s.entryBlock().NewValue1I(src.NoXPos, op, t, auxint, arg)
1149 // entryNewValue1A adds a new value with one argument and an aux value to the entry block.
1150 func (s *state) entryNewValue1A(op ssa.Op, t *types.Type, aux ssa.Aux, arg *ssa.Value) *ssa.Value {
1151 return s.entryBlock().NewValue1A(src.NoXPos, op, t, aux, arg)
1154 // entryNewValue2 adds a new value with two arguments to the entry block.
1155 func (s *state) entryNewValue2(op ssa.Op, t *types.Type, arg0, arg1 *ssa.Value) *ssa.Value {
1156 return s.entryBlock().NewValue2(src.NoXPos, op, t, arg0, arg1)
1159 // entryNewValue2A adds a new value with two arguments and an aux value to the entry block.
1160 func (s *state) entryNewValue2A(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1 *ssa.Value) *ssa.Value {
1161 return s.entryBlock().NewValue2A(src.NoXPos, op, t, aux, arg0, arg1)
1164 // const* routines add a new const value to the entry block.
1165 func (s *state) constSlice(t *types.Type) *ssa.Value {
1166 return s.f.ConstSlice(t)
1168 func (s *state) constInterface(t *types.Type) *ssa.Value {
1169 return s.f.ConstInterface(t)
1171 func (s *state) constNil(t *types.Type) *ssa.Value { return s.f.ConstNil(t) }
1172 func (s *state) constEmptyString(t *types.Type) *ssa.Value {
1173 return s.f.ConstEmptyString(t)
1175 func (s *state) constBool(c bool) *ssa.Value {
1176 return s.f.ConstBool(types.Types[types.TBOOL], c)
1178 func (s *state) constInt8(t *types.Type, c int8) *ssa.Value {
1179 return s.f.ConstInt8(t, c)
1181 func (s *state) constInt16(t *types.Type, c int16) *ssa.Value {
1182 return s.f.ConstInt16(t, c)
1184 func (s *state) constInt32(t *types.Type, c int32) *ssa.Value {
1185 return s.f.ConstInt32(t, c)
1187 func (s *state) constInt64(t *types.Type, c int64) *ssa.Value {
1188 return s.f.ConstInt64(t, c)
1190 func (s *state) constFloat32(t *types.Type, c float64) *ssa.Value {
1191 return s.f.ConstFloat32(t, c)
1193 func (s *state) constFloat64(t *types.Type, c float64) *ssa.Value {
1194 return s.f.ConstFloat64(t, c)
1196 func (s *state) constInt(t *types.Type, c int64) *ssa.Value {
1197 if s.config.PtrSize == 8 {
1198 return s.constInt64(t, c)
1200 if int64(int32(c)) != c {
1201 s.Fatalf("integer constant too big %d", c)
1203 return s.constInt32(t, int32(c))
1205 func (s *state) constOffPtrSP(t *types.Type, c int64) *ssa.Value {
1206 return s.f.ConstOffPtrSP(t, c, s.sp)
1209 // newValueOrSfCall* are wrappers around newValue*, which may create a call to a
1210 // soft-float runtime function instead (when emitting soft-float code).
1211 func (s *state) newValueOrSfCall1(op ssa.Op, t *types.Type, arg *ssa.Value) *ssa.Value {
1213 if c, ok := s.sfcall(op, arg); ok {
1217 return s.newValue1(op, t, arg)
1219 func (s *state) newValueOrSfCall2(op ssa.Op, t *types.Type, arg0, arg1 *ssa.Value) *ssa.Value {
1221 if c, ok := s.sfcall(op, arg0, arg1); ok {
1225 return s.newValue2(op, t, arg0, arg1)
1228 type instrumentKind uint8
1231 instrumentRead = iota
1236 func (s *state) instrument(t *types.Type, addr *ssa.Value, kind instrumentKind) {
1237 s.instrument2(t, addr, nil, kind)
1240 // instrumentFields instruments a read/write operation on addr.
1241 // If it is instrumenting for MSAN or ASAN and t is a struct type, it instruments
1242 // operation for each field, instead of for the whole struct.
1243 func (s *state) instrumentFields(t *types.Type, addr *ssa.Value, kind instrumentKind) {
1244 if !(base.Flag.MSan || base.Flag.ASan) || !t.IsStruct() {
1245 s.instrument(t, addr, kind)
1248 for _, f := range t.Fields().Slice() {
1249 if f.Sym.IsBlank() {
1252 offptr := s.newValue1I(ssa.OpOffPtr, types.NewPtr(f.Type), f.Offset, addr)
1253 s.instrumentFields(f.Type, offptr, kind)
1257 func (s *state) instrumentMove(t *types.Type, dst, src *ssa.Value) {
1259 s.instrument2(t, dst, src, instrumentMove)
1261 s.instrument(t, src, instrumentRead)
1262 s.instrument(t, dst, instrumentWrite)
1266 func (s *state) instrument2(t *types.Type, addr, addr2 *ssa.Value, kind instrumentKind) {
1267 if !s.curfn.InstrumentBody() {
1273 return // can't race on zero-sized things
1276 if ssa.IsSanitizerSafeAddr(addr) {
1283 if addr2 != nil && kind != instrumentMove {
1284 panic("instrument2: non-nil addr2 for non-move instrumentation")
1289 case instrumentRead:
1290 fn = ir.Syms.Msanread
1291 case instrumentWrite:
1292 fn = ir.Syms.Msanwrite
1293 case instrumentMove:
1294 fn = ir.Syms.Msanmove
1296 panic("unreachable")
1299 } else if base.Flag.Race && t.NumComponents(types.CountBlankFields) > 1 {
1300 // for composite objects we have to write every address
1301 // because a write might happen to any subobject.
1302 // composites with only one element don't have subobjects, though.
1304 case instrumentRead:
1305 fn = ir.Syms.Racereadrange
1306 case instrumentWrite:
1307 fn = ir.Syms.Racewriterange
1309 panic("unreachable")
1312 } else if base.Flag.Race {
1313 // for non-composite objects we can write just the start
1314 // address, as any write must write the first byte.
1316 case instrumentRead:
1317 fn = ir.Syms.Raceread
1318 case instrumentWrite:
1319 fn = ir.Syms.Racewrite
1321 panic("unreachable")
1323 } else if base.Flag.ASan {
1325 case instrumentRead:
1326 fn = ir.Syms.Asanread
1327 case instrumentWrite:
1328 fn = ir.Syms.Asanwrite
1330 panic("unreachable")
1334 panic("unreachable")
1337 args := []*ssa.Value{addr}
1339 args = append(args, addr2)
1342 args = append(args, s.constInt(types.Types[types.TUINTPTR], w))
1344 s.rtcall(fn, true, nil, args...)
1347 func (s *state) load(t *types.Type, src *ssa.Value) *ssa.Value {
1348 s.instrumentFields(t, src, instrumentRead)
1349 return s.rawLoad(t, src)
1352 func (s *state) rawLoad(t *types.Type, src *ssa.Value) *ssa.Value {
1353 return s.newValue2(ssa.OpLoad, t, src, s.mem())
1356 func (s *state) store(t *types.Type, dst, val *ssa.Value) {
1357 s.vars[memVar] = s.newValue3A(ssa.OpStore, types.TypeMem, t, dst, val, s.mem())
1360 func (s *state) zero(t *types.Type, dst *ssa.Value) {
1361 s.instrument(t, dst, instrumentWrite)
1362 store := s.newValue2I(ssa.OpZero, types.TypeMem, t.Size(), dst, s.mem())
1364 s.vars[memVar] = store
1367 func (s *state) move(t *types.Type, dst, src *ssa.Value) {
1368 s.moveWhichMayOverlap(t, dst, src, false)
1370 func (s *state) moveWhichMayOverlap(t *types.Type, dst, src *ssa.Value, mayOverlap bool) {
1371 s.instrumentMove(t, dst, src)
1372 if mayOverlap && t.IsArray() && t.NumElem() > 1 && !ssa.IsInlinableMemmove(dst, src, t.Size(), s.f.Config) {
1373 // Normally, when moving Go values of type T from one location to another,
1374 // we don't need to worry about partial overlaps. The two Ts must either be
1375 // in disjoint (nonoverlapping) memory or in exactly the same location.
1376 // There are 2 cases where this isn't true:
1377 // 1) Using unsafe you can arrange partial overlaps.
1378 // 2) Since Go 1.17, you can use a cast from a slice to a ptr-to-array.
1379 // https://go.dev/ref/spec#Conversions_from_slice_to_array_pointer
1380 // This feature can be used to construct partial overlaps of array types.
1382 // p := (*[2]int)(a[:])
1383 // q := (*[2]int)(a[1:])
1385 // We don't care about solving 1. Or at least, we haven't historically
1386 // and no one has complained.
1387 // For 2, we need to ensure that if there might be partial overlap,
1388 // then we can't use OpMove; we must use memmove instead.
1389 // (memmove handles partial overlap by copying in the correct
1390 // direction. OpMove does not.)
1392 // Note that we have to be careful here not to introduce a call when
1393 // we're marshaling arguments to a call or unmarshaling results from a call.
1394 // Cases where this is happening must pass mayOverlap to false.
1395 // (Currently this only happens when unmarshaling results of a call.)
1396 if t.HasPointers() {
1397 s.rtcall(ir.Syms.Typedmemmove, true, nil, s.reflectType(t), dst, src)
1398 // We would have otherwise implemented this move with straightline code,
1399 // including a write barrier. Pretend we issue a write barrier here,
1400 // so that the write barrier tests work. (Otherwise they'd need to know
1401 // the details of IsInlineableMemmove.)
1402 s.curfn.SetWBPos(s.peekPos())
1404 s.rtcall(ir.Syms.Memmove, true, nil, dst, src, s.constInt(types.Types[types.TUINTPTR], t.Size()))
1406 ssa.LogLargeCopy(s.f.Name, s.peekPos(), t.Size())
1409 store := s.newValue3I(ssa.OpMove, types.TypeMem, t.Size(), dst, src, s.mem())
1411 s.vars[memVar] = store
1414 // stmtList converts the statement list n to SSA and adds it to s.
1415 func (s *state) stmtList(l ir.Nodes) {
1416 for _, n := range l {
1421 // stmt converts the statement n to SSA and adds it to s.
1422 func (s *state) stmt(n ir.Node) {
1426 // If s.curBlock is nil, and n isn't a label (which might have an associated goto somewhere),
1427 // then this code is dead. Stop here.
1428 if s.curBlock == nil && n.Op() != ir.OLABEL {
1432 s.stmtList(n.Init())
1436 n := n.(*ir.BlockStmt)
1440 case ir.ODCLCONST, ir.ODCLTYPE, ir.OFALL:
1442 // Expression statements
1444 n := n.(*ir.CallExpr)
1445 if ir.IsIntrinsicCall(n) {
1452 n := n.(*ir.CallExpr)
1453 s.callResult(n, callNormal)
1454 if n.Op() == ir.OCALLFUNC && n.X.Op() == ir.ONAME && n.X.(*ir.Name).Class == ir.PFUNC {
1455 if fn := n.X.Sym().Name; base.Flag.CompilingRuntime && fn == "throw" ||
1456 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") {
1459 b.Kind = ssa.BlockExit
1461 // TODO: never rewrite OPANIC to OCALLFUNC in the
1462 // first place. Need to wait until all backends
1467 n := n.(*ir.GoDeferStmt)
1468 if base.Debug.Defer > 0 {
1469 var defertype string
1470 if s.hasOpenDefers {
1471 defertype = "open-coded"
1472 } else if n.Esc() == ir.EscNever {
1473 defertype = "stack-allocated"
1475 defertype = "heap-allocated"
1477 base.WarnfAt(n.Pos(), "%s defer", defertype)
1479 if s.hasOpenDefers {
1480 s.openDeferRecord(n.Call.(*ir.CallExpr))
1483 if n.Esc() == ir.EscNever {
1486 s.callResult(n.Call.(*ir.CallExpr), d)
1489 n := n.(*ir.GoDeferStmt)
1490 s.callResult(n.Call.(*ir.CallExpr), callGo)
1492 case ir.OAS2DOTTYPE:
1493 n := n.(*ir.AssignListStmt)
1494 var res, resok *ssa.Value
1495 if n.Rhs[0].Op() == ir.ODOTTYPE2 {
1496 res, resok = s.dottype(n.Rhs[0].(*ir.TypeAssertExpr), true)
1498 res, resok = s.dynamicDottype(n.Rhs[0].(*ir.DynamicTypeAssertExpr), true)
1501 if !TypeOK(n.Rhs[0].Type()) {
1502 if res.Op != ssa.OpLoad {
1503 s.Fatalf("dottype of non-load")
1506 if res.Args[1] != mem {
1507 s.Fatalf("memory no longer live from 2-result dottype load")
1512 s.assign(n.Lhs[0], res, deref, 0)
1513 s.assign(n.Lhs[1], resok, false, 0)
1517 // We come here only when it is an intrinsic call returning two values.
1518 n := n.(*ir.AssignListStmt)
1519 call := n.Rhs[0].(*ir.CallExpr)
1520 if !ir.IsIntrinsicCall(call) {
1521 s.Fatalf("non-intrinsic AS2FUNC not expanded %v", call)
1523 v := s.intrinsicCall(call)
1524 v1 := s.newValue1(ssa.OpSelect0, n.Lhs[0].Type(), v)
1525 v2 := s.newValue1(ssa.OpSelect1, n.Lhs[1].Type(), v)
1526 s.assign(n.Lhs[0], v1, false, 0)
1527 s.assign(n.Lhs[1], v2, false, 0)
1532 if v := n.X; v.Esc() == ir.EscHeap {
1537 n := n.(*ir.LabelStmt)
1540 // Nothing to do because the label isn't targetable. See issue 52278.
1545 // The label might already have a target block via a goto.
1546 if lab.target == nil {
1547 lab.target = s.f.NewBlock(ssa.BlockPlain)
1550 // Go to that label.
1551 // (We pretend "label:" is preceded by "goto label", unless the predecessor is unreachable.)
1552 if s.curBlock != nil {
1554 b.AddEdgeTo(lab.target)
1556 s.startBlock(lab.target)
1559 n := n.(*ir.BranchStmt)
1563 if lab.target == nil {
1564 lab.target = s.f.NewBlock(ssa.BlockPlain)
1568 b.Pos = s.lastPos.WithIsStmt() // Do this even if b is an empty block.
1569 b.AddEdgeTo(lab.target)
1572 n := n.(*ir.AssignStmt)
1573 if n.X == n.Y && n.X.Op() == ir.ONAME {
1574 // An x=x assignment. No point in doing anything
1575 // here. In addition, skipping this assignment
1576 // prevents generating:
1579 // which is bad because x is incorrectly considered
1580 // dead before the vardef. See issue #14904.
1584 // mayOverlap keeps track of whether the LHS and RHS might
1585 // refer to partially overlapping memory. Partial overlapping can
1586 // only happen for arrays, see the comment in moveWhichMayOverlap.
1588 // If both sides of the assignment are not dereferences, then partial
1589 // overlap can't happen. Partial overlap can only occur only when the
1590 // arrays referenced are strictly smaller parts of the same base array.
1591 // If one side of the assignment is a full array, then partial overlap
1592 // can't happen. (The arrays are either disjoint or identical.)
1593 mayOverlap := n.X.Op() == ir.ODEREF && (n.Y != nil && n.Y.Op() == ir.ODEREF)
1594 if n.Y != nil && n.Y.Op() == ir.ODEREF {
1595 p := n.Y.(*ir.StarExpr).X
1596 for p.Op() == ir.OCONVNOP {
1597 p = p.(*ir.ConvExpr).X
1599 if p.Op() == ir.OSPTR && p.(*ir.UnaryExpr).X.Type().IsString() {
1600 // Pointer fields of strings point to unmodifiable memory.
1601 // That memory can't overlap with the memory being written.
1610 case ir.OSTRUCTLIT, ir.OARRAYLIT, ir.OSLICELIT:
1611 // All literals with nonzero fields have already been
1612 // rewritten during walk. Any that remain are just T{}
1613 // or equivalents. Use the zero value.
1614 if !ir.IsZero(rhs) {
1615 s.Fatalf("literal with nonzero value in SSA: %v", rhs)
1619 rhs := rhs.(*ir.CallExpr)
1620 // Check whether we're writing the result of an append back to the same slice.
1621 // If so, we handle it specially to avoid write barriers on the fast
1622 // (non-growth) path.
1623 if !ir.SameSafeExpr(n.X, rhs.Args[0]) || base.Flag.N != 0 {
1626 // If the slice can be SSA'd, it'll be on the stack,
1627 // so there will be no write barriers,
1628 // so there's no need to attempt to prevent them.
1630 if base.Debug.Append > 0 { // replicating old diagnostic message
1631 base.WarnfAt(n.Pos(), "append: len-only update (in local slice)")
1635 if base.Debug.Append > 0 {
1636 base.WarnfAt(n.Pos(), "append: len-only update")
1643 if ir.IsBlank(n.X) {
1645 // Just evaluate rhs for side-effects.
1663 r = nil // Signal assign to use OpZero.
1676 if rhs != nil && (rhs.Op() == ir.OSLICE || rhs.Op() == ir.OSLICE3 || rhs.Op() == ir.OSLICESTR) && ir.SameSafeExpr(rhs.(*ir.SliceExpr).X, n.X) {
1677 // We're assigning a slicing operation back to its source.
1678 // Don't write back fields we aren't changing. See issue #14855.
1679 rhs := rhs.(*ir.SliceExpr)
1680 i, j, k := rhs.Low, rhs.High, rhs.Max
1681 if i != nil && (i.Op() == ir.OLITERAL && i.Val().Kind() == constant.Int && ir.Int64Val(i) == 0) {
1682 // [0:...] is the same as [:...]
1685 // TODO: detect defaults for len/cap also.
1686 // Currently doesn't really work because (*p)[:len(*p)] appears here as:
1689 // if j != nil && (j.Op == OLEN && SameSafeExpr(j.Left, n.Left)) {
1692 // if k != nil && (k.Op == OCAP && SameSafeExpr(k.Left, n.Left)) {
1706 s.assignWhichMayOverlap(n.X, r, deref, skip, mayOverlap)
1710 if ir.IsConst(n.Cond, constant.Bool) {
1711 s.stmtList(n.Cond.Init())
1712 if ir.BoolVal(n.Cond) {
1720 bEnd := s.f.NewBlock(ssa.BlockPlain)
1725 var bThen *ssa.Block
1726 if len(n.Body) != 0 {
1727 bThen = s.f.NewBlock(ssa.BlockPlain)
1731 var bElse *ssa.Block
1732 if len(n.Else) != 0 {
1733 bElse = s.f.NewBlock(ssa.BlockPlain)
1737 s.condBranch(n.Cond, bThen, bElse, likely)
1739 if len(n.Body) != 0 {
1742 if b := s.endBlock(); b != nil {
1746 if len(n.Else) != 0 {
1749 if b := s.endBlock(); b != nil {
1756 n := n.(*ir.ReturnStmt)
1757 s.stmtList(n.Results)
1759 b.Pos = s.lastPos.WithIsStmt()
1762 n := n.(*ir.TailCallStmt)
1763 s.callResult(n.Call, callTail)
1766 b.Kind = ssa.BlockRetJmp // could use BlockExit. BlockRetJmp is mostly for clarity.
1769 case ir.OCONTINUE, ir.OBREAK:
1770 n := n.(*ir.BranchStmt)
1773 // plain break/continue
1781 // labeled break/continue; look up the target
1786 to = lab.continueTarget
1788 to = lab.breakTarget
1793 b.Pos = s.lastPos.WithIsStmt() // Do this even if b is an empty block.
1797 // OFOR: for Ninit; Left; Right { Nbody }
1798 // cond (Left); body (Nbody); incr (Right)
1799 n := n.(*ir.ForStmt)
1800 base.Assert(!n.DistinctVars) // Should all be rewritten before escape analysis
1801 bCond := s.f.NewBlock(ssa.BlockPlain)
1802 bBody := s.f.NewBlock(ssa.BlockPlain)
1803 bIncr := s.f.NewBlock(ssa.BlockPlain)
1804 bEnd := s.f.NewBlock(ssa.BlockPlain)
1806 // ensure empty for loops have correct position; issue #30167
1809 // first, jump to condition test
1813 // generate code to test condition
1816 s.condBranch(n.Cond, bBody, bEnd, 1)
1819 b.Kind = ssa.BlockPlain
1823 // set up for continue/break in body
1824 prevContinue := s.continueTo
1825 prevBreak := s.breakTo
1826 s.continueTo = bIncr
1829 if sym := n.Label; sym != nil {
1832 lab.continueTarget = bIncr
1833 lab.breakTarget = bEnd
1840 // tear down continue/break
1841 s.continueTo = prevContinue
1842 s.breakTo = prevBreak
1844 lab.continueTarget = nil
1845 lab.breakTarget = nil
1848 // done with body, goto incr
1849 if b := s.endBlock(); b != nil {
1858 if b := s.endBlock(); b != nil {
1860 // It can happen that bIncr ends in a block containing only VARKILL,
1861 // and that muddles the debugging experience.
1862 if b.Pos == src.NoXPos {
1869 case ir.OSWITCH, ir.OSELECT:
1870 // These have been mostly rewritten by the front end into their Nbody fields.
1871 // Our main task is to correctly hook up any break statements.
1872 bEnd := s.f.NewBlock(ssa.BlockPlain)
1874 prevBreak := s.breakTo
1878 if n.Op() == ir.OSWITCH {
1879 n := n.(*ir.SwitchStmt)
1883 n := n.(*ir.SelectStmt)
1892 lab.breakTarget = bEnd
1895 // generate body code
1898 s.breakTo = prevBreak
1900 lab.breakTarget = nil
1903 // walk adds explicit OBREAK nodes to the end of all reachable code paths.
1904 // If we still have a current block here, then mark it unreachable.
1905 if s.curBlock != nil {
1908 b.Kind = ssa.BlockExit
1914 n := n.(*ir.JumpTableStmt)
1916 // Make blocks we'll need.
1917 jt := s.f.NewBlock(ssa.BlockJumpTable)
1918 bEnd := s.f.NewBlock(ssa.BlockPlain)
1920 // The only thing that needs evaluating is the index we're looking up.
1921 idx := s.expr(n.Idx)
1922 unsigned := idx.Type.IsUnsigned()
1924 // Extend so we can do everything in uintptr arithmetic.
1925 t := types.Types[types.TUINTPTR]
1926 idx = s.conv(nil, idx, idx.Type, t)
1928 // The ending condition for the current block decides whether we'll use
1929 // the jump table at all.
1930 // We check that min <= idx <= max and jump around the jump table
1931 // if that test fails.
1932 // We implement min <= idx <= max with 0 <= idx-min <= max-min, because
1933 // we'll need idx-min anyway as the control value for the jump table.
1936 min, _ = constant.Uint64Val(n.Cases[0])
1937 max, _ = constant.Uint64Val(n.Cases[len(n.Cases)-1])
1939 mn, _ := constant.Int64Val(n.Cases[0])
1940 mx, _ := constant.Int64Val(n.Cases[len(n.Cases)-1])
1944 // Compare idx-min with max-min, to see if we can use the jump table.
1945 idx = s.newValue2(s.ssaOp(ir.OSUB, t), t, idx, s.uintptrConstant(min))
1946 width := s.uintptrConstant(max - min)
1947 cmp := s.newValue2(s.ssaOp(ir.OLE, t), types.Types[types.TBOOL], idx, width)
1949 b.Kind = ssa.BlockIf
1951 b.AddEdgeTo(jt) // in range - use jump table
1952 b.AddEdgeTo(bEnd) // out of range - no case in the jump table will trigger
1953 b.Likely = ssa.BranchLikely // TODO: assumes missing the table entirely is unlikely. True?
1955 // Build jump table block.
1958 if base.Flag.Cfg.SpectreIndex {
1959 idx = s.newValue2(ssa.OpSpectreSliceIndex, t, idx, width)
1963 // Figure out where we should go for each index in the table.
1964 table := make([]*ssa.Block, max-min+1)
1965 for i := range table {
1966 table[i] = bEnd // default target
1968 for i := range n.Targets {
1970 lab := s.label(n.Targets[i])
1971 if lab.target == nil {
1972 lab.target = s.f.NewBlock(ssa.BlockPlain)
1976 val, _ = constant.Uint64Val(c)
1978 vl, _ := constant.Int64Val(c)
1981 // Overwrite the default target.
1982 table[val-min] = lab.target
1984 for _, t := range table {
1992 n := n.(*ir.UnaryExpr)
1997 n := n.(*ir.InlineMarkStmt)
1998 s.newValue1I(ssa.OpInlMark, types.TypeVoid, n.Index, s.mem())
2001 s.Fatalf("unhandled stmt %v", n.Op())
2005 // If true, share as many open-coded defer exits as possible (with the downside of
2006 // worse line-number information)
2007 const shareDeferExits = false
2009 // exit processes any code that needs to be generated just before returning.
2010 // It returns a BlockRet block that ends the control flow. Its control value
2011 // will be set to the final memory state.
2012 func (s *state) exit() *ssa.Block {
2014 if s.hasOpenDefers {
2015 if shareDeferExits && s.lastDeferExit != nil && len(s.openDefers) == s.lastDeferCount {
2016 if s.curBlock.Kind != ssa.BlockPlain {
2017 panic("Block for an exit should be BlockPlain")
2019 s.curBlock.AddEdgeTo(s.lastDeferExit)
2021 return s.lastDeferFinalBlock
2025 s.rtcall(ir.Syms.Deferreturn, true, nil)
2031 // Do actual return.
2032 // These currently turn into self-copies (in many cases).
2033 resultFields := s.curfn.Type().Results().FieldSlice()
2034 results := make([]*ssa.Value, len(resultFields)+1, len(resultFields)+1)
2035 m = s.newValue0(ssa.OpMakeResult, s.f.OwnAux.LateExpansionResultType())
2036 // Store SSAable and heap-escaped PPARAMOUT variables back to stack locations.
2037 for i, f := range resultFields {
2038 n := f.Nname.(*ir.Name)
2039 if s.canSSA(n) { // result is in some SSA variable
2040 if !n.IsOutputParamInRegisters() && n.Type().HasPointers() {
2041 // We are about to store to the result slot.
2042 s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, n, s.mem())
2044 results[i] = s.variable(n, n.Type())
2045 } else if !n.OnStack() { // result is actually heap allocated
2046 // We are about to copy the in-heap result to the result slot.
2047 if n.Type().HasPointers() {
2048 s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, n, s.mem())
2050 ha := s.expr(n.Heapaddr)
2051 s.instrumentFields(n.Type(), ha, instrumentRead)
2052 results[i] = s.newValue2(ssa.OpDereference, n.Type(), ha, s.mem())
2053 } else { // result is not SSA-able; not escaped, so not on heap, but too large for SSA.
2054 // Before register ABI this ought to be a self-move, home=dest,
2055 // With register ABI, it's still a self-move if parameter is on stack (i.e., too big or overflowed)
2056 // No VarDef, as the result slot is already holding live value.
2057 results[i] = s.newValue2(ssa.OpDereference, n.Type(), s.addr(n), s.mem())
2061 // Run exit code. Today, this is just racefuncexit, in -race mode.
2062 // TODO(register args) this seems risky here with a register-ABI, but not clear it is right to do it earlier either.
2063 // Spills in register allocation might just fix it.
2064 s.stmtList(s.curfn.Exit)
2066 results[len(results)-1] = s.mem()
2067 m.AddArgs(results...)
