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 // - Offset of the first closure slot (the rest are laid out consecutively).
286 func (s *state) emitOpenDeferInfo() {
287 firstOffset := s.openDefers[0].closureNode.FrameOffset()
289 // Verify that cmpstackvarlt laid out the slots in order.
290 for i, r := range s.openDefers {
291 have := r.closureNode.FrameOffset()
292 want := firstOffset + int64(i)*int64(types.PtrSize)
294 base.FatalfAt(s.curfn.Pos(), "unexpected frame offset for open-coded defer slot #%v: have %v, want %v", i, have, want)
298 x := base.Ctxt.Lookup(s.curfn.LSym.Name + ".opendefer")
299 x.Set(obj.AttrContentAddressable, true)
300 s.curfn.LSym.Func().OpenCodedDeferInfo = x
303 off = dvarint(x, off, -s.deferBitsTemp.FrameOffset())
304 off = dvarint(x, off, -firstOffset)
307 func okOffset(offset int64) int64 {
308 if offset == types.BOGUS_FUNARG_OFFSET {
309 panic(fmt.Errorf("Bogus offset %d", offset))
314 // buildssa builds an SSA function for fn.
315 // worker indicates which of the backend workers is doing the processing.
316 func buildssa(fn *ir.Func, worker int) *ssa.Func {
317 name := ir.FuncName(fn)
319 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"
320 pkgDotName := base.Ctxt.Pkgpath + "." + name
321 printssa = name == ssaDump ||
322 strings.HasSuffix(pkgDotName, ssaDump) && (pkgDotName == ssaDump || strings.HasSuffix(pkgDotName, "/"+ssaDump))
324 var astBuf *bytes.Buffer
326 astBuf = &bytes.Buffer{}
327 ir.FDumpList(astBuf, "buildssa-enter", fn.Enter)
328 ir.FDumpList(astBuf, "buildssa-body", fn.Body)
329 ir.FDumpList(astBuf, "buildssa-exit", fn.Exit)
331 fmt.Println("generating SSA for", name)
332 fmt.Print(astBuf.String())
340 s.hasdefer = fn.HasDefer()
341 if fn.Pragma&ir.CgoUnsafeArgs != 0 {
342 s.cgoUnsafeArgs = true
344 s.checkPtrEnabled = ir.ShouldCheckPtr(fn, 1)
348 log: printssa && ssaDumpStdout,
352 s.f = ssa.NewFunc(&fe)
355 s.f.Config = ssaConfig
356 s.f.Cache = &ssaCaches[worker]
359 s.f.PrintOrHtmlSSA = printssa
360 if fn.Pragma&ir.Nosplit != 0 {
363 s.f.ABI0 = ssaConfig.ABI0.Copy() // Make a copy to avoid racy map operations in type-register-width cache.
364 s.f.ABI1 = ssaConfig.ABI1.Copy()
365 s.f.ABIDefault = abiForFunc(nil, s.f.ABI0, s.f.ABI1)
366 s.f.ABISelf = abiForFunc(fn, s.f.ABI0, s.f.ABI1)
368 s.panics = map[funcLine]*ssa.Block{}
369 s.softFloat = s.config.SoftFloat
371 // Allocate starting block
372 s.f.Entry = s.f.NewBlock(ssa.BlockPlain)
373 s.f.Entry.Pos = fn.Pos()
378 ssaDF = filepath.Join(ssaDir, base.Ctxt.Pkgpath+"."+name+".html")
379 ssaD := filepath.Dir(ssaDF)
380 os.MkdirAll(ssaD, 0755)
382 s.f.HTMLWriter = ssa.NewHTMLWriter(ssaDF, s.f, ssaDumpCFG)
383 // TODO: generate and print a mapping from nodes to values and blocks
384 dumpSourcesColumn(s.f.HTMLWriter, fn)
385 s.f.HTMLWriter.WriteAST("AST", astBuf)
388 // Allocate starting values
389 s.labels = map[string]*ssaLabel{}
390 s.fwdVars = map[ir.Node]*ssa.Value{}
391 s.startmem = s.entryNewValue0(ssa.OpInitMem, types.TypeMem)
393 s.hasOpenDefers = base.Flag.N == 0 && s.hasdefer && !s.curfn.OpenCodedDeferDisallowed()
395 case base.Debug.NoOpenDefer != 0:
396 s.hasOpenDefers = false
397 case s.hasOpenDefers && (base.Ctxt.Flag_shared || base.Ctxt.Flag_dynlink) && base.Ctxt.Arch.Name == "386":
398 // Don't support open-coded defers for 386 ONLY when using shared
399 // libraries, because there is extra code (added by rewriteToUseGot())
400 // preceding the deferreturn/ret code that we don't track correctly.
401 s.hasOpenDefers = false
403 if s.hasOpenDefers && len(s.curfn.Exit) > 0 {
404 // Skip doing open defers if there is any extra exit code (likely
405 // race detection), since we will not generate that code in the
406 // case of the extra deferreturn/ret segment.
407 s.hasOpenDefers = false
410 // Similarly, skip if there are any heap-allocated result
411 // parameters that need to be copied back to their stack slots.
412 for _, f := range s.curfn.Type().Results().FieldSlice() {
413 if !f.Nname.(*ir.Name).OnStack() {
414 s.hasOpenDefers = false
419 if s.hasOpenDefers &&
420 s.curfn.NumReturns*s.curfn.NumDefers > 15 {
421 // Since we are generating defer calls at every exit for
422 // open-coded defers, skip doing open-coded defers if there are
423 // too many returns (especially if there are multiple defers).
424 // Open-coded defers are most important for improving performance
425 // for smaller functions (which don't have many returns).
426 s.hasOpenDefers = false
429 s.sp = s.entryNewValue0(ssa.OpSP, types.Types[types.TUINTPTR]) // TODO: use generic pointer type (unsafe.Pointer?) instead
430 s.sb = s.entryNewValue0(ssa.OpSB, types.Types[types.TUINTPTR])
432 s.startBlock(s.f.Entry)
433 s.vars[memVar] = s.startmem
435 // Create the deferBits variable and stack slot. deferBits is a
436 // bitmask showing which of the open-coded defers in this function
437 // have been activated.
438 deferBitsTemp := typecheck.TempAt(src.NoXPos, s.curfn, types.Types[types.TUINT8])
439 deferBitsTemp.SetAddrtaken(true)
440 s.deferBitsTemp = deferBitsTemp
441 // For this value, AuxInt is initialized to zero by default
442 startDeferBits := s.entryNewValue0(ssa.OpConst8, types.Types[types.TUINT8])
443 s.vars[deferBitsVar] = startDeferBits
444 s.deferBitsAddr = s.addr(deferBitsTemp)
445 s.store(types.Types[types.TUINT8], s.deferBitsAddr, startDeferBits)
446 // Make sure that the deferBits stack slot is kept alive (for use
447 // by panics) and stores to deferBits are not eliminated, even if
448 // all checking code on deferBits in the function exit can be
449 // eliminated, because the defer statements were all
451 s.vars[memVar] = s.newValue1Apos(ssa.OpVarLive, types.TypeMem, deferBitsTemp, s.mem(), false)
454 var params *abi.ABIParamResultInfo
455 params = s.f.ABISelf.ABIAnalyze(fn.Type(), true)
457 // The backend's stackframe pass prunes away entries from the fn's
458 // Dcl list, including PARAMOUT nodes that correspond to output
459 // params passed in registers. Walk the Dcl list and capture these
460 // nodes to a side list, so that we'll have them available during
461 // DWARF-gen later on. See issue 48573 for more details.
462 var debugInfo ssa.FuncDebug
463 for _, n := range fn.Dcl {
464 if n.Class == ir.PPARAMOUT && n.IsOutputParamInRegisters() {
465 debugInfo.RegOutputParams = append(debugInfo.RegOutputParams, n)
468 fn.DebugInfo = &debugInfo
470 // Generate addresses of local declarations
471 s.decladdrs = map[*ir.Name]*ssa.Value{}
472 for _, n := range fn.Dcl {
475 // Be aware that blank and unnamed input parameters will not appear here, but do appear in the type
476 s.decladdrs[n] = s.entryNewValue2A(ssa.OpLocalAddr, types.NewPtr(n.Type()), n, s.sp, s.startmem)
478 s.decladdrs[n] = s.entryNewValue2A(ssa.OpLocalAddr, types.NewPtr(n.Type()), n, s.sp, s.startmem)
480 // processed at each use, to prevent Addr coming
483 s.Fatalf("local variable with class %v unimplemented", n.Class)
487 s.f.OwnAux = ssa.OwnAuxCall(fn.LSym, params)
489 // Populate SSAable arguments.
490 for _, n := range fn.Dcl {
491 if n.Class == ir.PPARAM {
493 v := s.newValue0A(ssa.OpArg, n.Type(), n)
495 s.addNamedValue(n, v) // This helps with debugging information, not needed for compilation itself.
496 } else { // address was taken AND/OR too large for SSA
497 paramAssignment := ssa.ParamAssignmentForArgName(s.f, n)
498 if len(paramAssignment.Registers) > 0 {
499 if TypeOK(n.Type()) { // SSA-able type, so address was taken -- receive value in OpArg, DO NOT bind to var, store immediately to memory.
500 v := s.newValue0A(ssa.OpArg, n.Type(), n)
501 s.store(n.Type(), s.decladdrs[n], v)
502 } else { // Too big for SSA.
503 // Brute force, and early, do a bunch of stores from registers
504 // TODO fix the nasty storeArgOrLoad recursion in ssa/expand_calls.go so this Just Works with store of a big Arg.
505 s.storeParameterRegsToStack(s.f.ABISelf, paramAssignment, n, s.decladdrs[n], false)
512 // Populate closure variables.
514 clo := s.entryNewValue0(ssa.OpGetClosurePtr, s.f.Config.Types.BytePtr)
515 offset := int64(types.PtrSize) // PtrSize to skip past function entry PC field
516 for _, n := range fn.ClosureVars {
519 typ = types.NewPtr(typ)
522 offset = types.RoundUp(offset, typ.Alignment())
523 ptr := s.newValue1I(ssa.OpOffPtr, types.NewPtr(typ), offset, clo)
526 // If n is a small variable captured by value, promote
527 // it to PAUTO so it can be converted to SSA.
529 // Note: While we never capture a variable by value if
530 // the user took its address, we may have generated
531 // runtime calls that did (#43701). Since we don't
532 // convert Addrtaken variables to SSA anyway, no point
533 // in promoting them either.
534 if n.Byval() && !n.Addrtaken() && TypeOK(n.Type()) {
536 fn.Dcl = append(fn.Dcl, n)
537 s.assign(n, s.load(n.Type(), ptr), false, 0)
542 ptr = s.load(typ, ptr)
544 s.setHeapaddr(fn.Pos(), n, ptr)
548 // Convert the AST-based IR to the SSA-based IR
554 // fallthrough to exit
555 if s.curBlock != nil {
556 s.pushLine(fn.Endlineno)
561 for _, b := range s.f.Blocks {
562 if b.Pos != src.NoXPos {
563 s.updateUnsetPredPos(b)
567 s.f.HTMLWriter.WritePhase("before insert phis", "before insert phis")
571 // Main call to ssa package to compile function
574 if len(s.openDefers) != 0 {
575 s.emitOpenDeferInfo()
578 // Record incoming parameter spill information for morestack calls emitted in the assembler.
579 // This is done here, using all the parameters (used, partially used, and unused) because
580 // it mimics the behavior of the former ABI (everything stored) and because it's not 100%
581 // clear if naming conventions are respected in autogenerated code.
582 // TODO figure out exactly what's unused, don't spill it. Make liveness fine-grained, also.
583 for _, p := range params.InParams() {
584 typs, offs := p.RegisterTypesAndOffsets()
585 for i, t := range typs {
586 o := offs[i] // offset within parameter
587 fo := p.FrameOffset(params) // offset of parameter in frame
588 reg := ssa.ObjRegForAbiReg(p.Registers[i], s.f.Config)
589 s.f.RegArgs = append(s.f.RegArgs, ssa.Spill{Reg: reg, Offset: fo + o, Type: t})
596 func (s *state) storeParameterRegsToStack(abi *abi.ABIConfig, paramAssignment *abi.ABIParamAssignment, n *ir.Name, addr *ssa.Value, pointersOnly bool) {
597 typs, offs := paramAssignment.RegisterTypesAndOffsets()
598 for i, t := range typs {
599 if pointersOnly && !t.IsPtrShaped() {
602 r := paramAssignment.Registers[i]
604 op, reg := ssa.ArgOpAndRegisterFor(r, abi)
605 aux := &ssa.AuxNameOffset{Name: n, Offset: o}
606 v := s.newValue0I(op, t, reg)
608 p := s.newValue1I(ssa.OpOffPtr, types.NewPtr(t), o, addr)
613 // zeroResults zeros the return values at the start of the function.
614 // We need to do this very early in the function. Defer might stop a
615 // panic and show the return values as they exist at the time of
616 // panic. For precise stacks, the garbage collector assumes results
617 // are always live, so we need to zero them before any allocations,
618 // even allocations to move params/results to the heap.
619 func (s *state) zeroResults() {
620 for _, f := range s.curfn.Type().Results().FieldSlice() {
621 n := f.Nname.(*ir.Name)
623 // The local which points to the return value is the
624 // thing that needs zeroing. This is already handled
625 // by a Needzero annotation in plive.go:(*liveness).epilogue.
628 // Zero the stack location containing f.
629 if typ := n.Type(); TypeOK(typ) {
630 s.assign(n, s.zeroVal(typ), false, 0)
632 if typ.HasPointers() {
633 s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, n, s.mem())
635 s.zero(n.Type(), s.decladdrs[n])
640 // paramsToHeap produces code to allocate memory for heap-escaped parameters
641 // and to copy non-result parameters' values from the stack.
642 func (s *state) paramsToHeap() {
643 do := func(params *types.Type) {
644 for _, f := range params.FieldSlice() {
646 continue // anonymous or blank parameter
648 n := f.Nname.(*ir.Name)
649 if ir.IsBlank(n) || n.OnStack() {
653 if n.Class == ir.PPARAM {
654 s.move(n.Type(), s.expr(n.Heapaddr), s.decladdrs[n])
659 typ := s.curfn.Type()
665 // newHeapaddr allocates heap memory for n and sets its heap address.
666 func (s *state) newHeapaddr(n *ir.Name) {
667 s.setHeapaddr(n.Pos(), n, s.newObject(n.Type(), nil))
670 // setHeapaddr allocates a new PAUTO variable to store ptr (which must be non-nil)
671 // and then sets it as n's heap address.
672 func (s *state) setHeapaddr(pos src.XPos, n *ir.Name, ptr *ssa.Value) {
673 if !ptr.Type.IsPtr() || !types.Identical(n.Type(), ptr.Type.Elem()) {
674 base.FatalfAt(n.Pos(), "setHeapaddr %L with type %v", n, ptr.Type)
677 // Declare variable to hold address.
678 sym := &types.Sym{Name: "&" + n.Sym().Name, Pkg: types.LocalPkg}
679 addr := s.curfn.NewLocal(pos, sym, ir.PAUTO, types.NewPtr(n.Type()))
681 types.CalcSize(addr.Type())
683 if n.Class == ir.PPARAMOUT {
684 addr.SetIsOutputParamHeapAddr(true)
688 s.assign(addr, ptr, false, 0)
691 // newObject returns an SSA value denoting new(typ).
692 func (s *state) newObject(typ *types.Type, rtype *ssa.Value) *ssa.Value {
694 return s.newValue1A(ssa.OpAddr, types.NewPtr(typ), ir.Syms.Zerobase, s.sb)
697 rtype = s.reflectType(typ)
699 return s.rtcall(ir.Syms.Newobject, true, []*types.Type{types.NewPtr(typ)}, rtype)[0]
702 func (s *state) checkPtrAlignment(n *ir.ConvExpr, v *ssa.Value, count *ssa.Value) {
703 if !n.Type().IsPtr() {
704 s.Fatalf("expected pointer type: %v", n.Type())
706 elem, rtypeExpr := n.Type().Elem(), n.ElemRType
709 s.Fatalf("expected array type: %v", elem)
711 elem, rtypeExpr = elem.Elem(), n.ElemElemRType
714 // Casting from larger type to smaller one is ok, so for smallest type, do nothing.
715 if elem.Alignment() == 1 && (size == 0 || size == 1 || count == nil) {
719 count = s.constInt(types.Types[types.TUINTPTR], 1)
721 if count.Type.Size() != s.config.PtrSize {
722 s.Fatalf("expected count fit to a uintptr size, have: %d, want: %d", count.Type.Size(), s.config.PtrSize)
725 if rtypeExpr != nil {
726 rtype = s.expr(rtypeExpr)
728 rtype = s.reflectType(elem)
730 s.rtcall(ir.Syms.CheckPtrAlignment, true, nil, v, rtype, count)
733 // reflectType returns an SSA value representing a pointer to typ's
734 // reflection type descriptor.
735 func (s *state) reflectType(typ *types.Type) *ssa.Value {
736 // TODO(mdempsky): Make this Fatalf under Unified IR; frontend needs
737 // to supply RType expressions.
738 lsym := reflectdata.TypeLinksym(typ)
739 return s.entryNewValue1A(ssa.OpAddr, types.NewPtr(types.Types[types.TUINT8]), lsym, s.sb)
742 func dumpSourcesColumn(writer *ssa.HTMLWriter, fn *ir.Func) {
743 // Read sources of target function fn.
744 fname := base.Ctxt.PosTable.Pos(fn.Pos()).Filename()
745 targetFn, err := readFuncLines(fname, fn.Pos().Line(), fn.Endlineno.Line())
747 writer.Logf("cannot read sources for function %v: %v", fn, err)
750 // Read sources of inlined functions.
751 var inlFns []*ssa.FuncLines
752 for _, fi := range ssaDumpInlined {
754 fname := base.Ctxt.PosTable.Pos(fi.Pos()).Filename()
755 fnLines, err := readFuncLines(fname, fi.Pos().Line(), elno.Line())
757 writer.Logf("cannot read sources for inlined function %v: %v", fi, err)
760 inlFns = append(inlFns, fnLines)
763 sort.Sort(ssa.ByTopo(inlFns))
765 inlFns = append([]*ssa.FuncLines{targetFn}, inlFns...)
768 writer.WriteSources("sources", inlFns)
771 func readFuncLines(file string, start, end uint) (*ssa.FuncLines, error) {
772 f, err := os.Open(os.ExpandEnv(file))
779 scanner := bufio.NewScanner(f)
780 for scanner.Scan() && ln <= end {
782 lines = append(lines, scanner.Text())
786 return &ssa.FuncLines{Filename: file, StartLineno: start, Lines: lines}, nil
789 // updateUnsetPredPos propagates the earliest-value position information for b
790 // towards all of b's predecessors that need a position, and recurs on that
791 // predecessor if its position is updated. B should have a non-empty position.
792 func (s *state) updateUnsetPredPos(b *ssa.Block) {
793 if b.Pos == src.NoXPos {
794 s.Fatalf("Block %s should have a position", b)
796 bestPos := src.NoXPos
797 for _, e := range b.Preds {
802 if bestPos == src.NoXPos {
804 for _, v := range b.Values {
808 if v.Pos != src.NoXPos {
809 // Assume values are still in roughly textual order;
810 // TODO: could also seek minimum position?
817 s.updateUnsetPredPos(p) // We do not expect long chains of these, thus recursion is okay.
821 // Information about each open-coded defer.
822 type openDeferInfo struct {
823 // The node representing the call of the defer
825 // If defer call is closure call, the address of the argtmp where the
826 // closure is stored.
828 // The node representing the argtmp where the closure is stored - used for
829 // function, method, or interface call, to store a closure that panic
830 // processing can use for this defer.
835 // configuration (arch) information
838 // function we're building
845 labels map[string]*ssaLabel
847 // unlabeled break and continue statement tracking
848 breakTo *ssa.Block // current target for plain break statement
849 continueTo *ssa.Block // current target for plain continue statement
851 // current location where we're interpreting the AST
854 // variable assignments in the current block (map from variable symbol to ssa value)
855 // *Node is the unique identifier (an ONAME Node) for the variable.
856 // TODO: keep a single varnum map, then make all of these maps slices instead?
857 vars map[ir.Node]*ssa.Value
859 // fwdVars are variables that are used before they are defined in the current block.
860 // This map exists just to coalesce multiple references into a single FwdRef op.
861 // *Node is the unique identifier (an ONAME Node) for the variable.
862 fwdVars map[ir.Node]*ssa.Value
864 // all defined variables at the end of each block. Indexed by block ID.
865 defvars []map[ir.Node]*ssa.Value
867 // addresses of PPARAM and PPARAMOUT variables on the stack.
868 decladdrs map[*ir.Name]*ssa.Value
870 // starting values. Memory, stack pointer, and globals pointer
874 // value representing address of where deferBits autotmp is stored
875 deferBitsAddr *ssa.Value
876 deferBitsTemp *ir.Name
878 // line number stack. The current line number is top of stack
880 // the last line number processed; it may have been popped
883 // list of panic calls by function name and line number.
884 // Used to deduplicate panic calls.
885 panics map[funcLine]*ssa.Block
888 hasdefer bool // whether the function contains a defer statement
890 hasOpenDefers bool // whether we are doing open-coded defers
891 checkPtrEnabled bool // whether to insert checkptr instrumentation
893 // If doing open-coded defers, list of info about the defer calls in
894 // scanning order. Hence, at exit we should run these defers in reverse
895 // order of this list
896 openDefers []*openDeferInfo
897 // For open-coded defers, this is the beginning and end blocks of the last
898 // defer exit code that we have generated so far. We use these to share
899 // code between exits if the shareDeferExits option (disabled by default)
901 lastDeferExit *ssa.Block // Entry block of last defer exit code we generated
902 lastDeferFinalBlock *ssa.Block // Final block of last defer exit code we generated
903 lastDeferCount int // Number of defers encountered at that point
905 prevCall *ssa.Value // the previous call; use this to tie results to the call op.
908 type funcLine struct {
914 type ssaLabel struct {
915 target *ssa.Block // block identified by this label
916 breakTarget *ssa.Block // block to break to in control flow node identified by this label
917 continueTarget *ssa.Block // block to continue to in control flow node identified by this label
920 // label returns the label associated with sym, creating it if necessary.
921 func (s *state) label(sym *types.Sym) *ssaLabel {
922 lab := s.labels[sym.Name]
925 s.labels[sym.Name] = lab
930 func (s *state) Logf(msg string, args ...interface{}) { s.f.Logf(msg, args...) }
931 func (s *state) Log() bool { return s.f.Log() }
932 func (s *state) Fatalf(msg string, args ...interface{}) {
933 s.f.Frontend().Fatalf(s.peekPos(), msg, args...)
935 func (s *state) Warnl(pos src.XPos, msg string, args ...interface{}) { s.f.Warnl(pos, msg, args...) }
936 func (s *state) Debug_checknil() bool { return s.f.Frontend().Debug_checknil() }
938 func ssaMarker(name string) *ir.Name {
939 return ir.NewNameAt(base.Pos, &types.Sym{Name: name}, nil)
943 // marker node for the memory variable
944 memVar = ssaMarker("mem")
946 // marker nodes for temporary variables
947 ptrVar = ssaMarker("ptr")
948 lenVar = ssaMarker("len")
949 capVar = ssaMarker("cap")
950 typVar = ssaMarker("typ")
951 okVar = ssaMarker("ok")
952 deferBitsVar = ssaMarker("deferBits")
955 // startBlock sets the current block we're generating code in to b.
956 func (s *state) startBlock(b *ssa.Block) {
957 if s.curBlock != nil {
958 s.Fatalf("starting block %v when block %v has not ended", b, s.curBlock)
961 s.vars = map[ir.Node]*ssa.Value{}
962 for n := range s.fwdVars {
967 // endBlock marks the end of generating code for the current block.
968 // Returns the (former) current block. Returns nil if there is no current
969 // block, i.e. if no code flows to the current execution point.
970 func (s *state) endBlock() *ssa.Block {
975 for len(s.defvars) <= int(b.ID) {
976 s.defvars = append(s.defvars, nil)
978 s.defvars[b.ID] = s.vars
982 // Empty plain blocks get the line of their successor (handled after all blocks created),
983 // except for increment blocks in For statements (handled in ssa conversion of OFOR),
984 // and for blocks ending in GOTO/BREAK/CONTINUE.
992 // pushLine pushes a line number on the line number stack.
993 func (s *state) pushLine(line src.XPos) {
995 // the frontend may emit node with line number missing,
996 // use the parent line number in this case.
998 if base.Flag.K != 0 {
999 base.Warn("buildssa: unknown position (line 0)")
1005 s.line = append(s.line, line)
1008 // popLine pops the top of the line number stack.
1009 func (s *state) popLine() {
1010 s.line = s.line[:len(s.line)-1]
1013 // peekPos peeks the top of the line number stack.
1014 func (s *state) peekPos() src.XPos {
1015 return s.line[len(s.line)-1]
1018 // newValue0 adds a new value with no arguments to the current block.
1019 func (s *state) newValue0(op ssa.Op, t *types.Type) *ssa.Value {
1020 return s.curBlock.NewValue0(s.peekPos(), op, t)
1023 // newValue0A adds a new value with no arguments and an aux value to the current block.
1024 func (s *state) newValue0A(op ssa.Op, t *types.Type, aux ssa.Aux) *ssa.Value {
1025 return s.curBlock.NewValue0A(s.peekPos(), op, t, aux)
1028 // newValue0I adds a new value with no arguments and an auxint value to the current block.
1029 func (s *state) newValue0I(op ssa.Op, t *types.Type, auxint int64) *ssa.Value {
1030 return s.curBlock.NewValue0I(s.peekPos(), op, t, auxint)
1033 // newValue1 adds a new value with one argument to the current block.
1034 func (s *state) newValue1(op ssa.Op, t *types.Type, arg *ssa.Value) *ssa.Value {
1035 return s.curBlock.NewValue1(s.peekPos(), op, t, arg)
1038 // newValue1A adds a new value with one argument and an aux value to the current block.
1039 func (s *state) newValue1A(op ssa.Op, t *types.Type, aux ssa.Aux, arg *ssa.Value) *ssa.Value {
1040 return s.curBlock.NewValue1A(s.peekPos(), op, t, aux, arg)
1043 // newValue1Apos adds a new value with one argument and an aux value to the current block.
1044 // isStmt determines whether the created values may be a statement or not
1045 // (i.e., false means never, yes means maybe).
1046 func (s *state) newValue1Apos(op ssa.Op, t *types.Type, aux ssa.Aux, arg *ssa.Value, isStmt bool) *ssa.Value {
1048 return s.curBlock.NewValue1A(s.peekPos(), op, t, aux, arg)
1050 return s.curBlock.NewValue1A(s.peekPos().WithNotStmt(), op, t, aux, arg)
1053 // newValue1I adds a new value with one argument and an auxint value to the current block.
1054 func (s *state) newValue1I(op ssa.Op, t *types.Type, aux int64, arg *ssa.Value) *ssa.Value {
1055 return s.curBlock.NewValue1I(s.peekPos(), op, t, aux, arg)
1058 // newValue2 adds a new value with two arguments to the current block.
1059 func (s *state) newValue2(op ssa.Op, t *types.Type, arg0, arg1 *ssa.Value) *ssa.Value {
1060 return s.curBlock.NewValue2(s.peekPos(), op, t, arg0, arg1)
1063 // newValue2A adds a new value with two arguments and an aux value to the current block.
1064 func (s *state) newValue2A(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1 *ssa.Value) *ssa.Value {
1065 return s.curBlock.NewValue2A(s.peekPos(), op, t, aux, arg0, arg1)
1068 // newValue2Apos adds a new value with two arguments and an aux value to the current block.
1069 // isStmt determines whether the created values may be a statement or not
1070 // (i.e., false means never, yes means maybe).
1071 func (s *state) newValue2Apos(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1 *ssa.Value, isStmt bool) *ssa.Value {
1073 return s.curBlock.NewValue2A(s.peekPos(), op, t, aux, arg0, arg1)
1075 return s.curBlock.NewValue2A(s.peekPos().WithNotStmt(), op, t, aux, arg0, arg1)
1078 // newValue2I adds a new value with two arguments and an auxint value to the current block.
1079 func (s *state) newValue2I(op ssa.Op, t *types.Type, aux int64, arg0, arg1 *ssa.Value) *ssa.Value {
1080 return s.curBlock.NewValue2I(s.peekPos(), op, t, aux, arg0, arg1)
1083 // newValue3 adds a new value with three arguments to the current block.
1084 func (s *state) newValue3(op ssa.Op, t *types.Type, arg0, arg1, arg2 *ssa.Value) *ssa.Value {
1085 return s.curBlock.NewValue3(s.peekPos(), op, t, arg0, arg1, arg2)
1088 // newValue3I adds a new value with three arguments and an auxint value to the current block.
1089 func (s *state) newValue3I(op ssa.Op, t *types.Type, aux int64, arg0, arg1, arg2 *ssa.Value) *ssa.Value {
1090 return s.curBlock.NewValue3I(s.peekPos(), op, t, aux, arg0, arg1, arg2)
1093 // newValue3A adds a new value with three arguments and an aux value to the current block.
1094 func (s *state) newValue3A(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1, arg2 *ssa.Value) *ssa.Value {
1095 return s.curBlock.NewValue3A(s.peekPos(), op, t, aux, arg0, arg1, arg2)
1098 // newValue3Apos adds a new value with three arguments and an aux value to the current block.
1099 // isStmt determines whether the created values may be a statement or not
1100 // (i.e., false means never, yes means maybe).
1101 func (s *state) newValue3Apos(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1, arg2 *ssa.Value, isStmt bool) *ssa.Value {
1103 return s.curBlock.NewValue3A(s.peekPos(), op, t, aux, arg0, arg1, arg2)
1105 return s.curBlock.NewValue3A(s.peekPos().WithNotStmt(), op, t, aux, arg0, arg1, arg2)
1108 // newValue4 adds a new value with four arguments to the current block.
1109 func (s *state) newValue4(op ssa.Op, t *types.Type, arg0, arg1, arg2, arg3 *ssa.Value) *ssa.Value {
1110 return s.curBlock.NewValue4(s.peekPos(), op, t, arg0, arg1, arg2, arg3)
1113 // newValue4I adds a new value with four arguments and an auxint value to the current block.
1114 func (s *state) newValue4I(op ssa.Op, t *types.Type, aux int64, arg0, arg1, arg2, arg3 *ssa.Value) *ssa.Value {
1115 return s.curBlock.NewValue4I(s.peekPos(), op, t, aux, arg0, arg1, arg2, arg3)
1118 func (s *state) entryBlock() *ssa.Block {
1120 if base.Flag.N > 0 && s.curBlock != nil {
1121 // If optimizations are off, allocate in current block instead. Since with -N
1122 // we're not doing the CSE or tighten passes, putting lots of stuff in the
1123 // entry block leads to O(n^2) entries in the live value map during regalloc.
