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.NewPtr(reflectdata.MapType()) // *runtime.hmap
90 _ = types.NewPtr(deferstruct()) // *runtime._defer
91 types.NewPtrCacheEnabled = false
92 ssaConfig = ssa.NewConfig(base.Ctxt.Arch.Name, *types_, base.Ctxt, base.Flag.N == 0, Arch.SoftFloat)
93 ssaConfig.Race = base.Flag.Race
94 ssaCaches = make([]ssa.Cache, base.Flag.LowerC)
96 // Set up some runtime functions we'll need to call.
97 ir.Syms.AssertE2I = typecheck.LookupRuntimeFunc("assertE2I")
98 ir.Syms.AssertE2I2 = typecheck.LookupRuntimeFunc("assertE2I2")
99 ir.Syms.AssertI2I = typecheck.LookupRuntimeFunc("assertI2I")
100 ir.Syms.AssertI2I2 = typecheck.LookupRuntimeFunc("assertI2I2")
101 ir.Syms.CgoCheckMemmove = typecheck.LookupRuntimeFunc("cgoCheckMemmove")
102 ir.Syms.CgoCheckPtrWrite = typecheck.LookupRuntimeFunc("cgoCheckPtrWrite")
103 ir.Syms.CheckPtrAlignment = typecheck.LookupRuntimeFunc("checkptrAlignment")
104 ir.Syms.Deferproc = typecheck.LookupRuntimeFunc("deferproc")
105 ir.Syms.DeferprocStack = typecheck.LookupRuntimeFunc("deferprocStack")
106 ir.Syms.Deferreturn = typecheck.LookupRuntimeFunc("deferreturn")
107 ir.Syms.Duffcopy = typecheck.LookupRuntimeFunc("duffcopy")
108 ir.Syms.Duffzero = typecheck.LookupRuntimeFunc("duffzero")
109 ir.Syms.GCWriteBarrier[0] = typecheck.LookupRuntimeFunc("gcWriteBarrier1")
110 ir.Syms.GCWriteBarrier[1] = typecheck.LookupRuntimeFunc("gcWriteBarrier2")
111 ir.Syms.GCWriteBarrier[2] = typecheck.LookupRuntimeFunc("gcWriteBarrier3")
112 ir.Syms.GCWriteBarrier[3] = typecheck.LookupRuntimeFunc("gcWriteBarrier4")
113 ir.Syms.GCWriteBarrier[4] = typecheck.LookupRuntimeFunc("gcWriteBarrier5")
114 ir.Syms.GCWriteBarrier[5] = typecheck.LookupRuntimeFunc("gcWriteBarrier6")
115 ir.Syms.GCWriteBarrier[6] = typecheck.LookupRuntimeFunc("gcWriteBarrier7")
116 ir.Syms.GCWriteBarrier[7] = typecheck.LookupRuntimeFunc("gcWriteBarrier8")
117 ir.Syms.Goschedguarded = typecheck.LookupRuntimeFunc("goschedguarded")
118 ir.Syms.Growslice = typecheck.LookupRuntimeFunc("growslice")
119 ir.Syms.Memmove = typecheck.LookupRuntimeFunc("memmove")
120 ir.Syms.Msanread = typecheck.LookupRuntimeFunc("msanread")
121 ir.Syms.Msanwrite = typecheck.LookupRuntimeFunc("msanwrite")
122 ir.Syms.Msanmove = typecheck.LookupRuntimeFunc("msanmove")
123 ir.Syms.Asanread = typecheck.LookupRuntimeFunc("asanread")
124 ir.Syms.Asanwrite = typecheck.LookupRuntimeFunc("asanwrite")
125 ir.Syms.Newobject = typecheck.LookupRuntimeFunc("newobject")
126 ir.Syms.Newproc = typecheck.LookupRuntimeFunc("newproc")
127 ir.Syms.Panicdivide = typecheck.LookupRuntimeFunc("panicdivide")
128 ir.Syms.PanicdottypeE = typecheck.LookupRuntimeFunc("panicdottypeE")
129 ir.Syms.PanicdottypeI = typecheck.LookupRuntimeFunc("panicdottypeI")
130 ir.Syms.Panicnildottype = typecheck.LookupRuntimeFunc("panicnildottype")
131 ir.Syms.Panicoverflow = typecheck.LookupRuntimeFunc("panicoverflow")
132 ir.Syms.Panicshift = typecheck.LookupRuntimeFunc("panicshift")
133 ir.Syms.Raceread = typecheck.LookupRuntimeFunc("raceread")
134 ir.Syms.Racereadrange = typecheck.LookupRuntimeFunc("racereadrange")
135 ir.Syms.Racewrite = typecheck.LookupRuntimeFunc("racewrite")
136 ir.Syms.Racewriterange = typecheck.LookupRuntimeFunc("racewriterange")
137 ir.Syms.WBZero = typecheck.LookupRuntimeFunc("wbZero")
138 ir.Syms.WBMove = typecheck.LookupRuntimeFunc("wbMove")
139 ir.Syms.X86HasPOPCNT = typecheck.LookupRuntimeVar("x86HasPOPCNT") // bool
140 ir.Syms.X86HasSSE41 = typecheck.LookupRuntimeVar("x86HasSSE41") // bool
141 ir.Syms.X86HasFMA = typecheck.LookupRuntimeVar("x86HasFMA") // bool
142 ir.Syms.ARMHasVFPv4 = typecheck.LookupRuntimeVar("armHasVFPv4") // bool
143 ir.Syms.ARM64HasATOMICS = typecheck.LookupRuntimeVar("arm64HasATOMICS") // bool
144 ir.Syms.Staticuint64s = typecheck.LookupRuntimeVar("staticuint64s")
145 ir.Syms.Typedmemmove = typecheck.LookupRuntimeFunc("typedmemmove")
146 ir.Syms.Udiv = typecheck.LookupRuntimeVar("udiv") // asm func with special ABI
147 ir.Syms.WriteBarrier = typecheck.LookupRuntimeVar("writeBarrier") // struct { bool; ... }
148 ir.Syms.Zerobase = typecheck.LookupRuntimeVar("zerobase")
150 if Arch.LinkArch.Family == sys.Wasm {
151 BoundsCheckFunc[ssa.BoundsIndex] = typecheck.LookupRuntimeFunc("goPanicIndex")
152 BoundsCheckFunc[ssa.BoundsIndexU] = typecheck.LookupRuntimeFunc("goPanicIndexU")
153 BoundsCheckFunc[ssa.BoundsSliceAlen] = typecheck.LookupRuntimeFunc("goPanicSliceAlen")
154 BoundsCheckFunc[ssa.BoundsSliceAlenU] = typecheck.LookupRuntimeFunc("goPanicSliceAlenU")
155 BoundsCheckFunc[ssa.BoundsSliceAcap] = typecheck.LookupRuntimeFunc("goPanicSliceAcap")
156 BoundsCheckFunc[ssa.BoundsSliceAcapU] = typecheck.LookupRuntimeFunc("goPanicSliceAcapU")
157 BoundsCheckFunc[ssa.BoundsSliceB] = typecheck.LookupRuntimeFunc("goPanicSliceB")
158 BoundsCheckFunc[ssa.BoundsSliceBU] = typecheck.LookupRuntimeFunc("goPanicSliceBU")
159 BoundsCheckFunc[ssa.BoundsSlice3Alen] = typecheck.LookupRuntimeFunc("goPanicSlice3Alen")
160 BoundsCheckFunc[ssa.BoundsSlice3AlenU] = typecheck.LookupRuntimeFunc("goPanicSlice3AlenU")
161 BoundsCheckFunc[ssa.BoundsSlice3Acap] = typecheck.LookupRuntimeFunc("goPanicSlice3Acap")
162 BoundsCheckFunc[ssa.BoundsSlice3AcapU] = typecheck.LookupRuntimeFunc("goPanicSlice3AcapU")
163 BoundsCheckFunc[ssa.BoundsSlice3B] = typecheck.LookupRuntimeFunc("goPanicSlice3B")
164 BoundsCheckFunc[ssa.BoundsSlice3BU] = typecheck.LookupRuntimeFunc("goPanicSlice3BU")
165 BoundsCheckFunc[ssa.BoundsSlice3C] = typecheck.LookupRuntimeFunc("goPanicSlice3C")
166 BoundsCheckFunc[ssa.BoundsSlice3CU] = typecheck.LookupRuntimeFunc("goPanicSlice3CU")
167 BoundsCheckFunc[ssa.BoundsConvert] = typecheck.LookupRuntimeFunc("goPanicSliceConvert")
169 BoundsCheckFunc[ssa.BoundsIndex] = typecheck.LookupRuntimeFunc("panicIndex")
170 BoundsCheckFunc[ssa.BoundsIndexU] = typecheck.LookupRuntimeFunc("panicIndexU")
171 BoundsCheckFunc[ssa.BoundsSliceAlen] = typecheck.LookupRuntimeFunc("panicSliceAlen")
172 BoundsCheckFunc[ssa.BoundsSliceAlenU] = typecheck.LookupRuntimeFunc("panicSliceAlenU")
173 BoundsCheckFunc[ssa.BoundsSliceAcap] = typecheck.LookupRuntimeFunc("panicSliceAcap")
174 BoundsCheckFunc[ssa.BoundsSliceAcapU] = typecheck.LookupRuntimeFunc("panicSliceAcapU")
175 BoundsCheckFunc[ssa.BoundsSliceB] = typecheck.LookupRuntimeFunc("panicSliceB")
176 BoundsCheckFunc[ssa.BoundsSliceBU] = typecheck.LookupRuntimeFunc("panicSliceBU")
177 BoundsCheckFunc[ssa.BoundsSlice3Alen] = typecheck.LookupRuntimeFunc("panicSlice3Alen")
178 BoundsCheckFunc[ssa.BoundsSlice3AlenU] = typecheck.LookupRuntimeFunc("panicSlice3AlenU")
179 BoundsCheckFunc[ssa.BoundsSlice3Acap] = typecheck.LookupRuntimeFunc("panicSlice3Acap")
180 BoundsCheckFunc[ssa.BoundsSlice3AcapU] = typecheck.LookupRuntimeFunc("panicSlice3AcapU")
181 BoundsCheckFunc[ssa.BoundsSlice3B] = typecheck.LookupRuntimeFunc("panicSlice3B")
182 BoundsCheckFunc[ssa.BoundsSlice3BU] = typecheck.LookupRuntimeFunc("panicSlice3BU")
183 BoundsCheckFunc[ssa.BoundsSlice3C] = typecheck.LookupRuntimeFunc("panicSlice3C")
184 BoundsCheckFunc[ssa.BoundsSlice3CU] = typecheck.LookupRuntimeFunc("panicSlice3CU")
185 BoundsCheckFunc[ssa.BoundsConvert] = typecheck.LookupRuntimeFunc("panicSliceConvert")
187 if Arch.LinkArch.PtrSize == 4 {
188 ExtendCheckFunc[ssa.BoundsIndex] = typecheck.LookupRuntimeVar("panicExtendIndex")
189 ExtendCheckFunc[ssa.BoundsIndexU] = typecheck.LookupRuntimeVar("panicExtendIndexU")
190 ExtendCheckFunc[ssa.BoundsSliceAlen] = typecheck.LookupRuntimeVar("panicExtendSliceAlen")
191 ExtendCheckFunc[ssa.BoundsSliceAlenU] = typecheck.LookupRuntimeVar("panicExtendSliceAlenU")
192 ExtendCheckFunc[ssa.BoundsSliceAcap] = typecheck.LookupRuntimeVar("panicExtendSliceAcap")
193 ExtendCheckFunc[ssa.BoundsSliceAcapU] = typecheck.LookupRuntimeVar("panicExtendSliceAcapU")
194 ExtendCheckFunc[ssa.BoundsSliceB] = typecheck.LookupRuntimeVar("panicExtendSliceB")
195 ExtendCheckFunc[ssa.BoundsSliceBU] = typecheck.LookupRuntimeVar("panicExtendSliceBU")
196 ExtendCheckFunc[ssa.BoundsSlice3Alen] = typecheck.LookupRuntimeVar("panicExtendSlice3Alen")
197 ExtendCheckFunc[ssa.BoundsSlice3AlenU] = typecheck.LookupRuntimeVar("panicExtendSlice3AlenU")
198 ExtendCheckFunc[ssa.BoundsSlice3Acap] = typecheck.LookupRuntimeVar("panicExtendSlice3Acap")
199 ExtendCheckFunc[ssa.BoundsSlice3AcapU] = typecheck.LookupRuntimeVar("panicExtendSlice3AcapU")
200 ExtendCheckFunc[ssa.BoundsSlice3B] = typecheck.LookupRuntimeVar("panicExtendSlice3B")
201 ExtendCheckFunc[ssa.BoundsSlice3BU] = typecheck.LookupRuntimeVar("panicExtendSlice3BU")
202 ExtendCheckFunc[ssa.BoundsSlice3C] = typecheck.LookupRuntimeVar("panicExtendSlice3C")
203 ExtendCheckFunc[ssa.BoundsSlice3CU] = typecheck.LookupRuntimeVar("panicExtendSlice3CU")
206 // Wasm (all asm funcs with special ABIs)
207 ir.Syms.WasmDiv = typecheck.LookupRuntimeVar("wasmDiv")
208 ir.Syms.WasmTruncS = typecheck.LookupRuntimeVar("wasmTruncS")
209 ir.Syms.WasmTruncU = typecheck.LookupRuntimeVar("wasmTruncU")
210 ir.Syms.SigPanic = typecheck.LookupRuntimeFunc("sigpanic")
213 // AbiForBodylessFuncStackMap returns the ABI for a bodyless function's stack map.
214 // This is not necessarily the ABI used to call it.
215 // Currently (1.17 dev) such a stack map is always ABI0;
216 // any ABI wrapper that is present is nosplit, hence a precise
217 // stack map is not needed there (the parameters survive only long
218 // enough to call the wrapped assembly function).
219 // This always returns a freshly copied ABI.
220 func AbiForBodylessFuncStackMap(fn *ir.Func) *abi.ABIConfig {
221 return ssaConfig.ABI0.Copy() // No idea what races will result, be safe
224 // abiForFunc implements ABI policy for a function, but does not return a copy of the ABI.
225 // Passing a nil function returns the default ABI based on experiment configuration.
226 func abiForFunc(fn *ir.Func, abi0, abi1 *abi.ABIConfig) *abi.ABIConfig {
227 if buildcfg.Experiment.RegabiArgs {
228 // Select the ABI based on the function's defining ABI.
235 case obj.ABIInternal:
236 // TODO(austin): Clean up the nomenclature here.
237 // It's not clear that "abi1" is ABIInternal.
240 base.Fatalf("function %v has unknown ABI %v", fn, fn.ABI)
241 panic("not reachable")
246 if fn.Pragma&ir.RegisterParams != 0 { // TODO(register args) remove after register abi is working
253 // dvarint writes a varint v to the funcdata in symbol x and returns the new offset.
254 func dvarint(x *obj.LSym, off int, v int64) int {
255 if v < 0 || v > 1e9 {
256 panic(fmt.Sprintf("dvarint: bad offset for funcdata - %v", v))
259 return objw.Uint8(x, off, uint8(v))
261 off = objw.Uint8(x, off, uint8((v&127)|128))
263 return objw.Uint8(x, off, uint8(v>>7))
265 off = objw.Uint8(x, off, uint8(((v>>7)&127)|128))
267 return objw.Uint8(x, off, uint8(v>>14))
269 off = objw.Uint8(x, off, uint8(((v>>14)&127)|128))
271 return objw.Uint8(x, off, uint8(v>>21))
273 off = objw.Uint8(x, off, uint8(((v>>21)&127)|128))
274 return objw.Uint8(x, off, uint8(v>>28))
277 // emitOpenDeferInfo emits FUNCDATA information about the defers in a function
278 // that is using open-coded defers. This funcdata is used to determine the active
279 // defers in a function and execute those defers during panic processing.
281 // The funcdata is all encoded in varints (since values will almost always be less than
282 // 128, but stack offsets could potentially be up to 2Gbyte). All "locations" (offsets)
283 // for stack variables are specified as the number of bytes below varp (pointer to the
284 // top of the local variables) for their starting address. The format is:
286 // - Offset of the deferBits variable
287 // - Offset of the first closure slot (the rest are laid out consecutively).
288 func (s *state) emitOpenDeferInfo() {
289 firstOffset := s.openDefers[0].closureNode.FrameOffset()
291 // Verify that cmpstackvarlt laid out the slots in order.
292 for i, r := range s.openDefers {
293 have := r.closureNode.FrameOffset()
294 want := firstOffset + int64(i)*int64(types.PtrSize)
296 base.FatalfAt(s.curfn.Pos(), "unexpected frame offset for open-coded defer slot #%v: have %v, want %v", i, have, want)
300 x := base.Ctxt.Lookup(s.curfn.LSym.Name + ".opendefer")
301 x.Set(obj.AttrContentAddressable, true)
302 s.curfn.LSym.Func().OpenCodedDeferInfo = x
305 off = dvarint(x, off, -s.deferBitsTemp.FrameOffset())
306 off = dvarint(x, off, -firstOffset)
309 // buildssa builds an SSA function for fn.
310 // worker indicates which of the backend workers is doing the processing.
311 func buildssa(fn *ir.Func, worker int) *ssa.Func {
312 name := ir.FuncName(fn)
314 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"
315 pkgDotName := base.Ctxt.Pkgpath + "." + name
316 printssa = name == ssaDump ||
317 strings.HasSuffix(pkgDotName, ssaDump) && (pkgDotName == ssaDump || strings.HasSuffix(pkgDotName, "/"+ssaDump))
319 var astBuf *bytes.Buffer
321 astBuf = &bytes.Buffer{}
322 ir.FDumpList(astBuf, "buildssa-enter", fn.Enter)
323 ir.FDumpList(astBuf, "buildssa-body", fn.Body)
324 ir.FDumpList(astBuf, "buildssa-exit", fn.Exit)
326 fmt.Println("generating SSA for", name)
327 fmt.Print(astBuf.String())
335 s.hasdefer = fn.HasDefer()
336 if fn.Pragma&ir.CgoUnsafeArgs != 0 {
337 s.cgoUnsafeArgs = true
339 s.checkPtrEnabled = ir.ShouldCheckPtr(fn, 1)
343 log: printssa && ssaDumpStdout,
347 cache := &ssaCaches[worker]
350 s.f = ssaConfig.NewFunc(&fe, cache)
354 s.f.PrintOrHtmlSSA = printssa
355 if fn.Pragma&ir.Nosplit != 0 {
358 s.f.ABI0 = ssaConfig.ABI0.Copy() // Make a copy to avoid racy map operations in type-register-width cache.
359 s.f.ABI1 = ssaConfig.ABI1.Copy()
360 s.f.ABIDefault = abiForFunc(nil, s.f.ABI0, s.f.ABI1)
361 s.f.ABISelf = abiForFunc(fn, s.f.ABI0, s.f.ABI1)
363 s.panics = map[funcLine]*ssa.Block{}
364 s.softFloat = s.config.SoftFloat
366 // Allocate starting block
367 s.f.Entry = s.f.NewBlock(ssa.BlockPlain)
368 s.f.Entry.Pos = fn.Pos()
373 ssaDF = filepath.Join(ssaDir, base.Ctxt.Pkgpath+"."+name+".html")
374 ssaD := filepath.Dir(ssaDF)
375 os.MkdirAll(ssaD, 0755)
377 s.f.HTMLWriter = ssa.NewHTMLWriter(ssaDF, s.f, ssaDumpCFG)
378 // TODO: generate and print a mapping from nodes to values and blocks
379 dumpSourcesColumn(s.f.HTMLWriter, fn)
380 s.f.HTMLWriter.WriteAST("AST", astBuf)
383 // Allocate starting values
384 s.labels = map[string]*ssaLabel{}
385 s.fwdVars = map[ir.Node]*ssa.Value{}
386 s.startmem = s.entryNewValue0(ssa.OpInitMem, types.TypeMem)
388 s.hasOpenDefers = base.Flag.N == 0 && s.hasdefer && !s.curfn.OpenCodedDeferDisallowed()
390 case base.Debug.NoOpenDefer != 0:
391 s.hasOpenDefers = false
392 case s.hasOpenDefers && (base.Ctxt.Flag_shared || base.Ctxt.Flag_dynlink) && base.Ctxt.Arch.Name == "386":
393 // Don't support open-coded defers for 386 ONLY when using shared
394 // libraries, because there is extra code (added by rewriteToUseGot())
395 // preceding the deferreturn/ret code that we don't track correctly.
396 s.hasOpenDefers = false
398 if s.hasOpenDefers && len(s.curfn.Exit) > 0 {
399 // Skip doing open defers if there is any extra exit code (likely
400 // race detection), since we will not generate that code in the
401 // case of the extra deferreturn/ret segment.
402 s.hasOpenDefers = false
405 // Similarly, skip if there are any heap-allocated result
406 // parameters that need to be copied back to their stack slots.
407 for _, f := range s.curfn.Type().Results() {
408 if !f.Nname.(*ir.Name).OnStack() {
409 s.hasOpenDefers = false
414 if s.hasOpenDefers &&
415 s.curfn.NumReturns*s.curfn.NumDefers > 15 {
416 // Since we are generating defer calls at every exit for
417 // open-coded defers, skip doing open-coded defers if there are
418 // too many returns (especially if there are multiple defers).
419 // Open-coded defers are most important for improving performance
420 // for smaller functions (which don't have many returns).
421 s.hasOpenDefers = false
424 s.sp = s.entryNewValue0(ssa.OpSP, types.Types[types.TUINTPTR]) // TODO: use generic pointer type (unsafe.Pointer?) instead
425 s.sb = s.entryNewValue0(ssa.OpSB, types.Types[types.TUINTPTR])
427 s.startBlock(s.f.Entry)
428 s.vars[memVar] = s.startmem
430 // Create the deferBits variable and stack slot. deferBits is a
431 // bitmask showing which of the open-coded defers in this function
432 // have been activated.
433 deferBitsTemp := typecheck.TempAt(src.NoXPos, s.curfn, types.Types[types.TUINT8])
434 deferBitsTemp.SetAddrtaken(true)
435 s.deferBitsTemp = deferBitsTemp
436 // For this value, AuxInt is initialized to zero by default
437 startDeferBits := s.entryNewValue0(ssa.OpConst8, types.Types[types.TUINT8])
438 s.vars[deferBitsVar] = startDeferBits
439 s.deferBitsAddr = s.addr(deferBitsTemp)
440 s.store(types.Types[types.TUINT8], s.deferBitsAddr, startDeferBits)
441 // Make sure that the deferBits stack slot is kept alive (for use
442 // by panics) and stores to deferBits are not eliminated, even if
443 // all checking code on deferBits in the function exit can be
444 // eliminated, because the defer statements were all
446 s.vars[memVar] = s.newValue1Apos(ssa.OpVarLive, types.TypeMem, deferBitsTemp, s.mem(), false)
449 var params *abi.ABIParamResultInfo
450 params = s.f.ABISelf.ABIAnalyze(fn.Type(), true)
452 // The backend's stackframe pass prunes away entries from the fn's
453 // Dcl list, including PARAMOUT nodes that correspond to output
454 // params passed in registers. Walk the Dcl list and capture these
455 // nodes to a side list, so that we'll have them available during
456 // DWARF-gen later on. See issue 48573 for more details.
457 var debugInfo ssa.FuncDebug
458 for _, n := range fn.Dcl {
459 if n.Class == ir.PPARAMOUT && n.IsOutputParamInRegisters() {
460 debugInfo.RegOutputParams = append(debugInfo.RegOutputParams, n)
463 fn.DebugInfo = &debugInfo
465 // Generate addresses of local declarations
466 s.decladdrs = map[*ir.Name]*ssa.Value{}
467 for _, n := range fn.Dcl {
470 // Be aware that blank and unnamed input parameters will not appear here, but do appear in the type
471 s.decladdrs[n] = s.entryNewValue2A(ssa.OpLocalAddr, types.NewPtr(n.Type()), n, s.sp, s.startmem)
473 s.decladdrs[n] = s.entryNewValue2A(ssa.OpLocalAddr, types.NewPtr(n.Type()), n, s.sp, s.startmem)
475 // processed at each use, to prevent Addr coming
478 s.Fatalf("local variable with class %v unimplemented", n.Class)
482 s.f.OwnAux = ssa.OwnAuxCall(fn.LSym, params)
484 // Populate SSAable arguments.
485 for _, n := range fn.Dcl {
486 if n.Class == ir.PPARAM {
488 v := s.newValue0A(ssa.OpArg, n.Type(), n)
490 s.addNamedValue(n, v) // This helps with debugging information, not needed for compilation itself.
491 } else { // address was taken AND/OR too large for SSA
492 paramAssignment := ssa.ParamAssignmentForArgName(s.f, n)
493 if len(paramAssignment.Registers) > 0 {
494 if ssa.CanSSA(n.Type()) { // SSA-able type, so address was taken -- receive value in OpArg, DO NOT bind to var, store immediately to memory.
495 v := s.newValue0A(ssa.OpArg, n.Type(), n)
496 s.store(n.Type(), s.decladdrs[n], v)
497 } else { // Too big for SSA.
498 // Brute force, and early, do a bunch of stores from registers
499 // TODO fix the nasty storeArgOrLoad recursion in ssa/expand_calls.go so this Just Works with store of a big Arg.
500 s.storeParameterRegsToStack(s.f.ABISelf, paramAssignment, n, s.decladdrs[n], false)
507 // Populate closure variables.
509 clo := s.entryNewValue0(ssa.OpGetClosurePtr, s.f.Config.Types.BytePtr)
510 offset := int64(types.PtrSize) // PtrSize to skip past function entry PC field
511 for _, n := range fn.ClosureVars {
514 typ = types.NewPtr(typ)
517 offset = types.RoundUp(offset, typ.Alignment())
518 ptr := s.newValue1I(ssa.OpOffPtr, types.NewPtr(typ), offset, clo)
521 // If n is a small variable captured by value, promote
522 // it to PAUTO so it can be converted to SSA.
524 // Note: While we never capture a variable by value if
525 // the user took its address, we may have generated
526 // runtime calls that did (#43701). Since we don't
527 // convert Addrtaken variables to SSA anyway, no point
528 // in promoting them either.
529 if n.Byval() && !n.Addrtaken() && ssa.CanSSA(n.Type()) {
531 fn.Dcl = append(fn.Dcl, n)
532 s.assign(n, s.load(n.Type(), ptr), false, 0)
537 ptr = s.load(typ, ptr)
539 s.setHeapaddr(fn.Pos(), n, ptr)
543 // Convert the AST-based IR to the SSA-based IR
549 // fallthrough to exit
550 if s.curBlock != nil {
551 s.pushLine(fn.Endlineno)
556 for _, b := range s.f.Blocks {
557 if b.Pos != src.NoXPos {
558 s.updateUnsetPredPos(b)
562 s.f.HTMLWriter.WritePhase("before insert phis", "before insert phis")
566 // Main call to ssa package to compile function
571 if len(s.openDefers) != 0 {
572 s.emitOpenDeferInfo()
575 // Record incoming parameter spill information for morestack calls emitted in the assembler.
576 // This is done here, using all the parameters (used, partially used, and unused) because
577 // it mimics the behavior of the former ABI (everything stored) and because it's not 100%
578 // clear if naming conventions are respected in autogenerated code.
579 // TODO figure out exactly what's unused, don't spill it. Make liveness fine-grained, also.