2070 b.Kind = ssa.BlockRet
2072 if s.hasdefer && s.hasOpenDefers {
2073 s.lastDeferFinalBlock = b
2078 type opAndType struct {
2083 var opToSSA = map[opAndType]ssa.Op{
2084 {ir.OADD, types.TINT8}: ssa.OpAdd8,
2085 {ir.OADD, types.TUINT8}: ssa.OpAdd8,
2086 {ir.OADD, types.TINT16}: ssa.OpAdd16,
2087 {ir.OADD, types.TUINT16}: ssa.OpAdd16,
2088 {ir.OADD, types.TINT32}: ssa.OpAdd32,
2089 {ir.OADD, types.TUINT32}: ssa.OpAdd32,
2090 {ir.OADD, types.TINT64}: ssa.OpAdd64,
2091 {ir.OADD, types.TUINT64}: ssa.OpAdd64,
2092 {ir.OADD, types.TFLOAT32}: ssa.OpAdd32F,
2093 {ir.OADD, types.TFLOAT64}: ssa.OpAdd64F,
2095 {ir.OSUB, types.TINT8}: ssa.OpSub8,
2096 {ir.OSUB, types.TUINT8}: ssa.OpSub8,
2097 {ir.OSUB, types.TINT16}: ssa.OpSub16,
2098 {ir.OSUB, types.TUINT16}: ssa.OpSub16,
2099 {ir.OSUB, types.TINT32}: ssa.OpSub32,
2100 {ir.OSUB, types.TUINT32}: ssa.OpSub32,
2101 {ir.OSUB, types.TINT64}: ssa.OpSub64,
2102 {ir.OSUB, types.TUINT64}: ssa.OpSub64,
2103 {ir.OSUB, types.TFLOAT32}: ssa.OpSub32F,
2104 {ir.OSUB, types.TFLOAT64}: ssa.OpSub64F,
2106 {ir.ONOT, types.TBOOL}: ssa.OpNot,
2108 {ir.ONEG, types.TINT8}: ssa.OpNeg8,
2109 {ir.ONEG, types.TUINT8}: ssa.OpNeg8,
2110 {ir.ONEG, types.TINT16}: ssa.OpNeg16,
2111 {ir.ONEG, types.TUINT16}: ssa.OpNeg16,
2112 {ir.ONEG, types.TINT32}: ssa.OpNeg32,
2113 {ir.ONEG, types.TUINT32}: ssa.OpNeg32,
2114 {ir.ONEG, types.TINT64}: ssa.OpNeg64,
2115 {ir.ONEG, types.TUINT64}: ssa.OpNeg64,
2116 {ir.ONEG, types.TFLOAT32}: ssa.OpNeg32F,
2117 {ir.ONEG, types.TFLOAT64}: ssa.OpNeg64F,
2119 {ir.OBITNOT, types.TINT8}: ssa.OpCom8,
2120 {ir.OBITNOT, types.TUINT8}: ssa.OpCom8,
2121 {ir.OBITNOT, types.TINT16}: ssa.OpCom16,
2122 {ir.OBITNOT, types.TUINT16}: ssa.OpCom16,
2123 {ir.OBITNOT, types.TINT32}: ssa.OpCom32,
2124 {ir.OBITNOT, types.TUINT32}: ssa.OpCom32,
2125 {ir.OBITNOT, types.TINT64}: ssa.OpCom64,
2126 {ir.OBITNOT, types.TUINT64}: ssa.OpCom64,
2128 {ir.OIMAG, types.TCOMPLEX64}: ssa.OpComplexImag,
2129 {ir.OIMAG, types.TCOMPLEX128}: ssa.OpComplexImag,
2130 {ir.OREAL, types.TCOMPLEX64}: ssa.OpComplexReal,
2131 {ir.OREAL, types.TCOMPLEX128}: ssa.OpComplexReal,
2133 {ir.OMUL, types.TINT8}: ssa.OpMul8,
2134 {ir.OMUL, types.TUINT8}: ssa.OpMul8,
2135 {ir.OMUL, types.TINT16}: ssa.OpMul16,
2136 {ir.OMUL, types.TUINT16}: ssa.OpMul16,
2137 {ir.OMUL, types.TINT32}: ssa.OpMul32,
2138 {ir.OMUL, types.TUINT32}: ssa.OpMul32,
2139 {ir.OMUL, types.TINT64}: ssa.OpMul64,
2140 {ir.OMUL, types.TUINT64}: ssa.OpMul64,
2141 {ir.OMUL, types.TFLOAT32}: ssa.OpMul32F,
2142 {ir.OMUL, types.TFLOAT64}: ssa.OpMul64F,
2144 {ir.ODIV, types.TFLOAT32}: ssa.OpDiv32F,
2145 {ir.ODIV, types.TFLOAT64}: ssa.OpDiv64F,
2147 {ir.ODIV, types.TINT8}: ssa.OpDiv8,
2148 {ir.ODIV, types.TUINT8}: ssa.OpDiv8u,
2149 {ir.ODIV, types.TINT16}: ssa.OpDiv16,
2150 {ir.ODIV, types.TUINT16}: ssa.OpDiv16u,
2151 {ir.ODIV, types.TINT32}: ssa.OpDiv32,
2152 {ir.ODIV, types.TUINT32}: ssa.OpDiv32u,
2153 {ir.ODIV, types.TINT64}: ssa.OpDiv64,
2154 {ir.ODIV, types.TUINT64}: ssa.OpDiv64u,
2156 {ir.OMOD, types.TINT8}: ssa.OpMod8,
2157 {ir.OMOD, types.TUINT8}: ssa.OpMod8u,
2158 {ir.OMOD, types.TINT16}: ssa.OpMod16,
2159 {ir.OMOD, types.TUINT16}: ssa.OpMod16u,
2160 {ir.OMOD, types.TINT32}: ssa.OpMod32,
2161 {ir.OMOD, types.TUINT32}: ssa.OpMod32u,
2162 {ir.OMOD, types.TINT64}: ssa.OpMod64,
2163 {ir.OMOD, types.TUINT64}: ssa.OpMod64u,
2165 {ir.OAND, types.TINT8}: ssa.OpAnd8,
2166 {ir.OAND, types.TUINT8}: ssa.OpAnd8,
2167 {ir.OAND, types.TINT16}: ssa.OpAnd16,
2168 {ir.OAND, types.TUINT16}: ssa.OpAnd16,
2169 {ir.OAND, types.TINT32}: ssa.OpAnd32,
2170 {ir.OAND, types.TUINT32}: ssa.OpAnd32,
2171 {ir.OAND, types.TINT64}: ssa.OpAnd64,
2172 {ir.OAND, types.TUINT64}: ssa.OpAnd64,
2174 {ir.OOR, types.TINT8}: ssa.OpOr8,
2175 {ir.OOR, types.TUINT8}: ssa.OpOr8,
2176 {ir.OOR, types.TINT16}: ssa.OpOr16,
2177 {ir.OOR, types.TUINT16}: ssa.OpOr16,
2178 {ir.OOR, types.TINT32}: ssa.OpOr32,
2179 {ir.OOR, types.TUINT32}: ssa.OpOr32,
2180 {ir.OOR, types.TINT64}: ssa.OpOr64,
2181 {ir.OOR, types.TUINT64}: ssa.OpOr64,
2183 {ir.OXOR, types.TINT8}: ssa.OpXor8,
2184 {ir.OXOR, types.TUINT8}: ssa.OpXor8,
2185 {ir.OXOR, types.TINT16}: ssa.OpXor16,
2186 {ir.OXOR, types.TUINT16}: ssa.OpXor16,
2187 {ir.OXOR, types.TINT32}: ssa.OpXor32,
2188 {ir.OXOR, types.TUINT32}: ssa.OpXor32,
2189 {ir.OXOR, types.TINT64}: ssa.OpXor64,
2190 {ir.OXOR, types.TUINT64}: ssa.OpXor64,
2192 {ir.OEQ, types.TBOOL}: ssa.OpEqB,
2193 {ir.OEQ, types.TINT8}: ssa.OpEq8,
2194 {ir.OEQ, types.TUINT8}: ssa.OpEq8,
2195 {ir.OEQ, types.TINT16}: ssa.OpEq16,
2196 {ir.OEQ, types.TUINT16}: ssa.OpEq16,
2197 {ir.OEQ, types.TINT32}: ssa.OpEq32,
2198 {ir.OEQ, types.TUINT32}: ssa.OpEq32,
2199 {ir.OEQ, types.TINT64}: ssa.OpEq64,
2200 {ir.OEQ, types.TUINT64}: ssa.OpEq64,
2201 {ir.OEQ, types.TINTER}: ssa.OpEqInter,
2202 {ir.OEQ, types.TSLICE}: ssa.OpEqSlice,
2203 {ir.OEQ, types.TFUNC}: ssa.OpEqPtr,
2204 {ir.OEQ, types.TMAP}: ssa.OpEqPtr,
2205 {ir.OEQ, types.TCHAN}: ssa.OpEqPtr,
2206 {ir.OEQ, types.TPTR}: ssa.OpEqPtr,
2207 {ir.OEQ, types.TUINTPTR}: ssa.OpEqPtr,
2208 {ir.OEQ, types.TUNSAFEPTR}: ssa.OpEqPtr,
2209 {ir.OEQ, types.TFLOAT64}: ssa.OpEq64F,
2210 {ir.OEQ, types.TFLOAT32}: ssa.OpEq32F,
2212 {ir.ONE, types.TBOOL}: ssa.OpNeqB,
2213 {ir.ONE, types.TINT8}: ssa.OpNeq8,
2214 {ir.ONE, types.TUINT8}: ssa.OpNeq8,
2215 {ir.ONE, types.TINT16}: ssa.OpNeq16,
2216 {ir.ONE, types.TUINT16}: ssa.OpNeq16,
2217 {ir.ONE, types.TINT32}: ssa.OpNeq32,
2218 {ir.ONE, types.TUINT32}: ssa.OpNeq32,
2219 {ir.ONE, types.TINT64}: ssa.OpNeq64,
2220 {ir.ONE, types.TUINT64}: ssa.OpNeq64,
2221 {ir.ONE, types.TINTER}: ssa.OpNeqInter,
2222 {ir.ONE, types.TSLICE}: ssa.OpNeqSlice,
2223 {ir.ONE, types.TFUNC}: ssa.OpNeqPtr,
2224 {ir.ONE, types.TMAP}: ssa.OpNeqPtr,
2225 {ir.ONE, types.TCHAN}: ssa.OpNeqPtr,
2226 {ir.ONE, types.TPTR}: ssa.OpNeqPtr,
2227 {ir.ONE, types.TUINTPTR}: ssa.OpNeqPtr,
2228 {ir.ONE, types.TUNSAFEPTR}: ssa.OpNeqPtr,
2229 {ir.ONE, types.TFLOAT64}: ssa.OpNeq64F,
2230 {ir.ONE, types.TFLOAT32}: ssa.OpNeq32F,
2232 {ir.OLT, types.TINT8}: ssa.OpLess8,
2233 {ir.OLT, types.TUINT8}: ssa.OpLess8U,
2234 {ir.OLT, types.TINT16}: ssa.OpLess16,
2235 {ir.OLT, types.TUINT16}: ssa.OpLess16U,
2236 {ir.OLT, types.TINT32}: ssa.OpLess32,
2237 {ir.OLT, types.TUINT32}: ssa.OpLess32U,
2238 {ir.OLT, types.TINT64}: ssa.OpLess64,
2239 {ir.OLT, types.TUINT64}: ssa.OpLess64U,
2240 {ir.OLT, types.TFLOAT64}: ssa.OpLess64F,
2241 {ir.OLT, types.TFLOAT32}: ssa.OpLess32F,
2243 {ir.OLE, types.TINT8}: ssa.OpLeq8,
2244 {ir.OLE, types.TUINT8}: ssa.OpLeq8U,
2245 {ir.OLE, types.TINT16}: ssa.OpLeq16,
2246 {ir.OLE, types.TUINT16}: ssa.OpLeq16U,
2247 {ir.OLE, types.TINT32}: ssa.OpLeq32,
2248 {ir.OLE, types.TUINT32}: ssa.OpLeq32U,
2249 {ir.OLE, types.TINT64}: ssa.OpLeq64,
2250 {ir.OLE, types.TUINT64}: ssa.OpLeq64U,
2251 {ir.OLE, types.TFLOAT64}: ssa.OpLeq64F,
2252 {ir.OLE, types.TFLOAT32}: ssa.OpLeq32F,
2255 func (s *state) concreteEtype(t *types.Type) types.Kind {
2261 if s.config.PtrSize == 8 {
2266 if s.config.PtrSize == 8 {
2267 return types.TUINT64
2269 return types.TUINT32
2270 case types.TUINTPTR:
2271 if s.config.PtrSize == 8 {
2272 return types.TUINT64
2274 return types.TUINT32
2278 func (s *state) ssaOp(op ir.Op, t *types.Type) ssa.Op {
2279 etype := s.concreteEtype(t)
2280 x, ok := opToSSA[opAndType{op, etype}]
2282 s.Fatalf("unhandled binary op %v %s", op, etype)
2287 type opAndTwoTypes struct {
2293 type twoTypes struct {
2298 type twoOpsAndType struct {
2301 intermediateType types.Kind
2304 var fpConvOpToSSA = map[twoTypes]twoOpsAndType{
2306 {types.TINT8, types.TFLOAT32}: {ssa.OpSignExt8to32, ssa.OpCvt32to32F, types.TINT32},
2307 {types.TINT16, types.TFLOAT32}: {ssa.OpSignExt16to32, ssa.OpCvt32to32F, types.TINT32},
2308 {types.TINT32, types.TFLOAT32}: {ssa.OpCopy, ssa.OpCvt32to32F, types.TINT32},
2309 {types.TINT64, types.TFLOAT32}: {ssa.OpCopy, ssa.OpCvt64to32F, types.TINT64},
2311 {types.TINT8, types.TFLOAT64}: {ssa.OpSignExt8to32, ssa.OpCvt32to64F, types.TINT32},
2312 {types.TINT16, types.TFLOAT64}: {ssa.OpSignExt16to32, ssa.OpCvt32to64F, types.TINT32},
2313 {types.TINT32, types.TFLOAT64}: {ssa.OpCopy, ssa.OpCvt32to64F, types.TINT32},
2314 {types.TINT64, types.TFLOAT64}: {ssa.OpCopy, ssa.OpCvt64to64F, types.TINT64},
2316 {types.TFLOAT32, types.TINT8}: {ssa.OpCvt32Fto32, ssa.OpTrunc32to8, types.TINT32},
2317 {types.TFLOAT32, types.TINT16}: {ssa.OpCvt32Fto32, ssa.OpTrunc32to16, types.TINT32},
2318 {types.TFLOAT32, types.TINT32}: {ssa.OpCvt32Fto32, ssa.OpCopy, types.TINT32},
2319 {types.TFLOAT32, types.TINT64}: {ssa.OpCvt32Fto64, ssa.OpCopy, types.TINT64},
2321 {types.TFLOAT64, types.TINT8}: {ssa.OpCvt64Fto32, ssa.OpTrunc32to8, types.TINT32},
2322 {types.TFLOAT64, types.TINT16}: {ssa.OpCvt64Fto32, ssa.OpTrunc32to16, types.TINT32},
2323 {types.TFLOAT64, types.TINT32}: {ssa.OpCvt64Fto32, ssa.OpCopy, types.TINT32},
2324 {types.TFLOAT64, types.TINT64}: {ssa.OpCvt64Fto64, ssa.OpCopy, types.TINT64},
2326 {types.TUINT8, types.TFLOAT32}: {ssa.OpZeroExt8to32, ssa.OpCvt32to32F, types.TINT32},
2327 {types.TUINT16, types.TFLOAT32}: {ssa.OpZeroExt16to32, ssa.OpCvt32to32F, types.TINT32},
2328 {types.TUINT32, types.TFLOAT32}: {ssa.OpZeroExt32to64, ssa.OpCvt64to32F, types.TINT64}, // go wide to dodge unsigned
2329 {types.TUINT64, types.TFLOAT32}: {ssa.OpCopy, ssa.OpInvalid, types.TUINT64}, // Cvt64Uto32F, branchy code expansion instead
2331 {types.TUINT8, types.TFLOAT64}: {ssa.OpZeroExt8to32, ssa.OpCvt32to64F, types.TINT32},
2332 {types.TUINT16, types.TFLOAT64}: {ssa.OpZeroExt16to32, ssa.OpCvt32to64F, types.TINT32},
2333 {types.TUINT32, types.TFLOAT64}: {ssa.OpZeroExt32to64, ssa.OpCvt64to64F, types.TINT64}, // go wide to dodge unsigned
2334 {types.TUINT64, types.TFLOAT64}: {ssa.OpCopy, ssa.OpInvalid, types.TUINT64}, // Cvt64Uto64F, branchy code expansion instead
2336 {types.TFLOAT32, types.TUINT8}: {ssa.OpCvt32Fto32, ssa.OpTrunc32to8, types.TINT32},
2337 {types.TFLOAT32, types.TUINT16}: {ssa.OpCvt32Fto32, ssa.OpTrunc32to16, types.TINT32},
2338 {types.TFLOAT32, types.TUINT32}: {ssa.OpCvt32Fto64, ssa.OpTrunc64to32, types.TINT64}, // go wide to dodge unsigned
2339 {types.TFLOAT32, types.TUINT64}: {ssa.OpInvalid, ssa.OpCopy, types.TUINT64}, // Cvt32Fto64U, branchy code expansion instead
2341 {types.TFLOAT64, types.TUINT8}: {ssa.OpCvt64Fto32, ssa.OpTrunc32to8, types.TINT32},
2342 {types.TFLOAT64, types.TUINT16}: {ssa.OpCvt64Fto32, ssa.OpTrunc32to16, types.TINT32},
2343 {types.TFLOAT64, types.TUINT32}: {ssa.OpCvt64Fto64, ssa.OpTrunc64to32, types.TINT64}, // go wide to dodge unsigned
2344 {types.TFLOAT64, types.TUINT64}: {ssa.OpInvalid, ssa.OpCopy, types.TUINT64}, // Cvt64Fto64U, branchy code expansion instead
2347 {types.TFLOAT64, types.TFLOAT32}: {ssa.OpCvt64Fto32F, ssa.OpCopy, types.TFLOAT32},
2348 {types.TFLOAT64, types.TFLOAT64}: {ssa.OpRound64F, ssa.OpCopy, types.TFLOAT64},
2349 {types.TFLOAT32, types.TFLOAT32}: {ssa.OpRound32F, ssa.OpCopy, types.TFLOAT32},
2350 {types.TFLOAT32, types.TFLOAT64}: {ssa.OpCvt32Fto64F, ssa.OpCopy, types.TFLOAT64},
2353 // this map is used only for 32-bit arch, and only includes the difference
2354 // on 32-bit arch, don't use int64<->float conversion for uint32
2355 var fpConvOpToSSA32 = map[twoTypes]twoOpsAndType{
2356 {types.TUINT32, types.TFLOAT32}: {ssa.OpCopy, ssa.OpCvt32Uto32F, types.TUINT32},
2357 {types.TUINT32, types.TFLOAT64}: {ssa.OpCopy, ssa.OpCvt32Uto64F, types.TUINT32},
2358 {types.TFLOAT32, types.TUINT32}: {ssa.OpCvt32Fto32U, ssa.OpCopy, types.TUINT32},
2359 {types.TFLOAT64, types.TUINT32}: {ssa.OpCvt64Fto32U, ssa.OpCopy, types.TUINT32},
2362 // uint64<->float conversions, only on machines that have instructions for that
2363 var uint64fpConvOpToSSA = map[twoTypes]twoOpsAndType{
2364 {types.TUINT64, types.TFLOAT32}: {ssa.OpCopy, ssa.OpCvt64Uto32F, types.TUINT64},
2365 {types.TUINT64, types.TFLOAT64}: {ssa.OpCopy, ssa.OpCvt64Uto64F, types.TUINT64},
2366 {types.TFLOAT32, types.TUINT64}: {ssa.OpCvt32Fto64U, ssa.OpCopy, types.TUINT64},
2367 {types.TFLOAT64, types.TUINT64}: {ssa.OpCvt64Fto64U, ssa.OpCopy, types.TUINT64},
2370 var shiftOpToSSA = map[opAndTwoTypes]ssa.Op{
2371 {ir.OLSH, types.TINT8, types.TUINT8}: ssa.OpLsh8x8,
2372 {ir.OLSH, types.TUINT8, types.TUINT8}: ssa.OpLsh8x8,
2373 {ir.OLSH, types.TINT8, types.TUINT16}: ssa.OpLsh8x16,
2374 {ir.OLSH, types.TUINT8, types.TUINT16}: ssa.OpLsh8x16,
2375 {ir.OLSH, types.TINT8, types.TUINT32}: ssa.OpLsh8x32,
2376 {ir.OLSH, types.TUINT8, types.TUINT32}: ssa.OpLsh8x32,
2377 {ir.OLSH, types.TINT8, types.TUINT64}: ssa.OpLsh8x64,
2378 {ir.OLSH, types.TUINT8, types.TUINT64}: ssa.OpLsh8x64,
2380 {ir.OLSH, types.TINT16, types.TUINT8}: ssa.OpLsh16x8,
2381 {ir.OLSH, types.TUINT16, types.TUINT8}: ssa.OpLsh16x8,
2382 {ir.OLSH, types.TINT16, types.TUINT16}: ssa.OpLsh16x16,
2383 {ir.OLSH, types.TUINT16, types.TUINT16}: ssa.OpLsh16x16,
2384 {ir.OLSH, types.TINT16, types.TUINT32}: ssa.OpLsh16x32,
2385 {ir.OLSH, types.TUINT16, types.TUINT32}: ssa.OpLsh16x32,
2386 {ir.OLSH, types.TINT16, types.TUINT64}: ssa.OpLsh16x64,
2387 {ir.OLSH, types.TUINT16, types.TUINT64}: ssa.OpLsh16x64,
2389 {ir.OLSH, types.TINT32, types.TUINT8}: ssa.OpLsh32x8,
2390 {ir.OLSH, types.TUINT32, types.TUINT8}: ssa.OpLsh32x8,
2391 {ir.OLSH, types.TINT32, types.TUINT16}: ssa.OpLsh32x16,
2392 {ir.OLSH, types.TUINT32, types.TUINT16}: ssa.OpLsh32x16,
2393 {ir.OLSH, types.TINT32, types.TUINT32}: ssa.OpLsh32x32,
2394 {ir.OLSH, types.TUINT32, types.TUINT32}: ssa.OpLsh32x32,
2395 {ir.OLSH, types.TINT32, types.TUINT64}: ssa.OpLsh32x64,
2396 {ir.OLSH, types.TUINT32, types.TUINT64}: ssa.OpLsh32x64,
2398 {ir.OLSH, types.TINT64, types.TUINT8}: ssa.OpLsh64x8,
2399 {ir.OLSH, types.TUINT64, types.TUINT8}: ssa.OpLsh64x8,
2400 {ir.OLSH, types.TINT64, types.TUINT16}: ssa.OpLsh64x16,
2401 {ir.OLSH, types.TUINT64, types.TUINT16}: ssa.OpLsh64x16,
2402 {ir.OLSH, types.TINT64, types.TUINT32}: ssa.OpLsh64x32,
2403 {ir.OLSH, types.TUINT64, types.TUINT32}: ssa.OpLsh64x32,
2404 {ir.OLSH, types.TINT64, types.TUINT64}: ssa.OpLsh64x64,
2405 {ir.OLSH, types.TUINT64, types.TUINT64}: ssa.OpLsh64x64,
2407 {ir.ORSH, types.TINT8, types.TUINT8}: ssa.OpRsh8x8,
2408 {ir.ORSH, types.TUINT8, types.TUINT8}: ssa.OpRsh8Ux8,
2409 {ir.ORSH, types.TINT8, types.TUINT16}: ssa.OpRsh8x16,
2410 {ir.ORSH, types.TUINT8, types.TUINT16}: ssa.OpRsh8Ux16,
2411 {ir.ORSH, types.TINT8, types.TUINT32}: ssa.OpRsh8x32,
2412 {ir.ORSH, types.TUINT8, types.TUINT32}: ssa.OpRsh8Ux32,
2413 {ir.ORSH, types.TINT8, types.TUINT64}: ssa.OpRsh8x64,
2414 {ir.ORSH, types.TUINT8, types.TUINT64}: ssa.OpRsh8Ux64,
2416 {ir.ORSH, types.TINT16, types.TUINT8}: ssa.OpRsh16x8,
2417 {ir.ORSH, types.TUINT16, types.TUINT8}: ssa.OpRsh16Ux8,
2418 {ir.ORSH, types.TINT16, types.TUINT16}: ssa.OpRsh16x16,
2419 {ir.ORSH, types.TUINT16, types.TUINT16}: ssa.OpRsh16Ux16,
2420 {ir.ORSH, types.TINT16, types.TUINT32}: ssa.OpRsh16x32,
2421 {ir.ORSH, types.TUINT16, types.TUINT32}: ssa.OpRsh16Ux32,
2422 {ir.ORSH, types.TINT16, types.TUINT64}: ssa.OpRsh16x64,
2423 {ir.ORSH, types.TUINT16, types.TUINT64}: ssa.OpRsh16Ux64,
2425 {ir.ORSH, types.TINT32, types.TUINT8}: ssa.OpRsh32x8,
2426 {ir.ORSH, types.TUINT32, types.TUINT8}: ssa.OpRsh32Ux8,
2427 {ir.ORSH, types.TINT32, types.TUINT16}: ssa.OpRsh32x16,
2428 {ir.ORSH, types.TUINT32, types.TUINT16}: ssa.OpRsh32Ux16,
2429 {ir.ORSH, types.TINT32, types.TUINT32}: ssa.OpRsh32x32,
2430 {ir.ORSH, types.TUINT32, types.TUINT32}: ssa.OpRsh32Ux32,
2431 {ir.ORSH, types.TINT32, types.TUINT64}: ssa.OpRsh32x64,
2432 {ir.ORSH, types.TUINT32, types.TUINT64}: ssa.OpRsh32Ux64,
2434 {ir.ORSH, types.TINT64, types.TUINT8}: ssa.OpRsh64x8,
2435 {ir.ORSH, types.TUINT64, types.TUINT8}: ssa.OpRsh64Ux8,
2436 {ir.ORSH, types.TINT64, types.TUINT16}: ssa.OpRsh64x16,
2437 {ir.ORSH, types.TUINT64, types.TUINT16}: ssa.OpRsh64Ux16,
2438 {ir.ORSH, types.TINT64, types.TUINT32}: ssa.OpRsh64x32,
2439 {ir.ORSH, types.TUINT64, types.TUINT32}: ssa.OpRsh64Ux32,
2440 {ir.ORSH, types.TINT64, types.TUINT64}: ssa.OpRsh64x64,
2441 {ir.ORSH, types.TUINT64, types.TUINT64}: ssa.OpRsh64Ux64,
2444 func (s *state) ssaShiftOp(op ir.Op, t *types.Type, u *types.Type) ssa.Op {
2445 etype1 := s.concreteEtype(t)
2446 etype2 := s.concreteEtype(u)
2447 x, ok := shiftOpToSSA[opAndTwoTypes{op, etype1, etype2}]
2449 s.Fatalf("unhandled shift op %v etype=%s/%s", op, etype1, etype2)
2454 func (s *state) uintptrConstant(v uint64) *ssa.Value {
2455 if s.config.PtrSize == 4 {
2456 return s.newValue0I(ssa.OpConst32, types.Types[types.TUINTPTR], int64(v))
2458 return s.newValue0I(ssa.OpConst64, types.Types[types.TUINTPTR], int64(v))
2461 func (s *state) conv(n ir.Node, v *ssa.Value, ft, tt *types.Type) *ssa.Value {
2462 if ft.IsBoolean() && tt.IsKind(types.TUINT8) {
2463 // Bool -> uint8 is generated internally when indexing into runtime.staticbyte.
2464 return s.newValue1(ssa.OpCvtBoolToUint8, tt, v)
2466 if ft.IsInteger() && tt.IsInteger() {
2468 if tt.Size() == ft.Size() {
2470 } else if tt.Size() < ft.Size() {
2472 switch 10*ft.Size() + tt.Size() {
2474 op = ssa.OpTrunc16to8
2476 op = ssa.OpTrunc32to8
2478 op = ssa.OpTrunc32to16
2480 op = ssa.OpTrunc64to8
2482 op = ssa.OpTrunc64to16
2484 op = ssa.OpTrunc64to32
2486 s.Fatalf("weird integer truncation %v -> %v", ft, tt)
2488 } else if ft.IsSigned() {
2490 switch 10*ft.Size() + tt.Size() {
2492 op = ssa.OpSignExt8to16
2494 op = ssa.OpSignExt8to32
2496 op = ssa.OpSignExt8to64
2498 op = ssa.OpSignExt16to32
2500 op = ssa.OpSignExt16to64
2502 op = ssa.OpSignExt32to64
2504 s.Fatalf("bad integer sign extension %v -> %v", ft, tt)
2508 switch 10*ft.Size() + tt.Size() {
2510 op = ssa.OpZeroExt8to16
2512 op = ssa.OpZeroExt8to32
2514 op = ssa.OpZeroExt8to64
2516 op = ssa.OpZeroExt16to32
2518 op = ssa.OpZeroExt16to64
2520 op = ssa.OpZeroExt32to64
2522 s.Fatalf("weird integer sign extension %v -> %v", ft, tt)
2525 return s.newValue1(op, tt, v)
2528 if ft.IsComplex() && tt.IsComplex() {
2530 if ft.Size() == tt.Size() {
2537 s.Fatalf("weird complex conversion %v -> %v", ft, tt)
2539 } else if ft.Size() == 8 && tt.Size() == 16 {
2540 op = ssa.OpCvt32Fto64F
2541 } else if ft.Size() == 16 && tt.Size() == 8 {
2542 op = ssa.OpCvt64Fto32F
2544 s.Fatalf("weird complex conversion %v -> %v", ft, tt)
2546 ftp := types.FloatForComplex(ft)
2547 ttp := types.FloatForComplex(tt)
2548 return s.newValue2(ssa.OpComplexMake, tt,
2549 s.newValueOrSfCall1(op, ttp, s.newValue1(ssa.OpComplexReal, ftp, v)),
2550 s.newValueOrSfCall1(op, ttp, s.newValue1(ssa.OpComplexImag, ftp, v)))
2553 if tt.IsComplex() { // and ft is not complex
2554 // Needed for generics support - can't happen in normal Go code.
2555 et := types.FloatForComplex(tt)
2556 v = s.conv(n, v, ft, et)
2557 return s.newValue2(ssa.OpComplexMake, tt, v, s.zeroVal(et))
2560 if ft.IsFloat() || tt.IsFloat() {
2561 conv, ok := fpConvOpToSSA[twoTypes{s.concreteEtype(ft), s.concreteEtype(tt)}]
2562 if s.config.RegSize == 4 && Arch.LinkArch.Family != sys.MIPS && !s.softFloat {
2563 if conv1, ok1 := fpConvOpToSSA32[twoTypes{s.concreteEtype(ft), s.concreteEtype(tt)}]; ok1 {
2567 if Arch.LinkArch.Family == sys.ARM64 || Arch.LinkArch.Family == sys.Wasm || Arch.LinkArch.Family == sys.S390X || s.softFloat {
2568 if conv1, ok1 := uint64fpConvOpToSSA[twoTypes{s.concreteEtype(ft), s.concreteEtype(tt)}]; ok1 {
2573 if Arch.LinkArch.Family == sys.MIPS && !s.softFloat {
2574 if ft.Size() == 4 && ft.IsInteger() && !ft.IsSigned() {
2575 // tt is float32 or float64, and ft is also unsigned
2577 return s.uint32Tofloat32(n, v, ft, tt)
2580 return s.uint32Tofloat64(n, v, ft, tt)
2582 } else if tt.Size() == 4 && tt.IsInteger() && !tt.IsSigned() {
2583 // ft is float32 or float64, and tt is unsigned integer
2585 return s.float32ToUint32(n, v, ft, tt)
2588 return s.float64ToUint32(n, v, ft, tt)
2594 s.Fatalf("weird float conversion %v -> %v", ft, tt)
2596 op1, op2, it := conv.op1, conv.op2, conv.intermediateType
2598 if op1 != ssa.OpInvalid && op2 != ssa.OpInvalid {
2599 // normal case, not tripping over unsigned 64
2600 if op1 == ssa.OpCopy {
2601 if op2 == ssa.OpCopy {
2604 return s.newValueOrSfCall1(op2, tt, v)
2606 if op2 == ssa.OpCopy {
2607 return s.newValueOrSfCall1(op1, tt, v)
2609 return s.newValueOrSfCall1(op2, tt, s.newValueOrSfCall1(op1, types.Types[it], v))
2611 // Tricky 64-bit unsigned cases.
2613 // tt is float32 or float64, and ft is also unsigned
2615 return s.uint64Tofloat32(n, v, ft, tt)
2618 return s.uint64Tofloat64(n, v, ft, tt)
2620 s.Fatalf("weird unsigned integer to float conversion %v -> %v", ft, tt)
2622 // ft is float32 or float64, and tt is unsigned integer
2624 return s.float32ToUint64(n, v, ft, tt)
2627 return s.float64ToUint64(n, v, ft, tt)
2629 s.Fatalf("weird float to unsigned integer conversion %v -> %v", ft, tt)
2633 s.Fatalf("unhandled OCONV %s -> %s", ft.Kind(), tt.Kind())
2637 // expr converts the expression n to ssa, adds it to s and returns the ssa result.
2638 func (s *state) expr(n ir.Node) *ssa.Value {
2639 return s.exprCheckPtr(n, true)
2642 func (s *state) exprCheckPtr(n ir.Node, checkPtrOK bool) *ssa.Value {
2643 if ir.HasUniquePos(n) {
2644 // ONAMEs and named OLITERALs have the line number
2645 // of the decl, not the use. See issue 14742.
2650 s.stmtList(n.Init())
2652 case ir.OBYTES2STRTMP:
2653 n := n.(*ir.ConvExpr)
2654 slice := s.expr(n.X)
2655 ptr := s.newValue1(ssa.OpSlicePtr, s.f.Config.Types.BytePtr, slice)
2656 len := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], slice)
2657 return s.newValue2(ssa.OpStringMake, n.Type(), ptr, len)
2658 case ir.OSTR2BYTESTMP:
2659 n := n.(*ir.ConvExpr)
2661 ptr := s.newValue1(ssa.OpStringPtr, s.f.Config.Types.BytePtr, str)
2662 len := s.newValue1(ssa.OpStringLen, types.Types[types.TINT], str)
2663 return s.newValue3(ssa.OpSliceMake, n.Type(), ptr, len, len)
2665 n := n.(*ir.UnaryExpr)
2666 aux := n.X.(*ir.Name).Linksym()
2667 // OCFUNC is used to build function values, which must
2668 // always reference ABIInternal entry points.
2669 if aux.ABI() != obj.ABIInternal {
2670 s.Fatalf("expected ABIInternal: %v", aux.ABI())
2672 return s.entryNewValue1A(ssa.OpAddr, n.Type(), aux, s.sb)
2675 if n.Class == ir.PFUNC {
2676 // "value" of a function is the address of the function's closure
2677 sym := staticdata.FuncLinksym(n)
2678 return s.entryNewValue1A(ssa.OpAddr, types.NewPtr(n.Type()), sym, s.sb)
2681 return s.variable(n, n.Type())
2683 return s.load(n.Type(), s.addr(n))
2684 case ir.OLINKSYMOFFSET:
2685 n := n.(*ir.LinksymOffsetExpr)
2686 return s.load(n.Type(), s.addr(n))
2688 n := n.(*ir.NilExpr)
2692 return s.constSlice(t)
2693 case t.IsInterface():
2694 return s.constInterface(t)
2696 return s.constNil(t)
2699 switch u := n.Val(); u.Kind() {
2701 i := ir.IntVal(n.Type(), u)
2702 switch n.Type().Size() {
2704 return s.constInt8(n.Type(), int8(i))
2706 return s.constInt16(n.Type(), int16(i))
2708 return s.constInt32(n.Type(), int32(i))
2710 return s.constInt64(n.Type(), i)
2712 s.Fatalf("bad integer size %d", n.Type().Size())
2715 case constant.String:
2716 i := constant.StringVal(u)
2718 return s.constEmptyString(n.Type())
2720 return s.entryNewValue0A(ssa.OpConstString, n.Type(), ssa.StringToAux(i))
2722 return s.constBool(constant.BoolVal(u))
2723 case constant.Float:
2724 f, _ := constant.Float64Val(u)
2725 switch n.Type().Size() {
2727 return s.constFloat32(n.Type(), f)
2729 return s.constFloat64(n.Type(), f)
2731 s.Fatalf("bad float size %d", n.Type().Size())
2734 case constant.Complex:
2735 re, _ := constant.Float64Val(constant.Real(u))
2736 im, _ := constant.Float64Val(constant.Imag(u))
2737 switch n.Type().Size() {
2739 pt := types.Types[types.TFLOAT32]
2740 return s.newValue2(ssa.OpComplexMake, n.Type(),
2741 s.constFloat32(pt, re),
2742 s.constFloat32(pt, im))
2744 pt := types.Types[types.TFLOAT64]
2745 return s.newValue2(ssa.OpComplexMake, n.Type(),
2746 s.constFloat64(pt, re),
2747 s.constFloat64(pt, im))
2749 s.Fatalf("bad complex size %d", n.Type().Size())
2753 s.Fatalf("unhandled OLITERAL %v", u.Kind())
2757 n := n.(*ir.ConvExpr)
2761 // Assume everything will work out, so set up our return value.
2762 // Anything interesting that happens from here is a fatal.
2768 // Special case for not confusing GC and liveness.
2769 // We don't want pointers accidentally classified
2770 // as not-pointers or vice-versa because of copy
2772 if to.IsPtrShaped() != from.IsPtrShaped() {
2773 return s.newValue2(ssa.OpConvert, to, x, s.mem())
2776 v := s.newValue1(ssa.OpCopy, to, x) // ensure that v has the right type
2779 if to.Kind() == types.TFUNC && from.IsPtrShaped() {
2783 // named <--> unnamed type or typed <--> untyped const
2784 if from.Kind() == to.Kind() {
2788 // unsafe.Pointer <--> *T
2789 if to.IsUnsafePtr() && from.IsPtrShaped() || from.IsUnsafePtr() && to.IsPtrShaped() {
2790 if s.checkPtrEnabled && checkPtrOK && to.IsPtr() && from.IsUnsafePtr() {
2791 s.checkPtrAlignment(n, v, nil)
2797 if to.Kind() == types.TMAP && from.IsPtr() &&
2798 to.MapType().Hmap == from.Elem() {
2802 types.CalcSize(from)
2804 if from.Size() != to.Size() {
2805 s.Fatalf("CONVNOP width mismatch %v (%d) -> %v (%d)\n", from, from.Size(), to, to.Size())
2808 if etypesign(from.Kind()) != etypesign(to.Kind()) {
2809 s.Fatalf("CONVNOP sign mismatch %v (%s) -> %v (%s)\n", from, from.Kind(), to, to.Kind())
2813 if base.Flag.Cfg.Instrumenting {
2814 // These appear to be fine, but they fail the
2815 // integer constraint below, so okay them here.