1130 // entryNewValue0 adds a new value with no arguments to the entry block.
1131 func (s *state) entryNewValue0(op ssa.Op, t *types.Type) *ssa.Value {
1132 return s.entryBlock().NewValue0(src.NoXPos, op, t)
1135 // entryNewValue0A adds a new value with no arguments and an aux value to the entry block.
1136 func (s *state) entryNewValue0A(op ssa.Op, t *types.Type, aux ssa.Aux) *ssa.Value {
1137 return s.entryBlock().NewValue0A(src.NoXPos, op, t, aux)
1140 // entryNewValue1 adds a new value with one argument to the entry block.
1141 func (s *state) entryNewValue1(op ssa.Op, t *types.Type, arg *ssa.Value) *ssa.Value {
1142 return s.entryBlock().NewValue1(src.NoXPos, op, t, arg)
1145 // entryNewValue1I adds a new value with one argument and an auxint value to the entry block.
1146 func (s *state) entryNewValue1I(op ssa.Op, t *types.Type, auxint int64, arg *ssa.Value) *ssa.Value {
1147 return s.entryBlock().NewValue1I(src.NoXPos, op, t, auxint, arg)
1150 // entryNewValue1A adds a new value with one argument and an aux value to the entry block.
1151 func (s *state) entryNewValue1A(op ssa.Op, t *types.Type, aux ssa.Aux, arg *ssa.Value) *ssa.Value {
1152 return s.entryBlock().NewValue1A(src.NoXPos, op, t, aux, arg)
1155 // entryNewValue2 adds a new value with two arguments to the entry block.
1156 func (s *state) entryNewValue2(op ssa.Op, t *types.Type, arg0, arg1 *ssa.Value) *ssa.Value {
1157 return s.entryBlock().NewValue2(src.NoXPos, op, t, arg0, arg1)
1160 // entryNewValue2A adds a new value with two arguments and an aux value to the entry block.
1161 func (s *state) entryNewValue2A(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1 *ssa.Value) *ssa.Value {
1162 return s.entryBlock().NewValue2A(src.NoXPos, op, t, aux, arg0, arg1)
1165 // const* routines add a new const value to the entry block.
1166 func (s *state) constSlice(t *types.Type) *ssa.Value {
1167 return s.f.ConstSlice(t)
1169 func (s *state) constInterface(t *types.Type) *ssa.Value {
1170 return s.f.ConstInterface(t)
1172 func (s *state) constNil(t *types.Type) *ssa.Value { return s.f.ConstNil(t) }
1173 func (s *state) constEmptyString(t *types.Type) *ssa.Value {
1174 return s.f.ConstEmptyString(t)
1176 func (s *state) constBool(c bool) *ssa.Value {
1177 return s.f.ConstBool(types.Types[types.TBOOL], c)
1179 func (s *state) constInt8(t *types.Type, c int8) *ssa.Value {
1180 return s.f.ConstInt8(t, c)
1182 func (s *state) constInt16(t *types.Type, c int16) *ssa.Value {
1183 return s.f.ConstInt16(t, c)
1185 func (s *state) constInt32(t *types.Type, c int32) *ssa.Value {
1186 return s.f.ConstInt32(t, c)
1188 func (s *state) constInt64(t *types.Type, c int64) *ssa.Value {
1189 return s.f.ConstInt64(t, c)
1191 func (s *state) constFloat32(t *types.Type, c float64) *ssa.Value {
1192 return s.f.ConstFloat32(t, c)
1194 func (s *state) constFloat64(t *types.Type, c float64) *ssa.Value {
1195 return s.f.ConstFloat64(t, c)
1197 func (s *state) constInt(t *types.Type, c int64) *ssa.Value {
1198 if s.config.PtrSize == 8 {
1199 return s.constInt64(t, c)
1201 if int64(int32(c)) != c {
1202 s.Fatalf("integer constant too big %d", c)
1204 return s.constInt32(t, int32(c))
1206 func (s *state) constOffPtrSP(t *types.Type, c int64) *ssa.Value {
1207 return s.f.ConstOffPtrSP(t, c, s.sp)
1210 // newValueOrSfCall* are wrappers around newValue*, which may create a call to a
1211 // soft-float runtime function instead (when emitting soft-float code).
1212 func (s *state) newValueOrSfCall1(op ssa.Op, t *types.Type, arg *ssa.Value) *ssa.Value {
1214 if c, ok := s.sfcall(op, arg); ok {
1218 return s.newValue1(op, t, arg)
1220 func (s *state) newValueOrSfCall2(op ssa.Op, t *types.Type, arg0, arg1 *ssa.Value) *ssa.Value {
1222 if c, ok := s.sfcall(op, arg0, arg1); ok {
1226 return s.newValue2(op, t, arg0, arg1)
1229 type instrumentKind uint8
1232 instrumentRead = iota
1237 func (s *state) instrument(t *types.Type, addr *ssa.Value, kind instrumentKind) {
1238 s.instrument2(t, addr, nil, kind)
1241 // instrumentFields instruments a read/write operation on addr.
1242 // If it is instrumenting for MSAN or ASAN and t is a struct type, it instruments
1243 // operation for each field, instead of for the whole struct.
1244 func (s *state) instrumentFields(t *types.Type, addr *ssa.Value, kind instrumentKind) {
1245 if !(base.Flag.MSan || base.Flag.ASan) || !t.IsStruct() {
1246 s.instrument(t, addr, kind)
1249 for _, f := range t.Fields().Slice() {
1250 if f.Sym.IsBlank() {
1253 offptr := s.newValue1I(ssa.OpOffPtr, types.NewPtr(f.Type), f.Offset, addr)
1254 s.instrumentFields(f.Type, offptr, kind)
1258 func (s *state) instrumentMove(t *types.Type, dst, src *ssa.Value) {
1260 s.instrument2(t, dst, src, instrumentMove)
1262 s.instrument(t, src, instrumentRead)
1263 s.instrument(t, dst, instrumentWrite)
1267 func (s *state) instrument2(t *types.Type, addr, addr2 *ssa.Value, kind instrumentKind) {
1268 if !s.curfn.InstrumentBody() {
1274 return // can't race on zero-sized things
1277 if ssa.IsSanitizerSafeAddr(addr) {
1284 if addr2 != nil && kind != instrumentMove {
1285 panic("instrument2: non-nil addr2 for non-move instrumentation")
1290 case instrumentRead:
1291 fn = ir.Syms.Msanread
1292 case instrumentWrite:
1293 fn = ir.Syms.Msanwrite
1294 case instrumentMove:
1295 fn = ir.Syms.Msanmove
1297 panic("unreachable")
1300 } else if base.Flag.Race && t.NumComponents(types.CountBlankFields) > 1 {
1301 // for composite objects we have to write every address
1302 // because a write might happen to any subobject.
1303 // composites with only one element don't have subobjects, though.
1305 case instrumentRead:
1306 fn = ir.Syms.Racereadrange
1307 case instrumentWrite:
1308 fn = ir.Syms.Racewriterange
1310 panic("unreachable")
1313 } else if base.Flag.Race {
1314 // for non-composite objects we can write just the start
1315 // address, as any write must write the first byte.
1317 case instrumentRead:
1318 fn = ir.Syms.Raceread
1319 case instrumentWrite:
1320 fn = ir.Syms.Racewrite
1322 panic("unreachable")
1324 } else if base.Flag.ASan {
1326 case instrumentRead:
1327 fn = ir.Syms.Asanread
1328 case instrumentWrite:
1329 fn = ir.Syms.Asanwrite
1331 panic("unreachable")
1335 panic("unreachable")
1338 args := []*ssa.Value{addr}
1340 args = append(args, addr2)
1343 args = append(args, s.constInt(types.Types[types.TUINTPTR], w))
1345 s.rtcall(fn, true, nil, args...)
1348 func (s *state) load(t *types.Type, src *ssa.Value) *ssa.Value {
1349 s.instrumentFields(t, src, instrumentRead)
1350 return s.rawLoad(t, src)
1353 func (s *state) rawLoad(t *types.Type, src *ssa.Value) *ssa.Value {
1354 return s.newValue2(ssa.OpLoad, t, src, s.mem())
1357 func (s *state) store(t *types.Type, dst, val *ssa.Value) {
1358 s.vars[memVar] = s.newValue3A(ssa.OpStore, types.TypeMem, t, dst, val, s.mem())
1361 func (s *state) zero(t *types.Type, dst *ssa.Value) {
1362 s.instrument(t, dst, instrumentWrite)
1363 store := s.newValue2I(ssa.OpZero, types.TypeMem, t.Size(), dst, s.mem())
1365 s.vars[memVar] = store
1368 func (s *state) move(t *types.Type, dst, src *ssa.Value) {
1369 s.moveWhichMayOverlap(t, dst, src, false)
1371 func (s *state) moveWhichMayOverlap(t *types.Type, dst, src *ssa.Value, mayOverlap bool) {
1372 s.instrumentMove(t, dst, src)
1373 if mayOverlap && t.IsArray() && t.NumElem() > 1 && !ssa.IsInlinableMemmove(dst, src, t.Size(), s.f.Config) {
1374 // Normally, when moving Go values of type T from one location to another,
1375 // we don't need to worry about partial overlaps. The two Ts must either be
1376 // in disjoint (nonoverlapping) memory or in exactly the same location.
1377 // There are 2 cases where this isn't true:
1378 // 1) Using unsafe you can arrange partial overlaps.
1379 // 2) Since Go 1.17, you can use a cast from a slice to a ptr-to-array.
1380 // https://go.dev/ref/spec#Conversions_from_slice_to_array_pointer
1381 // This feature can be used to construct partial overlaps of array types.
1383 // p := (*[2]int)(a[:])
1384 // q := (*[2]int)(a[1:])
1386 // We don't care about solving 1. Or at least, we haven't historically
1387 // and no one has complained.
1388 // For 2, we need to ensure that if there might be partial overlap,
1389 // then we can't use OpMove; we must use memmove instead.
1390 // (memmove handles partial overlap by copying in the correct
1391 // direction. OpMove does not.)
1393 // Note that we have to be careful here not to introduce a call when
1394 // we're marshaling arguments to a call or unmarshaling results from a call.
1395 // Cases where this is happening must pass mayOverlap to false.
1396 // (Currently this only happens when unmarshaling results of a call.)
1397 if t.HasPointers() {
1398 s.rtcall(ir.Syms.Typedmemmove, true, nil, s.reflectType(t), dst, src)
1399 // We would have otherwise implemented this move with straightline code,
1400 // including a write barrier. Pretend we issue a write barrier here,
1401 // so that the write barrier tests work. (Otherwise they'd need to know
1402 // the details of IsInlineableMemmove.)
1403 s.curfn.SetWBPos(s.peekPos())
1405 s.rtcall(ir.Syms.Memmove, true, nil, dst, src, s.constInt(types.Types[types.TUINTPTR], t.Size()))
1407 ssa.LogLargeCopy(s.f.Name, s.peekPos(), t.Size())
1410 store := s.newValue3I(ssa.OpMove, types.TypeMem, t.Size(), dst, src, s.mem())
1412 s.vars[memVar] = store
1415 // stmtList converts the statement list n to SSA and adds it to s.
1416 func (s *state) stmtList(l ir.Nodes) {
1417 for _, n := range l {
1422 // stmt converts the statement n to SSA and adds it to s.
1423 func (s *state) stmt(n ir.Node) {
1427 // If s.curBlock is nil, and n isn't a label (which might have an associated goto somewhere),
1428 // then this code is dead. Stop here.
1429 if s.curBlock == nil && n.Op() != ir.OLABEL {
1433 s.stmtList(n.Init())
1437 n := n.(*ir.BlockStmt)
1440 case ir.OFALL: // no-op
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)
2663 // We need to ensure []byte("") evaluates to []byte{}, and not []byte(nil).
2665 // TODO(mdempsky): Investigate using "len != 0" instead of "ptr != nil".
2666 cond := s.newValue2(ssa.OpNeqPtr, types.Types[types.TBOOL], ptr, s.constNil(ptr.Type))
2667 zerobase := s.newValue1A(ssa.OpAddr, ptr.Type, ir.Syms.Zerobase, s.sb)
2668 ptr = s.ternary(cond, ptr, zerobase)
2670 len := s.newValue1(ssa.OpStringLen, types.Types[types.TINT], str)
2671 return s.newValue3(ssa.OpSliceMake, n.Type(), ptr, len, len)
2673 n := n.(*ir.UnaryExpr)
2674 aux := n.X.(*ir.Name).Linksym()
2675 // OCFUNC is used to build function values, which must
2676 // always reference ABIInternal entry points.
2677 if aux.ABI() != obj.ABIInternal {
2678 s.Fatalf("expected ABIInternal: %v", aux.ABI())
2680 return s.entryNewValue1A(ssa.OpAddr, n.Type(), aux, s.sb)
2683 if n.Class == ir.PFUNC {
2684 // "value" of a function is the address of the function's closure
2685 sym := staticdata.FuncLinksym(n)
2686 return s.entryNewValue1A(ssa.OpAddr, types.NewPtr(n.Type()), sym, s.sb)
2689 return s.variable(n, n.Type())
2691 return s.load(n.Type(), s.addr(n))
2692 case ir.OLINKSYMOFFSET:
2693 n := n.(*ir.LinksymOffsetExpr)
2694 return s.load(n.Type(), s.addr(n))
2696 n := n.(*ir.NilExpr)
2700 return s.constSlice(t)
2701 case t.IsInterface():
2702 return s.constInterface(t)
2704 return s.constNil(t)
2707 switch u := n.Val(); u.Kind() {
2709 i := ir.IntVal(n.Type(), u)
2710 switch n.Type().Size() {
2712 return s.constInt8(n.Type(), int8(i))
2714 return s.constInt16(n.Type(), int16(i))
2716 return s.constInt32(n.Type(), int32(i))
2718 return s.constInt64(n.Type(), i)
2720 s.Fatalf("bad integer size %d", n.Type().Size())
2723 case constant.String:
2724 i := constant.StringVal(u)
2726 return s.constEmptyString(n.Type())
2728 return s.entryNewValue0A(ssa.OpConstString, n.Type(), ssa.StringToAux(i))
2730 return s.constBool(constant.BoolVal(u))
2731 case constant.Float:
2732 f, _ := constant.Float64Val(u)
2733 switch n.Type().Size() {
2735 return s.constFloat32(n.Type(), f)
2737 return s.constFloat64(n.Type(), f)
2739 s.Fatalf("bad float size %d", n.Type().Size())
2742 case constant.Complex:
2743 re, _ := constant.Float64Val(constant.Real(u))
2744 im, _ := constant.Float64Val(constant.Imag(u))
2745 switch n.Type().Size() {
2747 pt := types.Types[types.TFLOAT32]
2748 return s.newValue2(ssa.OpComplexMake, n.Type(),
2749 s.constFloat32(pt, re),
2750 s.constFloat32(pt, im))
2752 pt := types.Types[types.TFLOAT64]
2753 return s.newValue2(ssa.OpComplexMake, n.Type(),
2754 s.constFloat64(pt, re),
2755 s.constFloat64(pt, im))
2757 s.Fatalf("bad complex size %d", n.Type().Size())
2761 s.Fatalf("unhandled OLITERAL %v", u.Kind())
2765 n := n.(*ir.ConvExpr)
2769 // Assume everything will work out, so set up our return value.
2770 // Anything interesting that happens from here is a fatal.
2776 // Special case for not confusing GC and liveness.
2777 // We don't want pointers accidentally classified
2778 // as not-pointers or vice-versa because of copy
2780 if to.IsPtrShaped() != from.IsPtrShaped() {
2781 return s.newValue2(ssa.OpConvert, to, x, s.mem())
2784 v := s.newValue1(ssa.OpCopy, to, x) // ensure that v has the right type
2787 if to.Kind() == types.TFUNC && from.IsPtrShaped() {
2791 // named <--> unnamed type or typed <--> untyped const
2792 if from.Kind() == to.Kind() {
2796 // unsafe.Pointer <--> *T
2797 if to.IsUnsafePtr() && from.IsPtrShaped() || from.IsUnsafePtr() && to.IsPtrShaped() {
2798 if s.checkPtrEnabled && checkPtrOK && to.IsPtr() && from.IsUnsafePtr() {
2799 s.checkPtrAlignment(n, v, nil)
2805 if to.Kind() == types.TMAP && from.IsPtr() &&
2806 to.MapType().Hmap == from.Elem() {
2810 types.CalcSize(from)
2812 if from.Size() != to.Size() {
2813 s.Fatalf("CONVNOP width mismatch %v (%d) -> %v (%d)\n", from, from.Size(), to, to.Size())
2816 if etypesign(from.Kind()) != etypesign(to.Kind()) {
2817 s.Fatalf("CONVNOP sign mismatch %v (%s) -> %v (%s)\n", from, from.Kind(), to, to.Kind())
2821 if base.Flag.Cfg.Instrumenting {
2822 // These appear to be fine, but they fail the
2823 // integer constraint below, so okay them here.
2824 // Sample non-integer conversion: map[string]string -> *uint8
2828 if etypesign(from.Kind()) == 0 {
2829 s.Fatalf("CONVNOP unrecognized non-integer %v -> %v\n", from, to)
2833 // integer, same width, same sign
2837 n := n.(*ir.ConvExpr)
2839 return s.conv(n, x, n.X.Type(), n.Type())
2842 n := n.(*ir.TypeAssertExpr)
2843 res, _ := s.dottype(n, false)
2846 case ir.ODYNAMICDOTTYPE:
2847 n := n.(*ir.DynamicTypeAssertExpr)
2848 res, _ := s.dynamicDottype(n, false)
2852 case ir.OLT, ir.OEQ, ir.ONE, ir.OLE, ir.OGE, ir.OGT:
2853 n := n.(*ir.BinaryExpr)
2856 if n.X.Type().IsComplex() {
2857 pt := types.FloatForComplex(n.X.Type())
2858 op := s.ssaOp(ir.OEQ, pt)
2859 r := s.newValueOrSfCall2(op, types.Types[types.TBOOL], s.newValue1(ssa.OpComplexReal, pt, a), s.newValue1(ssa.OpComplexReal, pt, b))
2860 i := s.newValueOrSfCall2(op, types.Types[types.TBOOL], s.newValue1(ssa.OpComplexImag, pt, a), s.newValue1(ssa.OpComplexImag, pt, b))
2861 c := s.newValue2(ssa.OpAndB, types.Types[types.TBOOL], r, i)
2866 return s.newValue1(ssa.OpNot, types.Types[types.TBOOL], c)
2868 s.Fatalf("ordered complex compare %v", n.Op())
2872 // Convert OGE and OGT into OLE and OLT.
2876 op, a, b = ir.OLE, b, a
2878 op, a, b = ir.OLT, b, a
2880 if n.X.Type().IsFloat() {
2882 return s.newValueOrSfCall2(s.ssaOp(op, n.X.Type()), types.Types[types.TBOOL], a, b)
2884 // integer comparison
2885 return s.newValue2(s.ssaOp(op, n.X.Type()), types.Types[types.TBOOL], a, b)
2887 n := n.(*ir.BinaryExpr)
2890 if n.Type().IsComplex() {
2891 mulop := ssa.OpMul64F
2892 addop := ssa.OpAdd64F
2893 subop := ssa.OpSub64F
2894 pt := types.FloatForComplex(n.Type()) // Could be Float32 or Float64
2895 wt := types.Types[types.TFLOAT64] // Compute in Float64 to minimize cancellation error
2897 areal := s.newValue1(ssa.OpComplexReal, pt, a)
2898 breal := s.newValue1(ssa.OpComplexReal, pt, b)
2899 aimag := s.newValue1(ssa.OpComplexImag, pt, a)
2900 bimag := s.newValue1(ssa.OpComplexImag, pt, b)
2902 if pt != wt { // Widen for calculation
2903 areal = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, areal)
2904 breal = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, breal)
2905 aimag = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, aimag)
2906 bimag = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, bimag)
2909 xreal := s.newValueOrSfCall2(subop, wt, s.newValueOrSfCall2(mulop, wt, areal, breal), s.newValueOrSfCall2(mulop, wt, aimag, bimag))
2910 ximag := s.newValueOrSfCall2(addop, wt, s.newValueOrSfCall2(mulop, wt, areal, bimag), s.newValueOrSfCall2(mulop, wt, aimag, breal))
2912 if pt != wt { // Narrow to store back
2913 xreal = s.newValueOrSfCall1(ssa.OpCvt64Fto32F, pt, xreal)
2914 ximag = s.newValueOrSfCall1(ssa.OpCvt64Fto32F, pt, ximag)
2917 return s.newValue2(ssa.OpComplexMake, n.Type(), xreal, ximag)
2920 if n.Type().IsFloat() {
2921 return s.newValueOrSfCall2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2924 return s.newValue2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2927 n := n.(*ir.BinaryExpr)
2930 if n.Type().IsComplex() {
2931 // TODO this is not executed because the front-end substitutes a runtime call.
2932 // That probably ought to change; with modest optimization the widen/narrow
2933 // conversions could all be elided in larger expression trees.
2934 mulop := ssa.OpMul64F
2935 addop := ssa.OpAdd64F
2936 subop := ssa.OpSub64F
2937 divop := ssa.OpDiv64F
2938 pt := types.FloatForComplex(n.Type()) // Could be Float32 or Float64
2939 wt := types.Types[types.TFLOAT64] // Compute in Float64 to minimize cancellation error
2941 areal := s.newValue1(ssa.OpComplexReal, pt, a)
2942 breal := s.newValue1(ssa.OpComplexReal, pt, b)
2943 aimag := s.newValue1(ssa.OpComplexImag, pt, a)
2944 bimag := s.newValue1(ssa.OpComplexImag, pt, b)
2946 if pt != wt { // Widen for calculation
2947 areal = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, areal)
2948 breal = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, breal)
2949 aimag = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, aimag)
2950 bimag = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, bimag)
2953 denom := s.newValueOrSfCall2(addop, wt, s.newValueOrSfCall2(mulop, wt, breal, breal), s.newValueOrSfCall2(mulop, wt, bimag, bimag))
2954 xreal := s.newValueOrSfCall2(addop, wt, s.newValueOrSfCall2(mulop, wt, areal, breal), s.newValueOrSfCall2(mulop, wt, aimag, bimag))
2955 ximag := s.newValueOrSfCall2(subop, wt, s.newValueOrSfCall2(mulop, wt, aimag, breal), s.newValueOrSfCall2(mulop, wt, areal, bimag))
2957 // TODO not sure if this is best done in wide precision or narrow
2958 // Double-rounding might be an issue.
2959 // Note that the pre-SSA implementation does the entire calculation
2960 // in wide format, so wide is compatible.
2961 xreal = s.newValueOrSfCall2(divop, wt, xreal, denom)
2962 ximag = s.newValueOrSfCall2(divop, wt, ximag, denom)
2964 if pt != wt { // Narrow to store back
2965 xreal = s.newValueOrSfCall1(ssa.OpCvt64Fto32F, pt, xreal)
2966 ximag = s.newValueOrSfCall1(ssa.OpCvt64Fto32F, pt, ximag)
2968 return s.newValue2(ssa.OpComplexMake, n.Type(), xreal, ximag)
2970 if n.Type().IsFloat() {
2971 return s.newValueOrSfCall2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2973 return s.intDivide(n, a, b)
2975 n := n.(*ir.BinaryExpr)
2978 return s.intDivide(n, a, b)
2979 case ir.OADD, ir.OSUB:
2980 n := n.(*ir.BinaryExpr)
2983 if n.Type().IsComplex() {
2984 pt := types.FloatForComplex(n.Type())
2985 op := s.ssaOp(n.Op(), pt)
2986 return s.newValue2(ssa.OpComplexMake, n.Type(),
2987 s.newValueOrSfCall2(op, pt, s.newValue1(ssa.OpComplexReal, pt, a), s.newValue1(ssa.OpComplexReal, pt, b)),
2988 s.newValueOrSfCall2(op, pt, s.newValue1(ssa.OpComplexImag, pt, a), s.newValue1(ssa.OpComplexImag, pt, b)))
2990 if n.Type().IsFloat() {
2991 return s.newValueOrSfCall2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2993 return s.newValue2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2994 case ir.OAND, ir.OOR, ir.OXOR:
2995 n := n.(*ir.BinaryExpr)
2998 return s.newValue2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
3000 n := n.(*ir.BinaryExpr)
3003 b = s.newValue1(s.ssaOp(ir.OBITNOT, b.Type), b.Type, b)
3004 return s.newValue2(s.ssaOp(ir.OAND, n.Type()), a.Type, a, b)
3005 case ir.OLSH, ir.ORSH:
3006 n := n.(*ir.BinaryExpr)
3011 cmp := s.newValue2(s.ssaOp(ir.OLE, bt), types.Types[types.TBOOL], s.zeroVal(bt), b)
3012 s.check(cmp, ir.Syms.Panicshift)
3013 bt = bt.ToUnsigned()
3015 return s.newValue2(s.ssaShiftOp(n.Op(), n.Type(), bt), a.Type, a, b)
3016 case ir.OANDAND, ir.OOROR:
3017 // To implement OANDAND (and OOROR), we introduce a
3018 // new temporary variable to hold the result. The
3019 // variable is associated with the OANDAND node in the
3020 // s.vars table (normally variables are only
3021 // associated with ONAME nodes). We convert
3028 // Using var in the subsequent block introduces the
3029 // necessary phi variable.
3030 n := n.(*ir.LogicalExpr)
3035 b.Kind = ssa.BlockIf
3037 // In theory, we should set b.Likely here based on context.
3038 // However, gc only gives us likeliness hints
3039 // in a single place, for plain OIF statements,
3040 // and passing around context is finnicky, so don't bother for now.
3042 bRight := s.f.NewBlock(ssa.BlockPlain)
3043 bResult := s.f.NewBlock(ssa.BlockPlain)
3044 if n.Op() == ir.OANDAND {
3046 b.AddEdgeTo(bResult)
3047 } else if n.Op() == ir.OOROR {
3048 b.AddEdgeTo(bResult)
3052 s.startBlock(bRight)
3057 b.AddEdgeTo(bResult)
3059 s.startBlock(bResult)
3060 return s.variable(n, types.Types[types.TBOOL])
3062 n := n.(*ir.BinaryExpr)
3065 return s.newValue2(ssa.OpComplexMake, n.Type(), r, i)
3069 n := n.(*ir.UnaryExpr)
3071 if n.Type().IsComplex() {
3072 tp := types.FloatForComplex(n.Type())
3073 negop := s.ssaOp(n.Op(), tp)
3074 return s.newValue2(ssa.OpComplexMake, n.Type(),
3075 s.newValue1(negop, tp, s.newValue1(ssa.OpComplexReal, tp, a)),
3076 s.newValue1(negop, tp, s.newValue1(ssa.OpComplexImag, tp, a)))
3078 return s.newValue1(s.ssaOp(n.Op(), n.Type()), a.Type, a)
3079 case ir.ONOT, ir.OBITNOT:
3080 n := n.(*ir.UnaryExpr)
3082 return s.newValue1(s.ssaOp(n.Op(), n.Type()), a.Type, a)
3083 case ir.OIMAG, ir.OREAL:
3084 n := n.(*ir.UnaryExpr)
3086 return s.newValue1(s.ssaOp(n.Op(), n.X.Type()), n.Type(), a)
3088 n := n.(*ir.UnaryExpr)
3092 n := n.(*ir.AddrExpr)
3096 n := n.(*ir.ResultExpr)
3097 if s.prevCall == nil || s.prevCall.Op != ssa.OpStaticLECall && s.prevCall.Op != ssa.OpInterLECall && s.prevCall.Op != ssa.OpClosureLECall {
3098 panic("Expected to see a previous call")
3102 panic(fmt.Errorf("ORESULT %v does not match call %s", n, s.prevCall))
3104 return s.resultOfCall(s.prevCall, which, n.Type())
3107 n := n.(*ir.StarExpr)
3108 p := s.exprPtr(n.X, n.Bounded(), n.Pos())
3109 return s.load(n.Type(), p)
3112 n := n.(*ir.SelectorExpr)
3113 if n.X.Op() == ir.OSTRUCTLIT {
3114 // All literals with nonzero fields have already been
3115 // rewritten during walk. Any that remain are just T{}
3116 // or equivalents. Use the zero value.
3117 if !ir.IsZero(n.X) {
3118 s.Fatalf("literal with nonzero value in SSA: %v", n.X)
3120 return s.zeroVal(n.Type())
3122 // If n is addressable and can't be represented in
3123 // SSA, then load just the selected field. This
3124 // prevents false memory dependencies in race/msan/asan
3126 if ir.IsAddressable(n) && !s.canSSA(n) {
3128 return s.load(n.Type(), p)
3131 return s.newValue1I(ssa.OpStructSelect, n.Type(), int64(fieldIdx(n)), v)
3134 n := n.(*ir.SelectorExpr)
3135 p := s.exprPtr(n.X, n.Bounded(), n.Pos())
3136 p = s.newValue1I(ssa.OpOffPtr, types.NewPtr(n.Type()), n.Offset(), p)
3137 return s.load(n.Type(), p)
3140 n := n.(*ir.IndexExpr)
3142 case n.X.Type().IsString():
3143 if n.Bounded() && ir.IsConst(n.X, constant.String) && ir.IsConst(n.Index, constant.Int) {
3144 // Replace "abc"[1] with 'b'.
3145 // Delayed until now because "abc"[1] is not an ideal constant.
3146 // See test/fixedbugs/issue11370.go.
3147 return s.newValue0I(ssa.OpConst8, types.Types[types.TUINT8], int64(int8(ir.StringVal(n.X)[ir.Int64Val(n.Index)])))
3150 i := s.expr(n.Index)
3151 len := s.newValue1(ssa.OpStringLen, types.Types[types.TINT], a)
3152 i = s.boundsCheck(i, len, ssa.BoundsIndex, n.Bounded())
3153 ptrtyp := s.f.Config.Types.BytePtr
3154 ptr := s.newValue1(ssa.OpStringPtr, ptrtyp, a)
3155 if ir.IsConst(n.Index, constant.Int) {
3156 ptr = s.newValue1I(ssa.OpOffPtr, ptrtyp, ir.Int64Val(n.Index), ptr)
3158 ptr = s.newValue2(ssa.OpAddPtr, ptrtyp, ptr, i)
3160 return s.load(types.Types[types.TUINT8], ptr)
3161 case n.X.Type().IsSlice():
3163 return s.load(n.X.Type().Elem(), p)
3164 case n.X.Type().IsArray():
3165 if TypeOK(n.X.Type()) {
3166 // SSA can handle arrays of length at most 1.