580 for _, p := range params.InParams() {
581 typs, offs := p.RegisterTypesAndOffsets()
582 for i, t := range typs {
583 o := offs[i] // offset within parameter
584 fo := p.FrameOffset(params) // offset of parameter in frame
585 reg := ssa.ObjRegForAbiReg(p.Registers[i], s.f.Config)
586 s.f.RegArgs = append(s.f.RegArgs, ssa.Spill{Reg: reg, Offset: fo + o, Type: t})
593 func (s *state) storeParameterRegsToStack(abi *abi.ABIConfig, paramAssignment *abi.ABIParamAssignment, n *ir.Name, addr *ssa.Value, pointersOnly bool) {
594 typs, offs := paramAssignment.RegisterTypesAndOffsets()
595 for i, t := range typs {
596 if pointersOnly && !t.IsPtrShaped() {
599 r := paramAssignment.Registers[i]
601 op, reg := ssa.ArgOpAndRegisterFor(r, abi)
602 aux := &ssa.AuxNameOffset{Name: n, Offset: o}
603 v := s.newValue0I(op, t, reg)
605 p := s.newValue1I(ssa.OpOffPtr, types.NewPtr(t), o, addr)
610 // zeroResults zeros the return values at the start of the function.
611 // We need to do this very early in the function. Defer might stop a
612 // panic and show the return values as they exist at the time of
613 // panic. For precise stacks, the garbage collector assumes results
614 // are always live, so we need to zero them before any allocations,
615 // even allocations to move params/results to the heap.
616 func (s *state) zeroResults() {
617 for _, f := range s.curfn.Type().Results() {
618 n := f.Nname.(*ir.Name)
620 // The local which points to the return value is the
621 // thing that needs zeroing. This is already handled
622 // by a Needzero annotation in plive.go:(*liveness).epilogue.
625 // Zero the stack location containing f.
626 if typ := n.Type(); ssa.CanSSA(typ) {
627 s.assign(n, s.zeroVal(typ), false, 0)
629 if typ.HasPointers() {
630 s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, n, s.mem())
632 s.zero(n.Type(), s.decladdrs[n])
637 // paramsToHeap produces code to allocate memory for heap-escaped parameters
638 // and to copy non-result parameters' values from the stack.
639 func (s *state) paramsToHeap() {
640 do := func(params []*types.Field) {
641 for _, f := range params {
643 continue // anonymous or blank parameter
645 n := f.Nname.(*ir.Name)
646 if ir.IsBlank(n) || n.OnStack() {
650 if n.Class == ir.PPARAM {
651 s.move(n.Type(), s.expr(n.Heapaddr), s.decladdrs[n])
656 typ := s.curfn.Type()
662 // newHeapaddr allocates heap memory for n and sets its heap address.
663 func (s *state) newHeapaddr(n *ir.Name) {
664 s.setHeapaddr(n.Pos(), n, s.newObject(n.Type(), nil))
667 // setHeapaddr allocates a new PAUTO variable to store ptr (which must be non-nil)
668 // and then sets it as n's heap address.
669 func (s *state) setHeapaddr(pos src.XPos, n *ir.Name, ptr *ssa.Value) {
670 if !ptr.Type.IsPtr() || !types.Identical(n.Type(), ptr.Type.Elem()) {
671 base.FatalfAt(n.Pos(), "setHeapaddr %L with type %v", n, ptr.Type)
674 // Declare variable to hold address.
675 sym := &types.Sym{Name: "&" + n.Sym().Name, Pkg: types.LocalPkg}
676 addr := s.curfn.NewLocal(pos, sym, types.NewPtr(n.Type()))
678 types.CalcSize(addr.Type())
680 if n.Class == ir.PPARAMOUT {
681 addr.SetIsOutputParamHeapAddr(true)
685 s.assign(addr, ptr, false, 0)
688 // newObject returns an SSA value denoting new(typ).
689 func (s *state) newObject(typ *types.Type, rtype *ssa.Value) *ssa.Value {
691 return s.newValue1A(ssa.OpAddr, types.NewPtr(typ), ir.Syms.Zerobase, s.sb)
694 rtype = s.reflectType(typ)
696 return s.rtcall(ir.Syms.Newobject, true, []*types.Type{types.NewPtr(typ)}, rtype)[0]
699 func (s *state) checkPtrAlignment(n *ir.ConvExpr, v *ssa.Value, count *ssa.Value) {
700 if !n.Type().IsPtr() {
701 s.Fatalf("expected pointer type: %v", n.Type())
703 elem, rtypeExpr := n.Type().Elem(), n.ElemRType
706 s.Fatalf("expected array type: %v", elem)
708 elem, rtypeExpr = elem.Elem(), n.ElemElemRType
711 // Casting from larger type to smaller one is ok, so for smallest type, do nothing.
712 if elem.Alignment() == 1 && (size == 0 || size == 1 || count == nil) {
716 count = s.constInt(types.Types[types.TUINTPTR], 1)
718 if count.Type.Size() != s.config.PtrSize {
719 s.Fatalf("expected count fit to a uintptr size, have: %d, want: %d", count.Type.Size(), s.config.PtrSize)
722 if rtypeExpr != nil {
723 rtype = s.expr(rtypeExpr)
725 rtype = s.reflectType(elem)
727 s.rtcall(ir.Syms.CheckPtrAlignment, true, nil, v, rtype, count)
730 // reflectType returns an SSA value representing a pointer to typ's
731 // reflection type descriptor.
732 func (s *state) reflectType(typ *types.Type) *ssa.Value {
733 // TODO(mdempsky): Make this Fatalf under Unified IR; frontend needs
734 // to supply RType expressions.
735 lsym := reflectdata.TypeLinksym(typ)
736 return s.entryNewValue1A(ssa.OpAddr, types.NewPtr(types.Types[types.TUINT8]), lsym, s.sb)
739 func dumpSourcesColumn(writer *ssa.HTMLWriter, fn *ir.Func) {
740 // Read sources of target function fn.
741 fname := base.Ctxt.PosTable.Pos(fn.Pos()).Filename()
742 targetFn, err := readFuncLines(fname, fn.Pos().Line(), fn.Endlineno.Line())
744 writer.Logf("cannot read sources for function %v: %v", fn, err)
747 // Read sources of inlined functions.
748 var inlFns []*ssa.FuncLines
749 for _, fi := range ssaDumpInlined {
751 fname := base.Ctxt.PosTable.Pos(fi.Pos()).Filename()
752 fnLines, err := readFuncLines(fname, fi.Pos().Line(), elno.Line())
754 writer.Logf("cannot read sources for inlined function %v: %v", fi, err)
757 inlFns = append(inlFns, fnLines)
760 sort.Sort(ssa.ByTopo(inlFns))
762 inlFns = append([]*ssa.FuncLines{targetFn}, inlFns...)
765 writer.WriteSources("sources", inlFns)
768 func readFuncLines(file string, start, end uint) (*ssa.FuncLines, error) {
769 f, err := os.Open(os.ExpandEnv(file))
776 scanner := bufio.NewScanner(f)
777 for scanner.Scan() && ln <= end {
779 lines = append(lines, scanner.Text())
783 return &ssa.FuncLines{Filename: file, StartLineno: start, Lines: lines}, nil
786 // updateUnsetPredPos propagates the earliest-value position information for b
787 // towards all of b's predecessors that need a position, and recurs on that
788 // predecessor if its position is updated. B should have a non-empty position.
789 func (s *state) updateUnsetPredPos(b *ssa.Block) {
790 if b.Pos == src.NoXPos {
791 s.Fatalf("Block %s should have a position", b)
793 bestPos := src.NoXPos
794 for _, e := range b.Preds {
799 if bestPos == src.NoXPos {
801 for _, v := range b.Values {
805 if v.Pos != src.NoXPos {
806 // Assume values are still in roughly textual order;
807 // TODO: could also seek minimum position?
814 s.updateUnsetPredPos(p) // We do not expect long chains of these, thus recursion is okay.
818 // Information about each open-coded defer.
819 type openDeferInfo struct {
820 // The node representing the call of the defer
822 // If defer call is closure call, the address of the argtmp where the
823 // closure is stored.
825 // The node representing the argtmp where the closure is stored - used for
826 // function, method, or interface call, to store a closure that panic
827 // processing can use for this defer.
832 // configuration (arch) information
835 // function we're building
842 labels map[string]*ssaLabel
844 // unlabeled break and continue statement tracking
845 breakTo *ssa.Block // current target for plain break statement
846 continueTo *ssa.Block // current target for plain continue statement
848 // current location where we're interpreting the AST
851 // variable assignments in the current block (map from variable symbol to ssa value)
852 // *Node is the unique identifier (an ONAME Node) for the variable.
853 // TODO: keep a single varnum map, then make all of these maps slices instead?
854 vars map[ir.Node]*ssa.Value
856 // fwdVars are variables that are used before they are defined in the current block.
857 // This map exists just to coalesce multiple references into a single FwdRef op.
858 // *Node is the unique identifier (an ONAME Node) for the variable.
859 fwdVars map[ir.Node]*ssa.Value
861 // all defined variables at the end of each block. Indexed by block ID.
862 defvars []map[ir.Node]*ssa.Value
864 // addresses of PPARAM and PPARAMOUT variables on the stack.
865 decladdrs map[*ir.Name]*ssa.Value
867 // starting values. Memory, stack pointer, and globals pointer
871 // value representing address of where deferBits autotmp is stored
872 deferBitsAddr *ssa.Value
873 deferBitsTemp *ir.Name
875 // line number stack. The current line number is top of stack
877 // the last line number processed; it may have been popped
880 // list of panic calls by function name and line number.
881 // Used to deduplicate panic calls.
882 panics map[funcLine]*ssa.Block
885 hasdefer bool // whether the function contains a defer statement
887 hasOpenDefers bool // whether we are doing open-coded defers
888 checkPtrEnabled bool // whether to insert checkptr instrumentation
890 // If doing open-coded defers, list of info about the defer calls in
891 // scanning order. Hence, at exit we should run these defers in reverse
892 // order of this list
893 openDefers []*openDeferInfo
894 // For open-coded defers, this is the beginning and end blocks of the last
895 // defer exit code that we have generated so far. We use these to share
896 // code between exits if the shareDeferExits option (disabled by default)
898 lastDeferExit *ssa.Block // Entry block of last defer exit code we generated
899 lastDeferFinalBlock *ssa.Block // Final block of last defer exit code we generated
900 lastDeferCount int // Number of defers encountered at that point
902 prevCall *ssa.Value // the previous call; use this to tie results to the call op.
905 type funcLine struct {
911 type ssaLabel struct {
912 target *ssa.Block // block identified by this label
913 breakTarget *ssa.Block // block to break to in control flow node identified by this label
914 continueTarget *ssa.Block // block to continue to in control flow node identified by this label
917 // label returns the label associated with sym, creating it if necessary.
918 func (s *state) label(sym *types.Sym) *ssaLabel {
919 lab := s.labels[sym.Name]
922 s.labels[sym.Name] = lab
927 func (s *state) Logf(msg string, args ...interface{}) { s.f.Logf(msg, args...) }
928 func (s *state) Log() bool { return s.f.Log() }
929 func (s *state) Fatalf(msg string, args ...interface{}) {
930 s.f.Frontend().Fatalf(s.peekPos(), msg, args...)
932 func (s *state) Warnl(pos src.XPos, msg string, args ...interface{}) { s.f.Warnl(pos, msg, args...) }
933 func (s *state) Debug_checknil() bool { return s.f.Frontend().Debug_checknil() }
935 func ssaMarker(name string) *ir.Name {
936 return ir.NewNameAt(base.Pos, &types.Sym{Name: name}, nil)
940 // marker node for the memory variable
941 memVar = ssaMarker("mem")
943 // marker nodes for temporary variables
944 ptrVar = ssaMarker("ptr")
945 lenVar = ssaMarker("len")
946 capVar = ssaMarker("cap")
947 typVar = ssaMarker("typ")
948 okVar = ssaMarker("ok")
949 deferBitsVar = ssaMarker("deferBits")
952 // startBlock sets the current block we're generating code in to b.
953 func (s *state) startBlock(b *ssa.Block) {
954 if s.curBlock != nil {
955 s.Fatalf("starting block %v when block %v has not ended", b, s.curBlock)
958 s.vars = map[ir.Node]*ssa.Value{}
959 for n := range s.fwdVars {
964 // endBlock marks the end of generating code for the current block.
965 // Returns the (former) current block. Returns nil if there is no current
966 // block, i.e. if no code flows to the current execution point.
967 func (s *state) endBlock() *ssa.Block {
972 for len(s.defvars) <= int(b.ID) {
973 s.defvars = append(s.defvars, nil)
975 s.defvars[b.ID] = s.vars
979 // Empty plain blocks get the line of their successor (handled after all blocks created),
980 // except for increment blocks in For statements (handled in ssa conversion of OFOR),
981 // and for blocks ending in GOTO/BREAK/CONTINUE.
989 // pushLine pushes a line number on the line number stack.
990 func (s *state) pushLine(line src.XPos) {
992 // the frontend may emit node with line number missing,
993 // use the parent line number in this case.
995 if base.Flag.K != 0 {
996 base.Warn("buildssa: unknown position (line 0)")
1002 s.line = append(s.line, line)
1005 // popLine pops the top of the line number stack.
1006 func (s *state) popLine() {
1007 s.line = s.line[:len(s.line)-1]
1010 // peekPos peeks the top of the line number stack.
1011 func (s *state) peekPos() src.XPos {
1012 return s.line[len(s.line)-1]
1015 // newValue0 adds a new value with no arguments to the current block.
1016 func (s *state) newValue0(op ssa.Op, t *types.Type) *ssa.Value {
1017 return s.curBlock.NewValue0(s.peekPos(), op, t)
1020 // newValue0A adds a new value with no arguments and an aux value to the current block.
1021 func (s *state) newValue0A(op ssa.Op, t *types.Type, aux ssa.Aux) *ssa.Value {
1022 return s.curBlock.NewValue0A(s.peekPos(), op, t, aux)
1025 // newValue0I adds a new value with no arguments and an auxint value to the current block.
1026 func (s *state) newValue0I(op ssa.Op, t *types.Type, auxint int64) *ssa.Value {
1027 return s.curBlock.NewValue0I(s.peekPos(), op, t, auxint)
1030 // newValue1 adds a new value with one argument to the current block.
1031 func (s *state) newValue1(op ssa.Op, t *types.Type, arg *ssa.Value) *ssa.Value {
1032 return s.curBlock.NewValue1(s.peekPos(), op, t, arg)
1035 // newValue1A adds a new value with one argument and an aux value to the current block.
1036 func (s *state) newValue1A(op ssa.Op, t *types.Type, aux ssa.Aux, arg *ssa.Value) *ssa.Value {
1037 return s.curBlock.NewValue1A(s.peekPos(), op, t, aux, arg)
1040 // newValue1Apos adds a new value with one argument and an aux value to the current block.
1041 // isStmt determines whether the created values may be a statement or not
1042 // (i.e., false means never, yes means maybe).
1043 func (s *state) newValue1Apos(op ssa.Op, t *types.Type, aux ssa.Aux, arg *ssa.Value, isStmt bool) *ssa.Value {
1045 return s.curBlock.NewValue1A(s.peekPos(), op, t, aux, arg)
1047 return s.curBlock.NewValue1A(s.peekPos().WithNotStmt(), op, t, aux, arg)
1050 // newValue1I adds a new value with one argument and an auxint value to the current block.
1051 func (s *state) newValue1I(op ssa.Op, t *types.Type, aux int64, arg *ssa.Value) *ssa.Value {
1052 return s.curBlock.NewValue1I(s.peekPos(), op, t, aux, arg)
1055 // newValue2 adds a new value with two arguments to the current block.
1056 func (s *state) newValue2(op ssa.Op, t *types.Type, arg0, arg1 *ssa.Value) *ssa.Value {
1057 return s.curBlock.NewValue2(s.peekPos(), op, t, arg0, arg1)
1060 // newValue2A adds a new value with two arguments and an aux value to the current block.
1061 func (s *state) newValue2A(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1 *ssa.Value) *ssa.Value {
1062 return s.curBlock.NewValue2A(s.peekPos(), op, t, aux, arg0, arg1)
1065 // newValue2Apos adds a new value with two arguments and an aux value to the current block.
1066 // isStmt determines whether the created values may be a statement or not
1067 // (i.e., false means never, yes means maybe).
1068 func (s *state) newValue2Apos(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1 *ssa.Value, isStmt bool) *ssa.Value {
1070 return s.curBlock.NewValue2A(s.peekPos(), op, t, aux, arg0, arg1)
1072 return s.curBlock.NewValue2A(s.peekPos().WithNotStmt(), op, t, aux, arg0, arg1)
1075 // newValue2I adds a new value with two arguments and an auxint value to the current block.
1076 func (s *state) newValue2I(op ssa.Op, t *types.Type, aux int64, arg0, arg1 *ssa.Value) *ssa.Value {
1077 return s.curBlock.NewValue2I(s.peekPos(), op, t, aux, arg0, arg1)
1080 // newValue3 adds a new value with three arguments to the current block.
1081 func (s *state) newValue3(op ssa.Op, t *types.Type, arg0, arg1, arg2 *ssa.Value) *ssa.Value {
1082 return s.curBlock.NewValue3(s.peekPos(), op, t, arg0, arg1, arg2)
1085 // newValue3I adds a new value with three arguments and an auxint value to the current block.
1086 func (s *state) newValue3I(op ssa.Op, t *types.Type, aux int64, arg0, arg1, arg2 *ssa.Value) *ssa.Value {
1087 return s.curBlock.NewValue3I(s.peekPos(), op, t, aux, arg0, arg1, arg2)
1090 // newValue3A adds a new value with three arguments and an aux value to the current block.
1091 func (s *state) newValue3A(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1, arg2 *ssa.Value) *ssa.Value {
1092 return s.curBlock.NewValue3A(s.peekPos(), op, t, aux, arg0, arg1, arg2)
1095 // newValue3Apos adds a new value with three arguments and an aux value to the current block.
1096 // isStmt determines whether the created values may be a statement or not
1097 // (i.e., false means never, yes means maybe).
1098 func (s *state) newValue3Apos(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1, arg2 *ssa.Value, isStmt bool) *ssa.Value {
1100 return s.curBlock.NewValue3A(s.peekPos(), op, t, aux, arg0, arg1, arg2)
1102 return s.curBlock.NewValue3A(s.peekPos().WithNotStmt(), op, t, aux, arg0, arg1, arg2)
1105 // newValue4 adds a new value with four arguments to the current block.
1106 func (s *state) newValue4(op ssa.Op, t *types.Type, arg0, arg1, arg2, arg3 *ssa.Value) *ssa.Value {
1107 return s.curBlock.NewValue4(s.peekPos(), op, t, arg0, arg1, arg2, arg3)
1110 // newValue4I adds a new value with four arguments and an auxint value to the current block.
1111 func (s *state) newValue4I(op ssa.Op, t *types.Type, aux int64, arg0, arg1, arg2, arg3 *ssa.Value) *ssa.Value {
1112 return s.curBlock.NewValue4I(s.peekPos(), op, t, aux, arg0, arg1, arg2, arg3)
1115 func (s *state) entryBlock() *ssa.Block {
1117 if base.Flag.N > 0 && s.curBlock != nil {
1118 // If optimizations are off, allocate in current block instead. Since with -N
1119 // we're not doing the CSE or tighten passes, putting lots of stuff in the
1120 // entry block leads to O(n^2) entries in the live value map during regalloc.
1127 // entryNewValue0 adds a new value with no arguments to the entry block.
1128 func (s *state) entryNewValue0(op ssa.Op, t *types.Type) *ssa.Value {
1129 return s.entryBlock().NewValue0(src.NoXPos, op, t)
1132 // entryNewValue0A adds a new value with no arguments and an aux value to the entry block.
1133 func (s *state) entryNewValue0A(op ssa.Op, t *types.Type, aux ssa.Aux) *ssa.Value {
1134 return s.entryBlock().NewValue0A(src.NoXPos, op, t, aux)
1137 // entryNewValue1 adds a new value with one argument to the entry block.
1138 func (s *state) entryNewValue1(op ssa.Op, t *types.Type, arg *ssa.Value) *ssa.Value {
1139 return s.entryBlock().NewValue1(src.NoXPos, op, t, arg)
1142 // entryNewValue1I adds a new value with one argument and an auxint value to the entry block.
1143 func (s *state) entryNewValue1I(op ssa.Op, t *types.Type, auxint int64, arg *ssa.Value) *ssa.Value {
1144 return s.entryBlock().NewValue1I(src.NoXPos, op, t, auxint, arg)
1147 // entryNewValue1A adds a new value with one argument and an aux value to the entry block.
1148 func (s *state) entryNewValue1A(op ssa.Op, t *types.Type, aux ssa.Aux, arg *ssa.Value) *ssa.Value {
1149 return s.entryBlock().NewValue1A(src.NoXPos, op, t, aux, arg)
1152 // entryNewValue2 adds a new value with two arguments to the entry block.
1153 func (s *state) entryNewValue2(op ssa.Op, t *types.Type, arg0, arg1 *ssa.Value) *ssa.Value {
1154 return s.entryBlock().NewValue2(src.NoXPos, op, t, arg0, arg1)
1157 // entryNewValue2A adds a new value with two arguments and an aux value to the entry block.
1158 func (s *state) entryNewValue2A(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1 *ssa.Value) *ssa.Value {
1159 return s.entryBlock().NewValue2A(src.NoXPos, op, t, aux, arg0, arg1)
1162 // const* routines add a new const value to the entry block.
1163 func (s *state) constSlice(t *types.Type) *ssa.Value {
1164 return s.f.ConstSlice(t)
1166 func (s *state) constInterface(t *types.Type) *ssa.Value {
1167 return s.f.ConstInterface(t)
1169 func (s *state) constNil(t *types.Type) *ssa.Value { return s.f.ConstNil(t) }
1170 func (s *state) constEmptyString(t *types.Type) *ssa.Value {
1171 return s.f.ConstEmptyString(t)
1173 func (s *state) constBool(c bool) *ssa.Value {
1174 return s.f.ConstBool(types.Types[types.TBOOL], c)
1176 func (s *state) constInt8(t *types.Type, c int8) *ssa.Value {
1177 return s.f.ConstInt8(t, c)
1179 func (s *state) constInt16(t *types.Type, c int16) *ssa.Value {
1180 return s.f.ConstInt16(t, c)
1182 func (s *state) constInt32(t *types.Type, c int32) *ssa.Value {
1183 return s.f.ConstInt32(t, c)
1185 func (s *state) constInt64(t *types.Type, c int64) *ssa.Value {
1186 return s.f.ConstInt64(t, c)
1188 func (s *state) constFloat32(t *types.Type, c float64) *ssa.Value {
1189 return s.f.ConstFloat32(t, c)
1191 func (s *state) constFloat64(t *types.Type, c float64) *ssa.Value {
1192 return s.f.ConstFloat64(t, c)
1194 func (s *state) constInt(t *types.Type, c int64) *ssa.Value {
1195 if s.config.PtrSize == 8 {
1196 return s.constInt64(t, c)
1198 if int64(int32(c)) != c {
1199 s.Fatalf("integer constant too big %d", c)
1201 return s.constInt32(t, int32(c))
1203 func (s *state) constOffPtrSP(t *types.Type, c int64) *ssa.Value {
1204 return s.f.ConstOffPtrSP(t, c, s.sp)
1207 // newValueOrSfCall* are wrappers around newValue*, which may create a call to a
1208 // soft-float runtime function instead (when emitting soft-float code).
1209 func (s *state) newValueOrSfCall1(op ssa.Op, t *types.Type, arg *ssa.Value) *ssa.Value {
1211 if c, ok := s.sfcall(op, arg); ok {
1215 return s.newValue1(op, t, arg)
1217 func (s *state) newValueOrSfCall2(op ssa.Op, t *types.Type, arg0, arg1 *ssa.Value) *ssa.Value {
1219 if c, ok := s.sfcall(op, arg0, arg1); ok {
1223 return s.newValue2(op, t, arg0, arg1)
1226 type instrumentKind uint8
1229 instrumentRead = iota
1234 func (s *state) instrument(t *types.Type, addr *ssa.Value, kind instrumentKind) {
1235 s.instrument2(t, addr, nil, kind)
1238 // instrumentFields instruments a read/write operation on addr.
1239 // If it is instrumenting for MSAN or ASAN and t is a struct type, it instruments
1240 // operation for each field, instead of for the whole struct.
1241 func (s *state) instrumentFields(t *types.Type, addr *ssa.Value, kind instrumentKind) {
1242 if !(base.Flag.MSan || base.Flag.ASan) || !t.IsStruct() {
1243 s.instrument(t, addr, kind)
1246 for _, f := range t.Fields() {
1247 if f.Sym.IsBlank() {
1250 offptr := s.newValue1I(ssa.OpOffPtr, types.NewPtr(f.Type), f.Offset, addr)
1251 s.instrumentFields(f.Type, offptr, kind)
1255 func (s *state) instrumentMove(t *types.Type, dst, src *ssa.Value) {
1257 s.instrument2(t, dst, src, instrumentMove)
1259 s.instrument(t, src, instrumentRead)
1260 s.instrument(t, dst, instrumentWrite)
1264 func (s *state) instrument2(t *types.Type, addr, addr2 *ssa.Value, kind instrumentKind) {
1265 if !s.curfn.InstrumentBody() {
1271 return // can't race on zero-sized things
1274 if ssa.IsSanitizerSafeAddr(addr) {
1281 if addr2 != nil && kind != instrumentMove {
1282 panic("instrument2: non-nil addr2 for non-move instrumentation")
1287 case instrumentRead:
1288 fn = ir.Syms.Msanread
1289 case instrumentWrite:
1290 fn = ir.Syms.Msanwrite
1291 case instrumentMove:
1292 fn = ir.Syms.Msanmove
1294 panic("unreachable")
1297 } else if base.Flag.Race && t.NumComponents(types.CountBlankFields) > 1 {
1298 // for composite objects we have to write every address
1299 // because a write might happen to any subobject.
1300 // composites with only one element don't have subobjects, though.
1302 case instrumentRead:
1303 fn = ir.Syms.Racereadrange
1304 case instrumentWrite:
1305 fn = ir.Syms.Racewriterange
1307 panic("unreachable")
1310 } else if base.Flag.Race {
1311 // for non-composite objects we can write just the start
1312 // address, as any write must write the first byte.
1314 case instrumentRead:
1315 fn = ir.Syms.Raceread
1316 case instrumentWrite:
1317 fn = ir.Syms.Racewrite
1319 panic("unreachable")
1321 } else if base.Flag.ASan {
1323 case instrumentRead:
1324 fn = ir.Syms.Asanread
1325 case instrumentWrite:
1326 fn = ir.Syms.Asanwrite
1328 panic("unreachable")
1332 panic("unreachable")
1335 args := []*ssa.Value{addr}
1337 args = append(args, addr2)
1340 args = append(args, s.constInt(types.Types[types.TUINTPTR], w))
1342 s.rtcall(fn, true, nil, args...)