2816 // Sample non-integer conversion: map[string]string -> *uint8
2820 if etypesign(from.Kind()) == 0 {
2821 s.Fatalf("CONVNOP unrecognized non-integer %v -> %v\n", from, to)
2825 // integer, same width, same sign
2829 n := n.(*ir.ConvExpr)
2831 return s.conv(n, x, n.X.Type(), n.Type())
2834 n := n.(*ir.TypeAssertExpr)
2835 res, _ := s.dottype(n, false)
2838 case ir.ODYNAMICDOTTYPE:
2839 n := n.(*ir.DynamicTypeAssertExpr)
2840 res, _ := s.dynamicDottype(n, false)
2844 case ir.OLT, ir.OEQ, ir.ONE, ir.OLE, ir.OGE, ir.OGT:
2845 n := n.(*ir.BinaryExpr)
2848 if n.X.Type().IsComplex() {
2849 pt := types.FloatForComplex(n.X.Type())
2850 op := s.ssaOp(ir.OEQ, pt)
2851 r := s.newValueOrSfCall2(op, types.Types[types.TBOOL], s.newValue1(ssa.OpComplexReal, pt, a), s.newValue1(ssa.OpComplexReal, pt, b))
2852 i := s.newValueOrSfCall2(op, types.Types[types.TBOOL], s.newValue1(ssa.OpComplexImag, pt, a), s.newValue1(ssa.OpComplexImag, pt, b))
2853 c := s.newValue2(ssa.OpAndB, types.Types[types.TBOOL], r, i)
2858 return s.newValue1(ssa.OpNot, types.Types[types.TBOOL], c)
2860 s.Fatalf("ordered complex compare %v", n.Op())
2864 // Convert OGE and OGT into OLE and OLT.
2868 op, a, b = ir.OLE, b, a
2870 op, a, b = ir.OLT, b, a
2872 if n.X.Type().IsFloat() {
2874 return s.newValueOrSfCall2(s.ssaOp(op, n.X.Type()), types.Types[types.TBOOL], a, b)
2876 // integer comparison
2877 return s.newValue2(s.ssaOp(op, n.X.Type()), types.Types[types.TBOOL], a, b)
2879 n := n.(*ir.BinaryExpr)
2882 if n.Type().IsComplex() {
2883 mulop := ssa.OpMul64F
2884 addop := ssa.OpAdd64F
2885 subop := ssa.OpSub64F
2886 pt := types.FloatForComplex(n.Type()) // Could be Float32 or Float64
2887 wt := types.Types[types.TFLOAT64] // Compute in Float64 to minimize cancellation error
2889 areal := s.newValue1(ssa.OpComplexReal, pt, a)
2890 breal := s.newValue1(ssa.OpComplexReal, pt, b)
2891 aimag := s.newValue1(ssa.OpComplexImag, pt, a)
2892 bimag := s.newValue1(ssa.OpComplexImag, pt, b)
2894 if pt != wt { // Widen for calculation
2895 areal = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, areal)
2896 breal = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, breal)
2897 aimag = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, aimag)
2898 bimag = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, bimag)
2901 xreal := s.newValueOrSfCall2(subop, wt, s.newValueOrSfCall2(mulop, wt, areal, breal), s.newValueOrSfCall2(mulop, wt, aimag, bimag))
2902 ximag := s.newValueOrSfCall2(addop, wt, s.newValueOrSfCall2(mulop, wt, areal, bimag), s.newValueOrSfCall2(mulop, wt, aimag, breal))
2904 if pt != wt { // Narrow to store back
2905 xreal = s.newValueOrSfCall1(ssa.OpCvt64Fto32F, pt, xreal)
2906 ximag = s.newValueOrSfCall1(ssa.OpCvt64Fto32F, pt, ximag)
2909 return s.newValue2(ssa.OpComplexMake, n.Type(), xreal, ximag)
2912 if n.Type().IsFloat() {
2913 return s.newValueOrSfCall2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2916 return s.newValue2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2919 n := n.(*ir.BinaryExpr)
2922 if n.Type().IsComplex() {
2923 // TODO this is not executed because the front-end substitutes a runtime call.
2924 // That probably ought to change; with modest optimization the widen/narrow
2925 // conversions could all be elided in larger expression trees.
2926 mulop := ssa.OpMul64F
2927 addop := ssa.OpAdd64F
2928 subop := ssa.OpSub64F
2929 divop := ssa.OpDiv64F
2930 pt := types.FloatForComplex(n.Type()) // Could be Float32 or Float64
2931 wt := types.Types[types.TFLOAT64] // Compute in Float64 to minimize cancellation error
2933 areal := s.newValue1(ssa.OpComplexReal, pt, a)
2934 breal := s.newValue1(ssa.OpComplexReal, pt, b)
2935 aimag := s.newValue1(ssa.OpComplexImag, pt, a)
2936 bimag := s.newValue1(ssa.OpComplexImag, pt, b)
2938 if pt != wt { // Widen for calculation
2939 areal = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, areal)
2940 breal = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, breal)
2941 aimag = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, aimag)
2942 bimag = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, bimag)
2945 denom := s.newValueOrSfCall2(addop, wt, s.newValueOrSfCall2(mulop, wt, breal, breal), s.newValueOrSfCall2(mulop, wt, bimag, bimag))
2946 xreal := s.newValueOrSfCall2(addop, wt, s.newValueOrSfCall2(mulop, wt, areal, breal), s.newValueOrSfCall2(mulop, wt, aimag, bimag))
2947 ximag := s.newValueOrSfCall2(subop, wt, s.newValueOrSfCall2(mulop, wt, aimag, breal), s.newValueOrSfCall2(mulop, wt, areal, bimag))
2949 // TODO not sure if this is best done in wide precision or narrow
2950 // Double-rounding might be an issue.
2951 // Note that the pre-SSA implementation does the entire calculation
2952 // in wide format, so wide is compatible.
2953 xreal = s.newValueOrSfCall2(divop, wt, xreal, denom)
2954 ximag = s.newValueOrSfCall2(divop, wt, ximag, denom)
2956 if pt != wt { // Narrow to store back
2957 xreal = s.newValueOrSfCall1(ssa.OpCvt64Fto32F, pt, xreal)
2958 ximag = s.newValueOrSfCall1(ssa.OpCvt64Fto32F, pt, ximag)
2960 return s.newValue2(ssa.OpComplexMake, n.Type(), xreal, ximag)
2962 if n.Type().IsFloat() {
2963 return s.newValueOrSfCall2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2965 return s.intDivide(n, a, b)
2967 n := n.(*ir.BinaryExpr)
2970 return s.intDivide(n, a, b)
2971 case ir.OADD, ir.OSUB:
2972 n := n.(*ir.BinaryExpr)
2975 if n.Type().IsComplex() {
2976 pt := types.FloatForComplex(n.Type())
2977 op := s.ssaOp(n.Op(), pt)
2978 return s.newValue2(ssa.OpComplexMake, n.Type(),
2979 s.newValueOrSfCall2(op, pt, s.newValue1(ssa.OpComplexReal, pt, a), s.newValue1(ssa.OpComplexReal, pt, b)),
2980 s.newValueOrSfCall2(op, pt, s.newValue1(ssa.OpComplexImag, pt, a), s.newValue1(ssa.OpComplexImag, pt, b)))
2982 if n.Type().IsFloat() {
2983 return s.newValueOrSfCall2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2985 return s.newValue2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2986 case ir.OAND, ir.OOR, ir.OXOR:
2987 n := n.(*ir.BinaryExpr)
2990 return s.newValue2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2992 n := n.(*ir.BinaryExpr)
2995 b = s.newValue1(s.ssaOp(ir.OBITNOT, b.Type), b.Type, b)
2996 return s.newValue2(s.ssaOp(ir.OAND, n.Type()), a.Type, a, b)
2997 case ir.OLSH, ir.ORSH:
2998 n := n.(*ir.BinaryExpr)
3003 cmp := s.newValue2(s.ssaOp(ir.OLE, bt), types.Types[types.TBOOL], s.zeroVal(bt), b)
3004 s.check(cmp, ir.Syms.Panicshift)
3005 bt = bt.ToUnsigned()
3007 return s.newValue2(s.ssaShiftOp(n.Op(), n.Type(), bt), a.Type, a, b)
3008 case ir.OANDAND, ir.OOROR:
3009 // To implement OANDAND (and OOROR), we introduce a
3010 // new temporary variable to hold the result. The
3011 // variable is associated with the OANDAND node in the
3012 // s.vars table (normally variables are only
3013 // associated with ONAME nodes). We convert
3020 // Using var in the subsequent block introduces the
3021 // necessary phi variable.
3022 n := n.(*ir.LogicalExpr)
3027 b.Kind = ssa.BlockIf
3029 // In theory, we should set b.Likely here based on context.
3030 // However, gc only gives us likeliness hints
3031 // in a single place, for plain OIF statements,
3032 // and passing around context is finnicky, so don't bother for now.
3034 bRight := s.f.NewBlock(ssa.BlockPlain)
3035 bResult := s.f.NewBlock(ssa.BlockPlain)
3036 if n.Op() == ir.OANDAND {
3038 b.AddEdgeTo(bResult)
3039 } else if n.Op() == ir.OOROR {
3040 b.AddEdgeTo(bResult)
3044 s.startBlock(bRight)
3049 b.AddEdgeTo(bResult)
3051 s.startBlock(bResult)
3052 return s.variable(n, types.Types[types.TBOOL])
3054 n := n.(*ir.BinaryExpr)
3057 return s.newValue2(ssa.OpComplexMake, n.Type(), r, i)
3061 n := n.(*ir.UnaryExpr)
3063 if n.Type().IsComplex() {
3064 tp := types.FloatForComplex(n.Type())
3065 negop := s.ssaOp(n.Op(), tp)
3066 return s.newValue2(ssa.OpComplexMake, n.Type(),
3067 s.newValue1(negop, tp, s.newValue1(ssa.OpComplexReal, tp, a)),
3068 s.newValue1(negop, tp, s.newValue1(ssa.OpComplexImag, tp, a)))
3070 return s.newValue1(s.ssaOp(n.Op(), n.Type()), a.Type, a)
3071 case ir.ONOT, ir.OBITNOT:
3072 n := n.(*ir.UnaryExpr)
3074 return s.newValue1(s.ssaOp(n.Op(), n.Type()), a.Type, a)
3075 case ir.OIMAG, ir.OREAL:
3076 n := n.(*ir.UnaryExpr)
3078 return s.newValue1(s.ssaOp(n.Op(), n.X.Type()), n.Type(), a)
3080 n := n.(*ir.UnaryExpr)
3084 n := n.(*ir.AddrExpr)
3088 n := n.(*ir.ResultExpr)
3089 if s.prevCall == nil || s.prevCall.Op != ssa.OpStaticLECall && s.prevCall.Op != ssa.OpInterLECall && s.prevCall.Op != ssa.OpClosureLECall {
3090 panic("Expected to see a previous call")
3094 panic(fmt.Errorf("ORESULT %v does not match call %s", n, s.prevCall))
3096 return s.resultOfCall(s.prevCall, which, n.Type())
3099 n := n.(*ir.StarExpr)
3100 p := s.exprPtr(n.X, n.Bounded(), n.Pos())
3101 return s.load(n.Type(), p)
3104 n := n.(*ir.SelectorExpr)
3105 if n.X.Op() == ir.OSTRUCTLIT {
3106 // All literals with nonzero fields have already been
3107 // rewritten during walk. Any that remain are just T{}
3108 // or equivalents. Use the zero value.
3109 if !ir.IsZero(n.X) {
3110 s.Fatalf("literal with nonzero value in SSA: %v", n.X)
3112 return s.zeroVal(n.Type())
3114 // If n is addressable and can't be represented in
3115 // SSA, then load just the selected field. This
3116 // prevents false memory dependencies in race/msan/asan
3118 if ir.IsAddressable(n) && !s.canSSA(n) {
3120 return s.load(n.Type(), p)
3123 return s.newValue1I(ssa.OpStructSelect, n.Type(), int64(fieldIdx(n)), v)
3126 n := n.(*ir.SelectorExpr)
3127 p := s.exprPtr(n.X, n.Bounded(), n.Pos())
3128 p = s.newValue1I(ssa.OpOffPtr, types.NewPtr(n.Type()), n.Offset(), p)
3129 return s.load(n.Type(), p)
3132 n := n.(*ir.IndexExpr)
3134 case n.X.Type().IsString():
3135 if n.Bounded() && ir.IsConst(n.X, constant.String) && ir.IsConst(n.Index, constant.Int) {
3136 // Replace "abc"[1] with 'b'.
3137 // Delayed until now because "abc"[1] is not an ideal constant.
3138 // See test/fixedbugs/issue11370.go.
3139 return s.newValue0I(ssa.OpConst8, types.Types[types.TUINT8], int64(int8(ir.StringVal(n.X)[ir.Int64Val(n.Index)])))
3142 i := s.expr(n.Index)
3143 len := s.newValue1(ssa.OpStringLen, types.Types[types.TINT], a)
3144 i = s.boundsCheck(i, len, ssa.BoundsIndex, n.Bounded())
3145 ptrtyp := s.f.Config.Types.BytePtr
3146 ptr := s.newValue1(ssa.OpStringPtr, ptrtyp, a)
3147 if ir.IsConst(n.Index, constant.Int) {
3148 ptr = s.newValue1I(ssa.OpOffPtr, ptrtyp, ir.Int64Val(n.Index), ptr)
3150 ptr = s.newValue2(ssa.OpAddPtr, ptrtyp, ptr, i)
3152 return s.load(types.Types[types.TUINT8], ptr)
3153 case n.X.Type().IsSlice():
3155 return s.load(n.X.Type().Elem(), p)
3156 case n.X.Type().IsArray():
3157 if TypeOK(n.X.Type()) {
3158 // SSA can handle arrays of length at most 1.
3159 bound := n.X.Type().NumElem()
3161 i := s.expr(n.Index)
3163 // Bounds check will never succeed. Might as well
3164 // use constants for the bounds check.
3165 z := s.constInt(types.Types[types.TINT], 0)
3166 s.boundsCheck(z, z, ssa.BoundsIndex, false)
3167 // The return value won't be live, return junk.
3168 // But not quite junk, in case bounds checks are turned off. See issue 48092.
3169 return s.zeroVal(n.Type())
3171 len := s.constInt(types.Types[types.TINT], bound)
3172 s.boundsCheck(i, len, ssa.BoundsIndex, n.Bounded()) // checks i == 0
3173 return s.newValue1I(ssa.OpArraySelect, n.Type(), 0, a)
3176 return s.load(n.X.Type().Elem(), p)
3178 s.Fatalf("bad type for index %v", n.X.Type())
3182 case ir.OLEN, ir.OCAP:
3183 n := n.(*ir.UnaryExpr)
3185 case n.X.Type().IsSlice():
3186 op := ssa.OpSliceLen
3187 if n.Op() == ir.OCAP {
3190 return s.newValue1(op, types.Types[types.TINT], s.expr(n.X))
3191 case n.X.Type().IsString(): // string; not reachable for OCAP
3192 return s.newValue1(ssa.OpStringLen, types.Types[types.TINT], s.expr(n.X))
3193 case n.X.Type().IsMap(), n.X.Type().IsChan():
3194 return s.referenceTypeBuiltin(n, s.expr(n.X))
3196 return s.constInt(types.Types[types.TINT], n.X.Type().NumElem())
3200 n := n.(*ir.UnaryExpr)
3202 if n.X.Type().IsSlice() {
3204 return s.newValue1(ssa.OpSlicePtr, n.Type(), a)
3206 return s.newValue1(ssa.OpSlicePtrUnchecked, n.Type(), a)
3208 return s.newValue1(ssa.OpStringPtr, n.Type(), a)
3212 n := n.(*ir.UnaryExpr)
3214 return s.newValue1(ssa.OpITab, n.Type(), a)
3217 n := n.(*ir.UnaryExpr)
3219 return s.newValue1(ssa.OpIData, n.Type(), a)
3222 n := n.(*ir.BinaryExpr)
3225 return s.newValue2(ssa.OpIMake, n.Type(), tab, data)
3227 case ir.OSLICEHEADER:
3228 n := n.(*ir.SliceHeaderExpr)
3232 return s.newValue3(ssa.OpSliceMake, n.Type(), p, l, c)
3234 case ir.OSTRINGHEADER:
3235 n := n.(*ir.StringHeaderExpr)
3238 return s.newValue2(ssa.OpStringMake, n.Type(), p, l)
3240 case ir.OSLICE, ir.OSLICEARR, ir.OSLICE3, ir.OSLICE3ARR:
3241 n := n.(*ir.SliceExpr)
3242 check := s.checkPtrEnabled && n.Op() == ir.OSLICE3ARR && n.X.Op() == ir.OCONVNOP && n.X.(*ir.ConvExpr).X.Type().IsUnsafePtr()
3243 v := s.exprCheckPtr(n.X, !check)
3244 var i, j, k *ssa.Value
3254 p, l, c := s.slice(v, i, j, k, n.Bounded())
3256 // Emit checkptr instrumentation after bound check to prevent false positive, see #46938.
3257 s.checkPtrAlignment(n.X.(*ir.ConvExpr), v, s.conv(n.Max, k, k.Type, types.Types[types.TUINTPTR]))
3259 return s.newValue3(ssa.OpSliceMake, n.Type(), p, l, c)
3262 n := n.(*ir.SliceExpr)
3271 p, l, _ := s.slice(v, i, j, nil, n.Bounded())
3272 return s.newValue2(ssa.OpStringMake, n.Type(), p, l)
3274 case ir.OSLICE2ARRPTR:
3275 // if arrlen > slice.len {
3279 n := n.(*ir.ConvExpr)
3281 nelem := n.Type().Elem().NumElem()
3282 arrlen := s.constInt(types.Types[types.TINT], nelem)
3283 cap := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], v)
3284 s.boundsCheck(arrlen, cap, ssa.BoundsConvert, false)
3285 op := ssa.OpSlicePtr
3287 op = ssa.OpSlicePtrUnchecked
3289 return s.newValue1(op, n.Type(), v)
3292 n := n.(*ir.CallExpr)
3293 if ir.IsIntrinsicCall(n) {
3294 return s.intrinsicCall(n)
3299 n := n.(*ir.CallExpr)
3300 return s.callResult(n, callNormal)
3303 n := n.(*ir.CallExpr)
3304 return s.newValue1(ssa.OpGetG, n.Type(), s.mem())
3306 case ir.OGETCALLERPC:
3307 n := n.(*ir.CallExpr)
3308 return s.newValue0(ssa.OpGetCallerPC, n.Type())
3310 case ir.OGETCALLERSP:
3311 n := n.(*ir.CallExpr)
3312 return s.newValue1(ssa.OpGetCallerSP, n.Type(), s.mem())
3315 return s.append(n.(*ir.CallExpr), false)
3317 case ir.OMIN, ir.OMAX:
3318 return s.minMax(n.(*ir.CallExpr))
3320 case ir.OSTRUCTLIT, ir.OARRAYLIT:
3321 // All literals with nonzero fields have already been
3322 // rewritten during walk. Any that remain are just T{}
3323 // or equivalents. Use the zero value.
3324 n := n.(*ir.CompLitExpr)
3326 s.Fatalf("literal with nonzero value in SSA: %v", n)
3328 return s.zeroVal(n.Type())
3331 n := n.(*ir.UnaryExpr)
3332 var rtype *ssa.Value
3333 if x, ok := n.X.(*ir.DynamicType); ok && x.Op() == ir.ODYNAMICTYPE {
3334 rtype = s.expr(x.RType)
3336 return s.newObject(n.Type().Elem(), rtype)
3339 n := n.(*ir.BinaryExpr)
3343 // Force len to uintptr to prevent misuse of garbage bits in the
3344 // upper part of the register (#48536).
3345 len = s.conv(n, len, len.Type, types.Types[types.TUINTPTR])
3347 return s.newValue2(ssa.OpAddPtr, n.Type(), ptr, len)
3350 s.Fatalf("unhandled expr %v", n.Op())
3355 func (s *state) resultOfCall(c *ssa.Value, which int64, t *types.Type) *ssa.Value {
3356 aux := c.Aux.(*ssa.AuxCall)
3357 pa := aux.ParamAssignmentForResult(which)
3358 // TODO(register args) determine if in-memory TypeOK is better loaded early from SelectNAddr or later when SelectN is expanded.
3359 // SelectN is better for pattern-matching and possible call-aware analysis we might want to do in the future.
3360 if len(pa.Registers) == 0 && !TypeOK(t) {
3361 addr := s.newValue1I(ssa.OpSelectNAddr, types.NewPtr(t), which, c)
3362 return s.rawLoad(t, addr)
3364 return s.newValue1I(ssa.OpSelectN, t, which, c)
3367 func (s *state) resultAddrOfCall(c *ssa.Value, which int64, t *types.Type) *ssa.Value {
3368 aux := c.Aux.(*ssa.AuxCall)
3369 pa := aux.ParamAssignmentForResult(which)
3370 if len(pa.Registers) == 0 {
3371 return s.newValue1I(ssa.OpSelectNAddr, types.NewPtr(t), which, c)
3373 _, addr := s.temp(c.Pos, t)
3374 rval := s.newValue1I(ssa.OpSelectN, t, which, c)
3375 s.vars[memVar] = s.newValue3Apos(ssa.OpStore, types.TypeMem, t, addr, rval, s.mem(), false)
3379 // append converts an OAPPEND node to SSA.
3380 // If inplace is false, it converts the OAPPEND expression n to an ssa.Value,
3381 // adds it to s, and returns the Value.
3382 // If inplace is true, it writes the result of the OAPPEND expression n
3383 // back to the slice being appended to, and returns nil.
3384 // inplace MUST be set to false if the slice can be SSA'd.
3385 // Note: this code only handles fixed-count appends. Dotdotdot appends
3386 // have already been rewritten at this point (by walk).
3387 func (s *state) append(n *ir.CallExpr, inplace bool) *ssa.Value {
3388 // If inplace is false, process as expression "append(s, e1, e2, e3)":
3390 // ptr, len, cap := s
3392 // if uint(len) > uint(cap) {
3393 // ptr, len, cap = growslice(ptr, len, cap, 3, typ)
3394 // Note that len is unmodified by growslice.
3396 // // with write barriers, if needed:
3397 // *(ptr+(len-3)) = e1
3398 // *(ptr+(len-2)) = e2
3399 // *(ptr+(len-1)) = e3
3400 // return makeslice(ptr, len, cap)
3403 // If inplace is true, process as statement "s = append(s, e1, e2, e3)":
3406 // ptr, len, cap := s
3408 // if uint(len) > uint(cap) {
3409 // ptr, len, cap = growslice(ptr, len, cap, 3, typ)
3410 // vardef(a) // if necessary, advise liveness we are writing a new a
3411 // *a.cap = cap // write before ptr to avoid a spill
3412 // *a.ptr = ptr // with write barrier
3415 // // with write barriers, if needed:
3416 // *(ptr+(len-3)) = e1
3417 // *(ptr+(len-2)) = e2
3418 // *(ptr+(len-1)) = e3
3420 et := n.Type().Elem()
3421 pt := types.NewPtr(et)
3424 sn := n.Args[0] // the slice node is the first in the list
3425 var slice, addr *ssa.Value
3428 slice = s.load(n.Type(), addr)
3433 // Allocate new blocks
3434 grow := s.f.NewBlock(ssa.BlockPlain)
3435 assign := s.f.NewBlock(ssa.BlockPlain)
3437 // Decomposse input slice.
3438 p := s.newValue1(ssa.OpSlicePtr, pt, slice)
3439 l := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], slice)
3440 c := s.newValue1(ssa.OpSliceCap, types.Types[types.TINT], slice)
3442 // Add number of new elements to length.
3443 nargs := s.constInt(types.Types[types.TINT], int64(len(n.Args)-1))
3444 l = s.newValue2(s.ssaOp(ir.OADD, types.Types[types.TINT]), types.Types[types.TINT], l, nargs)
3446 // Decide if we need to grow
3447 cmp := s.newValue2(s.ssaOp(ir.OLT, types.Types[types.TUINT]), types.Types[types.TBOOL], c, l)
3449 // Record values of ptr/len/cap before branch.
3457 b.Kind = ssa.BlockIf
3458 b.Likely = ssa.BranchUnlikely
3465 taddr := s.expr(n.X)
3466 r := s.rtcall(ir.Syms.Growslice, true, []*types.Type{n.Type()}, p, l, c, nargs, taddr)
3468 // Decompose output slice
3469 p = s.newValue1(ssa.OpSlicePtr, pt, r[0])
3470 l = s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], r[0])
3471 c = s.newValue1(ssa.OpSliceCap, types.Types[types.TINT], r[0])
3477 if sn.Op() == ir.ONAME {
3479 if sn.Class != ir.PEXTERN {
3480 // Tell liveness we're about to build a new slice
3481 s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, sn, s.mem())
3484 capaddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, types.SliceCapOffset, addr)
3485 s.store(types.Types[types.TINT], capaddr, c)
3486 s.store(pt, addr, p)
3492 // assign new elements to slots
3493 s.startBlock(assign)
3494 p = s.variable(ptrVar, pt) // generates phi for ptr
3495 l = s.variable(lenVar, types.Types[types.TINT]) // generates phi for len
3497 c = s.variable(capVar, types.Types[types.TINT]) // generates phi for cap
3501 // Update length in place.
3502 // We have to wait until here to make sure growslice succeeded.
3503 lenaddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, types.SliceLenOffset, addr)
3504 s.store(types.Types[types.TINT], lenaddr, l)
3508 type argRec struct {
3509 // if store is true, we're appending the value v. If false, we're appending the
3514 args := make([]argRec, 0, len(n.Args[1:]))
3515 for _, n := range n.Args[1:] {
3516 if TypeOK(n.Type()) {
3517 args = append(args, argRec{v: s.expr(n), store: true})
3520 args = append(args, argRec{v: v})
3524 // Write args into slice.
3525 oldLen := s.newValue2(s.ssaOp(ir.OSUB, types.Types[types.TINT]), types.Types[types.TINT], l, nargs)
3526 p2 := s.newValue2(ssa.OpPtrIndex, pt, p, oldLen)
3527 for i, arg := range args {
3528 addr := s.newValue2(ssa.OpPtrIndex, pt, p2, s.constInt(types.Types[types.TINT], int64(i)))
3530 s.storeType(et, addr, arg.v, 0, true)
3532 s.move(et, addr, arg.v)
3536 // The following deletions have no practical effect at this time
3537 // because state.vars has been reset by the preceding state.startBlock.
3538 // They only enforce the fact that these variables are no longer need in
3539 // the current scope.
3540 delete(s.vars, ptrVar)
3541 delete(s.vars, lenVar)
3543 delete(s.vars, capVar)
3550 return s.newValue3(ssa.OpSliceMake, n.Type(), p, l, c)
3553 // minMax converts an OMIN/OMAX builtin call into SSA.
3554 func (s *state) minMax(n *ir.CallExpr) *ssa.Value {
3555 // The OMIN/OMAX builtin is variadic, but its semantics are
3556 // equivalent to left-folding a binary min/max operation across the
3558 fold := func(op func(x, a *ssa.Value) *ssa.Value) *ssa.Value {
3559 x := s.expr(n.Args[0])
3560 for _, arg := range n.Args[1:] {
3561 x = op(x, s.expr(arg))
3568 if typ.IsFloat() || typ.IsString() {
3569 // min/max semantics for floats are tricky because of NaNs and
3570 // negative zero, so we let the runtime handle this instead.
3572 // Strings are conceptually simpler, but we currently desugar
3573 // string comparisons during walk, not ssagen.
3577 case types.TFLOAT32:
3584 case types.TFLOAT64:
3599 fn := typecheck.LookupRuntimeFunc(name)
3601 return fold(func(x, a *ssa.Value) *ssa.Value {
3602 return s.rtcall(fn, true, []*types.Type{typ}, x, a)[0]
3606 lt := s.ssaOp(ir.OLT, typ)
3608 return fold(func(x, a *ssa.Value) *ssa.Value {
3612 return s.ternary(s.newValue2(lt, types.Types[types.TBOOL], a, x), a, x)
3615 return s.ternary(s.newValue2(lt, types.Types[types.TBOOL], x, a), a, x)
3617 panic("unreachable")
3621 // ternary emits code to evaluate cond ? x : y.
3622 func (s *state) ternary(cond, x, y *ssa.Value) *ssa.Value {
3623 // Note that we need a new ternaryVar each time (unlike okVar where we can
3624 // reuse the variable) because it might have a different type every time.
3625 ternaryVar := ssaMarker("ternary")
3627 bThen := s.f.NewBlock(ssa.BlockPlain)
3628 bElse := s.f.NewBlock(ssa.BlockPlain)
3629 bEnd := s.f.NewBlock(ssa.BlockPlain)
3632 b.Kind = ssa.BlockIf
3638 s.vars[ternaryVar] = x
3639 s.endBlock().AddEdgeTo(bEnd)
3642 s.vars[ternaryVar] = y
3643 s.endBlock().AddEdgeTo(bEnd)
3646 r := s.variable(ternaryVar, x.Type)
3647 delete(s.vars, ternaryVar)
3651 // condBranch evaluates the boolean expression cond and branches to yes
3652 // if cond is true and no if cond is false.
3653 // This function is intended to handle && and || better than just calling
3654 // s.expr(cond) and branching on the result.
3655 func (s *state) condBranch(cond ir.Node, yes, no *ssa.Block, likely int8) {
3658 cond := cond.(*ir.LogicalExpr)
3659 mid := s.f.NewBlock(ssa.BlockPlain)
3660 s.stmtList(cond.Init())
3661 s.condBranch(cond.X, mid, no, max8(likely, 0))
3663 s.condBranch(cond.Y, yes, no, likely)
3665 // Note: if likely==1, then both recursive calls pass 1.
3666 // If likely==-1, then we don't have enough information to decide
3667 // whether the first branch is likely or not. So we pass 0 for
3668 // the likeliness of the first branch.
3669 // TODO: have the frontend give us branch prediction hints for
3670 // OANDAND and OOROR nodes (if it ever has such info).
3672 cond := cond.(*ir.LogicalExpr)
3673 mid := s.f.NewBlock(ssa.BlockPlain)
3674 s.stmtList(cond.Init())
3675 s.condBranch(cond.X, yes, mid, min8(likely, 0))
3677 s.condBranch(cond.Y, yes, no, likely)
3679 // Note: if likely==-1, then both recursive calls pass -1.
3680 // If likely==1, then we don't have enough info to decide
3681 // the likelihood of the first branch.
3683 cond := cond.(*ir.UnaryExpr)
3684 s.stmtList(cond.Init())
3685 s.condBranch(cond.X, no, yes, -likely)
3688 cond := cond.(*ir.ConvExpr)
3689 s.stmtList(cond.Init())
3690 s.condBranch(cond.X, yes, no, likely)
3695 b.Kind = ssa.BlockIf
3697 b.Likely = ssa.BranchPrediction(likely) // gc and ssa both use -1/0/+1 for likeliness
3705 skipPtr skipMask = 1 << iota
3710 // assign does left = right.
3711 // Right has already been evaluated to ssa, left has not.
3712 // If deref is true, then we do left = *right instead (and right has already been nil-checked).
3713 // If deref is true and right == nil, just do left = 0.
3714 // skip indicates assignments (at the top level) that can be avoided.
3715 // mayOverlap indicates whether left&right might partially overlap in memory. Default is false.
3716 func (s *state) assign(left ir.Node, right *ssa.Value, deref bool, skip skipMask) {
3717 s.assignWhichMayOverlap(left, right, deref, skip, false)
3719 func (s *state) assignWhichMayOverlap(left ir.Node, right *ssa.Value, deref bool, skip skipMask, mayOverlap bool) {
3720 if left.Op() == ir.ONAME && ir.IsBlank(left) {
3727 s.Fatalf("can SSA LHS %v but not RHS %s", left, right)
3729 if left.Op() == ir.ODOT {
3730 // We're assigning to a field of an ssa-able value.
3731 // We need to build a new structure with the new value for the
3732 // field we're assigning and the old values for the other fields.
3734 // type T struct {a, b, c int}
3737 // For the x.b = 5 assignment we want to generate x = T{x.a, 5, x.c}
3739 // Grab information about the structure type.