3167 bound := n.X.Type().NumElem()
3169 i := s.expr(n.Index)
3171 // Bounds check will never succeed. Might as well
3172 // use constants for the bounds check.
3173 z := s.constInt(types.Types[types.TINT], 0)
3174 s.boundsCheck(z, z, ssa.BoundsIndex, false)
3175 // The return value won't be live, return junk.
3176 // But not quite junk, in case bounds checks are turned off. See issue 48092.
3177 return s.zeroVal(n.Type())
3179 len := s.constInt(types.Types[types.TINT], bound)
3180 s.boundsCheck(i, len, ssa.BoundsIndex, n.Bounded()) // checks i == 0
3181 return s.newValue1I(ssa.OpArraySelect, n.Type(), 0, a)
3184 return s.load(n.X.Type().Elem(), p)
3186 s.Fatalf("bad type for index %v", n.X.Type())
3190 case ir.OLEN, ir.OCAP:
3191 n := n.(*ir.UnaryExpr)
3193 case n.X.Type().IsSlice():
3194 op := ssa.OpSliceLen
3195 if n.Op() == ir.OCAP {
3198 return s.newValue1(op, types.Types[types.TINT], s.expr(n.X))
3199 case n.X.Type().IsString(): // string; not reachable for OCAP
3200 return s.newValue1(ssa.OpStringLen, types.Types[types.TINT], s.expr(n.X))
3201 case n.X.Type().IsMap(), n.X.Type().IsChan():
3202 return s.referenceTypeBuiltin(n, s.expr(n.X))
3204 return s.constInt(types.Types[types.TINT], n.X.Type().NumElem())
3208 n := n.(*ir.UnaryExpr)
3210 if n.X.Type().IsSlice() {
3212 return s.newValue1(ssa.OpSlicePtr, n.Type(), a)
3214 return s.newValue1(ssa.OpSlicePtrUnchecked, n.Type(), a)
3216 return s.newValue1(ssa.OpStringPtr, n.Type(), a)
3220 n := n.(*ir.UnaryExpr)
3222 return s.newValue1(ssa.OpITab, n.Type(), a)
3225 n := n.(*ir.UnaryExpr)
3227 return s.newValue1(ssa.OpIData, n.Type(), a)
3230 n := n.(*ir.BinaryExpr)
3233 return s.newValue2(ssa.OpIMake, n.Type(), tab, data)
3235 case ir.OSLICEHEADER:
3236 n := n.(*ir.SliceHeaderExpr)
3240 return s.newValue3(ssa.OpSliceMake, n.Type(), p, l, c)
3242 case ir.OSTRINGHEADER:
3243 n := n.(*ir.StringHeaderExpr)
3246 return s.newValue2(ssa.OpStringMake, n.Type(), p, l)
3248 case ir.OSLICE, ir.OSLICEARR, ir.OSLICE3, ir.OSLICE3ARR:
3249 n := n.(*ir.SliceExpr)
3250 check := s.checkPtrEnabled && n.Op() == ir.OSLICE3ARR && n.X.Op() == ir.OCONVNOP && n.X.(*ir.ConvExpr).X.Type().IsUnsafePtr()
3251 v := s.exprCheckPtr(n.X, !check)
3252 var i, j, k *ssa.Value
3262 p, l, c := s.slice(v, i, j, k, n.Bounded())
3264 // Emit checkptr instrumentation after bound check to prevent false positive, see #46938.
3265 s.checkPtrAlignment(n.X.(*ir.ConvExpr), v, s.conv(n.Max, k, k.Type, types.Types[types.TUINTPTR]))
3267 return s.newValue3(ssa.OpSliceMake, n.Type(), p, l, c)
3270 n := n.(*ir.SliceExpr)
3279 p, l, _ := s.slice(v, i, j, nil, n.Bounded())
3280 return s.newValue2(ssa.OpStringMake, n.Type(), p, l)
3282 case ir.OSLICE2ARRPTR:
3283 // if arrlen > slice.len {
3287 n := n.(*ir.ConvExpr)
3289 nelem := n.Type().Elem().NumElem()
3290 arrlen := s.constInt(types.Types[types.TINT], nelem)
3291 cap := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], v)
3292 s.boundsCheck(arrlen, cap, ssa.BoundsConvert, false)
3293 op := ssa.OpSlicePtr
3295 op = ssa.OpSlicePtrUnchecked
3297 return s.newValue1(op, n.Type(), v)
3300 n := n.(*ir.CallExpr)
3301 if ir.IsIntrinsicCall(n) {
3302 return s.intrinsicCall(n)
3307 n := n.(*ir.CallExpr)
3308 return s.callResult(n, callNormal)
3311 n := n.(*ir.CallExpr)
3312 return s.newValue1(ssa.OpGetG, n.Type(), s.mem())
3314 case ir.OGETCALLERPC:
3315 n := n.(*ir.CallExpr)
3316 return s.newValue0(ssa.OpGetCallerPC, n.Type())
3318 case ir.OGETCALLERSP:
3319 n := n.(*ir.CallExpr)
3320 return s.newValue1(ssa.OpGetCallerSP, n.Type(), s.mem())
3323 return s.append(n.(*ir.CallExpr), false)
3325 case ir.OMIN, ir.OMAX:
3326 return s.minMax(n.(*ir.CallExpr))
3328 case ir.OSTRUCTLIT, ir.OARRAYLIT:
3329 // All literals with nonzero fields have already been
3330 // rewritten during walk. Any that remain are just T{}
3331 // or equivalents. Use the zero value.
3332 n := n.(*ir.CompLitExpr)
3334 s.Fatalf("literal with nonzero value in SSA: %v", n)
3336 return s.zeroVal(n.Type())
3339 n := n.(*ir.UnaryExpr)
3340 var rtype *ssa.Value
3341 if x, ok := n.X.(*ir.DynamicType); ok && x.Op() == ir.ODYNAMICTYPE {
3342 rtype = s.expr(x.RType)
3344 return s.newObject(n.Type().Elem(), rtype)
3347 n := n.(*ir.BinaryExpr)
3351 // Force len to uintptr to prevent misuse of garbage bits in the
3352 // upper part of the register (#48536).
3353 len = s.conv(n, len, len.Type, types.Types[types.TUINTPTR])
3355 return s.newValue2(ssa.OpAddPtr, n.Type(), ptr, len)
3358 s.Fatalf("unhandled expr %v", n.Op())
3363 func (s *state) resultOfCall(c *ssa.Value, which int64, t *types.Type) *ssa.Value {
3364 aux := c.Aux.(*ssa.AuxCall)
3365 pa := aux.ParamAssignmentForResult(which)
3366 // TODO(register args) determine if in-memory TypeOK is better loaded early from SelectNAddr or later when SelectN is expanded.
3367 // SelectN is better for pattern-matching and possible call-aware analysis we might want to do in the future.
3368 if len(pa.Registers) == 0 && !TypeOK(t) {
3369 addr := s.newValue1I(ssa.OpSelectNAddr, types.NewPtr(t), which, c)
3370 return s.rawLoad(t, addr)
3372 return s.newValue1I(ssa.OpSelectN, t, which, c)
3375 func (s *state) resultAddrOfCall(c *ssa.Value, which int64, t *types.Type) *ssa.Value {
3376 aux := c.Aux.(*ssa.AuxCall)
3377 pa := aux.ParamAssignmentForResult(which)
3378 if len(pa.Registers) == 0 {
3379 return s.newValue1I(ssa.OpSelectNAddr, types.NewPtr(t), which, c)
3381 _, addr := s.temp(c.Pos, t)
3382 rval := s.newValue1I(ssa.OpSelectN, t, which, c)
3383 s.vars[memVar] = s.newValue3Apos(ssa.OpStore, types.TypeMem, t, addr, rval, s.mem(), false)
3387 // append converts an OAPPEND node to SSA.
3388 // If inplace is false, it converts the OAPPEND expression n to an ssa.Value,
3389 // adds it to s, and returns the Value.
3390 // If inplace is true, it writes the result of the OAPPEND expression n
3391 // back to the slice being appended to, and returns nil.
3392 // inplace MUST be set to false if the slice can be SSA'd.
3393 // Note: this code only handles fixed-count appends. Dotdotdot appends
3394 // have already been rewritten at this point (by walk).
3395 func (s *state) append(n *ir.CallExpr, inplace bool) *ssa.Value {
3396 // If inplace is false, process as expression "append(s, e1, e2, e3)":
3398 // ptr, len, cap := s
3400 // if uint(len) > uint(cap) {
3401 // ptr, len, cap = growslice(ptr, len, cap, 3, typ)
3402 // Note that len is unmodified by growslice.
3404 // // with write barriers, if needed:
3405 // *(ptr+(len-3)) = e1
3406 // *(ptr+(len-2)) = e2
3407 // *(ptr+(len-1)) = e3
3408 // return makeslice(ptr, len, cap)
3411 // If inplace is true, process as statement "s = append(s, e1, e2, e3)":
3414 // ptr, len, cap := s
3416 // if uint(len) > uint(cap) {
3417 // ptr, len, cap = growslice(ptr, len, cap, 3, typ)
3418 // vardef(a) // if necessary, advise liveness we are writing a new a
3419 // *a.cap = cap // write before ptr to avoid a spill
3420 // *a.ptr = ptr // with write barrier
3423 // // with write barriers, if needed:
3424 // *(ptr+(len-3)) = e1
3425 // *(ptr+(len-2)) = e2
3426 // *(ptr+(len-1)) = e3
3428 et := n.Type().Elem()
3429 pt := types.NewPtr(et)
3432 sn := n.Args[0] // the slice node is the first in the list
3433 var slice, addr *ssa.Value
3436 slice = s.load(n.Type(), addr)
3441 // Allocate new blocks
3442 grow := s.f.NewBlock(ssa.BlockPlain)
3443 assign := s.f.NewBlock(ssa.BlockPlain)
3445 // Decomposse input slice.
3446 p := s.newValue1(ssa.OpSlicePtr, pt, slice)
3447 l := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], slice)
3448 c := s.newValue1(ssa.OpSliceCap, types.Types[types.TINT], slice)
3450 // Add number of new elements to length.
3451 nargs := s.constInt(types.Types[types.TINT], int64(len(n.Args)-1))
3452 l = s.newValue2(s.ssaOp(ir.OADD, types.Types[types.TINT]), types.Types[types.TINT], l, nargs)
3454 // Decide if we need to grow
3455 cmp := s.newValue2(s.ssaOp(ir.OLT, types.Types[types.TUINT]), types.Types[types.TBOOL], c, l)
3457 // Record values of ptr/len/cap before branch.
3465 b.Kind = ssa.BlockIf
3466 b.Likely = ssa.BranchUnlikely
3473 taddr := s.expr(n.X)
3474 r := s.rtcall(ir.Syms.Growslice, true, []*types.Type{n.Type()}, p, l, c, nargs, taddr)
3476 // Decompose output slice
3477 p = s.newValue1(ssa.OpSlicePtr, pt, r[0])
3478 l = s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], r[0])
3479 c = s.newValue1(ssa.OpSliceCap, types.Types[types.TINT], r[0])
3485 if sn.Op() == ir.ONAME {
3487 if sn.Class != ir.PEXTERN {
3488 // Tell liveness we're about to build a new slice
3489 s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, sn, s.mem())
3492 capaddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, types.SliceCapOffset, addr)
3493 s.store(types.Types[types.TINT], capaddr, c)
3494 s.store(pt, addr, p)
3500 // assign new elements to slots
3501 s.startBlock(assign)
3502 p = s.variable(ptrVar, pt) // generates phi for ptr
3503 l = s.variable(lenVar, types.Types[types.TINT]) // generates phi for len
3505 c = s.variable(capVar, types.Types[types.TINT]) // generates phi for cap
3509 // Update length in place.
3510 // We have to wait until here to make sure growslice succeeded.
3511 lenaddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, types.SliceLenOffset, addr)
3512 s.store(types.Types[types.TINT], lenaddr, l)
3516 type argRec struct {
3517 // if store is true, we're appending the value v. If false, we're appending the
3522 args := make([]argRec, 0, len(n.Args[1:]))
3523 for _, n := range n.Args[1:] {
3524 if TypeOK(n.Type()) {
3525 args = append(args, argRec{v: s.expr(n), store: true})
3528 args = append(args, argRec{v: v})
3532 // Write args into slice.
3533 oldLen := s.newValue2(s.ssaOp(ir.OSUB, types.Types[types.TINT]), types.Types[types.TINT], l, nargs)
3534 p2 := s.newValue2(ssa.OpPtrIndex, pt, p, oldLen)
3535 for i, arg := range args {
3536 addr := s.newValue2(ssa.OpPtrIndex, pt, p2, s.constInt(types.Types[types.TINT], int64(i)))
3538 s.storeType(et, addr, arg.v, 0, true)
3540 s.move(et, addr, arg.v)
3544 // The following deletions have no practical effect at this time
3545 // because state.vars has been reset by the preceding state.startBlock.
3546 // They only enforce the fact that these variables are no longer need in
3547 // the current scope.
3548 delete(s.vars, ptrVar)
3549 delete(s.vars, lenVar)
3551 delete(s.vars, capVar)
3558 return s.newValue3(ssa.OpSliceMake, n.Type(), p, l, c)
3561 // minMax converts an OMIN/OMAX builtin call into SSA.
3562 func (s *state) minMax(n *ir.CallExpr) *ssa.Value {
3563 // The OMIN/OMAX builtin is variadic, but its semantics are
3564 // equivalent to left-folding a binary min/max operation across the
3566 fold := func(op func(x, a *ssa.Value) *ssa.Value) *ssa.Value {
3567 x := s.expr(n.Args[0])
3568 for _, arg := range n.Args[1:] {
3569 x = op(x, s.expr(arg))
3576 if typ.IsFloat() || typ.IsString() {
3577 // min/max semantics for floats are tricky because of NaNs and
3578 // negative zero. Some architectures have instructions which
3579 // we can use to generate the right result. For others we must
3580 // call into the runtime instead.
3582 // Strings are conceptually simpler, but we currently desugar
3583 // string comparisons during walk, not ssagen.
3586 switch Arch.LinkArch.Family {
3587 case sys.AMD64, sys.ARM64:
3590 case typ.Kind() == types.TFLOAT64 && n.Op() == ir.OMIN:
3592 case typ.Kind() == types.TFLOAT64 && n.Op() == ir.OMAX:
3594 case typ.Kind() == types.TFLOAT32 && n.Op() == ir.OMIN:
3596 case typ.Kind() == types.TFLOAT32 && n.Op() == ir.OMAX:
3599 return fold(func(x, a *ssa.Value) *ssa.Value {
3600 return s.newValue2(op, typ, x, a)
3606 case types.TFLOAT32:
3613 case types.TFLOAT64:
3628 fn := typecheck.LookupRuntimeFunc(name)
3630 return fold(func(x, a *ssa.Value) *ssa.Value {
3631 return s.rtcall(fn, true, []*types.Type{typ}, x, a)[0]
3635 lt := s.ssaOp(ir.OLT, typ)
3637 return fold(func(x, a *ssa.Value) *ssa.Value {
3641 return s.ternary(s.newValue2(lt, types.Types[types.TBOOL], a, x), a, x)
3644 return s.ternary(s.newValue2(lt, types.Types[types.TBOOL], x, a), a, x)
3646 panic("unreachable")
3650 // ternary emits code to evaluate cond ? x : y.
3651 func (s *state) ternary(cond, x, y *ssa.Value) *ssa.Value {
3652 // Note that we need a new ternaryVar each time (unlike okVar where we can
3653 // reuse the variable) because it might have a different type every time.
3654 ternaryVar := ssaMarker("ternary")
3656 bThen := s.f.NewBlock(ssa.BlockPlain)
3657 bElse := s.f.NewBlock(ssa.BlockPlain)
3658 bEnd := s.f.NewBlock(ssa.BlockPlain)
3661 b.Kind = ssa.BlockIf
3667 s.vars[ternaryVar] = x
3668 s.endBlock().AddEdgeTo(bEnd)
3671 s.vars[ternaryVar] = y
3672 s.endBlock().AddEdgeTo(bEnd)
3675 r := s.variable(ternaryVar, x.Type)
3676 delete(s.vars, ternaryVar)
3680 // condBranch evaluates the boolean expression cond and branches to yes
3681 // if cond is true and no if cond is false.
3682 // This function is intended to handle && and || better than just calling
3683 // s.expr(cond) and branching on the result.
3684 func (s *state) condBranch(cond ir.Node, yes, no *ssa.Block, likely int8) {
3687 cond := cond.(*ir.LogicalExpr)
3688 mid := s.f.NewBlock(ssa.BlockPlain)
3689 s.stmtList(cond.Init())
3690 s.condBranch(cond.X, mid, no, max8(likely, 0))
3692 s.condBranch(cond.Y, yes, no, likely)
3694 // Note: if likely==1, then both recursive calls pass 1.
3695 // If likely==-1, then we don't have enough information to decide
3696 // whether the first branch is likely or not. So we pass 0 for
3697 // the likeliness of the first branch.
3698 // TODO: have the frontend give us branch prediction hints for
3699 // OANDAND and OOROR nodes (if it ever has such info).
3701 cond := cond.(*ir.LogicalExpr)
3702 mid := s.f.NewBlock(ssa.BlockPlain)
3703 s.stmtList(cond.Init())
3704 s.condBranch(cond.X, yes, mid, min8(likely, 0))
3706 s.condBranch(cond.Y, yes, no, likely)
3708 // Note: if likely==-1, then both recursive calls pass -1.
3709 // If likely==1, then we don't have enough info to decide
3710 // the likelihood of the first branch.
3712 cond := cond.(*ir.UnaryExpr)
3713 s.stmtList(cond.Init())
3714 s.condBranch(cond.X, no, yes, -likely)
3717 cond := cond.(*ir.ConvExpr)
3718 s.stmtList(cond.Init())
3719 s.condBranch(cond.X, yes, no, likely)
3724 b.Kind = ssa.BlockIf
3726 b.Likely = ssa.BranchPrediction(likely) // gc and ssa both use -1/0/+1 for likeliness
3734 skipPtr skipMask = 1 << iota
3739 // assign does left = right.
3740 // Right has already been evaluated to ssa, left has not.
3741 // If deref is true, then we do left = *right instead (and right has already been nil-checked).
3742 // If deref is true and right == nil, just do left = 0.
3743 // skip indicates assignments (at the top level) that can be avoided.
3744 // mayOverlap indicates whether left&right might partially overlap in memory. Default is false.
3745 func (s *state) assign(left ir.Node, right *ssa.Value, deref bool, skip skipMask) {
3746 s.assignWhichMayOverlap(left, right, deref, skip, false)
3748 func (s *state) assignWhichMayOverlap(left ir.Node, right *ssa.Value, deref bool, skip skipMask, mayOverlap bool) {
3749 if left.Op() == ir.ONAME && ir.IsBlank(left) {
3756 s.Fatalf("can SSA LHS %v but not RHS %s", left, right)
3758 if left.Op() == ir.ODOT {
3759 // We're assigning to a field of an ssa-able value.
3760 // We need to build a new structure with the new value for the
3761 // field we're assigning and the old values for the other fields.
3763 // type T struct {a, b, c int}
3766 // For the x.b = 5 assignment we want to generate x = T{x.a, 5, x.c}
3768 // Grab information about the structure type.
3769 left := left.(*ir.SelectorExpr)
3772 idx := fieldIdx(left)
3774 // Grab old value of structure.
3775 old := s.expr(left.X)
3777 // Make new structure.
3778 new := s.newValue0(ssa.StructMakeOp(t.NumFields()), t)
3780 // Add fields as args.
3781 for i := 0; i < nf; i++ {
3785 new.AddArg(s.newValue1I(ssa.OpStructSelect, t.FieldType(i), int64(i), old))
3789 // Recursively assign the new value we've made to the base of the dot op.
3790 s.assign(left.X, new, false, 0)
3791 // TODO: do we need to update named values here?
3794 if left.Op() == ir.OINDEX && left.(*ir.IndexExpr).X.Type().IsArray() {
3795 left := left.(*ir.IndexExpr)
3796 s.pushLine(left.Pos())
3798 // We're assigning to an element of an ssa-able array.
3803 i := s.expr(left.Index) // index
3805 // The bounds check must fail. Might as well
3806 // ignore the actual index and just use zeros.
3807 z := s.constInt(types.Types[types.TINT], 0)
3808 s.boundsCheck(z, z, ssa.BoundsIndex, false)
3812 s.Fatalf("assigning to non-1-length array")
3814 // Rewrite to a = [1]{v}
3815 len := s.constInt(types.Types[types.TINT], 1)
3816 s.boundsCheck(i, len, ssa.BoundsIndex, false) // checks i == 0
3817 v := s.newValue1(ssa.OpArrayMake1, t, right)
3818 s.assign(left.X, v, false, 0)
3821 left := left.(*ir.Name)
3822 // Update variable assignment.
3823 s.vars[left] = right
3824 s.addNamedValue(left, right)
3828 // If this assignment clobbers an entire local variable, then emit
3829 // OpVarDef so liveness analysis knows the variable is redefined.
3830 if base, ok := clobberBase(left).(*ir.Name); ok && base.OnStack() && skip == 0 && t.HasPointers() {
3831 s.vars[memVar] = s.newValue1Apos(ssa.OpVarDef, types.TypeMem, base, s.mem(), !ir.IsAutoTmp(base))
3834 // Left is not ssa-able. Compute its address.
3835 addr := s.addr(left)
3836 if ir.IsReflectHeaderDataField(left) {
3837 // Package unsafe's documentation says storing pointers into
3838 // reflect.SliceHeader and reflect.StringHeader's Data fields
3839 // is valid, even though they have type uintptr (#19168).
3840 // Mark it pointer type to signal the writebarrier pass to
3841 // insert a write barrier.
3842 t = types.Types[types.TUNSAFEPTR]
3845 // Treat as a mem->mem move.
3849 s.moveWhichMayOverlap(t, addr, right, mayOverlap)
3853 // Treat as a store.
3854 s.storeType(t, addr, right, skip, !ir.IsAutoTmp(left))
3857 // zeroVal returns the zero value for type t.
3858 func (s *state) zeroVal(t *types.Type) *ssa.Value {
3863 return s.constInt8(t, 0)
3865 return s.constInt16(t, 0)
3867 return s.constInt32(t, 0)
3869 return s.constInt64(t, 0)
3871 s.Fatalf("bad sized integer type %v", t)
3876 return s.constFloat32(t, 0)
3878 return s.constFloat64(t, 0)
3880 s.Fatalf("bad sized float type %v", t)
3885 z := s.constFloat32(types.Types[types.TFLOAT32], 0)
3886 return s.entryNewValue2(ssa.OpComplexMake, t, z, z)
3888 z := s.constFloat64(types.Types[types.TFLOAT64], 0)
3889 return s.entryNewValue2(ssa.OpComplexMake, t, z, z)
3891 s.Fatalf("bad sized complex type %v", t)
3895 return s.constEmptyString(t)
3896 case t.IsPtrShaped():
3897 return s.constNil(t)
3899 return s.constBool(false)
3900 case t.IsInterface():
3901 return s.constInterface(t)
3903 return s.constSlice(t)
3906 v := s.entryNewValue0(ssa.StructMakeOp(t.NumFields()), t)
3907 for i := 0; i < n; i++ {
3908 v.AddArg(s.zeroVal(t.FieldType(i)))
3912 switch t.NumElem() {
3914 return s.entryNewValue0(ssa.OpArrayMake0, t)
3916 return s.entryNewValue1(ssa.OpArrayMake1, t, s.zeroVal(t.Elem()))
3919 s.Fatalf("zero for type %v not implemented", t)
3926 callNormal callKind = iota
3933 type sfRtCallDef struct {
3938 var softFloatOps map[ssa.Op]sfRtCallDef
3940 func softfloatInit() {
3941 // Some of these operations get transformed by sfcall.
3942 softFloatOps = map[ssa.Op]sfRtCallDef{
3943 ssa.OpAdd32F: {typecheck.LookupRuntimeFunc("fadd32"), types.TFLOAT32},
3944 ssa.OpAdd64F: {typecheck.LookupRuntimeFunc("fadd64"), types.TFLOAT64},
3945 ssa.OpSub32F: {typecheck.LookupRuntimeFunc("fadd32"), types.TFLOAT32},
3946 ssa.OpSub64F: {typecheck.LookupRuntimeFunc("fadd64"), types.TFLOAT64},
3947 ssa.OpMul32F: {typecheck.LookupRuntimeFunc("fmul32"), types.TFLOAT32},
3948 ssa.OpMul64F: {typecheck.LookupRuntimeFunc("fmul64"), types.TFLOAT64},
3949 ssa.OpDiv32F: {typecheck.LookupRuntimeFunc("fdiv32"), types.TFLOAT32},
3950 ssa.OpDiv64F: {typecheck.LookupRuntimeFunc("fdiv64"), types.TFLOAT64},
3952 ssa.OpEq64F: {typecheck.LookupRuntimeFunc("feq64"), types.TBOOL},
3953 ssa.OpEq32F: {typecheck.LookupRuntimeFunc("feq32"), types.TBOOL},
3954 ssa.OpNeq64F: {typecheck.LookupRuntimeFunc("feq64"), types.TBOOL},
3955 ssa.OpNeq32F: {typecheck.LookupRuntimeFunc("feq32"), types.TBOOL},
3956 ssa.OpLess64F: {typecheck.LookupRuntimeFunc("fgt64"), types.TBOOL},
3957 ssa.OpLess32F: {typecheck.LookupRuntimeFunc("fgt32"), types.TBOOL},
3958 ssa.OpLeq64F: {typecheck.LookupRuntimeFunc("fge64"), types.TBOOL},
3959 ssa.OpLeq32F: {typecheck.LookupRuntimeFunc("fge32"), types.TBOOL},
3961 ssa.OpCvt32to32F: {typecheck.LookupRuntimeFunc("fint32to32"), types.TFLOAT32},
3962 ssa.OpCvt32Fto32: {typecheck.LookupRuntimeFunc("f32toint32"), types.TINT32},
3963 ssa.OpCvt64to32F: {typecheck.LookupRuntimeFunc("fint64to32"), types.TFLOAT32},
3964 ssa.OpCvt32Fto64: {typecheck.LookupRuntimeFunc("f32toint64"), types.TINT64},
3965 ssa.OpCvt64Uto32F: {typecheck.LookupRuntimeFunc("fuint64to32"), types.TFLOAT32},
3966 ssa.OpCvt32Fto64U: {typecheck.LookupRuntimeFunc("f32touint64"), types.TUINT64},
3967 ssa.OpCvt32to64F: {typecheck.LookupRuntimeFunc("fint32to64"), types.TFLOAT64},
3968 ssa.OpCvt64Fto32: {typecheck.LookupRuntimeFunc("f64toint32"), types.TINT32},
3969 ssa.OpCvt64to64F: {typecheck.LookupRuntimeFunc("fint64to64"), types.TFLOAT64},
3970 ssa.OpCvt64Fto64: {typecheck.LookupRuntimeFunc("f64toint64"), types.TINT64},
3971 ssa.OpCvt64Uto64F: {typecheck.LookupRuntimeFunc("fuint64to64"), types.TFLOAT64},
3972 ssa.OpCvt64Fto64U: {typecheck.LookupRuntimeFunc("f64touint64"), types.TUINT64},
3973 ssa.OpCvt32Fto64F: {typecheck.LookupRuntimeFunc("f32to64"), types.TFLOAT64},
3974 ssa.OpCvt64Fto32F: {typecheck.LookupRuntimeFunc("f64to32"), types.TFLOAT32},
3978 // TODO: do not emit sfcall if operation can be optimized to constant in later
3980 func (s *state) sfcall(op ssa.Op, args ...*ssa.Value) (*ssa.Value, bool) {
3981 f2i := func(t *types.Type) *types.Type {
3983 case types.TFLOAT32:
3984 return types.Types[types.TUINT32]
3985 case types.TFLOAT64:
3986 return types.Types[types.TUINT64]
3991 if callDef, ok := softFloatOps[op]; ok {
3997 args[0], args[1] = args[1], args[0]
4000 args[1] = s.newValue1(s.ssaOp(ir.ONEG, types.Types[callDef.rtype]), args[1].Type, args[1])
4003 // runtime functions take uints for floats and returns uints.
4004 // Convert to uints so we use the right calling convention.
4005 for i, a := range args {
4006 if a.Type.IsFloat() {
4007 args[i] = s.newValue1(ssa.OpCopy, f2i(a.Type), a)
4011 rt := types.Types[callDef.rtype]
4012 result := s.rtcall(callDef.rtfn, true, []*types.Type{f2i(rt)}, args...)[0]
4014 result = s.newValue1(ssa.OpCopy, rt, result)
4016 if op == ssa.OpNeq32F || op == ssa.OpNeq64F {
4017 result = s.newValue1(ssa.OpNot, result.Type, result)
4024 var intrinsics map[intrinsicKey]intrinsicBuilder
4026 // An intrinsicBuilder converts a call node n into an ssa value that
4027 // implements that call as an intrinsic. args is a list of arguments to the func.
4028 type intrinsicBuilder func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value
4030 type intrinsicKey struct {
4037 intrinsics = map[intrinsicKey]intrinsicBuilder{}
4042 var lwatomics []*sys.Arch
4043 for _, a := range &sys.Archs {
4044 all = append(all, a)
4050 if a.Family != sys.PPC64 {
4051 lwatomics = append(lwatomics, a)
4055 // add adds the intrinsic b for pkg.fn for the given list of architectures.
4056 add := func(pkg, fn string, b intrinsicBuilder, archs ...*sys.Arch) {
4057 for _, a := range archs {
4058 intrinsics[intrinsicKey{a, pkg, fn}] = b
4061 // addF does the same as add but operates on architecture families.
4062 addF := func(pkg, fn string, b intrinsicBuilder, archFamilies ...sys.ArchFamily) {
4064 for _, f := range archFamilies {
4066 panic("too many architecture families")
4070 for _, a := range all {
4071 if m>>uint(a.Family)&1 != 0 {
4072 intrinsics[intrinsicKey{a, pkg, fn}] = b
4076 // alias defines pkg.fn = pkg2.fn2 for all architectures in archs for which pkg2.fn2 exists.
4077 alias := func(pkg, fn, pkg2, fn2 string, archs ...*sys.Arch) {
4079 for _, a := range archs {
4080 if b, ok := intrinsics[intrinsicKey{a, pkg2, fn2}]; ok {
4081 intrinsics[intrinsicKey{a, pkg, fn}] = b
4086 panic(fmt.Sprintf("attempted to alias undefined intrinsic: %s.%s", pkg, fn))
4090 /******** runtime ********/
4091 if !base.Flag.Cfg.Instrumenting {
4092 add("runtime", "slicebytetostringtmp",
4093 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4094 // Compiler frontend optimizations emit OBYTES2STRTMP nodes
4095 // for the backend instead of slicebytetostringtmp calls
4096 // when not instrumenting.