1345 func (s *state) load(t *types.Type, src *ssa.Value) *ssa.Value {
1346 s.instrumentFields(t, src, instrumentRead)
1347 return s.rawLoad(t, src)
1350 func (s *state) rawLoad(t *types.Type, src *ssa.Value) *ssa.Value {
1351 return s.newValue2(ssa.OpLoad, t, src, s.mem())
1354 func (s *state) store(t *types.Type, dst, val *ssa.Value) {
1355 s.vars[memVar] = s.newValue3A(ssa.OpStore, types.TypeMem, t, dst, val, s.mem())
1358 func (s *state) zero(t *types.Type, dst *ssa.Value) {
1359 s.instrument(t, dst, instrumentWrite)
1360 store := s.newValue2I(ssa.OpZero, types.TypeMem, t.Size(), dst, s.mem())
1362 s.vars[memVar] = store
1365 func (s *state) move(t *types.Type, dst, src *ssa.Value) {
1366 s.moveWhichMayOverlap(t, dst, src, false)
1368 func (s *state) moveWhichMayOverlap(t *types.Type, dst, src *ssa.Value, mayOverlap bool) {
1369 s.instrumentMove(t, dst, src)
1370 if mayOverlap && t.IsArray() && t.NumElem() > 1 && !ssa.IsInlinableMemmove(dst, src, t.Size(), s.f.Config) {
1371 // Normally, when moving Go values of type T from one location to another,
1372 // we don't need to worry about partial overlaps. The two Ts must either be
1373 // in disjoint (nonoverlapping) memory or in exactly the same location.
1374 // There are 2 cases where this isn't true:
1375 // 1) Using unsafe you can arrange partial overlaps.
1376 // 2) Since Go 1.17, you can use a cast from a slice to a ptr-to-array.
1377 // https://go.dev/ref/spec#Conversions_from_slice_to_array_pointer
1378 // This feature can be used to construct partial overlaps of array types.
1380 // p := (*[2]int)(a[:])
1381 // q := (*[2]int)(a[1:])
1383 // We don't care about solving 1. Or at least, we haven't historically
1384 // and no one has complained.
1385 // For 2, we need to ensure that if there might be partial overlap,
1386 // then we can't use OpMove; we must use memmove instead.
1387 // (memmove handles partial overlap by copying in the correct
1388 // direction. OpMove does not.)
1390 // Note that we have to be careful here not to introduce a call when
1391 // we're marshaling arguments to a call or unmarshaling results from a call.
1392 // Cases where this is happening must pass mayOverlap to false.
1393 // (Currently this only happens when unmarshaling results of a call.)
1394 if t.HasPointers() {
1395 s.rtcall(ir.Syms.Typedmemmove, true, nil, s.reflectType(t), dst, src)
1396 // We would have otherwise implemented this move with straightline code,
1397 // including a write barrier. Pretend we issue a write barrier here,
1398 // so that the write barrier tests work. (Otherwise they'd need to know
1399 // the details of IsInlineableMemmove.)
1400 s.curfn.SetWBPos(s.peekPos())
1402 s.rtcall(ir.Syms.Memmove, true, nil, dst, src, s.constInt(types.Types[types.TUINTPTR], t.Size()))
1404 ssa.LogLargeCopy(s.f.Name, s.peekPos(), t.Size())
1407 store := s.newValue3I(ssa.OpMove, types.TypeMem, t.Size(), dst, src, s.mem())
1409 s.vars[memVar] = store
1412 // stmtList converts the statement list n to SSA and adds it to s.
1413 func (s *state) stmtList(l ir.Nodes) {
1414 for _, n := range l {
1419 // stmt converts the statement n to SSA and adds it to s.
1420 func (s *state) stmt(n ir.Node) {
1424 // If s.curBlock is nil, and n isn't a label (which might have an associated goto somewhere),
1425 // then this code is dead. Stop here.
1426 if s.curBlock == nil && n.Op() != ir.OLABEL {
1430 s.stmtList(n.Init())
1434 n := n.(*ir.BlockStmt)
1437 case ir.OFALL: // no-op
1439 // Expression statements
1441 n := n.(*ir.CallExpr)
1442 if ir.IsIntrinsicCall(n) {
1449 n := n.(*ir.CallExpr)
1450 s.callResult(n, callNormal)
1451 if n.Op() == ir.OCALLFUNC && n.X.Op() == ir.ONAME && n.X.(*ir.Name).Class == ir.PFUNC {
1452 if fn := n.X.Sym().Name; base.Flag.CompilingRuntime && fn == "throw" ||
1453 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") {
1456 b.Kind = ssa.BlockExit
1458 // TODO: never rewrite OPANIC to OCALLFUNC in the
1459 // first place. Need to wait until all backends
1464 n := n.(*ir.GoDeferStmt)
1465 if base.Debug.Defer > 0 {
1466 var defertype string
1467 if s.hasOpenDefers {
1468 defertype = "open-coded"
1469 } else if n.Esc() == ir.EscNever {
1470 defertype = "stack-allocated"
1472 defertype = "heap-allocated"
1474 base.WarnfAt(n.Pos(), "%s defer", defertype)
1476 if s.hasOpenDefers {
1477 s.openDeferRecord(n.Call.(*ir.CallExpr))
1480 if n.Esc() == ir.EscNever {
1483 s.callResult(n.Call.(*ir.CallExpr), d)
1486 n := n.(*ir.GoDeferStmt)
1487 s.callResult(n.Call.(*ir.CallExpr), callGo)
1489 case ir.OAS2DOTTYPE:
1490 n := n.(*ir.AssignListStmt)
1491 var res, resok *ssa.Value
1492 if n.Rhs[0].Op() == ir.ODOTTYPE2 {
1493 res, resok = s.dottype(n.Rhs[0].(*ir.TypeAssertExpr), true)
1495 res, resok = s.dynamicDottype(n.Rhs[0].(*ir.DynamicTypeAssertExpr), true)
1498 if !ssa.CanSSA(n.Rhs[0].Type()) {
1499 if res.Op != ssa.OpLoad {
1500 s.Fatalf("dottype of non-load")
1503 if res.Args[1] != mem {
1504 s.Fatalf("memory no longer live from 2-result dottype load")
1509 s.assign(n.Lhs[0], res, deref, 0)
1510 s.assign(n.Lhs[1], resok, false, 0)
1514 // We come here only when it is an intrinsic call returning two values.
1515 n := n.(*ir.AssignListStmt)
1516 call := n.Rhs[0].(*ir.CallExpr)
1517 if !ir.IsIntrinsicCall(call) {
1518 s.Fatalf("non-intrinsic AS2FUNC not expanded %v", call)
1520 v := s.intrinsicCall(call)
1521 v1 := s.newValue1(ssa.OpSelect0, n.Lhs[0].Type(), v)
1522 v2 := s.newValue1(ssa.OpSelect1, n.Lhs[1].Type(), v)
1523 s.assign(n.Lhs[0], v1, false, 0)
1524 s.assign(n.Lhs[1], v2, false, 0)
1529 if v := n.X; v.Esc() == ir.EscHeap {
1534 n := n.(*ir.LabelStmt)
1537 // Nothing to do because the label isn't targetable. See issue 52278.
1542 // The label might already have a target block via a goto.
1543 if lab.target == nil {
1544 lab.target = s.f.NewBlock(ssa.BlockPlain)
1547 // Go to that label.
1548 // (We pretend "label:" is preceded by "goto label", unless the predecessor is unreachable.)
1549 if s.curBlock != nil {
1551 b.AddEdgeTo(lab.target)
1553 s.startBlock(lab.target)
1556 n := n.(*ir.BranchStmt)
1560 if lab.target == nil {
1561 lab.target = s.f.NewBlock(ssa.BlockPlain)
1565 b.Pos = s.lastPos.WithIsStmt() // Do this even if b is an empty block.
1566 b.AddEdgeTo(lab.target)
1569 n := n.(*ir.AssignStmt)
1570 if n.X == n.Y && n.X.Op() == ir.ONAME {
1571 // An x=x assignment. No point in doing anything
1572 // here. In addition, skipping this assignment
1573 // prevents generating:
1576 // which is bad because x is incorrectly considered
1577 // dead before the vardef. See issue #14904.
1581 // mayOverlap keeps track of whether the LHS and RHS might
1582 // refer to partially overlapping memory. Partial overlapping can
1583 // only happen for arrays, see the comment in moveWhichMayOverlap.
1585 // If both sides of the assignment are not dereferences, then partial
1586 // overlap can't happen. Partial overlap can only occur only when the
1587 // arrays referenced are strictly smaller parts of the same base array.
1588 // If one side of the assignment is a full array, then partial overlap
1589 // can't happen. (The arrays are either disjoint or identical.)
1590 mayOverlap := n.X.Op() == ir.ODEREF && (n.Y != nil && n.Y.Op() == ir.ODEREF)
1591 if n.Y != nil && n.Y.Op() == ir.ODEREF {
1592 p := n.Y.(*ir.StarExpr).X
1593 for p.Op() == ir.OCONVNOP {
1594 p = p.(*ir.ConvExpr).X
1596 if p.Op() == ir.OSPTR && p.(*ir.UnaryExpr).X.Type().IsString() {
1597 // Pointer fields of strings point to unmodifiable memory.
1598 // That memory can't overlap with the memory being written.
1607 case ir.OSTRUCTLIT, ir.OARRAYLIT, ir.OSLICELIT:
1608 // All literals with nonzero fields have already been
1609 // rewritten during walk. Any that remain are just T{}
1610 // or equivalents. Use the zero value.
1611 if !ir.IsZero(rhs) {
1612 s.Fatalf("literal with nonzero value in SSA: %v", rhs)
1616 rhs := rhs.(*ir.CallExpr)
1617 // Check whether we're writing the result of an append back to the same slice.
1618 // If so, we handle it specially to avoid write barriers on the fast
1619 // (non-growth) path.
1620 if !ir.SameSafeExpr(n.X, rhs.Args[0]) || base.Flag.N != 0 {
1623 // If the slice can be SSA'd, it'll be on the stack,
1624 // so there will be no write barriers,
1625 // so there's no need to attempt to prevent them.
1627 if base.Debug.Append > 0 { // replicating old diagnostic message
1628 base.WarnfAt(n.Pos(), "append: len-only update (in local slice)")
1632 if base.Debug.Append > 0 {
1633 base.WarnfAt(n.Pos(), "append: len-only update")
1640 if ir.IsBlank(n.X) {
1642 // Just evaluate rhs for side-effects.
1657 deref := !ssa.CanSSA(t)
1660 r = nil // Signal assign to use OpZero.
1673 if rhs != nil && (rhs.Op() == ir.OSLICE || rhs.Op() == ir.OSLICE3 || rhs.Op() == ir.OSLICESTR) && ir.SameSafeExpr(rhs.(*ir.SliceExpr).X, n.X) {
1674 // We're assigning a slicing operation back to its source.
1675 // Don't write back fields we aren't changing. See issue #14855.
1676 rhs := rhs.(*ir.SliceExpr)
1677 i, j, k := rhs.Low, rhs.High, rhs.Max
1678 if i != nil && (i.Op() == ir.OLITERAL && i.Val().Kind() == constant.Int && ir.Int64Val(i) == 0) {
1679 // [0:...] is the same as [:...]
1682 // TODO: detect defaults for len/cap also.
1683 // Currently doesn't really work because (*p)[:len(*p)] appears here as:
1686 // if j != nil && (j.Op == OLEN && SameSafeExpr(j.Left, n.Left)) {
1689 // if k != nil && (k.Op == OCAP && SameSafeExpr(k.Left, n.Left)) {
1703 s.assignWhichMayOverlap(n.X, r, deref, skip, mayOverlap)
1707 if ir.IsConst(n.Cond, constant.Bool) {
1708 s.stmtList(n.Cond.Init())
1709 if ir.BoolVal(n.Cond) {
1717 bEnd := s.f.NewBlock(ssa.BlockPlain)
1722 var bThen *ssa.Block
1723 if len(n.Body) != 0 {
1724 bThen = s.f.NewBlock(ssa.BlockPlain)
1728 var bElse *ssa.Block
1729 if len(n.Else) != 0 {
1730 bElse = s.f.NewBlock(ssa.BlockPlain)
1734 s.condBranch(n.Cond, bThen, bElse, likely)
1736 if len(n.Body) != 0 {
1739 if b := s.endBlock(); b != nil {
1743 if len(n.Else) != 0 {
1746 if b := s.endBlock(); b != nil {
1753 n := n.(*ir.ReturnStmt)
1754 s.stmtList(n.Results)
1756 b.Pos = s.lastPos.WithIsStmt()
1759 n := n.(*ir.TailCallStmt)
1760 s.callResult(n.Call, callTail)
1763 b.Kind = ssa.BlockRetJmp // could use BlockExit. BlockRetJmp is mostly for clarity.
1766 case ir.OCONTINUE, ir.OBREAK:
1767 n := n.(*ir.BranchStmt)
1770 // plain break/continue
1778 // labeled break/continue; look up the target
1783 to = lab.continueTarget
1785 to = lab.breakTarget
1790 b.Pos = s.lastPos.WithIsStmt() // Do this even if b is an empty block.
1794 // OFOR: for Ninit; Left; Right { Nbody }
1795 // cond (Left); body (Nbody); incr (Right)
1796 n := n.(*ir.ForStmt)
1797 base.Assert(!n.DistinctVars) // Should all be rewritten before escape analysis
1798 bCond := s.f.NewBlock(ssa.BlockPlain)
1799 bBody := s.f.NewBlock(ssa.BlockPlain)
1800 bIncr := s.f.NewBlock(ssa.BlockPlain)
1801 bEnd := s.f.NewBlock(ssa.BlockPlain)
1803 // ensure empty for loops have correct position; issue #30167
1806 // first, jump to condition test
1810 // generate code to test condition
1813 s.condBranch(n.Cond, bBody, bEnd, 1)
1816 b.Kind = ssa.BlockPlain
1820 // set up for continue/break in body
1821 prevContinue := s.continueTo
1822 prevBreak := s.breakTo
1823 s.continueTo = bIncr
1826 if sym := n.Label; sym != nil {
1829 lab.continueTarget = bIncr
1830 lab.breakTarget = bEnd
1837 // tear down continue/break
1838 s.continueTo = prevContinue
1839 s.breakTo = prevBreak
1841 lab.continueTarget = nil
1842 lab.breakTarget = nil
1845 // done with body, goto incr
1846 if b := s.endBlock(); b != nil {
1855 if b := s.endBlock(); b != nil {
1857 // It can happen that bIncr ends in a block containing only VARKILL,
1858 // and that muddles the debugging experience.
1859 if b.Pos == src.NoXPos {
1866 case ir.OSWITCH, ir.OSELECT:
1867 // These have been mostly rewritten by the front end into their Nbody fields.
1868 // Our main task is to correctly hook up any break statements.
1869 bEnd := s.f.NewBlock(ssa.BlockPlain)
1871 prevBreak := s.breakTo
1875 if n.Op() == ir.OSWITCH {
1876 n := n.(*ir.SwitchStmt)
1880 n := n.(*ir.SelectStmt)
1889 lab.breakTarget = bEnd
1892 // generate body code
1895 s.breakTo = prevBreak
1897 lab.breakTarget = nil
1900 // walk adds explicit OBREAK nodes to the end of all reachable code paths.
1901 // If we still have a current block here, then mark it unreachable.
1902 if s.curBlock != nil {
1905 b.Kind = ssa.BlockExit
1911 n := n.(*ir.JumpTableStmt)
1913 // Make blocks we'll need.
1914 jt := s.f.NewBlock(ssa.BlockJumpTable)
1915 bEnd := s.f.NewBlock(ssa.BlockPlain)
1917 // The only thing that needs evaluating is the index we're looking up.
1918 idx := s.expr(n.Idx)
1919 unsigned := idx.Type.IsUnsigned()
1921 // Extend so we can do everything in uintptr arithmetic.
1922 t := types.Types[types.TUINTPTR]
1923 idx = s.conv(nil, idx, idx.Type, t)
1925 // The ending condition for the current block decides whether we'll use
1926 // the jump table at all.
1927 // We check that min <= idx <= max and jump around the jump table
1928 // if that test fails.
1929 // We implement min <= idx <= max with 0 <= idx-min <= max-min, because
1930 // we'll need idx-min anyway as the control value for the jump table.
1933 min, _ = constant.Uint64Val(n.Cases[0])
1934 max, _ = constant.Uint64Val(n.Cases[len(n.Cases)-1])
1936 mn, _ := constant.Int64Val(n.Cases[0])
1937 mx, _ := constant.Int64Val(n.Cases[len(n.Cases)-1])
1941 // Compare idx-min with max-min, to see if we can use the jump table.
1942 idx = s.newValue2(s.ssaOp(ir.OSUB, t), t, idx, s.uintptrConstant(min))
1943 width := s.uintptrConstant(max - min)
1944 cmp := s.newValue2(s.ssaOp(ir.OLE, t), types.Types[types.TBOOL], idx, width)
1946 b.Kind = ssa.BlockIf
1948 b.AddEdgeTo(jt) // in range - use jump table
1949 b.AddEdgeTo(bEnd) // out of range - no case in the jump table will trigger
1950 b.Likely = ssa.BranchLikely // TODO: assumes missing the table entirely is unlikely. True?
1952 // Build jump table block.
1955 if base.Flag.Cfg.SpectreIndex {
1956 idx = s.newValue2(ssa.OpSpectreSliceIndex, t, idx, width)
1960 // Figure out where we should go for each index in the table.
1961 table := make([]*ssa.Block, max-min+1)
1962 for i := range table {
1963 table[i] = bEnd // default target
1965 for i := range n.Targets {
1967 lab := s.label(n.Targets[i])
1968 if lab.target == nil {
1969 lab.target = s.f.NewBlock(ssa.BlockPlain)
1973 val, _ = constant.Uint64Val(c)
1975 vl, _ := constant.Int64Val(c)
1978 // Overwrite the default target.
1979 table[val-min] = lab.target
1981 for _, t := range table {
1989 n := n.(*ir.UnaryExpr)
1994 n := n.(*ir.InlineMarkStmt)
1995 s.newValue1I(ssa.OpInlMark, types.TypeVoid, n.Index, s.mem())
1998 s.Fatalf("unhandled stmt %v", n.Op())
2002 // If true, share as many open-coded defer exits as possible (with the downside of
2003 // worse line-number information)
2004 const shareDeferExits = false
2006 // exit processes any code that needs to be generated just before returning.
2007 // It returns a BlockRet block that ends the control flow. Its control value
2008 // will be set to the final memory state.
2009 func (s *state) exit() *ssa.Block {
2011 if s.hasOpenDefers {
2012 if shareDeferExits && s.lastDeferExit != nil && len(s.openDefers) == s.lastDeferCount {
2013 if s.curBlock.Kind != ssa.BlockPlain {
2014 panic("Block for an exit should be BlockPlain")
2016 s.curBlock.AddEdgeTo(s.lastDeferExit)
2018 return s.lastDeferFinalBlock
2022 s.rtcall(ir.Syms.Deferreturn, true, nil)
2028 // Do actual return.
2029 // These currently turn into self-copies (in many cases).
2030 resultFields := s.curfn.Type().Results()
2031 results := make([]*ssa.Value, len(resultFields)+1, len(resultFields)+1)
2032 m = s.newValue0(ssa.OpMakeResult, s.f.OwnAux.LateExpansionResultType())
2033 // Store SSAable and heap-escaped PPARAMOUT variables back to stack locations.
2034 for i, f := range resultFields {
2035 n := f.Nname.(*ir.Name)
2036 if s.canSSA(n) { // result is in some SSA variable
2037 if !n.IsOutputParamInRegisters() && n.Type().HasPointers() {
2038 // We are about to store to the result slot.
2039 s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, n, s.mem())
2041 results[i] = s.variable(n, n.Type())
2042 } else if !n.OnStack() { // result is actually heap allocated
2043 // We are about to copy the in-heap result to the result slot.
2044 if n.Type().HasPointers() {
2045 s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, n, s.mem())
2047 ha := s.expr(n.Heapaddr)
2048 s.instrumentFields(n.Type(), ha, instrumentRead)
2049 results[i] = s.newValue2(ssa.OpDereference, n.Type(), ha, s.mem())
2050 } else { // result is not SSA-able; not escaped, so not on heap, but too large for SSA.
2051 // Before register ABI this ought to be a self-move, home=dest,
2052 // With register ABI, it's still a self-move if parameter is on stack (i.e., too big or overflowed)
2053 // No VarDef, as the result slot is already holding live value.
2054 results[i] = s.newValue2(ssa.OpDereference, n.Type(), s.addr(n), s.mem())
2058 // Run exit code. Today, this is just racefuncexit, in -race mode.
2059 // TODO(register args) this seems risky here with a register-ABI, but not clear it is right to do it earlier either.
2060 // Spills in register allocation might just fix it.
2061 s.stmtList(s.curfn.Exit)
2063 results[len(results)-1] = s.mem()
2064 m.AddArgs(results...)