3740 left := left.(*ir.SelectorExpr)
3743 idx := fieldIdx(left)
3745 // Grab old value of structure.
3746 old := s.expr(left.X)
3748 // Make new structure.
3749 new := s.newValue0(ssa.StructMakeOp(t.NumFields()), t)
3751 // Add fields as args.
3752 for i := 0; i < nf; i++ {
3756 new.AddArg(s.newValue1I(ssa.OpStructSelect, t.FieldType(i), int64(i), old))
3760 // Recursively assign the new value we've made to the base of the dot op.
3761 s.assign(left.X, new, false, 0)
3762 // TODO: do we need to update named values here?
3765 if left.Op() == ir.OINDEX && left.(*ir.IndexExpr).X.Type().IsArray() {
3766 left := left.(*ir.IndexExpr)
3767 s.pushLine(left.Pos())
3769 // We're assigning to an element of an ssa-able array.
3774 i := s.expr(left.Index) // index
3776 // The bounds check must fail. Might as well
3777 // ignore the actual index and just use zeros.
3778 z := s.constInt(types.Types[types.TINT], 0)
3779 s.boundsCheck(z, z, ssa.BoundsIndex, false)
3783 s.Fatalf("assigning to non-1-length array")
3785 // Rewrite to a = [1]{v}
3786 len := s.constInt(types.Types[types.TINT], 1)
3787 s.boundsCheck(i, len, ssa.BoundsIndex, false) // checks i == 0
3788 v := s.newValue1(ssa.OpArrayMake1, t, right)
3789 s.assign(left.X, v, false, 0)
3792 left := left.(*ir.Name)
3793 // Update variable assignment.
3794 s.vars[left] = right
3795 s.addNamedValue(left, right)
3799 // If this assignment clobbers an entire local variable, then emit
3800 // OpVarDef so liveness analysis knows the variable is redefined.
3801 if base, ok := clobberBase(left).(*ir.Name); ok && base.OnStack() && skip == 0 && t.HasPointers() {
3802 s.vars[memVar] = s.newValue1Apos(ssa.OpVarDef, types.TypeMem, base, s.mem(), !ir.IsAutoTmp(base))
3805 // Left is not ssa-able. Compute its address.
3806 addr := s.addr(left)
3807 if ir.IsReflectHeaderDataField(left) {
3808 // Package unsafe's documentation says storing pointers into
3809 // reflect.SliceHeader and reflect.StringHeader's Data fields
3810 // is valid, even though they have type uintptr (#19168).
3811 // Mark it pointer type to signal the writebarrier pass to
3812 // insert a write barrier.
3813 t = types.Types[types.TUNSAFEPTR]
3816 // Treat as a mem->mem move.
3820 s.moveWhichMayOverlap(t, addr, right, mayOverlap)
3824 // Treat as a store.
3825 s.storeType(t, addr, right, skip, !ir.IsAutoTmp(left))
3828 // zeroVal returns the zero value for type t.
3829 func (s *state) zeroVal(t *types.Type) *ssa.Value {
3834 return s.constInt8(t, 0)
3836 return s.constInt16(t, 0)
3838 return s.constInt32(t, 0)
3840 return s.constInt64(t, 0)
3842 s.Fatalf("bad sized integer type %v", t)
3847 return s.constFloat32(t, 0)
3849 return s.constFloat64(t, 0)
3851 s.Fatalf("bad sized float type %v", t)
3856 z := s.constFloat32(types.Types[types.TFLOAT32], 0)
3857 return s.entryNewValue2(ssa.OpComplexMake, t, z, z)
3859 z := s.constFloat64(types.Types[types.TFLOAT64], 0)
3860 return s.entryNewValue2(ssa.OpComplexMake, t, z, z)
3862 s.Fatalf("bad sized complex type %v", t)
3866 return s.constEmptyString(t)
3867 case t.IsPtrShaped():
3868 return s.constNil(t)
3870 return s.constBool(false)
3871 case t.IsInterface():
3872 return s.constInterface(t)
3874 return s.constSlice(t)
3877 v := s.entryNewValue0(ssa.StructMakeOp(t.NumFields()), t)
3878 for i := 0; i < n; i++ {
3879 v.AddArg(s.zeroVal(t.FieldType(i)))
3883 switch t.NumElem() {
3885 return s.entryNewValue0(ssa.OpArrayMake0, t)
3887 return s.entryNewValue1(ssa.OpArrayMake1, t, s.zeroVal(t.Elem()))
3890 s.Fatalf("zero for type %v not implemented", t)
3897 callNormal callKind = iota
3904 type sfRtCallDef struct {
3909 var softFloatOps map[ssa.Op]sfRtCallDef
3911 func softfloatInit() {
3912 // Some of these operations get transformed by sfcall.
3913 softFloatOps = map[ssa.Op]sfRtCallDef{
3914 ssa.OpAdd32F: {typecheck.LookupRuntimeFunc("fadd32"), types.TFLOAT32},
3915 ssa.OpAdd64F: {typecheck.LookupRuntimeFunc("fadd64"), types.TFLOAT64},
3916 ssa.OpSub32F: {typecheck.LookupRuntimeFunc("fadd32"), types.TFLOAT32},
3917 ssa.OpSub64F: {typecheck.LookupRuntimeFunc("fadd64"), types.TFLOAT64},
3918 ssa.OpMul32F: {typecheck.LookupRuntimeFunc("fmul32"), types.TFLOAT32},
3919 ssa.OpMul64F: {typecheck.LookupRuntimeFunc("fmul64"), types.TFLOAT64},
3920 ssa.OpDiv32F: {typecheck.LookupRuntimeFunc("fdiv32"), types.TFLOAT32},
3921 ssa.OpDiv64F: {typecheck.LookupRuntimeFunc("fdiv64"), types.TFLOAT64},
3923 ssa.OpEq64F: {typecheck.LookupRuntimeFunc("feq64"), types.TBOOL},
3924 ssa.OpEq32F: {typecheck.LookupRuntimeFunc("feq32"), types.TBOOL},
3925 ssa.OpNeq64F: {typecheck.LookupRuntimeFunc("feq64"), types.TBOOL},
3926 ssa.OpNeq32F: {typecheck.LookupRuntimeFunc("feq32"), types.TBOOL},
3927 ssa.OpLess64F: {typecheck.LookupRuntimeFunc("fgt64"), types.TBOOL},
3928 ssa.OpLess32F: {typecheck.LookupRuntimeFunc("fgt32"), types.TBOOL},
3929 ssa.OpLeq64F: {typecheck.LookupRuntimeFunc("fge64"), types.TBOOL},
3930 ssa.OpLeq32F: {typecheck.LookupRuntimeFunc("fge32"), types.TBOOL},
3932 ssa.OpCvt32to32F: {typecheck.LookupRuntimeFunc("fint32to32"), types.TFLOAT32},
3933 ssa.OpCvt32Fto32: {typecheck.LookupRuntimeFunc("f32toint32"), types.TINT32},
3934 ssa.OpCvt64to32F: {typecheck.LookupRuntimeFunc("fint64to32"), types.TFLOAT32},
3935 ssa.OpCvt32Fto64: {typecheck.LookupRuntimeFunc("f32toint64"), types.TINT64},
3936 ssa.OpCvt64Uto32F: {typecheck.LookupRuntimeFunc("fuint64to32"), types.TFLOAT32},
3937 ssa.OpCvt32Fto64U: {typecheck.LookupRuntimeFunc("f32touint64"), types.TUINT64},
3938 ssa.OpCvt32to64F: {typecheck.LookupRuntimeFunc("fint32to64"), types.TFLOAT64},
3939 ssa.OpCvt64Fto32: {typecheck.LookupRuntimeFunc("f64toint32"), types.TINT32},
3940 ssa.OpCvt64to64F: {typecheck.LookupRuntimeFunc("fint64to64"), types.TFLOAT64},
3941 ssa.OpCvt64Fto64: {typecheck.LookupRuntimeFunc("f64toint64"), types.TINT64},
3942 ssa.OpCvt64Uto64F: {typecheck.LookupRuntimeFunc("fuint64to64"), types.TFLOAT64},
3943 ssa.OpCvt64Fto64U: {typecheck.LookupRuntimeFunc("f64touint64"), types.TUINT64},
3944 ssa.OpCvt32Fto64F: {typecheck.LookupRuntimeFunc("f32to64"), types.TFLOAT64},
3945 ssa.OpCvt64Fto32F: {typecheck.LookupRuntimeFunc("f64to32"), types.TFLOAT32},
3949 // TODO: do not emit sfcall if operation can be optimized to constant in later
3951 func (s *state) sfcall(op ssa.Op, args ...*ssa.Value) (*ssa.Value, bool) {
3952 f2i := func(t *types.Type) *types.Type {
3954 case types.TFLOAT32:
3955 return types.Types[types.TUINT32]
3956 case types.TFLOAT64:
3957 return types.Types[types.TUINT64]
3962 if callDef, ok := softFloatOps[op]; ok {
3968 args[0], args[1] = args[1], args[0]
3971 args[1] = s.newValue1(s.ssaOp(ir.ONEG, types.Types[callDef.rtype]), args[1].Type, args[1])
3974 // runtime functions take uints for floats and returns uints.
3975 // Convert to uints so we use the right calling convention.
3976 for i, a := range args {
3977 if a.Type.IsFloat() {
3978 args[i] = s.newValue1(ssa.OpCopy, f2i(a.Type), a)
3982 rt := types.Types[callDef.rtype]
3983 result := s.rtcall(callDef.rtfn, true, []*types.Type{f2i(rt)}, args...)[0]
3985 result = s.newValue1(ssa.OpCopy, rt, result)
3987 if op == ssa.OpNeq32F || op == ssa.OpNeq64F {
3988 result = s.newValue1(ssa.OpNot, result.Type, result)
3995 var intrinsics map[intrinsicKey]intrinsicBuilder
3997 // An intrinsicBuilder converts a call node n into an ssa value that
3998 // implements that call as an intrinsic. args is a list of arguments to the func.
3999 type intrinsicBuilder func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value
4001 type intrinsicKey struct {
4008 intrinsics = map[intrinsicKey]intrinsicBuilder{}
4013 var lwatomics []*sys.Arch
4014 for _, a := range &sys.Archs {
4015 all = append(all, a)
4021 if a.Family != sys.PPC64 {
4022 lwatomics = append(lwatomics, a)
4026 // add adds the intrinsic b for pkg.fn for the given list of architectures.
4027 add := func(pkg, fn string, b intrinsicBuilder, archs ...*sys.Arch) {
4028 for _, a := range archs {
4029 intrinsics[intrinsicKey{a, pkg, fn}] = b
4032 // addF does the same as add but operates on architecture families.
4033 addF := func(pkg, fn string, b intrinsicBuilder, archFamilies ...sys.ArchFamily) {
4035 for _, f := range archFamilies {
4037 panic("too many architecture families")
4041 for _, a := range all {
4042 if m>>uint(a.Family)&1 != 0 {
4043 intrinsics[intrinsicKey{a, pkg, fn}] = b
4047 // alias defines pkg.fn = pkg2.fn2 for all architectures in archs for which pkg2.fn2 exists.
4048 alias := func(pkg, fn, pkg2, fn2 string, archs ...*sys.Arch) {
4050 for _, a := range archs {
4051 if b, ok := intrinsics[intrinsicKey{a, pkg2, fn2}]; ok {
4052 intrinsics[intrinsicKey{a, pkg, fn}] = b
4057 panic(fmt.Sprintf("attempted to alias undefined intrinsic: %s.%s", pkg, fn))
4061 /******** runtime ********/
4062 if !base.Flag.Cfg.Instrumenting {
4063 add("runtime", "slicebytetostringtmp",
4064 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4065 // Compiler frontend optimizations emit OBYTES2STRTMP nodes
4066 // for the backend instead of slicebytetostringtmp calls
4067 // when not instrumenting.
4068 return s.newValue2(ssa.OpStringMake, n.Type(), args[0], args[1])
4072 addF("runtime/internal/math", "MulUintptr",
4073 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4074 if s.config.PtrSize == 4 {
4075 return s.newValue2(ssa.OpMul32uover, types.NewTuple(types.Types[types.TUINT], types.Types[types.TUINT]), args[0], args[1])
4077 return s.newValue2(ssa.OpMul64uover, types.NewTuple(types.Types[types.TUINT], types.Types[types.TUINT]), args[0], args[1])
4079 sys.AMD64, sys.I386, sys.Loong64, sys.MIPS64, sys.RISCV64, sys.ARM64)
4080 alias("runtime", "mulUintptr", "runtime/internal/math", "MulUintptr", all...)
4081 add("runtime", "KeepAlive",
4082 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4083 data := s.newValue1(ssa.OpIData, s.f.Config.Types.BytePtr, args[0])
4084 s.vars[memVar] = s.newValue2(ssa.OpKeepAlive, types.TypeMem, data, s.mem())
4088 add("runtime", "getclosureptr",
4089 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4090 return s.newValue0(ssa.OpGetClosurePtr, s.f.Config.Types.Uintptr)
4094 add("runtime", "getcallerpc",
4095 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4096 return s.newValue0(ssa.OpGetCallerPC, s.f.Config.Types.Uintptr)
4100 add("runtime", "getcallersp",
4101 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4102 return s.newValue1(ssa.OpGetCallerSP, s.f.Config.Types.Uintptr, s.mem())
4106 addF("runtime", "publicationBarrier",
4107 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4108 s.vars[memVar] = s.newValue1(ssa.OpPubBarrier, types.TypeMem, s.mem())
4111 sys.ARM64, sys.PPC64)
4113 brev_arch := []sys.ArchFamily{sys.AMD64, sys.I386, sys.ARM64, sys.ARM, sys.S390X}
4114 if buildcfg.GOPPC64 >= 10 {
4115 // Use only on Power10 as the new byte reverse instructions that Power10 provide
4116 // make it worthwhile as an intrinsic
4117 brev_arch = append(brev_arch, sys.PPC64)
4119 /******** runtime/internal/sys ********/
4120 addF("runtime/internal/sys", "Bswap32",
4121 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4122 return s.newValue1(ssa.OpBswap32, types.Types[types.TUINT32], args[0])
4125 addF("runtime/internal/sys", "Bswap64",
4126 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4127 return s.newValue1(ssa.OpBswap64, types.Types[types.TUINT64], args[0])
4131 /****** Prefetch ******/
4132 makePrefetchFunc := func(op ssa.Op) func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4133 return func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4134 s.vars[memVar] = s.newValue2(op, types.TypeMem, args[0], s.mem())
4139 // Make Prefetch intrinsics for supported platforms
4140 // On the unsupported platforms stub function will be eliminated
4141 addF("runtime/internal/sys", "Prefetch", makePrefetchFunc(ssa.OpPrefetchCache),
4142 sys.AMD64, sys.ARM64, sys.PPC64)
4143 addF("runtime/internal/sys", "PrefetchStreamed", makePrefetchFunc(ssa.OpPrefetchCacheStreamed),
4144 sys.AMD64, sys.ARM64, sys.PPC64)
4146 /******** runtime/internal/atomic ********/
4147 addF("runtime/internal/atomic", "Load",
4148 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4149 v := s.newValue2(ssa.OpAtomicLoad32, types.NewTuple(types.Types[types.TUINT32], types.TypeMem), args[0], s.mem())
4150 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4151 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT32], v)
4153 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4154 addF("runtime/internal/atomic", "Load8",
4155 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4156 v := s.newValue2(ssa.OpAtomicLoad8, types.NewTuple(types.Types[types.TUINT8], types.TypeMem), args[0], s.mem())
4157 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4158 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT8], v)
4160 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4161 addF("runtime/internal/atomic", "Load64",
4162 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4163 v := s.newValue2(ssa.OpAtomicLoad64, types.NewTuple(types.Types[types.TUINT64], types.TypeMem), args[0], s.mem())
4164 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4165 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT64], v)
4167 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4168 addF("runtime/internal/atomic", "LoadAcq",
4169 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4170 v := s.newValue2(ssa.OpAtomicLoadAcq32, types.NewTuple(types.Types[types.TUINT32], types.TypeMem), args[0], s.mem())
4171 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4172 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT32], v)
4174 sys.PPC64, sys.S390X)
4175 addF("runtime/internal/atomic", "LoadAcq64",
4176 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4177 v := s.newValue2(ssa.OpAtomicLoadAcq64, types.NewTuple(types.Types[types.TUINT64], types.TypeMem), args[0], s.mem())
4178 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4179 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT64], v)
4182 addF("runtime/internal/atomic", "Loadp",
4183 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4184 v := s.newValue2(ssa.OpAtomicLoadPtr, types.NewTuple(s.f.Config.Types.BytePtr, types.TypeMem), args[0], s.mem())
4185 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4186 return s.newValue1(ssa.OpSelect0, s.f.Config.Types.BytePtr, v)
4188 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4190 addF("runtime/internal/atomic", "Store",
4191 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4192 s.vars[memVar] = s.newValue3(ssa.OpAtomicStore32, types.TypeMem, args[0], args[1], s.mem())
4195 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4196 addF("runtime/internal/atomic", "Store8",
4197 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4198 s.vars[memVar] = s.newValue3(ssa.OpAtomicStore8, types.TypeMem, args[0], args[1], s.mem())
4201 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4202 addF("runtime/internal/atomic", "Store64",
4203 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4204 s.vars[memVar] = s.newValue3(ssa.OpAtomicStore64, types.TypeMem, args[0], args[1], s.mem())
4207 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4208 addF("runtime/internal/atomic", "StorepNoWB",
4209 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4210 s.vars[memVar] = s.newValue3(ssa.OpAtomicStorePtrNoWB, types.TypeMem, args[0], args[1], s.mem())
4213 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.RISCV64, sys.S390X)
4214 addF("runtime/internal/atomic", "StoreRel",
4215 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4216 s.vars[memVar] = s.newValue3(ssa.OpAtomicStoreRel32, types.TypeMem, args[0], args[1], s.mem())
4219 sys.PPC64, sys.S390X)
4220 addF("runtime/internal/atomic", "StoreRel64",
4221 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4222 s.vars[memVar] = s.newValue3(ssa.OpAtomicStoreRel64, types.TypeMem, args[0], args[1], s.mem())
4227 addF("runtime/internal/atomic", "Xchg",
4228 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4229 v := s.newValue3(ssa.OpAtomicExchange32, types.NewTuple(types.Types[types.TUINT32], types.TypeMem), args[0], args[1], s.mem())
4230 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4231 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT32], v)
4233 sys.AMD64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4234 addF("runtime/internal/atomic", "Xchg64",
4235 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4236 v := s.newValue3(ssa.OpAtomicExchange64, types.NewTuple(types.Types[types.TUINT64], types.TypeMem), args[0], args[1], s.mem())
4237 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4238 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT64], v)
4240 sys.AMD64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4242 type atomicOpEmitter func(s *state, n *ir.CallExpr, args []*ssa.Value, op ssa.Op, typ types.Kind)
4244 makeAtomicGuardedIntrinsicARM64 := func(op0, op1 ssa.Op, typ, rtyp types.Kind, emit atomicOpEmitter) intrinsicBuilder {
4246 return func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4247 // Target Atomic feature is identified by dynamic detection
4248 addr := s.entryNewValue1A(ssa.OpAddr, types.Types[types.TBOOL].PtrTo(), ir.Syms.ARM64HasATOMICS, s.sb)
4249 v := s.load(types.Types[types.TBOOL], addr)
4251 b.Kind = ssa.BlockIf
4253 bTrue := s.f.NewBlock(ssa.BlockPlain)
4254 bFalse := s.f.NewBlock(ssa.BlockPlain)
4255 bEnd := s.f.NewBlock(ssa.BlockPlain)
4258 b.Likely = ssa.BranchLikely
4260 // We have atomic instructions - use it directly.
4262 emit(s, n, args, op1, typ)
4263 s.endBlock().AddEdgeTo(bEnd)
4265 // Use original instruction sequence.
4266 s.startBlock(bFalse)
4267 emit(s, n, args, op0, typ)
4268 s.endBlock().AddEdgeTo(bEnd)
4272 if rtyp == types.TNIL {
4275 return s.variable(n, types.Types[rtyp])
4280 atomicXchgXaddEmitterARM64 := func(s *state, n *ir.CallExpr, args []*ssa.Value, op ssa.Op, typ types.Kind) {
4281 v := s.newValue3(op, types.NewTuple(types.Types[typ], types.TypeMem), args[0], args[1], s.mem())
4282 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4283 s.vars[n] = s.newValue1(ssa.OpSelect0, types.Types[typ], v)
4285 addF("runtime/internal/atomic", "Xchg",
4286 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicExchange32, ssa.OpAtomicExchange32Variant, types.TUINT32, types.TUINT32, atomicXchgXaddEmitterARM64),
4288 addF("runtime/internal/atomic", "Xchg64",
4289 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicExchange64, ssa.OpAtomicExchange64Variant, types.TUINT64, types.TUINT64, atomicXchgXaddEmitterARM64),
4292 addF("runtime/internal/atomic", "Xadd",
4293 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4294 v := s.newValue3(ssa.OpAtomicAdd32, types.NewTuple(types.Types[types.TUINT32], types.TypeMem), args[0], args[1], s.mem())
4295 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4296 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT32], v)
4298 sys.AMD64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4299 addF("runtime/internal/atomic", "Xadd64",
4300 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4301 v := s.newValue3(ssa.OpAtomicAdd64, types.NewTuple(types.Types[types.TUINT64], types.TypeMem), args[0], args[1], s.mem())
4302 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4303 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT64], v)
4305 sys.AMD64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4307 addF("runtime/internal/atomic", "Xadd",
4308 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAdd32, ssa.OpAtomicAdd32Variant, types.TUINT32, types.TUINT32, atomicXchgXaddEmitterARM64),
4310 addF("runtime/internal/atomic", "Xadd64",
4311 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAdd64, ssa.OpAtomicAdd64Variant, types.TUINT64, types.TUINT64, atomicXchgXaddEmitterARM64),
4314 addF("runtime/internal/atomic", "Cas",
4315 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4316 v := s.newValue4(ssa.OpAtomicCompareAndSwap32, types.NewTuple(types.Types[types.TBOOL], types.TypeMem), args[0], args[1], args[2], s.mem())
4317 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4318 return s.newValue1(ssa.OpSelect0, types.Types[types.TBOOL], v)
4320 sys.AMD64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4321 addF("runtime/internal/atomic", "Cas64",
4322 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4323 v := s.newValue4(ssa.OpAtomicCompareAndSwap64, types.NewTuple(types.Types[types.TBOOL], types.TypeMem), args[0], args[1], args[2], s.mem())
4324 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4325 return s.newValue1(ssa.OpSelect0, types.Types[types.TBOOL], v)
4327 sys.AMD64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4328 addF("runtime/internal/atomic", "CasRel",
4329 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4330 v := s.newValue4(ssa.OpAtomicCompareAndSwap32, types.NewTuple(types.Types[types.TBOOL], types.TypeMem), args[0], args[1], args[2], s.mem())
4331 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4332 return s.newValue1(ssa.OpSelect0, types.Types[types.TBOOL], v)
4336 atomicCasEmitterARM64 := func(s *state, n *ir.CallExpr, args []*ssa.Value, op ssa.Op, typ types.Kind) {
4337 v := s.newValue4(op, types.NewTuple(types.Types[types.TBOOL], types.TypeMem), args[0], args[1], args[2], s.mem())
4338 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4339 s.vars[n] = s.newValue1(ssa.OpSelect0, types.Types[typ], v)
4342 addF("runtime/internal/atomic", "Cas",
4343 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicCompareAndSwap32, ssa.OpAtomicCompareAndSwap32Variant, types.TUINT32, types.TBOOL, atomicCasEmitterARM64),
4345 addF("runtime/internal/atomic", "Cas64",
4346 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicCompareAndSwap64, ssa.OpAtomicCompareAndSwap64Variant, types.TUINT64, types.TBOOL, atomicCasEmitterARM64),
4349 addF("runtime/internal/atomic", "And8",
4350 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4351 s.vars[memVar] = s.newValue3(ssa.OpAtomicAnd8, types.TypeMem, args[0], args[1], s.mem())
4354 sys.AMD64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4355 addF("runtime/internal/atomic", "And",
4356 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4357 s.vars[memVar] = s.newValue3(ssa.OpAtomicAnd32, types.TypeMem, args[0], args[1], s.mem())
4360 sys.AMD64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4361 addF("runtime/internal/atomic", "Or8",
4362 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4363 s.vars[memVar] = s.newValue3(ssa.OpAtomicOr8, types.TypeMem, args[0], args[1], s.mem())
4366 sys.AMD64, sys.ARM64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4367 addF("runtime/internal/atomic", "Or",
4368 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4369 s.vars[memVar] = s.newValue3(ssa.OpAtomicOr32, types.TypeMem, args[0], args[1], s.mem())
4372 sys.AMD64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4374 atomicAndOrEmitterARM64 := func(s *state, n *ir.CallExpr, args []*ssa.Value, op ssa.Op, typ types.Kind) {
4375 s.vars[memVar] = s.newValue3(op, types.TypeMem, args[0], args[1], s.mem())
4378 addF("runtime/internal/atomic", "And8",
4379 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAnd8, ssa.OpAtomicAnd8Variant, types.TNIL, types.TNIL, atomicAndOrEmitterARM64),
4381 addF("runtime/internal/atomic", "And",
4382 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAnd32, ssa.OpAtomicAnd32Variant, types.TNIL, types.TNIL, atomicAndOrEmitterARM64),
4384 addF("runtime/internal/atomic", "Or8",
4385 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicOr8, ssa.OpAtomicOr8Variant, types.TNIL, types.TNIL, atomicAndOrEmitterARM64),
4387 addF("runtime/internal/atomic", "Or",
4388 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicOr32, ssa.OpAtomicOr32Variant, types.TNIL, types.TNIL, atomicAndOrEmitterARM64),
4391 // Aliases for atomic load operations
4392 alias("runtime/internal/atomic", "Loadint32", "runtime/internal/atomic", "Load", all...)
4393 alias("runtime/internal/atomic", "Loadint64", "runtime/internal/atomic", "Load64", all...)
4394 alias("runtime/internal/atomic", "Loaduintptr", "runtime/internal/atomic", "Load", p4...)
4395 alias("runtime/internal/atomic", "Loaduintptr", "runtime/internal/atomic", "Load64", p8...)
4396 alias("runtime/internal/atomic", "Loaduint", "runtime/internal/atomic", "Load", p4...)
4397 alias("runtime/internal/atomic", "Loaduint", "runtime/internal/atomic", "Load64", p8...)
4398 alias("runtime/internal/atomic", "LoadAcq", "runtime/internal/atomic", "Load", lwatomics...)
4399 alias("runtime/internal/atomic", "LoadAcq64", "runtime/internal/atomic", "Load64", lwatomics...)
4400 alias("runtime/internal/atomic", "LoadAcquintptr", "runtime/internal/atomic", "LoadAcq", p4...)
4401 alias("sync", "runtime_LoadAcquintptr", "runtime/internal/atomic", "LoadAcq", p4...) // linknamed
4402 alias("runtime/internal/atomic", "LoadAcquintptr", "runtime/internal/atomic", "LoadAcq64", p8...)
4403 alias("sync", "runtime_LoadAcquintptr", "runtime/internal/atomic", "LoadAcq64", p8...) // linknamed
4405 // Aliases for atomic store operations
4406 alias("runtime/internal/atomic", "Storeint32", "runtime/internal/atomic", "Store", all...)
4407 alias("runtime/internal/atomic", "Storeint64", "runtime/internal/atomic", "Store64", all...)
4408 alias("runtime/internal/atomic", "Storeuintptr", "runtime/internal/atomic", "Store", p4...)
4409 alias("runtime/internal/atomic", "Storeuintptr", "runtime/internal/atomic", "Store64", p8...)
4410 alias("runtime/internal/atomic", "StoreRel", "runtime/internal/atomic", "Store", lwatomics...)
4411 alias("runtime/internal/atomic", "StoreRel64", "runtime/internal/atomic", "Store64", lwatomics...)
4412 alias("runtime/internal/atomic", "StoreReluintptr", "runtime/internal/atomic", "StoreRel", p4...)
4413 alias("sync", "runtime_StoreReluintptr", "runtime/internal/atomic", "StoreRel", p4...) // linknamed
4414 alias("runtime/internal/atomic", "StoreReluintptr", "runtime/internal/atomic", "StoreRel64", p8...)
4415 alias("sync", "runtime_StoreReluintptr", "runtime/internal/atomic", "StoreRel64", p8...) // linknamed
4417 // Aliases for atomic swap operations
4418 alias("runtime/internal/atomic", "Xchgint32", "runtime/internal/atomic", "Xchg", all...)
4419 alias("runtime/internal/atomic", "Xchgint64", "runtime/internal/atomic", "Xchg64", all...)
4420 alias("runtime/internal/atomic", "Xchguintptr", "runtime/internal/atomic", "Xchg", p4...)
4421 alias("runtime/internal/atomic", "Xchguintptr", "runtime/internal/atomic", "Xchg64", p8...)
4423 // Aliases for atomic add operations
4424 alias("runtime/internal/atomic", "Xaddint32", "runtime/internal/atomic", "Xadd", all...)
4425 alias("runtime/internal/atomic", "Xaddint64", "runtime/internal/atomic", "Xadd64", all...)
4426 alias("runtime/internal/atomic", "Xadduintptr", "runtime/internal/atomic", "Xadd", p4...)
4427 alias("runtime/internal/atomic", "Xadduintptr", "runtime/internal/atomic", "Xadd64", p8...)
4429 // Aliases for atomic CAS operations
4430 alias("runtime/internal/atomic", "Casint32", "runtime/internal/atomic", "Cas", all...)
4431 alias("runtime/internal/atomic", "Casint64", "runtime/internal/atomic", "Cas64", all...)
4432 alias("runtime/internal/atomic", "Casuintptr", "runtime/internal/atomic", "Cas", p4...)
4433 alias("runtime/internal/atomic", "Casuintptr", "runtime/internal/atomic", "Cas64", p8...)
4434 alias("runtime/internal/atomic", "Casp1", "runtime/internal/atomic", "Cas", p4...)
4435 alias("runtime/internal/atomic", "Casp1", "runtime/internal/atomic", "Cas64", p8...)
4436 alias("runtime/internal/atomic", "CasRel", "runtime/internal/atomic", "Cas", lwatomics...)
4438 /******** math ********/
4439 addF("math", "sqrt",
4440 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4441 return s.newValue1(ssa.OpSqrt, types.Types[types.TFLOAT64], args[0])
4443 sys.I386, sys.AMD64, sys.ARM, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X, sys.Wasm)
4444 addF("math", "Trunc",
4445 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4446 return s.newValue1(ssa.OpTrunc, types.Types[types.TFLOAT64], args[0])
4448 sys.ARM64, sys.PPC64, sys.S390X, sys.Wasm)
4449 addF("math", "Ceil",
4450 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4451 return s.newValue1(ssa.OpCeil, types.Types[types.TFLOAT64], args[0])
4453 sys.ARM64, sys.PPC64, sys.S390X, sys.Wasm)
4454 addF("math", "Floor",
4455 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4456 return s.newValue1(ssa.OpFloor, types.Types[types.TFLOAT64], args[0])
4458 sys.ARM64, sys.PPC64, sys.S390X, sys.Wasm)
4459 addF("math", "Round",
4460 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4461 return s.newValue1(ssa.OpRound, types.Types[types.TFLOAT64], args[0])
4463 sys.ARM64, sys.PPC64, sys.S390X)
4464 addF("math", "RoundToEven",
4465 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4466 return s.newValue1(ssa.OpRoundToEven, types.Types[types.TFLOAT64], args[0])
4468 sys.ARM64, sys.S390X, sys.Wasm)
4470 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4471 return s.newValue1(ssa.OpAbs, types.Types[types.TFLOAT64], args[0])
4473 sys.ARM64, sys.ARM, sys.PPC64, sys.RISCV64, sys.Wasm, sys.MIPS, sys.MIPS64)
4474 addF("math", "Copysign",
4475 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4476 return s.newValue2(ssa.OpCopysign, types.Types[types.TFLOAT64], args[0], args[1])
4478 sys.PPC64, sys.RISCV64, sys.Wasm)
4480 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4481 return s.newValue3(ssa.OpFMA, types.Types[types.TFLOAT64], args[0], args[1], args[2])
4483 sys.ARM64, sys.PPC64, sys.RISCV64, sys.S390X)
4485 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4486 if !s.config.UseFMA {
4487 s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
4488 return s.variable(n, types.Types[types.TFLOAT64])
4491 if buildcfg.GOAMD64 >= 3 {
4492 return s.newValue3(ssa.OpFMA, types.Types[types.TFLOAT64], args[0], args[1], args[2])
4495 v := s.entryNewValue0A(ssa.OpHasCPUFeature, types.Types[types.TBOOL], ir.Syms.X86HasFMA)
4497 b.Kind = ssa.BlockIf
4499 bTrue := s.f.NewBlock(ssa.BlockPlain)
4500 bFalse := s.f.NewBlock(ssa.BlockPlain)
4501 bEnd := s.f.NewBlock(ssa.BlockPlain)
4504 b.Likely = ssa.BranchLikely // >= haswell cpus are common
4506 // We have the intrinsic - use it directly.