4097 return s.newValue2(ssa.OpStringMake, n.Type(), args[0], args[1])
4101 addF("runtime/internal/math", "MulUintptr",
4102 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4103 if s.config.PtrSize == 4 {
4104 return s.newValue2(ssa.OpMul32uover, types.NewTuple(types.Types[types.TUINT], types.Types[types.TUINT]), args[0], args[1])
4106 return s.newValue2(ssa.OpMul64uover, types.NewTuple(types.Types[types.TUINT], types.Types[types.TUINT]), args[0], args[1])
4108 sys.AMD64, sys.I386, sys.Loong64, sys.MIPS64, sys.RISCV64, sys.ARM64)
4109 alias("runtime", "mulUintptr", "runtime/internal/math", "MulUintptr", all...)
4110 add("runtime", "KeepAlive",
4111 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4112 data := s.newValue1(ssa.OpIData, s.f.Config.Types.BytePtr, args[0])
4113 s.vars[memVar] = s.newValue2(ssa.OpKeepAlive, types.TypeMem, data, s.mem())
4117 add("runtime", "getclosureptr",
4118 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4119 return s.newValue0(ssa.OpGetClosurePtr, s.f.Config.Types.Uintptr)
4123 add("runtime", "getcallerpc",
4124 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4125 return s.newValue0(ssa.OpGetCallerPC, s.f.Config.Types.Uintptr)
4129 add("runtime", "getcallersp",
4130 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4131 return s.newValue1(ssa.OpGetCallerSP, s.f.Config.Types.Uintptr, s.mem())
4135 addF("runtime", "publicationBarrier",
4136 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4137 s.vars[memVar] = s.newValue1(ssa.OpPubBarrier, types.TypeMem, s.mem())
4140 sys.ARM64, sys.PPC64)
4142 brev_arch := []sys.ArchFamily{sys.AMD64, sys.I386, sys.ARM64, sys.ARM, sys.S390X}
4143 if buildcfg.GOPPC64 >= 10 {
4144 // Use only on Power10 as the new byte reverse instructions that Power10 provide
4145 // make it worthwhile as an intrinsic
4146 brev_arch = append(brev_arch, sys.PPC64)
4148 /******** runtime/internal/sys ********/
4149 addF("runtime/internal/sys", "Bswap32",
4150 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4151 return s.newValue1(ssa.OpBswap32, types.Types[types.TUINT32], args[0])
4154 addF("runtime/internal/sys", "Bswap64",
4155 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4156 return s.newValue1(ssa.OpBswap64, types.Types[types.TUINT64], args[0])
4160 /****** Prefetch ******/
4161 makePrefetchFunc := func(op ssa.Op) func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4162 return func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4163 s.vars[memVar] = s.newValue2(op, types.TypeMem, args[0], s.mem())
4168 // Make Prefetch intrinsics for supported platforms
4169 // On the unsupported platforms stub function will be eliminated
4170 addF("runtime/internal/sys", "Prefetch", makePrefetchFunc(ssa.OpPrefetchCache),
4171 sys.AMD64, sys.ARM64, sys.PPC64)
4172 addF("runtime/internal/sys", "PrefetchStreamed", makePrefetchFunc(ssa.OpPrefetchCacheStreamed),
4173 sys.AMD64, sys.ARM64, sys.PPC64)
4175 /******** runtime/internal/atomic ********/
4176 addF("runtime/internal/atomic", "Load",
4177 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4178 v := s.newValue2(ssa.OpAtomicLoad32, types.NewTuple(types.Types[types.TUINT32], types.TypeMem), args[0], s.mem())
4179 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4180 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT32], v)
4182 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4183 addF("runtime/internal/atomic", "Load8",
4184 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4185 v := s.newValue2(ssa.OpAtomicLoad8, types.NewTuple(types.Types[types.TUINT8], types.TypeMem), args[0], s.mem())
4186 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4187 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT8], v)
4189 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4190 addF("runtime/internal/atomic", "Load64",
4191 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4192 v := s.newValue2(ssa.OpAtomicLoad64, types.NewTuple(types.Types[types.TUINT64], types.TypeMem), args[0], s.mem())
4193 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4194 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT64], v)
4196 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4197 addF("runtime/internal/atomic", "LoadAcq",
4198 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4199 v := s.newValue2(ssa.OpAtomicLoadAcq32, types.NewTuple(types.Types[types.TUINT32], types.TypeMem), args[0], s.mem())
4200 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4201 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT32], v)
4203 sys.PPC64, sys.S390X)
4204 addF("runtime/internal/atomic", "LoadAcq64",
4205 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4206 v := s.newValue2(ssa.OpAtomicLoadAcq64, types.NewTuple(types.Types[types.TUINT64], types.TypeMem), args[0], s.mem())
4207 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4208 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT64], v)
4211 addF("runtime/internal/atomic", "Loadp",
4212 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4213 v := s.newValue2(ssa.OpAtomicLoadPtr, types.NewTuple(s.f.Config.Types.BytePtr, types.TypeMem), args[0], s.mem())
4214 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4215 return s.newValue1(ssa.OpSelect0, s.f.Config.Types.BytePtr, v)
4217 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4219 addF("runtime/internal/atomic", "Store",
4220 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4221 s.vars[memVar] = s.newValue3(ssa.OpAtomicStore32, types.TypeMem, args[0], args[1], s.mem())
4224 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4225 addF("runtime/internal/atomic", "Store8",
4226 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4227 s.vars[memVar] = s.newValue3(ssa.OpAtomicStore8, types.TypeMem, args[0], args[1], s.mem())
4230 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4231 addF("runtime/internal/atomic", "Store64",
4232 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4233 s.vars[memVar] = s.newValue3(ssa.OpAtomicStore64, types.TypeMem, args[0], args[1], s.mem())
4236 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4237 addF("runtime/internal/atomic", "StorepNoWB",
4238 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4239 s.vars[memVar] = s.newValue3(ssa.OpAtomicStorePtrNoWB, types.TypeMem, args[0], args[1], s.mem())
4242 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.RISCV64, sys.S390X)
4243 addF("runtime/internal/atomic", "StoreRel",
4244 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4245 s.vars[memVar] = s.newValue3(ssa.OpAtomicStoreRel32, types.TypeMem, args[0], args[1], s.mem())
4248 sys.PPC64, sys.S390X)
4249 addF("runtime/internal/atomic", "StoreRel64",
4250 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4251 s.vars[memVar] = s.newValue3(ssa.OpAtomicStoreRel64, types.TypeMem, args[0], args[1], s.mem())
4256 addF("runtime/internal/atomic", "Xchg",
4257 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4258 v := s.newValue3(ssa.OpAtomicExchange32, types.NewTuple(types.Types[types.TUINT32], types.TypeMem), args[0], args[1], s.mem())
4259 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4260 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT32], v)
4262 sys.AMD64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4263 addF("runtime/internal/atomic", "Xchg64",
4264 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4265 v := s.newValue3(ssa.OpAtomicExchange64, types.NewTuple(types.Types[types.TUINT64], types.TypeMem), args[0], args[1], s.mem())
4266 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4267 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT64], v)
4269 sys.AMD64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4271 type atomicOpEmitter func(s *state, n *ir.CallExpr, args []*ssa.Value, op ssa.Op, typ types.Kind)
4273 makeAtomicGuardedIntrinsicARM64 := func(op0, op1 ssa.Op, typ, rtyp types.Kind, emit atomicOpEmitter) intrinsicBuilder {
4275 return func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4276 // Target Atomic feature is identified by dynamic detection
4277 addr := s.entryNewValue1A(ssa.OpAddr, types.Types[types.TBOOL].PtrTo(), ir.Syms.ARM64HasATOMICS, s.sb)
4278 v := s.load(types.Types[types.TBOOL], addr)
4280 b.Kind = ssa.BlockIf
4282 bTrue := s.f.NewBlock(ssa.BlockPlain)
4283 bFalse := s.f.NewBlock(ssa.BlockPlain)
4284 bEnd := s.f.NewBlock(ssa.BlockPlain)
4287 b.Likely = ssa.BranchLikely
4289 // We have atomic instructions - use it directly.
4291 emit(s, n, args, op1, typ)
4292 s.endBlock().AddEdgeTo(bEnd)
4294 // Use original instruction sequence.
4295 s.startBlock(bFalse)
4296 emit(s, n, args, op0, typ)
4297 s.endBlock().AddEdgeTo(bEnd)
4301 if rtyp == types.TNIL {
4304 return s.variable(n, types.Types[rtyp])
4309 atomicXchgXaddEmitterARM64 := func(s *state, n *ir.CallExpr, args []*ssa.Value, op ssa.Op, typ types.Kind) {
4310 v := s.newValue3(op, types.NewTuple(types.Types[typ], types.TypeMem), args[0], args[1], s.mem())
4311 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4312 s.vars[n] = s.newValue1(ssa.OpSelect0, types.Types[typ], v)
4314 addF("runtime/internal/atomic", "Xchg",
4315 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicExchange32, ssa.OpAtomicExchange32Variant, types.TUINT32, types.TUINT32, atomicXchgXaddEmitterARM64),
4317 addF("runtime/internal/atomic", "Xchg64",
4318 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicExchange64, ssa.OpAtomicExchange64Variant, types.TUINT64, types.TUINT64, atomicXchgXaddEmitterARM64),
4321 addF("runtime/internal/atomic", "Xadd",
4322 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4323 v := s.newValue3(ssa.OpAtomicAdd32, types.NewTuple(types.Types[types.TUINT32], types.TypeMem), args[0], args[1], s.mem())
4324 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4325 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT32], v)
4327 sys.AMD64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4328 addF("runtime/internal/atomic", "Xadd64",
4329 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4330 v := s.newValue3(ssa.OpAtomicAdd64, types.NewTuple(types.Types[types.TUINT64], types.TypeMem), args[0], args[1], s.mem())
4331 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4332 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT64], v)
4334 sys.AMD64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4336 addF("runtime/internal/atomic", "Xadd",
4337 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAdd32, ssa.OpAtomicAdd32Variant, types.TUINT32, types.TUINT32, atomicXchgXaddEmitterARM64),
4339 addF("runtime/internal/atomic", "Xadd64",
4340 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAdd64, ssa.OpAtomicAdd64Variant, types.TUINT64, types.TUINT64, atomicXchgXaddEmitterARM64),
4343 addF("runtime/internal/atomic", "Cas",
4344 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4345 v := s.newValue4(ssa.OpAtomicCompareAndSwap32, types.NewTuple(types.Types[types.TBOOL], types.TypeMem), args[0], args[1], args[2], s.mem())
4346 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4347 return s.newValue1(ssa.OpSelect0, types.Types[types.TBOOL], v)
4349 sys.AMD64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4350 addF("runtime/internal/atomic", "Cas64",
4351 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4352 v := s.newValue4(ssa.OpAtomicCompareAndSwap64, types.NewTuple(types.Types[types.TBOOL], types.TypeMem), args[0], args[1], args[2], s.mem())
4353 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4354 return s.newValue1(ssa.OpSelect0, types.Types[types.TBOOL], v)
4356 sys.AMD64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4357 addF("runtime/internal/atomic", "CasRel",
4358 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4359 v := s.newValue4(ssa.OpAtomicCompareAndSwap32, types.NewTuple(types.Types[types.TBOOL], types.TypeMem), args[0], args[1], args[2], s.mem())
4360 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4361 return s.newValue1(ssa.OpSelect0, types.Types[types.TBOOL], v)
4365 atomicCasEmitterARM64 := func(s *state, n *ir.CallExpr, args []*ssa.Value, op ssa.Op, typ types.Kind) {
4366 v := s.newValue4(op, types.NewTuple(types.Types[types.TBOOL], types.TypeMem), args[0], args[1], args[2], s.mem())
4367 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4368 s.vars[n] = s.newValue1(ssa.OpSelect0, types.Types[typ], v)
4371 addF("runtime/internal/atomic", "Cas",
4372 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicCompareAndSwap32, ssa.OpAtomicCompareAndSwap32Variant, types.TUINT32, types.TBOOL, atomicCasEmitterARM64),
4374 addF("runtime/internal/atomic", "Cas64",
4375 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicCompareAndSwap64, ssa.OpAtomicCompareAndSwap64Variant, types.TUINT64, types.TBOOL, atomicCasEmitterARM64),
4378 addF("runtime/internal/atomic", "And8",
4379 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4380 s.vars[memVar] = s.newValue3(ssa.OpAtomicAnd8, types.TypeMem, args[0], args[1], s.mem())
4383 sys.AMD64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4384 addF("runtime/internal/atomic", "And",
4385 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4386 s.vars[memVar] = s.newValue3(ssa.OpAtomicAnd32, types.TypeMem, args[0], args[1], s.mem())
4389 sys.AMD64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4390 addF("runtime/internal/atomic", "Or8",
4391 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4392 s.vars[memVar] = s.newValue3(ssa.OpAtomicOr8, types.TypeMem, args[0], args[1], s.mem())
4395 sys.AMD64, sys.ARM64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4396 addF("runtime/internal/atomic", "Or",
4397 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4398 s.vars[memVar] = s.newValue3(ssa.OpAtomicOr32, types.TypeMem, args[0], args[1], s.mem())
4401 sys.AMD64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4403 atomicAndOrEmitterARM64 := func(s *state, n *ir.CallExpr, args []*ssa.Value, op ssa.Op, typ types.Kind) {
4404 s.vars[memVar] = s.newValue3(op, types.TypeMem, args[0], args[1], s.mem())
4407 addF("runtime/internal/atomic", "And8",
4408 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAnd8, ssa.OpAtomicAnd8Variant, types.TNIL, types.TNIL, atomicAndOrEmitterARM64),
4410 addF("runtime/internal/atomic", "And",
4411 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAnd32, ssa.OpAtomicAnd32Variant, types.TNIL, types.TNIL, atomicAndOrEmitterARM64),
4413 addF("runtime/internal/atomic", "Or8",
4414 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicOr8, ssa.OpAtomicOr8Variant, types.TNIL, types.TNIL, atomicAndOrEmitterARM64),
4416 addF("runtime/internal/atomic", "Or",
4417 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicOr32, ssa.OpAtomicOr32Variant, types.TNIL, types.TNIL, atomicAndOrEmitterARM64),
4420 // Aliases for atomic load operations
4421 alias("runtime/internal/atomic", "Loadint32", "runtime/internal/atomic", "Load", all...)
4422 alias("runtime/internal/atomic", "Loadint64", "runtime/internal/atomic", "Load64", all...)
4423 alias("runtime/internal/atomic", "Loaduintptr", "runtime/internal/atomic", "Load", p4...)
4424 alias("runtime/internal/atomic", "Loaduintptr", "runtime/internal/atomic", "Load64", p8...)
4425 alias("runtime/internal/atomic", "Loaduint", "runtime/internal/atomic", "Load", p4...)
4426 alias("runtime/internal/atomic", "Loaduint", "runtime/internal/atomic", "Load64", p8...)
4427 alias("runtime/internal/atomic", "LoadAcq", "runtime/internal/atomic", "Load", lwatomics...)
4428 alias("runtime/internal/atomic", "LoadAcq64", "runtime/internal/atomic", "Load64", lwatomics...)
4429 alias("runtime/internal/atomic", "LoadAcquintptr", "runtime/internal/atomic", "LoadAcq", p4...)
4430 alias("sync", "runtime_LoadAcquintptr", "runtime/internal/atomic", "LoadAcq", p4...) // linknamed
4431 alias("runtime/internal/atomic", "LoadAcquintptr", "runtime/internal/atomic", "LoadAcq64", p8...)
4432 alias("sync", "runtime_LoadAcquintptr", "runtime/internal/atomic", "LoadAcq64", p8...) // linknamed
4434 // Aliases for atomic store operations
4435 alias("runtime/internal/atomic", "Storeint32", "runtime/internal/atomic", "Store", all...)
4436 alias("runtime/internal/atomic", "Storeint64", "runtime/internal/atomic", "Store64", all...)
4437 alias("runtime/internal/atomic", "Storeuintptr", "runtime/internal/atomic", "Store", p4...)
4438 alias("runtime/internal/atomic", "Storeuintptr", "runtime/internal/atomic", "Store64", p8...)
4439 alias("runtime/internal/atomic", "StoreRel", "runtime/internal/atomic", "Store", lwatomics...)
4440 alias("runtime/internal/atomic", "StoreRel64", "runtime/internal/atomic", "Store64", lwatomics...)
4441 alias("runtime/internal/atomic", "StoreReluintptr", "runtime/internal/atomic", "StoreRel", p4...)
4442 alias("sync", "runtime_StoreReluintptr", "runtime/internal/atomic", "StoreRel", p4...) // linknamed
4443 alias("runtime/internal/atomic", "StoreReluintptr", "runtime/internal/atomic", "StoreRel64", p8...)
4444 alias("sync", "runtime_StoreReluintptr", "runtime/internal/atomic", "StoreRel64", p8...) // linknamed
4446 // Aliases for atomic swap operations
4447 alias("runtime/internal/atomic", "Xchgint32", "runtime/internal/atomic", "Xchg", all...)
4448 alias("runtime/internal/atomic", "Xchgint64", "runtime/internal/atomic", "Xchg64", all...)
4449 alias("runtime/internal/atomic", "Xchguintptr", "runtime/internal/atomic", "Xchg", p4...)
4450 alias("runtime/internal/atomic", "Xchguintptr", "runtime/internal/atomic", "Xchg64", p8...)
4452 // Aliases for atomic add operations
4453 alias("runtime/internal/atomic", "Xaddint32", "runtime/internal/atomic", "Xadd", all...)
4454 alias("runtime/internal/atomic", "Xaddint64", "runtime/internal/atomic", "Xadd64", all...)
4455 alias("runtime/internal/atomic", "Xadduintptr", "runtime/internal/atomic", "Xadd", p4...)
4456 alias("runtime/internal/atomic", "Xadduintptr", "runtime/internal/atomic", "Xadd64", p8...)
4458 // Aliases for atomic CAS operations
4459 alias("runtime/internal/atomic", "Casint32", "runtime/internal/atomic", "Cas", all...)
4460 alias("runtime/internal/atomic", "Casint64", "runtime/internal/atomic", "Cas64", all...)
4461 alias("runtime/internal/atomic", "Casuintptr", "runtime/internal/atomic", "Cas", p4...)
4462 alias("runtime/internal/atomic", "Casuintptr", "runtime/internal/atomic", "Cas64", p8...)
4463 alias("runtime/internal/atomic", "Casp1", "runtime/internal/atomic", "Cas", p4...)
4464 alias("runtime/internal/atomic", "Casp1", "runtime/internal/atomic", "Cas64", p8...)
4465 alias("runtime/internal/atomic", "CasRel", "runtime/internal/atomic", "Cas", lwatomics...)
4467 /******** math ********/
4468 addF("math", "sqrt",
4469 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4470 return s.newValue1(ssa.OpSqrt, types.Types[types.TFLOAT64], args[0])
4472 sys.I386, sys.AMD64, sys.ARM, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X, sys.Wasm)
4473 addF("math", "Trunc",
4474 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4475 return s.newValue1(ssa.OpTrunc, types.Types[types.TFLOAT64], args[0])
4477 sys.ARM64, sys.PPC64, sys.S390X, sys.Wasm)
4478 addF("math", "Ceil",
4479 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4480 return s.newValue1(ssa.OpCeil, types.Types[types.TFLOAT64], args[0])
4482 sys.ARM64, sys.PPC64, sys.S390X, sys.Wasm)
4483 addF("math", "Floor",
4484 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4485 return s.newValue1(ssa.OpFloor, types.Types[types.TFLOAT64], args[0])
4487 sys.ARM64, sys.PPC64, sys.S390X, sys.Wasm)
4488 addF("math", "Round",
4489 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4490 return s.newValue1(ssa.OpRound, types.Types[types.TFLOAT64], args[0])
4492 sys.ARM64, sys.PPC64, sys.S390X)
4493 addF("math", "RoundToEven",
4494 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4495 return s.newValue1(ssa.OpRoundToEven, types.Types[types.TFLOAT64], args[0])
4497 sys.ARM64, sys.S390X, sys.Wasm)
4499 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4500 return s.newValue1(ssa.OpAbs, types.Types[types.TFLOAT64], args[0])
4502 sys.ARM64, sys.ARM, sys.PPC64, sys.RISCV64, sys.Wasm, sys.MIPS, sys.MIPS64)
4503 addF("math", "Copysign",
4504 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4505 return s.newValue2(ssa.OpCopysign, types.Types[types.TFLOAT64], args[0], args[1])
4507 sys.PPC64, sys.RISCV64, sys.Wasm)
4509 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4510 return s.newValue3(ssa.OpFMA, types.Types[types.TFLOAT64], args[0], args[1], args[2])
4512 sys.ARM64, sys.PPC64, sys.RISCV64, sys.S390X)
4514 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4515 if !s.config.UseFMA {
4516 s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
4517 return s.variable(n, types.Types[types.TFLOAT64])
4520 if buildcfg.GOAMD64 >= 3 {
4521 return s.newValue3(ssa.OpFMA, types.Types[types.TFLOAT64], args[0], args[1], args[2])
4524 v := s.entryNewValue0A(ssa.OpHasCPUFeature, types.Types[types.TBOOL], ir.Syms.X86HasFMA)
4526 b.Kind = ssa.BlockIf
4528 bTrue := s.f.NewBlock(ssa.BlockPlain)
4529 bFalse := s.f.NewBlock(ssa.BlockPlain)
4530 bEnd := s.f.NewBlock(ssa.BlockPlain)
4533 b.Likely = ssa.BranchLikely // >= haswell cpus are common
4535 // We have the intrinsic - use it directly.
4537 s.vars[n] = s.newValue3(ssa.OpFMA, types.Types[types.TFLOAT64], args[0], args[1], args[2])
4538 s.endBlock().AddEdgeTo(bEnd)
4540 // Call the pure Go version.
4541 s.startBlock(bFalse)
4542 s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
4543 s.endBlock().AddEdgeTo(bEnd)
4547 return s.variable(n, types.Types[types.TFLOAT64])
4551 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4552 if !s.config.UseFMA {
4553 s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
4554 return s.variable(n, types.Types[types.TFLOAT64])
4556 addr := s.entryNewValue1A(ssa.OpAddr, types.Types[types.TBOOL].PtrTo(), ir.Syms.ARMHasVFPv4, s.sb)
4557 v := s.load(types.Types[types.TBOOL], addr)
4559 b.Kind = ssa.BlockIf
4561 bTrue := s.f.NewBlock(ssa.BlockPlain)
4562 bFalse := s.f.NewBlock(ssa.BlockPlain)
4563 bEnd := s.f.NewBlock(ssa.BlockPlain)
4566 b.Likely = ssa.BranchLikely
4568 // We have the intrinsic - use it directly.
4570 s.vars[n] = s.newValue3(ssa.OpFMA, types.Types[types.TFLOAT64], args[0], args[1], args[2])
4571 s.endBlock().AddEdgeTo(bEnd)
4573 // Call the pure Go version.
4574 s.startBlock(bFalse)
4575 s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
4576 s.endBlock().AddEdgeTo(bEnd)
4580 return s.variable(n, types.Types[types.TFLOAT64])
4584 makeRoundAMD64 := func(op ssa.Op) func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4585 return func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4586 if buildcfg.GOAMD64 >= 2 {
4587 return s.newValue1(op, types.Types[types.TFLOAT64], args[0])
4590 v := s.entryNewValue0A(ssa.OpHasCPUFeature, types.Types[types.TBOOL], ir.Syms.X86HasSSE41)
4592 b.Kind = ssa.BlockIf
4594 bTrue := s.f.NewBlock(ssa.BlockPlain)
4595 bFalse := s.f.NewBlock(ssa.BlockPlain)
4596 bEnd := s.f.NewBlock(ssa.BlockPlain)
4599 b.Likely = ssa.BranchLikely // most machines have sse4.1 nowadays
4601 // We have the intrinsic - use it directly.
4603 s.vars[n] = s.newValue1(op, types.Types[types.TFLOAT64], args[0])
4604 s.endBlock().AddEdgeTo(bEnd)
4606 // Call the pure Go version.
4607 s.startBlock(bFalse)
4608 s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
4609 s.endBlock().AddEdgeTo(bEnd)
4613 return s.variable(n, types.Types[types.TFLOAT64])
4616 addF("math", "RoundToEven",
4617 makeRoundAMD64(ssa.OpRoundToEven),
4619 addF("math", "Floor",
4620 makeRoundAMD64(ssa.OpFloor),
4622 addF("math", "Ceil",
4623 makeRoundAMD64(ssa.OpCeil),
4625 addF("math", "Trunc",
4626 makeRoundAMD64(ssa.OpTrunc),
4629 /******** math/bits ********/
4630 addF("math/bits", "TrailingZeros64",
4631 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4632 return s.newValue1(ssa.OpCtz64, types.Types[types.TINT], args[0])
4634 sys.AMD64, sys.I386, sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4635 addF("math/bits", "TrailingZeros32",
4636 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4637 return s.newValue1(ssa.OpCtz32, types.Types[types.TINT], args[0])
4639 sys.AMD64, sys.I386, sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4640 addF("math/bits", "TrailingZeros16",
4641 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4642 x := s.newValue1(ssa.OpZeroExt16to32, types.Types[types.TUINT32], args[0])
4643 c := s.constInt32(types.Types[types.TUINT32], 1<<16)
4644 y := s.newValue2(ssa.OpOr32, types.Types[types.TUINT32], x, c)
4645 return s.newValue1(ssa.OpCtz32, types.Types[types.TINT], y)
4648 addF("math/bits", "TrailingZeros16",
4649 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4650 return s.newValue1(ssa.OpCtz16, types.Types[types.TINT], args[0])
4652 sys.AMD64, sys.I386, sys.ARM, sys.ARM64, sys.Wasm)
4653 addF("math/bits", "TrailingZeros16",
4654 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4655 x := s.newValue1(ssa.OpZeroExt16to64, types.Types[types.TUINT64], args[0])
4656 c := s.constInt64(types.Types[types.TUINT64], 1<<16)
4657 y := s.newValue2(ssa.OpOr64, types.Types[types.TUINT64], x, c)
4658 return s.newValue1(ssa.OpCtz64, types.Types[types.TINT], y)
4660 sys.S390X, sys.PPC64)
4661 addF("math/bits", "TrailingZeros8",
4662 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4663 x := s.newValue1(ssa.OpZeroExt8to32, types.Types[types.TUINT32], args[0])
4664 c := s.constInt32(types.Types[types.TUINT32], 1<<8)
4665 y := s.newValue2(ssa.OpOr32, types.Types[types.TUINT32], x, c)
4666 return s.newValue1(ssa.OpCtz32, types.Types[types.TINT], y)
4669 addF("math/bits", "TrailingZeros8",
4670 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4671 return s.newValue1(ssa.OpCtz8, types.Types[types.TINT], args[0])
4673 sys.AMD64, sys.I386, sys.ARM, sys.ARM64, sys.Wasm)
4674 addF("math/bits", "TrailingZeros8",
4675 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4676 x := s.newValue1(ssa.OpZeroExt8to64, types.Types[types.TUINT64], args[0])
4677 c := s.constInt64(types.Types[types.TUINT64], 1<<8)
4678 y := s.newValue2(ssa.OpOr64, types.Types[types.TUINT64], x, c)
4679 return s.newValue1(ssa.OpCtz64, types.Types[types.TINT], y)
4682 alias("math/bits", "ReverseBytes64", "runtime/internal/sys", "Bswap64", all...)
4683 alias("math/bits", "ReverseBytes32", "runtime/internal/sys", "Bswap32", all...)
4684 // ReverseBytes inlines correctly, no need to intrinsify it.
4685 // Nothing special is needed for targets where ReverseBytes16 lowers to a rotate
4686 // On Power10, 16-bit rotate is not available so use BRH instruction
4687 if buildcfg.GOPPC64 >= 10 {
4688 addF("math/bits", "ReverseBytes16",
4689 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4690 return s.newValue1(ssa.OpBswap16, types.Types[types.TUINT], args[0])
4695 addF("math/bits", "Len64",
4696 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4697 return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], args[0])
4699 sys.AMD64, sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4700 addF("math/bits", "Len32",
4701 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4702 return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], args[0])
4704 sys.AMD64, sys.ARM64, sys.PPC64)
4705 addF("math/bits", "Len32",
4706 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4707 if s.config.PtrSize == 4 {
4708 return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], args[0])
4710 x := s.newValue1(ssa.OpZeroExt32to64, types.Types[types.TUINT64], args[0])
4711 return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], x)
4713 sys.ARM, sys.S390X, sys.MIPS, sys.Wasm)
4714 addF("math/bits", "Len16",
4715 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4716 if s.config.PtrSize == 4 {
4717 x := s.newValue1(ssa.OpZeroExt16to32, types.Types[types.TUINT32], args[0])
4718 return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], x)
4720 x := s.newValue1(ssa.OpZeroExt16to64, types.Types[types.TUINT64], args[0])
4721 return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], x)
4723 sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4724 addF("math/bits", "Len16",
4725 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4726 return s.newValue1(ssa.OpBitLen16, types.Types[types.TINT], args[0])
4729 addF("math/bits", "Len8",
4730 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4731 if s.config.PtrSize == 4 {
4732 x := s.newValue1(ssa.OpZeroExt8to32, types.Types[types.TUINT32], args[0])
4733 return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], x)
4735 x := s.newValue1(ssa.OpZeroExt8to64, types.Types[types.TUINT64], args[0])
4736 return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], x)
4738 sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4739 addF("math/bits", "Len8",
4740 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4741 return s.newValue1(ssa.OpBitLen8, types.Types[types.TINT], args[0])
4744 addF("math/bits", "Len",
4745 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4746 if s.config.PtrSize == 4 {
4747 return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], args[0])
4749 return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], args[0])
4751 sys.AMD64, sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4752 // LeadingZeros is handled because it trivially calls Len.