2067 b.Kind = ssa.BlockRet
2069 if s.hasdefer && s.hasOpenDefers {
2070 s.lastDeferFinalBlock = b
2075 type opAndType struct {
2080 var opToSSA = map[opAndType]ssa.Op{
2081 {ir.OADD, types.TINT8}: ssa.OpAdd8,
2082 {ir.OADD, types.TUINT8}: ssa.OpAdd8,
2083 {ir.OADD, types.TINT16}: ssa.OpAdd16,
2084 {ir.OADD, types.TUINT16}: ssa.OpAdd16,
2085 {ir.OADD, types.TINT32}: ssa.OpAdd32,
2086 {ir.OADD, types.TUINT32}: ssa.OpAdd32,
2087 {ir.OADD, types.TINT64}: ssa.OpAdd64,
2088 {ir.OADD, types.TUINT64}: ssa.OpAdd64,
2089 {ir.OADD, types.TFLOAT32}: ssa.OpAdd32F,
2090 {ir.OADD, types.TFLOAT64}: ssa.OpAdd64F,
2092 {ir.OSUB, types.TINT8}: ssa.OpSub8,
2093 {ir.OSUB, types.TUINT8}: ssa.OpSub8,
2094 {ir.OSUB, types.TINT16}: ssa.OpSub16,
2095 {ir.OSUB, types.TUINT16}: ssa.OpSub16,
2096 {ir.OSUB, types.TINT32}: ssa.OpSub32,
2097 {ir.OSUB, types.TUINT32}: ssa.OpSub32,
2098 {ir.OSUB, types.TINT64}: ssa.OpSub64,
2099 {ir.OSUB, types.TUINT64}: ssa.OpSub64,
2100 {ir.OSUB, types.TFLOAT32}: ssa.OpSub32F,
2101 {ir.OSUB, types.TFLOAT64}: ssa.OpSub64F,
2103 {ir.ONOT, types.TBOOL}: ssa.OpNot,
2105 {ir.ONEG, types.TINT8}: ssa.OpNeg8,
2106 {ir.ONEG, types.TUINT8}: ssa.OpNeg8,
2107 {ir.ONEG, types.TINT16}: ssa.OpNeg16,
2108 {ir.ONEG, types.TUINT16}: ssa.OpNeg16,
2109 {ir.ONEG, types.TINT32}: ssa.OpNeg32,
2110 {ir.ONEG, types.TUINT32}: ssa.OpNeg32,
2111 {ir.ONEG, types.TINT64}: ssa.OpNeg64,
2112 {ir.ONEG, types.TUINT64}: ssa.OpNeg64,
2113 {ir.ONEG, types.TFLOAT32}: ssa.OpNeg32F,
2114 {ir.ONEG, types.TFLOAT64}: ssa.OpNeg64F,
2116 {ir.OBITNOT, types.TINT8}: ssa.OpCom8,
2117 {ir.OBITNOT, types.TUINT8}: ssa.OpCom8,
2118 {ir.OBITNOT, types.TINT16}: ssa.OpCom16,
2119 {ir.OBITNOT, types.TUINT16}: ssa.OpCom16,
2120 {ir.OBITNOT, types.TINT32}: ssa.OpCom32,
2121 {ir.OBITNOT, types.TUINT32}: ssa.OpCom32,
2122 {ir.OBITNOT, types.TINT64}: ssa.OpCom64,
2123 {ir.OBITNOT, types.TUINT64}: ssa.OpCom64,
2125 {ir.OIMAG, types.TCOMPLEX64}: ssa.OpComplexImag,
2126 {ir.OIMAG, types.TCOMPLEX128}: ssa.OpComplexImag,
2127 {ir.OREAL, types.TCOMPLEX64}: ssa.OpComplexReal,
2128 {ir.OREAL, types.TCOMPLEX128}: ssa.OpComplexReal,
2130 {ir.OMUL, types.TINT8}: ssa.OpMul8,
2131 {ir.OMUL, types.TUINT8}: ssa.OpMul8,
2132 {ir.OMUL, types.TINT16}: ssa.OpMul16,
2133 {ir.OMUL, types.TUINT16}: ssa.OpMul16,
2134 {ir.OMUL, types.TINT32}: ssa.OpMul32,
2135 {ir.OMUL, types.TUINT32}: ssa.OpMul32,
2136 {ir.OMUL, types.TINT64}: ssa.OpMul64,
2137 {ir.OMUL, types.TUINT64}: ssa.OpMul64,
2138 {ir.OMUL, types.TFLOAT32}: ssa.OpMul32F,
2139 {ir.OMUL, types.TFLOAT64}: ssa.OpMul64F,
2141 {ir.ODIV, types.TFLOAT32}: ssa.OpDiv32F,
2142 {ir.ODIV, types.TFLOAT64}: ssa.OpDiv64F,
2144 {ir.ODIV, types.TINT8}: ssa.OpDiv8,
2145 {ir.ODIV, types.TUINT8}: ssa.OpDiv8u,
2146 {ir.ODIV, types.TINT16}: ssa.OpDiv16,
2147 {ir.ODIV, types.TUINT16}: ssa.OpDiv16u,
2148 {ir.ODIV, types.TINT32}: ssa.OpDiv32,
2149 {ir.ODIV, types.TUINT32}: ssa.OpDiv32u,
2150 {ir.ODIV, types.TINT64}: ssa.OpDiv64,
2151 {ir.ODIV, types.TUINT64}: ssa.OpDiv64u,
2153 {ir.OMOD, types.TINT8}: ssa.OpMod8,
2154 {ir.OMOD, types.TUINT8}: ssa.OpMod8u,
2155 {ir.OMOD, types.TINT16}: ssa.OpMod16,
2156 {ir.OMOD, types.TUINT16}: ssa.OpMod16u,
2157 {ir.OMOD, types.TINT32}: ssa.OpMod32,
2158 {ir.OMOD, types.TUINT32}: ssa.OpMod32u,
2159 {ir.OMOD, types.TINT64}: ssa.OpMod64,
2160 {ir.OMOD, types.TUINT64}: ssa.OpMod64u,
2162 {ir.OAND, types.TINT8}: ssa.OpAnd8,
2163 {ir.OAND, types.TUINT8}: ssa.OpAnd8,
2164 {ir.OAND, types.TINT16}: ssa.OpAnd16,
2165 {ir.OAND, types.TUINT16}: ssa.OpAnd16,
2166 {ir.OAND, types.TINT32}: ssa.OpAnd32,
2167 {ir.OAND, types.TUINT32}: ssa.OpAnd32,
2168 {ir.OAND, types.TINT64}: ssa.OpAnd64,
2169 {ir.OAND, types.TUINT64}: ssa.OpAnd64,
2171 {ir.OOR, types.TINT8}: ssa.OpOr8,
2172 {ir.OOR, types.TUINT8}: ssa.OpOr8,
2173 {ir.OOR, types.TINT16}: ssa.OpOr16,
2174 {ir.OOR, types.TUINT16}: ssa.OpOr16,
2175 {ir.OOR, types.TINT32}: ssa.OpOr32,
2176 {ir.OOR, types.TUINT32}: ssa.OpOr32,
2177 {ir.OOR, types.TINT64}: ssa.OpOr64,
2178 {ir.OOR, types.TUINT64}: ssa.OpOr64,
2180 {ir.OXOR, types.TINT8}: ssa.OpXor8,
2181 {ir.OXOR, types.TUINT8}: ssa.OpXor8,
2182 {ir.OXOR, types.TINT16}: ssa.OpXor16,
2183 {ir.OXOR, types.TUINT16}: ssa.OpXor16,
2184 {ir.OXOR, types.TINT32}: ssa.OpXor32,
2185 {ir.OXOR, types.TUINT32}: ssa.OpXor32,
2186 {ir.OXOR, types.TINT64}: ssa.OpXor64,
2187 {ir.OXOR, types.TUINT64}: ssa.OpXor64,
2189 {ir.OEQ, types.TBOOL}: ssa.OpEqB,
2190 {ir.OEQ, types.TINT8}: ssa.OpEq8,
2191 {ir.OEQ, types.TUINT8}: ssa.OpEq8,
2192 {ir.OEQ, types.TINT16}: ssa.OpEq16,
2193 {ir.OEQ, types.TUINT16}: ssa.OpEq16,
2194 {ir.OEQ, types.TINT32}: ssa.OpEq32,
2195 {ir.OEQ, types.TUINT32}: ssa.OpEq32,
2196 {ir.OEQ, types.TINT64}: ssa.OpEq64,
2197 {ir.OEQ, types.TUINT64}: ssa.OpEq64,
2198 {ir.OEQ, types.TINTER}: ssa.OpEqInter,
2199 {ir.OEQ, types.TSLICE}: ssa.OpEqSlice,
2200 {ir.OEQ, types.TFUNC}: ssa.OpEqPtr,
2201 {ir.OEQ, types.TMAP}: ssa.OpEqPtr,
2202 {ir.OEQ, types.TCHAN}: ssa.OpEqPtr,
2203 {ir.OEQ, types.TPTR}: ssa.OpEqPtr,
2204 {ir.OEQ, types.TUINTPTR}: ssa.OpEqPtr,
2205 {ir.OEQ, types.TUNSAFEPTR}: ssa.OpEqPtr,
2206 {ir.OEQ, types.TFLOAT64}: ssa.OpEq64F,
2207 {ir.OEQ, types.TFLOAT32}: ssa.OpEq32F,
2209 {ir.ONE, types.TBOOL}: ssa.OpNeqB,
2210 {ir.ONE, types.TINT8}: ssa.OpNeq8,
2211 {ir.ONE, types.TUINT8}: ssa.OpNeq8,
2212 {ir.ONE, types.TINT16}: ssa.OpNeq16,
2213 {ir.ONE, types.TUINT16}: ssa.OpNeq16,
2214 {ir.ONE, types.TINT32}: ssa.OpNeq32,
2215 {ir.ONE, types.TUINT32}: ssa.OpNeq32,
2216 {ir.ONE, types.TINT64}: ssa.OpNeq64,
2217 {ir.ONE, types.TUINT64}: ssa.OpNeq64,
2218 {ir.ONE, types.TINTER}: ssa.OpNeqInter,
2219 {ir.ONE, types.TSLICE}: ssa.OpNeqSlice,
2220 {ir.ONE, types.TFUNC}: ssa.OpNeqPtr,
2221 {ir.ONE, types.TMAP}: ssa.OpNeqPtr,
2222 {ir.ONE, types.TCHAN}: ssa.OpNeqPtr,
2223 {ir.ONE, types.TPTR}: ssa.OpNeqPtr,
2224 {ir.ONE, types.TUINTPTR}: ssa.OpNeqPtr,
2225 {ir.ONE, types.TUNSAFEPTR}: ssa.OpNeqPtr,
2226 {ir.ONE, types.TFLOAT64}: ssa.OpNeq64F,
2227 {ir.ONE, types.TFLOAT32}: ssa.OpNeq32F,
2229 {ir.OLT, types.TINT8}: ssa.OpLess8,
2230 {ir.OLT, types.TUINT8}: ssa.OpLess8U,
2231 {ir.OLT, types.TINT16}: ssa.OpLess16,
2232 {ir.OLT, types.TUINT16}: ssa.OpLess16U,
2233 {ir.OLT, types.TINT32}: ssa.OpLess32,
2234 {ir.OLT, types.TUINT32}: ssa.OpLess32U,
2235 {ir.OLT, types.TINT64}: ssa.OpLess64,
2236 {ir.OLT, types.TUINT64}: ssa.OpLess64U,
2237 {ir.OLT, types.TFLOAT64}: ssa.OpLess64F,
2238 {ir.OLT, types.TFLOAT32}: ssa.OpLess32F,
2240 {ir.OLE, types.TINT8}: ssa.OpLeq8,
2241 {ir.OLE, types.TUINT8}: ssa.OpLeq8U,
2242 {ir.OLE, types.TINT16}: ssa.OpLeq16,
2243 {ir.OLE, types.TUINT16}: ssa.OpLeq16U,
2244 {ir.OLE, types.TINT32}: ssa.OpLeq32,
2245 {ir.OLE, types.TUINT32}: ssa.OpLeq32U,
2246 {ir.OLE, types.TINT64}: ssa.OpLeq64,
2247 {ir.OLE, types.TUINT64}: ssa.OpLeq64U,
2248 {ir.OLE, types.TFLOAT64}: ssa.OpLeq64F,
2249 {ir.OLE, types.TFLOAT32}: ssa.OpLeq32F,
2252 func (s *state) concreteEtype(t *types.Type) types.Kind {
2258 if s.config.PtrSize == 8 {
2263 if s.config.PtrSize == 8 {
2264 return types.TUINT64
2266 return types.TUINT32
2267 case types.TUINTPTR:
2268 if s.config.PtrSize == 8 {
2269 return types.TUINT64
2271 return types.TUINT32
2275 func (s *state) ssaOp(op ir.Op, t *types.Type) ssa.Op {
2276 etype := s.concreteEtype(t)
2277 x, ok := opToSSA[opAndType{op, etype}]
2279 s.Fatalf("unhandled binary op %v %s", op, etype)
2284 type opAndTwoTypes struct {
2290 type twoTypes struct {
2295 type twoOpsAndType struct {
2298 intermediateType types.Kind
2301 var fpConvOpToSSA = map[twoTypes]twoOpsAndType{
2303 {types.TINT8, types.TFLOAT32}: {ssa.OpSignExt8to32, ssa.OpCvt32to32F, types.TINT32},
2304 {types.TINT16, types.TFLOAT32}: {ssa.OpSignExt16to32, ssa.OpCvt32to32F, types.TINT32},
2305 {types.TINT32, types.TFLOAT32}: {ssa.OpCopy, ssa.OpCvt32to32F, types.TINT32},
2306 {types.TINT64, types.TFLOAT32}: {ssa.OpCopy, ssa.OpCvt64to32F, types.TINT64},
2308 {types.TINT8, types.TFLOAT64}: {ssa.OpSignExt8to32, ssa.OpCvt32to64F, types.TINT32},
2309 {types.TINT16, types.TFLOAT64}: {ssa.OpSignExt16to32, ssa.OpCvt32to64F, types.TINT32},
2310 {types.TINT32, types.TFLOAT64}: {ssa.OpCopy, ssa.OpCvt32to64F, types.TINT32},
2311 {types.TINT64, types.TFLOAT64}: {ssa.OpCopy, ssa.OpCvt64to64F, types.TINT64},
2313 {types.TFLOAT32, types.TINT8}: {ssa.OpCvt32Fto32, ssa.OpTrunc32to8, types.TINT32},
2314 {types.TFLOAT32, types.TINT16}: {ssa.OpCvt32Fto32, ssa.OpTrunc32to16, types.TINT32},
2315 {types.TFLOAT32, types.TINT32}: {ssa.OpCvt32Fto32, ssa.OpCopy, types.TINT32},
2316 {types.TFLOAT32, types.TINT64}: {ssa.OpCvt32Fto64, ssa.OpCopy, types.TINT64},
2318 {types.TFLOAT64, types.TINT8}: {ssa.OpCvt64Fto32, ssa.OpTrunc32to8, types.TINT32},
2319 {types.TFLOAT64, types.TINT16}: {ssa.OpCvt64Fto32, ssa.OpTrunc32to16, types.TINT32},
2320 {types.TFLOAT64, types.TINT32}: {ssa.OpCvt64Fto32, ssa.OpCopy, types.TINT32},
2321 {types.TFLOAT64, types.TINT64}: {ssa.OpCvt64Fto64, ssa.OpCopy, types.TINT64},
2323 {types.TUINT8, types.TFLOAT32}: {ssa.OpZeroExt8to32, ssa.OpCvt32to32F, types.TINT32},
2324 {types.TUINT16, types.TFLOAT32}: {ssa.OpZeroExt16to32, ssa.OpCvt32to32F, types.TINT32},
2325 {types.TUINT32, types.TFLOAT32}: {ssa.OpZeroExt32to64, ssa.OpCvt64to32F, types.TINT64}, // go wide to dodge unsigned
2326 {types.TUINT64, types.TFLOAT32}: {ssa.OpCopy, ssa.OpInvalid, types.TUINT64}, // Cvt64Uto32F, branchy code expansion instead
2328 {types.TUINT8, types.TFLOAT64}: {ssa.OpZeroExt8to32, ssa.OpCvt32to64F, types.TINT32},
2329 {types.TUINT16, types.TFLOAT64}: {ssa.OpZeroExt16to32, ssa.OpCvt32to64F, types.TINT32},
2330 {types.TUINT32, types.TFLOAT64}: {ssa.OpZeroExt32to64, ssa.OpCvt64to64F, types.TINT64}, // go wide to dodge unsigned
2331 {types.TUINT64, types.TFLOAT64}: {ssa.OpCopy, ssa.OpInvalid, types.TUINT64}, // Cvt64Uto64F, branchy code expansion instead
2333 {types.TFLOAT32, types.TUINT8}: {ssa.OpCvt32Fto32, ssa.OpTrunc32to8, types.TINT32},
2334 {types.TFLOAT32, types.TUINT16}: {ssa.OpCvt32Fto32, ssa.OpTrunc32to16, types.TINT32},
2335 {types.TFLOAT32, types.TUINT32}: {ssa.OpCvt32Fto64, ssa.OpTrunc64to32, types.TINT64}, // go wide to dodge unsigned
2336 {types.TFLOAT32, types.TUINT64}: {ssa.OpInvalid, ssa.OpCopy, types.TUINT64}, // Cvt32Fto64U, branchy code expansion instead
2338 {types.TFLOAT64, types.TUINT8}: {ssa.OpCvt64Fto32, ssa.OpTrunc32to8, types.TINT32},
2339 {types.TFLOAT64, types.TUINT16}: {ssa.OpCvt64Fto32, ssa.OpTrunc32to16, types.TINT32},
2340 {types.TFLOAT64, types.TUINT32}: {ssa.OpCvt64Fto64, ssa.OpTrunc64to32, types.TINT64}, // go wide to dodge unsigned
2341 {types.TFLOAT64, types.TUINT64}: {ssa.OpInvalid, ssa.OpCopy, types.TUINT64}, // Cvt64Fto64U, branchy code expansion instead
2344 {types.TFLOAT64, types.TFLOAT32}: {ssa.OpCvt64Fto32F, ssa.OpCopy, types.TFLOAT32},
2345 {types.TFLOAT64, types.TFLOAT64}: {ssa.OpRound64F, ssa.OpCopy, types.TFLOAT64},
2346 {types.TFLOAT32, types.TFLOAT32}: {ssa.OpRound32F, ssa.OpCopy, types.TFLOAT32},
2347 {types.TFLOAT32, types.TFLOAT64}: {ssa.OpCvt32Fto64F, ssa.OpCopy, types.TFLOAT64},
2350 // this map is used only for 32-bit arch, and only includes the difference
2351 // on 32-bit arch, don't use int64<->float conversion for uint32
2352 var fpConvOpToSSA32 = map[twoTypes]twoOpsAndType{
2353 {types.TUINT32, types.TFLOAT32}: {ssa.OpCopy, ssa.OpCvt32Uto32F, types.TUINT32},
2354 {types.TUINT32, types.TFLOAT64}: {ssa.OpCopy, ssa.OpCvt32Uto64F, types.TUINT32},
2355 {types.TFLOAT32, types.TUINT32}: {ssa.OpCvt32Fto32U, ssa.OpCopy, types.TUINT32},
2356 {types.TFLOAT64, types.TUINT32}: {ssa.OpCvt64Fto32U, ssa.OpCopy, types.TUINT32},
2359 // uint64<->float conversions, only on machines that have instructions for that
2360 var uint64fpConvOpToSSA = map[twoTypes]twoOpsAndType{
2361 {types.TUINT64, types.TFLOAT32}: {ssa.OpCopy, ssa.OpCvt64Uto32F, types.TUINT64},
2362 {types.TUINT64, types.TFLOAT64}: {ssa.OpCopy, ssa.OpCvt64Uto64F, types.TUINT64},
2363 {types.TFLOAT32, types.TUINT64}: {ssa.OpCvt32Fto64U, ssa.OpCopy, types.TUINT64},
2364 {types.TFLOAT64, types.TUINT64}: {ssa.OpCvt64Fto64U, ssa.OpCopy, types.TUINT64},
2367 var shiftOpToSSA = map[opAndTwoTypes]ssa.Op{
2368 {ir.OLSH, types.TINT8, types.TUINT8}: ssa.OpLsh8x8,
2369 {ir.OLSH, types.TUINT8, types.TUINT8}: ssa.OpLsh8x8,
2370 {ir.OLSH, types.TINT8, types.TUINT16}: ssa.OpLsh8x16,
2371 {ir.OLSH, types.TUINT8, types.TUINT16}: ssa.OpLsh8x16,
2372 {ir.OLSH, types.TINT8, types.TUINT32}: ssa.OpLsh8x32,
2373 {ir.OLSH, types.TUINT8, types.TUINT32}: ssa.OpLsh8x32,
2374 {ir.OLSH, types.TINT8, types.TUINT64}: ssa.OpLsh8x64,
2375 {ir.OLSH, types.TUINT8, types.TUINT64}: ssa.OpLsh8x64,
2377 {ir.OLSH, types.TINT16, types.TUINT8}: ssa.OpLsh16x8,
2378 {ir.OLSH, types.TUINT16, types.TUINT8}: ssa.OpLsh16x8,
2379 {ir.OLSH, types.TINT16, types.TUINT16}: ssa.OpLsh16x16,
2380 {ir.OLSH, types.TUINT16, types.TUINT16}: ssa.OpLsh16x16,
2381 {ir.OLSH, types.TINT16, types.TUINT32}: ssa.OpLsh16x32,
2382 {ir.OLSH, types.TUINT16, types.TUINT32}: ssa.OpLsh16x32,
2383 {ir.OLSH, types.TINT16, types.TUINT64}: ssa.OpLsh16x64,
2384 {ir.OLSH, types.TUINT16, types.TUINT64}: ssa.OpLsh16x64,
2386 {ir.OLSH, types.TINT32, types.TUINT8}: ssa.OpLsh32x8,
2387 {ir.OLSH, types.TUINT32, types.TUINT8}: ssa.OpLsh32x8,
2388 {ir.OLSH, types.TINT32, types.TUINT16}: ssa.OpLsh32x16,
2389 {ir.OLSH, types.TUINT32, types.TUINT16}: ssa.OpLsh32x16,
2390 {ir.OLSH, types.TINT32, types.TUINT32}: ssa.OpLsh32x32,
2391 {ir.OLSH, types.TUINT32, types.TUINT32}: ssa.OpLsh32x32,
2392 {ir.OLSH, types.TINT32, types.TUINT64}: ssa.OpLsh32x64,
2393 {ir.OLSH, types.TUINT32, types.TUINT64}: ssa.OpLsh32x64,
2395 {ir.OLSH, types.TINT64, types.TUINT8}: ssa.OpLsh64x8,
2396 {ir.OLSH, types.TUINT64, types.TUINT8}: ssa.OpLsh64x8,
2397 {ir.OLSH, types.TINT64, types.TUINT16}: ssa.OpLsh64x16,
2398 {ir.OLSH, types.TUINT64, types.TUINT16}: ssa.OpLsh64x16,
2399 {ir.OLSH, types.TINT64, types.TUINT32}: ssa.OpLsh64x32,
2400 {ir.OLSH, types.TUINT64, types.TUINT32}: ssa.OpLsh64x32,
2401 {ir.OLSH, types.TINT64, types.TUINT64}: ssa.OpLsh64x64,
2402 {ir.OLSH, types.TUINT64, types.TUINT64}: ssa.OpLsh64x64,
2404 {ir.ORSH, types.TINT8, types.TUINT8}: ssa.OpRsh8x8,
2405 {ir.ORSH, types.TUINT8, types.TUINT8}: ssa.OpRsh8Ux8,
2406 {ir.ORSH, types.TINT8, types.TUINT16}: ssa.OpRsh8x16,
2407 {ir.ORSH, types.TUINT8, types.TUINT16}: ssa.OpRsh8Ux16,
2408 {ir.ORSH, types.TINT8, types.TUINT32}: ssa.OpRsh8x32,
2409 {ir.ORSH, types.TUINT8, types.TUINT32}: ssa.OpRsh8Ux32,
2410 {ir.ORSH, types.TINT8, types.TUINT64}: ssa.OpRsh8x64,
2411 {ir.ORSH, types.TUINT8, types.TUINT64}: ssa.OpRsh8Ux64,
2413 {ir.ORSH, types.TINT16, types.TUINT8}: ssa.OpRsh16x8,
2414 {ir.ORSH, types.TUINT16, types.TUINT8}: ssa.OpRsh16Ux8,
2415 {ir.ORSH, types.TINT16, types.TUINT16}: ssa.OpRsh16x16,
2416 {ir.ORSH, types.TUINT16, types.TUINT16}: ssa.OpRsh16Ux16,
2417 {ir.ORSH, types.TINT16, types.TUINT32}: ssa.OpRsh16x32,
2418 {ir.ORSH, types.TUINT16, types.TUINT32}: ssa.OpRsh16Ux32,
2419 {ir.ORSH, types.TINT16, types.TUINT64}: ssa.OpRsh16x64,
2420 {ir.ORSH, types.TUINT16, types.TUINT64}: ssa.OpRsh16Ux64,
2422 {ir.ORSH, types.TINT32, types.TUINT8}: ssa.OpRsh32x8,
2423 {ir.ORSH, types.TUINT32, types.TUINT8}: ssa.OpRsh32Ux8,
2424 {ir.ORSH, types.TINT32, types.TUINT16}: ssa.OpRsh32x16,
2425 {ir.ORSH, types.TUINT32, types.TUINT16}: ssa.OpRsh32Ux16,
2426 {ir.ORSH, types.TINT32, types.TUINT32}: ssa.OpRsh32x32,
2427 {ir.ORSH, types.TUINT32, types.TUINT32}: ssa.OpRsh32Ux32,
2428 {ir.ORSH, types.TINT32, types.TUINT64}: ssa.OpRsh32x64,
2429 {ir.ORSH, types.TUINT32, types.TUINT64}: ssa.OpRsh32Ux64,
2431 {ir.ORSH, types.TINT64, types.TUINT8}: ssa.OpRsh64x8,
2432 {ir.ORSH, types.TUINT64, types.TUINT8}: ssa.OpRsh64Ux8,
2433 {ir.ORSH, types.TINT64, types.TUINT16}: ssa.OpRsh64x16,
2434 {ir.ORSH, types.TUINT64, types.TUINT16}: ssa.OpRsh64Ux16,
2435 {ir.ORSH, types.TINT64, types.TUINT32}: ssa.OpRsh64x32,
2436 {ir.ORSH, types.TUINT64, types.TUINT32}: ssa.OpRsh64Ux32,
2437 {ir.ORSH, types.TINT64, types.TUINT64}: ssa.OpRsh64x64,
2438 {ir.ORSH, types.TUINT64, types.TUINT64}: ssa.OpRsh64Ux64,
2441 func (s *state) ssaShiftOp(op ir.Op, t *types.Type, u *types.Type) ssa.Op {
2442 etype1 := s.concreteEtype(t)
2443 etype2 := s.concreteEtype(u)
2444 x, ok := shiftOpToSSA[opAndTwoTypes{op, etype1, etype2}]
2446 s.Fatalf("unhandled shift op %v etype=%s/%s", op, etype1, etype2)
2451 func (s *state) uintptrConstant(v uint64) *ssa.Value {
2452 if s.config.PtrSize == 4 {
2453 return s.newValue0I(ssa.OpConst32, types.Types[types.TUINTPTR], int64(v))
2455 return s.newValue0I(ssa.OpConst64, types.Types[types.TUINTPTR], int64(v))
2458 func (s *state) conv(n ir.Node, v *ssa.Value, ft, tt *types.Type) *ssa.Value {
2459 if ft.IsBoolean() && tt.IsKind(types.TUINT8) {
2460 // Bool -> uint8 is generated internally when indexing into runtime.staticbyte.
2461 return s.newValue1(ssa.OpCvtBoolToUint8, tt, v)
2463 if ft.IsInteger() && tt.IsInteger() {
2465 if tt.Size() == ft.Size() {
2467 } else if tt.Size() < ft.Size() {
2469 switch 10*ft.Size() + tt.Size() {
2471 op = ssa.OpTrunc16to8
2473 op = ssa.OpTrunc32to8
2475 op = ssa.OpTrunc32to16
2477 op = ssa.OpTrunc64to8
2479 op = ssa.OpTrunc64to16
2481 op = ssa.OpTrunc64to32
2483 s.Fatalf("weird integer truncation %v -> %v", ft, tt)
2485 } else if ft.IsSigned() {
2487 switch 10*ft.Size() + tt.Size() {
2489 op = ssa.OpSignExt8to16
2491 op = ssa.OpSignExt8to32
2493 op = ssa.OpSignExt8to64
2495 op = ssa.OpSignExt16to32
2497 op = ssa.OpSignExt16to64
2499 op = ssa.OpSignExt32to64
2501 s.Fatalf("bad integer sign extension %v -> %v", ft, tt)
2505 switch 10*ft.Size() + tt.Size() {
2507 op = ssa.OpZeroExt8to16
2509 op = ssa.OpZeroExt8to32
2511 op = ssa.OpZeroExt8to64
2513 op = ssa.OpZeroExt16to32
2515 op = ssa.OpZeroExt16to64
2517 op = ssa.OpZeroExt32to64
2519 s.Fatalf("weird integer sign extension %v -> %v", ft, tt)
2522 return s.newValue1(op, tt, v)
2525 if ft.IsComplex() && tt.IsComplex() {
2527 if ft.Size() == tt.Size() {
2534 s.Fatalf("weird complex conversion %v -> %v", ft, tt)
2536 } else if ft.Size() == 8 && tt.Size() == 16 {
2537 op = ssa.OpCvt32Fto64F
2538 } else if ft.Size() == 16 && tt.Size() == 8 {
2539 op = ssa.OpCvt64Fto32F
2541 s.Fatalf("weird complex conversion %v -> %v", ft, tt)
2543 ftp := types.FloatForComplex(ft)
2544 ttp := types.FloatForComplex(tt)
2545 return s.newValue2(ssa.OpComplexMake, tt,
2546 s.newValueOrSfCall1(op, ttp, s.newValue1(ssa.OpComplexReal, ftp, v)),
2547 s.newValueOrSfCall1(op, ttp, s.newValue1(ssa.OpComplexImag, ftp, v)))
2550 if tt.IsComplex() { // and ft is not complex
2551 // Needed for generics support - can't happen in normal Go code.
2552 et := types.FloatForComplex(tt)
2553 v = s.conv(n, v, ft, et)
2554 return s.newValue2(ssa.OpComplexMake, tt, v, s.zeroVal(et))
2557 if ft.IsFloat() || tt.IsFloat() {
2558 conv, ok := fpConvOpToSSA[twoTypes{s.concreteEtype(ft), s.concreteEtype(tt)}]
2559 if s.config.RegSize == 4 && Arch.LinkArch.Family != sys.MIPS && !s.softFloat {
2560 if conv1, ok1 := fpConvOpToSSA32[twoTypes{s.concreteEtype(ft), s.concreteEtype(tt)}]; ok1 {
2564 if Arch.LinkArch.Family == sys.ARM64 || Arch.LinkArch.Family == sys.Wasm || Arch.LinkArch.Family == sys.S390X || s.softFloat {
2565 if conv1, ok1 := uint64fpConvOpToSSA[twoTypes{s.concreteEtype(ft), s.concreteEtype(tt)}]; ok1 {
2570 if Arch.LinkArch.Family == sys.MIPS && !s.softFloat {
2571 if ft.Size() == 4 && ft.IsInteger() && !ft.IsSigned() {
2572 // tt is float32 or float64, and ft is also unsigned
2574 return s.uint32Tofloat32(n, v, ft, tt)
2577 return s.uint32Tofloat64(n, v, ft, tt)
2579 } else if tt.Size() == 4 && tt.IsInteger() && !tt.IsSigned() {
2580 // ft is float32 or float64, and tt is unsigned integer
2582 return s.float32ToUint32(n, v, ft, tt)
2585 return s.float64ToUint32(n, v, ft, tt)
2591 s.Fatalf("weird float conversion %v -> %v", ft, tt)
2593 op1, op2, it := conv.op1, conv.op2, conv.intermediateType
2595 if op1 != ssa.OpInvalid && op2 != ssa.OpInvalid {
2596 // normal case, not tripping over unsigned 64
2597 if op1 == ssa.OpCopy {
2598 if op2 == ssa.OpCopy {
2601 return s.newValueOrSfCall1(op2, tt, v)
2603 if op2 == ssa.OpCopy {
2604 return s.newValueOrSfCall1(op1, tt, v)
2606 return s.newValueOrSfCall1(op2, tt, s.newValueOrSfCall1(op1, types.Types[it], v))
2608 // Tricky 64-bit unsigned cases.
2610 // tt is float32 or float64, and ft is also unsigned
2612 return s.uint64Tofloat32(n, v, ft, tt)
2615 return s.uint64Tofloat64(n, v, ft, tt)
2617 s.Fatalf("weird unsigned integer to float conversion %v -> %v", ft, tt)
2619 // ft is float32 or float64, and tt is unsigned integer
2621 return s.float32ToUint64(n, v, ft, tt)
2624 return s.float64ToUint64(n, v, ft, tt)
2626 s.Fatalf("weird float to unsigned integer conversion %v -> %v", ft, tt)
2630 s.Fatalf("unhandled OCONV %s -> %s", ft.Kind(), tt.Kind())
2634 // expr converts the expression n to ssa, adds it to s and returns the ssa result.
2635 func (s *state) expr(n ir.Node) *ssa.Value {
2636 return s.exprCheckPtr(n, true)
2639 func (s *state) exprCheckPtr(n ir.Node, checkPtrOK bool) *ssa.Value {
2640 if ir.HasUniquePos(n) {
2641 // ONAMEs and named OLITERALs have the line number
2642 // of the decl, not the use. See issue 14742.
2647 s.stmtList(n.Init())
2649 case ir.OBYTES2STRTMP:
2650 n := n.(*ir.ConvExpr)
2651 slice := s.expr(n.X)
2652 ptr := s.newValue1(ssa.OpSlicePtr, s.f.Config.Types.BytePtr, slice)
2653 len := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], slice)
2654 return s.newValue2(ssa.OpStringMake, n.Type(), ptr, len)
2655 case ir.OSTR2BYTESTMP:
2656 n := n.(*ir.ConvExpr)
2658 ptr := s.newValue1(ssa.OpStringPtr, s.f.Config.Types.BytePtr, str)
2660 // We need to ensure []byte("") evaluates to []byte{}, and not []byte(nil).