4508 s.vars[n] = s.newValue3(ssa.OpFMA, types.Types[types.TFLOAT64], args[0], args[1], args[2])
4509 s.endBlock().AddEdgeTo(bEnd)
4511 // Call the pure Go version.
4512 s.startBlock(bFalse)
4513 s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
4514 s.endBlock().AddEdgeTo(bEnd)
4518 return s.variable(n, types.Types[types.TFLOAT64])
4522 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4523 if !s.config.UseFMA {
4524 s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
4525 return s.variable(n, types.Types[types.TFLOAT64])
4527 addr := s.entryNewValue1A(ssa.OpAddr, types.Types[types.TBOOL].PtrTo(), ir.Syms.ARMHasVFPv4, s.sb)
4528 v := s.load(types.Types[types.TBOOL], addr)
4530 b.Kind = ssa.BlockIf
4532 bTrue := s.f.NewBlock(ssa.BlockPlain)
4533 bFalse := s.f.NewBlock(ssa.BlockPlain)
4534 bEnd := s.f.NewBlock(ssa.BlockPlain)
4537 b.Likely = ssa.BranchLikely
4539 // We have the intrinsic - use it directly.
4541 s.vars[n] = s.newValue3(ssa.OpFMA, types.Types[types.TFLOAT64], args[0], args[1], args[2])
4542 s.endBlock().AddEdgeTo(bEnd)
4544 // Call the pure Go version.
4545 s.startBlock(bFalse)
4546 s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
4547 s.endBlock().AddEdgeTo(bEnd)
4551 return s.variable(n, types.Types[types.TFLOAT64])
4555 makeRoundAMD64 := func(op ssa.Op) func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4556 return func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4557 if buildcfg.GOAMD64 >= 2 {
4558 return s.newValue1(op, types.Types[types.TFLOAT64], args[0])
4561 v := s.entryNewValue0A(ssa.OpHasCPUFeature, types.Types[types.TBOOL], ir.Syms.X86HasSSE41)
4563 b.Kind = ssa.BlockIf
4565 bTrue := s.f.NewBlock(ssa.BlockPlain)
4566 bFalse := s.f.NewBlock(ssa.BlockPlain)
4567 bEnd := s.f.NewBlock(ssa.BlockPlain)
4570 b.Likely = ssa.BranchLikely // most machines have sse4.1 nowadays
4572 // We have the intrinsic - use it directly.
4574 s.vars[n] = s.newValue1(op, types.Types[types.TFLOAT64], args[0])
4575 s.endBlock().AddEdgeTo(bEnd)
4577 // Call the pure Go version.
4578 s.startBlock(bFalse)
4579 s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
4580 s.endBlock().AddEdgeTo(bEnd)
4584 return s.variable(n, types.Types[types.TFLOAT64])
4587 addF("math", "RoundToEven",
4588 makeRoundAMD64(ssa.OpRoundToEven),
4590 addF("math", "Floor",
4591 makeRoundAMD64(ssa.OpFloor),
4593 addF("math", "Ceil",
4594 makeRoundAMD64(ssa.OpCeil),
4596 addF("math", "Trunc",
4597 makeRoundAMD64(ssa.OpTrunc),
4600 /******** math/bits ********/
4601 addF("math/bits", "TrailingZeros64",
4602 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4603 return s.newValue1(ssa.OpCtz64, types.Types[types.TINT], args[0])
4605 sys.AMD64, sys.I386, sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4606 addF("math/bits", "TrailingZeros32",
4607 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4608 return s.newValue1(ssa.OpCtz32, types.Types[types.TINT], args[0])
4610 sys.AMD64, sys.I386, sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4611 addF("math/bits", "TrailingZeros16",
4612 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4613 x := s.newValue1(ssa.OpZeroExt16to32, types.Types[types.TUINT32], args[0])
4614 c := s.constInt32(types.Types[types.TUINT32], 1<<16)
4615 y := s.newValue2(ssa.OpOr32, types.Types[types.TUINT32], x, c)
4616 return s.newValue1(ssa.OpCtz32, types.Types[types.TINT], y)
4619 addF("math/bits", "TrailingZeros16",
4620 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4621 return s.newValue1(ssa.OpCtz16, types.Types[types.TINT], args[0])
4623 sys.AMD64, sys.I386, sys.ARM, sys.ARM64, sys.Wasm)
4624 addF("math/bits", "TrailingZeros16",
4625 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4626 x := s.newValue1(ssa.OpZeroExt16to64, types.Types[types.TUINT64], args[0])
4627 c := s.constInt64(types.Types[types.TUINT64], 1<<16)
4628 y := s.newValue2(ssa.OpOr64, types.Types[types.TUINT64], x, c)
4629 return s.newValue1(ssa.OpCtz64, types.Types[types.TINT], y)
4631 sys.S390X, sys.PPC64)
4632 addF("math/bits", "TrailingZeros8",
4633 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4634 x := s.newValue1(ssa.OpZeroExt8to32, types.Types[types.TUINT32], args[0])
4635 c := s.constInt32(types.Types[types.TUINT32], 1<<8)
4636 y := s.newValue2(ssa.OpOr32, types.Types[types.TUINT32], x, c)
4637 return s.newValue1(ssa.OpCtz32, types.Types[types.TINT], y)
4640 addF("math/bits", "TrailingZeros8",
4641 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4642 return s.newValue1(ssa.OpCtz8, types.Types[types.TINT], args[0])
4644 sys.AMD64, sys.I386, sys.ARM, sys.ARM64, sys.Wasm)
4645 addF("math/bits", "TrailingZeros8",
4646 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4647 x := s.newValue1(ssa.OpZeroExt8to64, types.Types[types.TUINT64], args[0])
4648 c := s.constInt64(types.Types[types.TUINT64], 1<<8)
4649 y := s.newValue2(ssa.OpOr64, types.Types[types.TUINT64], x, c)
4650 return s.newValue1(ssa.OpCtz64, types.Types[types.TINT], y)
4653 alias("math/bits", "ReverseBytes64", "runtime/internal/sys", "Bswap64", all...)
4654 alias("math/bits", "ReverseBytes32", "runtime/internal/sys", "Bswap32", all...)
4655 // ReverseBytes inlines correctly, no need to intrinsify it.
4656 // Nothing special is needed for targets where ReverseBytes16 lowers to a rotate
4657 // On Power10, 16-bit rotate is not available so use BRH instruction
4658 if buildcfg.GOPPC64 >= 10 {
4659 addF("math/bits", "ReverseBytes16",
4660 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4661 return s.newValue1(ssa.OpBswap16, types.Types[types.TUINT], args[0])
4666 addF("math/bits", "Len64",
4667 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4668 return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], args[0])
4670 sys.AMD64, sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4671 addF("math/bits", "Len32",
4672 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4673 return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], args[0])
4675 sys.AMD64, sys.ARM64, sys.PPC64)
4676 addF("math/bits", "Len32",
4677 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4678 if s.config.PtrSize == 4 {
4679 return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], args[0])
4681 x := s.newValue1(ssa.OpZeroExt32to64, types.Types[types.TUINT64], args[0])
4682 return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], x)
4684 sys.ARM, sys.S390X, sys.MIPS, sys.Wasm)
4685 addF("math/bits", "Len16",
4686 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4687 if s.config.PtrSize == 4 {
4688 x := s.newValue1(ssa.OpZeroExt16to32, types.Types[types.TUINT32], args[0])
4689 return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], x)
4691 x := s.newValue1(ssa.OpZeroExt16to64, types.Types[types.TUINT64], args[0])
4692 return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], x)
4694 sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4695 addF("math/bits", "Len16",
4696 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4697 return s.newValue1(ssa.OpBitLen16, types.Types[types.TINT], args[0])
4700 addF("math/bits", "Len8",
4701 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4702 if s.config.PtrSize == 4 {
4703 x := s.newValue1(ssa.OpZeroExt8to32, types.Types[types.TUINT32], args[0])
4704 return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], x)
4706 x := s.newValue1(ssa.OpZeroExt8to64, types.Types[types.TUINT64], args[0])
4707 return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], x)
4709 sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4710 addF("math/bits", "Len8",
4711 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4712 return s.newValue1(ssa.OpBitLen8, types.Types[types.TINT], args[0])
4715 addF("math/bits", "Len",
4716 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4717 if s.config.PtrSize == 4 {
4718 return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], args[0])
4720 return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], args[0])
4722 sys.AMD64, sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4723 // LeadingZeros is handled because it trivially calls Len.
4724 addF("math/bits", "Reverse64",
4725 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4726 return s.newValue1(ssa.OpBitRev64, types.Types[types.TINT], args[0])
4729 addF("math/bits", "Reverse32",
4730 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4731 return s.newValue1(ssa.OpBitRev32, types.Types[types.TINT], args[0])
4734 addF("math/bits", "Reverse16",
4735 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4736 return s.newValue1(ssa.OpBitRev16, types.Types[types.TINT], args[0])
4739 addF("math/bits", "Reverse8",
4740 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4741 return s.newValue1(ssa.OpBitRev8, types.Types[types.TINT], args[0])
4744 addF("math/bits", "Reverse",
4745 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4746 return s.newValue1(ssa.OpBitRev64, types.Types[types.TINT], args[0])
4749 addF("math/bits", "RotateLeft8",
4750 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4751 return s.newValue2(ssa.OpRotateLeft8, types.Types[types.TUINT8], args[0], args[1])
4754 addF("math/bits", "RotateLeft16",
4755 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4756 return s.newValue2(ssa.OpRotateLeft16, types.Types[types.TUINT16], args[0], args[1])
4759 addF("math/bits", "RotateLeft32",
4760 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4761 return s.newValue2(ssa.OpRotateLeft32, types.Types[types.TUINT32], args[0], args[1])
4763 sys.AMD64, sys.ARM, sys.ARM64, sys.S390X, sys.PPC64, sys.Wasm, sys.Loong64)
4764 addF("math/bits", "RotateLeft64",
4765 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4766 return s.newValue2(ssa.OpRotateLeft64, types.Types[types.TUINT64], args[0], args[1])
4768 sys.AMD64, sys.ARM64, sys.S390X, sys.PPC64, sys.Wasm, sys.Loong64)
4769 alias("math/bits", "RotateLeft", "math/bits", "RotateLeft64", p8...)
4771 makeOnesCountAMD64 := func(op ssa.Op) func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4772 return func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4773 if buildcfg.GOAMD64 >= 2 {
4774 return s.newValue1(op, types.Types[types.TINT], args[0])
4777 v := s.entryNewValue0A(ssa.OpHasCPUFeature, types.Types[types.TBOOL], ir.Syms.X86HasPOPCNT)
4779 b.Kind = ssa.BlockIf
4781 bTrue := s.f.NewBlock(ssa.BlockPlain)
4782 bFalse := s.f.NewBlock(ssa.BlockPlain)
4783 bEnd := s.f.NewBlock(ssa.BlockPlain)
4786 b.Likely = ssa.BranchLikely // most machines have popcnt nowadays
4788 // We have the intrinsic - use it directly.
4790 s.vars[n] = s.newValue1(op, types.Types[types.TINT], args[0])
4791 s.endBlock().AddEdgeTo(bEnd)
4793 // Call the pure Go version.
4794 s.startBlock(bFalse)
4795 s.vars[n] = s.callResult(n, callNormal) // types.Types[TINT]
4796 s.endBlock().AddEdgeTo(bEnd)
4800 return s.variable(n, types.Types[types.TINT])
4803 addF("math/bits", "OnesCount64",
4804 makeOnesCountAMD64(ssa.OpPopCount64),
4806 addF("math/bits", "OnesCount64",
4807 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4808 return s.newValue1(ssa.OpPopCount64, types.Types[types.TINT], args[0])
4810 sys.PPC64, sys.ARM64, sys.S390X, sys.Wasm)
4811 addF("math/bits", "OnesCount32",
4812 makeOnesCountAMD64(ssa.OpPopCount32),
4814 addF("math/bits", "OnesCount32",
4815 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4816 return s.newValue1(ssa.OpPopCount32, types.Types[types.TINT], args[0])
4818 sys.PPC64, sys.ARM64, sys.S390X, sys.Wasm)
4819 addF("math/bits", "OnesCount16",
4820 makeOnesCountAMD64(ssa.OpPopCount16),
4822 addF("math/bits", "OnesCount16",
4823 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4824 return s.newValue1(ssa.OpPopCount16, types.Types[types.TINT], args[0])
4826 sys.ARM64, sys.S390X, sys.PPC64, sys.Wasm)
4827 addF("math/bits", "OnesCount8",
4828 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4829 return s.newValue1(ssa.OpPopCount8, types.Types[types.TINT], args[0])
4831 sys.S390X, sys.PPC64, sys.Wasm)
4832 addF("math/bits", "OnesCount",
4833 makeOnesCountAMD64(ssa.OpPopCount64),
4835 addF("math/bits", "Mul64",
4836 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4837 return s.newValue2(ssa.OpMul64uhilo, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1])
4839 sys.AMD64, sys.ARM64, sys.PPC64, sys.S390X, sys.MIPS64, sys.RISCV64, sys.Loong64)
4840 alias("math/bits", "Mul", "math/bits", "Mul64", p8...)
4841 alias("runtime/internal/math", "Mul64", "math/bits", "Mul64", p8...)
4842 addF("math/bits", "Add64",
4843 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4844 return s.newValue3(ssa.OpAdd64carry, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1], args[2])
4846 sys.AMD64, sys.ARM64, sys.PPC64, sys.S390X, sys.RISCV64, sys.Loong64, sys.MIPS64)
4847 alias("math/bits", "Add", "math/bits", "Add64", p8...)
4848 addF("math/bits", "Sub64",
4849 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4850 return s.newValue3(ssa.OpSub64borrow, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1], args[2])
4852 sys.AMD64, sys.ARM64, sys.PPC64, sys.S390X, sys.RISCV64, sys.Loong64)
4853 alias("math/bits", "Sub", "math/bits", "Sub64", p8...)
4854 addF("math/bits", "Div64",
4855 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4856 // check for divide-by-zero/overflow and panic with appropriate message
4857 cmpZero := s.newValue2(s.ssaOp(ir.ONE, types.Types[types.TUINT64]), types.Types[types.TBOOL], args[2], s.zeroVal(types.Types[types.TUINT64]))
4858 s.check(cmpZero, ir.Syms.Panicdivide)
4859 cmpOverflow := s.newValue2(s.ssaOp(ir.OLT, types.Types[types.TUINT64]), types.Types[types.TBOOL], args[0], args[2])
4860 s.check(cmpOverflow, ir.Syms.Panicoverflow)
4861 return s.newValue3(ssa.OpDiv128u, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1], args[2])
4864 alias("math/bits", "Div", "math/bits", "Div64", sys.ArchAMD64)
4866 alias("runtime/internal/sys", "TrailingZeros8", "math/bits", "TrailingZeros8", all...)
4867 alias("runtime/internal/sys", "TrailingZeros32", "math/bits", "TrailingZeros32", all...)
4868 alias("runtime/internal/sys", "TrailingZeros64", "math/bits", "TrailingZeros64", all...)
4869 alias("runtime/internal/sys", "Len8", "math/bits", "Len8", all...)
4870 alias("runtime/internal/sys", "Len64", "math/bits", "Len64", all...)
4871 alias("runtime/internal/sys", "OnesCount64", "math/bits", "OnesCount64", all...)
4873 /******** sync/atomic ********/
4875 // Note: these are disabled by flag_race in findIntrinsic below.
4876 alias("sync/atomic", "LoadInt32", "runtime/internal/atomic", "Load", all...)
4877 alias("sync/atomic", "LoadInt64", "runtime/internal/atomic", "Load64", all...)
4878 alias("sync/atomic", "LoadPointer", "runtime/internal/atomic", "Loadp", all...)
4879 alias("sync/atomic", "LoadUint32", "runtime/internal/atomic", "Load", all...)
4880 alias("sync/atomic", "LoadUint64", "runtime/internal/atomic", "Load64", all...)
4881 alias("sync/atomic", "LoadUintptr", "runtime/internal/atomic", "Load", p4...)
4882 alias("sync/atomic", "LoadUintptr", "runtime/internal/atomic", "Load64", p8...)
4884 alias("sync/atomic", "StoreInt32", "runtime/internal/atomic", "Store", all...)
4885 alias("sync/atomic", "StoreInt64", "runtime/internal/atomic", "Store64", all...)
4886 // Note: not StorePointer, that needs a write barrier. Same below for {CompareAnd}Swap.
4887 alias("sync/atomic", "StoreUint32", "runtime/internal/atomic", "Store", all...)
4888 alias("sync/atomic", "StoreUint64", "runtime/internal/atomic", "Store64", all...)
4889 alias("sync/atomic", "StoreUintptr", "runtime/internal/atomic", "Store", p4...)
4890 alias("sync/atomic", "StoreUintptr", "runtime/internal/atomic", "Store64", p8...)
4892 alias("sync/atomic", "SwapInt32", "runtime/internal/atomic", "Xchg", all...)
4893 alias("sync/atomic", "SwapInt64", "runtime/internal/atomic", "Xchg64", all...)
4894 alias("sync/atomic", "SwapUint32", "runtime/internal/atomic", "Xchg", all...)
4895 alias("sync/atomic", "SwapUint64", "runtime/internal/atomic", "Xchg64", all...)
4896 alias("sync/atomic", "SwapUintptr", "runtime/internal/atomic", "Xchg", p4...)
4897 alias("sync/atomic", "SwapUintptr", "runtime/internal/atomic", "Xchg64", p8...)
4899 alias("sync/atomic", "CompareAndSwapInt32", "runtime/internal/atomic", "Cas", all...)
4900 alias("sync/atomic", "CompareAndSwapInt64", "runtime/internal/atomic", "Cas64", all...)
4901 alias("sync/atomic", "CompareAndSwapUint32", "runtime/internal/atomic", "Cas", all...)
4902 alias("sync/atomic", "CompareAndSwapUint64", "runtime/internal/atomic", "Cas64", all...)
4903 alias("sync/atomic", "CompareAndSwapUintptr", "runtime/internal/atomic", "Cas", p4...)
4904 alias("sync/atomic", "CompareAndSwapUintptr", "runtime/internal/atomic", "Cas64", p8...)
4906 alias("sync/atomic", "AddInt32", "runtime/internal/atomic", "Xadd", all...)
4907 alias("sync/atomic", "AddInt64", "runtime/internal/atomic", "Xadd64", all...)
4908 alias("sync/atomic", "AddUint32", "runtime/internal/atomic", "Xadd", all...)
4909 alias("sync/atomic", "AddUint64", "runtime/internal/atomic", "Xadd64", all...)
4910 alias("sync/atomic", "AddUintptr", "runtime/internal/atomic", "Xadd", p4...)
4911 alias("sync/atomic", "AddUintptr", "runtime/internal/atomic", "Xadd64", p8...)
4913 /******** math/big ********/
4914 alias("math/big", "mulWW", "math/bits", "Mul64", p8...)
4917 // findIntrinsic returns a function which builds the SSA equivalent of the
4918 // function identified by the symbol sym. If sym is not an intrinsic call, returns nil.
4919 func findIntrinsic(sym *types.Sym) intrinsicBuilder {
4920 if sym == nil || sym.Pkg == nil {
4924 if sym.Pkg == ir.Pkgs.Runtime {
4927 if base.Flag.Race && pkg == "sync/atomic" {
4928 // The race detector needs to be able to intercept these calls.
4929 // We can't intrinsify them.
4932 // Skip intrinsifying math functions (which may contain hard-float
4933 // instructions) when soft-float
4934 if Arch.SoftFloat && pkg == "math" {
4939 if ssa.IntrinsicsDisable {
4940 if pkg == "runtime" && (fn == "getcallerpc" || fn == "getcallersp" || fn == "getclosureptr") {
4941 // These runtime functions don't have definitions, must be intrinsics.
4946 return intrinsics[intrinsicKey{Arch.LinkArch.Arch, pkg, fn}]
4949 func IsIntrinsicCall(n *ir.CallExpr) bool {
4953 name, ok := n.X.(*ir.Name)
4957 return findIntrinsic(name.Sym()) != nil
4960 // intrinsicCall converts a call to a recognized intrinsic function into the intrinsic SSA operation.
4961 func (s *state) intrinsicCall(n *ir.CallExpr) *ssa.Value {
4962 v := findIntrinsic(n.X.Sym())(s, n, s.intrinsicArgs(n))
4963 if ssa.IntrinsicsDebug > 0 {
4968 if x.Op == ssa.OpSelect0 || x.Op == ssa.OpSelect1 {
4971 base.WarnfAt(n.Pos(), "intrinsic substitution for %v with %s", n.X.Sym().Name, x.LongString())
4976 // intrinsicArgs extracts args from n, evaluates them to SSA values, and returns them.
4977 func (s *state) intrinsicArgs(n *ir.CallExpr) []*ssa.Value {
4978 args := make([]*ssa.Value, len(n.Args))
4979 for i, n := range n.Args {
4985 // openDeferRecord adds code to evaluate and store the function for an open-code defer
4986 // call, and records info about the defer, so we can generate proper code on the
4987 // exit paths. n is the sub-node of the defer node that is the actual function
4988 // call. We will also record funcdata information on where the function is stored
4989 // (as well as the deferBits variable), and this will enable us to run the proper
4990 // defer calls during panics.
4991 func (s *state) openDeferRecord(n *ir.CallExpr) {
4992 if len(n.Args) != 0 || n.Op() != ir.OCALLFUNC || n.X.Type().NumResults() != 0 {
4993 s.Fatalf("defer call with arguments or results: %v", n)
4996 opendefer := &openDeferInfo{
5000 // We must always store the function value in a stack slot for the
5001 // runtime panic code to use. But in the defer exit code, we will
5002 // call the function directly if it is a static function.
5003 closureVal := s.expr(fn)
5004 closure := s.openDeferSave(fn.Type(), closureVal)
5005 opendefer.closureNode = closure.Aux.(*ir.Name)
5006 if !(fn.Op() == ir.ONAME && fn.(*ir.Name).Class == ir.PFUNC) {
5007 opendefer.closure = closure
5009 index := len(s.openDefers)
5010 s.openDefers = append(s.openDefers, opendefer)
5012 // Update deferBits only after evaluation and storage to stack of
5013 // the function is successful.
5014 bitvalue := s.constInt8(types.Types[types.TUINT8], 1<<uint(index))
5015 newDeferBits := s.newValue2(ssa.OpOr8, types.Types[types.TUINT8], s.variable(deferBitsVar, types.Types[types.TUINT8]), bitvalue)
5016 s.vars[deferBitsVar] = newDeferBits
5017 s.store(types.Types[types.TUINT8], s.deferBitsAddr, newDeferBits)
5020 // openDeferSave generates SSA nodes to store a value (with type t) for an
5021 // open-coded defer at an explicit autotmp location on the stack, so it can be
5022 // reloaded and used for the appropriate call on exit. Type t must be a function type
5023 // (therefore SSAable). val is the value to be stored. The function returns an SSA
5024 // value representing a pointer to the autotmp location.
5025 func (s *state) openDeferSave(t *types.Type, val *ssa.Value) *ssa.Value {
5027 s.Fatalf("openDeferSave of non-SSA-able type %v val=%v", t, val)
5029 if !t.HasPointers() {
5030 s.Fatalf("openDeferSave of pointerless type %v val=%v", t, val)
5033 temp := typecheck.TempAt(pos.WithNotStmt(), s.curfn, t)
5034 temp.SetOpenDeferSlot(true)
5035 var addrTemp *ssa.Value
5036 // Use OpVarLive to make sure stack slot for the closure is not removed by
5037 // dead-store elimination
5038 if s.curBlock.ID != s.f.Entry.ID {
5039 // Force the tmp storing this defer function to be declared in the entry
5040 // block, so that it will be live for the defer exit code (which will
5041 // actually access it only if the associated defer call has been activated).
5042 if t.HasPointers() {
5043 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])
5045 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])
5046 addrTemp = s.f.Entry.NewValue2A(src.NoXPos, ssa.OpLocalAddr, types.NewPtr(temp.Type()), temp, s.sp, s.defvars[s.f.Entry.ID][memVar])
5048 // Special case if we're still in the entry block. We can't use
5049 // the above code, since s.defvars[s.f.Entry.ID] isn't defined
5050 // until we end the entry block with s.endBlock().
5051 if t.HasPointers() {
5052 s.vars[memVar] = s.newValue1Apos(ssa.OpVarDef, types.TypeMem, temp, s.mem(), false)
5054 s.vars[memVar] = s.newValue1Apos(ssa.OpVarLive, types.TypeMem, temp, s.mem(), false)
5055 addrTemp = s.newValue2Apos(ssa.OpLocalAddr, types.NewPtr(temp.Type()), temp, s.sp, s.mem(), false)
5057 // Since we may use this temp during exit depending on the
5058 // deferBits, we must define it unconditionally on entry.
5059 // Therefore, we must make sure it is zeroed out in the entry
5060 // block if it contains pointers, else GC may wrongly follow an
5061 // uninitialized pointer value.
5062 temp.SetNeedzero(true)
5063 // We are storing to the stack, hence we can avoid the full checks in
5064 // storeType() (no write barrier) and do a simple store().
5065 s.store(t, addrTemp, val)
5069 // openDeferExit generates SSA for processing all the open coded defers at exit.
5070 // The code involves loading deferBits, and checking each of the bits to see if
5071 // the corresponding defer statement was executed. For each bit that is turned
5072 // on, the associated defer call is made.
5073 func (s *state) openDeferExit() {
5074 deferExit := s.f.NewBlock(ssa.BlockPlain)
5075 s.endBlock().AddEdgeTo(deferExit)
5076 s.startBlock(deferExit)
5077 s.lastDeferExit = deferExit
5078 s.lastDeferCount = len(s.openDefers)
5079 zeroval := s.constInt8(types.Types[types.TUINT8], 0)
5080 // Test for and run defers in reverse order
5081 for i := len(s.openDefers) - 1; i >= 0; i-- {
5082 r := s.openDefers[i]
5083 bCond := s.f.NewBlock(ssa.BlockPlain)
5084 bEnd := s.f.NewBlock(ssa.BlockPlain)
5086 deferBits := s.variable(deferBitsVar, types.Types[types.TUINT8])
5087 // Generate code to check if the bit associated with the current
5089 bitval := s.constInt8(types.Types[types.TUINT8], 1<<uint(i))
5090 andval := s.newValue2(ssa.OpAnd8, types.Types[types.TUINT8], deferBits, bitval)
5091 eqVal := s.newValue2(ssa.OpEq8, types.Types[types.TBOOL], andval, zeroval)
5093 b.Kind = ssa.BlockIf
5097 bCond.AddEdgeTo(bEnd)
5100 // Clear this bit in deferBits and force store back to stack, so
5101 // we will not try to re-run this defer call if this defer call panics.
5102 nbitval := s.newValue1(ssa.OpCom8, types.Types[types.TUINT8], bitval)
5103 maskedval := s.newValue2(ssa.OpAnd8, types.Types[types.TUINT8], deferBits, nbitval)
5104 s.store(types.Types[types.TUINT8], s.deferBitsAddr, maskedval)
5105 // Use this value for following tests, so we keep previous
5107 s.vars[deferBitsVar] = maskedval
5109 // Generate code to call the function call of the defer, using the
5110 // closure that were stored in argtmps at the point of the defer
5113 stksize := fn.Type().ArgWidth()
5114 var callArgs []*ssa.Value
5116 if r.closure != nil {
5117 v := s.load(r.closure.Type.Elem(), r.closure)
5118 s.maybeNilCheckClosure(v, callDefer)
5119 codeptr := s.rawLoad(types.Types[types.TUINTPTR], v)
5120 aux := ssa.ClosureAuxCall(s.f.ABIDefault.ABIAnalyzeTypes(nil, nil, nil))
5121 call = s.newValue2A(ssa.OpClosureLECall, aux.LateExpansionResultType(), aux, codeptr, v)
5123 aux := ssa.StaticAuxCall(fn.(*ir.Name).Linksym(), s.f.ABIDefault.ABIAnalyzeTypes(nil, nil, nil))
5124 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
5126 callArgs = append(callArgs, s.mem())
5127 call.AddArgs(callArgs...)
5128 call.AuxInt = stksize
5129 s.vars[memVar] = s.newValue1I(ssa.OpSelectN, types.TypeMem, 0, call)
5130 // Make sure that the stack slots with pointers are kept live
5131 // through the call (which is a pre-emption point). Also, we will
5132 // use the first call of the last defer exit to compute liveness
5133 // for the deferreturn, so we want all stack slots to be live.
5134 if r.closureNode != nil {
5135 s.vars[memVar] = s.newValue1Apos(ssa.OpVarLive, types.TypeMem, r.closureNode, s.mem(), false)
5143 func (s *state) callResult(n *ir.CallExpr, k callKind) *ssa.Value {
5144 return s.call(n, k, false)
5147 func (s *state) callAddr(n *ir.CallExpr, k callKind) *ssa.Value {
5148 return s.call(n, k, true)
5151 // Calls the function n using the specified call type.
5152 // Returns the address of the return value (or nil if none).
5153 func (s *state) call(n *ir.CallExpr, k callKind, returnResultAddr bool) *ssa.Value {
5155 var callee *ir.Name // target function (if static)
5156 var closure *ssa.Value // ptr to closure to run (if dynamic)
5157 var codeptr *ssa.Value // ptr to target code (if dynamic)
5158 var rcvr *ssa.Value // receiver to set
5160 var ACArgs []*types.Type // AuxCall args
5161 var ACResults []*types.Type // AuxCall results
5162 var callArgs []*ssa.Value // For late-expansion, the args themselves (not stored, args to the call instead).
5164 callABI := s.f.ABIDefault
5166 if k != callNormal && k != callTail && (len(n.Args) != 0 || n.Op() == ir.OCALLINTER || n.X.Type().NumResults() != 0) {
5167 s.Fatalf("go/defer call with arguments: %v", n)
5172 if (k == callNormal || k == callTail) && fn.Op() == ir.ONAME && fn.(*ir.Name).Class == ir.PFUNC {
5175 if buildcfg.Experiment.RegabiArgs {
5176 // This is a static call, so it may be
5177 // a direct call to a non-ABIInternal
5178 // function. fn.Func may be nil for
5179 // some compiler-generated functions,
5180 // but those are all ABIInternal.
5182 callABI = abiForFunc(fn.Func, s.f.ABI0, s.f.ABI1)
5185 // TODO(register args) remove after register abi is working
5186 inRegistersImported := fn.Pragma()&ir.RegisterParams != 0
5187 inRegistersSamePackage := fn.Func != nil && fn.Func.Pragma&ir.RegisterParams != 0
5188 if inRegistersImported || inRegistersSamePackage {
5194 closure = s.expr(fn)
5195 if k != callDefer && k != callDeferStack {
5196 // Deferred nil function needs to panic when the function is invoked,
5197 // not the point of defer statement.
5198 s.maybeNilCheckClosure(closure, k)
5201 if fn.Op() != ir.ODOTINTER {
5202 s.Fatalf("OCALLINTER: n.Left not an ODOTINTER: %v", fn.Op())
5204 fn := fn.(*ir.SelectorExpr)
5205 var iclosure *ssa.Value
5206 iclosure, rcvr = s.getClosureAndRcvr(fn)
5207 if k == callNormal {
5208 codeptr = s.load(types.Types[types.TUINTPTR], iclosure)
5214 params := callABI.ABIAnalyze(n.X.Type(), false /* Do not set (register) nNames from caller side -- can cause races. */)
5215 types.CalcSize(fn.Type())
5216 stksize := params.ArgWidth() // includes receiver, args, and results
5218 res := n.X.Type().Results()
5219 if k == callNormal || k == callTail {
5220 for _, p := range params.OutParams() {
5221 ACResults = append(ACResults, p.Type)
5226 if k == callDeferStack {
5227 // Make a defer struct d on the stack.
5229 s.Fatalf("deferprocStack with non-zero stack size %d: %v", stksize, n)
5233 d := typecheck.TempAt(n.Pos(), s.curfn, t)
5235 if t.HasPointers() {
5236 s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, d, s.mem())
5240 // Must match deferstruct() below and src/runtime/runtime2.go:_defer.
5241 // 0: started, set in deferprocStack
5242 // 1: heap, set in deferprocStack
5244 // 3: sp, set in deferprocStack
5245 // 4: pc, set in deferprocStack
5247 s.store(closure.Type,
5248 s.newValue1I(ssa.OpOffPtr, closure.Type.PtrTo(), t.FieldOff(5), addr),
5250 // 6: panic, set in deferprocStack
5251 // 7: link, set in deferprocStack
5256 // Call runtime.deferprocStack with pointer to _defer record.
5257 ACArgs = append(ACArgs, types.Types[types.TUINTPTR])
5258 aux := ssa.StaticAuxCall(ir.Syms.DeferprocStack, s.f.ABIDefault.ABIAnalyzeTypes(nil, ACArgs, ACResults))
5259 callArgs = append(callArgs, addr, s.mem())
5260 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
5261 call.AddArgs(callArgs...)