4753 addF("math/bits", "Reverse64",
4754 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4755 return s.newValue1(ssa.OpBitRev64, types.Types[types.TINT], args[0])
4758 addF("math/bits", "Reverse32",
4759 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4760 return s.newValue1(ssa.OpBitRev32, types.Types[types.TINT], args[0])
4763 addF("math/bits", "Reverse16",
4764 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4765 return s.newValue1(ssa.OpBitRev16, types.Types[types.TINT], args[0])
4768 addF("math/bits", "Reverse8",
4769 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4770 return s.newValue1(ssa.OpBitRev8, types.Types[types.TINT], args[0])
4773 addF("math/bits", "Reverse",
4774 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4775 return s.newValue1(ssa.OpBitRev64, types.Types[types.TINT], args[0])
4778 addF("math/bits", "RotateLeft8",
4779 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4780 return s.newValue2(ssa.OpRotateLeft8, types.Types[types.TUINT8], args[0], args[1])
4783 addF("math/bits", "RotateLeft16",
4784 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4785 return s.newValue2(ssa.OpRotateLeft16, types.Types[types.TUINT16], args[0], args[1])
4788 addF("math/bits", "RotateLeft32",
4789 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4790 return s.newValue2(ssa.OpRotateLeft32, types.Types[types.TUINT32], args[0], args[1])
4792 sys.AMD64, sys.ARM, sys.ARM64, sys.S390X, sys.PPC64, sys.Wasm, sys.Loong64)
4793 addF("math/bits", "RotateLeft64",
4794 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4795 return s.newValue2(ssa.OpRotateLeft64, types.Types[types.TUINT64], args[0], args[1])
4797 sys.AMD64, sys.ARM64, sys.S390X, sys.PPC64, sys.Wasm, sys.Loong64)
4798 alias("math/bits", "RotateLeft", "math/bits", "RotateLeft64", p8...)
4800 makeOnesCountAMD64 := func(op ssa.Op) func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4801 return func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4802 if buildcfg.GOAMD64 >= 2 {
4803 return s.newValue1(op, types.Types[types.TINT], args[0])
4806 v := s.entryNewValue0A(ssa.OpHasCPUFeature, types.Types[types.TBOOL], ir.Syms.X86HasPOPCNT)
4808 b.Kind = ssa.BlockIf
4810 bTrue := s.f.NewBlock(ssa.BlockPlain)
4811 bFalse := s.f.NewBlock(ssa.BlockPlain)
4812 bEnd := s.f.NewBlock(ssa.BlockPlain)
4815 b.Likely = ssa.BranchLikely // most machines have popcnt nowadays
4817 // We have the intrinsic - use it directly.
4819 s.vars[n] = s.newValue1(op, types.Types[types.TINT], args[0])
4820 s.endBlock().AddEdgeTo(bEnd)
4822 // Call the pure Go version.
4823 s.startBlock(bFalse)
4824 s.vars[n] = s.callResult(n, callNormal) // types.Types[TINT]
4825 s.endBlock().AddEdgeTo(bEnd)
4829 return s.variable(n, types.Types[types.TINT])
4832 addF("math/bits", "OnesCount64",
4833 makeOnesCountAMD64(ssa.OpPopCount64),
4835 addF("math/bits", "OnesCount64",
4836 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4837 return s.newValue1(ssa.OpPopCount64, types.Types[types.TINT], args[0])
4839 sys.PPC64, sys.ARM64, sys.S390X, sys.Wasm)
4840 addF("math/bits", "OnesCount32",
4841 makeOnesCountAMD64(ssa.OpPopCount32),
4843 addF("math/bits", "OnesCount32",
4844 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4845 return s.newValue1(ssa.OpPopCount32, types.Types[types.TINT], args[0])
4847 sys.PPC64, sys.ARM64, sys.S390X, sys.Wasm)
4848 addF("math/bits", "OnesCount16",
4849 makeOnesCountAMD64(ssa.OpPopCount16),
4851 addF("math/bits", "OnesCount16",
4852 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4853 return s.newValue1(ssa.OpPopCount16, types.Types[types.TINT], args[0])
4855 sys.ARM64, sys.S390X, sys.PPC64, sys.Wasm)
4856 addF("math/bits", "OnesCount8",
4857 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4858 return s.newValue1(ssa.OpPopCount8, types.Types[types.TINT], args[0])
4860 sys.S390X, sys.PPC64, sys.Wasm)
4861 addF("math/bits", "OnesCount",
4862 makeOnesCountAMD64(ssa.OpPopCount64),
4864 addF("math/bits", "Mul64",
4865 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4866 return s.newValue2(ssa.OpMul64uhilo, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1])
4868 sys.AMD64, sys.ARM64, sys.PPC64, sys.S390X, sys.MIPS64, sys.RISCV64, sys.Loong64)
4869 alias("math/bits", "Mul", "math/bits", "Mul64", p8...)
4870 alias("runtime/internal/math", "Mul64", "math/bits", "Mul64", p8...)
4871 addF("math/bits", "Add64",
4872 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4873 return s.newValue3(ssa.OpAdd64carry, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1], args[2])
4875 sys.AMD64, sys.ARM64, sys.PPC64, sys.S390X, sys.RISCV64, sys.Loong64, sys.MIPS64)
4876 alias("math/bits", "Add", "math/bits", "Add64", p8...)
4877 alias("runtime/internal/math", "Add64", "math/bits", "Add64", all...)
4878 addF("math/bits", "Sub64",
4879 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4880 return s.newValue3(ssa.OpSub64borrow, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1], args[2])
4882 sys.AMD64, sys.ARM64, sys.PPC64, sys.S390X, sys.RISCV64, sys.Loong64, sys.MIPS64)
4883 alias("math/bits", "Sub", "math/bits", "Sub64", p8...)
4884 addF("math/bits", "Div64",
4885 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4886 // check for divide-by-zero/overflow and panic with appropriate message
4887 cmpZero := s.newValue2(s.ssaOp(ir.ONE, types.Types[types.TUINT64]), types.Types[types.TBOOL], args[2], s.zeroVal(types.Types[types.TUINT64]))
4888 s.check(cmpZero, ir.Syms.Panicdivide)
4889 cmpOverflow := s.newValue2(s.ssaOp(ir.OLT, types.Types[types.TUINT64]), types.Types[types.TBOOL], args[0], args[2])
4890 s.check(cmpOverflow, ir.Syms.Panicoverflow)
4891 return s.newValue3(ssa.OpDiv128u, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1], args[2])
4894 alias("math/bits", "Div", "math/bits", "Div64", sys.ArchAMD64)
4896 alias("runtime/internal/sys", "TrailingZeros8", "math/bits", "TrailingZeros8", all...)
4897 alias("runtime/internal/sys", "TrailingZeros32", "math/bits", "TrailingZeros32", all...)
4898 alias("runtime/internal/sys", "TrailingZeros64", "math/bits", "TrailingZeros64", all...)
4899 alias("runtime/internal/sys", "Len8", "math/bits", "Len8", all...)
4900 alias("runtime/internal/sys", "Len64", "math/bits", "Len64", all...)
4901 alias("runtime/internal/sys", "OnesCount64", "math/bits", "OnesCount64", all...)
4903 /******** sync/atomic ********/
4905 // Note: these are disabled by flag_race in findIntrinsic below.
4906 alias("sync/atomic", "LoadInt32", "runtime/internal/atomic", "Load", all...)
4907 alias("sync/atomic", "LoadInt64", "runtime/internal/atomic", "Load64", all...)
4908 alias("sync/atomic", "LoadPointer", "runtime/internal/atomic", "Loadp", all...)
4909 alias("sync/atomic", "LoadUint32", "runtime/internal/atomic", "Load", all...)
4910 alias("sync/atomic", "LoadUint64", "runtime/internal/atomic", "Load64", all...)
4911 alias("sync/atomic", "LoadUintptr", "runtime/internal/atomic", "Load", p4...)
4912 alias("sync/atomic", "LoadUintptr", "runtime/internal/atomic", "Load64", p8...)
4914 alias("sync/atomic", "StoreInt32", "runtime/internal/atomic", "Store", all...)
4915 alias("sync/atomic", "StoreInt64", "runtime/internal/atomic", "Store64", all...)
4916 // Note: not StorePointer, that needs a write barrier. Same below for {CompareAnd}Swap.
4917 alias("sync/atomic", "StoreUint32", "runtime/internal/atomic", "Store", all...)
4918 alias("sync/atomic", "StoreUint64", "runtime/internal/atomic", "Store64", all...)
4919 alias("sync/atomic", "StoreUintptr", "runtime/internal/atomic", "Store", p4...)
4920 alias("sync/atomic", "StoreUintptr", "runtime/internal/atomic", "Store64", p8...)
4922 alias("sync/atomic", "SwapInt32", "runtime/internal/atomic", "Xchg", all...)
4923 alias("sync/atomic", "SwapInt64", "runtime/internal/atomic", "Xchg64", all...)
4924 alias("sync/atomic", "SwapUint32", "runtime/internal/atomic", "Xchg", all...)
4925 alias("sync/atomic", "SwapUint64", "runtime/internal/atomic", "Xchg64", all...)
4926 alias("sync/atomic", "SwapUintptr", "runtime/internal/atomic", "Xchg", p4...)
4927 alias("sync/atomic", "SwapUintptr", "runtime/internal/atomic", "Xchg64", p8...)
4929 alias("sync/atomic", "CompareAndSwapInt32", "runtime/internal/atomic", "Cas", all...)
4930 alias("sync/atomic", "CompareAndSwapInt64", "runtime/internal/atomic", "Cas64", all...)
4931 alias("sync/atomic", "CompareAndSwapUint32", "runtime/internal/atomic", "Cas", all...)
4932 alias("sync/atomic", "CompareAndSwapUint64", "runtime/internal/atomic", "Cas64", all...)
4933 alias("sync/atomic", "CompareAndSwapUintptr", "runtime/internal/atomic", "Cas", p4...)
4934 alias("sync/atomic", "CompareAndSwapUintptr", "runtime/internal/atomic", "Cas64", p8...)
4936 alias("sync/atomic", "AddInt32", "runtime/internal/atomic", "Xadd", all...)
4937 alias("sync/atomic", "AddInt64", "runtime/internal/atomic", "Xadd64", all...)
4938 alias("sync/atomic", "AddUint32", "runtime/internal/atomic", "Xadd", all...)
4939 alias("sync/atomic", "AddUint64", "runtime/internal/atomic", "Xadd64", all...)
4940 alias("sync/atomic", "AddUintptr", "runtime/internal/atomic", "Xadd", p4...)
4941 alias("sync/atomic", "AddUintptr", "runtime/internal/atomic", "Xadd64", p8...)
4943 /******** math/big ********/
4944 alias("math/big", "mulWW", "math/bits", "Mul64", p8...)
4947 // findIntrinsic returns a function which builds the SSA equivalent of the
4948 // function identified by the symbol sym. If sym is not an intrinsic call, returns nil.
4949 func findIntrinsic(sym *types.Sym) intrinsicBuilder {
4950 if sym == nil || sym.Pkg == nil {
4954 if sym.Pkg == ir.Pkgs.Runtime {
4957 if base.Flag.Race && pkg == "sync/atomic" {
4958 // The race detector needs to be able to intercept these calls.
4959 // We can't intrinsify them.
4962 // Skip intrinsifying math functions (which may contain hard-float
4963 // instructions) when soft-float
4964 if Arch.SoftFloat && pkg == "math" {
4969 if ssa.IntrinsicsDisable {
4970 if pkg == "runtime" && (fn == "getcallerpc" || fn == "getcallersp" || fn == "getclosureptr") {
4971 // These runtime functions don't have definitions, must be intrinsics.
4976 return intrinsics[intrinsicKey{Arch.LinkArch.Arch, pkg, fn}]
4979 func IsIntrinsicCall(n *ir.CallExpr) bool {
4983 name, ok := n.X.(*ir.Name)
4987 return findIntrinsic(name.Sym()) != nil
4990 // intrinsicCall converts a call to a recognized intrinsic function into the intrinsic SSA operation.
4991 func (s *state) intrinsicCall(n *ir.CallExpr) *ssa.Value {
4992 v := findIntrinsic(n.X.Sym())(s, n, s.intrinsicArgs(n))
4993 if ssa.IntrinsicsDebug > 0 {
4998 if x.Op == ssa.OpSelect0 || x.Op == ssa.OpSelect1 {
5001 base.WarnfAt(n.Pos(), "intrinsic substitution for %v with %s", n.X.Sym().Name, x.LongString())
5006 // intrinsicArgs extracts args from n, evaluates them to SSA values, and returns them.
5007 func (s *state) intrinsicArgs(n *ir.CallExpr) []*ssa.Value {
5008 args := make([]*ssa.Value, len(n.Args))
5009 for i, n := range n.Args {
5015 // openDeferRecord adds code to evaluate and store the function for an open-code defer
5016 // call, and records info about the defer, so we can generate proper code on the
5017 // exit paths. n is the sub-node of the defer node that is the actual function
5018 // call. We will also record funcdata information on where the function is stored
5019 // (as well as the deferBits variable), and this will enable us to run the proper
5020 // defer calls during panics.
5021 func (s *state) openDeferRecord(n *ir.CallExpr) {
5022 if len(n.Args) != 0 || n.Op() != ir.OCALLFUNC || n.X.Type().NumResults() != 0 {
5023 s.Fatalf("defer call with arguments or results: %v", n)
5026 opendefer := &openDeferInfo{
5030 // We must always store the function value in a stack slot for the
5031 // runtime panic code to use. But in the defer exit code, we will
5032 // call the function directly if it is a static function.
5033 closureVal := s.expr(fn)
5034 closure := s.openDeferSave(fn.Type(), closureVal)
5035 opendefer.closureNode = closure.Aux.(*ir.Name)
5036 if !(fn.Op() == ir.ONAME && fn.(*ir.Name).Class == ir.PFUNC) {
5037 opendefer.closure = closure
5039 index := len(s.openDefers)
5040 s.openDefers = append(s.openDefers, opendefer)
5042 // Update deferBits only after evaluation and storage to stack of
5043 // the function is successful.
5044 bitvalue := s.constInt8(types.Types[types.TUINT8], 1<<uint(index))
5045 newDeferBits := s.newValue2(ssa.OpOr8, types.Types[types.TUINT8], s.variable(deferBitsVar, types.Types[types.TUINT8]), bitvalue)
5046 s.vars[deferBitsVar] = newDeferBits
5047 s.store(types.Types[types.TUINT8], s.deferBitsAddr, newDeferBits)
5050 // openDeferSave generates SSA nodes to store a value (with type t) for an
5051 // open-coded defer at an explicit autotmp location on the stack, so it can be
5052 // reloaded and used for the appropriate call on exit. Type t must be a function type
5053 // (therefore SSAable). val is the value to be stored. The function returns an SSA
5054 // value representing a pointer to the autotmp location.
5055 func (s *state) openDeferSave(t *types.Type, val *ssa.Value) *ssa.Value {
5057 s.Fatalf("openDeferSave of non-SSA-able type %v val=%v", t, val)
5059 if !t.HasPointers() {
5060 s.Fatalf("openDeferSave of pointerless type %v val=%v", t, val)
5063 temp := typecheck.TempAt(pos.WithNotStmt(), s.curfn, t)
5064 temp.SetOpenDeferSlot(true)
5065 temp.SetFrameOffset(int64(len(s.openDefers))) // so cmpstackvarlt can order them
5066 var addrTemp *ssa.Value
5067 // Use OpVarLive to make sure stack slot for the closure is not removed by
5068 // dead-store elimination
5069 if s.curBlock.ID != s.f.Entry.ID {
5070 // Force the tmp storing this defer function to be declared in the entry
5071 // block, so that it will be live for the defer exit code (which will
5072 // actually access it only if the associated defer call has been activated).
5073 if t.HasPointers() {
5074 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])
5076 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])
5077 addrTemp = s.f.Entry.NewValue2A(src.NoXPos, ssa.OpLocalAddr, types.NewPtr(temp.Type()), temp, s.sp, s.defvars[s.f.Entry.ID][memVar])
5079 // Special case if we're still in the entry block. We can't use
5080 // the above code, since s.defvars[s.f.Entry.ID] isn't defined
5081 // until we end the entry block with s.endBlock().
5082 if t.HasPointers() {
5083 s.vars[memVar] = s.newValue1Apos(ssa.OpVarDef, types.TypeMem, temp, s.mem(), false)
5085 s.vars[memVar] = s.newValue1Apos(ssa.OpVarLive, types.TypeMem, temp, s.mem(), false)
5086 addrTemp = s.newValue2Apos(ssa.OpLocalAddr, types.NewPtr(temp.Type()), temp, s.sp, s.mem(), false)
5088 // Since we may use this temp during exit depending on the
5089 // deferBits, we must define it unconditionally on entry.
5090 // Therefore, we must make sure it is zeroed out in the entry
5091 // block if it contains pointers, else GC may wrongly follow an
5092 // uninitialized pointer value.
5093 temp.SetNeedzero(true)
5094 // We are storing to the stack, hence we can avoid the full checks in
5095 // storeType() (no write barrier) and do a simple store().
5096 s.store(t, addrTemp, val)
5100 // openDeferExit generates SSA for processing all the open coded defers at exit.
5101 // The code involves loading deferBits, and checking each of the bits to see if
5102 // the corresponding defer statement was executed. For each bit that is turned
5103 // on, the associated defer call is made.
5104 func (s *state) openDeferExit() {
5105 deferExit := s.f.NewBlock(ssa.BlockPlain)
5106 s.endBlock().AddEdgeTo(deferExit)
5107 s.startBlock(deferExit)
5108 s.lastDeferExit = deferExit
5109 s.lastDeferCount = len(s.openDefers)
5110 zeroval := s.constInt8(types.Types[types.TUINT8], 0)
5111 // Test for and run defers in reverse order
5112 for i := len(s.openDefers) - 1; i >= 0; i-- {
5113 r := s.openDefers[i]
5114 bCond := s.f.NewBlock(ssa.BlockPlain)
5115 bEnd := s.f.NewBlock(ssa.BlockPlain)
5117 deferBits := s.variable(deferBitsVar, types.Types[types.TUINT8])
5118 // Generate code to check if the bit associated with the current
5120 bitval := s.constInt8(types.Types[types.TUINT8], 1<<uint(i))
5121 andval := s.newValue2(ssa.OpAnd8, types.Types[types.TUINT8], deferBits, bitval)
5122 eqVal := s.newValue2(ssa.OpEq8, types.Types[types.TBOOL], andval, zeroval)
5124 b.Kind = ssa.BlockIf
5128 bCond.AddEdgeTo(bEnd)
5131 // Clear this bit in deferBits and force store back to stack, so
5132 // we will not try to re-run this defer call if this defer call panics.
5133 nbitval := s.newValue1(ssa.OpCom8, types.Types[types.TUINT8], bitval)
5134 maskedval := s.newValue2(ssa.OpAnd8, types.Types[types.TUINT8], deferBits, nbitval)
5135 s.store(types.Types[types.TUINT8], s.deferBitsAddr, maskedval)
5136 // Use this value for following tests, so we keep previous
5138 s.vars[deferBitsVar] = maskedval
5140 // Generate code to call the function call of the defer, using the
5141 // closure that were stored in argtmps at the point of the defer
5144 stksize := fn.Type().ArgWidth()
5145 var callArgs []*ssa.Value
5147 if r.closure != nil {
5148 v := s.load(r.closure.Type.Elem(), r.closure)
5149 s.maybeNilCheckClosure(v, callDefer)
5150 codeptr := s.rawLoad(types.Types[types.TUINTPTR], v)
5151 aux := ssa.ClosureAuxCall(s.f.ABIDefault.ABIAnalyzeTypes(nil, nil, nil))
5152 call = s.newValue2A(ssa.OpClosureLECall, aux.LateExpansionResultType(), aux, codeptr, v)
5154 aux := ssa.StaticAuxCall(fn.(*ir.Name).Linksym(), s.f.ABIDefault.ABIAnalyzeTypes(nil, nil, nil))
5155 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
5157 callArgs = append(callArgs, s.mem())
5158 call.AddArgs(callArgs...)
5159 call.AuxInt = stksize
5160 s.vars[memVar] = s.newValue1I(ssa.OpSelectN, types.TypeMem, 0, call)
5161 // Make sure that the stack slots with pointers are kept live
5162 // through the call (which is a pre-emption point). Also, we will
5163 // use the first call of the last defer exit to compute liveness
5164 // for the deferreturn, so we want all stack slots to be live.
5165 if r.closureNode != nil {
5166 s.vars[memVar] = s.newValue1Apos(ssa.OpVarLive, types.TypeMem, r.closureNode, s.mem(), false)
5174 func (s *state) callResult(n *ir.CallExpr, k callKind) *ssa.Value {
5175 return s.call(n, k, false)
5178 func (s *state) callAddr(n *ir.CallExpr, k callKind) *ssa.Value {
5179 return s.call(n, k, true)
5182 // Calls the function n using the specified call type.
5183 // Returns the address of the return value (or nil if none).
5184 func (s *state) call(n *ir.CallExpr, k callKind, returnResultAddr bool) *ssa.Value {
5186 var callee *ir.Name // target function (if static)
5187 var closure *ssa.Value // ptr to closure to run (if dynamic)
5188 var codeptr *ssa.Value // ptr to target code (if dynamic)
5189 var rcvr *ssa.Value // receiver to set
5191 var ACArgs []*types.Type // AuxCall args
5192 var ACResults []*types.Type // AuxCall results
5193 var callArgs []*ssa.Value // For late-expansion, the args themselves (not stored, args to the call instead).
5195 callABI := s.f.ABIDefault
5197 if k != callNormal && k != callTail && (len(n.Args) != 0 || n.Op() == ir.OCALLINTER || n.X.Type().NumResults() != 0) {
5198 s.Fatalf("go/defer call with arguments: %v", n)
5203 if (k == callNormal || k == callTail) && fn.Op() == ir.ONAME && fn.(*ir.Name).Class == ir.PFUNC {
5206 if buildcfg.Experiment.RegabiArgs {
5207 // This is a static call, so it may be
5208 // a direct call to a non-ABIInternal
5209 // function. fn.Func may be nil for
5210 // some compiler-generated functions,
5211 // but those are all ABIInternal.
5213 callABI = abiForFunc(fn.Func, s.f.ABI0, s.f.ABI1)
5216 // TODO(register args) remove after register abi is working
5217 inRegistersImported := fn.Pragma()&ir.RegisterParams != 0
5218 inRegistersSamePackage := fn.Func != nil && fn.Func.Pragma&ir.RegisterParams != 0
5219 if inRegistersImported || inRegistersSamePackage {
5225 closure = s.expr(fn)
5226 if k != callDefer && k != callDeferStack {
5227 // Deferred nil function needs to panic when the function is invoked,
5228 // not the point of defer statement.
5229 s.maybeNilCheckClosure(closure, k)
5232 if fn.Op() != ir.ODOTINTER {
5233 s.Fatalf("OCALLINTER: n.Left not an ODOTINTER: %v", fn.Op())
5235 fn := fn.(*ir.SelectorExpr)
5236 var iclosure *ssa.Value
5237 iclosure, rcvr = s.getClosureAndRcvr(fn)
5238 if k == callNormal {
5239 codeptr = s.load(types.Types[types.TUINTPTR], iclosure)
5245 params := callABI.ABIAnalyze(n.X.Type(), false /* Do not set (register) nNames from caller side -- can cause races. */)
5246 types.CalcSize(fn.Type())
5247 stksize := params.ArgWidth() // includes receiver, args, and results
5249 res := n.X.Type().Results()
5250 if k == callNormal || k == callTail {
5251 for _, p := range params.OutParams() {
5252 ACResults = append(ACResults, p.Type)
5257 if k == callDeferStack {
5258 // Make a defer struct d on the stack.
5260 s.Fatalf("deferprocStack with non-zero stack size %d: %v", stksize, n)
5264 d := typecheck.TempAt(n.Pos(), s.curfn, t)
5266 if t.HasPointers() {
5267 s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, d, s.mem())
5271 s.store(closure.Type,
5272 s.newValue1I(ssa.OpOffPtr, closure.Type.PtrTo(), t.FieldOff(deferStructFnField), addr),
5275 // Call runtime.deferprocStack with pointer to _defer record.
5276 ACArgs = append(ACArgs, types.Types[types.TUINTPTR])
5277 aux := ssa.StaticAuxCall(ir.Syms.DeferprocStack, s.f.ABIDefault.ABIAnalyzeTypes(nil, ACArgs, ACResults))
5278 callArgs = append(callArgs, addr, s.mem())
5279 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
5280 call.AddArgs(callArgs...)
5281 call.AuxInt = int64(types.PtrSize) // deferprocStack takes a *_defer arg
5283 // Store arguments to stack, including defer/go arguments and receiver for method calls.
5284 // These are written in SP-offset order.
5285 argStart := base.Ctxt.Arch.FixedFrameSize
5287 if k != callNormal && k != callTail {
5288 // Write closure (arg to newproc/deferproc).
5289 ACArgs = append(ACArgs, types.Types[types.TUINTPTR]) // not argExtra
5290 callArgs = append(callArgs, closure)
5291 stksize += int64(types.PtrSize)
5292 argStart += int64(types.PtrSize)
5295 // Set receiver (for interface calls).
5297 callArgs = append(callArgs, rcvr)
5304 for _, p := range params.InParams() { // includes receiver for interface calls
5305 ACArgs = append(ACArgs, p.Type)
5308 // Split the entry block if there are open defers, because later calls to
5309 // openDeferSave may cause a mismatch between the mem for an OpDereference
5310 // and the call site which uses it. See #49282.
5311 if s.curBlock.ID == s.f.Entry.ID && s.hasOpenDefers {
5313 b.Kind = ssa.BlockPlain
5314 curb := s.f.NewBlock(ssa.BlockPlain)
5319 for i, n := range args {
5320 callArgs = append(callArgs, s.putArg(n, t.Params().Field(i).Type))
5323 callArgs = append(callArgs, s.mem())
5327 case k == callDefer:
5328 aux := ssa.StaticAuxCall(ir.Syms.Deferproc, s.f.ABIDefault.ABIAnalyzeTypes(nil, ACArgs, ACResults)) // TODO paramResultInfo for DeferProc
5329 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
5331 aux := ssa.StaticAuxCall(ir.Syms.Newproc, s.f.ABIDefault.ABIAnalyzeTypes(nil, ACArgs, ACResults))
5332 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux) // TODO paramResultInfo for NewProc
5333 case closure != nil:
5334 // rawLoad because loading the code pointer from a
5335 // closure is always safe, but IsSanitizerSafeAddr
5336 // can't always figure that out currently, and it's
5337 // critical that we not clobber any arguments already
5338 // stored onto the stack.
5339 codeptr = s.rawLoad(types.Types[types.TUINTPTR], closure)
5340 aux := ssa.ClosureAuxCall(callABI.ABIAnalyzeTypes(nil, ACArgs, ACResults))
5341 call = s.newValue2A(ssa.OpClosureLECall, aux.LateExpansionResultType(), aux, codeptr, closure)
5342 case codeptr != nil:
5343 // Note that the "receiver" parameter is nil because the actual receiver is the first input parameter.
5344 aux := ssa.InterfaceAuxCall(params)
5345 call = s.newValue1A(ssa.OpInterLECall, aux.LateExpansionResultType(), aux, codeptr)
5347 aux := ssa.StaticAuxCall(callTargetLSym(callee), params)
5348 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
5350 call.Op = ssa.OpTailLECall
5351 stksize = 0 // Tail call does not use stack. We reuse caller's frame.
5354 s.Fatalf("bad call type %v %v", n.Op(), n)
5356 call.AddArgs(callArgs...)
5357 call.AuxInt = stksize // Call operations carry the argsize of the callee along with them
5360 s.vars[memVar] = s.newValue1I(ssa.OpSelectN, types.TypeMem, int64(len(ACResults)), call)
5361 // Insert VarLive opcodes.
5362 for _, v := range n.KeepAlive {
5364 s.Fatalf("KeepAlive variable %v must have Addrtaken set", v)
5367 case ir.PAUTO, ir.PPARAM, ir.PPARAMOUT:
5369 s.Fatalf("KeepAlive variable %v must be Auto or Arg", v)
5371 s.vars[memVar] = s.newValue1A(ssa.OpVarLive, types.TypeMem, v, s.mem())
5374 // Finish block for defers
5375 if k == callDefer || k == callDeferStack {
5377 b.Kind = ssa.BlockDefer
5379 bNext := s.f.NewBlock(ssa.BlockPlain)
5381 // Add recover edge to exit code.
5382 r := s.f.NewBlock(ssa.BlockPlain)
5386 b.Likely = ssa.BranchLikely
5390 if res.NumFields() == 0 || k != callNormal {
5391 // call has no return value. Continue with the next statement.
5395 if returnResultAddr {
5396 return s.resultAddrOfCall(call, 0, fp.Type)
5398 return s.newValue1I(ssa.OpSelectN, fp.Type, 0, call)
5401 // maybeNilCheckClosure checks if a nil check of a closure is needed in some
5402 // architecture-dependent situations and, if so, emits the nil check.
5403 func (s *state) maybeNilCheckClosure(closure *ssa.Value, k callKind) {
5404 if Arch.LinkArch.Family == sys.Wasm || buildcfg.GOOS == "aix" && k != callGo {
5405 // 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.
5406 // TODO(neelance): On other architectures this should be eliminated by the optimization steps
5411 // getClosureAndRcvr returns values for the appropriate closure and receiver of an
5413 func (s *state) getClosureAndRcvr(fn *ir.SelectorExpr) (*ssa.Value, *ssa.Value) {
5415 itab := s.newValue1(ssa.OpITab, types.Types[types.TUINTPTR], i)
5417 itabidx := fn.Offset() + 2*int64(types.PtrSize) + 8 // offset of fun field in runtime.itab
5418 closure := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.UintptrPtr, itabidx, itab)
5419 rcvr := s.newValue1(ssa.OpIData, s.f.Config.Types.BytePtr, i)
5420 return closure, rcvr
5423 // etypesign returns the signed-ness of e, for integer/pointer etypes.
5424 // -1 means signed, +1 means unsigned, 0 means non-integer/non-pointer.
5425 func etypesign(e types.Kind) int8 {
5427 case types.TINT8, types.TINT16, types.TINT32, types.TINT64, types.TINT:
5429 case types.TUINT8, types.TUINT16, types.TUINT32, types.TUINT64, types.TUINT, types.TUINTPTR, types.TUNSAFEPTR:
5435 // addr converts the address of the expression n to SSA, adds it to s and returns the SSA result.
5436 // The value that the returned Value represents is guaranteed to be non-nil.
5437 func (s *state) addr(n ir.Node) *ssa.Value {
5438 if n.Op() != ir.ONAME {
5444 s.Fatalf("addr of canSSA expression: %+v", n)
5447 t := types.NewPtr(n.Type())
5448 linksymOffset := func(lsym *obj.LSym, offset int64) *ssa.Value {
5449 v := s.entryNewValue1A(ssa.OpAddr, t, lsym, s.sb)
5450 // TODO: Make OpAddr use AuxInt as well as Aux.