2662 // TODO(mdempsky): Investigate using "len != 0" instead of "ptr != nil".
2663 cond := s.newValue2(ssa.OpNeqPtr, types.Types[types.TBOOL], ptr, s.constNil(ptr.Type))
2664 zerobase := s.newValue1A(ssa.OpAddr, ptr.Type, ir.Syms.Zerobase, s.sb)
2665 ptr = s.ternary(cond, ptr, zerobase)
2667 len := s.newValue1(ssa.OpStringLen, types.Types[types.TINT], str)
2668 return s.newValue3(ssa.OpSliceMake, n.Type(), ptr, len, len)
2670 n := n.(*ir.UnaryExpr)
2671 aux := n.X.(*ir.Name).Linksym()
2672 // OCFUNC is used to build function values, which must
2673 // always reference ABIInternal entry points.
2674 if aux.ABI() != obj.ABIInternal {
2675 s.Fatalf("expected ABIInternal: %v", aux.ABI())
2677 return s.entryNewValue1A(ssa.OpAddr, n.Type(), aux, s.sb)
2680 if n.Class == ir.PFUNC {
2681 // "value" of a function is the address of the function's closure
2682 sym := staticdata.FuncLinksym(n)
2683 return s.entryNewValue1A(ssa.OpAddr, types.NewPtr(n.Type()), sym, s.sb)
2686 return s.variable(n, n.Type())
2688 return s.load(n.Type(), s.addr(n))
2689 case ir.OLINKSYMOFFSET:
2690 n := n.(*ir.LinksymOffsetExpr)
2691 return s.load(n.Type(), s.addr(n))
2693 n := n.(*ir.NilExpr)
2697 return s.constSlice(t)
2698 case t.IsInterface():
2699 return s.constInterface(t)
2701 return s.constNil(t)
2704 switch u := n.Val(); u.Kind() {
2706 i := ir.IntVal(n.Type(), u)
2707 switch n.Type().Size() {
2709 return s.constInt8(n.Type(), int8(i))
2711 return s.constInt16(n.Type(), int16(i))
2713 return s.constInt32(n.Type(), int32(i))
2715 return s.constInt64(n.Type(), i)
2717 s.Fatalf("bad integer size %d", n.Type().Size())
2720 case constant.String:
2721 i := constant.StringVal(u)
2723 return s.constEmptyString(n.Type())
2725 return s.entryNewValue0A(ssa.OpConstString, n.Type(), ssa.StringToAux(i))
2727 return s.constBool(constant.BoolVal(u))
2728 case constant.Float:
2729 f, _ := constant.Float64Val(u)
2730 switch n.Type().Size() {
2732 return s.constFloat32(n.Type(), f)
2734 return s.constFloat64(n.Type(), f)
2736 s.Fatalf("bad float size %d", n.Type().Size())
2739 case constant.Complex:
2740 re, _ := constant.Float64Val(constant.Real(u))
2741 im, _ := constant.Float64Val(constant.Imag(u))
2742 switch n.Type().Size() {
2744 pt := types.Types[types.TFLOAT32]
2745 return s.newValue2(ssa.OpComplexMake, n.Type(),
2746 s.constFloat32(pt, re),
2747 s.constFloat32(pt, im))
2749 pt := types.Types[types.TFLOAT64]
2750 return s.newValue2(ssa.OpComplexMake, n.Type(),
2751 s.constFloat64(pt, re),
2752 s.constFloat64(pt, im))
2754 s.Fatalf("bad complex size %d", n.Type().Size())
2758 s.Fatalf("unhandled OLITERAL %v", u.Kind())
2762 n := n.(*ir.ConvExpr)
2766 // Assume everything will work out, so set up our return value.
2767 // Anything interesting that happens from here is a fatal.
2773 // Special case for not confusing GC and liveness.
2774 // We don't want pointers accidentally classified
2775 // as not-pointers or vice-versa because of copy
2777 if to.IsPtrShaped() != from.IsPtrShaped() {
2778 return s.newValue2(ssa.OpConvert, to, x, s.mem())
2781 v := s.newValue1(ssa.OpCopy, to, x) // ensure that v has the right type
2784 if to.Kind() == types.TFUNC && from.IsPtrShaped() {
2788 // named <--> unnamed type or typed <--> untyped const
2789 if from.Kind() == to.Kind() {
2793 // unsafe.Pointer <--> *T
2794 if to.IsUnsafePtr() && from.IsPtrShaped() || from.IsUnsafePtr() && to.IsPtrShaped() {
2795 if s.checkPtrEnabled && checkPtrOK && to.IsPtr() && from.IsUnsafePtr() {
2796 s.checkPtrAlignment(n, v, nil)
2802 if to.Kind() == types.TMAP && from == types.NewPtr(reflectdata.MapType()) {
2806 types.CalcSize(from)
2808 if from.Size() != to.Size() {
2809 s.Fatalf("CONVNOP width mismatch %v (%d) -> %v (%d)\n", from, from.Size(), to, to.Size())
2812 if etypesign(from.Kind()) != etypesign(to.Kind()) {
2813 s.Fatalf("CONVNOP sign mismatch %v (%s) -> %v (%s)\n", from, from.Kind(), to, to.Kind())
2817 if base.Flag.Cfg.Instrumenting {
2818 // These appear to be fine, but they fail the
2819 // integer constraint below, so okay them here.
2820 // Sample non-integer conversion: map[string]string -> *uint8
2824 if etypesign(from.Kind()) == 0 {
2825 s.Fatalf("CONVNOP unrecognized non-integer %v -> %v\n", from, to)
2829 // integer, same width, same sign
2833 n := n.(*ir.ConvExpr)
2835 return s.conv(n, x, n.X.Type(), n.Type())
2838 n := n.(*ir.TypeAssertExpr)
2839 res, _ := s.dottype(n, false)
2842 case ir.ODYNAMICDOTTYPE:
2843 n := n.(*ir.DynamicTypeAssertExpr)
2844 res, _ := s.dynamicDottype(n, false)
2848 case ir.OLT, ir.OEQ, ir.ONE, ir.OLE, ir.OGE, ir.OGT:
2849 n := n.(*ir.BinaryExpr)
2852 if n.X.Type().IsComplex() {
2853 pt := types.FloatForComplex(n.X.Type())
2854 op := s.ssaOp(ir.OEQ, pt)
2855 r := s.newValueOrSfCall2(op, types.Types[types.TBOOL], s.newValue1(ssa.OpComplexReal, pt, a), s.newValue1(ssa.OpComplexReal, pt, b))
2856 i := s.newValueOrSfCall2(op, types.Types[types.TBOOL], s.newValue1(ssa.OpComplexImag, pt, a), s.newValue1(ssa.OpComplexImag, pt, b))
2857 c := s.newValue2(ssa.OpAndB, types.Types[types.TBOOL], r, i)
2862 return s.newValue1(ssa.OpNot, types.Types[types.TBOOL], c)
2864 s.Fatalf("ordered complex compare %v", n.Op())
2868 // Convert OGE and OGT into OLE and OLT.
2872 op, a, b = ir.OLE, b, a
2874 op, a, b = ir.OLT, b, a
2876 if n.X.Type().IsFloat() {
2878 return s.newValueOrSfCall2(s.ssaOp(op, n.X.Type()), types.Types[types.TBOOL], a, b)
2880 // integer comparison
2881 return s.newValue2(s.ssaOp(op, n.X.Type()), types.Types[types.TBOOL], a, b)
2883 n := n.(*ir.BinaryExpr)
2886 if n.Type().IsComplex() {
2887 mulop := ssa.OpMul64F
2888 addop := ssa.OpAdd64F
2889 subop := ssa.OpSub64F
2890 pt := types.FloatForComplex(n.Type()) // Could be Float32 or Float64
2891 wt := types.Types[types.TFLOAT64] // Compute in Float64 to minimize cancellation error
2893 areal := s.newValue1(ssa.OpComplexReal, pt, a)
2894 breal := s.newValue1(ssa.OpComplexReal, pt, b)
2895 aimag := s.newValue1(ssa.OpComplexImag, pt, a)
2896 bimag := s.newValue1(ssa.OpComplexImag, pt, b)
2898 if pt != wt { // Widen for calculation
2899 areal = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, areal)
2900 breal = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, breal)
2901 aimag = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, aimag)
2902 bimag = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, bimag)
2905 xreal := s.newValueOrSfCall2(subop, wt, s.newValueOrSfCall2(mulop, wt, areal, breal), s.newValueOrSfCall2(mulop, wt, aimag, bimag))
2906 ximag := s.newValueOrSfCall2(addop, wt, s.newValueOrSfCall2(mulop, wt, areal, bimag), s.newValueOrSfCall2(mulop, wt, aimag, breal))
2908 if pt != wt { // Narrow to store back
2909 xreal = s.newValueOrSfCall1(ssa.OpCvt64Fto32F, pt, xreal)
2910 ximag = s.newValueOrSfCall1(ssa.OpCvt64Fto32F, pt, ximag)
2913 return s.newValue2(ssa.OpComplexMake, n.Type(), xreal, ximag)
2916 if n.Type().IsFloat() {
2917 return s.newValueOrSfCall2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2920 return s.newValue2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2923 n := n.(*ir.BinaryExpr)
2926 if n.Type().IsComplex() {
2927 // TODO this is not executed because the front-end substitutes a runtime call.
2928 // That probably ought to change; with modest optimization the widen/narrow
2929 // conversions could all be elided in larger expression trees.
2930 mulop := ssa.OpMul64F
2931 addop := ssa.OpAdd64F
2932 subop := ssa.OpSub64F
2933 divop := ssa.OpDiv64F
2934 pt := types.FloatForComplex(n.Type()) // Could be Float32 or Float64
2935 wt := types.Types[types.TFLOAT64] // Compute in Float64 to minimize cancellation error
2937 areal := s.newValue1(ssa.OpComplexReal, pt, a)
2938 breal := s.newValue1(ssa.OpComplexReal, pt, b)
2939 aimag := s.newValue1(ssa.OpComplexImag, pt, a)
2940 bimag := s.newValue1(ssa.OpComplexImag, pt, b)
2942 if pt != wt { // Widen for calculation
2943 areal = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, areal)
2944 breal = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, breal)
2945 aimag = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, aimag)
2946 bimag = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, bimag)
2949 denom := s.newValueOrSfCall2(addop, wt, s.newValueOrSfCall2(mulop, wt, breal, breal), s.newValueOrSfCall2(mulop, wt, bimag, bimag))
2950 xreal := s.newValueOrSfCall2(addop, wt, s.newValueOrSfCall2(mulop, wt, areal, breal), s.newValueOrSfCall2(mulop, wt, aimag, bimag))
2951 ximag := s.newValueOrSfCall2(subop, wt, s.newValueOrSfCall2(mulop, wt, aimag, breal), s.newValueOrSfCall2(mulop, wt, areal, bimag))
2953 // TODO not sure if this is best done in wide precision or narrow
2954 // Double-rounding might be an issue.
2955 // Note that the pre-SSA implementation does the entire calculation
2956 // in wide format, so wide is compatible.
2957 xreal = s.newValueOrSfCall2(divop, wt, xreal, denom)
2958 ximag = s.newValueOrSfCall2(divop, wt, ximag, denom)
2960 if pt != wt { // Narrow to store back
2961 xreal = s.newValueOrSfCall1(ssa.OpCvt64Fto32F, pt, xreal)
2962 ximag = s.newValueOrSfCall1(ssa.OpCvt64Fto32F, pt, ximag)
2964 return s.newValue2(ssa.OpComplexMake, n.Type(), xreal, ximag)
2966 if n.Type().IsFloat() {
2967 return s.newValueOrSfCall2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2969 return s.intDivide(n, a, b)
2971 n := n.(*ir.BinaryExpr)
2974 return s.intDivide(n, a, b)
2975 case ir.OADD, ir.OSUB:
2976 n := n.(*ir.BinaryExpr)
2979 if n.Type().IsComplex() {
2980 pt := types.FloatForComplex(n.Type())
2981 op := s.ssaOp(n.Op(), pt)
2982 return s.newValue2(ssa.OpComplexMake, n.Type(),
2983 s.newValueOrSfCall2(op, pt, s.newValue1(ssa.OpComplexReal, pt, a), s.newValue1(ssa.OpComplexReal, pt, b)),
2984 s.newValueOrSfCall2(op, pt, s.newValue1(ssa.OpComplexImag, pt, a), s.newValue1(ssa.OpComplexImag, pt, b)))
2986 if n.Type().IsFloat() {
2987 return s.newValueOrSfCall2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2989 return s.newValue2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2990 case ir.OAND, ir.OOR, ir.OXOR:
2991 n := n.(*ir.BinaryExpr)
2994 return s.newValue2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
2996 n := n.(*ir.BinaryExpr)
2999 b = s.newValue1(s.ssaOp(ir.OBITNOT, b.Type), b.Type, b)
3000 return s.newValue2(s.ssaOp(ir.OAND, n.Type()), a.Type, a, b)
3001 case ir.OLSH, ir.ORSH:
3002 n := n.(*ir.BinaryExpr)
3007 cmp := s.newValue2(s.ssaOp(ir.OLE, bt), types.Types[types.TBOOL], s.zeroVal(bt), b)
3008 s.check(cmp, ir.Syms.Panicshift)
3009 bt = bt.ToUnsigned()
3011 return s.newValue2(s.ssaShiftOp(n.Op(), n.Type(), bt), a.Type, a, b)
3012 case ir.OANDAND, ir.OOROR:
3013 // To implement OANDAND (and OOROR), we introduce a
3014 // new temporary variable to hold the result. The
3015 // variable is associated with the OANDAND node in the
3016 // s.vars table (normally variables are only
3017 // associated with ONAME nodes). We convert
3024 // Using var in the subsequent block introduces the
3025 // necessary phi variable.
3026 n := n.(*ir.LogicalExpr)
3031 b.Kind = ssa.BlockIf
3033 // In theory, we should set b.Likely here based on context.
3034 // However, gc only gives us likeliness hints
3035 // in a single place, for plain OIF statements,
3036 // and passing around context is finnicky, so don't bother for now.
3038 bRight := s.f.NewBlock(ssa.BlockPlain)
3039 bResult := s.f.NewBlock(ssa.BlockPlain)
3040 if n.Op() == ir.OANDAND {
3042 b.AddEdgeTo(bResult)
3043 } else if n.Op() == ir.OOROR {
3044 b.AddEdgeTo(bResult)
3048 s.startBlock(bRight)
3053 b.AddEdgeTo(bResult)
3055 s.startBlock(bResult)
3056 return s.variable(n, types.Types[types.TBOOL])
3058 n := n.(*ir.BinaryExpr)
3061 return s.newValue2(ssa.OpComplexMake, n.Type(), r, i)
3065 n := n.(*ir.UnaryExpr)
3067 if n.Type().IsComplex() {
3068 tp := types.FloatForComplex(n.Type())
3069 negop := s.ssaOp(n.Op(), tp)
3070 return s.newValue2(ssa.OpComplexMake, n.Type(),
3071 s.newValue1(negop, tp, s.newValue1(ssa.OpComplexReal, tp, a)),
3072 s.newValue1(negop, tp, s.newValue1(ssa.OpComplexImag, tp, a)))
3074 return s.newValue1(s.ssaOp(n.Op(), n.Type()), a.Type, a)
3075 case ir.ONOT, ir.OBITNOT:
3076 n := n.(*ir.UnaryExpr)
3078 return s.newValue1(s.ssaOp(n.Op(), n.Type()), a.Type, a)
3079 case ir.OIMAG, ir.OREAL:
3080 n := n.(*ir.UnaryExpr)
3082 return s.newValue1(s.ssaOp(n.Op(), n.X.Type()), n.Type(), a)
3084 n := n.(*ir.UnaryExpr)
3088 n := n.(*ir.AddrExpr)
3092 n := n.(*ir.ResultExpr)
3093 if s.prevCall == nil || s.prevCall.Op != ssa.OpStaticLECall && s.prevCall.Op != ssa.OpInterLECall && s.prevCall.Op != ssa.OpClosureLECall {
3094 panic("Expected to see a previous call")
3098 panic(fmt.Errorf("ORESULT %v does not match call %s", n, s.prevCall))
3100 return s.resultOfCall(s.prevCall, which, n.Type())
3103 n := n.(*ir.StarExpr)
3104 p := s.exprPtr(n.X, n.Bounded(), n.Pos())
3105 return s.load(n.Type(), p)
3108 n := n.(*ir.SelectorExpr)
3109 if n.X.Op() == ir.OSTRUCTLIT {
3110 // All literals with nonzero fields have already been
3111 // rewritten during walk. Any that remain are just T{}
3112 // or equivalents. Use the zero value.
3113 if !ir.IsZero(n.X) {
3114 s.Fatalf("literal with nonzero value in SSA: %v", n.X)
3116 return s.zeroVal(n.Type())
3118 // If n is addressable and can't be represented in
3119 // SSA, then load just the selected field. This
3120 // prevents false memory dependencies in race/msan/asan
3122 if ir.IsAddressable(n) && !s.canSSA(n) {
3124 return s.load(n.Type(), p)
3127 return s.newValue1I(ssa.OpStructSelect, n.Type(), int64(fieldIdx(n)), v)
3130 n := n.(*ir.SelectorExpr)
3131 p := s.exprPtr(n.X, n.Bounded(), n.Pos())
3132 p = s.newValue1I(ssa.OpOffPtr, types.NewPtr(n.Type()), n.Offset(), p)
3133 return s.load(n.Type(), p)
3136 n := n.(*ir.IndexExpr)
3138 case n.X.Type().IsString():
3139 if n.Bounded() && ir.IsConst(n.X, constant.String) && ir.IsConst(n.Index, constant.Int) {
3140 // Replace "abc"[1] with 'b'.
3141 // Delayed until now because "abc"[1] is not an ideal constant.
3142 // See test/fixedbugs/issue11370.go.
3143 return s.newValue0I(ssa.OpConst8, types.Types[types.TUINT8], int64(int8(ir.StringVal(n.X)[ir.Int64Val(n.Index)])))
3146 i := s.expr(n.Index)
3147 len := s.newValue1(ssa.OpStringLen, types.Types[types.TINT], a)
3148 i = s.boundsCheck(i, len, ssa.BoundsIndex, n.Bounded())
3149 ptrtyp := s.f.Config.Types.BytePtr
3150 ptr := s.newValue1(ssa.OpStringPtr, ptrtyp, a)
3151 if ir.IsConst(n.Index, constant.Int) {
3152 ptr = s.newValue1I(ssa.OpOffPtr, ptrtyp, ir.Int64Val(n.Index), ptr)
3154 ptr = s.newValue2(ssa.OpAddPtr, ptrtyp, ptr, i)
3156 return s.load(types.Types[types.TUINT8], ptr)
3157 case n.X.Type().IsSlice():
3159 return s.load(n.X.Type().Elem(), p)
3160 case n.X.Type().IsArray():
3161 if ssa.CanSSA(n.X.Type()) {
3162 // SSA can handle arrays of length at most 1.
3163 bound := n.X.Type().NumElem()
3165 i := s.expr(n.Index)
3167 // Bounds check will never succeed. Might as well
3168 // use constants for the bounds check.
3169 z := s.constInt(types.Types[types.TINT], 0)
3170 s.boundsCheck(z, z, ssa.BoundsIndex, false)
3171 // The return value won't be live, return junk.
3172 // But not quite junk, in case bounds checks are turned off. See issue 48092.
3173 return s.zeroVal(n.Type())
3175 len := s.constInt(types.Types[types.TINT], bound)
3176 s.boundsCheck(i, len, ssa.BoundsIndex, n.Bounded()) // checks i == 0
3177 return s.newValue1I(ssa.OpArraySelect, n.Type(), 0, a)
3180 return s.load(n.X.Type().Elem(), p)
3182 s.Fatalf("bad type for index %v", n.X.Type())
3186 case ir.OLEN, ir.OCAP:
3187 n := n.(*ir.UnaryExpr)
3189 case n.X.Type().IsSlice():
3190 op := ssa.OpSliceLen
3191 if n.Op() == ir.OCAP {
3194 return s.newValue1(op, types.Types[types.TINT], s.expr(n.X))
3195 case n.X.Type().IsString(): // string; not reachable for OCAP
3196 return s.newValue1(ssa.OpStringLen, types.Types[types.TINT], s.expr(n.X))
3197 case n.X.Type().IsMap(), n.X.Type().IsChan():
3198 return s.referenceTypeBuiltin(n, s.expr(n.X))
3200 return s.constInt(types.Types[types.TINT], n.X.Type().NumElem())
3204 n := n.(*ir.UnaryExpr)
3206 if n.X.Type().IsSlice() {
3208 return s.newValue1(ssa.OpSlicePtr, n.Type(), a)
3210 return s.newValue1(ssa.OpSlicePtrUnchecked, n.Type(), a)
3212 return s.newValue1(ssa.OpStringPtr, n.Type(), a)
3216 n := n.(*ir.UnaryExpr)
3218 return s.newValue1(ssa.OpITab, n.Type(), a)
3221 n := n.(*ir.UnaryExpr)
3223 return s.newValue1(ssa.OpIData, n.Type(), a)
3226 n := n.(*ir.BinaryExpr)
3229 return s.newValue2(ssa.OpIMake, n.Type(), tab, data)
3231 case ir.OSLICEHEADER:
3232 n := n.(*ir.SliceHeaderExpr)
3236 return s.newValue3(ssa.OpSliceMake, n.Type(), p, l, c)
3238 case ir.OSTRINGHEADER:
3239 n := n.(*ir.StringHeaderExpr)
3242 return s.newValue2(ssa.OpStringMake, n.Type(), p, l)
3244 case ir.OSLICE, ir.OSLICEARR, ir.OSLICE3, ir.OSLICE3ARR:
3245 n := n.(*ir.SliceExpr)
3246 check := s.checkPtrEnabled && n.Op() == ir.OSLICE3ARR && n.X.Op() == ir.OCONVNOP && n.X.(*ir.ConvExpr).X.Type().IsUnsafePtr()
3247 v := s.exprCheckPtr(n.X, !check)
3248 var i, j, k *ssa.Value
3258 p, l, c := s.slice(v, i, j, k, n.Bounded())
3260 // Emit checkptr instrumentation after bound check to prevent false positive, see #46938.
3261 s.checkPtrAlignment(n.X.(*ir.ConvExpr), v, s.conv(n.Max, k, k.Type, types.Types[types.TUINTPTR]))
3263 return s.newValue3(ssa.OpSliceMake, n.Type(), p, l, c)
3266 n := n.(*ir.SliceExpr)
3275 p, l, _ := s.slice(v, i, j, nil, n.Bounded())
3276 return s.newValue2(ssa.OpStringMake, n.Type(), p, l)
3278 case ir.OSLICE2ARRPTR:
3279 // if arrlen > slice.len {
3283 n := n.(*ir.ConvExpr)
3285 nelem := n.Type().Elem().NumElem()
3286 arrlen := s.constInt(types.Types[types.TINT], nelem)
3287 cap := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], v)
3288 s.boundsCheck(arrlen, cap, ssa.BoundsConvert, false)
3289 op := ssa.OpSlicePtr
3291 op = ssa.OpSlicePtrUnchecked
3293 return s.newValue1(op, n.Type(), v)
3296 n := n.(*ir.CallExpr)
3297 if ir.IsIntrinsicCall(n) {
3298 return s.intrinsicCall(n)
3303 n := n.(*ir.CallExpr)
3304 return s.callResult(n, callNormal)
3307 n := n.(*ir.CallExpr)
3308 return s.newValue1(ssa.OpGetG, n.Type(), s.mem())
3310 case ir.OGETCALLERPC:
3311 n := n.(*ir.CallExpr)
3312 return s.newValue0(ssa.OpGetCallerPC, n.Type())
3314 case ir.OGETCALLERSP:
3315 n := n.(*ir.CallExpr)
3316 return s.newValue1(ssa.OpGetCallerSP, n.Type(), s.mem())
3319 return s.append(n.(*ir.CallExpr), false)
3321 case ir.OMIN, ir.OMAX:
3322 return s.minMax(n.(*ir.CallExpr))
3324 case ir.OSTRUCTLIT, ir.OARRAYLIT:
3325 // All literals with nonzero fields have already been
3326 // rewritten during walk. Any that remain are just T{}
3327 // or equivalents. Use the zero value.
3328 n := n.(*ir.CompLitExpr)
3330 s.Fatalf("literal with nonzero value in SSA: %v", n)
3332 return s.zeroVal(n.Type())
3335 n := n.(*ir.UnaryExpr)
3336 var rtype *ssa.Value
3337 if x, ok := n.X.(*ir.DynamicType); ok && x.Op() == ir.ODYNAMICTYPE {
3338 rtype = s.expr(x.RType)
3340 return s.newObject(n.Type().Elem(), rtype)
3343 n := n.(*ir.BinaryExpr)
3347 // Force len to uintptr to prevent misuse of garbage bits in the
3348 // upper part of the register (#48536).
3349 len = s.conv(n, len, len.Type, types.Types[types.TUINTPTR])
3351 return s.newValue2(ssa.OpAddPtr, n.Type(), ptr, len)
3354 s.Fatalf("unhandled expr %v", n.Op())
3359 func (s *state) resultOfCall(c *ssa.Value, which int64, t *types.Type) *ssa.Value {
3360 aux := c.Aux.(*ssa.AuxCall)
3361 pa := aux.ParamAssignmentForResult(which)
3362 // TODO(register args) determine if in-memory TypeOK is better loaded early from SelectNAddr or later when SelectN is expanded.
3363 // SelectN is better for pattern-matching and possible call-aware analysis we might want to do in the future.
3364 if len(pa.Registers) == 0 && !ssa.CanSSA(t) {
3365 addr := s.newValue1I(ssa.OpSelectNAddr, types.NewPtr(t), which, c)
3366 return s.rawLoad(t, addr)
3368 return s.newValue1I(ssa.OpSelectN, t, which, c)
3371 func (s *state) resultAddrOfCall(c *ssa.Value, which int64, t *types.Type) *ssa.Value {
3372 aux := c.Aux.(*ssa.AuxCall)
3373 pa := aux.ParamAssignmentForResult(which)
3374 if len(pa.Registers) == 0 {
3375 return s.newValue1I(ssa.OpSelectNAddr, types.NewPtr(t), which, c)
3377 _, addr := s.temp(c.Pos, t)
3378 rval := s.newValue1I(ssa.OpSelectN, t, which, c)
3379 s.vars[memVar] = s.newValue3Apos(ssa.OpStore, types.TypeMem, t, addr, rval, s.mem(), false)
3383 // append converts an OAPPEND node to SSA.
3384 // If inplace is false, it converts the OAPPEND expression n to an ssa.Value,
3385 // adds it to s, and returns the Value.
3386 // If inplace is true, it writes the result of the OAPPEND expression n
3387 // back to the slice being appended to, and returns nil.
3388 // inplace MUST be set to false if the slice can be SSA'd.
3389 // Note: this code only handles fixed-count appends. Dotdotdot appends
3390 // have already been rewritten at this point (by walk).
3391 func (s *state) append(n *ir.CallExpr, inplace bool) *ssa.Value {
3392 // If inplace is false, process as expression "append(s, e1, e2, e3)":
3394 // ptr, len, cap := s
3396 // if uint(len) > uint(cap) {
3397 // ptr, len, cap = growslice(ptr, len, cap, 3, typ)
3398 // Note that len is unmodified by growslice.