5262 call.AuxInt = int64(types.PtrSize) // deferprocStack takes a *_defer arg
5264 // Store arguments to stack, including defer/go arguments and receiver for method calls.
5265 // These are written in SP-offset order.
5266 argStart := base.Ctxt.Arch.FixedFrameSize
5268 if k != callNormal && k != callTail {
5269 // Write closure (arg to newproc/deferproc).
5270 ACArgs = append(ACArgs, types.Types[types.TUINTPTR]) // not argExtra
5271 callArgs = append(callArgs, closure)
5272 stksize += int64(types.PtrSize)
5273 argStart += int64(types.PtrSize)
5276 // Set receiver (for interface calls).
5278 callArgs = append(callArgs, rcvr)
5285 for _, p := range params.InParams() { // includes receiver for interface calls
5286 ACArgs = append(ACArgs, p.Type)
5289 // Split the entry block if there are open defers, because later calls to
5290 // openDeferSave may cause a mismatch between the mem for an OpDereference
5291 // and the call site which uses it. See #49282.
5292 if s.curBlock.ID == s.f.Entry.ID && s.hasOpenDefers {
5294 b.Kind = ssa.BlockPlain
5295 curb := s.f.NewBlock(ssa.BlockPlain)
5300 for i, n := range args {
5301 callArgs = append(callArgs, s.putArg(n, t.Params().Field(i).Type))
5304 callArgs = append(callArgs, s.mem())
5308 case k == callDefer:
5309 aux := ssa.StaticAuxCall(ir.Syms.Deferproc, s.f.ABIDefault.ABIAnalyzeTypes(nil, ACArgs, ACResults)) // TODO paramResultInfo for DeferProc
5310 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
5312 aux := ssa.StaticAuxCall(ir.Syms.Newproc, s.f.ABIDefault.ABIAnalyzeTypes(nil, ACArgs, ACResults))
5313 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux) // TODO paramResultInfo for NewProc
5314 case closure != nil:
5315 // rawLoad because loading the code pointer from a
5316 // closure is always safe, but IsSanitizerSafeAddr
5317 // can't always figure that out currently, and it's
5318 // critical that we not clobber any arguments already
5319 // stored onto the stack.
5320 codeptr = s.rawLoad(types.Types[types.TUINTPTR], closure)
5321 aux := ssa.ClosureAuxCall(callABI.ABIAnalyzeTypes(nil, ACArgs, ACResults))
5322 call = s.newValue2A(ssa.OpClosureLECall, aux.LateExpansionResultType(), aux, codeptr, closure)
5323 case codeptr != nil:
5324 // Note that the "receiver" parameter is nil because the actual receiver is the first input parameter.
5325 aux := ssa.InterfaceAuxCall(params)
5326 call = s.newValue1A(ssa.OpInterLECall, aux.LateExpansionResultType(), aux, codeptr)
5328 aux := ssa.StaticAuxCall(callTargetLSym(callee), params)
5329 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
5331 call.Op = ssa.OpTailLECall
5332 stksize = 0 // Tail call does not use stack. We reuse caller's frame.
5335 s.Fatalf("bad call type %v %v", n.Op(), n)
5337 call.AddArgs(callArgs...)
5338 call.AuxInt = stksize // Call operations carry the argsize of the callee along with them
5341 s.vars[memVar] = s.newValue1I(ssa.OpSelectN, types.TypeMem, int64(len(ACResults)), call)
5342 // Insert VarLive opcodes.
5343 for _, v := range n.KeepAlive {
5345 s.Fatalf("KeepAlive variable %v must have Addrtaken set", v)
5348 case ir.PAUTO, ir.PPARAM, ir.PPARAMOUT:
5350 s.Fatalf("KeepAlive variable %v must be Auto or Arg", v)
5352 s.vars[memVar] = s.newValue1A(ssa.OpVarLive, types.TypeMem, v, s.mem())
5355 // Finish block for defers
5356 if k == callDefer || k == callDeferStack {
5358 b.Kind = ssa.BlockDefer
5360 bNext := s.f.NewBlock(ssa.BlockPlain)
5362 // Add recover edge to exit code.
5363 r := s.f.NewBlock(ssa.BlockPlain)
5367 b.Likely = ssa.BranchLikely
5371 if res.NumFields() == 0 || k != callNormal {
5372 // call has no return value. Continue with the next statement.
5376 if returnResultAddr {
5377 return s.resultAddrOfCall(call, 0, fp.Type)
5379 return s.newValue1I(ssa.OpSelectN, fp.Type, 0, call)
5382 // maybeNilCheckClosure checks if a nil check of a closure is needed in some
5383 // architecture-dependent situations and, if so, emits the nil check.
5384 func (s *state) maybeNilCheckClosure(closure *ssa.Value, k callKind) {
5385 if Arch.LinkArch.Family == sys.Wasm || buildcfg.GOOS == "aix" && k != callGo {
5386 // 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.
5387 // TODO(neelance): On other architectures this should be eliminated by the optimization steps
5392 // getClosureAndRcvr returns values for the appropriate closure and receiver of an
5394 func (s *state) getClosureAndRcvr(fn *ir.SelectorExpr) (*ssa.Value, *ssa.Value) {
5396 itab := s.newValue1(ssa.OpITab, types.Types[types.TUINTPTR], i)
5398 itabidx := fn.Offset() + 2*int64(types.PtrSize) + 8 // offset of fun field in runtime.itab
5399 closure := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.UintptrPtr, itabidx, itab)
5400 rcvr := s.newValue1(ssa.OpIData, s.f.Config.Types.BytePtr, i)
5401 return closure, rcvr
5404 // etypesign returns the signed-ness of e, for integer/pointer etypes.
5405 // -1 means signed, +1 means unsigned, 0 means non-integer/non-pointer.
5406 func etypesign(e types.Kind) int8 {
5408 case types.TINT8, types.TINT16, types.TINT32, types.TINT64, types.TINT:
5410 case types.TUINT8, types.TUINT16, types.TUINT32, types.TUINT64, types.TUINT, types.TUINTPTR, types.TUNSAFEPTR:
5416 // addr converts the address of the expression n to SSA, adds it to s and returns the SSA result.
5417 // The value that the returned Value represents is guaranteed to be non-nil.
5418 func (s *state) addr(n ir.Node) *ssa.Value {
5419 if n.Op() != ir.ONAME {
5425 s.Fatalf("addr of canSSA expression: %+v", n)
5428 t := types.NewPtr(n.Type())
5429 linksymOffset := func(lsym *obj.LSym, offset int64) *ssa.Value {
5430 v := s.entryNewValue1A(ssa.OpAddr, t, lsym, s.sb)
5431 // TODO: Make OpAddr use AuxInt as well as Aux.
5433 v = s.entryNewValue1I(ssa.OpOffPtr, v.Type, offset, v)
5438 case ir.OLINKSYMOFFSET:
5439 no := n.(*ir.LinksymOffsetExpr)
5440 return linksymOffset(no.Linksym, no.Offset_)
5443 if n.Heapaddr != nil {
5444 return s.expr(n.Heapaddr)
5449 return linksymOffset(n.Linksym(), 0)
5456 s.Fatalf("addr of undeclared ONAME %v. declared: %v", n, s.decladdrs)
5459 return s.newValue2Apos(ssa.OpLocalAddr, t, n, s.sp, s.mem(), !ir.IsAutoTmp(n))
5461 case ir.PPARAMOUT: // Same as PAUTO -- cannot generate LEA early.
5462 // ensure that we reuse symbols for out parameters so
5463 // that cse works on their addresses
5464 return s.newValue2Apos(ssa.OpLocalAddr, t, n, s.sp, s.mem(), true)
5466 s.Fatalf("variable address class %v not implemented", n.Class)
5470 // load return from callee
5471 n := n.(*ir.ResultExpr)
5472 return s.resultAddrOfCall(s.prevCall, n.Index, n.Type())
5474 n := n.(*ir.IndexExpr)
5475 if n.X.Type().IsSlice() {
5477 i := s.expr(n.Index)
5478 len := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], a)
5479 i = s.boundsCheck(i, len, ssa.BoundsIndex, n.Bounded())
5480 p := s.newValue1(ssa.OpSlicePtr, t, a)
5481 return s.newValue2(ssa.OpPtrIndex, t, p, i)
5484 i := s.expr(n.Index)
5485 len := s.constInt(types.Types[types.TINT], n.X.Type().NumElem())
5486 i = s.boundsCheck(i, len, ssa.BoundsIndex, n.Bounded())
5487 return s.newValue2(ssa.OpPtrIndex, types.NewPtr(n.X.Type().Elem()), a, i)
5490 n := n.(*ir.StarExpr)
5491 return s.exprPtr(n.X, n.Bounded(), n.Pos())
5493 n := n.(*ir.SelectorExpr)
5495 return s.newValue1I(ssa.OpOffPtr, t, n.Offset(), p)
5497 n := n.(*ir.SelectorExpr)
5498 p := s.exprPtr(n.X, n.Bounded(), n.Pos())
5499 return s.newValue1I(ssa.OpOffPtr, t, n.Offset(), p)
5501 n := n.(*ir.ConvExpr)
5502 if n.Type() == n.X.Type() {
5506 return s.newValue1(ssa.OpCopy, t, addr) // ensure that addr has the right type
5507 case ir.OCALLFUNC, ir.OCALLINTER:
5508 n := n.(*ir.CallExpr)
5509 return s.callAddr(n, callNormal)
5510 case ir.ODOTTYPE, ir.ODYNAMICDOTTYPE:
5512 if n.Op() == ir.ODOTTYPE {
5513 v, _ = s.dottype(n.(*ir.TypeAssertExpr), false)
5515 v, _ = s.dynamicDottype(n.(*ir.DynamicTypeAssertExpr), false)
5517 if v.Op != ssa.OpLoad {
5518 s.Fatalf("dottype of non-load")
5520 if v.Args[1] != s.mem() {
5521 s.Fatalf("memory no longer live from dottype load")
5525 s.Fatalf("unhandled addr %v", n.Op())
5530 // canSSA reports whether n is SSA-able.
5531 // n must be an ONAME (or an ODOT sequence with an ONAME base).
5532 func (s *state) canSSA(n ir.Node) bool {
5533 if base.Flag.N != 0 {
5538 if nn.Op() == ir.ODOT {
5539 nn := nn.(*ir.SelectorExpr)
5543 if nn.Op() == ir.OINDEX {
5544 nn := nn.(*ir.IndexExpr)
5545 if nn.X.Type().IsArray() {
5552 if n.Op() != ir.ONAME {
5555 return s.canSSAName(n.(*ir.Name)) && TypeOK(n.Type())
5558 func (s *state) canSSAName(name *ir.Name) bool {
5559 if name.Addrtaken() || !name.OnStack() {
5565 // TODO: handle this case? Named return values must be
5566 // in memory so that the deferred function can see them.
5567 // Maybe do: if !strings.HasPrefix(n.String(), "~") { return false }
5568 // Or maybe not, see issue 18860. Even unnamed return values
5569 // must be written back so if a defer recovers, the caller can see them.
5572 if s.cgoUnsafeArgs {
5573 // Cgo effectively takes the address of all result args,
5574 // but the compiler can't see that.
5579 // TODO: try to make more variables SSAable?
5582 // TypeOK reports whether variables of type t are SSA-able.
5583 func TypeOK(t *types.Type) bool {
5585 if t.Size() > int64(4*types.PtrSize) {
5586 // 4*Widthptr is an arbitrary constant. We want it
5587 // to be at least 3*Widthptr so slices can be registerized.
5588 // Too big and we'll introduce too much register pressure.
5593 // We can't do larger arrays because dynamic indexing is
5594 // not supported on SSA variables.
5595 // TODO: allow if all indexes are constant.
5596 if t.NumElem() <= 1 {
5597 return TypeOK(t.Elem())
5601 if t.NumFields() > ssa.MaxStruct {
5604 for _, t1 := range t.Fields().Slice() {
5605 if !TypeOK(t1.Type) {
5615 // exprPtr evaluates n to a pointer and nil-checks it.
5616 func (s *state) exprPtr(n ir.Node, bounded bool, lineno src.XPos) *ssa.Value {
5618 if bounded || n.NonNil() {
5619 if s.f.Frontend().Debug_checknil() && lineno.Line() > 1 {
5620 s.f.Warnl(lineno, "removed nil check")
5628 // nilCheck generates nil pointer checking code.
5629 // Used only for automatically inserted nil checks,
5630 // not for user code like 'x != nil'.
5631 func (s *state) nilCheck(ptr *ssa.Value) {
5632 if base.Debug.DisableNil != 0 || s.curfn.NilCheckDisabled() {
5635 s.newValue2(ssa.OpNilCheck, types.TypeVoid, ptr, s.mem())
5638 // boundsCheck generates bounds checking code. Checks if 0 <= idx <[=] len, branches to exit if not.
5639 // Starts a new block on return.
5640 // On input, len must be converted to full int width and be nonnegative.
5641 // Returns idx converted to full int width.
5642 // If bounded is true then caller guarantees the index is not out of bounds
5643 // (but boundsCheck will still extend the index to full int width).
5644 func (s *state) boundsCheck(idx, len *ssa.Value, kind ssa.BoundsKind, bounded bool) *ssa.Value {
5645 idx = s.extendIndex(idx, len, kind, bounded)
5647 if bounded || base.Flag.B != 0 {
5648 // If bounded or bounds checking is flag-disabled, then no check necessary,
5649 // just return the extended index.
5651 // Here, bounded == true if the compiler generated the index itself,
5652 // such as in the expansion of a slice initializer. These indexes are
5653 // compiler-generated, not Go program variables, so they cannot be
5654 // attacker-controlled, so we can omit Spectre masking as well.
5656 // Note that we do not want to omit Spectre masking in code like:
5658 // if 0 <= i && i < len(x) {
5662 // Lucky for us, bounded==false for that code.
5663 // In that case (handled below), we emit a bound check (and Spectre mask)
5664 // and then the prove pass will remove the bounds check.
5665 // In theory the prove pass could potentially remove certain
5666 // Spectre masks, but it's very delicate and probably better
5667 // to be conservative and leave them all in.
5671 bNext := s.f.NewBlock(ssa.BlockPlain)
5672 bPanic := s.f.NewBlock(ssa.BlockExit)
5674 if !idx.Type.IsSigned() {
5676 case ssa.BoundsIndex:
5677 kind = ssa.BoundsIndexU
5678 case ssa.BoundsSliceAlen:
5679 kind = ssa.BoundsSliceAlenU
5680 case ssa.BoundsSliceAcap:
5681 kind = ssa.BoundsSliceAcapU
5682 case ssa.BoundsSliceB:
5683 kind = ssa.BoundsSliceBU
5684 case ssa.BoundsSlice3Alen:
5685 kind = ssa.BoundsSlice3AlenU
5686 case ssa.BoundsSlice3Acap:
5687 kind = ssa.BoundsSlice3AcapU
5688 case ssa.BoundsSlice3B:
5689 kind = ssa.BoundsSlice3BU
5690 case ssa.BoundsSlice3C:
5691 kind = ssa.BoundsSlice3CU
5696 if kind == ssa.BoundsIndex || kind == ssa.BoundsIndexU {
5697 cmp = s.newValue2(ssa.OpIsInBounds, types.Types[types.TBOOL], idx, len)
5699 cmp = s.newValue2(ssa.OpIsSliceInBounds, types.Types[types.TBOOL], idx, len)
5702 b.Kind = ssa.BlockIf
5704 b.Likely = ssa.BranchLikely
5708 s.startBlock(bPanic)
5709 if Arch.LinkArch.Family == sys.Wasm {
5710 // TODO(khr): figure out how to do "register" based calling convention for bounds checks.
5711 // Should be similar to gcWriteBarrier, but I can't make it work.
5712 s.rtcall(BoundsCheckFunc[kind], false, nil, idx, len)
5714 mem := s.newValue3I(ssa.OpPanicBounds, types.TypeMem, int64(kind), idx, len, s.mem())
5715 s.endBlock().SetControl(mem)
5719 // In Spectre index mode, apply an appropriate mask to avoid speculative out-of-bounds accesses.
5720 if base.Flag.Cfg.SpectreIndex {
5721 op := ssa.OpSpectreIndex
5722 if kind != ssa.BoundsIndex && kind != ssa.BoundsIndexU {
5723 op = ssa.OpSpectreSliceIndex
5725 idx = s.newValue2(op, types.Types[types.TINT], idx, len)
5731 // If cmp (a bool) is false, panic using the given function.
5732 func (s *state) check(cmp *ssa.Value, fn *obj.LSym) {
5734 b.Kind = ssa.BlockIf
5736 b.Likely = ssa.BranchLikely
5737 bNext := s.f.NewBlock(ssa.BlockPlain)
5739 pos := base.Ctxt.PosTable.Pos(line)
5740 fl := funcLine{f: fn, base: pos.Base(), line: pos.Line()}
5741 bPanic := s.panics[fl]
5743 bPanic = s.f.NewBlock(ssa.BlockPlain)
5744 s.panics[fl] = bPanic
5745 s.startBlock(bPanic)
5746 // The panic call takes/returns memory to ensure that the right
5747 // memory state is observed if the panic happens.
5748 s.rtcall(fn, false, nil)
5755 func (s *state) intDivide(n ir.Node, a, b *ssa.Value) *ssa.Value {
5758 case ssa.OpConst8, ssa.OpConst16, ssa.OpConst32, ssa.OpConst64:
5764 // do a size-appropriate check for zero
5765 cmp := s.newValue2(s.ssaOp(ir.ONE, n.Type()), types.Types[types.TBOOL], b, s.zeroVal(n.Type()))
5766 s.check(cmp, ir.Syms.Panicdivide)
5768 return s.newValue2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
5771 // rtcall issues a call to the given runtime function fn with the listed args.
5772 // Returns a slice of results of the given result types.
5773 // The call is added to the end of the current block.
5774 // If returns is false, the block is marked as an exit block.
5775 func (s *state) rtcall(fn *obj.LSym, returns bool, results []*types.Type, args ...*ssa.Value) []*ssa.Value {
5777 // Write args to the stack
5778 off := base.Ctxt.Arch.FixedFrameSize
5779 var callArgs []*ssa.Value
5780 var callArgTypes []*types.Type
5782 for _, arg := range args {
5784 off = types.RoundUp(off, t.Alignment())
5786 callArgs = append(callArgs, arg)
5787 callArgTypes = append(callArgTypes, t)
5790 off = types.RoundUp(off, int64(types.RegSize))
5794 aux := ssa.StaticAuxCall(fn, s.f.ABIDefault.ABIAnalyzeTypes(nil, callArgTypes, results))
5795 callArgs = append(callArgs, s.mem())
5796 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
5797 call.AddArgs(callArgs...)
5798 s.vars[memVar] = s.newValue1I(ssa.OpSelectN, types.TypeMem, int64(len(results)), call)
5803 b.Kind = ssa.BlockExit
5805 call.AuxInt = off - base.Ctxt.Arch.FixedFrameSize
5806 if len(results) > 0 {
5807 s.Fatalf("panic call can't have results")
5813 res := make([]*ssa.Value, len(results))
5814 for i, t := range results {
5815 off = types.RoundUp(off, t.Alignment())
5816 res[i] = s.resultOfCall(call, int64(i), t)
5819 off = types.RoundUp(off, int64(types.PtrSize))
5821 // Remember how much callee stack space we needed.
5827 // do *left = right for type t.
5828 func (s *state) storeType(t *types.Type, left, right *ssa.Value, skip skipMask, leftIsStmt bool) {
5829 s.instrument(t, left, instrumentWrite)
5831 if skip == 0 && (!t.HasPointers() || ssa.IsStackAddr(left)) {
5832 // Known to not have write barrier. Store the whole type.
5833 s.vars[memVar] = s.newValue3Apos(ssa.OpStore, types.TypeMem, t, left, right, s.mem(), leftIsStmt)
5837 // store scalar fields first, so write barrier stores for
5838 // pointer fields can be grouped together, and scalar values
5839 // don't need to be live across the write barrier call.
5840 // TODO: if the writebarrier pass knows how to reorder stores,
5841 // we can do a single store here as long as skip==0.
5842 s.storeTypeScalars(t, left, right, skip)
5843 if skip&skipPtr == 0 && t.HasPointers() {
5844 s.storeTypePtrs(t, left, right)
5848 // do *left = right for all scalar (non-pointer) parts of t.
5849 func (s *state) storeTypeScalars(t *types.Type, left, right *ssa.Value, skip skipMask) {
5851 case t.IsBoolean() || t.IsInteger() || t.IsFloat() || t.IsComplex():
5852 s.store(t, left, right)
5853 case t.IsPtrShaped():
5854 if t.IsPtr() && t.Elem().NotInHeap() {
5855 s.store(t, left, right) // see issue 42032
5857 // otherwise, no scalar fields.
5859 if skip&skipLen != 0 {
5862 len := s.newValue1(ssa.OpStringLen, types.Types[types.TINT], right)
5863 lenAddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, s.config.PtrSize, left)
5864 s.store(types.Types[types.TINT], lenAddr, len)
5866 if skip&skipLen == 0 {
5867 len := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], right)
5868 lenAddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, s.config.PtrSize, left)
5869 s.store(types.Types[types.TINT], lenAddr, len)
5871 if skip&skipCap == 0 {
5872 cap := s.newValue1(ssa.OpSliceCap, types.Types[types.TINT], right)
5873 capAddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, 2*s.config.PtrSize, left)
5874 s.store(types.Types[types.TINT], capAddr, cap)
5876 case t.IsInterface():
5877 // itab field doesn't need a write barrier (even though it is a pointer).
5878 itab := s.newValue1(ssa.OpITab, s.f.Config.Types.BytePtr, right)
5879 s.store(types.Types[types.TUINTPTR], left, itab)
5882 for i := 0; i < n; i++ {
5883 ft := t.FieldType(i)
5884 addr := s.newValue1I(ssa.OpOffPtr, ft.PtrTo(), t.FieldOff(i), left)
5885 val := s.newValue1I(ssa.OpStructSelect, ft, int64(i), right)
5886 s.storeTypeScalars(ft, addr, val, 0)
5888 case t.IsArray() && t.NumElem() == 0:
5890 case t.IsArray() && t.NumElem() == 1:
5891 s.storeTypeScalars(t.Elem(), left, s.newValue1I(ssa.OpArraySelect, t.Elem(), 0, right), 0)
5893 s.Fatalf("bad write barrier type %v", t)
5897 // do *left = right for all pointer parts of t.
5898 func (s *state) storeTypePtrs(t *types.Type, left, right *ssa.Value) {
5900 case t.IsPtrShaped():
5901 if t.IsPtr() && t.Elem().NotInHeap() {
5902 break // see issue 42032
5904 s.store(t, left, right)
5906 ptr := s.newValue1(ssa.OpStringPtr, s.f.Config.Types.BytePtr, right)
5907 s.store(s.f.Config.Types.BytePtr, left, ptr)
5909 elType := types.NewPtr(t.Elem())
5910 ptr := s.newValue1(ssa.OpSlicePtr, elType, right)
5911 s.store(elType, left, ptr)
5912 case t.IsInterface():
5913 // itab field is treated as a scalar.
5914 idata := s.newValue1(ssa.OpIData, s.f.Config.Types.BytePtr, right)
5915 idataAddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.BytePtrPtr, s.config.PtrSize, left)
5916 s.store(s.f.Config.Types.BytePtr, idataAddr, idata)
5919 for i := 0; i < n; i++ {
5920 ft := t.FieldType(i)
5921 if !ft.HasPointers() {
5924 addr := s.newValue1I(ssa.OpOffPtr, ft.PtrTo(), t.FieldOff(i), left)
5925 val := s.newValue1I(ssa.OpStructSelect, ft, int64(i), right)
5926 s.storeTypePtrs(ft, addr, val)
5928 case t.IsArray() && t.NumElem() == 0:
5930 case t.IsArray() && t.NumElem() == 1:
5931 s.storeTypePtrs(t.Elem(), left, s.newValue1I(ssa.OpArraySelect, t.Elem(), 0, right))
5933 s.Fatalf("bad write barrier type %v", t)
5937 // putArg evaluates n for the purpose of passing it as an argument to a function and returns the value for the call.
5938 func (s *state) putArg(n ir.Node, t *types.Type) *ssa.Value {
5941 a = s.newValue2(ssa.OpDereference, t, s.addr(n), s.mem())
5948 func (s *state) storeArgWithBase(n ir.Node, t *types.Type, base *ssa.Value, off int64) {
5949 pt := types.NewPtr(t)
5952 // Use special routine that avoids allocation on duplicate offsets.
5953 addr = s.constOffPtrSP(pt, off)
5955 addr = s.newValue1I(ssa.OpOffPtr, pt, off, base)
5965 s.storeType(t, addr, a, 0, false)
5968 // slice computes the slice v[i:j:k] and returns ptr, len, and cap of result.
5969 // i,j,k may be nil, in which case they are set to their default value.
5970 // v may be a slice, string or pointer to an array.
5971 func (s *state) slice(v, i, j, k *ssa.Value, bounded bool) (p, l, c *ssa.Value) {
5973 var ptr, len, cap *ssa.Value
5976 ptr = s.newValue1(ssa.OpSlicePtr, types.NewPtr(t.Elem()), v)
5977 len = s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], v)
5978 cap = s.newValue1(ssa.OpSliceCap, types.Types[types.TINT], v)
5980 ptr = s.newValue1(ssa.OpStringPtr, types.NewPtr(types.Types[types.TUINT8]), v)
5981 len = s.newValue1(ssa.OpStringLen, types.Types[types.TINT], v)
5984 if !t.Elem().IsArray() {
5985 s.Fatalf("bad ptr to array in slice %v\n", t)
5988 ptr = s.newValue1(ssa.OpCopy, types.NewPtr(t.Elem().Elem()), v)
5989 len = s.constInt(types.Types[types.TINT], t.Elem().NumElem())
5992 s.Fatalf("bad type in slice %v\n", t)
5995 // Set default values
5997 i = s.constInt(types.Types[types.TINT], 0)
6008 // Panic if slice indices are not in bounds.
6009 // Make sure we check these in reverse order so that we're always
6010 // comparing against a value known to be nonnegative. See issue 28797.
6013 kind := ssa.BoundsSlice3Alen
6015 kind = ssa.BoundsSlice3Acap
6017 k = s.boundsCheck(k, cap, kind, bounded)
6020 j = s.boundsCheck(j, k, ssa.BoundsSlice3B, bounded)
6022 i = s.boundsCheck(i, j, ssa.BoundsSlice3C, bounded)
6025 kind := ssa.BoundsSliceAlen
6027 kind = ssa.BoundsSliceAcap
6029 j = s.boundsCheck(j, k, kind, bounded)
6031 i = s.boundsCheck(i, j, ssa.BoundsSliceB, bounded)
6034 // Word-sized integer operations.
6035 subOp := s.ssaOp(ir.OSUB, types.Types[types.TINT])
6036 mulOp := s.ssaOp(ir.OMUL, types.Types[types.TINT])
6037 andOp := s.ssaOp(ir.OAND, types.Types[types.TINT])
6039 // Calculate the length (rlen) and capacity (rcap) of the new slice.
6040 // For strings the capacity of the result is unimportant. However,
6041 // we use rcap to test if we've generated a zero-length slice.
6042 // Use length of strings for that.
6043 rlen := s.newValue2(subOp, types.Types[types.TINT], j, i)
6045 if j != k && !t.IsString() {
6046 rcap = s.newValue2(subOp, types.Types[types.TINT], k, i)
6049 if (i.Op == ssa.OpConst64 || i.Op == ssa.OpConst32) && i.AuxInt == 0 {
6050 // No pointer arithmetic necessary.
6051 return ptr, rlen, rcap
6054 // Calculate the base pointer (rptr) for the new slice.
6056 // Generate the following code assuming that indexes are in bounds.
6057 // The masking is to make sure that we don't generate a slice
6058 // that points to the next object in memory. We cannot just set
6059 // the pointer to nil because then we would create a nil slice or
6064 // rptr = ptr + (mask(rcap) & (i * stride))
6066 // Where mask(x) is 0 if x==0 and -1 if x>0 and stride is the width
6067 // of the element type.
6068 stride := s.constInt(types.Types[types.TINT], ptr.Type.Elem().Size())
6070 // The delta is the number of bytes to offset ptr by.
6071 delta := s.newValue2(mulOp, types.Types[types.TINT], i, stride)
6073 // If we're slicing to the point where the capacity is zero,
6074 // zero out the delta.
6075 mask := s.newValue1(ssa.OpSlicemask, types.Types[types.TINT], rcap)
6076 delta = s.newValue2(andOp, types.Types[types.TINT], delta, mask)
6078 // Compute rptr = ptr + delta.
6079 rptr := s.newValue2(ssa.OpAddPtr, ptr.Type, ptr, delta)
6081 return rptr, rlen, rcap
6084 type u642fcvtTab struct {
6085 leq, cvt2F, and, rsh, or, add ssa.Op
6086 one func(*state, *types.Type, int64) *ssa.Value
6089 var u64_f64 = u642fcvtTab{
6091 cvt2F: ssa.OpCvt64to64F,
6093 rsh: ssa.OpRsh64Ux64,
6096 one: (*state).constInt64,
6099 var u64_f32 = u642fcvtTab{
6101 cvt2F: ssa.OpCvt64to32F,
6103 rsh: ssa.OpRsh64Ux64,
6106 one: (*state).constInt64,
6109 func (s *state) uint64Tofloat64(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6110 return s.uint64Tofloat(&u64_f64, n, x, ft, tt)
6113 func (s *state) uint64Tofloat32(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6114 return s.uint64Tofloat(&u64_f32, n, x, ft, tt)
6117 func (s *state) uint64Tofloat(cvttab *u642fcvtTab, n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6119 // result = (floatY) x
6121 // y = uintX(x) ; y = x & 1
6122 // z = uintX(x) ; z = z >> 1
6124 // result = floatY(z)
6125 // result = result + result
6128 // Code borrowed from old code generator.
6129 // What's going on: large 64-bit "unsigned" looks like
6130 // negative number to hardware's integer-to-float
6131 // conversion. However, because the mantissa is only
6132 // 63 bits, we don't need the LSB, so instead we do an
6133 // unsigned right shift (divide by two), convert, and
6134 // double. However, before we do that, we need to be
6135 // sure that we do not lose a "1" if that made the
6136 // difference in the resulting rounding. Therefore, we
6137 // preserve it, and OR (not ADD) it back in. The case
6138 // that matters is when the eleven discarded bits are
6139 // equal to 10000000001; that rounds up, and the 1 cannot
6140 // be lost else it would round down if the LSB of the
6141 // candidate mantissa is 0.
6142 cmp := s.newValue2(cvttab.leq, types.Types[types.TBOOL], s.zeroVal(ft), x)
6144 b.Kind = ssa.BlockIf
6146 b.Likely = ssa.BranchLikely
6148 bThen := s.f.NewBlock(ssa.BlockPlain)
6149 bElse := s.f.NewBlock(ssa.BlockPlain)
6150 bAfter := s.f.NewBlock(ssa.BlockPlain)
6154 a0 := s.newValue1(cvttab.cvt2F, tt, x)
6157 bThen.AddEdgeTo(bAfter)
6161 one := cvttab.one(s, ft, 1)
6162 y := s.newValue2(cvttab.and, ft, x, one)
6163 z := s.newValue2(cvttab.rsh, ft, x, one)
6164 z = s.newValue2(cvttab.or, ft, z, y)
6165 a := s.newValue1(cvttab.cvt2F, tt, z)
6166 a1 := s.newValue2(cvttab.add, tt, a, a)
6169 bElse.AddEdgeTo(bAfter)
6171 s.startBlock(bAfter)
6172 return s.variable(n, n.Type())
6175 type u322fcvtTab struct {
6176 cvtI2F, cvtF2F ssa.Op
6179 var u32_f64 = u322fcvtTab{
6180 cvtI2F: ssa.OpCvt32to64F,
6184 var u32_f32 = u322fcvtTab{
6185 cvtI2F: ssa.OpCvt32to32F,
6186 cvtF2F: ssa.OpCvt64Fto32F,
6189 func (s *state) uint32Tofloat64(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6190 return s.uint32Tofloat(&u32_f64, n, x, ft, tt)
6193 func (s *state) uint32Tofloat32(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6194 return s.uint32Tofloat(&u32_f32, n, x, ft, tt)
6197 func (s *state) uint32Tofloat(cvttab *u322fcvtTab, n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6199 // result = floatY(x)
6201 // result = floatY(float64(x) + (1<<32))
6203 cmp := s.newValue2(ssa.OpLeq32, types.Types[types.TBOOL], s.zeroVal(ft), x)
6205 b.Kind = ssa.BlockIf
6207 b.Likely = ssa.BranchLikely
6209 bThen := s.f.NewBlock(ssa.BlockPlain)
6210 bElse := s.f.NewBlock(ssa.BlockPlain)
6211 bAfter := s.f.NewBlock(ssa.BlockPlain)
6215 a0 := s.newValue1(cvttab.cvtI2F, tt, x)
6218 bThen.AddEdgeTo(bAfter)
6222 a1 := s.newValue1(ssa.OpCvt32to64F, types.Types[types.TFLOAT64], x)
6223 twoToThe32 := s.constFloat64(types.Types[types.TFLOAT64], float64(1<<32))
6224 a2 := s.newValue2(ssa.OpAdd64F, types.Types[types.TFLOAT64], a1, twoToThe32)
6225 a3 := s.newValue1(cvttab.cvtF2F, tt, a2)
6229 bElse.AddEdgeTo(bAfter)
6231 s.startBlock(bAfter)
6232 return s.variable(n, n.Type())
6235 // referenceTypeBuiltin generates code for the len/cap builtins for maps and channels.