5452 v = s.entryNewValue1I(ssa.OpOffPtr, v.Type, offset, v)
5457 case ir.OLINKSYMOFFSET:
5458 no := n.(*ir.LinksymOffsetExpr)
5459 return linksymOffset(no.Linksym, no.Offset_)
5462 if n.Heapaddr != nil {
5463 return s.expr(n.Heapaddr)
5468 return linksymOffset(n.Linksym(), 0)
5475 s.Fatalf("addr of undeclared ONAME %v. declared: %v", n, s.decladdrs)
5478 return s.newValue2Apos(ssa.OpLocalAddr, t, n, s.sp, s.mem(), !ir.IsAutoTmp(n))
5480 case ir.PPARAMOUT: // Same as PAUTO -- cannot generate LEA early.
5481 // ensure that we reuse symbols for out parameters so
5482 // that cse works on their addresses
5483 return s.newValue2Apos(ssa.OpLocalAddr, t, n, s.sp, s.mem(), true)
5485 s.Fatalf("variable address class %v not implemented", n.Class)
5489 // load return from callee
5490 n := n.(*ir.ResultExpr)
5491 return s.resultAddrOfCall(s.prevCall, n.Index, n.Type())
5493 n := n.(*ir.IndexExpr)
5494 if n.X.Type().IsSlice() {
5496 i := s.expr(n.Index)
5497 len := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], a)
5498 i = s.boundsCheck(i, len, ssa.BoundsIndex, n.Bounded())
5499 p := s.newValue1(ssa.OpSlicePtr, t, a)
5500 return s.newValue2(ssa.OpPtrIndex, t, p, i)
5503 i := s.expr(n.Index)
5504 len := s.constInt(types.Types[types.TINT], n.X.Type().NumElem())
5505 i = s.boundsCheck(i, len, ssa.BoundsIndex, n.Bounded())
5506 return s.newValue2(ssa.OpPtrIndex, types.NewPtr(n.X.Type().Elem()), a, i)
5509 n := n.(*ir.StarExpr)
5510 return s.exprPtr(n.X, n.Bounded(), n.Pos())
5512 n := n.(*ir.SelectorExpr)
5514 return s.newValue1I(ssa.OpOffPtr, t, n.Offset(), p)
5516 n := n.(*ir.SelectorExpr)
5517 p := s.exprPtr(n.X, n.Bounded(), n.Pos())
5518 return s.newValue1I(ssa.OpOffPtr, t, n.Offset(), p)
5520 n := n.(*ir.ConvExpr)
5521 if n.Type() == n.X.Type() {
5525 return s.newValue1(ssa.OpCopy, t, addr) // ensure that addr has the right type
5526 case ir.OCALLFUNC, ir.OCALLINTER:
5527 n := n.(*ir.CallExpr)
5528 return s.callAddr(n, callNormal)
5529 case ir.ODOTTYPE, ir.ODYNAMICDOTTYPE:
5531 if n.Op() == ir.ODOTTYPE {
5532 v, _ = s.dottype(n.(*ir.TypeAssertExpr), false)
5534 v, _ = s.dynamicDottype(n.(*ir.DynamicTypeAssertExpr), false)
5536 if v.Op != ssa.OpLoad {
5537 s.Fatalf("dottype of non-load")
5539 if v.Args[1] != s.mem() {
5540 s.Fatalf("memory no longer live from dottype load")
5544 s.Fatalf("unhandled addr %v", n.Op())
5549 // canSSA reports whether n is SSA-able.
5550 // n must be an ONAME (or an ODOT sequence with an ONAME base).
5551 func (s *state) canSSA(n ir.Node) bool {
5552 if base.Flag.N != 0 {
5557 if nn.Op() == ir.ODOT {
5558 nn := nn.(*ir.SelectorExpr)
5562 if nn.Op() == ir.OINDEX {
5563 nn := nn.(*ir.IndexExpr)
5564 if nn.X.Type().IsArray() {
5571 if n.Op() != ir.ONAME {
5574 return s.canSSAName(n.(*ir.Name)) && TypeOK(n.Type())
5577 func (s *state) canSSAName(name *ir.Name) bool {
5578 if name.Addrtaken() || !name.OnStack() {
5584 // TODO: handle this case? Named return values must be
5585 // in memory so that the deferred function can see them.
5586 // Maybe do: if !strings.HasPrefix(n.String(), "~") { return false }
5587 // Or maybe not, see issue 18860. Even unnamed return values
5588 // must be written back so if a defer recovers, the caller can see them.
5591 if s.cgoUnsafeArgs {
5592 // Cgo effectively takes the address of all result args,
5593 // but the compiler can't see that.
5598 // TODO: try to make more variables SSAable?
5601 // TypeOK reports whether variables of type t are SSA-able.
5602 func TypeOK(t *types.Type) bool {
5604 if t.Size() > int64(4*types.PtrSize) {
5605 // 4*Widthptr is an arbitrary constant. We want it
5606 // to be at least 3*Widthptr so slices can be registerized.
5607 // Too big and we'll introduce too much register pressure.
5612 // We can't do larger arrays because dynamic indexing is
5613 // not supported on SSA variables.
5614 // TODO: allow if all indexes are constant.
5615 if t.NumElem() <= 1 {
5616 return TypeOK(t.Elem())
5620 if t.NumFields() > ssa.MaxStruct {
5623 for _, t1 := range t.Fields().Slice() {
5624 if !TypeOK(t1.Type) {
5634 // exprPtr evaluates n to a pointer and nil-checks it.
5635 func (s *state) exprPtr(n ir.Node, bounded bool, lineno src.XPos) *ssa.Value {
5637 if bounded || n.NonNil() {
5638 if s.f.Frontend().Debug_checknil() && lineno.Line() > 1 {
5639 s.f.Warnl(lineno, "removed nil check")
5647 // nilCheck generates nil pointer checking code.
5648 // Used only for automatically inserted nil checks,
5649 // not for user code like 'x != nil'.
5650 func (s *state) nilCheck(ptr *ssa.Value) {
5651 if base.Debug.DisableNil != 0 || s.curfn.NilCheckDisabled() {
5654 s.newValue2(ssa.OpNilCheck, types.TypeVoid, ptr, s.mem())
5657 // boundsCheck generates bounds checking code. Checks if 0 <= idx <[=] len, branches to exit if not.
5658 // Starts a new block on return.
5659 // On input, len must be converted to full int width and be nonnegative.
5660 // Returns idx converted to full int width.
5661 // If bounded is true then caller guarantees the index is not out of bounds
5662 // (but boundsCheck will still extend the index to full int width).
5663 func (s *state) boundsCheck(idx, len *ssa.Value, kind ssa.BoundsKind, bounded bool) *ssa.Value {
5664 idx = s.extendIndex(idx, len, kind, bounded)
5666 if bounded || base.Flag.B != 0 {
5667 // If bounded or bounds checking is flag-disabled, then no check necessary,
5668 // just return the extended index.
5670 // Here, bounded == true if the compiler generated the index itself,
5671 // such as in the expansion of a slice initializer. These indexes are
5672 // compiler-generated, not Go program variables, so they cannot be
5673 // attacker-controlled, so we can omit Spectre masking as well.
5675 // Note that we do not want to omit Spectre masking in code like:
5677 // if 0 <= i && i < len(x) {
5681 // Lucky for us, bounded==false for that code.
5682 // In that case (handled below), we emit a bound check (and Spectre mask)
5683 // and then the prove pass will remove the bounds check.
5684 // In theory the prove pass could potentially remove certain
5685 // Spectre masks, but it's very delicate and probably better
5686 // to be conservative and leave them all in.
5690 bNext := s.f.NewBlock(ssa.BlockPlain)
5691 bPanic := s.f.NewBlock(ssa.BlockExit)
5693 if !idx.Type.IsSigned() {
5695 case ssa.BoundsIndex:
5696 kind = ssa.BoundsIndexU
5697 case ssa.BoundsSliceAlen:
5698 kind = ssa.BoundsSliceAlenU
5699 case ssa.BoundsSliceAcap:
5700 kind = ssa.BoundsSliceAcapU
5701 case ssa.BoundsSliceB:
5702 kind = ssa.BoundsSliceBU
5703 case ssa.BoundsSlice3Alen:
5704 kind = ssa.BoundsSlice3AlenU
5705 case ssa.BoundsSlice3Acap:
5706 kind = ssa.BoundsSlice3AcapU
5707 case ssa.BoundsSlice3B:
5708 kind = ssa.BoundsSlice3BU
5709 case ssa.BoundsSlice3C:
5710 kind = ssa.BoundsSlice3CU
5715 if kind == ssa.BoundsIndex || kind == ssa.BoundsIndexU {
5716 cmp = s.newValue2(ssa.OpIsInBounds, types.Types[types.TBOOL], idx, len)
5718 cmp = s.newValue2(ssa.OpIsSliceInBounds, types.Types[types.TBOOL], idx, len)
5721 b.Kind = ssa.BlockIf
5723 b.Likely = ssa.BranchLikely
5727 s.startBlock(bPanic)
5728 if Arch.LinkArch.Family == sys.Wasm {
5729 // TODO(khr): figure out how to do "register" based calling convention for bounds checks.
5730 // Should be similar to gcWriteBarrier, but I can't make it work.
5731 s.rtcall(BoundsCheckFunc[kind], false, nil, idx, len)
5733 mem := s.newValue3I(ssa.OpPanicBounds, types.TypeMem, int64(kind), idx, len, s.mem())
5734 s.endBlock().SetControl(mem)
5738 // In Spectre index mode, apply an appropriate mask to avoid speculative out-of-bounds accesses.
5739 if base.Flag.Cfg.SpectreIndex {
5740 op := ssa.OpSpectreIndex
5741 if kind != ssa.BoundsIndex && kind != ssa.BoundsIndexU {
5742 op = ssa.OpSpectreSliceIndex
5744 idx = s.newValue2(op, types.Types[types.TINT], idx, len)
5750 // If cmp (a bool) is false, panic using the given function.
5751 func (s *state) check(cmp *ssa.Value, fn *obj.LSym) {
5753 b.Kind = ssa.BlockIf
5755 b.Likely = ssa.BranchLikely
5756 bNext := s.f.NewBlock(ssa.BlockPlain)
5758 pos := base.Ctxt.PosTable.Pos(line)
5759 fl := funcLine{f: fn, base: pos.Base(), line: pos.Line()}
5760 bPanic := s.panics[fl]
5762 bPanic = s.f.NewBlock(ssa.BlockPlain)
5763 s.panics[fl] = bPanic
5764 s.startBlock(bPanic)
5765 // The panic call takes/returns memory to ensure that the right
5766 // memory state is observed if the panic happens.
5767 s.rtcall(fn, false, nil)
5774 func (s *state) intDivide(n ir.Node, a, b *ssa.Value) *ssa.Value {
5777 case ssa.OpConst8, ssa.OpConst16, ssa.OpConst32, ssa.OpConst64:
5783 // do a size-appropriate check for zero
5784 cmp := s.newValue2(s.ssaOp(ir.ONE, n.Type()), types.Types[types.TBOOL], b, s.zeroVal(n.Type()))
5785 s.check(cmp, ir.Syms.Panicdivide)
5787 return s.newValue2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
5790 // rtcall issues a call to the given runtime function fn with the listed args.
5791 // Returns a slice of results of the given result types.
5792 // The call is added to the end of the current block.
5793 // If returns is false, the block is marked as an exit block.
5794 func (s *state) rtcall(fn *obj.LSym, returns bool, results []*types.Type, args ...*ssa.Value) []*ssa.Value {
5796 // Write args to the stack
5797 off := base.Ctxt.Arch.FixedFrameSize
5798 var callArgs []*ssa.Value
5799 var callArgTypes []*types.Type
5801 for _, arg := range args {
5803 off = types.RoundUp(off, t.Alignment())
5805 callArgs = append(callArgs, arg)
5806 callArgTypes = append(callArgTypes, t)
5809 off = types.RoundUp(off, int64(types.RegSize))
5813 aux := ssa.StaticAuxCall(fn, s.f.ABIDefault.ABIAnalyzeTypes(nil, callArgTypes, results))
5814 callArgs = append(callArgs, s.mem())
5815 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
5816 call.AddArgs(callArgs...)
5817 s.vars[memVar] = s.newValue1I(ssa.OpSelectN, types.TypeMem, int64(len(results)), call)
5822 b.Kind = ssa.BlockExit
5824 call.AuxInt = off - base.Ctxt.Arch.FixedFrameSize
5825 if len(results) > 0 {
5826 s.Fatalf("panic call can't have results")
5832 res := make([]*ssa.Value, len(results))
5833 for i, t := range results {
5834 off = types.RoundUp(off, t.Alignment())
5835 res[i] = s.resultOfCall(call, int64(i), t)
5838 off = types.RoundUp(off, int64(types.PtrSize))
5840 // Remember how much callee stack space we needed.
5846 // do *left = right for type t.
5847 func (s *state) storeType(t *types.Type, left, right *ssa.Value, skip skipMask, leftIsStmt bool) {
5848 s.instrument(t, left, instrumentWrite)
5850 if skip == 0 && (!t.HasPointers() || ssa.IsStackAddr(left)) {
5851 // Known to not have write barrier. Store the whole type.
5852 s.vars[memVar] = s.newValue3Apos(ssa.OpStore, types.TypeMem, t, left, right, s.mem(), leftIsStmt)
5856 // store scalar fields first, so write barrier stores for
5857 // pointer fields can be grouped together, and scalar values
5858 // don't need to be live across the write barrier call.
5859 // TODO: if the writebarrier pass knows how to reorder stores,
5860 // we can do a single store here as long as skip==0.
5861 s.storeTypeScalars(t, left, right, skip)
5862 if skip&skipPtr == 0 && t.HasPointers() {
5863 s.storeTypePtrs(t, left, right)
5867 // do *left = right for all scalar (non-pointer) parts of t.
5868 func (s *state) storeTypeScalars(t *types.Type, left, right *ssa.Value, skip skipMask) {
5870 case t.IsBoolean() || t.IsInteger() || t.IsFloat() || t.IsComplex():
5871 s.store(t, left, right)
5872 case t.IsPtrShaped():
5873 if t.IsPtr() && t.Elem().NotInHeap() {
5874 s.store(t, left, right) // see issue 42032
5876 // otherwise, no scalar fields.
5878 if skip&skipLen != 0 {
5881 len := s.newValue1(ssa.OpStringLen, types.Types[types.TINT], right)
5882 lenAddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, s.config.PtrSize, left)
5883 s.store(types.Types[types.TINT], lenAddr, len)
5885 if skip&skipLen == 0 {
5886 len := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], right)
5887 lenAddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, s.config.PtrSize, left)
5888 s.store(types.Types[types.TINT], lenAddr, len)
5890 if skip&skipCap == 0 {
5891 cap := s.newValue1(ssa.OpSliceCap, types.Types[types.TINT], right)
5892 capAddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, 2*s.config.PtrSize, left)
5893 s.store(types.Types[types.TINT], capAddr, cap)
5895 case t.IsInterface():
5896 // itab field doesn't need a write barrier (even though it is a pointer).
5897 itab := s.newValue1(ssa.OpITab, s.f.Config.Types.BytePtr, right)
5898 s.store(types.Types[types.TUINTPTR], left, itab)
5901 for i := 0; i < n; i++ {
5902 ft := t.FieldType(i)
5903 addr := s.newValue1I(ssa.OpOffPtr, ft.PtrTo(), t.FieldOff(i), left)
5904 val := s.newValue1I(ssa.OpStructSelect, ft, int64(i), right)
5905 s.storeTypeScalars(ft, addr, val, 0)
5907 case t.IsArray() && t.NumElem() == 0:
5909 case t.IsArray() && t.NumElem() == 1:
5910 s.storeTypeScalars(t.Elem(), left, s.newValue1I(ssa.OpArraySelect, t.Elem(), 0, right), 0)
5912 s.Fatalf("bad write barrier type %v", t)
5916 // do *left = right for all pointer parts of t.
5917 func (s *state) storeTypePtrs(t *types.Type, left, right *ssa.Value) {
5919 case t.IsPtrShaped():
5920 if t.IsPtr() && t.Elem().NotInHeap() {
5921 break // see issue 42032
5923 s.store(t, left, right)
5925 ptr := s.newValue1(ssa.OpStringPtr, s.f.Config.Types.BytePtr, right)
5926 s.store(s.f.Config.Types.BytePtr, left, ptr)
5928 elType := types.NewPtr(t.Elem())
5929 ptr := s.newValue1(ssa.OpSlicePtr, elType, right)
5930 s.store(elType, left, ptr)
5931 case t.IsInterface():
5932 // itab field is treated as a scalar.
5933 idata := s.newValue1(ssa.OpIData, s.f.Config.Types.BytePtr, right)
5934 idataAddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.BytePtrPtr, s.config.PtrSize, left)
5935 s.store(s.f.Config.Types.BytePtr, idataAddr, idata)
5938 for i := 0; i < n; i++ {
5939 ft := t.FieldType(i)
5940 if !ft.HasPointers() {
5943 addr := s.newValue1I(ssa.OpOffPtr, ft.PtrTo(), t.FieldOff(i), left)
5944 val := s.newValue1I(ssa.OpStructSelect, ft, int64(i), right)
5945 s.storeTypePtrs(ft, addr, val)
5947 case t.IsArray() && t.NumElem() == 0:
5949 case t.IsArray() && t.NumElem() == 1:
5950 s.storeTypePtrs(t.Elem(), left, s.newValue1I(ssa.OpArraySelect, t.Elem(), 0, right))
5952 s.Fatalf("bad write barrier type %v", t)
5956 // putArg evaluates n for the purpose of passing it as an argument to a function and returns the value for the call.
5957 func (s *state) putArg(n ir.Node, t *types.Type) *ssa.Value {
5960 a = s.newValue2(ssa.OpDereference, t, s.addr(n), s.mem())
5967 func (s *state) storeArgWithBase(n ir.Node, t *types.Type, base *ssa.Value, off int64) {
5968 pt := types.NewPtr(t)
5971 // Use special routine that avoids allocation on duplicate offsets.
5972 addr = s.constOffPtrSP(pt, off)
5974 addr = s.newValue1I(ssa.OpOffPtr, pt, off, base)
5984 s.storeType(t, addr, a, 0, false)
5987 // slice computes the slice v[i:j:k] and returns ptr, len, and cap of result.
5988 // i,j,k may be nil, in which case they are set to their default value.
5989 // v may be a slice, string or pointer to an array.
5990 func (s *state) slice(v, i, j, k *ssa.Value, bounded bool) (p, l, c *ssa.Value) {
5992 var ptr, len, cap *ssa.Value
5995 ptr = s.newValue1(ssa.OpSlicePtr, types.NewPtr(t.Elem()), v)
5996 len = s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], v)
5997 cap = s.newValue1(ssa.OpSliceCap, types.Types[types.TINT], v)
5999 ptr = s.newValue1(ssa.OpStringPtr, types.NewPtr(types.Types[types.TUINT8]), v)
6000 len = s.newValue1(ssa.OpStringLen, types.Types[types.TINT], v)
6003 if !t.Elem().IsArray() {
6004 s.Fatalf("bad ptr to array in slice %v\n", t)
6007 ptr = s.newValue1(ssa.OpCopy, types.NewPtr(t.Elem().Elem()), v)
6008 len = s.constInt(types.Types[types.TINT], t.Elem().NumElem())
6011 s.Fatalf("bad type in slice %v\n", t)
6014 // Set default values
6016 i = s.constInt(types.Types[types.TINT], 0)
6027 // Panic if slice indices are not in bounds.
6028 // Make sure we check these in reverse order so that we're always
6029 // comparing against a value known to be nonnegative. See issue 28797.
6032 kind := ssa.BoundsSlice3Alen
6034 kind = ssa.BoundsSlice3Acap
6036 k = s.boundsCheck(k, cap, kind, bounded)
6039 j = s.boundsCheck(j, k, ssa.BoundsSlice3B, bounded)
6041 i = s.boundsCheck(i, j, ssa.BoundsSlice3C, bounded)
6044 kind := ssa.BoundsSliceAlen
6046 kind = ssa.BoundsSliceAcap
6048 j = s.boundsCheck(j, k, kind, bounded)
6050 i = s.boundsCheck(i, j, ssa.BoundsSliceB, bounded)
6053 // Word-sized integer operations.
6054 subOp := s.ssaOp(ir.OSUB, types.Types[types.TINT])
6055 mulOp := s.ssaOp(ir.OMUL, types.Types[types.TINT])
6056 andOp := s.ssaOp(ir.OAND, types.Types[types.TINT])
6058 // Calculate the length (rlen) and capacity (rcap) of the new slice.
6059 // For strings the capacity of the result is unimportant. However,
6060 // we use rcap to test if we've generated a zero-length slice.
6061 // Use length of strings for that.
6062 rlen := s.newValue2(subOp, types.Types[types.TINT], j, i)
6064 if j != k && !t.IsString() {
6065 rcap = s.newValue2(subOp, types.Types[types.TINT], k, i)
6068 if (i.Op == ssa.OpConst64 || i.Op == ssa.OpConst32) && i.AuxInt == 0 {
6069 // No pointer arithmetic necessary.
6070 return ptr, rlen, rcap
6073 // Calculate the base pointer (rptr) for the new slice.
6075 // Generate the following code assuming that indexes are in bounds.
6076 // The masking is to make sure that we don't generate a slice
6077 // that points to the next object in memory. We cannot just set
6078 // the pointer to nil because then we would create a nil slice or
6083 // rptr = ptr + (mask(rcap) & (i * stride))
6085 // Where mask(x) is 0 if x==0 and -1 if x>0 and stride is the width
6086 // of the element type.
6087 stride := s.constInt(types.Types[types.TINT], ptr.Type.Elem().Size())
6089 // The delta is the number of bytes to offset ptr by.
6090 delta := s.newValue2(mulOp, types.Types[types.TINT], i, stride)
6092 // If we're slicing to the point where the capacity is zero,
6093 // zero out the delta.
6094 mask := s.newValue1(ssa.OpSlicemask, types.Types[types.TINT], rcap)
6095 delta = s.newValue2(andOp, types.Types[types.TINT], delta, mask)
6097 // Compute rptr = ptr + delta.
6098 rptr := s.newValue2(ssa.OpAddPtr, ptr.Type, ptr, delta)
6100 return rptr, rlen, rcap
6103 type u642fcvtTab struct {
6104 leq, cvt2F, and, rsh, or, add ssa.Op
6105 one func(*state, *types.Type, int64) *ssa.Value
6108 var u64_f64 = u642fcvtTab{
6110 cvt2F: ssa.OpCvt64to64F,
6112 rsh: ssa.OpRsh64Ux64,
6115 one: (*state).constInt64,
6118 var u64_f32 = u642fcvtTab{
6120 cvt2F: ssa.OpCvt64to32F,
6122 rsh: ssa.OpRsh64Ux64,
6125 one: (*state).constInt64,
6128 func (s *state) uint64Tofloat64(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6129 return s.uint64Tofloat(&u64_f64, n, x, ft, tt)
6132 func (s *state) uint64Tofloat32(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6133 return s.uint64Tofloat(&u64_f32, n, x, ft, tt)
6136 func (s *state) uint64Tofloat(cvttab *u642fcvtTab, n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6138 // result = (floatY) x
6140 // y = uintX(x) ; y = x & 1
6141 // z = uintX(x) ; z = z >> 1
6143 // result = floatY(z)
6144 // result = result + result
6147 // Code borrowed from old code generator.
6148 // What's going on: large 64-bit "unsigned" looks like
6149 // negative number to hardware's integer-to-float
6150 // conversion. However, because the mantissa is only
6151 // 63 bits, we don't need the LSB, so instead we do an
6152 // unsigned right shift (divide by two), convert, and
6153 // double. However, before we do that, we need to be
6154 // sure that we do not lose a "1" if that made the
6155 // difference in the resulting rounding. Therefore, we
6156 // preserve it, and OR (not ADD) it back in. The case
6157 // that matters is when the eleven discarded bits are
6158 // equal to 10000000001; that rounds up, and the 1 cannot
6159 // be lost else it would round down if the LSB of the
6160 // candidate mantissa is 0.
6161 cmp := s.newValue2(cvttab.leq, types.Types[types.TBOOL], s.zeroVal(ft), x)
6163 b.Kind = ssa.BlockIf
6165 b.Likely = ssa.BranchLikely
6167 bThen := s.f.NewBlock(ssa.BlockPlain)
6168 bElse := s.f.NewBlock(ssa.BlockPlain)
6169 bAfter := s.f.NewBlock(ssa.BlockPlain)
6173 a0 := s.newValue1(cvttab.cvt2F, tt, x)
6176 bThen.AddEdgeTo(bAfter)
6180 one := cvttab.one(s, ft, 1)
6181 y := s.newValue2(cvttab.and, ft, x, one)
6182 z := s.newValue2(cvttab.rsh, ft, x, one)
6183 z = s.newValue2(cvttab.or, ft, z, y)
6184 a := s.newValue1(cvttab.cvt2F, tt, z)
6185 a1 := s.newValue2(cvttab.add, tt, a, a)
6188 bElse.AddEdgeTo(bAfter)
6190 s.startBlock(bAfter)
6191 return s.variable(n, n.Type())
6194 type u322fcvtTab struct {
6195 cvtI2F, cvtF2F ssa.Op
6198 var u32_f64 = u322fcvtTab{
6199 cvtI2F: ssa.OpCvt32to64F,
6203 var u32_f32 = u322fcvtTab{
6204 cvtI2F: ssa.OpCvt32to32F,
6205 cvtF2F: ssa.OpCvt64Fto32F,
6208 func (s *state) uint32Tofloat64(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6209 return s.uint32Tofloat(&u32_f64, n, x, ft, tt)
6212 func (s *state) uint32Tofloat32(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6213 return s.uint32Tofloat(&u32_f32, n, x, ft, tt)
6216 func (s *state) uint32Tofloat(cvttab *u322fcvtTab, n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6218 // result = floatY(x)
6220 // result = floatY(float64(x) + (1<<32))
6222 cmp := s.newValue2(ssa.OpLeq32, types.Types[types.TBOOL], s.zeroVal(ft), x)
6224 b.Kind = ssa.BlockIf
6226 b.Likely = ssa.BranchLikely
6228 bThen := s.f.NewBlock(ssa.BlockPlain)
6229 bElse := s.f.NewBlock(ssa.BlockPlain)
6230 bAfter := s.f.NewBlock(ssa.BlockPlain)
6234 a0 := s.newValue1(cvttab.cvtI2F, tt, x)
6237 bThen.AddEdgeTo(bAfter)
6241 a1 := s.newValue1(ssa.OpCvt32to64F, types.Types[types.TFLOAT64], x)
6242 twoToThe32 := s.constFloat64(types.Types[types.TFLOAT64], float64(1<<32))
6243 a2 := s.newValue2(ssa.OpAdd64F, types.Types[types.TFLOAT64], a1, twoToThe32)
6244 a3 := s.newValue1(cvttab.cvtF2F, tt, a2)
6248 bElse.AddEdgeTo(bAfter)
6250 s.startBlock(bAfter)
6251 return s.variable(n, n.Type())
6254 // referenceTypeBuiltin generates code for the len/cap builtins for maps and channels.
6255 func (s *state) referenceTypeBuiltin(n *ir.UnaryExpr, x *ssa.Value) *ssa.Value {
6256 if !n.X.Type().IsMap() && !n.X.Type().IsChan() {
6257 s.Fatalf("node must be a map or a channel")
6263 // return *((*int)n)
6265 // return *(((*int)n)+1)
6268 nilValue := s.constNil(types.Types[types.TUINTPTR])
6269 cmp := s.newValue2(ssa.OpEqPtr, types.Types[types.TBOOL], x, nilValue)
6271 b.Kind = ssa.BlockIf
6273 b.Likely = ssa.BranchUnlikely
6275 bThen := s.f.NewBlock(ssa.BlockPlain)
6276 bElse := s.f.NewBlock(ssa.BlockPlain)
6277 bAfter := s.f.NewBlock(ssa.BlockPlain)
6279 // length/capacity of a nil map/chan is zero
6282 s.vars[n] = s.zeroVal(lenType)
6284 bThen.AddEdgeTo(bAfter)
6290 // length is stored in the first word for map/chan
6291 s.vars[n] = s.load(lenType, x)
6293 // capacity is stored in the second word for chan
6294 sw := s.newValue1I(ssa.OpOffPtr, lenType.PtrTo(), lenType.Size(), x)
6295 s.vars[n] = s.load(lenType, sw)
6297 s.Fatalf("op must be OLEN or OCAP")
6300 bElse.AddEdgeTo(bAfter)
6302 s.startBlock(bAfter)
6303 return s.variable(n, lenType)
6306 type f2uCvtTab struct {
6307 ltf, cvt2U, subf, or ssa.Op
6308 floatValue func(*state, *types.Type, float64) *ssa.Value
6309 intValue func(*state, *types.Type, int64) *ssa.Value
6313 var f32_u64 = f2uCvtTab{
6315 cvt2U: ssa.OpCvt32Fto64,
6318 floatValue: (*state).constFloat32,
6319 intValue: (*state).constInt64,
6323 var f64_u64 = f2uCvtTab{
6325 cvt2U: ssa.OpCvt64Fto64,
6328 floatValue: (*state).constFloat64,
6329 intValue: (*state).constInt64,
6333 var f32_u32 = f2uCvtTab{
6335 cvt2U: ssa.OpCvt32Fto32,
6338 floatValue: (*state).constFloat32,
6339 intValue: func(s *state, t *types.Type, v int64) *ssa.Value { return s.constInt32(t, int32(v)) },
6343 var f64_u32 = f2uCvtTab{
6345 cvt2U: ssa.OpCvt64Fto32,
6348 floatValue: (*state).constFloat64,
6349 intValue: func(s *state, t *types.Type, v int64) *ssa.Value { return s.constInt32(t, int32(v)) },
6353 func (s *state) float32ToUint64(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6354 return s.floatToUint(&f32_u64, n, x, ft, tt)
6356 func (s *state) float64ToUint64(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6357 return s.floatToUint(&f64_u64, n, x, ft, tt)
6360 func (s *state) float32ToUint32(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6361 return s.floatToUint(&f32_u32, n, x, ft, tt)
6364 func (s *state) float64ToUint32(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6365 return s.floatToUint(&f64_u32, n, x, ft, tt)
6368 func (s *state) floatToUint(cvttab *f2uCvtTab, n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6369 // cutoff:=1<<(intY_Size-1)
6370 // if x < floatX(cutoff) {
6371 // result = uintY(x)
6373 // y = x - floatX(cutoff)
6375 // result = z | -(cutoff)
6377 cutoff := cvttab.floatValue(s, ft, float64(cvttab.cutoff))
6378 cmp := s.newValue2(cvttab.ltf, types.Types[types.TBOOL], x, cutoff)
6380 b.Kind = ssa.BlockIf
6382 b.Likely = ssa.BranchLikely
6384 bThen := s.f.NewBlock(ssa.BlockPlain)
6385 bElse := s.f.NewBlock(ssa.BlockPlain)
6386 bAfter := s.f.NewBlock(ssa.BlockPlain)
6390 a0 := s.newValue1(cvttab.cvt2U, tt, x)
6393 bThen.AddEdgeTo(bAfter)
6397 y := s.newValue2(cvttab.subf, ft, x, cutoff)
6398 y = s.newValue1(cvttab.cvt2U, tt, y)
6399 z := cvttab.intValue(s, tt, int64(-cvttab.cutoff))
6400 a1 := s.newValue2(cvttab.or, tt, y, z)
6403 bElse.AddEdgeTo(bAfter)
6405 s.startBlock(bAfter)
6406 return s.variable(n, n.Type())
6409 // dottype generates SSA for a type assertion node.