3400 // // with write barriers, if needed:
3401 // *(ptr+(len-3)) = e1
3402 // *(ptr+(len-2)) = e2
3403 // *(ptr+(len-1)) = e3
3404 // return makeslice(ptr, len, cap)
3407 // If inplace is true, process as statement "s = append(s, e1, e2, e3)":
3410 // ptr, len, cap := s
3412 // if uint(len) > uint(cap) {
3413 // ptr, len, cap = growslice(ptr, len, cap, 3, typ)
3414 // vardef(a) // if necessary, advise liveness we are writing a new a
3415 // *a.cap = cap // write before ptr to avoid a spill
3416 // *a.ptr = ptr // with write barrier
3419 // // with write barriers, if needed:
3420 // *(ptr+(len-3)) = e1
3421 // *(ptr+(len-2)) = e2
3422 // *(ptr+(len-1)) = e3
3424 et := n.Type().Elem()
3425 pt := types.NewPtr(et)
3428 sn := n.Args[0] // the slice node is the first in the list
3429 var slice, addr *ssa.Value
3432 slice = s.load(n.Type(), addr)
3437 // Allocate new blocks
3438 grow := s.f.NewBlock(ssa.BlockPlain)
3439 assign := s.f.NewBlock(ssa.BlockPlain)
3441 // Decomposse input slice.
3442 p := s.newValue1(ssa.OpSlicePtr, pt, slice)
3443 l := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], slice)
3444 c := s.newValue1(ssa.OpSliceCap, types.Types[types.TINT], slice)
3446 // Add number of new elements to length.
3447 nargs := s.constInt(types.Types[types.TINT], int64(len(n.Args)-1))
3448 l = s.newValue2(s.ssaOp(ir.OADD, types.Types[types.TINT]), types.Types[types.TINT], l, nargs)
3450 // Decide if we need to grow
3451 cmp := s.newValue2(s.ssaOp(ir.OLT, types.Types[types.TUINT]), types.Types[types.TBOOL], c, l)
3453 // Record values of ptr/len/cap before branch.
3461 b.Kind = ssa.BlockIf
3462 b.Likely = ssa.BranchUnlikely
3469 taddr := s.expr(n.X)
3470 r := s.rtcall(ir.Syms.Growslice, true, []*types.Type{n.Type()}, p, l, c, nargs, taddr)
3472 // Decompose output slice
3473 p = s.newValue1(ssa.OpSlicePtr, pt, r[0])
3474 l = s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], r[0])
3475 c = s.newValue1(ssa.OpSliceCap, types.Types[types.TINT], r[0])
3481 if sn.Op() == ir.ONAME {
3483 if sn.Class != ir.PEXTERN {
3484 // Tell liveness we're about to build a new slice
3485 s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, sn, s.mem())
3488 capaddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, types.SliceCapOffset, addr)
3489 s.store(types.Types[types.TINT], capaddr, c)
3490 s.store(pt, addr, p)
3496 // assign new elements to slots
3497 s.startBlock(assign)
3498 p = s.variable(ptrVar, pt) // generates phi for ptr
3499 l = s.variable(lenVar, types.Types[types.TINT]) // generates phi for len
3501 c = s.variable(capVar, types.Types[types.TINT]) // generates phi for cap
3505 // Update length in place.
3506 // We have to wait until here to make sure growslice succeeded.
3507 lenaddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, types.SliceLenOffset, addr)
3508 s.store(types.Types[types.TINT], lenaddr, l)
3512 type argRec struct {
3513 // if store is true, we're appending the value v. If false, we're appending the
3518 args := make([]argRec, 0, len(n.Args[1:]))
3519 for _, n := range n.Args[1:] {
3520 if ssa.CanSSA(n.Type()) {
3521 args = append(args, argRec{v: s.expr(n), store: true})
3524 args = append(args, argRec{v: v})
3528 // Write args into slice.
3529 oldLen := s.newValue2(s.ssaOp(ir.OSUB, types.Types[types.TINT]), types.Types[types.TINT], l, nargs)
3530 p2 := s.newValue2(ssa.OpPtrIndex, pt, p, oldLen)
3531 for i, arg := range args {
3532 addr := s.newValue2(ssa.OpPtrIndex, pt, p2, s.constInt(types.Types[types.TINT], int64(i)))
3534 s.storeType(et, addr, arg.v, 0, true)
3536 s.move(et, addr, arg.v)
3540 // The following deletions have no practical effect at this time
3541 // because state.vars has been reset by the preceding state.startBlock.
3542 // They only enforce the fact that these variables are no longer need in
3543 // the current scope.
3544 delete(s.vars, ptrVar)
3545 delete(s.vars, lenVar)
3547 delete(s.vars, capVar)
3554 return s.newValue3(ssa.OpSliceMake, n.Type(), p, l, c)
3557 // minMax converts an OMIN/OMAX builtin call into SSA.
3558 func (s *state) minMax(n *ir.CallExpr) *ssa.Value {
3559 // The OMIN/OMAX builtin is variadic, but its semantics are
3560 // equivalent to left-folding a binary min/max operation across the
3562 fold := func(op func(x, a *ssa.Value) *ssa.Value) *ssa.Value {
3563 x := s.expr(n.Args[0])
3564 for _, arg := range n.Args[1:] {
3565 x = op(x, s.expr(arg))
3572 if typ.IsFloat() || typ.IsString() {
3573 // min/max semantics for floats are tricky because of NaNs and
3574 // negative zero. Some architectures have instructions which
3575 // we can use to generate the right result. For others we must
3576 // call into the runtime instead.
3578 // Strings are conceptually simpler, but we currently desugar
3579 // string comparisons during walk, not ssagen.
3582 switch Arch.LinkArch.Family {
3583 case sys.AMD64, sys.ARM64:
3586 case typ.Kind() == types.TFLOAT64 && n.Op() == ir.OMIN:
3588 case typ.Kind() == types.TFLOAT64 && n.Op() == ir.OMAX:
3590 case typ.Kind() == types.TFLOAT32 && n.Op() == ir.OMIN:
3592 case typ.Kind() == types.TFLOAT32 && n.Op() == ir.OMAX:
3595 return fold(func(x, a *ssa.Value) *ssa.Value {
3596 return s.newValue2(op, typ, x, a)
3602 case types.TFLOAT32:
3609 case types.TFLOAT64:
3624 fn := typecheck.LookupRuntimeFunc(name)
3626 return fold(func(x, a *ssa.Value) *ssa.Value {
3627 return s.rtcall(fn, true, []*types.Type{typ}, x, a)[0]
3631 lt := s.ssaOp(ir.OLT, typ)
3633 return fold(func(x, a *ssa.Value) *ssa.Value {
3637 return s.ternary(s.newValue2(lt, types.Types[types.TBOOL], a, x), a, x)
3640 return s.ternary(s.newValue2(lt, types.Types[types.TBOOL], x, a), a, x)
3642 panic("unreachable")
3646 // ternary emits code to evaluate cond ? x : y.
3647 func (s *state) ternary(cond, x, y *ssa.Value) *ssa.Value {
3648 // Note that we need a new ternaryVar each time (unlike okVar where we can
3649 // reuse the variable) because it might have a different type every time.
3650 ternaryVar := ssaMarker("ternary")
3652 bThen := s.f.NewBlock(ssa.BlockPlain)
3653 bElse := s.f.NewBlock(ssa.BlockPlain)
3654 bEnd := s.f.NewBlock(ssa.BlockPlain)
3657 b.Kind = ssa.BlockIf
3663 s.vars[ternaryVar] = x
3664 s.endBlock().AddEdgeTo(bEnd)
3667 s.vars[ternaryVar] = y
3668 s.endBlock().AddEdgeTo(bEnd)
3671 r := s.variable(ternaryVar, x.Type)
3672 delete(s.vars, ternaryVar)
3676 // condBranch evaluates the boolean expression cond and branches to yes
3677 // if cond is true and no if cond is false.
3678 // This function is intended to handle && and || better than just calling
3679 // s.expr(cond) and branching on the result.
3680 func (s *state) condBranch(cond ir.Node, yes, no *ssa.Block, likely int8) {
3683 cond := cond.(*ir.LogicalExpr)
3684 mid := s.f.NewBlock(ssa.BlockPlain)
3685 s.stmtList(cond.Init())
3686 s.condBranch(cond.X, mid, no, max8(likely, 0))
3688 s.condBranch(cond.Y, yes, no, likely)
3690 // Note: if likely==1, then both recursive calls pass 1.
3691 // If likely==-1, then we don't have enough information to decide
3692 // whether the first branch is likely or not. So we pass 0 for
3693 // the likeliness of the first branch.
3694 // TODO: have the frontend give us branch prediction hints for
3695 // OANDAND and OOROR nodes (if it ever has such info).
3697 cond := cond.(*ir.LogicalExpr)
3698 mid := s.f.NewBlock(ssa.BlockPlain)
3699 s.stmtList(cond.Init())
3700 s.condBranch(cond.X, yes, mid, min8(likely, 0))
3702 s.condBranch(cond.Y, yes, no, likely)
3704 // Note: if likely==-1, then both recursive calls pass -1.
3705 // If likely==1, then we don't have enough info to decide
3706 // the likelihood of the first branch.
3708 cond := cond.(*ir.UnaryExpr)
3709 s.stmtList(cond.Init())
3710 s.condBranch(cond.X, no, yes, -likely)
3713 cond := cond.(*ir.ConvExpr)
3714 s.stmtList(cond.Init())
3715 s.condBranch(cond.X, yes, no, likely)
3720 b.Kind = ssa.BlockIf
3722 b.Likely = ssa.BranchPrediction(likely) // gc and ssa both use -1/0/+1 for likeliness
3730 skipPtr skipMask = 1 << iota
3735 // assign does left = right.
3736 // Right has already been evaluated to ssa, left has not.
3737 // If deref is true, then we do left = *right instead (and right has already been nil-checked).
3738 // If deref is true and right == nil, just do left = 0.
3739 // skip indicates assignments (at the top level) that can be avoided.
3740 // mayOverlap indicates whether left&right might partially overlap in memory. Default is false.
3741 func (s *state) assign(left ir.Node, right *ssa.Value, deref bool, skip skipMask) {
3742 s.assignWhichMayOverlap(left, right, deref, skip, false)
3744 func (s *state) assignWhichMayOverlap(left ir.Node, right *ssa.Value, deref bool, skip skipMask, mayOverlap bool) {
3745 if left.Op() == ir.ONAME && ir.IsBlank(left) {
3752 s.Fatalf("can SSA LHS %v but not RHS %s", left, right)
3754 if left.Op() == ir.ODOT {
3755 // We're assigning to a field of an ssa-able value.
3756 // We need to build a new structure with the new value for the
3757 // field we're assigning and the old values for the other fields.
3759 // type T struct {a, b, c int}
3762 // For the x.b = 5 assignment we want to generate x = T{x.a, 5, x.c}
3764 // Grab information about the structure type.
3765 left := left.(*ir.SelectorExpr)
3768 idx := fieldIdx(left)
3770 // Grab old value of structure.
3771 old := s.expr(left.X)
3773 // Make new structure.
3774 new := s.newValue0(ssa.StructMakeOp(t.NumFields()), t)
3776 // Add fields as args.
3777 for i := 0; i < nf; i++ {
3781 new.AddArg(s.newValue1I(ssa.OpStructSelect, t.FieldType(i), int64(i), old))
3785 // Recursively assign the new value we've made to the base of the dot op.
3786 s.assign(left.X, new, false, 0)
3787 // TODO: do we need to update named values here?
3790 if left.Op() == ir.OINDEX && left.(*ir.IndexExpr).X.Type().IsArray() {
3791 left := left.(*ir.IndexExpr)
3792 s.pushLine(left.Pos())
3794 // We're assigning to an element of an ssa-able array.
3799 i := s.expr(left.Index) // index
3801 // The bounds check must fail. Might as well
3802 // ignore the actual index and just use zeros.
3803 z := s.constInt(types.Types[types.TINT], 0)
3804 s.boundsCheck(z, z, ssa.BoundsIndex, false)
3808 s.Fatalf("assigning to non-1-length array")
3810 // Rewrite to a = [1]{v}
3811 len := s.constInt(types.Types[types.TINT], 1)
3812 s.boundsCheck(i, len, ssa.BoundsIndex, false) // checks i == 0
3813 v := s.newValue1(ssa.OpArrayMake1, t, right)
3814 s.assign(left.X, v, false, 0)
3817 left := left.(*ir.Name)
3818 // Update variable assignment.
3819 s.vars[left] = right
3820 s.addNamedValue(left, right)
3824 // If this assignment clobbers an entire local variable, then emit
3825 // OpVarDef so liveness analysis knows the variable is redefined.
3826 if base, ok := clobberBase(left).(*ir.Name); ok && base.OnStack() && skip == 0 && t.HasPointers() {
3827 s.vars[memVar] = s.newValue1Apos(ssa.OpVarDef, types.TypeMem, base, s.mem(), !ir.IsAutoTmp(base))
3830 // Left is not ssa-able. Compute its address.
3831 addr := s.addr(left)
3832 if ir.IsReflectHeaderDataField(left) {
3833 // Package unsafe's documentation says storing pointers into
3834 // reflect.SliceHeader and reflect.StringHeader's Data fields
3835 // is valid, even though they have type uintptr (#19168).
3836 // Mark it pointer type to signal the writebarrier pass to
3837 // insert a write barrier.
3838 t = types.Types[types.TUNSAFEPTR]
3841 // Treat as a mem->mem move.
3845 s.moveWhichMayOverlap(t, addr, right, mayOverlap)
3849 // Treat as a store.
3850 s.storeType(t, addr, right, skip, !ir.IsAutoTmp(left))
3853 // zeroVal returns the zero value for type t.
3854 func (s *state) zeroVal(t *types.Type) *ssa.Value {
3859 return s.constInt8(t, 0)
3861 return s.constInt16(t, 0)
3863 return s.constInt32(t, 0)
3865 return s.constInt64(t, 0)
3867 s.Fatalf("bad sized integer type %v", t)
3872 return s.constFloat32(t, 0)
3874 return s.constFloat64(t, 0)
3876 s.Fatalf("bad sized float type %v", t)
3881 z := s.constFloat32(types.Types[types.TFLOAT32], 0)
3882 return s.entryNewValue2(ssa.OpComplexMake, t, z, z)
3884 z := s.constFloat64(types.Types[types.TFLOAT64], 0)
3885 return s.entryNewValue2(ssa.OpComplexMake, t, z, z)
3887 s.Fatalf("bad sized complex type %v", t)
3891 return s.constEmptyString(t)
3892 case t.IsPtrShaped():
3893 return s.constNil(t)
3895 return s.constBool(false)
3896 case t.IsInterface():
3897 return s.constInterface(t)
3899 return s.constSlice(t)
3902 v := s.entryNewValue0(ssa.StructMakeOp(t.NumFields()), t)
3903 for i := 0; i < n; i++ {
3904 v.AddArg(s.zeroVal(t.FieldType(i)))
3908 switch t.NumElem() {
3910 return s.entryNewValue0(ssa.OpArrayMake0, t)
3912 return s.entryNewValue1(ssa.OpArrayMake1, t, s.zeroVal(t.Elem()))
3915 s.Fatalf("zero for type %v not implemented", t)
3922 callNormal callKind = iota
3929 type sfRtCallDef struct {
3934 var softFloatOps map[ssa.Op]sfRtCallDef
3936 func softfloatInit() {
3937 // Some of these operations get transformed by sfcall.
3938 softFloatOps = map[ssa.Op]sfRtCallDef{
3939 ssa.OpAdd32F: {typecheck.LookupRuntimeFunc("fadd32"), types.TFLOAT32},
3940 ssa.OpAdd64F: {typecheck.LookupRuntimeFunc("fadd64"), types.TFLOAT64},
3941 ssa.OpSub32F: {typecheck.LookupRuntimeFunc("fadd32"), types.TFLOAT32},
3942 ssa.OpSub64F: {typecheck.LookupRuntimeFunc("fadd64"), types.TFLOAT64},
3943 ssa.OpMul32F: {typecheck.LookupRuntimeFunc("fmul32"), types.TFLOAT32},
3944 ssa.OpMul64F: {typecheck.LookupRuntimeFunc("fmul64"), types.TFLOAT64},
3945 ssa.OpDiv32F: {typecheck.LookupRuntimeFunc("fdiv32"), types.TFLOAT32},
3946 ssa.OpDiv64F: {typecheck.LookupRuntimeFunc("fdiv64"), types.TFLOAT64},
3948 ssa.OpEq64F: {typecheck.LookupRuntimeFunc("feq64"), types.TBOOL},
3949 ssa.OpEq32F: {typecheck.LookupRuntimeFunc("feq32"), types.TBOOL},
3950 ssa.OpNeq64F: {typecheck.LookupRuntimeFunc("feq64"), types.TBOOL},
3951 ssa.OpNeq32F: {typecheck.LookupRuntimeFunc("feq32"), types.TBOOL},
3952 ssa.OpLess64F: {typecheck.LookupRuntimeFunc("fgt64"), types.TBOOL},
3953 ssa.OpLess32F: {typecheck.LookupRuntimeFunc("fgt32"), types.TBOOL},
3954 ssa.OpLeq64F: {typecheck.LookupRuntimeFunc("fge64"), types.TBOOL},
3955 ssa.OpLeq32F: {typecheck.LookupRuntimeFunc("fge32"), types.TBOOL},
3957 ssa.OpCvt32to32F: {typecheck.LookupRuntimeFunc("fint32to32"), types.TFLOAT32},
3958 ssa.OpCvt32Fto32: {typecheck.LookupRuntimeFunc("f32toint32"), types.TINT32},
3959 ssa.OpCvt64to32F: {typecheck.LookupRuntimeFunc("fint64to32"), types.TFLOAT32},
3960 ssa.OpCvt32Fto64: {typecheck.LookupRuntimeFunc("f32toint64"), types.TINT64},
3961 ssa.OpCvt64Uto32F: {typecheck.LookupRuntimeFunc("fuint64to32"), types.TFLOAT32},
3962 ssa.OpCvt32Fto64U: {typecheck.LookupRuntimeFunc("f32touint64"), types.TUINT64},
3963 ssa.OpCvt32to64F: {typecheck.LookupRuntimeFunc("fint32to64"), types.TFLOAT64},
3964 ssa.OpCvt64Fto32: {typecheck.LookupRuntimeFunc("f64toint32"), types.TINT32},
3965 ssa.OpCvt64to64F: {typecheck.LookupRuntimeFunc("fint64to64"), types.TFLOAT64},
3966 ssa.OpCvt64Fto64: {typecheck.LookupRuntimeFunc("f64toint64"), types.TINT64},
3967 ssa.OpCvt64Uto64F: {typecheck.LookupRuntimeFunc("fuint64to64"), types.TFLOAT64},
3968 ssa.OpCvt64Fto64U: {typecheck.LookupRuntimeFunc("f64touint64"), types.TUINT64},
3969 ssa.OpCvt32Fto64F: {typecheck.LookupRuntimeFunc("f32to64"), types.TFLOAT64},
3970 ssa.OpCvt64Fto32F: {typecheck.LookupRuntimeFunc("f64to32"), types.TFLOAT32},
3974 // TODO: do not emit sfcall if operation can be optimized to constant in later
3976 func (s *state) sfcall(op ssa.Op, args ...*ssa.Value) (*ssa.Value, bool) {
3977 f2i := func(t *types.Type) *types.Type {
3979 case types.TFLOAT32:
3980 return types.Types[types.TUINT32]
3981 case types.TFLOAT64:
3982 return types.Types[types.TUINT64]
3987 if callDef, ok := softFloatOps[op]; ok {
3993 args[0], args[1] = args[1], args[0]
3996 args[1] = s.newValue1(s.ssaOp(ir.ONEG, types.Types[callDef.rtype]), args[1].Type, args[1])
3999 // runtime functions take uints for floats and returns uints.
4000 // Convert to uints so we use the right calling convention.
4001 for i, a := range args {
4002 if a.Type.IsFloat() {
4003 args[i] = s.newValue1(ssa.OpCopy, f2i(a.Type), a)
4007 rt := types.Types[callDef.rtype]
4008 result := s.rtcall(callDef.rtfn, true, []*types.Type{f2i(rt)}, args...)[0]
4010 result = s.newValue1(ssa.OpCopy, rt, result)
4012 if op == ssa.OpNeq32F || op == ssa.OpNeq64F {
4013 result = s.newValue1(ssa.OpNot, result.Type, result)
4020 var intrinsics map[intrinsicKey]intrinsicBuilder
4022 // An intrinsicBuilder converts a call node n into an ssa value that
4023 // implements that call as an intrinsic. args is a list of arguments to the func.
4024 type intrinsicBuilder func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value
4026 type intrinsicKey struct {
4033 intrinsics = map[intrinsicKey]intrinsicBuilder{}
4038 var lwatomics []*sys.Arch
4039 for _, a := range &sys.Archs {
4040 all = append(all, a)
4046 if a.Family != sys.PPC64 {
4047 lwatomics = append(lwatomics, a)
4051 // add adds the intrinsic b for pkg.fn for the given list of architectures.
4052 add := func(pkg, fn string, b intrinsicBuilder, archs ...*sys.Arch) {
4053 for _, a := range archs {
4054 intrinsics[intrinsicKey{a, pkg, fn}] = b
4057 // addF does the same as add but operates on architecture families.
4058 addF := func(pkg, fn string, b intrinsicBuilder, archFamilies ...sys.ArchFamily) {
4060 for _, f := range archFamilies {
4062 panic("too many architecture families")
4066 for _, a := range all {
4067 if m>>uint(a.Family)&1 != 0 {
4068 intrinsics[intrinsicKey{a, pkg, fn}] = b
4072 // alias defines pkg.fn = pkg2.fn2 for all architectures in archs for which pkg2.fn2 exists.
4073 alias := func(pkg, fn, pkg2, fn2 string, archs ...*sys.Arch) {
4075 for _, a := range archs {
4076 if b, ok := intrinsics[intrinsicKey{a, pkg2, fn2}]; ok {
4077 intrinsics[intrinsicKey{a, pkg, fn}] = b
4082 panic(fmt.Sprintf("attempted to alias undefined intrinsic: %s.%s", pkg, fn))
4086 /******** runtime ********/
4087 if !base.Flag.Cfg.Instrumenting {
4088 add("runtime", "slicebytetostringtmp",
4089 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4090 // Compiler frontend optimizations emit OBYTES2STRTMP nodes
4091 // for the backend instead of slicebytetostringtmp calls
4092 // when not instrumenting.
4093 return s.newValue2(ssa.OpStringMake, n.Type(), args[0], args[1])
4097 addF("runtime/internal/math", "MulUintptr",
4098 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4099 if s.config.PtrSize == 4 {
4100 return s.newValue2(ssa.OpMul32uover, types.NewTuple(types.Types[types.TUINT], types.Types[types.TUINT]), args[0], args[1])
4102 return s.newValue2(ssa.OpMul64uover, types.NewTuple(types.Types[types.TUINT], types.Types[types.TUINT]), args[0], args[1])
4104 sys.AMD64, sys.I386, sys.Loong64, sys.MIPS64, sys.RISCV64, sys.ARM64)
4105 alias("runtime", "mulUintptr", "runtime/internal/math", "MulUintptr", all...)
4106 add("runtime", "KeepAlive",
4107 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4108 data := s.newValue1(ssa.OpIData, s.f.Config.Types.BytePtr, args[0])
4109 s.vars[memVar] = s.newValue2(ssa.OpKeepAlive, types.TypeMem, data, s.mem())
4113 add("runtime", "getclosureptr",
4114 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4115 return s.newValue0(ssa.OpGetClosurePtr, s.f.Config.Types.Uintptr)
4119 add("runtime", "getcallerpc",
4120 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4121 return s.newValue0(ssa.OpGetCallerPC, s.f.Config.Types.Uintptr)
4125 add("runtime", "getcallersp",
4126 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4127 return s.newValue1(ssa.OpGetCallerSP, s.f.Config.Types.Uintptr, s.mem())
4131 addF("runtime", "publicationBarrier",
4132 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4133 s.vars[memVar] = s.newValue1(ssa.OpPubBarrier, types.TypeMem, s.mem())
4136 sys.ARM64, sys.PPC64)
4138 brev_arch := []sys.ArchFamily{sys.AMD64, sys.I386, sys.ARM64, sys.ARM, sys.S390X}
4139 if buildcfg.GOPPC64 >= 10 {
4140 // Use only on Power10 as the new byte reverse instructions that Power10 provide
4141 // make it worthwhile as an intrinsic
4142 brev_arch = append(brev_arch, sys.PPC64)
4144 /******** runtime/internal/sys ********/
4145 addF("runtime/internal/sys", "Bswap32",
4146 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4147 return s.newValue1(ssa.OpBswap32, types.Types[types.TUINT32], args[0])
4150 addF("runtime/internal/sys", "Bswap64",
4151 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4152 return s.newValue1(ssa.OpBswap64, types.Types[types.TUINT64], args[0])
4156 /****** Prefetch ******/
4157 makePrefetchFunc := func(op ssa.Op) func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4158 return func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4159 s.vars[memVar] = s.newValue2(op, types.TypeMem, args[0], s.mem())
4164 // Make Prefetch intrinsics for supported platforms
4165 // On the unsupported platforms stub function will be eliminated
4166 addF("runtime/internal/sys", "Prefetch", makePrefetchFunc(ssa.OpPrefetchCache),
4167 sys.AMD64, sys.ARM64, sys.PPC64)
4168 addF("runtime/internal/sys", "PrefetchStreamed", makePrefetchFunc(ssa.OpPrefetchCacheStreamed),
4169 sys.AMD64, sys.ARM64, sys.PPC64)
4171 /******** runtime/internal/atomic ********/
4172 addF("runtime/internal/atomic", "Load",
4173 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4174 v := s.newValue2(ssa.OpAtomicLoad32, types.NewTuple(types.Types[types.TUINT32], types.TypeMem), args[0], s.mem())
4175 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4176 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT32], v)
4178 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4179 addF("runtime/internal/atomic", "Load8",
4180 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4181 v := s.newValue2(ssa.OpAtomicLoad8, types.NewTuple(types.Types[types.TUINT8], types.TypeMem), args[0], s.mem())
4182 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4183 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT8], v)
4185 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4186 addF("runtime/internal/atomic", "Load64",
4187 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4188 v := s.newValue2(ssa.OpAtomicLoad64, types.NewTuple(types.Types[types.TUINT64], types.TypeMem), args[0], s.mem())
4189 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4190 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT64], v)
4192 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4193 addF("runtime/internal/atomic", "LoadAcq",
4194 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4195 v := s.newValue2(ssa.OpAtomicLoadAcq32, types.NewTuple(types.Types[types.TUINT32], types.TypeMem), args[0], s.mem())
4196 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4197 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT32], v)
4199 sys.PPC64, sys.S390X)
4200 addF("runtime/internal/atomic", "LoadAcq64",
4201 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4202 v := s.newValue2(ssa.OpAtomicLoadAcq64, types.NewTuple(types.Types[types.TUINT64], types.TypeMem), args[0], s.mem())
4203 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4204 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT64], v)
4207 addF("runtime/internal/atomic", "Loadp",
4208 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4209 v := s.newValue2(ssa.OpAtomicLoadPtr, types.NewTuple(s.f.Config.Types.BytePtr, types.TypeMem), args[0], s.mem())
4210 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4211 return s.newValue1(ssa.OpSelect0, s.f.Config.Types.BytePtr, v)
4213 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4215 addF("runtime/internal/atomic", "Store",
4216 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4217 s.vars[memVar] = s.newValue3(ssa.OpAtomicStore32, types.TypeMem, args[0], args[1], s.mem())
4220 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4221 addF("runtime/internal/atomic", "Store8",
4222 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4223 s.vars[memVar] = s.newValue3(ssa.OpAtomicStore8, types.TypeMem, args[0], args[1], s.mem())
4226 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4227 addF("runtime/internal/atomic", "Store64",
4228 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4229 s.vars[memVar] = s.newValue3(ssa.OpAtomicStore64, types.TypeMem, args[0], args[1], s.mem())
4232 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4233 addF("runtime/internal/atomic", "StorepNoWB",
4234 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4235 s.vars[memVar] = s.newValue3(ssa.OpAtomicStorePtrNoWB, types.TypeMem, args[0], args[1], s.mem())
4238 sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.RISCV64, sys.S390X)
4239 addF("runtime/internal/atomic", "StoreRel",
4240 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4241 s.vars[memVar] = s.newValue3(ssa.OpAtomicStoreRel32, types.TypeMem, args[0], args[1], s.mem())
4244 sys.PPC64, sys.S390X)
4245 addF("runtime/internal/atomic", "StoreRel64",
4246 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4247 s.vars[memVar] = s.newValue3(ssa.OpAtomicStoreRel64, types.TypeMem, args[0], args[1], s.mem())
4252 addF("runtime/internal/atomic", "Xchg",
4253 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4254 v := s.newValue3(ssa.OpAtomicExchange32, types.NewTuple(types.Types[types.TUINT32], types.TypeMem), args[0], args[1], s.mem())
4255 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4256 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT32], v)
4258 sys.AMD64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4259 addF("runtime/internal/atomic", "Xchg64",
4260 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4261 v := s.newValue3(ssa.OpAtomicExchange64, types.NewTuple(types.Types[types.TUINT64], types.TypeMem), args[0], args[1], s.mem())
4262 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4263 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT64], v)
4265 sys.AMD64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4267 type atomicOpEmitter func(s *state, n *ir.CallExpr, args []*ssa.Value, op ssa.Op, typ types.Kind)
4269 makeAtomicGuardedIntrinsicARM64 := func(op0, op1 ssa.Op, typ, rtyp types.Kind, emit atomicOpEmitter) intrinsicBuilder {
4271 return func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4272 // Target Atomic feature is identified by dynamic detection
4273 addr := s.entryNewValue1A(ssa.OpAddr, types.Types[types.TBOOL].PtrTo(), ir.Syms.ARM64HasATOMICS, s.sb)
4274 v := s.load(types.Types[types.TBOOL], addr)
4276 b.Kind = ssa.BlockIf
4278 bTrue := s.f.NewBlock(ssa.BlockPlain)
4279 bFalse := s.f.NewBlock(ssa.BlockPlain)
4280 bEnd := s.f.NewBlock(ssa.BlockPlain)
4283 b.Likely = ssa.BranchLikely
4285 // We have atomic instructions - use it directly.