6236 func (s *state) referenceTypeBuiltin(n *ir.UnaryExpr, x *ssa.Value) *ssa.Value {
6237 if !n.X.Type().IsMap() && !n.X.Type().IsChan() {
6238 s.Fatalf("node must be a map or a channel")
6244 // return *((*int)n)
6246 // return *(((*int)n)+1)
6249 nilValue := s.constNil(types.Types[types.TUINTPTR])
6250 cmp := s.newValue2(ssa.OpEqPtr, types.Types[types.TBOOL], x, nilValue)
6252 b.Kind = ssa.BlockIf
6254 b.Likely = ssa.BranchUnlikely
6256 bThen := s.f.NewBlock(ssa.BlockPlain)
6257 bElse := s.f.NewBlock(ssa.BlockPlain)
6258 bAfter := s.f.NewBlock(ssa.BlockPlain)
6260 // length/capacity of a nil map/chan is zero
6263 s.vars[n] = s.zeroVal(lenType)
6265 bThen.AddEdgeTo(bAfter)
6271 // length is stored in the first word for map/chan
6272 s.vars[n] = s.load(lenType, x)
6274 // capacity is stored in the second word for chan
6275 sw := s.newValue1I(ssa.OpOffPtr, lenType.PtrTo(), lenType.Size(), x)
6276 s.vars[n] = s.load(lenType, sw)
6278 s.Fatalf("op must be OLEN or OCAP")
6281 bElse.AddEdgeTo(bAfter)
6283 s.startBlock(bAfter)
6284 return s.variable(n, lenType)
6287 type f2uCvtTab struct {
6288 ltf, cvt2U, subf, or ssa.Op
6289 floatValue func(*state, *types.Type, float64) *ssa.Value
6290 intValue func(*state, *types.Type, int64) *ssa.Value
6294 var f32_u64 = f2uCvtTab{
6296 cvt2U: ssa.OpCvt32Fto64,
6299 floatValue: (*state).constFloat32,
6300 intValue: (*state).constInt64,
6304 var f64_u64 = f2uCvtTab{
6306 cvt2U: ssa.OpCvt64Fto64,
6309 floatValue: (*state).constFloat64,
6310 intValue: (*state).constInt64,
6314 var f32_u32 = f2uCvtTab{
6316 cvt2U: ssa.OpCvt32Fto32,
6319 floatValue: (*state).constFloat32,
6320 intValue: func(s *state, t *types.Type, v int64) *ssa.Value { return s.constInt32(t, int32(v)) },
6324 var f64_u32 = f2uCvtTab{
6326 cvt2U: ssa.OpCvt64Fto32,
6329 floatValue: (*state).constFloat64,
6330 intValue: func(s *state, t *types.Type, v int64) *ssa.Value { return s.constInt32(t, int32(v)) },
6334 func (s *state) float32ToUint64(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6335 return s.floatToUint(&f32_u64, n, x, ft, tt)
6337 func (s *state) float64ToUint64(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6338 return s.floatToUint(&f64_u64, n, x, ft, tt)
6341 func (s *state) float32ToUint32(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6342 return s.floatToUint(&f32_u32, n, x, ft, tt)
6345 func (s *state) float64ToUint32(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6346 return s.floatToUint(&f64_u32, n, x, ft, tt)
6349 func (s *state) floatToUint(cvttab *f2uCvtTab, n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6350 // cutoff:=1<<(intY_Size-1)
6351 // if x < floatX(cutoff) {
6352 // result = uintY(x)
6354 // y = x - floatX(cutoff)
6356 // result = z | -(cutoff)
6358 cutoff := cvttab.floatValue(s, ft, float64(cvttab.cutoff))
6359 cmp := s.newValue2(cvttab.ltf, types.Types[types.TBOOL], x, cutoff)
6361 b.Kind = ssa.BlockIf
6363 b.Likely = ssa.BranchLikely
6365 bThen := s.f.NewBlock(ssa.BlockPlain)
6366 bElse := s.f.NewBlock(ssa.BlockPlain)
6367 bAfter := s.f.NewBlock(ssa.BlockPlain)
6371 a0 := s.newValue1(cvttab.cvt2U, tt, x)
6374 bThen.AddEdgeTo(bAfter)
6378 y := s.newValue2(cvttab.subf, ft, x, cutoff)
6379 y = s.newValue1(cvttab.cvt2U, tt, y)
6380 z := cvttab.intValue(s, tt, int64(-cvttab.cutoff))
6381 a1 := s.newValue2(cvttab.or, tt, y, z)
6384 bElse.AddEdgeTo(bAfter)
6386 s.startBlock(bAfter)
6387 return s.variable(n, n.Type())
6390 // dottype generates SSA for a type assertion node.
6391 // commaok indicates whether to panic or return a bool.
6392 // If commaok is false, resok will be nil.
6393 func (s *state) dottype(n *ir.TypeAssertExpr, commaok bool) (res, resok *ssa.Value) {
6394 iface := s.expr(n.X) // input interface
6395 target := s.reflectType(n.Type()) // target type
6396 var targetItab *ssa.Value
6398 targetItab = s.expr(n.ITab)
6400 return s.dottype1(n.Pos(), n.X.Type(), n.Type(), iface, nil, target, targetItab, commaok)
6403 func (s *state) dynamicDottype(n *ir.DynamicTypeAssertExpr, commaok bool) (res, resok *ssa.Value) {
6404 iface := s.expr(n.X)
6405 var source, target, targetItab *ssa.Value
6406 if n.SrcRType != nil {
6407 source = s.expr(n.SrcRType)
6409 if !n.X.Type().IsEmptyInterface() && !n.Type().IsInterface() {
6410 byteptr := s.f.Config.Types.BytePtr
6411 targetItab = s.expr(n.ITab)
6412 // TODO(mdempsky): Investigate whether compiling n.RType could be
6413 // better than loading itab.typ.
6414 target = s.load(byteptr, s.newValue1I(ssa.OpOffPtr, byteptr, int64(types.PtrSize), targetItab)) // itab.typ
6416 target = s.expr(n.RType)
6418 return s.dottype1(n.Pos(), n.X.Type(), n.Type(), iface, source, target, targetItab, commaok)
6421 // dottype1 implements a x.(T) operation. iface is the argument (x), dst is the type we're asserting to (T)
6422 // and src is the type we're asserting from.
6423 // source is the *runtime._type of src
6424 // target is the *runtime._type of dst.
6425 // If src is a nonempty interface and dst is not an interface, targetItab is an itab representing (dst, src). Otherwise it is nil.
6426 // commaok is true if the caller wants a boolean success value. Otherwise, the generated code panics if the conversion fails.
6427 func (s *state) dottype1(pos src.XPos, src, dst *types.Type, iface, source, target, targetItab *ssa.Value, commaok bool) (res, resok *ssa.Value) {
6428 byteptr := s.f.Config.Types.BytePtr
6429 if dst.IsInterface() {
6430 if dst.IsEmptyInterface() {
6431 // Converting to an empty interface.
6432 // Input could be an empty or nonempty interface.
6433 if base.Debug.TypeAssert > 0 {
6434 base.WarnfAt(pos, "type assertion inlined")
6437 // Get itab/type field from input.
6438 itab := s.newValue1(ssa.OpITab, byteptr, iface)
6439 // Conversion succeeds iff that field is not nil.
6440 cond := s.newValue2(ssa.OpNeqPtr, types.Types[types.TBOOL], itab, s.constNil(byteptr))
6442 if src.IsEmptyInterface() && commaok {
6443 // Converting empty interface to empty interface with ,ok is just a nil check.
6447 // Branch on nilness.
6449 b.Kind = ssa.BlockIf
6451 b.Likely = ssa.BranchLikely
6452 bOk := s.f.NewBlock(ssa.BlockPlain)
6453 bFail := s.f.NewBlock(ssa.BlockPlain)
6458 // On failure, panic by calling panicnildottype.
6460 s.rtcall(ir.Syms.Panicnildottype, false, nil, target)
6462 // On success, return (perhaps modified) input interface.
6464 if src.IsEmptyInterface() {
6465 res = iface // Use input interface unchanged.
6468 // Load type out of itab, build interface with existing idata.
6469 off := s.newValue1I(ssa.OpOffPtr, byteptr, int64(types.PtrSize), itab)
6470 typ := s.load(byteptr, off)
6471 idata := s.newValue1(ssa.OpIData, byteptr, iface)
6472 res = s.newValue2(ssa.OpIMake, dst, typ, idata)
6477 // nonempty -> empty
6478 // Need to load type from itab
6479 off := s.newValue1I(ssa.OpOffPtr, byteptr, int64(types.PtrSize), itab)
6480 s.vars[typVar] = s.load(byteptr, off)
6483 // itab is nil, might as well use that as the nil result.
6485 s.vars[typVar] = itab
6489 bEnd := s.f.NewBlock(ssa.BlockPlain)
6491 bFail.AddEdgeTo(bEnd)
6493 idata := s.newValue1(ssa.OpIData, byteptr, iface)
6494 res = s.newValue2(ssa.OpIMake, dst, s.variable(typVar, byteptr), idata)
6496 delete(s.vars, typVar) // no practical effect, just to indicate typVar is no longer live.
6499 // converting to a nonempty interface needs a runtime call.
6500 if base.Debug.TypeAssert > 0 {
6501 base.WarnfAt(pos, "type assertion not inlined")
6504 fn := ir.Syms.AssertI2I
6505 if src.IsEmptyInterface() {
6506 fn = ir.Syms.AssertE2I
6508 data := s.newValue1(ssa.OpIData, types.Types[types.TUNSAFEPTR], iface)
6509 tab := s.newValue1(ssa.OpITab, byteptr, iface)
6510 tab = s.rtcall(fn, true, []*types.Type{byteptr}, target, tab)[0]
6511 return s.newValue2(ssa.OpIMake, dst, tab, data), nil
6513 fn := ir.Syms.AssertI2I2
6514 if src.IsEmptyInterface() {
6515 fn = ir.Syms.AssertE2I2
6517 res = s.rtcall(fn, true, []*types.Type{dst}, target, iface)[0]
6518 resok = s.newValue2(ssa.OpNeqInter, types.Types[types.TBOOL], res, s.constInterface(dst))
6522 if base.Debug.TypeAssert > 0 {
6523 base.WarnfAt(pos, "type assertion inlined")
6526 // Converting to a concrete type.
6527 direct := types.IsDirectIface(dst)
6528 itab := s.newValue1(ssa.OpITab, byteptr, iface) // type word of interface
6529 if base.Debug.TypeAssert > 0 {
6530 base.WarnfAt(pos, "type assertion inlined")
6532 var wantedFirstWord *ssa.Value
6533 if src.IsEmptyInterface() {
6534 // Looking for pointer to target type.
6535 wantedFirstWord = target
6537 // Looking for pointer to itab for target type and source interface.
6538 wantedFirstWord = targetItab
6541 var tmp ir.Node // temporary for use with large types
6542 var addr *ssa.Value // address of tmp
6543 if commaok && !TypeOK(dst) {
6544 // unSSAable type, use temporary.
6545 // TODO: get rid of some of these temporaries.
6546 tmp, addr = s.temp(pos, dst)
6549 cond := s.newValue2(ssa.OpEqPtr, types.Types[types.TBOOL], itab, wantedFirstWord)
6551 b.Kind = ssa.BlockIf
6553 b.Likely = ssa.BranchLikely
6555 bOk := s.f.NewBlock(ssa.BlockPlain)
6556 bFail := s.f.NewBlock(ssa.BlockPlain)
6561 // on failure, panic by calling panicdottype
6565 taddr = s.reflectType(src)
6567 if src.IsEmptyInterface() {
6568 s.rtcall(ir.Syms.PanicdottypeE, false, nil, itab, target, taddr)
6570 s.rtcall(ir.Syms.PanicdottypeI, false, nil, itab, target, taddr)
6573 // on success, return data from interface
6576 return s.newValue1(ssa.OpIData, dst, iface), nil
6578 p := s.newValue1(ssa.OpIData, types.NewPtr(dst), iface)
6579 return s.load(dst, p), nil
6582 // commaok is the more complicated case because we have
6583 // a control flow merge point.
6584 bEnd := s.f.NewBlock(ssa.BlockPlain)
6585 // Note that we need a new valVar each time (unlike okVar where we can
6586 // reuse the variable) because it might have a different type every time.
6587 valVar := ssaMarker("val")
6589 // type assertion succeeded
6593 s.vars[valVar] = s.newValue1(ssa.OpIData, dst, iface)
6595 p := s.newValue1(ssa.OpIData, types.NewPtr(dst), iface)
6596 s.vars[valVar] = s.load(dst, p)
6599 p := s.newValue1(ssa.OpIData, types.NewPtr(dst), iface)
6600 s.move(dst, addr, p)
6602 s.vars[okVar] = s.constBool(true)
6606 // type assertion failed
6609 s.vars[valVar] = s.zeroVal(dst)
6613 s.vars[okVar] = s.constBool(false)
6615 bFail.AddEdgeTo(bEnd)
6620 res = s.variable(valVar, dst)
6621 delete(s.vars, valVar) // no practical effect, just to indicate typVar is no longer live.
6623 res = s.load(dst, addr)
6625 resok = s.variable(okVar, types.Types[types.TBOOL])
6626 delete(s.vars, okVar) // ditto
6630 // temp allocates a temp of type t at position pos
6631 func (s *state) temp(pos src.XPos, t *types.Type) (*ir.Name, *ssa.Value) {
6632 tmp := typecheck.TempAt(pos, s.curfn, t)
6633 if t.HasPointers() {
6634 s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, tmp, s.mem())
6640 // variable returns the value of a variable at the current location.
6641 func (s *state) variable(n ir.Node, t *types.Type) *ssa.Value {
6651 if s.curBlock == s.f.Entry {
6652 // No variable should be live at entry.
6653 s.f.Fatalf("value %v (%v) incorrectly live at entry", n, v)
6655 // Make a FwdRef, which records a value that's live on block input.
6656 // We'll find the matching definition as part of insertPhis.
6657 v = s.newValue0A(ssa.OpFwdRef, t, fwdRefAux{N: n})
6659 if n.Op() == ir.ONAME {
6660 s.addNamedValue(n.(*ir.Name), v)
6665 func (s *state) mem() *ssa.Value {
6666 return s.variable(memVar, types.TypeMem)
6669 func (s *state) addNamedValue(n *ir.Name, v *ssa.Value) {
6670 if n.Class == ir.Pxxx {
6671 // Don't track our marker nodes (memVar etc.).
6674 if ir.IsAutoTmp(n) {
6675 // Don't track temporary variables.
6678 if n.Class == ir.PPARAMOUT {
6679 // Don't track named output values. This prevents return values
6680 // from being assigned too early. See #14591 and #14762. TODO: allow this.
6683 loc := ssa.LocalSlot{N: n, Type: n.Type(), Off: 0}
6684 values, ok := s.f.NamedValues[loc]
6686 s.f.Names = append(s.f.Names, &loc)
6687 s.f.CanonicalLocalSlots[loc] = &loc
6689 s.f.NamedValues[loc] = append(values, v)
6692 // Branch is an unresolved branch.
6693 type Branch struct {
6694 P *obj.Prog // branch instruction
6695 B *ssa.Block // target
6698 // State contains state needed during Prog generation.
6704 // Branches remembers all the branch instructions we've seen
6705 // and where they would like to go.
6708 // JumpTables remembers all the jump tables we've seen.
6709 JumpTables []*ssa.Block
6711 // bstart remembers where each block starts (indexed by block ID)
6714 maxarg int64 // largest frame size for arguments to calls made by the function
6716 // Map from GC safe points to liveness index, generated by
6717 // liveness analysis.
6718 livenessMap liveness.Map
6720 // partLiveArgs includes arguments that may be partially live, for which we
6721 // need to generate instructions that spill the argument registers.
6722 partLiveArgs map[*ir.Name]bool
6724 // lineRunStart records the beginning of the current run of instructions
6725 // within a single block sharing the same line number
6726 // Used to move statement marks to the beginning of such runs.
6727 lineRunStart *obj.Prog
6729 // wasm: The number of values on the WebAssembly stack. This is only used as a safeguard.
6730 OnWasmStackSkipped int
6733 func (s *State) FuncInfo() *obj.FuncInfo {
6734 return s.pp.CurFunc.LSym.Func()
6737 // Prog appends a new Prog.
6738 func (s *State) Prog(as obj.As) *obj.Prog {
6740 if objw.LosesStmtMark(as) {
6743 // Float a statement start to the beginning of any same-line run.
6744 // lineRunStart is reset at block boundaries, which appears to work well.
6745 if s.lineRunStart == nil || s.lineRunStart.Pos.Line() != p.Pos.Line() {
6747 } else if p.Pos.IsStmt() == src.PosIsStmt {
6748 s.lineRunStart.Pos = s.lineRunStart.Pos.WithIsStmt()
6749 p.Pos = p.Pos.WithNotStmt()
6754 // Pc returns the current Prog.
6755 func (s *State) Pc() *obj.Prog {
6759 // SetPos sets the current source position.
6760 func (s *State) SetPos(pos src.XPos) {
6764 // Br emits a single branch instruction and returns the instruction.
6765 // Not all architectures need the returned instruction, but otherwise
6766 // the boilerplate is common to all.
6767 func (s *State) Br(op obj.As, target *ssa.Block) *obj.Prog {
6769 p.To.Type = obj.TYPE_BRANCH
6770 s.Branches = append(s.Branches, Branch{P: p, B: target})
6774 // DebugFriendlySetPosFrom adjusts Pos.IsStmt subject to heuristics
6775 // that reduce "jumpy" line number churn when debugging.
6776 // Spill/fill/copy instructions from the register allocator,
6777 // phi functions, and instructions with a no-pos position
6778 // are examples of instructions that can cause churn.
6779 func (s *State) DebugFriendlySetPosFrom(v *ssa.Value) {
6781 case ssa.OpPhi, ssa.OpCopy, ssa.OpLoadReg, ssa.OpStoreReg:
6782 // These are not statements
6783 s.SetPos(v.Pos.WithNotStmt())
6786 if p != src.NoXPos {
6787 // If the position is defined, update the position.
6788 // Also convert default IsStmt to NotStmt; only
6789 // explicit statement boundaries should appear
6790 // in the generated code.
6791 if p.IsStmt() != src.PosIsStmt {
6792 if s.pp.Pos.IsStmt() == src.PosIsStmt && s.pp.Pos.SameFileAndLine(p) {
6793 // If s.pp.Pos already has a statement mark, then it was set here (below) for
6794 // the previous value. If an actual instruction had been emitted for that
6795 // value, then the statement mark would have been reset. Since the statement
6796 // mark of s.pp.Pos was not reset, this position (file/line) still needs a
6797 // statement mark on an instruction. If file and line for this value are
6798 // the same as the previous value, then the first instruction for this
6799 // value will work to take the statement mark. Return early to avoid
6800 // resetting the statement mark.
6802 // The reset of s.pp.Pos occurs in (*Progs).Prog() -- if it emits
6803 // an instruction, and the instruction's statement mark was set,
6804 // and it is not one of the LosesStmtMark instructions,
6805 // then Prog() resets the statement mark on the (*Progs).Pos.
6809 // Calls use the pos attached to v, but copy the statement mark from State
6813 s.SetPos(s.pp.Pos.WithNotStmt())
6818 // emit argument info (locations on stack) for traceback.
6819 func emitArgInfo(e *ssafn, f *ssa.Func, pp *objw.Progs) {
6820 ft := e.curfn.Type()
6821 if ft.NumRecvs() == 0 && ft.NumParams() == 0 {
6825 x := EmitArgInfo(e.curfn, f.OwnAux.ABIInfo())
6826 x.Set(obj.AttrContentAddressable, true)
6827 e.curfn.LSym.Func().ArgInfo = x
6829 // Emit a funcdata pointing at the arg info data.
6830 p := pp.Prog(obj.AFUNCDATA)
6831 p.From.SetConst(rtabi.FUNCDATA_ArgInfo)
6832 p.To.Type = obj.TYPE_MEM
6833 p.To.Name = obj.NAME_EXTERN
6837 // emit argument info (locations on stack) of f for traceback.
6838 func EmitArgInfo(f *ir.Func, abiInfo *abi.ABIParamResultInfo) *obj.LSym {
6839 x := base.Ctxt.Lookup(fmt.Sprintf("%s.arginfo%d", f.LSym.Name, f.ABI))
6840 // NOTE: do not set ContentAddressable here. This may be referenced from
6841 // assembly code by name (in this case f is a declaration).
6842 // Instead, set it in emitArgInfo above.
6844 PtrSize := int64(types.PtrSize)
6845 uintptrTyp := types.Types[types.TUINTPTR]
6847 isAggregate := func(t *types.Type) bool {
6848 return t.IsStruct() || t.IsArray() || t.IsComplex() || t.IsInterface() || t.IsString() || t.IsSlice()
6851 // Populate the data.
6852 // The data is a stream of bytes, which contains the offsets and sizes of the
6853 // non-aggregate arguments or non-aggregate fields/elements of aggregate-typed
6854 // arguments, along with special "operators". Specifically,
6855 // - for each non-aggrgate arg/field/element, its offset from FP (1 byte) and
6857 // - special operators:
6858 // - 0xff - end of sequence
6859 // - 0xfe - print { (at the start of an aggregate-typed argument)
6860 // - 0xfd - print } (at the end of an aggregate-typed argument)
6861 // - 0xfc - print ... (more args/fields/elements)
6862 // - 0xfb - print _ (offset too large)
6863 // These constants need to be in sync with runtime.traceback.go:printArgs.
6869 _offsetTooLarge = 0xfb
6870 _special = 0xf0 // above this are operators, below this are ordinary offsets
6874 limit = 10 // print no more than 10 args/components
6875 maxDepth = 5 // no more than 5 layers of nesting
6877 // maxLen is a (conservative) upper bound of the byte stream length. For
6878 // each arg/component, it has no more than 2 bytes of data (size, offset),
6879 // and no more than one {, }, ... at each level (it cannot have both the
6880 // data and ... unless it is the last one, just be conservative). Plus 1
6882 maxLen = (maxDepth*3+2)*limit + 1
6887 writebyte := func(o uint8) { wOff = objw.Uint8(x, wOff, o) }
6889 // Write one non-aggrgate arg/field/element.
6890 write1 := func(sz, offset int64) {
6891 if offset >= _special {
6892 writebyte(_offsetTooLarge)
6894 writebyte(uint8(offset))
6895 writebyte(uint8(sz))
6900 // Visit t recursively and write it out.
6901 // Returns whether to continue visiting.
6902 var visitType func(baseOffset int64, t *types.Type, depth int) bool
6903 visitType = func(baseOffset int64, t *types.Type, depth int) bool {
6905 writebyte(_dotdotdot)
6908 if !isAggregate(t) {
6909 write1(t.Size(), baseOffset)
6912 writebyte(_startAgg)
6914 if depth >= maxDepth {
6915 writebyte(_dotdotdot)
6921 case t.IsInterface(), t.IsString():
6922 _ = visitType(baseOffset, uintptrTyp, depth) &&
6923 visitType(baseOffset+PtrSize, uintptrTyp, depth)
6925 _ = visitType(baseOffset, uintptrTyp, depth) &&
6926 visitType(baseOffset+PtrSize, uintptrTyp, depth) &&
6927 visitType(baseOffset+PtrSize*2, uintptrTyp, depth)
6929 _ = visitType(baseOffset, types.FloatForComplex(t), depth) &&
6930 visitType(baseOffset+t.Size()/2, types.FloatForComplex(t), depth)
6932 if t.NumElem() == 0 {
6933 n++ // {} counts as a component
6936 for i := int64(0); i < t.NumElem(); i++ {
6937 if !visitType(baseOffset, t.Elem(), depth) {
6940 baseOffset += t.Elem().Size()
6943 if t.NumFields() == 0 {
6944 n++ // {} counts as a component
6947 for _, field := range t.Fields().Slice() {
6948 if !visitType(baseOffset+field.Offset, field.Type, depth) {
6958 if strings.Contains(f.LSym.Name, "[") {
6959 // Skip the dictionary argument - it is implicit and the user doesn't need to see it.
6963 for _, a := range abiInfo.InParams()[start:] {
6964 if !visitType(a.FrameOffset(abiInfo), a.Type, 0) {
6970 base.Fatalf("ArgInfo too large")
6976 // for wrapper, emit info of wrapped function.
6977 func emitWrappedFuncInfo(e *ssafn, pp *objw.Progs) {
6978 if base.Ctxt.Flag_linkshared {
6979 // Relative reference (SymPtrOff) to another shared object doesn't work.
6984 wfn := e.curfn.WrappedFunc
6989 wsym := wfn.Linksym()
6990 x := base.Ctxt.LookupInit(fmt.Sprintf("%s.wrapinfo", wsym.Name), func(x *obj.LSym) {
6991 objw.SymPtrOff(x, 0, wsym)
6992 x.Set(obj.AttrContentAddressable, true)
6994 e.curfn.LSym.Func().WrapInfo = x
6996 // Emit a funcdata pointing at the wrap info data.
6997 p := pp.Prog(obj.AFUNCDATA)
6998 p.From.SetConst(rtabi.FUNCDATA_WrapInfo)
6999 p.To.Type = obj.TYPE_MEM
7000 p.To.Name = obj.NAME_EXTERN
7004 // genssa appends entries to pp for each instruction in f.
7005 func genssa(f *ssa.Func, pp *objw.Progs) {
7007 s.ABI = f.OwnAux.Fn.ABI()
7009 e := f.Frontend().(*ssafn)
7011 s.livenessMap, s.partLiveArgs = liveness.Compute(e.curfn, f, e.stkptrsize, pp)
7012 emitArgInfo(e, f, pp)
7013 argLiveBlockMap, argLiveValueMap := liveness.ArgLiveness(e.curfn, f, pp)
7015 openDeferInfo := e.curfn.LSym.Func().OpenCodedDeferInfo
7016 if openDeferInfo != nil {
7017 // This function uses open-coded defers -- write out the funcdata
7018 // info that we computed at the end of genssa.
7019 p := pp.Prog(obj.AFUNCDATA)
7020 p.From.SetConst(rtabi.FUNCDATA_OpenCodedDeferInfo)
7021 p.To.Type = obj.TYPE_MEM
7022 p.To.Name = obj.NAME_EXTERN
7023 p.To.Sym = openDeferInfo
7026 emitWrappedFuncInfo(e, pp)
7028 // Remember where each block starts.
7029 s.bstart = make([]*obj.Prog, f.NumBlocks())
7031 var progToValue map[*obj.Prog]*ssa.Value
7032 var progToBlock map[*obj.Prog]*ssa.Block
7033 var valueToProgAfter []*obj.Prog // The first Prog following computation of a value v; v is visible at this point.
7034 gatherPrintInfo := f.PrintOrHtmlSSA || ssa.GenssaDump[f.Name]
7035 if gatherPrintInfo {
7036 progToValue = make(map[*obj.Prog]*ssa.Value, f.NumValues())
7037 progToBlock = make(map[*obj.Prog]*ssa.Block, f.NumBlocks())
7038 f.Logf("genssa %s\n", f.Name)
7039 progToBlock[s.pp.Next] = f.Blocks[0]
7042 if base.Ctxt.Flag_locationlists {
7043 if cap(f.Cache.ValueToProgAfter) < f.NumValues() {
7044 f.Cache.ValueToProgAfter = make([]*obj.Prog, f.NumValues())
7046 valueToProgAfter = f.Cache.ValueToProgAfter[:f.NumValues()]
7047 for i := range valueToProgAfter {
7048 valueToProgAfter[i] = nil
7052 // If the very first instruction is not tagged as a statement,
7053 // debuggers may attribute it to previous function in program.
7054 firstPos := src.NoXPos
7055 for _, v := range f.Entry.Values {
7056 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 {
7058 v.Pos = firstPos.WithDefaultStmt()
7063 // inlMarks has an entry for each Prog that implements an inline mark.
7064 // It maps from that Prog to the global inlining id of the inlined body
7065 // which should unwind to this Prog's location.
7066 var inlMarks map[*obj.Prog]int32
7067 var inlMarkList []*obj.Prog
7069 // inlMarksByPos maps from a (column 1) source position to the set of
7070 // Progs that are in the set above and have that source position.
7071 var inlMarksByPos map[src.XPos][]*obj.Prog
7073 var argLiveIdx int = -1 // argument liveness info index
7075 // Emit basic blocks
7076 for i, b := range f.Blocks {
7077 s.bstart[b.ID] = s.pp.Next
7078 s.lineRunStart = nil
7079 s.SetPos(s.pp.Pos.WithNotStmt()) // It needs a non-empty Pos, but cannot be a statement boundary (yet).
7081 // Attach a "default" liveness info. Normally this will be
7082 // overwritten in the Values loop below for each Value. But
7083 // for an empty block this will be used for its control
7084 // instruction. We won't use the actual liveness map on a
7085 // control instruction. Just mark it something that is
7086 // preemptible, unless this function is "all unsafe".
7087 s.pp.NextLive = objw.LivenessIndex{StackMapIndex: -1, IsUnsafePoint: liveness.IsUnsafe(f)}
7089 if idx, ok := argLiveBlockMap[b.ID]; ok && idx != argLiveIdx {
7091 p := s.pp.Prog(obj.APCDATA)
7092 p.From.SetConst(rtabi.PCDATA_ArgLiveIndex)
7093 p.To.SetConst(int64(idx))
7096 // Emit values in block
7097 Arch.SSAMarkMoves(&s, b)
7098 for _, v := range b.Values {
7100 s.DebugFriendlySetPosFrom(v)
7102 if v.Op.ResultInArg0() && v.ResultReg() != v.Args[0].Reg() {
7103 v.Fatalf("input[0] and output not in same register %s", v.LongString())
7108 // memory arg needs no code
7110 // input args need no code
7111 case ssa.OpSP, ssa.OpSB:
7113 case ssa.OpSelect0, ssa.OpSelect1, ssa.OpSelectN, ssa.OpMakeResult:
7116 // nothing to do when there's a g register,
7117 // and checkLower complains if there's not
7118 case ssa.OpVarDef, ssa.OpVarLive, ssa.OpKeepAlive, ssa.OpWBend:
7119 // nothing to do; already used by liveness
7123 // nothing to do; no-op conversion for liveness
7124 if v.Args[0].Reg() != v.Reg() {
7125 v.Fatalf("OpConvert should be a no-op: %s; %s", v.Args[0].LongString(), v.LongString())
7128 p := Arch.Ginsnop(s.pp)
7129 if inlMarks == nil {
7130 inlMarks = map[*obj.Prog]int32{}
7131 inlMarksByPos = map[src.XPos][]*obj.Prog{}
7133 inlMarks[p] = v.AuxInt32()
7134 inlMarkList = append(inlMarkList, p)
7135 pos := v.Pos.AtColumn1()
7136 inlMarksByPos[pos] = append(inlMarksByPos[pos], p)
7137 firstPos = src.NoXPos
7140 // Special case for first line in function; move it to the start (which cannot be a register-valued instruction)
7141 if firstPos != src.NoXPos && v.Op != ssa.OpArgIntReg && v.Op != ssa.OpArgFloatReg && v.Op != ssa.OpLoadReg && v.Op != ssa.OpStoreReg {
7143 firstPos = src.NoXPos
7145 // Attach this safe point to the next
7147 s.pp.NextLive = s.livenessMap.Get(v)
7149 // let the backend handle it
7150 Arch.SSAGenValue(&s, v)
7153 if idx, ok := argLiveValueMap[v.ID]; ok && idx != argLiveIdx {
7155 p := s.pp.Prog(obj.APCDATA)
7156 p.From.SetConst(rtabi.PCDATA_ArgLiveIndex)
7157 p.To.SetConst(int64(idx))
7160 if base.Ctxt.Flag_locationlists {
7161 valueToProgAfter[v.ID] = s.pp.Next
7164 if gatherPrintInfo {
7165 for ; x != s.pp.Next; x = x.Link {
7170 // If this is an empty infinite loop, stick a hardware NOP in there so that debuggers are less confused.
7171 if s.bstart[b.ID] == s.pp.Next && len(b.Succs) == 1 && b.Succs[0].Block() == b {
7172 p := Arch.Ginsnop(s.pp)
7173 p.Pos = p.Pos.WithIsStmt()
7174 if b.Pos == src.NoXPos {
7175 b.Pos = p.Pos // It needs a file, otherwise a no-file non-zero line causes confusion. See #35652.