6410 // commaok indicates whether to panic or return a bool.
6411 // If commaok is false, resok will be nil.
6412 func (s *state) dottype(n *ir.TypeAssertExpr, commaok bool) (res, resok *ssa.Value) {
6413 iface := s.expr(n.X) // input interface
6414 target := s.reflectType(n.Type()) // target type
6415 var targetItab *ssa.Value
6417 targetItab = s.expr(n.ITab)
6419 return s.dottype1(n.Pos(), n.X.Type(), n.Type(), iface, nil, target, targetItab, commaok)
6422 func (s *state) dynamicDottype(n *ir.DynamicTypeAssertExpr, commaok bool) (res, resok *ssa.Value) {
6423 iface := s.expr(n.X)
6424 var source, target, targetItab *ssa.Value
6425 if n.SrcRType != nil {
6426 source = s.expr(n.SrcRType)
6428 if !n.X.Type().IsEmptyInterface() && !n.Type().IsInterface() {
6429 byteptr := s.f.Config.Types.BytePtr
6430 targetItab = s.expr(n.ITab)
6431 // TODO(mdempsky): Investigate whether compiling n.RType could be
6432 // better than loading itab.typ.
6433 target = s.load(byteptr, s.newValue1I(ssa.OpOffPtr, byteptr, int64(types.PtrSize), targetItab)) // itab.typ
6435 target = s.expr(n.RType)
6437 return s.dottype1(n.Pos(), n.X.Type(), n.Type(), iface, source, target, targetItab, commaok)
6440 // dottype1 implements a x.(T) operation. iface is the argument (x), dst is the type we're asserting to (T)
6441 // and src is the type we're asserting from.
6442 // source is the *runtime._type of src
6443 // target is the *runtime._type of dst.
6444 // If src is a nonempty interface and dst is not an interface, targetItab is an itab representing (dst, src). Otherwise it is nil.
6445 // commaok is true if the caller wants a boolean success value. Otherwise, the generated code panics if the conversion fails.
6446 func (s *state) dottype1(pos src.XPos, src, dst *types.Type, iface, source, target, targetItab *ssa.Value, commaok bool) (res, resok *ssa.Value) {
6447 byteptr := s.f.Config.Types.BytePtr
6448 if dst.IsInterface() {
6449 if dst.IsEmptyInterface() {
6450 // Converting to an empty interface.
6451 // Input could be an empty or nonempty interface.
6452 if base.Debug.TypeAssert > 0 {
6453 base.WarnfAt(pos, "type assertion inlined")
6456 // Get itab/type field from input.
6457 itab := s.newValue1(ssa.OpITab, byteptr, iface)
6458 // Conversion succeeds iff that field is not nil.
6459 cond := s.newValue2(ssa.OpNeqPtr, types.Types[types.TBOOL], itab, s.constNil(byteptr))
6461 if src.IsEmptyInterface() && commaok {
6462 // Converting empty interface to empty interface with ,ok is just a nil check.
6466 // Branch on nilness.
6468 b.Kind = ssa.BlockIf
6470 b.Likely = ssa.BranchLikely
6471 bOk := s.f.NewBlock(ssa.BlockPlain)
6472 bFail := s.f.NewBlock(ssa.BlockPlain)
6477 // On failure, panic by calling panicnildottype.
6479 s.rtcall(ir.Syms.Panicnildottype, false, nil, target)
6481 // On success, return (perhaps modified) input interface.
6483 if src.IsEmptyInterface() {
6484 res = iface // Use input interface unchanged.
6487 // Load type out of itab, build interface with existing idata.
6488 off := s.newValue1I(ssa.OpOffPtr, byteptr, int64(types.PtrSize), itab)
6489 typ := s.load(byteptr, off)
6490 idata := s.newValue1(ssa.OpIData, byteptr, iface)
6491 res = s.newValue2(ssa.OpIMake, dst, typ, idata)
6496 // nonempty -> empty
6497 // Need to load type from itab
6498 off := s.newValue1I(ssa.OpOffPtr, byteptr, int64(types.PtrSize), itab)
6499 s.vars[typVar] = s.load(byteptr, off)
6502 // itab is nil, might as well use that as the nil result.
6504 s.vars[typVar] = itab
6508 bEnd := s.f.NewBlock(ssa.BlockPlain)
6510 bFail.AddEdgeTo(bEnd)
6512 idata := s.newValue1(ssa.OpIData, byteptr, iface)
6513 res = s.newValue2(ssa.OpIMake, dst, s.variable(typVar, byteptr), idata)
6515 delete(s.vars, typVar) // no practical effect, just to indicate typVar is no longer live.
6518 // converting to a nonempty interface needs a runtime call.
6519 if base.Debug.TypeAssert > 0 {
6520 base.WarnfAt(pos, "type assertion not inlined")
6523 fn := ir.Syms.AssertI2I
6524 if src.IsEmptyInterface() {
6525 fn = ir.Syms.AssertE2I
6527 data := s.newValue1(ssa.OpIData, types.Types[types.TUNSAFEPTR], iface)
6528 tab := s.newValue1(ssa.OpITab, byteptr, iface)
6529 tab = s.rtcall(fn, true, []*types.Type{byteptr}, target, tab)[0]
6530 return s.newValue2(ssa.OpIMake, dst, tab, data), nil
6532 fn := ir.Syms.AssertI2I2
6533 if src.IsEmptyInterface() {
6534 fn = ir.Syms.AssertE2I2
6536 res = s.rtcall(fn, true, []*types.Type{dst}, target, iface)[0]
6537 resok = s.newValue2(ssa.OpNeqInter, types.Types[types.TBOOL], res, s.constInterface(dst))
6541 if base.Debug.TypeAssert > 0 {
6542 base.WarnfAt(pos, "type assertion inlined")
6545 // Converting to a concrete type.
6546 direct := types.IsDirectIface(dst)
6547 itab := s.newValue1(ssa.OpITab, byteptr, iface) // type word of interface
6548 if base.Debug.TypeAssert > 0 {
6549 base.WarnfAt(pos, "type assertion inlined")
6551 var wantedFirstWord *ssa.Value
6552 if src.IsEmptyInterface() {
6553 // Looking for pointer to target type.
6554 wantedFirstWord = target
6556 // Looking for pointer to itab for target type and source interface.
6557 wantedFirstWord = targetItab
6560 var tmp ir.Node // temporary for use with large types
6561 var addr *ssa.Value // address of tmp
6562 if commaok && !TypeOK(dst) {
6563 // unSSAable type, use temporary.
6564 // TODO: get rid of some of these temporaries.
6565 tmp, addr = s.temp(pos, dst)
6568 cond := s.newValue2(ssa.OpEqPtr, types.Types[types.TBOOL], itab, wantedFirstWord)
6570 b.Kind = ssa.BlockIf
6572 b.Likely = ssa.BranchLikely
6574 bOk := s.f.NewBlock(ssa.BlockPlain)
6575 bFail := s.f.NewBlock(ssa.BlockPlain)
6580 // on failure, panic by calling panicdottype
6584 taddr = s.reflectType(src)
6586 if src.IsEmptyInterface() {
6587 s.rtcall(ir.Syms.PanicdottypeE, false, nil, itab, target, taddr)
6589 s.rtcall(ir.Syms.PanicdottypeI, false, nil, itab, target, taddr)
6592 // on success, return data from interface
6595 return s.newValue1(ssa.OpIData, dst, iface), nil
6597 p := s.newValue1(ssa.OpIData, types.NewPtr(dst), iface)
6598 return s.load(dst, p), nil
6601 // commaok is the more complicated case because we have
6602 // a control flow merge point.
6603 bEnd := s.f.NewBlock(ssa.BlockPlain)
6604 // Note that we need a new valVar each time (unlike okVar where we can
6605 // reuse the variable) because it might have a different type every time.
6606 valVar := ssaMarker("val")
6608 // type assertion succeeded
6612 s.vars[valVar] = s.newValue1(ssa.OpIData, dst, iface)
6614 p := s.newValue1(ssa.OpIData, types.NewPtr(dst), iface)
6615 s.vars[valVar] = s.load(dst, p)
6618 p := s.newValue1(ssa.OpIData, types.NewPtr(dst), iface)
6619 s.move(dst, addr, p)
6621 s.vars[okVar] = s.constBool(true)
6625 // type assertion failed
6628 s.vars[valVar] = s.zeroVal(dst)
6632 s.vars[okVar] = s.constBool(false)
6634 bFail.AddEdgeTo(bEnd)
6639 res = s.variable(valVar, dst)
6640 delete(s.vars, valVar) // no practical effect, just to indicate typVar is no longer live.
6642 res = s.load(dst, addr)
6644 resok = s.variable(okVar, types.Types[types.TBOOL])
6645 delete(s.vars, okVar) // ditto
6649 // temp allocates a temp of type t at position pos
6650 func (s *state) temp(pos src.XPos, t *types.Type) (*ir.Name, *ssa.Value) {
6651 tmp := typecheck.TempAt(pos, s.curfn, t)
6652 if t.HasPointers() {
6653 s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, tmp, s.mem())
6659 // variable returns the value of a variable at the current location.
6660 func (s *state) variable(n ir.Node, t *types.Type) *ssa.Value {
6670 if s.curBlock == s.f.Entry {
6671 // No variable should be live at entry.
6672 s.f.Fatalf("value %v (%v) incorrectly live at entry", n, v)
6674 // Make a FwdRef, which records a value that's live on block input.
6675 // We'll find the matching definition as part of insertPhis.
6676 v = s.newValue0A(ssa.OpFwdRef, t, fwdRefAux{N: n})
6678 if n.Op() == ir.ONAME {
6679 s.addNamedValue(n.(*ir.Name), v)
6684 func (s *state) mem() *ssa.Value {
6685 return s.variable(memVar, types.TypeMem)
6688 func (s *state) addNamedValue(n *ir.Name, v *ssa.Value) {
6689 if n.Class == ir.Pxxx {
6690 // Don't track our marker nodes (memVar etc.).
6693 if ir.IsAutoTmp(n) {
6694 // Don't track temporary variables.
6697 if n.Class == ir.PPARAMOUT {
6698 // Don't track named output values. This prevents return values
6699 // from being assigned too early. See #14591 and #14762. TODO: allow this.
6702 loc := ssa.LocalSlot{N: n, Type: n.Type(), Off: 0}
6703 values, ok := s.f.NamedValues[loc]
6705 s.f.Names = append(s.f.Names, &loc)
6706 s.f.CanonicalLocalSlots[loc] = &loc
6708 s.f.NamedValues[loc] = append(values, v)
6711 // Branch is an unresolved branch.
6712 type Branch struct {
6713 P *obj.Prog // branch instruction
6714 B *ssa.Block // target
6717 // State contains state needed during Prog generation.
6723 // Branches remembers all the branch instructions we've seen
6724 // and where they would like to go.
6727 // JumpTables remembers all the jump tables we've seen.
6728 JumpTables []*ssa.Block
6730 // bstart remembers where each block starts (indexed by block ID)
6733 maxarg int64 // largest frame size for arguments to calls made by the function
6735 // Map from GC safe points to liveness index, generated by
6736 // liveness analysis.
6737 livenessMap liveness.Map
6739 // partLiveArgs includes arguments that may be partially live, for which we
6740 // need to generate instructions that spill the argument registers.
6741 partLiveArgs map[*ir.Name]bool
6743 // lineRunStart records the beginning of the current run of instructions
6744 // within a single block sharing the same line number
6745 // Used to move statement marks to the beginning of such runs.
6746 lineRunStart *obj.Prog
6748 // wasm: The number of values on the WebAssembly stack. This is only used as a safeguard.
6749 OnWasmStackSkipped int
6752 func (s *State) FuncInfo() *obj.FuncInfo {
6753 return s.pp.CurFunc.LSym.Func()
6756 // Prog appends a new Prog.
6757 func (s *State) Prog(as obj.As) *obj.Prog {
6759 if objw.LosesStmtMark(as) {
6762 // Float a statement start to the beginning of any same-line run.
6763 // lineRunStart is reset at block boundaries, which appears to work well.
6764 if s.lineRunStart == nil || s.lineRunStart.Pos.Line() != p.Pos.Line() {
6766 } else if p.Pos.IsStmt() == src.PosIsStmt {
6767 s.lineRunStart.Pos = s.lineRunStart.Pos.WithIsStmt()
6768 p.Pos = p.Pos.WithNotStmt()
6773 // Pc returns the current Prog.
6774 func (s *State) Pc() *obj.Prog {
6778 // SetPos sets the current source position.
6779 func (s *State) SetPos(pos src.XPos) {
6783 // Br emits a single branch instruction and returns the instruction.
6784 // Not all architectures need the returned instruction, but otherwise
6785 // the boilerplate is common to all.
6786 func (s *State) Br(op obj.As, target *ssa.Block) *obj.Prog {
6788 p.To.Type = obj.TYPE_BRANCH
6789 s.Branches = append(s.Branches, Branch{P: p, B: target})
6793 // DebugFriendlySetPosFrom adjusts Pos.IsStmt subject to heuristics
6794 // that reduce "jumpy" line number churn when debugging.
6795 // Spill/fill/copy instructions from the register allocator,
6796 // phi functions, and instructions with a no-pos position
6797 // are examples of instructions that can cause churn.
6798 func (s *State) DebugFriendlySetPosFrom(v *ssa.Value) {
6800 case ssa.OpPhi, ssa.OpCopy, ssa.OpLoadReg, ssa.OpStoreReg:
6801 // These are not statements
6802 s.SetPos(v.Pos.WithNotStmt())
6805 if p != src.NoXPos {
6806 // If the position is defined, update the position.
6807 // Also convert default IsStmt to NotStmt; only
6808 // explicit statement boundaries should appear
6809 // in the generated code.
6810 if p.IsStmt() != src.PosIsStmt {
6811 if s.pp.Pos.IsStmt() == src.PosIsStmt && s.pp.Pos.SameFileAndLine(p) {
6812 // If s.pp.Pos already has a statement mark, then it was set here (below) for
6813 // the previous value. If an actual instruction had been emitted for that
6814 // value, then the statement mark would have been reset. Since the statement
6815 // mark of s.pp.Pos was not reset, this position (file/line) still needs a
6816 // statement mark on an instruction. If file and line for this value are
6817 // the same as the previous value, then the first instruction for this
6818 // value will work to take the statement mark. Return early to avoid
6819 // resetting the statement mark.
6821 // The reset of s.pp.Pos occurs in (*Progs).Prog() -- if it emits
6822 // an instruction, and the instruction's statement mark was set,
6823 // and it is not one of the LosesStmtMark instructions,
6824 // then Prog() resets the statement mark on the (*Progs).Pos.
6828 // Calls use the pos attached to v, but copy the statement mark from State
6832 s.SetPos(s.pp.Pos.WithNotStmt())
6837 // emit argument info (locations on stack) for traceback.
6838 func emitArgInfo(e *ssafn, f *ssa.Func, pp *objw.Progs) {
6839 ft := e.curfn.Type()
6840 if ft.NumRecvs() == 0 && ft.NumParams() == 0 {
6844 x := EmitArgInfo(e.curfn, f.OwnAux.ABIInfo())
6845 x.Set(obj.AttrContentAddressable, true)
6846 e.curfn.LSym.Func().ArgInfo = x
6848 // Emit a funcdata pointing at the arg info data.
6849 p := pp.Prog(obj.AFUNCDATA)
6850 p.From.SetConst(rtabi.FUNCDATA_ArgInfo)
6851 p.To.Type = obj.TYPE_MEM
6852 p.To.Name = obj.NAME_EXTERN
6856 // emit argument info (locations on stack) of f for traceback.
6857 func EmitArgInfo(f *ir.Func, abiInfo *abi.ABIParamResultInfo) *obj.LSym {
6858 x := base.Ctxt.Lookup(fmt.Sprintf("%s.arginfo%d", f.LSym.Name, f.ABI))
6859 // NOTE: do not set ContentAddressable here. This may be referenced from
6860 // assembly code by name (in this case f is a declaration).
6861 // Instead, set it in emitArgInfo above.
6863 PtrSize := int64(types.PtrSize)
6864 uintptrTyp := types.Types[types.TUINTPTR]
6866 isAggregate := func(t *types.Type) bool {
6867 return t.IsStruct() || t.IsArray() || t.IsComplex() || t.IsInterface() || t.IsString() || t.IsSlice()
6870 // Populate the data.
6871 // The data is a stream of bytes, which contains the offsets and sizes of the
6872 // non-aggregate arguments or non-aggregate fields/elements of aggregate-typed
6873 // arguments, along with special "operators". Specifically,
6874 // - for each non-aggrgate arg/field/element, its offset from FP (1 byte) and
6876 // - special operators:
6877 // - 0xff - end of sequence
6878 // - 0xfe - print { (at the start of an aggregate-typed argument)
6879 // - 0xfd - print } (at the end of an aggregate-typed argument)
6880 // - 0xfc - print ... (more args/fields/elements)
6881 // - 0xfb - print _ (offset too large)
6882 // These constants need to be in sync with runtime.traceback.go:printArgs.
6888 _offsetTooLarge = 0xfb
6889 _special = 0xf0 // above this are operators, below this are ordinary offsets
6893 limit = 10 // print no more than 10 args/components
6894 maxDepth = 5 // no more than 5 layers of nesting
6896 // maxLen is a (conservative) upper bound of the byte stream length. For
6897 // each arg/component, it has no more than 2 bytes of data (size, offset),
6898 // and no more than one {, }, ... at each level (it cannot have both the
6899 // data and ... unless it is the last one, just be conservative). Plus 1
6901 maxLen = (maxDepth*3+2)*limit + 1
6906 writebyte := func(o uint8) { wOff = objw.Uint8(x, wOff, o) }
6908 // Write one non-aggrgate arg/field/element.
6909 write1 := func(sz, offset int64) {
6910 if offset >= _special {
6911 writebyte(_offsetTooLarge)
6913 writebyte(uint8(offset))
6914 writebyte(uint8(sz))
6919 // Visit t recursively and write it out.
6920 // Returns whether to continue visiting.
6921 var visitType func(baseOffset int64, t *types.Type, depth int) bool
6922 visitType = func(baseOffset int64, t *types.Type, depth int) bool {
6924 writebyte(_dotdotdot)
6927 if !isAggregate(t) {
6928 write1(t.Size(), baseOffset)
6931 writebyte(_startAgg)
6933 if depth >= maxDepth {
6934 writebyte(_dotdotdot)
6940 case t.IsInterface(), t.IsString():
6941 _ = visitType(baseOffset, uintptrTyp, depth) &&
6942 visitType(baseOffset+PtrSize, uintptrTyp, depth)
6944 _ = visitType(baseOffset, uintptrTyp, depth) &&
6945 visitType(baseOffset+PtrSize, uintptrTyp, depth) &&
6946 visitType(baseOffset+PtrSize*2, uintptrTyp, depth)
6948 _ = visitType(baseOffset, types.FloatForComplex(t), depth) &&
6949 visitType(baseOffset+t.Size()/2, types.FloatForComplex(t), depth)
6951 if t.NumElem() == 0 {
6952 n++ // {} counts as a component
6955 for i := int64(0); i < t.NumElem(); i++ {
6956 if !visitType(baseOffset, t.Elem(), depth) {
6959 baseOffset += t.Elem().Size()
6962 if t.NumFields() == 0 {
6963 n++ // {} counts as a component
6966 for _, field := range t.Fields().Slice() {
6967 if !visitType(baseOffset+field.Offset, field.Type, depth) {
6977 if strings.Contains(f.LSym.Name, "[") {
6978 // Skip the dictionary argument - it is implicit and the user doesn't need to see it.
6982 for _, a := range abiInfo.InParams()[start:] {
6983 if !visitType(a.FrameOffset(abiInfo), a.Type, 0) {
6989 base.Fatalf("ArgInfo too large")
6995 // for wrapper, emit info of wrapped function.
6996 func emitWrappedFuncInfo(e *ssafn, pp *objw.Progs) {
6997 if base.Ctxt.Flag_linkshared {
6998 // Relative reference (SymPtrOff) to another shared object doesn't work.
7003 wfn := e.curfn.WrappedFunc
7008 wsym := wfn.Linksym()
7009 x := base.Ctxt.LookupInit(fmt.Sprintf("%s.wrapinfo", wsym.Name), func(x *obj.LSym) {
7010 objw.SymPtrOff(x, 0, wsym)
7011 x.Set(obj.AttrContentAddressable, true)
7013 e.curfn.LSym.Func().WrapInfo = x
7015 // Emit a funcdata pointing at the wrap info data.
7016 p := pp.Prog(obj.AFUNCDATA)
7017 p.From.SetConst(rtabi.FUNCDATA_WrapInfo)
7018 p.To.Type = obj.TYPE_MEM
7019 p.To.Name = obj.NAME_EXTERN
7023 // genssa appends entries to pp for each instruction in f.
7024 func genssa(f *ssa.Func, pp *objw.Progs) {
7026 s.ABI = f.OwnAux.Fn.ABI()
7028 e := f.Frontend().(*ssafn)
7030 s.livenessMap, s.partLiveArgs = liveness.Compute(e.curfn, f, e.stkptrsize, pp)
7031 emitArgInfo(e, f, pp)
7032 argLiveBlockMap, argLiveValueMap := liveness.ArgLiveness(e.curfn, f, pp)
7034 openDeferInfo := e.curfn.LSym.Func().OpenCodedDeferInfo
7035 if openDeferInfo != nil {
7036 // This function uses open-coded defers -- write out the funcdata
7037 // info that we computed at the end of genssa.
7038 p := pp.Prog(obj.AFUNCDATA)
7039 p.From.SetConst(rtabi.FUNCDATA_OpenCodedDeferInfo)
7040 p.To.Type = obj.TYPE_MEM
7041 p.To.Name = obj.NAME_EXTERN
7042 p.To.Sym = openDeferInfo
7045 emitWrappedFuncInfo(e, pp)
7047 // Remember where each block starts.
7048 s.bstart = make([]*obj.Prog, f.NumBlocks())
7050 var progToValue map[*obj.Prog]*ssa.Value
7051 var progToBlock map[*obj.Prog]*ssa.Block
7052 var valueToProgAfter []*obj.Prog // The first Prog following computation of a value v; v is visible at this point.
7053 gatherPrintInfo := f.PrintOrHtmlSSA || ssa.GenssaDump[f.Name]
7054 if gatherPrintInfo {
7055 progToValue = make(map[*obj.Prog]*ssa.Value, f.NumValues())
7056 progToBlock = make(map[*obj.Prog]*ssa.Block, f.NumBlocks())
7057 f.Logf("genssa %s\n", f.Name)
7058 progToBlock[s.pp.Next] = f.Blocks[0]
7061 if base.Ctxt.Flag_locationlists {
7062 if cap(f.Cache.ValueToProgAfter) < f.NumValues() {
7063 f.Cache.ValueToProgAfter = make([]*obj.Prog, f.NumValues())
7065 valueToProgAfter = f.Cache.ValueToProgAfter[:f.NumValues()]
7066 for i := range valueToProgAfter {
7067 valueToProgAfter[i] = nil
7071 // If the very first instruction is not tagged as a statement,
7072 // debuggers may attribute it to previous function in program.
7073 firstPos := src.NoXPos
7074 for _, v := range f.Entry.Values {
7075 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 {
7077 v.Pos = firstPos.WithDefaultStmt()
7082 // inlMarks has an entry for each Prog that implements an inline mark.
7083 // It maps from that Prog to the global inlining id of the inlined body
7084 // which should unwind to this Prog's location.
7085 var inlMarks map[*obj.Prog]int32
7086 var inlMarkList []*obj.Prog
7088 // inlMarksByPos maps from a (column 1) source position to the set of
7089 // Progs that are in the set above and have that source position.
7090 var inlMarksByPos map[src.XPos][]*obj.Prog
7092 var argLiveIdx int = -1 // argument liveness info index
7094 // Emit basic blocks
7095 for i, b := range f.Blocks {
7096 s.bstart[b.ID] = s.pp.Next
7097 s.lineRunStart = nil
7098 s.SetPos(s.pp.Pos.WithNotStmt()) // It needs a non-empty Pos, but cannot be a statement boundary (yet).
7100 if idx, ok := argLiveBlockMap[b.ID]; ok && idx != argLiveIdx {
7102 p := s.pp.Prog(obj.APCDATA)
7103 p.From.SetConst(rtabi.PCDATA_ArgLiveIndex)
7104 p.To.SetConst(int64(idx))
7107 // Emit values in block
7108 Arch.SSAMarkMoves(&s, b)
7109 for _, v := range b.Values {
7111 s.DebugFriendlySetPosFrom(v)
7113 if v.Op.ResultInArg0() && v.ResultReg() != v.Args[0].Reg() {
7114 v.Fatalf("input[0] and output not in same register %s", v.LongString())
7119 // memory arg needs no code
7121 // input args need no code
7122 case ssa.OpSP, ssa.OpSB:
7124 case ssa.OpSelect0, ssa.OpSelect1, ssa.OpSelectN, ssa.OpMakeResult:
7127 // nothing to do when there's a g register,
7128 // and checkLower complains if there's not
7129 case ssa.OpVarDef, ssa.OpVarLive, ssa.OpKeepAlive, ssa.OpWBend:
7130 // nothing to do; already used by liveness
7134 // nothing to do; no-op conversion for liveness
7135 if v.Args[0].Reg() != v.Reg() {
7136 v.Fatalf("OpConvert should be a no-op: %s; %s", v.Args[0].LongString(), v.LongString())
7139 p := Arch.Ginsnop(s.pp)
7140 if inlMarks == nil {
7141 inlMarks = map[*obj.Prog]int32{}
7142 inlMarksByPos = map[src.XPos][]*obj.Prog{}
7144 inlMarks[p] = v.AuxInt32()
7145 inlMarkList = append(inlMarkList, p)
7146 pos := v.Pos.AtColumn1()
7147 inlMarksByPos[pos] = append(inlMarksByPos[pos], p)
7148 firstPos = src.NoXPos
7151 // Special case for first line in function; move it to the start (which cannot be a register-valued instruction)
7152 if firstPos != src.NoXPos && v.Op != ssa.OpArgIntReg && v.Op != ssa.OpArgFloatReg && v.Op != ssa.OpLoadReg && v.Op != ssa.OpStoreReg {
7154 firstPos = src.NoXPos
7156 // Attach this safe point to the next
7158 s.pp.NextLive = s.livenessMap.Get(v)
7159 s.pp.NextUnsafe = s.livenessMap.GetUnsafe(v)
7161 // let the backend handle it
7162 Arch.SSAGenValue(&s, v)
7165 if idx, ok := argLiveValueMap[v.ID]; ok && idx != argLiveIdx {
7167 p := s.pp.Prog(obj.APCDATA)
7168 p.From.SetConst(rtabi.PCDATA_ArgLiveIndex)
7169 p.To.SetConst(int64(idx))
7172 if base.Ctxt.Flag_locationlists {
7173 valueToProgAfter[v.ID] = s.pp.Next
7176 if gatherPrintInfo {
7177 for ; x != s.pp.Next; x = x.Link {
7182 // If this is an empty infinite loop, stick a hardware NOP in there so that debuggers are less confused.
7183 if s.bstart[b.ID] == s.pp.Next && len(b.Succs) == 1 && b.Succs[0].Block() == b {
7184 p := Arch.Ginsnop(s.pp)
7185 p.Pos = p.Pos.WithIsStmt()
7186 if b.Pos == src.NoXPos {
7187 b.Pos = p.Pos // It needs a file, otherwise a no-file non-zero line causes confusion. See #35652.
7188 if b.Pos == src.NoXPos {
7189 b.Pos = pp.Text.Pos // Sometimes p.Pos is empty. See #35695.
7192 b.Pos = b.Pos.WithBogusLine() // Debuggers are not good about infinite loops, force a change in line number
7195 // Set unsafe mark for any end-of-block generated instructions
7196 // (normally, conditional or unconditional branches).
7197 // This is particularly important for empty blocks, as there
7198 // are no values to inherit the unsafe mark from.
7199 s.pp.NextUnsafe = s.livenessMap.GetUnsafeBlock(b)
7201 // Emit control flow instructions for block
7203 if i < len(f.Blocks)-1 && base.Flag.N == 0 {
7204 // If -N, leave next==nil so every block with successors
7205 // ends in a JMP (except call blocks - plive doesn't like
7206 // select{send,recv} followed by a JMP call). Helps keep
7207 // line numbers for otherwise empty blocks.
7208 next = f.Blocks[i+1]
7212 Arch.SSAGenBlock(&s, b, next)
7213 if gatherPrintInfo {
7214 for ; x != s.pp.Next; x = x.Link {
7219 if f.Blocks[len(f.Blocks)-1].Kind == ssa.BlockExit {
7220 // We need the return address of a panic call to
7221 // still be inside the function in question. So if
7222 // it ends in a call which doesn't return, add a
7223 // nop (which will never execute) after the call.
7226 if openDeferInfo != nil {
7227 // When doing open-coded defers, generate a disconnected call to
7228 // deferreturn and a return. This will be used to during panic
7229 // recovery to unwind the stack and return back to the runtime.
7230 s.pp.NextLive = s.livenessMap.DeferReturn
7231 p := pp.Prog(obj.ACALL)
7232 p.To.Type = obj.TYPE_MEM
7233 p.To.Name = obj.NAME_EXTERN
7234 p.To.Sym = ir.Syms.Deferreturn
7236 // Load results into registers. So when a deferred function
7237 // recovers a panic, it will return to caller with right results.
7238 // The results are already in memory, because they are not SSA'd
7239 // when the function has defers (see canSSAName).
7240 for _, o := range f.OwnAux.ABIInfo().OutParams() {
7241 n := o.Name.(*ir.Name)
7242 rts, offs := o.RegisterTypesAndOffsets()
7243 for i := range o.Registers {
7244 Arch.LoadRegResult(&s, f, rts[i], ssa.ObjRegForAbiReg(o.Registers[i], f.Config), n, offs[i])
7251 if inlMarks != nil {
7254 // We have some inline marks. Try to find other instructions we're
7255 // going to emit anyway, and use those instructions instead of the
7257 for p := pp.Text; p != nil; p = p.Link {
7258 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 {
7259 // Don't use 0-sized instructions as inline marks, because we need
7260 // to identify inline mark instructions by pc offset.