4287 emit(s, n, args, op1, typ)
4288 s.endBlock().AddEdgeTo(bEnd)
4290 // Use original instruction sequence.
4291 s.startBlock(bFalse)
4292 emit(s, n, args, op0, typ)
4293 s.endBlock().AddEdgeTo(bEnd)
4297 if rtyp == types.TNIL {
4300 return s.variable(n, types.Types[rtyp])
4305 atomicXchgXaddEmitterARM64 := func(s *state, n *ir.CallExpr, args []*ssa.Value, op ssa.Op, typ types.Kind) {
4306 v := s.newValue3(op, types.NewTuple(types.Types[typ], types.TypeMem), args[0], args[1], s.mem())
4307 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4308 s.vars[n] = s.newValue1(ssa.OpSelect0, types.Types[typ], v)
4310 addF("runtime/internal/atomic", "Xchg",
4311 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicExchange32, ssa.OpAtomicExchange32Variant, types.TUINT32, types.TUINT32, atomicXchgXaddEmitterARM64),
4313 addF("runtime/internal/atomic", "Xchg64",
4314 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicExchange64, ssa.OpAtomicExchange64Variant, types.TUINT64, types.TUINT64, atomicXchgXaddEmitterARM64),
4317 addF("runtime/internal/atomic", "Xadd",
4318 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4319 v := s.newValue3(ssa.OpAtomicAdd32, types.NewTuple(types.Types[types.TUINT32], types.TypeMem), args[0], args[1], s.mem())
4320 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4321 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT32], v)
4323 sys.AMD64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4324 addF("runtime/internal/atomic", "Xadd64",
4325 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4326 v := s.newValue3(ssa.OpAtomicAdd64, types.NewTuple(types.Types[types.TUINT64], types.TypeMem), args[0], args[1], s.mem())
4327 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4328 return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT64], v)
4330 sys.AMD64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4332 addF("runtime/internal/atomic", "Xadd",
4333 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAdd32, ssa.OpAtomicAdd32Variant, types.TUINT32, types.TUINT32, atomicXchgXaddEmitterARM64),
4335 addF("runtime/internal/atomic", "Xadd64",
4336 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAdd64, ssa.OpAtomicAdd64Variant, types.TUINT64, types.TUINT64, atomicXchgXaddEmitterARM64),
4339 addF("runtime/internal/atomic", "Cas",
4340 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4341 v := s.newValue4(ssa.OpAtomicCompareAndSwap32, types.NewTuple(types.Types[types.TBOOL], types.TypeMem), args[0], args[1], args[2], s.mem())
4342 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4343 return s.newValue1(ssa.OpSelect0, types.Types[types.TBOOL], v)
4345 sys.AMD64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4346 addF("runtime/internal/atomic", "Cas64",
4347 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4348 v := s.newValue4(ssa.OpAtomicCompareAndSwap64, types.NewTuple(types.Types[types.TBOOL], types.TypeMem), args[0], args[1], args[2], s.mem())
4349 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4350 return s.newValue1(ssa.OpSelect0, types.Types[types.TBOOL], v)
4352 sys.AMD64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4353 addF("runtime/internal/atomic", "CasRel",
4354 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4355 v := s.newValue4(ssa.OpAtomicCompareAndSwap32, types.NewTuple(types.Types[types.TBOOL], types.TypeMem), args[0], args[1], args[2], s.mem())
4356 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4357 return s.newValue1(ssa.OpSelect0, types.Types[types.TBOOL], v)
4361 atomicCasEmitterARM64 := func(s *state, n *ir.CallExpr, args []*ssa.Value, op ssa.Op, typ types.Kind) {
4362 v := s.newValue4(op, types.NewTuple(types.Types[types.TBOOL], types.TypeMem), args[0], args[1], args[2], s.mem())
4363 s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
4364 s.vars[n] = s.newValue1(ssa.OpSelect0, types.Types[typ], v)
4367 addF("runtime/internal/atomic", "Cas",
4368 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicCompareAndSwap32, ssa.OpAtomicCompareAndSwap32Variant, types.TUINT32, types.TBOOL, atomicCasEmitterARM64),
4370 addF("runtime/internal/atomic", "Cas64",
4371 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicCompareAndSwap64, ssa.OpAtomicCompareAndSwap64Variant, types.TUINT64, types.TBOOL, atomicCasEmitterARM64),
4374 addF("runtime/internal/atomic", "And8",
4375 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4376 s.vars[memVar] = s.newValue3(ssa.OpAtomicAnd8, types.TypeMem, args[0], args[1], s.mem())
4379 sys.AMD64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4380 addF("runtime/internal/atomic", "And",
4381 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4382 s.vars[memVar] = s.newValue3(ssa.OpAtomicAnd32, types.TypeMem, args[0], args[1], s.mem())
4385 sys.AMD64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4386 addF("runtime/internal/atomic", "Or8",
4387 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4388 s.vars[memVar] = s.newValue3(ssa.OpAtomicOr8, types.TypeMem, args[0], args[1], s.mem())
4391 sys.AMD64, sys.ARM64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4392 addF("runtime/internal/atomic", "Or",
4393 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4394 s.vars[memVar] = s.newValue3(ssa.OpAtomicOr32, types.TypeMem, args[0], args[1], s.mem())
4397 sys.AMD64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
4399 atomicAndOrEmitterARM64 := func(s *state, n *ir.CallExpr, args []*ssa.Value, op ssa.Op, typ types.Kind) {
4400 s.vars[memVar] = s.newValue3(op, types.TypeMem, args[0], args[1], s.mem())
4403 addF("runtime/internal/atomic", "And8",
4404 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAnd8, ssa.OpAtomicAnd8Variant, types.TNIL, types.TNIL, atomicAndOrEmitterARM64),
4406 addF("runtime/internal/atomic", "And",
4407 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAnd32, ssa.OpAtomicAnd32Variant, types.TNIL, types.TNIL, atomicAndOrEmitterARM64),
4409 addF("runtime/internal/atomic", "Or8",
4410 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicOr8, ssa.OpAtomicOr8Variant, types.TNIL, types.TNIL, atomicAndOrEmitterARM64),
4412 addF("runtime/internal/atomic", "Or",
4413 makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicOr32, ssa.OpAtomicOr32Variant, types.TNIL, types.TNIL, atomicAndOrEmitterARM64),
4416 // Aliases for atomic load operations
4417 alias("runtime/internal/atomic", "Loadint32", "runtime/internal/atomic", "Load", all...)
4418 alias("runtime/internal/atomic", "Loadint64", "runtime/internal/atomic", "Load64", all...)
4419 alias("runtime/internal/atomic", "Loaduintptr", "runtime/internal/atomic", "Load", p4...)
4420 alias("runtime/internal/atomic", "Loaduintptr", "runtime/internal/atomic", "Load64", p8...)
4421 alias("runtime/internal/atomic", "Loaduint", "runtime/internal/atomic", "Load", p4...)
4422 alias("runtime/internal/atomic", "Loaduint", "runtime/internal/atomic", "Load64", p8...)
4423 alias("runtime/internal/atomic", "LoadAcq", "runtime/internal/atomic", "Load", lwatomics...)
4424 alias("runtime/internal/atomic", "LoadAcq64", "runtime/internal/atomic", "Load64", lwatomics...)
4425 alias("runtime/internal/atomic", "LoadAcquintptr", "runtime/internal/atomic", "LoadAcq", p4...)
4426 alias("sync", "runtime_LoadAcquintptr", "runtime/internal/atomic", "LoadAcq", p4...) // linknamed
4427 alias("runtime/internal/atomic", "LoadAcquintptr", "runtime/internal/atomic", "LoadAcq64", p8...)
4428 alias("sync", "runtime_LoadAcquintptr", "runtime/internal/atomic", "LoadAcq64", p8...) // linknamed
4430 // Aliases for atomic store operations
4431 alias("runtime/internal/atomic", "Storeint32", "runtime/internal/atomic", "Store", all...)
4432 alias("runtime/internal/atomic", "Storeint64", "runtime/internal/atomic", "Store64", all...)
4433 alias("runtime/internal/atomic", "Storeuintptr", "runtime/internal/atomic", "Store", p4...)
4434 alias("runtime/internal/atomic", "Storeuintptr", "runtime/internal/atomic", "Store64", p8...)
4435 alias("runtime/internal/atomic", "StoreRel", "runtime/internal/atomic", "Store", lwatomics...)
4436 alias("runtime/internal/atomic", "StoreRel64", "runtime/internal/atomic", "Store64", lwatomics...)
4437 alias("runtime/internal/atomic", "StoreReluintptr", "runtime/internal/atomic", "StoreRel", p4...)
4438 alias("sync", "runtime_StoreReluintptr", "runtime/internal/atomic", "StoreRel", p4...) // linknamed
4439 alias("runtime/internal/atomic", "StoreReluintptr", "runtime/internal/atomic", "StoreRel64", p8...)
4440 alias("sync", "runtime_StoreReluintptr", "runtime/internal/atomic", "StoreRel64", p8...) // linknamed
4442 // Aliases for atomic swap operations
4443 alias("runtime/internal/atomic", "Xchgint32", "runtime/internal/atomic", "Xchg", all...)
4444 alias("runtime/internal/atomic", "Xchgint64", "runtime/internal/atomic", "Xchg64", all...)
4445 alias("runtime/internal/atomic", "Xchguintptr", "runtime/internal/atomic", "Xchg", p4...)
4446 alias("runtime/internal/atomic", "Xchguintptr", "runtime/internal/atomic", "Xchg64", p8...)
4448 // Aliases for atomic add operations
4449 alias("runtime/internal/atomic", "Xaddint32", "runtime/internal/atomic", "Xadd", all...)
4450 alias("runtime/internal/atomic", "Xaddint64", "runtime/internal/atomic", "Xadd64", all...)
4451 alias("runtime/internal/atomic", "Xadduintptr", "runtime/internal/atomic", "Xadd", p4...)
4452 alias("runtime/internal/atomic", "Xadduintptr", "runtime/internal/atomic", "Xadd64", p8...)
4454 // Aliases for atomic CAS operations
4455 alias("runtime/internal/atomic", "Casint32", "runtime/internal/atomic", "Cas", all...)
4456 alias("runtime/internal/atomic", "Casint64", "runtime/internal/atomic", "Cas64", all...)
4457 alias("runtime/internal/atomic", "Casuintptr", "runtime/internal/atomic", "Cas", p4...)
4458 alias("runtime/internal/atomic", "Casuintptr", "runtime/internal/atomic", "Cas64", p8...)
4459 alias("runtime/internal/atomic", "Casp1", "runtime/internal/atomic", "Cas", p4...)
4460 alias("runtime/internal/atomic", "Casp1", "runtime/internal/atomic", "Cas64", p8...)
4461 alias("runtime/internal/atomic", "CasRel", "runtime/internal/atomic", "Cas", lwatomics...)
4463 /******** math ********/
4464 addF("math", "sqrt",
4465 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4466 return s.newValue1(ssa.OpSqrt, types.Types[types.TFLOAT64], args[0])
4468 sys.I386, sys.AMD64, sys.ARM, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X, sys.Wasm)
4469 addF("math", "Trunc",
4470 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4471 return s.newValue1(ssa.OpTrunc, types.Types[types.TFLOAT64], args[0])
4473 sys.ARM64, sys.PPC64, sys.S390X, sys.Wasm)
4474 addF("math", "Ceil",
4475 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4476 return s.newValue1(ssa.OpCeil, types.Types[types.TFLOAT64], args[0])
4478 sys.ARM64, sys.PPC64, sys.S390X, sys.Wasm)
4479 addF("math", "Floor",
4480 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4481 return s.newValue1(ssa.OpFloor, types.Types[types.TFLOAT64], args[0])
4483 sys.ARM64, sys.PPC64, sys.S390X, sys.Wasm)
4484 addF("math", "Round",
4485 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4486 return s.newValue1(ssa.OpRound, types.Types[types.TFLOAT64], args[0])
4488 sys.ARM64, sys.PPC64, sys.S390X)
4489 addF("math", "RoundToEven",
4490 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4491 return s.newValue1(ssa.OpRoundToEven, types.Types[types.TFLOAT64], args[0])
4493 sys.ARM64, sys.S390X, sys.Wasm)
4495 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4496 return s.newValue1(ssa.OpAbs, types.Types[types.TFLOAT64], args[0])
4498 sys.ARM64, sys.ARM, sys.PPC64, sys.RISCV64, sys.Wasm, sys.MIPS, sys.MIPS64)
4499 addF("math", "Copysign",
4500 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4501 return s.newValue2(ssa.OpCopysign, types.Types[types.TFLOAT64], args[0], args[1])
4503 sys.PPC64, sys.RISCV64, sys.Wasm)
4505 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4506 return s.newValue3(ssa.OpFMA, types.Types[types.TFLOAT64], args[0], args[1], args[2])
4508 sys.ARM64, sys.PPC64, sys.RISCV64, sys.S390X)
4510 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4511 if !s.config.UseFMA {
4512 s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
4513 return s.variable(n, types.Types[types.TFLOAT64])
4516 if buildcfg.GOAMD64 >= 3 {
4517 return s.newValue3(ssa.OpFMA, types.Types[types.TFLOAT64], args[0], args[1], args[2])
4520 v := s.entryNewValue0A(ssa.OpHasCPUFeature, types.Types[types.TBOOL], ir.Syms.X86HasFMA)
4522 b.Kind = ssa.BlockIf
4524 bTrue := s.f.NewBlock(ssa.BlockPlain)
4525 bFalse := s.f.NewBlock(ssa.BlockPlain)
4526 bEnd := s.f.NewBlock(ssa.BlockPlain)
4529 b.Likely = ssa.BranchLikely // >= haswell cpus are common
4531 // We have the intrinsic - use it directly.
4533 s.vars[n] = s.newValue3(ssa.OpFMA, types.Types[types.TFLOAT64], args[0], args[1], args[2])
4534 s.endBlock().AddEdgeTo(bEnd)
4536 // Call the pure Go version.
4537 s.startBlock(bFalse)
4538 s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
4539 s.endBlock().AddEdgeTo(bEnd)
4543 return s.variable(n, types.Types[types.TFLOAT64])
4547 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4548 if !s.config.UseFMA {
4549 s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
4550 return s.variable(n, types.Types[types.TFLOAT64])
4552 addr := s.entryNewValue1A(ssa.OpAddr, types.Types[types.TBOOL].PtrTo(), ir.Syms.ARMHasVFPv4, s.sb)
4553 v := s.load(types.Types[types.TBOOL], addr)
4555 b.Kind = ssa.BlockIf
4557 bTrue := s.f.NewBlock(ssa.BlockPlain)
4558 bFalse := s.f.NewBlock(ssa.BlockPlain)
4559 bEnd := s.f.NewBlock(ssa.BlockPlain)
4562 b.Likely = ssa.BranchLikely
4564 // We have the intrinsic - use it directly.
4566 s.vars[n] = s.newValue3(ssa.OpFMA, types.Types[types.TFLOAT64], args[0], args[1], args[2])
4567 s.endBlock().AddEdgeTo(bEnd)
4569 // Call the pure Go version.
4570 s.startBlock(bFalse)
4571 s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
4572 s.endBlock().AddEdgeTo(bEnd)
4576 return s.variable(n, types.Types[types.TFLOAT64])
4580 makeRoundAMD64 := func(op ssa.Op) func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4581 return func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4582 if buildcfg.GOAMD64 >= 2 {
4583 return s.newValue1(op, types.Types[types.TFLOAT64], args[0])
4586 v := s.entryNewValue0A(ssa.OpHasCPUFeature, types.Types[types.TBOOL], ir.Syms.X86HasSSE41)
4588 b.Kind = ssa.BlockIf
4590 bTrue := s.f.NewBlock(ssa.BlockPlain)
4591 bFalse := s.f.NewBlock(ssa.BlockPlain)
4592 bEnd := s.f.NewBlock(ssa.BlockPlain)
4595 b.Likely = ssa.BranchLikely // most machines have sse4.1 nowadays
4597 // We have the intrinsic - use it directly.
4599 s.vars[n] = s.newValue1(op, types.Types[types.TFLOAT64], args[0])
4600 s.endBlock().AddEdgeTo(bEnd)
4602 // Call the pure Go version.
4603 s.startBlock(bFalse)
4604 s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
4605 s.endBlock().AddEdgeTo(bEnd)
4609 return s.variable(n, types.Types[types.TFLOAT64])
4612 addF("math", "RoundToEven",
4613 makeRoundAMD64(ssa.OpRoundToEven),
4615 addF("math", "Floor",
4616 makeRoundAMD64(ssa.OpFloor),
4618 addF("math", "Ceil",
4619 makeRoundAMD64(ssa.OpCeil),
4621 addF("math", "Trunc",
4622 makeRoundAMD64(ssa.OpTrunc),
4625 /******** math/bits ********/
4626 addF("math/bits", "TrailingZeros64",
4627 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4628 return s.newValue1(ssa.OpCtz64, types.Types[types.TINT], args[0])
4630 sys.AMD64, sys.I386, sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4631 addF("math/bits", "TrailingZeros32",
4632 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4633 return s.newValue1(ssa.OpCtz32, types.Types[types.TINT], args[0])
4635 sys.AMD64, sys.I386, sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4636 addF("math/bits", "TrailingZeros16",
4637 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4638 x := s.newValue1(ssa.OpZeroExt16to32, types.Types[types.TUINT32], args[0])
4639 c := s.constInt32(types.Types[types.TUINT32], 1<<16)
4640 y := s.newValue2(ssa.OpOr32, types.Types[types.TUINT32], x, c)
4641 return s.newValue1(ssa.OpCtz32, types.Types[types.TINT], y)
4644 addF("math/bits", "TrailingZeros16",
4645 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4646 return s.newValue1(ssa.OpCtz16, types.Types[types.TINT], args[0])
4648 sys.AMD64, sys.I386, sys.ARM, sys.ARM64, sys.Wasm)
4649 addF("math/bits", "TrailingZeros16",
4650 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4651 x := s.newValue1(ssa.OpZeroExt16to64, types.Types[types.TUINT64], args[0])
4652 c := s.constInt64(types.Types[types.TUINT64], 1<<16)
4653 y := s.newValue2(ssa.OpOr64, types.Types[types.TUINT64], x, c)
4654 return s.newValue1(ssa.OpCtz64, types.Types[types.TINT], y)
4656 sys.S390X, sys.PPC64)
4657 addF("math/bits", "TrailingZeros8",
4658 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4659 x := s.newValue1(ssa.OpZeroExt8to32, types.Types[types.TUINT32], args[0])
4660 c := s.constInt32(types.Types[types.TUINT32], 1<<8)
4661 y := s.newValue2(ssa.OpOr32, types.Types[types.TUINT32], x, c)
4662 return s.newValue1(ssa.OpCtz32, types.Types[types.TINT], y)
4665 addF("math/bits", "TrailingZeros8",
4666 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4667 return s.newValue1(ssa.OpCtz8, types.Types[types.TINT], args[0])
4669 sys.AMD64, sys.I386, sys.ARM, sys.ARM64, sys.Wasm)
4670 addF("math/bits", "TrailingZeros8",
4671 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4672 x := s.newValue1(ssa.OpZeroExt8to64, types.Types[types.TUINT64], args[0])
4673 c := s.constInt64(types.Types[types.TUINT64], 1<<8)
4674 y := s.newValue2(ssa.OpOr64, types.Types[types.TUINT64], x, c)
4675 return s.newValue1(ssa.OpCtz64, types.Types[types.TINT], y)
4678 alias("math/bits", "ReverseBytes64", "runtime/internal/sys", "Bswap64", all...)
4679 alias("math/bits", "ReverseBytes32", "runtime/internal/sys", "Bswap32", all...)
4680 // ReverseBytes inlines correctly, no need to intrinsify it.
4681 // Nothing special is needed for targets where ReverseBytes16 lowers to a rotate
4682 // On Power10, 16-bit rotate is not available so use BRH instruction
4683 if buildcfg.GOPPC64 >= 10 {
4684 addF("math/bits", "ReverseBytes16",
4685 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4686 return s.newValue1(ssa.OpBswap16, types.Types[types.TUINT], args[0])
4691 addF("math/bits", "Len64",
4692 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4693 return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], args[0])
4695 sys.AMD64, sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4696 addF("math/bits", "Len32",
4697 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4698 return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], args[0])
4700 sys.AMD64, sys.ARM64, sys.PPC64)
4701 addF("math/bits", "Len32",
4702 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4703 if s.config.PtrSize == 4 {
4704 return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], args[0])
4706 x := s.newValue1(ssa.OpZeroExt32to64, types.Types[types.TUINT64], args[0])
4707 return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], x)
4709 sys.ARM, sys.S390X, sys.MIPS, sys.Wasm)
4710 addF("math/bits", "Len16",
4711 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4712 if s.config.PtrSize == 4 {
4713 x := s.newValue1(ssa.OpZeroExt16to32, types.Types[types.TUINT32], args[0])
4714 return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], x)
4716 x := s.newValue1(ssa.OpZeroExt16to64, types.Types[types.TUINT64], args[0])
4717 return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], x)
4719 sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4720 addF("math/bits", "Len16",
4721 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4722 return s.newValue1(ssa.OpBitLen16, types.Types[types.TINT], args[0])
4725 addF("math/bits", "Len8",
4726 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4727 if s.config.PtrSize == 4 {
4728 x := s.newValue1(ssa.OpZeroExt8to32, types.Types[types.TUINT32], args[0])
4729 return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], x)
4731 x := s.newValue1(ssa.OpZeroExt8to64, types.Types[types.TUINT64], args[0])
4732 return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], x)
4734 sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4735 addF("math/bits", "Len8",
4736 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4737 return s.newValue1(ssa.OpBitLen8, types.Types[types.TINT], args[0])
4740 addF("math/bits", "Len",
4741 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4742 if s.config.PtrSize == 4 {
4743 return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], args[0])
4745 return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], args[0])
4747 sys.AMD64, sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
4748 // LeadingZeros is handled because it trivially calls Len.
4749 addF("math/bits", "Reverse64",
4750 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4751 return s.newValue1(ssa.OpBitRev64, types.Types[types.TINT], args[0])
4754 addF("math/bits", "Reverse32",
4755 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4756 return s.newValue1(ssa.OpBitRev32, types.Types[types.TINT], args[0])
4759 addF("math/bits", "Reverse16",
4760 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4761 return s.newValue1(ssa.OpBitRev16, types.Types[types.TINT], args[0])
4764 addF("math/bits", "Reverse8",
4765 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4766 return s.newValue1(ssa.OpBitRev8, types.Types[types.TINT], args[0])
4769 addF("math/bits", "Reverse",
4770 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4771 return s.newValue1(ssa.OpBitRev64, types.Types[types.TINT], args[0])
4774 addF("math/bits", "RotateLeft8",
4775 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4776 return s.newValue2(ssa.OpRotateLeft8, types.Types[types.TUINT8], args[0], args[1])
4779 addF("math/bits", "RotateLeft16",
4780 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4781 return s.newValue2(ssa.OpRotateLeft16, types.Types[types.TUINT16], args[0], args[1])
4784 addF("math/bits", "RotateLeft32",
4785 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4786 return s.newValue2(ssa.OpRotateLeft32, types.Types[types.TUINT32], args[0], args[1])
4788 sys.AMD64, sys.ARM, sys.ARM64, sys.S390X, sys.PPC64, sys.Wasm, sys.Loong64)
4789 addF("math/bits", "RotateLeft64",
4790 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4791 return s.newValue2(ssa.OpRotateLeft64, types.Types[types.TUINT64], args[0], args[1])
4793 sys.AMD64, sys.ARM64, sys.S390X, sys.PPC64, sys.Wasm, sys.Loong64)
4794 alias("math/bits", "RotateLeft", "math/bits", "RotateLeft64", p8...)
4796 makeOnesCountAMD64 := func(op ssa.Op) func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4797 return func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4798 if buildcfg.GOAMD64 >= 2 {
4799 return s.newValue1(op, types.Types[types.TINT], args[0])
4802 v := s.entryNewValue0A(ssa.OpHasCPUFeature, types.Types[types.TBOOL], ir.Syms.X86HasPOPCNT)
4804 b.Kind = ssa.BlockIf
4806 bTrue := s.f.NewBlock(ssa.BlockPlain)
4807 bFalse := s.f.NewBlock(ssa.BlockPlain)
4808 bEnd := s.f.NewBlock(ssa.BlockPlain)
4811 b.Likely = ssa.BranchLikely // most machines have popcnt nowadays
4813 // We have the intrinsic - use it directly.