7176 if b.Pos == src.NoXPos {
7177 b.Pos = pp.Text.Pos // Sometimes p.Pos is empty. See #35695.
7180 b.Pos = b.Pos.WithBogusLine() // Debuggers are not good about infinite loops, force a change in line number
7182 // Emit control flow instructions for block
7184 if i < len(f.Blocks)-1 && base.Flag.N == 0 {
7185 // If -N, leave next==nil so every block with successors
7186 // ends in a JMP (except call blocks - plive doesn't like
7187 // select{send,recv} followed by a JMP call). Helps keep
7188 // line numbers for otherwise empty blocks.
7189 next = f.Blocks[i+1]
7193 Arch.SSAGenBlock(&s, b, next)
7194 if gatherPrintInfo {
7195 for ; x != s.pp.Next; x = x.Link {
7200 if f.Blocks[len(f.Blocks)-1].Kind == ssa.BlockExit {
7201 // We need the return address of a panic call to
7202 // still be inside the function in question. So if
7203 // it ends in a call which doesn't return, add a
7204 // nop (which will never execute) after the call.
7207 if openDeferInfo != nil {
7208 // When doing open-coded defers, generate a disconnected call to
7209 // deferreturn and a return. This will be used to during panic
7210 // recovery to unwind the stack and return back to the runtime.
7211 s.pp.NextLive = s.livenessMap.DeferReturn
7212 p := pp.Prog(obj.ACALL)
7213 p.To.Type = obj.TYPE_MEM
7214 p.To.Name = obj.NAME_EXTERN
7215 p.To.Sym = ir.Syms.Deferreturn
7217 // Load results into registers. So when a deferred function
7218 // recovers a panic, it will return to caller with right results.
7219 // The results are already in memory, because they are not SSA'd
7220 // when the function has defers (see canSSAName).
7221 for _, o := range f.OwnAux.ABIInfo().OutParams() {
7222 n := o.Name.(*ir.Name)
7223 rts, offs := o.RegisterTypesAndOffsets()
7224 for i := range o.Registers {
7225 Arch.LoadRegResult(&s, f, rts[i], ssa.ObjRegForAbiReg(o.Registers[i], f.Config), n, offs[i])
7232 if inlMarks != nil {
7235 // We have some inline marks. Try to find other instructions we're
7236 // going to emit anyway, and use those instructions instead of the
7238 for p := pp.Text; p != nil; p = p.Link {
7239 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 {
7240 // Don't use 0-sized instructions as inline marks, because we need
7241 // to identify inline mark instructions by pc offset.
7242 // (Some of these instructions are sometimes zero-sized, sometimes not.
7243 // We must not use anything that even might be zero-sized.)
7244 // TODO: are there others?
7247 if _, ok := inlMarks[p]; ok {
7248 // Don't use inline marks themselves. We don't know
7249 // whether they will be zero-sized or not yet.
7252 if p.As == obj.ACALL || p.As == obj.ADUFFCOPY || p.As == obj.ADUFFZERO {
7255 pos := p.Pos.AtColumn1()
7256 s := inlMarksByPos[pos]
7260 for _, m := range s {
7261 // We found an instruction with the same source position as
7262 // some of the inline marks.
7263 // Use this instruction instead.
7264 p.Pos = p.Pos.WithIsStmt() // promote position to a statement
7265 pp.CurFunc.LSym.Func().AddInlMark(p, inlMarks[m])
7266 // Make the inline mark a real nop, so it doesn't generate any code.
7272 delete(inlMarksByPos, pos)
7274 // Any unmatched inline marks now need to be added to the inlining tree (and will generate a nop instruction).
7275 for _, p := range inlMarkList {
7276 if p.As != obj.ANOP {
7277 pp.CurFunc.LSym.Func().AddInlMark(p, inlMarks[p])
7281 if e.stksize == 0 && !hasCall {
7282 // Frameless leaf function. It doesn't need any preamble,
7283 // so make sure its first instruction isn't from an inlined callee.
7284 // If it is, add a nop at the start of the function with a position
7285 // equal to the start of the function.
7286 // This ensures that runtime.FuncForPC(uintptr(reflect.ValueOf(fn).Pointer())).Name()
7287 // returns the right answer. See issue 58300.
7288 for p := pp.Text; p != nil; p = p.Link {
7289 if p.As == obj.AFUNCDATA || p.As == obj.APCDATA || p.As == obj.ATEXT || p.As == obj.ANOP {
7292 if base.Ctxt.PosTable.Pos(p.Pos).Base().InliningIndex() >= 0 {
7293 // Make a real (not 0-sized) nop.
7294 nop := Arch.Ginsnop(pp)
7295 nop.Pos = e.curfn.Pos().WithIsStmt()
7297 // Unfortunately, Ginsnop puts the instruction at the
7298 // end of the list. Move it up to just before p.
7300 // Unlink from the current list.
7301 for x := pp.Text; x != nil; x = x.Link {
7307 // Splice in right before p.
7308 for x := pp.Text; x != nil; x = x.Link {
7321 if base.Ctxt.Flag_locationlists {
7322 var debugInfo *ssa.FuncDebug
7323 debugInfo = e.curfn.DebugInfo.(*ssa.FuncDebug)
7324 if e.curfn.ABI == obj.ABIInternal && base.Flag.N != 0 {
7325 ssa.BuildFuncDebugNoOptimized(base.Ctxt, f, base.Debug.LocationLists > 1, StackOffset, debugInfo)
7327 ssa.BuildFuncDebug(base.Ctxt, f, base.Debug.LocationLists, StackOffset, debugInfo)
7330 idToIdx := make([]int, f.NumBlocks())
7331 for i, b := range f.Blocks {
7334 // Note that at this moment, Prog.Pc is a sequence number; it's
7335 // not a real PC until after assembly, so this mapping has to
7337 debugInfo.GetPC = func(b, v ssa.ID) int64 {
7339 case ssa.BlockStart.ID:
7340 if b == f.Entry.ID {
7341 return 0 // Start at the very beginning, at the assembler-generated prologue.
7342 // this should only happen for function args (ssa.OpArg)
7345 case ssa.BlockEnd.ID:
7346 blk := f.Blocks[idToIdx[b]]
7347 nv := len(blk.Values)
7348 return valueToProgAfter[blk.Values[nv-1].ID].Pc
7349 case ssa.FuncEnd.ID:
7350 return e.curfn.LSym.Size
7352 return valueToProgAfter[v].Pc
7357 // Resolve branches, and relax DefaultStmt into NotStmt
7358 for _, br := range s.Branches {
7359 br.P.To.SetTarget(s.bstart[br.B.ID])
7360 if br.P.Pos.IsStmt() != src.PosIsStmt {
7361 br.P.Pos = br.P.Pos.WithNotStmt()
7362 } else if v0 := br.B.FirstPossibleStmtValue(); v0 != nil && v0.Pos.Line() == br.P.Pos.Line() && v0.Pos.IsStmt() == src.PosIsStmt {
7363 br.P.Pos = br.P.Pos.WithNotStmt()
7368 // Resolve jump table destinations.
7369 for _, jt := range s.JumpTables {
7370 // Convert from *Block targets to *Prog targets.
7371 targets := make([]*obj.Prog, len(jt.Succs))
7372 for i, e := range jt.Succs {
7373 targets[i] = s.bstart[e.Block().ID]
7375 // Add to list of jump tables to be resolved at assembly time.
7376 // The assembler converts from *Prog entries to absolute addresses
7377 // once it knows instruction byte offsets.
7378 fi := pp.CurFunc.LSym.Func()
7379 fi.JumpTables = append(fi.JumpTables, obj.JumpTable{Sym: jt.Aux.(*obj.LSym), Targets: targets})
7382 if e.log { // spew to stdout
7384 for p := pp.Text; p != nil; p = p.Link {
7385 if p.Pos.IsKnown() && p.InnermostFilename() != filename {
7386 filename = p.InnermostFilename()
7387 f.Logf("# %s\n", filename)
7391 if v, ok := progToValue[p]; ok {
7393 } else if b, ok := progToBlock[p]; ok {
7396 s = " " // most value and branch strings are 2-3 characters long
7398 f.Logf(" %-6s\t%.5d (%s)\t%s\n", s, p.Pc, p.InnermostLineNumber(), p.InstructionString())
7401 if f.HTMLWriter != nil { // spew to ssa.html
7402 var buf strings.Builder
7403 buf.WriteString("<code>")
7404 buf.WriteString("<dl class=\"ssa-gen\">")
7406 for p := pp.Text; p != nil; p = p.Link {
7407 // Don't spam every line with the file name, which is often huge.
7408 // Only print changes, and "unknown" is not a change.
7409 if p.Pos.IsKnown() && p.InnermostFilename() != filename {
7410 filename = p.InnermostFilename()
7411 buf.WriteString("<dt class=\"ssa-prog-src\"></dt><dd class=\"ssa-prog\">")
7412 buf.WriteString(html.EscapeString("# " + filename))
7413 buf.WriteString("</dd>")
7416 buf.WriteString("<dt class=\"ssa-prog-src\">")
7417 if v, ok := progToValue[p]; ok {
7418 buf.WriteString(v.HTML())
7419 } else if b, ok := progToBlock[p]; ok {
7420 buf.WriteString("<b>" + b.HTML() + "</b>")
7422 buf.WriteString("</dt>")
7423 buf.WriteString("<dd class=\"ssa-prog\">")
7424 fmt.Fprintf(&buf, "%.5d <span class=\"l%v line-number\">(%s)</span> %s", p.Pc, p.InnermostLineNumber(), p.InnermostLineNumberHTML(), html.EscapeString(p.InstructionString()))
7425 buf.WriteString("</dd>")
7427 buf.WriteString("</dl>")
7428 buf.WriteString("</code>")
7429 f.HTMLWriter.WriteColumn("genssa", "genssa", "ssa-prog", buf.String())
7431 if ssa.GenssaDump[f.Name] {
7432 fi := f.DumpFileForPhase("genssa")
7435 // inliningDiffers if any filename changes or if any line number except the innermost (last index) changes.
7436 inliningDiffers := func(a, b []src.Pos) bool {
7437 if len(a) != len(b) {
7441 if a[i].Filename() != b[i].Filename() {
7444 if i != len(a)-1 && a[i].Line() != b[i].Line() {
7451 var allPosOld []src.Pos
7452 var allPos []src.Pos
7454 for p := pp.Text; p != nil; p = p.Link {
7455 if p.Pos.IsKnown() {
7457 p.Ctxt.AllPos(p.Pos, func(pos src.Pos) { allPos = append(allPos, pos) })
7458 if inliningDiffers(allPos, allPosOld) {
7459 for _, pos := range allPos {
7460 fmt.Fprintf(fi, "# %s:%d\n", pos.Filename(), pos.Line())
7462 allPos, allPosOld = allPosOld, allPos // swap, not copy, so that they do not share slice storage.
7467 if v, ok := progToValue[p]; ok {
7469 } else if b, ok := progToBlock[p]; ok {
7472 s = " " // most value and branch strings are 2-3 characters long
7474 fmt.Fprintf(fi, " %-6s\t%.5d %s\t%s\n", s, p.Pc, ssa.StmtString(p.Pos), p.InstructionString())
7482 f.HTMLWriter.Close()
7486 func defframe(s *State, e *ssafn, f *ssa.Func) {
7489 s.maxarg = types.RoundUp(s.maxarg, e.stkalign)
7490 frame := s.maxarg + e.stksize
7491 if Arch.PadFrame != nil {
7492 frame = Arch.PadFrame(frame)
7495 // Fill in argument and frame size.
7496 pp.Text.To.Type = obj.TYPE_TEXTSIZE
7497 pp.Text.To.Val = int32(types.RoundUp(f.OwnAux.ArgWidth(), int64(types.RegSize)))
7498 pp.Text.To.Offset = frame
7502 // Insert code to spill argument registers if the named slot may be partially
7503 // live. That is, the named slot is considered live by liveness analysis,
7504 // (because a part of it is live), but we may not spill all parts into the
7505 // slot. This can only happen with aggregate-typed arguments that are SSA-able
7506 // and not address-taken (for non-SSA-able or address-taken arguments we always
7508 // Note: spilling is unnecessary in the -N/no-optimize case, since all values
7509 // will be considered non-SSAable and spilled up front.
7510 // TODO(register args) Make liveness more fine-grained to that partial spilling is okay.
7511 if f.OwnAux.ABIInfo().InRegistersUsed() != 0 && base.Flag.N == 0 {
7512 // First, see if it is already spilled before it may be live. Look for a spill
7513 // in the entry block up to the first safepoint.
7514 type nameOff struct {
7518 partLiveArgsSpilled := make(map[nameOff]bool)
7519 for _, v := range f.Entry.Values {
7523 if v.Op != ssa.OpStoreReg || v.Args[0].Op != ssa.OpArgIntReg {
7526 n, off := ssa.AutoVar(v)
7527 if n.Class != ir.PPARAM || n.Addrtaken() || !TypeOK(n.Type()) || !s.partLiveArgs[n] {
7530 partLiveArgsSpilled[nameOff{n, off}] = true
7533 // Then, insert code to spill registers if not already.
7534 for _, a := range f.OwnAux.ABIInfo().InParams() {
7535 n, ok := a.Name.(*ir.Name)
7536 if !ok || n.Addrtaken() || !TypeOK(n.Type()) || !s.partLiveArgs[n] || len(a.Registers) <= 1 {
7539 rts, offs := a.RegisterTypesAndOffsets()
7540 for i := range a.Registers {
7541 if !rts[i].HasPointers() {
7544 if partLiveArgsSpilled[nameOff{n, offs[i]}] {
7545 continue // already spilled
7547 reg := ssa.ObjRegForAbiReg(a.Registers[i], f.Config)
7548 p = Arch.SpillArgReg(pp, p, f, rts[i], reg, n, offs[i])
7553 // Insert code to zero ambiguously live variables so that the
7554 // garbage collector only sees initialized values when it
7555 // looks for pointers.
7558 // Opaque state for backend to use. Current backends use it to
7559 // keep track of which helper registers have been zeroed.
7562 // Iterate through declarations. Autos are sorted in decreasing
7563 // frame offset order.
7564 for _, n := range e.curfn.Dcl {
7568 if n.Class != ir.PAUTO {
7569 e.Fatalf(n.Pos(), "needzero class %d", n.Class)
7571 if n.Type().Size()%int64(types.PtrSize) != 0 || n.FrameOffset()%int64(types.PtrSize) != 0 || n.Type().Size() == 0 {
7572 e.Fatalf(n.Pos(), "var %L has size %d offset %d", n, n.Type().Size(), n.Offset_)
7575 if lo != hi && n.FrameOffset()+n.Type().Size() >= lo-int64(2*types.RegSize) {
7576 // Merge with range we already have.
7577 lo = n.FrameOffset()
7582 p = Arch.ZeroRange(pp, p, frame+lo, hi-lo, &state)
7585 lo = n.FrameOffset()
7586 hi = lo + n.Type().Size()
7589 // Zero final range.
7590 Arch.ZeroRange(pp, p, frame+lo, hi-lo, &state)
7593 // For generating consecutive jump instructions to model a specific branching
7594 type IndexJump struct {
7599 func (s *State) oneJump(b *ssa.Block, jump *IndexJump) {
7600 p := s.Br(jump.Jump, b.Succs[jump.Index].Block())
7604 // CombJump generates combinational instructions (2 at present) for a block jump,
7605 // thereby the behaviour of non-standard condition codes could be simulated
7606 func (s *State) CombJump(b, next *ssa.Block, jumps *[2][2]IndexJump) {
7608 case b.Succs[0].Block():
7609 s.oneJump(b, &jumps[0][0])
7610 s.oneJump(b, &jumps[0][1])
7611 case b.Succs[1].Block():
7612 s.oneJump(b, &jumps[1][0])
7613 s.oneJump(b, &jumps[1][1])
7616 if b.Likely != ssa.BranchUnlikely {
7617 s.oneJump(b, &jumps[1][0])
7618 s.oneJump(b, &jumps[1][1])
7619 q = s.Br(obj.AJMP, b.Succs[1].Block())
7621 s.oneJump(b, &jumps[0][0])
7622 s.oneJump(b, &jumps[0][1])
7623 q = s.Br(obj.AJMP, b.Succs[0].Block())
7629 // AddAux adds the offset in the aux fields (AuxInt and Aux) of v to a.
7630 func AddAux(a *obj.Addr, v *ssa.Value) {
7631 AddAux2(a, v, v.AuxInt)
7633 func AddAux2(a *obj.Addr, v *ssa.Value, offset int64) {
7634 if a.Type != obj.TYPE_MEM && a.Type != obj.TYPE_ADDR {
7635 v.Fatalf("bad AddAux addr %v", a)
7637 // add integer offset
7640 // If no additional symbol offset, we're done.
7644 // Add symbol's offset from its base register.
7645 switch n := v.Aux.(type) {
7647 a.Name = obj.NAME_EXTERN
7650 a.Name = obj.NAME_EXTERN
7653 if n.Class == ir.PPARAM || (n.Class == ir.PPARAMOUT && !n.IsOutputParamInRegisters()) {
7654 a.Name = obj.NAME_PARAM
7655 a.Sym = ir.Orig(n).(*ir.Name).Linksym()
7656 a.Offset += n.FrameOffset()
7659 a.Name = obj.NAME_AUTO
7660 if n.Class == ir.PPARAMOUT {
7661 a.Sym = ir.Orig(n).(*ir.Name).Linksym()
7665 a.Offset += n.FrameOffset()
7667 v.Fatalf("aux in %s not implemented %#v", v, v.Aux)
7671 // extendIndex extends v to a full int width.
7672 // panic with the given kind if v does not fit in an int (only on 32-bit archs).
7673 func (s *state) extendIndex(idx, len *ssa.Value, kind ssa.BoundsKind, bounded bool) *ssa.Value {
7674 size := idx.Type.Size()
7675 if size == s.config.PtrSize {
7678 if size > s.config.PtrSize {
7679 // truncate 64-bit indexes on 32-bit pointer archs. Test the
7680 // high word and branch to out-of-bounds failure if it is not 0.
7682 if idx.Type.IsSigned() {
7683 lo = s.newValue1(ssa.OpInt64Lo, types.Types[types.TINT], idx)
7685 lo = s.newValue1(ssa.OpInt64Lo, types.Types[types.TUINT], idx)
7687 if bounded || base.Flag.B != 0 {
7690 bNext := s.f.NewBlock(ssa.BlockPlain)
7691 bPanic := s.f.NewBlock(ssa.BlockExit)
7692 hi := s.newValue1(ssa.OpInt64Hi, types.Types[types.TUINT32], idx)
7693 cmp := s.newValue2(ssa.OpEq32, types.Types[types.TBOOL], hi, s.constInt32(types.Types[types.TUINT32], 0))
7694 if !idx.Type.IsSigned() {
7696 case ssa.BoundsIndex:
7697 kind = ssa.BoundsIndexU
7698 case ssa.BoundsSliceAlen:
7699 kind = ssa.BoundsSliceAlenU
7700 case ssa.BoundsSliceAcap:
7701 kind = ssa.BoundsSliceAcapU
7702 case ssa.BoundsSliceB:
7703 kind = ssa.BoundsSliceBU
7704 case ssa.BoundsSlice3Alen:
7705 kind = ssa.BoundsSlice3AlenU
7706 case ssa.BoundsSlice3Acap:
7707 kind = ssa.BoundsSlice3AcapU
7708 case ssa.BoundsSlice3B:
7709 kind = ssa.BoundsSlice3BU
7710 case ssa.BoundsSlice3C:
7711 kind = ssa.BoundsSlice3CU
7715 b.Kind = ssa.BlockIf
7717 b.Likely = ssa.BranchLikely
7721 s.startBlock(bPanic)
7722 mem := s.newValue4I(ssa.OpPanicExtend, types.TypeMem, int64(kind), hi, lo, len, s.mem())
7723 s.endBlock().SetControl(mem)
7729 // Extend value to the required size
7731 if idx.Type.IsSigned() {
7732 switch 10*size + s.config.PtrSize {
7734 op = ssa.OpSignExt8to32
7736 op = ssa.OpSignExt8to64
7738 op = ssa.OpSignExt16to32
7740 op = ssa.OpSignExt16to64
7742 op = ssa.OpSignExt32to64
7744 s.Fatalf("bad signed index extension %s", idx.Type)
7747 switch 10*size + s.config.PtrSize {
7749 op = ssa.OpZeroExt8to32
7751 op = ssa.OpZeroExt8to64
7753 op = ssa.OpZeroExt16to32
7755 op = ssa.OpZeroExt16to64
7757 op = ssa.OpZeroExt32to64
7759 s.Fatalf("bad unsigned index extension %s", idx.Type)
7762 return s.newValue1(op, types.Types[types.TINT], idx)
7765 // CheckLoweredPhi checks that regalloc and stackalloc correctly handled phi values.
7766 // Called during ssaGenValue.
7767 func CheckLoweredPhi(v *ssa.Value) {
7768 if v.Op != ssa.OpPhi {
7769 v.Fatalf("CheckLoweredPhi called with non-phi value: %v", v.LongString())
7771 if v.Type.IsMemory() {
7775 loc := f.RegAlloc[v.ID]
7776 for _, a := range v.Args {
7777 if aloc := f.RegAlloc[a.ID]; aloc != loc { // TODO: .Equal() instead?
7778 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)
7783 // CheckLoweredGetClosurePtr checks that v is the first instruction in the function's entry block,
7784 // except for incoming in-register arguments.
7785 // The output of LoweredGetClosurePtr is generally hardwired to the correct register.
7786 // That register contains the closure pointer on closure entry.
7787 func CheckLoweredGetClosurePtr(v *ssa.Value) {
7788 entry := v.Block.Func.Entry
7789 if entry != v.Block {
7790 base.Fatalf("in %s, badly placed LoweredGetClosurePtr: %v %v", v.Block.Func.Name, v.Block, v)
7792 for _, w := range entry.Values {
7797 case ssa.OpArgIntReg, ssa.OpArgFloatReg:
7800 base.Fatalf("in %s, badly placed LoweredGetClosurePtr: %v %v", v.Block.Func.Name, v.Block, v)
7805 // CheckArgReg ensures that v is in the function's entry block.
7806 func CheckArgReg(v *ssa.Value) {
7807 entry := v.Block.Func.Entry
7808 if entry != v.Block {
7809 base.Fatalf("in %s, badly placed ArgIReg or ArgFReg: %v %v", v.Block.Func.Name, v.Block, v)
7813 func AddrAuto(a *obj.Addr, v *ssa.Value) {
7814 n, off := ssa.AutoVar(v)
7815 a.Type = obj.TYPE_MEM
7817 a.Reg = int16(Arch.REGSP)
7818 a.Offset = n.FrameOffset() + off
7819 if n.Class == ir.PPARAM || (n.Class == ir.PPARAMOUT && !n.IsOutputParamInRegisters()) {
7820 a.Name = obj.NAME_PARAM
7822 a.Name = obj.NAME_AUTO
7826 // Call returns a new CALL instruction for the SSA value v.
7827 // It uses PrepareCall to prepare the call.
7828 func (s *State) Call(v *ssa.Value) *obj.Prog {
7829 pPosIsStmt := s.pp.Pos.IsStmt() // The statement-ness fo the call comes from ssaGenState
7832 p := s.Prog(obj.ACALL)
7833 if pPosIsStmt == src.PosIsStmt {
7834 p.Pos = v.Pos.WithIsStmt()
7836 p.Pos = v.Pos.WithNotStmt()
7838 if sym, ok := v.Aux.(*ssa.AuxCall); ok && sym.Fn != nil {
7839 p.To.Type = obj.TYPE_MEM
7840 p.To.Name = obj.NAME_EXTERN
7843 // TODO(mdempsky): Can these differences be eliminated?
7844 switch Arch.LinkArch.Family {
7845 case sys.AMD64, sys.I386, sys.PPC64, sys.RISCV64, sys.S390X, sys.Wasm:
7846 p.To.Type = obj.TYPE_REG
7847 case sys.ARM, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64:
7848 p.To.Type = obj.TYPE_MEM
7850 base.Fatalf("unknown indirect call family")
7852 p.To.Reg = v.Args[0].Reg()
7857 // TailCall returns a new tail call instruction for the SSA value v.
7858 // It is like Call, but for a tail call.
7859 func (s *State) TailCall(v *ssa.Value) *obj.Prog {
7865 // PrepareCall prepares to emit a CALL instruction for v and does call-related bookkeeping.
7866 // It must be called immediately before emitting the actual CALL instruction,
7867 // since it emits PCDATA for the stack map at the call (calls are safe points).
7868 func (s *State) PrepareCall(v *ssa.Value) {
7869 idx := s.livenessMap.Get(v)
7870 if !idx.StackMapValid() {
7871 // See Liveness.hasStackMap.
7872 if sym, ok := v.Aux.(*ssa.AuxCall); !ok || !(sym.Fn == ir.Syms.WBZero || sym.Fn == ir.Syms.WBMove) {
7873 base.Fatalf("missing stack map index for %v", v.LongString())
7877 call, ok := v.Aux.(*ssa.AuxCall)
7880 // Record call graph information for nowritebarrierrec
7882 if nowritebarrierrecCheck != nil {
7883 nowritebarrierrecCheck.recordCall(s.pp.CurFunc, call.Fn, v.Pos)
7887 if s.maxarg < v.AuxInt {
7892 // UseArgs records the fact that an instruction needs a certain amount of
7893 // callee args space for its use.
7894 func (s *State) UseArgs(n int64) {
7900 // fieldIdx finds the index of the field referred to by the ODOT node n.
7901 func fieldIdx(n *ir.SelectorExpr) int {
7904 panic("ODOT's LHS is not a struct")
7907 for i, f := range t.Fields().Slice() {
7909 if f.Offset != n.Offset() {
7910 panic("field offset doesn't match")
7915 panic(fmt.Sprintf("can't find field in expr %v\n", n))
7917 // TODO: keep the result of this function somewhere in the ODOT Node
7918 // so we don't have to recompute it each time we need it.
7921 // ssafn holds frontend information about a function that the backend is processing.
7922 // It also exports a bunch of compiler services for the ssa backend.
7925 strings map[string]*obj.LSym // map from constant string to data symbols
7926 stksize int64 // stack size for current frame
7927 stkptrsize int64 // prefix of stack containing pointers
7929 // alignment for current frame.
7930 // NOTE: when stkalign > PtrSize, currently this only ensures the offsets of
7931 // objects in the stack frame are aligned. The stack pointer is still aligned
7935 log bool // print ssa debug to the stdout
7938 // StringData returns a symbol which
7939 // is the data component of a global string constant containing s.
7940 func (e *ssafn) StringData(s string) *obj.LSym {
7941 if aux, ok := e.strings[s]; ok {
7944 if e.strings == nil {
7945 e.strings = make(map[string]*obj.LSym)
7947 data := staticdata.StringSym(e.curfn.Pos(), s)
7952 func (e *ssafn) Auto(pos src.XPos, t *types.Type) *ir.Name {
7953 return typecheck.TempAt(pos, e.curfn, t) // Note: adds new auto to e.curfn.Func.Dcl list
7956 // SplitSlot returns a slot representing the data of parent starting at offset.
7957 func (e *ssafn) SplitSlot(parent *ssa.LocalSlot, suffix string, offset int64, t *types.Type) ssa.LocalSlot {
7960 if node.Class != ir.PAUTO || node.Addrtaken() {
7961 // addressed things and non-autos retain their parents (i.e., cannot truly be split)
7962 return ssa.LocalSlot{N: node, Type: t, Off: parent.Off + offset}
7965 s := &types.Sym{Name: node.Sym().Name + suffix, Pkg: types.LocalPkg}
7966 n := ir.NewNameAt(parent.N.Pos(), s)
7968 ir.AsNode(s.Def).Name().SetUsed(true)
7971 n.SetEsc(ir.EscNever)
7973 e.curfn.Dcl = append(e.curfn.Dcl, n)
7975 return ssa.LocalSlot{N: n, Type: t, Off: 0, SplitOf: parent, SplitOffset: offset}
7978 func (e *ssafn) CanSSA(t *types.Type) bool {
7982 // Logf logs a message from the compiler.
7983 func (e *ssafn) Logf(msg string, args ...interface{}) {
7985 fmt.Printf(msg, args...)
7989 func (e *ssafn) Log() bool {
7993 // Fatalf reports a compiler error and exits.
7994 func (e *ssafn) Fatalf(pos src.XPos, msg string, args ...interface{}) {
7996 nargs := append([]interface{}{ir.FuncName(e.curfn)}, args...)
7997 base.Fatalf("'%s': "+msg, nargs...)
8000 // Warnl reports a "warning", which is usually flag-triggered
8001 // logging output for the benefit of tests.
8002 func (e *ssafn) Warnl(pos src.XPos, fmt_ string, args ...interface{}) {
8003 base.WarnfAt(pos, fmt_, args...)
8006 func (e *ssafn) Debug_checknil() bool {
8007 return base.Debug.Nil != 0
8010 func (e *ssafn) UseWriteBarrier() bool {
8014 func (e *ssafn) Syslook(name string) *obj.LSym {
8016 case "goschedguarded":
8017 return ir.Syms.Goschedguarded
8018 case "writeBarrier":
8019 return ir.Syms.WriteBarrier
8021 return ir.Syms.WBZero
8023 return ir.Syms.WBMove
8024 case "cgoCheckMemmove":
8025 return ir.Syms.CgoCheckMemmove
8026 case "cgoCheckPtrWrite":
8027 return ir.Syms.CgoCheckPtrWrite
8029 e.Fatalf(src.NoXPos, "unknown Syslook func %v", name)
8033 func (e *ssafn) MyImportPath() string {
8034 return base.Ctxt.Pkgpath
8037 func (e *ssafn) Func() *ir.Func {
8041 func clobberBase(n ir.Node) ir.Node {
8042 if n.Op() == ir.ODOT {
8043 n := n.(*ir.SelectorExpr)
8044 if n.X.Type().NumFields() == 1 {
8045 return clobberBase(n.X)
8048 if n.Op() == ir.OINDEX {
8049 n := n.(*ir.IndexExpr)
8050 if n.X.Type().IsArray() && n.X.Type().NumElem() == 1 {
8051 return clobberBase(n.X)
8057 // callTargetLSym returns the correct LSym to call 'callee' using its ABI.
8058 func callTargetLSym(callee *ir.Name) *obj.LSym {
8059 if callee.Func == nil {
8060 // TODO(austin): This happens in case of interface method I.M from imported package.
8061 // It's ABIInternal, and would be better if callee.Func was never nil and we didn't
8063 return callee.Linksym()
8066 return callee.LinksymABI(callee.Func.ABI)
8069 func min8(a, b int8) int8 {
8076 func max8(a, b int8) int8 {
8083 // deferstruct makes a runtime._defer structure.
8084 func deferstruct() *types.Type {
8085 makefield := func(name string, typ *types.Type) *types.Field {
8086 // Unlike the global makefield function, this one needs to set Pkg
8087 // because these types might be compared (in SSA CSE sorting).
8088 // TODO: unify this makefield and the global one above.
8089 sym := &types.Sym{Name: name, Pkg: types.LocalPkg}
8090 return types.NewField(src.NoXPos, sym, typ)
8092 // These fields must match the ones in runtime/runtime2.go:_defer and
8093 // (*state).call above.
8094 fields := []*types.Field{
8095 makefield("started", types.Types[types.TBOOL]),
8096 makefield("heap", types.Types[types.TBOOL]),
8097 makefield("openDefer", types.Types[types.TBOOL]),
8098 makefield("sp", types.Types[types.TUINTPTR]),
8099 makefield("pc", types.Types[types.TUINTPTR]),
8100 // Note: the types here don't really matter. Defer structures
8101 // are always scanned explicitly during stack copying and GC,
8102 // so we make them uintptr type even though they are real pointers.
8103 makefield("fn", types.Types[types.TUINTPTR]),
8104 makefield("_panic", types.Types[types.TUINTPTR]),
8105 makefield("link", types.Types[types.TUINTPTR]),
8106 makefield("fd", types.Types[types.TUINTPTR]),
8107 makefield("varp", types.Types[types.TUINTPTR]),
8108 makefield("framepc", types.Types[types.TUINTPTR]),
8111 // build struct holding the above fields
8112 s := types.NewStruct(fields)
8114 types.CalcStructSize(s)
8118 // SpillSlotAddr uses LocalSlot information to initialize an obj.Addr
8119 // The resulting addr is used in a non-standard context -- in the prologue
8120 // of a function, before the frame has been constructed, so the standard
8121 // addressing for the parameters will be wrong.
8122 func SpillSlotAddr(spill ssa.Spill, baseReg int16, extraOffset int64) obj.Addr {
8124 Name: obj.NAME_NONE,
8127 Offset: spill.Offset + extraOffset,
8132 BoundsCheckFunc [ssa.BoundsKindCount]*obj.LSym
8133 ExtendCheckFunc [ssa.BoundsKindCount]*obj.LSym