7261 // (Some of these instructions are sometimes zero-sized, sometimes not.
7262 // We must not use anything that even might be zero-sized.)
7263 // TODO: are there others?
7266 if _, ok := inlMarks[p]; ok {
7267 // Don't use inline marks themselves. We don't know
7268 // whether they will be zero-sized or not yet.
7271 if p.As == obj.ACALL || p.As == obj.ADUFFCOPY || p.As == obj.ADUFFZERO {
7274 pos := p.Pos.AtColumn1()
7275 s := inlMarksByPos[pos]
7279 for _, m := range s {
7280 // We found an instruction with the same source position as
7281 // some of the inline marks.
7282 // Use this instruction instead.
7283 p.Pos = p.Pos.WithIsStmt() // promote position to a statement
7284 pp.CurFunc.LSym.Func().AddInlMark(p, inlMarks[m])
7285 // Make the inline mark a real nop, so it doesn't generate any code.
7291 delete(inlMarksByPos, pos)
7293 // Any unmatched inline marks now need to be added to the inlining tree (and will generate a nop instruction).
7294 for _, p := range inlMarkList {
7295 if p.As != obj.ANOP {
7296 pp.CurFunc.LSym.Func().AddInlMark(p, inlMarks[p])
7300 if e.stksize == 0 && !hasCall {
7301 // Frameless leaf function. It doesn't need any preamble,
7302 // so make sure its first instruction isn't from an inlined callee.
7303 // If it is, add a nop at the start of the function with a position
7304 // equal to the start of the function.
7305 // This ensures that runtime.FuncForPC(uintptr(reflect.ValueOf(fn).Pointer())).Name()
7306 // returns the right answer. See issue 58300.
7307 for p := pp.Text; p != nil; p = p.Link {
7308 if p.As == obj.AFUNCDATA || p.As == obj.APCDATA || p.As == obj.ATEXT || p.As == obj.ANOP {
7311 if base.Ctxt.PosTable.Pos(p.Pos).Base().InliningIndex() >= 0 {
7312 // Make a real (not 0-sized) nop.
7313 nop := Arch.Ginsnop(pp)
7314 nop.Pos = e.curfn.Pos().WithIsStmt()
7316 // Unfortunately, Ginsnop puts the instruction at the
7317 // end of the list. Move it up to just before p.
7319 // Unlink from the current list.
7320 for x := pp.Text; x != nil; x = x.Link {
7326 // Splice in right before p.
7327 for x := pp.Text; x != nil; x = x.Link {
7340 if base.Ctxt.Flag_locationlists {
7341 var debugInfo *ssa.FuncDebug
7342 debugInfo = e.curfn.DebugInfo.(*ssa.FuncDebug)
7343 if e.curfn.ABI == obj.ABIInternal && base.Flag.N != 0 {
7344 ssa.BuildFuncDebugNoOptimized(base.Ctxt, f, base.Debug.LocationLists > 1, StackOffset, debugInfo)
7346 ssa.BuildFuncDebug(base.Ctxt, f, base.Debug.LocationLists, StackOffset, debugInfo)
7349 idToIdx := make([]int, f.NumBlocks())
7350 for i, b := range f.Blocks {
7353 // Note that at this moment, Prog.Pc is a sequence number; it's
7354 // not a real PC until after assembly, so this mapping has to
7356 debugInfo.GetPC = func(b, v ssa.ID) int64 {
7358 case ssa.BlockStart.ID:
7359 if b == f.Entry.ID {
7360 return 0 // Start at the very beginning, at the assembler-generated prologue.
7361 // this should only happen for function args (ssa.OpArg)
7364 case ssa.BlockEnd.ID:
7365 blk := f.Blocks[idToIdx[b]]
7366 nv := len(blk.Values)
7367 return valueToProgAfter[blk.Values[nv-1].ID].Pc
7368 case ssa.FuncEnd.ID:
7369 return e.curfn.LSym.Size
7371 return valueToProgAfter[v].Pc
7376 // Resolve branches, and relax DefaultStmt into NotStmt
7377 for _, br := range s.Branches {
7378 br.P.To.SetTarget(s.bstart[br.B.ID])
7379 if br.P.Pos.IsStmt() != src.PosIsStmt {
7380 br.P.Pos = br.P.Pos.WithNotStmt()
7381 } else if v0 := br.B.FirstPossibleStmtValue(); v0 != nil && v0.Pos.Line() == br.P.Pos.Line() && v0.Pos.IsStmt() == src.PosIsStmt {
7382 br.P.Pos = br.P.Pos.WithNotStmt()
7387 // Resolve jump table destinations.
7388 for _, jt := range s.JumpTables {
7389 // Convert from *Block targets to *Prog targets.
7390 targets := make([]*obj.Prog, len(jt.Succs))
7391 for i, e := range jt.Succs {
7392 targets[i] = s.bstart[e.Block().ID]
7394 // Add to list of jump tables to be resolved at assembly time.
7395 // The assembler converts from *Prog entries to absolute addresses
7396 // once it knows instruction byte offsets.
7397 fi := pp.CurFunc.LSym.Func()
7398 fi.JumpTables = append(fi.JumpTables, obj.JumpTable{Sym: jt.Aux.(*obj.LSym), Targets: targets})
7401 if e.log { // spew to stdout
7403 for p := pp.Text; p != nil; p = p.Link {
7404 if p.Pos.IsKnown() && p.InnermostFilename() != filename {
7405 filename = p.InnermostFilename()
7406 f.Logf("# %s\n", filename)
7410 if v, ok := progToValue[p]; ok {
7412 } else if b, ok := progToBlock[p]; ok {
7415 s = " " // most value and branch strings are 2-3 characters long
7417 f.Logf(" %-6s\t%.5d (%s)\t%s\n", s, p.Pc, p.InnermostLineNumber(), p.InstructionString())
7420 if f.HTMLWriter != nil { // spew to ssa.html
7421 var buf strings.Builder
7422 buf.WriteString("<code>")
7423 buf.WriteString("<dl class=\"ssa-gen\">")
7425 for p := pp.Text; p != nil; p = p.Link {
7426 // Don't spam every line with the file name, which is often huge.
7427 // Only print changes, and "unknown" is not a change.
7428 if p.Pos.IsKnown() && p.InnermostFilename() != filename {
7429 filename = p.InnermostFilename()
7430 buf.WriteString("<dt class=\"ssa-prog-src\"></dt><dd class=\"ssa-prog\">")
7431 buf.WriteString(html.EscapeString("# " + filename))
7432 buf.WriteString("</dd>")
7435 buf.WriteString("<dt class=\"ssa-prog-src\">")
7436 if v, ok := progToValue[p]; ok {
7437 buf.WriteString(v.HTML())
7438 } else if b, ok := progToBlock[p]; ok {
7439 buf.WriteString("<b>" + b.HTML() + "</b>")
7441 buf.WriteString("</dt>")
7442 buf.WriteString("<dd class=\"ssa-prog\">")
7443 fmt.Fprintf(&buf, "%.5d <span class=\"l%v line-number\">(%s)</span> %s", p.Pc, p.InnermostLineNumber(), p.InnermostLineNumberHTML(), html.EscapeString(p.InstructionString()))
7444 buf.WriteString("</dd>")
7446 buf.WriteString("</dl>")
7447 buf.WriteString("</code>")
7448 f.HTMLWriter.WriteColumn("genssa", "genssa", "ssa-prog", buf.String())
7450 if ssa.GenssaDump[f.Name] {
7451 fi := f.DumpFileForPhase("genssa")
7454 // inliningDiffers if any filename changes or if any line number except the innermost (last index) changes.
7455 inliningDiffers := func(a, b []src.Pos) bool {
7456 if len(a) != len(b) {
7460 if a[i].Filename() != b[i].Filename() {
7463 if i != len(a)-1 && a[i].Line() != b[i].Line() {
7470 var allPosOld []src.Pos
7471 var allPos []src.Pos
7473 for p := pp.Text; p != nil; p = p.Link {
7474 if p.Pos.IsKnown() {
7476 p.Ctxt.AllPos(p.Pos, func(pos src.Pos) { allPos = append(allPos, pos) })
7477 if inliningDiffers(allPos, allPosOld) {
7478 for _, pos := range allPos {
7479 fmt.Fprintf(fi, "# %s:%d\n", pos.Filename(), pos.Line())
7481 allPos, allPosOld = allPosOld, allPos // swap, not copy, so that they do not share slice storage.
7486 if v, ok := progToValue[p]; ok {
7488 } else if b, ok := progToBlock[p]; ok {
7491 s = " " // most value and branch strings are 2-3 characters long
7493 fmt.Fprintf(fi, " %-6s\t%.5d %s\t%s\n", s, p.Pc, ssa.StmtString(p.Pos), p.InstructionString())
7501 f.HTMLWriter.Close()
7505 func defframe(s *State, e *ssafn, f *ssa.Func) {
7508 s.maxarg = types.RoundUp(s.maxarg, e.stkalign)
7509 frame := s.maxarg + e.stksize
7510 if Arch.PadFrame != nil {
7511 frame = Arch.PadFrame(frame)
7514 // Fill in argument and frame size.
7515 pp.Text.To.Type = obj.TYPE_TEXTSIZE
7516 pp.Text.To.Val = int32(types.RoundUp(f.OwnAux.ArgWidth(), int64(types.RegSize)))
7517 pp.Text.To.Offset = frame
7521 // Insert code to spill argument registers if the named slot may be partially
7522 // live. That is, the named slot is considered live by liveness analysis,
7523 // (because a part of it is live), but we may not spill all parts into the
7524 // slot. This can only happen with aggregate-typed arguments that are SSA-able
7525 // and not address-taken (for non-SSA-able or address-taken arguments we always
7527 // Note: spilling is unnecessary in the -N/no-optimize case, since all values
7528 // will be considered non-SSAable and spilled up front.
7529 // TODO(register args) Make liveness more fine-grained to that partial spilling is okay.
7530 if f.OwnAux.ABIInfo().InRegistersUsed() != 0 && base.Flag.N == 0 {
7531 // First, see if it is already spilled before it may be live. Look for a spill
7532 // in the entry block up to the first safepoint.
7533 type nameOff struct {
7537 partLiveArgsSpilled := make(map[nameOff]bool)
7538 for _, v := range f.Entry.Values {
7542 if v.Op != ssa.OpStoreReg || v.Args[0].Op != ssa.OpArgIntReg {
7545 n, off := ssa.AutoVar(v)
7546 if n.Class != ir.PPARAM || n.Addrtaken() || !TypeOK(n.Type()) || !s.partLiveArgs[n] {
7549 partLiveArgsSpilled[nameOff{n, off}] = true
7552 // Then, insert code to spill registers if not already.
7553 for _, a := range f.OwnAux.ABIInfo().InParams() {
7554 n, ok := a.Name.(*ir.Name)
7555 if !ok || n.Addrtaken() || !TypeOK(n.Type()) || !s.partLiveArgs[n] || len(a.Registers) <= 1 {
7558 rts, offs := a.RegisterTypesAndOffsets()
7559 for i := range a.Registers {
7560 if !rts[i].HasPointers() {
7563 if partLiveArgsSpilled[nameOff{n, offs[i]}] {
7564 continue // already spilled
7566 reg := ssa.ObjRegForAbiReg(a.Registers[i], f.Config)
7567 p = Arch.SpillArgReg(pp, p, f, rts[i], reg, n, offs[i])
7572 // Insert code to zero ambiguously live variables so that the
7573 // garbage collector only sees initialized values when it
7574 // looks for pointers.
7577 // Opaque state for backend to use. Current backends use it to
7578 // keep track of which helper registers have been zeroed.
7581 // Iterate through declarations. Autos are sorted in decreasing
7582 // frame offset order.
7583 for _, n := range e.curfn.Dcl {
7587 if n.Class != ir.PAUTO {
7588 e.Fatalf(n.Pos(), "needzero class %d", n.Class)
7590 if n.Type().Size()%int64(types.PtrSize) != 0 || n.FrameOffset()%int64(types.PtrSize) != 0 || n.Type().Size() == 0 {
7591 e.Fatalf(n.Pos(), "var %L has size %d offset %d", n, n.Type().Size(), n.Offset_)
7594 if lo != hi && n.FrameOffset()+n.Type().Size() >= lo-int64(2*types.RegSize) {
7595 // Merge with range we already have.
7596 lo = n.FrameOffset()
7601 p = Arch.ZeroRange(pp, p, frame+lo, hi-lo, &state)
7604 lo = n.FrameOffset()
7605 hi = lo + n.Type().Size()
7608 // Zero final range.
7609 Arch.ZeroRange(pp, p, frame+lo, hi-lo, &state)
7612 // For generating consecutive jump instructions to model a specific branching
7613 type IndexJump struct {
7618 func (s *State) oneJump(b *ssa.Block, jump *IndexJump) {
7619 p := s.Br(jump.Jump, b.Succs[jump.Index].Block())
7623 // CombJump generates combinational instructions (2 at present) for a block jump,
7624 // thereby the behaviour of non-standard condition codes could be simulated
7625 func (s *State) CombJump(b, next *ssa.Block, jumps *[2][2]IndexJump) {
7627 case b.Succs[0].Block():
7628 s.oneJump(b, &jumps[0][0])
7629 s.oneJump(b, &jumps[0][1])
7630 case b.Succs[1].Block():
7631 s.oneJump(b, &jumps[1][0])
7632 s.oneJump(b, &jumps[1][1])
7635 if b.Likely != ssa.BranchUnlikely {
7636 s.oneJump(b, &jumps[1][0])
7637 s.oneJump(b, &jumps[1][1])
7638 q = s.Br(obj.AJMP, b.Succs[1].Block())
7640 s.oneJump(b, &jumps[0][0])
7641 s.oneJump(b, &jumps[0][1])
7642 q = s.Br(obj.AJMP, b.Succs[0].Block())
7648 // AddAux adds the offset in the aux fields (AuxInt and Aux) of v to a.
7649 func AddAux(a *obj.Addr, v *ssa.Value) {
7650 AddAux2(a, v, v.AuxInt)
7652 func AddAux2(a *obj.Addr, v *ssa.Value, offset int64) {
7653 if a.Type != obj.TYPE_MEM && a.Type != obj.TYPE_ADDR {
7654 v.Fatalf("bad AddAux addr %v", a)
7656 // add integer offset
7659 // If no additional symbol offset, we're done.
7663 // Add symbol's offset from its base register.
7664 switch n := v.Aux.(type) {
7666 a.Name = obj.NAME_EXTERN
7669 a.Name = obj.NAME_EXTERN
7672 if n.Class == ir.PPARAM || (n.Class == ir.PPARAMOUT && !n.IsOutputParamInRegisters()) {
7673 a.Name = obj.NAME_PARAM
7674 a.Sym = ir.Orig(n).(*ir.Name).Linksym()
7675 a.Offset += n.FrameOffset()
7678 a.Name = obj.NAME_AUTO
7679 if n.Class == ir.PPARAMOUT {
7680 a.Sym = ir.Orig(n).(*ir.Name).Linksym()
7684 a.Offset += n.FrameOffset()
7686 v.Fatalf("aux in %s not implemented %#v", v, v.Aux)
7690 // extendIndex extends v to a full int width.
7691 // panic with the given kind if v does not fit in an int (only on 32-bit archs).
7692 func (s *state) extendIndex(idx, len *ssa.Value, kind ssa.BoundsKind, bounded bool) *ssa.Value {
7693 size := idx.Type.Size()
7694 if size == s.config.PtrSize {
7697 if size > s.config.PtrSize {
7698 // truncate 64-bit indexes on 32-bit pointer archs. Test the
7699 // high word and branch to out-of-bounds failure if it is not 0.
7701 if idx.Type.IsSigned() {
7702 lo = s.newValue1(ssa.OpInt64Lo, types.Types[types.TINT], idx)
7704 lo = s.newValue1(ssa.OpInt64Lo, types.Types[types.TUINT], idx)
7706 if bounded || base.Flag.B != 0 {
7709 bNext := s.f.NewBlock(ssa.BlockPlain)
7710 bPanic := s.f.NewBlock(ssa.BlockExit)
7711 hi := s.newValue1(ssa.OpInt64Hi, types.Types[types.TUINT32], idx)
7712 cmp := s.newValue2(ssa.OpEq32, types.Types[types.TBOOL], hi, s.constInt32(types.Types[types.TUINT32], 0))
7713 if !idx.Type.IsSigned() {
7715 case ssa.BoundsIndex:
7716 kind = ssa.BoundsIndexU
7717 case ssa.BoundsSliceAlen:
7718 kind = ssa.BoundsSliceAlenU
7719 case ssa.BoundsSliceAcap:
7720 kind = ssa.BoundsSliceAcapU
7721 case ssa.BoundsSliceB:
7722 kind = ssa.BoundsSliceBU
7723 case ssa.BoundsSlice3Alen:
7724 kind = ssa.BoundsSlice3AlenU
7725 case ssa.BoundsSlice3Acap:
7726 kind = ssa.BoundsSlice3AcapU
7727 case ssa.BoundsSlice3B:
7728 kind = ssa.BoundsSlice3BU
7729 case ssa.BoundsSlice3C:
7730 kind = ssa.BoundsSlice3CU
7734 b.Kind = ssa.BlockIf
7736 b.Likely = ssa.BranchLikely
7740 s.startBlock(bPanic)
7741 mem := s.newValue4I(ssa.OpPanicExtend, types.TypeMem, int64(kind), hi, lo, len, s.mem())
7742 s.endBlock().SetControl(mem)
7748 // Extend value to the required size
7750 if idx.Type.IsSigned() {
7751 switch 10*size + s.config.PtrSize {
7753 op = ssa.OpSignExt8to32
7755 op = ssa.OpSignExt8to64
7757 op = ssa.OpSignExt16to32
7759 op = ssa.OpSignExt16to64
7761 op = ssa.OpSignExt32to64
7763 s.Fatalf("bad signed index extension %s", idx.Type)
7766 switch 10*size + s.config.PtrSize {
7768 op = ssa.OpZeroExt8to32
7770 op = ssa.OpZeroExt8to64
7772 op = ssa.OpZeroExt16to32
7774 op = ssa.OpZeroExt16to64
7776 op = ssa.OpZeroExt32to64
7778 s.Fatalf("bad unsigned index extension %s", idx.Type)
7781 return s.newValue1(op, types.Types[types.TINT], idx)
7784 // CheckLoweredPhi checks that regalloc and stackalloc correctly handled phi values.
7785 // Called during ssaGenValue.
7786 func CheckLoweredPhi(v *ssa.Value) {
7787 if v.Op != ssa.OpPhi {
7788 v.Fatalf("CheckLoweredPhi called with non-phi value: %v", v.LongString())
7790 if v.Type.IsMemory() {
7794 loc := f.RegAlloc[v.ID]
7795 for _, a := range v.Args {
7796 if aloc := f.RegAlloc[a.ID]; aloc != loc { // TODO: .Equal() instead?
7797 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)
7802 // CheckLoweredGetClosurePtr checks that v is the first instruction in the function's entry block,
7803 // except for incoming in-register arguments.
7804 // The output of LoweredGetClosurePtr is generally hardwired to the correct register.
7805 // That register contains the closure pointer on closure entry.
7806 func CheckLoweredGetClosurePtr(v *ssa.Value) {
7807 entry := v.Block.Func.Entry
7808 if entry != v.Block {
7809 base.Fatalf("in %s, badly placed LoweredGetClosurePtr: %v %v", v.Block.Func.Name, v.Block, v)
7811 for _, w := range entry.Values {
7816 case ssa.OpArgIntReg, ssa.OpArgFloatReg:
7819 base.Fatalf("in %s, badly placed LoweredGetClosurePtr: %v %v", v.Block.Func.Name, v.Block, v)
7824 // CheckArgReg ensures that v is in the function's entry block.
7825 func CheckArgReg(v *ssa.Value) {
7826 entry := v.Block.Func.Entry
7827 if entry != v.Block {
7828 base.Fatalf("in %s, badly placed ArgIReg or ArgFReg: %v %v", v.Block.Func.Name, v.Block, v)
7832 func AddrAuto(a *obj.Addr, v *ssa.Value) {
7833 n, off := ssa.AutoVar(v)
7834 a.Type = obj.TYPE_MEM
7836 a.Reg = int16(Arch.REGSP)
7837 a.Offset = n.FrameOffset() + off
7838 if n.Class == ir.PPARAM || (n.Class == ir.PPARAMOUT && !n.IsOutputParamInRegisters()) {
7839 a.Name = obj.NAME_PARAM
7841 a.Name = obj.NAME_AUTO
7845 // Call returns a new CALL instruction for the SSA value v.
7846 // It uses PrepareCall to prepare the call.
7847 func (s *State) Call(v *ssa.Value) *obj.Prog {
7848 pPosIsStmt := s.pp.Pos.IsStmt() // The statement-ness fo the call comes from ssaGenState
7851 p := s.Prog(obj.ACALL)
7852 if pPosIsStmt == src.PosIsStmt {
7853 p.Pos = v.Pos.WithIsStmt()
7855 p.Pos = v.Pos.WithNotStmt()
7857 if sym, ok := v.Aux.(*ssa.AuxCall); ok && sym.Fn != nil {
7858 p.To.Type = obj.TYPE_MEM
7859 p.To.Name = obj.NAME_EXTERN
7862 // TODO(mdempsky): Can these differences be eliminated?
7863 switch Arch.LinkArch.Family {
7864 case sys.AMD64, sys.I386, sys.PPC64, sys.RISCV64, sys.S390X, sys.Wasm:
7865 p.To.Type = obj.TYPE_REG
7866 case sys.ARM, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64:
7867 p.To.Type = obj.TYPE_MEM
7869 base.Fatalf("unknown indirect call family")
7871 p.To.Reg = v.Args[0].Reg()
7876 // TailCall returns a new tail call instruction for the SSA value v.
7877 // It is like Call, but for a tail call.
7878 func (s *State) TailCall(v *ssa.Value) *obj.Prog {
7884 // PrepareCall prepares to emit a CALL instruction for v and does call-related bookkeeping.
7885 // It must be called immediately before emitting the actual CALL instruction,
7886 // since it emits PCDATA for the stack map at the call (calls are safe points).
7887 func (s *State) PrepareCall(v *ssa.Value) {
7888 idx := s.livenessMap.Get(v)
7889 if !idx.StackMapValid() {
7890 // See Liveness.hasStackMap.
7891 if sym, ok := v.Aux.(*ssa.AuxCall); !ok || !(sym.Fn == ir.Syms.WBZero || sym.Fn == ir.Syms.WBMove) {
7892 base.Fatalf("missing stack map index for %v", v.LongString())
7896 call, ok := v.Aux.(*ssa.AuxCall)
7899 // Record call graph information for nowritebarrierrec
7901 if nowritebarrierrecCheck != nil {
7902 nowritebarrierrecCheck.recordCall(s.pp.CurFunc, call.Fn, v.Pos)
7906 if s.maxarg < v.AuxInt {
7911 // UseArgs records the fact that an instruction needs a certain amount of
7912 // callee args space for its use.
7913 func (s *State) UseArgs(n int64) {
7919 // fieldIdx finds the index of the field referred to by the ODOT node n.
7920 func fieldIdx(n *ir.SelectorExpr) int {
7923 panic("ODOT's LHS is not a struct")
7926 for i, f := range t.Fields().Slice() {
7928 if f.Offset != n.Offset() {
7929 panic("field offset doesn't match")
7934 panic(fmt.Sprintf("can't find field in expr %v\n", n))
7936 // TODO: keep the result of this function somewhere in the ODOT Node
7937 // so we don't have to recompute it each time we need it.
7940 // ssafn holds frontend information about a function that the backend is processing.
7941 // It also exports a bunch of compiler services for the ssa backend.
7944 strings map[string]*obj.LSym // map from constant string to data symbols
7945 stksize int64 // stack size for current frame
7946 stkptrsize int64 // prefix of stack containing pointers
7948 // alignment for current frame.
7949 // NOTE: when stkalign > PtrSize, currently this only ensures the offsets of
7950 // objects in the stack frame are aligned. The stack pointer is still aligned
7954 log bool // print ssa debug to the stdout
7957 // StringData returns a symbol which
7958 // is the data component of a global string constant containing s.
7959 func (e *ssafn) StringData(s string) *obj.LSym {
7960 if aux, ok := e.strings[s]; ok {
7963 if e.strings == nil {
7964 e.strings = make(map[string]*obj.LSym)
7966 data := staticdata.StringSym(e.curfn.Pos(), s)
7971 func (e *ssafn) Auto(pos src.XPos, t *types.Type) *ir.Name {
7972 return typecheck.TempAt(pos, e.curfn, t) // Note: adds new auto to e.curfn.Func.Dcl list
7975 // SplitSlot returns a slot representing the data of parent starting at offset.
7976 func (e *ssafn) SplitSlot(parent *ssa.LocalSlot, suffix string, offset int64, t *types.Type) ssa.LocalSlot {
7979 if node.Class != ir.PAUTO || node.Addrtaken() {
7980 // addressed things and non-autos retain their parents (i.e., cannot truly be split)
7981 return ssa.LocalSlot{N: node, Type: t, Off: parent.Off + offset}
7984 sym := &types.Sym{Name: node.Sym().Name + suffix, Pkg: types.LocalPkg}
7985 n := e.curfn.NewLocal(parent.N.Pos(), sym, ir.PAUTO, t)
7987 n.SetEsc(ir.EscNever)
7989 return ssa.LocalSlot{N: n, Type: t, Off: 0, SplitOf: parent, SplitOffset: offset}
7992 func (e *ssafn) CanSSA(t *types.Type) bool {
7996 // Logf logs a message from the compiler.
7997 func (e *ssafn) Logf(msg string, args ...interface{}) {
7999 fmt.Printf(msg, args...)
8003 func (e *ssafn) Log() bool {
8007 // Fatalf reports a compiler error and exits.
8008 func (e *ssafn) Fatalf(pos src.XPos, msg string, args ...interface{}) {
8010 nargs := append([]interface{}{ir.FuncName(e.curfn)}, args...)
8011 base.Fatalf("'%s': "+msg, nargs...)
8014 // Warnl reports a "warning", which is usually flag-triggered
8015 // logging output for the benefit of tests.
8016 func (e *ssafn) Warnl(pos src.XPos, fmt_ string, args ...interface{}) {
8017 base.WarnfAt(pos, fmt_, args...)
8020 func (e *ssafn) Debug_checknil() bool {
8021 return base.Debug.Nil != 0
8024 func (e *ssafn) UseWriteBarrier() bool {
8028 func (e *ssafn) Syslook(name string) *obj.LSym {
8030 case "goschedguarded":
8031 return ir.Syms.Goschedguarded
8032 case "writeBarrier":
8033 return ir.Syms.WriteBarrier
8035 return ir.Syms.WBZero
8037 return ir.Syms.WBMove
8038 case "cgoCheckMemmove":
8039 return ir.Syms.CgoCheckMemmove
8040 case "cgoCheckPtrWrite":
8041 return ir.Syms.CgoCheckPtrWrite
8043 e.Fatalf(src.NoXPos, "unknown Syslook func %v", name)
8047 func (e *ssafn) MyImportPath() string {
8048 return base.Ctxt.Pkgpath
8051 func (e *ssafn) Func() *ir.Func {
8055 func clobberBase(n ir.Node) ir.Node {
8056 if n.Op() == ir.ODOT {
8057 n := n.(*ir.SelectorExpr)
8058 if n.X.Type().NumFields() == 1 {
8059 return clobberBase(n.X)
8062 if n.Op() == ir.OINDEX {
8063 n := n.(*ir.IndexExpr)
8064 if n.X.Type().IsArray() && n.X.Type().NumElem() == 1 {
8065 return clobberBase(n.X)
8071 // callTargetLSym returns the correct LSym to call 'callee' using its ABI.
8072 func callTargetLSym(callee *ir.Name) *obj.LSym {
8073 if callee.Func == nil {
8074 // TODO(austin): This happens in case of interface method I.M from imported package.
8075 // It's ABIInternal, and would be better if callee.Func was never nil and we didn't
8077 return callee.Linksym()
8080 return callee.LinksymABI(callee.Func.ABI)
8083 func min8(a, b int8) int8 {
8090 func max8(a, b int8) int8 {
8097 var deferStructFnField = -1
8099 // deferstruct makes a runtime._defer structure.
8100 func deferstruct() *types.Type {
8101 makefield := func(name string, typ *types.Type) *types.Field {
8102 // Unlike the global makefield function, this one needs to set Pkg
8103 // because these types might be compared (in SSA CSE sorting).
8104 // TODO: unify this makefield and the global one above.
8105 sym := &types.Sym{Name: name, Pkg: types.LocalPkg}
8106 return types.NewField(src.NoXPos, sym, typ)
8108 // These fields must match the ones in runtime/runtime2.go:_defer and
8109 // (*state).call above.
8110 fields := []*types.Field{
8111 makefield("heap", types.Types[types.TBOOL]),
8112 makefield("rangefunc", types.Types[types.TBOOL]),
8113 makefield("sp", types.Types[types.TUINTPTR]),
8114 makefield("pc", types.Types[types.TUINTPTR]),
8115 // Note: the types here don't really matter. Defer structures
8116 // are always scanned explicitly during stack copying and GC,
8117 // so we make them uintptr type even though they are real pointers.
8118 makefield("fn", types.Types[types.TUINTPTR]),
8119 makefield("link", types.Types[types.TUINTPTR]),
8120 makefield("head", types.Types[types.TUINTPTR]),
8122 for i, f := range fields {
8123 if f.Sym.Name == "fn" {
8124 deferStructFnField = i
8128 if deferStructFnField < 0 {
8129 base.Fatalf("deferstruct has no fn field")
8132 // build struct holding the above fields
8133 s := types.NewStruct(fields)
8135 types.CalcStructSize(s)
8139 // SpillSlotAddr uses LocalSlot information to initialize an obj.Addr
8140 // The resulting addr is used in a non-standard context -- in the prologue
8141 // of a function, before the frame has been constructed, so the standard
8142 // addressing for the parameters will be wrong.
8143 func SpillSlotAddr(spill ssa.Spill, baseReg int16, extraOffset int64) obj.Addr {
8145 Name: obj.NAME_NONE,
8148 Offset: spill.Offset + extraOffset,
8153 BoundsCheckFunc [ssa.BoundsKindCount]*obj.LSym
8154 ExtendCheckFunc [ssa.BoundsKindCount]*obj.LSym