4815 s.vars[n] = s.newValue1(op, types.Types[types.TINT], args[0])
4816 s.endBlock().AddEdgeTo(bEnd)
4818 // Call the pure Go version.
4819 s.startBlock(bFalse)
4820 s.vars[n] = s.callResult(n, callNormal) // types.Types[TINT]
4821 s.endBlock().AddEdgeTo(bEnd)
4825 return s.variable(n, types.Types[types.TINT])
4828 addF("math/bits", "OnesCount64",
4829 makeOnesCountAMD64(ssa.OpPopCount64),
4831 addF("math/bits", "OnesCount64",
4832 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4833 return s.newValue1(ssa.OpPopCount64, types.Types[types.TINT], args[0])
4835 sys.PPC64, sys.ARM64, sys.S390X, sys.Wasm)
4836 addF("math/bits", "OnesCount32",
4837 makeOnesCountAMD64(ssa.OpPopCount32),
4839 addF("math/bits", "OnesCount32",
4840 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4841 return s.newValue1(ssa.OpPopCount32, types.Types[types.TINT], args[0])
4843 sys.PPC64, sys.ARM64, sys.S390X, sys.Wasm)
4844 addF("math/bits", "OnesCount16",
4845 makeOnesCountAMD64(ssa.OpPopCount16),
4847 addF("math/bits", "OnesCount16",
4848 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4849 return s.newValue1(ssa.OpPopCount16, types.Types[types.TINT], args[0])
4851 sys.ARM64, sys.S390X, sys.PPC64, sys.Wasm)
4852 addF("math/bits", "OnesCount8",
4853 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4854 return s.newValue1(ssa.OpPopCount8, types.Types[types.TINT], args[0])
4856 sys.S390X, sys.PPC64, sys.Wasm)
4857 addF("math/bits", "OnesCount",
4858 makeOnesCountAMD64(ssa.OpPopCount64),
4860 addF("math/bits", "Mul64",
4861 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4862 return s.newValue2(ssa.OpMul64uhilo, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1])
4864 sys.AMD64, sys.ARM64, sys.PPC64, sys.S390X, sys.MIPS64, sys.RISCV64, sys.Loong64)
4865 alias("math/bits", "Mul", "math/bits", "Mul64", p8...)
4866 alias("runtime/internal/math", "Mul64", "math/bits", "Mul64", p8...)
4867 addF("math/bits", "Add64",
4868 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4869 return s.newValue3(ssa.OpAdd64carry, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1], args[2])
4871 sys.AMD64, sys.ARM64, sys.PPC64, sys.S390X, sys.RISCV64, sys.Loong64, sys.MIPS64)
4872 alias("math/bits", "Add", "math/bits", "Add64", p8...)
4873 alias("runtime/internal/math", "Add64", "math/bits", "Add64", all...)
4874 addF("math/bits", "Sub64",
4875 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4876 return s.newValue3(ssa.OpSub64borrow, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1], args[2])
4878 sys.AMD64, sys.ARM64, sys.PPC64, sys.S390X, sys.RISCV64, sys.Loong64, sys.MIPS64)
4879 alias("math/bits", "Sub", "math/bits", "Sub64", p8...)
4880 addF("math/bits", "Div64",
4881 func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
4882 // check for divide-by-zero/overflow and panic with appropriate message
4883 cmpZero := s.newValue2(s.ssaOp(ir.ONE, types.Types[types.TUINT64]), types.Types[types.TBOOL], args[2], s.zeroVal(types.Types[types.TUINT64]))
4884 s.check(cmpZero, ir.Syms.Panicdivide)
4885 cmpOverflow := s.newValue2(s.ssaOp(ir.OLT, types.Types[types.TUINT64]), types.Types[types.TBOOL], args[0], args[2])
4886 s.check(cmpOverflow, ir.Syms.Panicoverflow)
4887 return s.newValue3(ssa.OpDiv128u, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1], args[2])
4890 alias("math/bits", "Div", "math/bits", "Div64", sys.ArchAMD64)
4892 alias("runtime/internal/sys", "TrailingZeros8", "math/bits", "TrailingZeros8", all...)
4893 alias("runtime/internal/sys", "TrailingZeros32", "math/bits", "TrailingZeros32", all...)
4894 alias("runtime/internal/sys", "TrailingZeros64", "math/bits", "TrailingZeros64", all...)
4895 alias("runtime/internal/sys", "Len8", "math/bits", "Len8", all...)
4896 alias("runtime/internal/sys", "Len64", "math/bits", "Len64", all...)
4897 alias("runtime/internal/sys", "OnesCount64", "math/bits", "OnesCount64", all...)
4899 /******** sync/atomic ********/
4901 // Note: these are disabled by flag_race in findIntrinsic below.
4902 alias("sync/atomic", "LoadInt32", "runtime/internal/atomic", "Load", all...)
4903 alias("sync/atomic", "LoadInt64", "runtime/internal/atomic", "Load64", all...)
4904 alias("sync/atomic", "LoadPointer", "runtime/internal/atomic", "Loadp", all...)
4905 alias("sync/atomic", "LoadUint32", "runtime/internal/atomic", "Load", all...)
4906 alias("sync/atomic", "LoadUint64", "runtime/internal/atomic", "Load64", all...)
4907 alias("sync/atomic", "LoadUintptr", "runtime/internal/atomic", "Load", p4...)
4908 alias("sync/atomic", "LoadUintptr", "runtime/internal/atomic", "Load64", p8...)
4910 alias("sync/atomic", "StoreInt32", "runtime/internal/atomic", "Store", all...)
4911 alias("sync/atomic", "StoreInt64", "runtime/internal/atomic", "Store64", all...)
4912 // Note: not StorePointer, that needs a write barrier. Same below for {CompareAnd}Swap.
4913 alias("sync/atomic", "StoreUint32", "runtime/internal/atomic", "Store", all...)
4914 alias("sync/atomic", "StoreUint64", "runtime/internal/atomic", "Store64", all...)
4915 alias("sync/atomic", "StoreUintptr", "runtime/internal/atomic", "Store", p4...)
4916 alias("sync/atomic", "StoreUintptr", "runtime/internal/atomic", "Store64", p8...)
4918 alias("sync/atomic", "SwapInt32", "runtime/internal/atomic", "Xchg", all...)
4919 alias("sync/atomic", "SwapInt64", "runtime/internal/atomic", "Xchg64", all...)
4920 alias("sync/atomic", "SwapUint32", "runtime/internal/atomic", "Xchg", all...)
4921 alias("sync/atomic", "SwapUint64", "runtime/internal/atomic", "Xchg64", all...)
4922 alias("sync/atomic", "SwapUintptr", "runtime/internal/atomic", "Xchg", p4...)
4923 alias("sync/atomic", "SwapUintptr", "runtime/internal/atomic", "Xchg64", p8...)
4925 alias("sync/atomic", "CompareAndSwapInt32", "runtime/internal/atomic", "Cas", all...)
4926 alias("sync/atomic", "CompareAndSwapInt64", "runtime/internal/atomic", "Cas64", all...)
4927 alias("sync/atomic", "CompareAndSwapUint32", "runtime/internal/atomic", "Cas", all...)
4928 alias("sync/atomic", "CompareAndSwapUint64", "runtime/internal/atomic", "Cas64", all...)
4929 alias("sync/atomic", "CompareAndSwapUintptr", "runtime/internal/atomic", "Cas", p4...)
4930 alias("sync/atomic", "CompareAndSwapUintptr", "runtime/internal/atomic", "Cas64", p8...)
4932 alias("sync/atomic", "AddInt32", "runtime/internal/atomic", "Xadd", all...)
4933 alias("sync/atomic", "AddInt64", "runtime/internal/atomic", "Xadd64", all...)
4934 alias("sync/atomic", "AddUint32", "runtime/internal/atomic", "Xadd", all...)
4935 alias("sync/atomic", "AddUint64", "runtime/internal/atomic", "Xadd64", all...)
4936 alias("sync/atomic", "AddUintptr", "runtime/internal/atomic", "Xadd", p4...)
4937 alias("sync/atomic", "AddUintptr", "runtime/internal/atomic", "Xadd64", p8...)
4939 /******** math/big ********/
4940 alias("math/big", "mulWW", "math/bits", "Mul64", p8...)
4943 // findIntrinsic returns a function which builds the SSA equivalent of the
4944 // function identified by the symbol sym. If sym is not an intrinsic call, returns nil.
4945 func findIntrinsic(sym *types.Sym) intrinsicBuilder {
4946 if sym == nil || sym.Pkg == nil {
4950 if sym.Pkg == ir.Pkgs.Runtime {
4953 if base.Flag.Race && pkg == "sync/atomic" {
4954 // The race detector needs to be able to intercept these calls.
4955 // We can't intrinsify them.
4958 // Skip intrinsifying math functions (which may contain hard-float
4959 // instructions) when soft-float
4960 if Arch.SoftFloat && pkg == "math" {
4965 if ssa.IntrinsicsDisable {
4966 if pkg == "runtime" && (fn == "getcallerpc" || fn == "getcallersp" || fn == "getclosureptr") {
4967 // These runtime functions don't have definitions, must be intrinsics.
4972 return intrinsics[intrinsicKey{Arch.LinkArch.Arch, pkg, fn}]
4975 func IsIntrinsicCall(n *ir.CallExpr) bool {
4979 name, ok := n.X.(*ir.Name)
4983 return findIntrinsic(name.Sym()) != nil
4986 // intrinsicCall converts a call to a recognized intrinsic function into the intrinsic SSA operation.
4987 func (s *state) intrinsicCall(n *ir.CallExpr) *ssa.Value {
4988 v := findIntrinsic(n.X.Sym())(s, n, s.intrinsicArgs(n))
4989 if ssa.IntrinsicsDebug > 0 {
4994 if x.Op == ssa.OpSelect0 || x.Op == ssa.OpSelect1 {
4997 base.WarnfAt(n.Pos(), "intrinsic substitution for %v with %s", n.X.Sym().Name, x.LongString())
5002 // intrinsicArgs extracts args from n, evaluates them to SSA values, and returns them.
5003 func (s *state) intrinsicArgs(n *ir.CallExpr) []*ssa.Value {
5004 args := make([]*ssa.Value, len(n.Args))
5005 for i, n := range n.Args {
5011 // openDeferRecord adds code to evaluate and store the function for an open-code defer
5012 // call, and records info about the defer, so we can generate proper code on the
5013 // exit paths. n is the sub-node of the defer node that is the actual function
5014 // call. We will also record funcdata information on where the function is stored
5015 // (as well as the deferBits variable), and this will enable us to run the proper
5016 // defer calls during panics.
5017 func (s *state) openDeferRecord(n *ir.CallExpr) {
5018 if len(n.Args) != 0 || n.Op() != ir.OCALLFUNC || n.X.Type().NumResults() != 0 {
5019 s.Fatalf("defer call with arguments or results: %v", n)
5022 opendefer := &openDeferInfo{
5026 // We must always store the function value in a stack slot for the
5027 // runtime panic code to use. But in the defer exit code, we will
5028 // call the function directly if it is a static function.
5029 closureVal := s.expr(fn)
5030 closure := s.openDeferSave(fn.Type(), closureVal)
5031 opendefer.closureNode = closure.Aux.(*ir.Name)
5032 if !(fn.Op() == ir.ONAME && fn.(*ir.Name).Class == ir.PFUNC) {
5033 opendefer.closure = closure
5035 index := len(s.openDefers)
5036 s.openDefers = append(s.openDefers, opendefer)
5038 // Update deferBits only after evaluation and storage to stack of
5039 // the function is successful.
5040 bitvalue := s.constInt8(types.Types[types.TUINT8], 1<<uint(index))
5041 newDeferBits := s.newValue2(ssa.OpOr8, types.Types[types.TUINT8], s.variable(deferBitsVar, types.Types[types.TUINT8]), bitvalue)
5042 s.vars[deferBitsVar] = newDeferBits
5043 s.store(types.Types[types.TUINT8], s.deferBitsAddr, newDeferBits)
5046 // openDeferSave generates SSA nodes to store a value (with type t) for an
5047 // open-coded defer at an explicit autotmp location on the stack, so it can be
5048 // reloaded and used for the appropriate call on exit. Type t must be a function type
5049 // (therefore SSAable). val is the value to be stored. The function returns an SSA
5050 // value representing a pointer to the autotmp location.
5051 func (s *state) openDeferSave(t *types.Type, val *ssa.Value) *ssa.Value {
5053 s.Fatalf("openDeferSave of non-SSA-able type %v val=%v", t, val)
5055 if !t.HasPointers() {
5056 s.Fatalf("openDeferSave of pointerless type %v val=%v", t, val)
5059 temp := typecheck.TempAt(pos.WithNotStmt(), s.curfn, t)
5060 temp.SetOpenDeferSlot(true)
5061 temp.SetFrameOffset(int64(len(s.openDefers))) // so cmpstackvarlt can order them
5062 var addrTemp *ssa.Value
5063 // Use OpVarLive to make sure stack slot for the closure is not removed by
5064 // dead-store elimination
5065 if s.curBlock.ID != s.f.Entry.ID {
5066 // Force the tmp storing this defer function to be declared in the entry
5067 // block, so that it will be live for the defer exit code (which will
5068 // actually access it only if the associated defer call has been activated).
5069 if t.HasPointers() {
5070 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])
5072 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])
5073 addrTemp = s.f.Entry.NewValue2A(src.NoXPos, ssa.OpLocalAddr, types.NewPtr(temp.Type()), temp, s.sp, s.defvars[s.f.Entry.ID][memVar])
5075 // Special case if we're still in the entry block. We can't use
5076 // the above code, since s.defvars[s.f.Entry.ID] isn't defined
5077 // until we end the entry block with s.endBlock().
5078 if t.HasPointers() {
5079 s.vars[memVar] = s.newValue1Apos(ssa.OpVarDef, types.TypeMem, temp, s.mem(), false)
5081 s.vars[memVar] = s.newValue1Apos(ssa.OpVarLive, types.TypeMem, temp, s.mem(), false)
5082 addrTemp = s.newValue2Apos(ssa.OpLocalAddr, types.NewPtr(temp.Type()), temp, s.sp, s.mem(), false)
5084 // Since we may use this temp during exit depending on the
5085 // deferBits, we must define it unconditionally on entry.
5086 // Therefore, we must make sure it is zeroed out in the entry
5087 // block if it contains pointers, else GC may wrongly follow an
5088 // uninitialized pointer value.
5089 temp.SetNeedzero(true)
5090 // We are storing to the stack, hence we can avoid the full checks in
5091 // storeType() (no write barrier) and do a simple store().
5092 s.store(t, addrTemp, val)
5096 // openDeferExit generates SSA for processing all the open coded defers at exit.
5097 // The code involves loading deferBits, and checking each of the bits to see if
5098 // the corresponding defer statement was executed. For each bit that is turned
5099 // on, the associated defer call is made.
5100 func (s *state) openDeferExit() {
5101 deferExit := s.f.NewBlock(ssa.BlockPlain)
5102 s.endBlock().AddEdgeTo(deferExit)
5103 s.startBlock(deferExit)
5104 s.lastDeferExit = deferExit
5105 s.lastDeferCount = len(s.openDefers)
5106 zeroval := s.constInt8(types.Types[types.TUINT8], 0)
5107 // Test for and run defers in reverse order
5108 for i := len(s.openDefers) - 1; i >= 0; i-- {
5109 r := s.openDefers[i]
5110 bCond := s.f.NewBlock(ssa.BlockPlain)
5111 bEnd := s.f.NewBlock(ssa.BlockPlain)
5113 deferBits := s.variable(deferBitsVar, types.Types[types.TUINT8])
5114 // Generate code to check if the bit associated with the current
5116 bitval := s.constInt8(types.Types[types.TUINT8], 1<<uint(i))
5117 andval := s.newValue2(ssa.OpAnd8, types.Types[types.TUINT8], deferBits, bitval)
5118 eqVal := s.newValue2(ssa.OpEq8, types.Types[types.TBOOL], andval, zeroval)
5120 b.Kind = ssa.BlockIf
5124 bCond.AddEdgeTo(bEnd)
5127 // Clear this bit in deferBits and force store back to stack, so
5128 // we will not try to re-run this defer call if this defer call panics.
5129 nbitval := s.newValue1(ssa.OpCom8, types.Types[types.TUINT8], bitval)
5130 maskedval := s.newValue2(ssa.OpAnd8, types.Types[types.TUINT8], deferBits, nbitval)
5131 s.store(types.Types[types.TUINT8], s.deferBitsAddr, maskedval)
5132 // Use this value for following tests, so we keep previous
5134 s.vars[deferBitsVar] = maskedval
5136 // Generate code to call the function call of the defer, using the
5137 // closure that were stored in argtmps at the point of the defer
5140 stksize := fn.Type().ArgWidth()
5141 var callArgs []*ssa.Value
5143 if r.closure != nil {
5144 v := s.load(r.closure.Type.Elem(), r.closure)
5145 s.maybeNilCheckClosure(v, callDefer)
5146 codeptr := s.rawLoad(types.Types[types.TUINTPTR], v)
5147 aux := ssa.ClosureAuxCall(s.f.ABIDefault.ABIAnalyzeTypes(nil, nil))
5148 call = s.newValue2A(ssa.OpClosureLECall, aux.LateExpansionResultType(), aux, codeptr, v)
5150 aux := ssa.StaticAuxCall(fn.(*ir.Name).Linksym(), s.f.ABIDefault.ABIAnalyzeTypes(nil, nil))
5151 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
5153 callArgs = append(callArgs, s.mem())
5154 call.AddArgs(callArgs...)
5155 call.AuxInt = stksize
5156 s.vars[memVar] = s.newValue1I(ssa.OpSelectN, types.TypeMem, 0, call)
5157 // Make sure that the stack slots with pointers are kept live
5158 // through the call (which is a pre-emption point). Also, we will
5159 // use the first call of the last defer exit to compute liveness
5160 // for the deferreturn, so we want all stack slots to be live.
5161 if r.closureNode != nil {
5162 s.vars[memVar] = s.newValue1Apos(ssa.OpVarLive, types.TypeMem, r.closureNode, s.mem(), false)
5170 func (s *state) callResult(n *ir.CallExpr, k callKind) *ssa.Value {
5171 return s.call(n, k, false)
5174 func (s *state) callAddr(n *ir.CallExpr, k callKind) *ssa.Value {
5175 return s.call(n, k, true)
5178 // Calls the function n using the specified call type.
5179 // Returns the address of the return value (or nil if none).
5180 func (s *state) call(n *ir.CallExpr, k callKind, returnResultAddr bool) *ssa.Value {
5182 var callee *ir.Name // target function (if static)
5183 var closure *ssa.Value // ptr to closure to run (if dynamic)
5184 var codeptr *ssa.Value // ptr to target code (if dynamic)
5185 var rcvr *ssa.Value // receiver to set
5187 var ACArgs []*types.Type // AuxCall args
5188 var ACResults []*types.Type // AuxCall results
5189 var callArgs []*ssa.Value // For late-expansion, the args themselves (not stored, args to the call instead).
5191 callABI := s.f.ABIDefault
5193 if k != callNormal && k != callTail && (len(n.Args) != 0 || n.Op() == ir.OCALLINTER || n.X.Type().NumResults() != 0) {
5194 s.Fatalf("go/defer call with arguments: %v", n)
5199 if (k == callNormal || k == callTail) && fn.Op() == ir.ONAME && fn.(*ir.Name).Class == ir.PFUNC {
5202 if buildcfg.Experiment.RegabiArgs {
5203 // This is a static call, so it may be
5204 // a direct call to a non-ABIInternal
5205 // function. fn.Func may be nil for
5206 // some compiler-generated functions,
5207 // but those are all ABIInternal.
5209 callABI = abiForFunc(fn.Func, s.f.ABI0, s.f.ABI1)
5212 // TODO(register args) remove after register abi is working
5213 inRegistersImported := fn.Pragma()&ir.RegisterParams != 0
5214 inRegistersSamePackage := fn.Func != nil && fn.Func.Pragma&ir.RegisterParams != 0
5215 if inRegistersImported || inRegistersSamePackage {
5221 closure = s.expr(fn)
5222 if k != callDefer && k != callDeferStack {
5223 // Deferred nil function needs to panic when the function is invoked,
5224 // not the point of defer statement.
5225 s.maybeNilCheckClosure(closure, k)
5228 if fn.Op() != ir.ODOTINTER {
5229 s.Fatalf("OCALLINTER: n.Left not an ODOTINTER: %v", fn.Op())
5231 fn := fn.(*ir.SelectorExpr)
5232 var iclosure *ssa.Value
5233 iclosure, rcvr = s.getClosureAndRcvr(fn)
5234 if k == callNormal {
5235 codeptr = s.load(types.Types[types.TUINTPTR], iclosure)
5241 params := callABI.ABIAnalyze(n.X.Type(), false /* Do not set (register) nNames from caller side -- can cause races. */)
5242 types.CalcSize(fn.Type())
5243 stksize := params.ArgWidth() // includes receiver, args, and results
5245 res := n.X.Type().Results()
5246 if k == callNormal || k == callTail {
5247 for _, p := range params.OutParams() {
5248 ACResults = append(ACResults, p.Type)
5253 if k == callDeferStack {
5255 s.Fatalf("deferprocStack with non-zero stack size %d: %v", stksize, n)
5257 // Make a defer struct on the stack.
5259 _, addr := s.temp(n.Pos(), t)
5260 s.store(closure.Type,
5261 s.newValue1I(ssa.OpOffPtr, closure.Type.PtrTo(), t.FieldOff(deferStructFnField), addr),
5264 // Call runtime.deferprocStack with pointer to _defer record.
5265 ACArgs = append(ACArgs, types.Types[types.TUINTPTR])
5266 aux := ssa.StaticAuxCall(ir.Syms.DeferprocStack, s.f.ABIDefault.ABIAnalyzeTypes(ACArgs, ACResults))
5267 callArgs = append(callArgs, addr, s.mem())
5268 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
5269 call.AddArgs(callArgs...)
5270 call.AuxInt = int64(types.PtrSize) // deferprocStack takes a *_defer arg
5272 // Store arguments to stack, including defer/go arguments and receiver for method calls.
5273 // These are written in SP-offset order.
5274 argStart := base.Ctxt.Arch.FixedFrameSize
5276 if k != callNormal && k != callTail {
5277 // Write closure (arg to newproc/deferproc).
5278 ACArgs = append(ACArgs, types.Types[types.TUINTPTR]) // not argExtra
5279 callArgs = append(callArgs, closure)
5280 stksize += int64(types.PtrSize)
5281 argStart += int64(types.PtrSize)
5284 // Set receiver (for interface calls).
5286 callArgs = append(callArgs, rcvr)
5293 for _, p := range params.InParams() { // includes receiver for interface calls
5294 ACArgs = append(ACArgs, p.Type)
5297 // Split the entry block if there are open defers, because later calls to
5298 // openDeferSave may cause a mismatch between the mem for an OpDereference
5299 // and the call site which uses it. See #49282.
5300 if s.curBlock.ID == s.f.Entry.ID && s.hasOpenDefers {
5302 b.Kind = ssa.BlockPlain
5303 curb := s.f.NewBlock(ssa.BlockPlain)
5308 for i, n := range args {
5309 callArgs = append(callArgs, s.putArg(n, t.Param(i).Type))
5312 callArgs = append(callArgs, s.mem())
5316 case k == callDefer:
5317 aux := ssa.StaticAuxCall(ir.Syms.Deferproc, s.f.ABIDefault.ABIAnalyzeTypes(ACArgs, ACResults)) // TODO paramResultInfo for DeferProc
5318 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
5320 aux := ssa.StaticAuxCall(ir.Syms.Newproc, s.f.ABIDefault.ABIAnalyzeTypes(ACArgs, ACResults))
5321 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux) // TODO paramResultInfo for NewProc
5322 case closure != nil:
5323 // rawLoad because loading the code pointer from a
5324 // closure is always safe, but IsSanitizerSafeAddr
5325 // can't always figure that out currently, and it's
5326 // critical that we not clobber any arguments already
5327 // stored onto the stack.
5328 codeptr = s.rawLoad(types.Types[types.TUINTPTR], closure)
5329 aux := ssa.ClosureAuxCall(callABI.ABIAnalyzeTypes(ACArgs, ACResults))
5330 call = s.newValue2A(ssa.OpClosureLECall, aux.LateExpansionResultType(), aux, codeptr, closure)
5331 case codeptr != nil:
5332 // Note that the "receiver" parameter is nil because the actual receiver is the first input parameter.
5333 aux := ssa.InterfaceAuxCall(params)
5334 call = s.newValue1A(ssa.OpInterLECall, aux.LateExpansionResultType(), aux, codeptr)
5336 aux := ssa.StaticAuxCall(callTargetLSym(callee), params)
5337 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
5339 call.Op = ssa.OpTailLECall
5340 stksize = 0 // Tail call does not use stack. We reuse caller's frame.
5343 s.Fatalf("bad call type %v %v", n.Op(), n)
5345 call.AddArgs(callArgs...)
5346 call.AuxInt = stksize // Call operations carry the argsize of the callee along with them
5349 s.vars[memVar] = s.newValue1I(ssa.OpSelectN, types.TypeMem, int64(len(ACResults)), call)
5350 // Insert VarLive opcodes.
5351 for _, v := range n.KeepAlive {
5353 s.Fatalf("KeepAlive variable %v must have Addrtaken set", v)
5356 case ir.PAUTO, ir.PPARAM, ir.PPARAMOUT:
5358 s.Fatalf("KeepAlive variable %v must be Auto or Arg", v)
5360 s.vars[memVar] = s.newValue1A(ssa.OpVarLive, types.TypeMem, v, s.mem())
5363 // Finish block for defers
5364 if k == callDefer || k == callDeferStack {
5366 b.Kind = ssa.BlockDefer
5368 bNext := s.f.NewBlock(ssa.BlockPlain)
5370 // Add recover edge to exit code.
5371 r := s.f.NewBlock(ssa.BlockPlain)
5375 b.Likely = ssa.BranchLikely
5379 if len(res) == 0 || k != callNormal {
5380 // call has no return value. Continue with the next statement.
5384 if returnResultAddr {
5385 return s.resultAddrOfCall(call, 0, fp.Type)
5387 return s.newValue1I(ssa.OpSelectN, fp.Type, 0, call)
5390 // maybeNilCheckClosure checks if a nil check of a closure is needed in some
5391 // architecture-dependent situations and, if so, emits the nil check.
5392 func (s *state) maybeNilCheckClosure(closure *ssa.Value, k callKind) {
5393 if Arch.LinkArch.Family == sys.Wasm || buildcfg.GOOS == "aix" && k != callGo {
5394 // 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.
5395 // TODO(neelance): On other architectures this should be eliminated by the optimization steps
5400 // getClosureAndRcvr returns values for the appropriate closure and receiver of an
5402 func (s *state) getClosureAndRcvr(fn *ir.SelectorExpr) (*ssa.Value, *ssa.Value) {
5404 itab := s.newValue1(ssa.OpITab, types.Types[types.TUINTPTR], i)
5406 itabidx := fn.Offset() + 2*int64(types.PtrSize) + 8 // offset of fun field in runtime.itab
5407 closure := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.UintptrPtr, itabidx, itab)
5408 rcvr := s.newValue1(ssa.OpIData, s.f.Config.Types.BytePtr, i)
5409 return closure, rcvr
5412 // etypesign returns the signed-ness of e, for integer/pointer etypes.
5413 // -1 means signed, +1 means unsigned, 0 means non-integer/non-pointer.
5414 func etypesign(e types.Kind) int8 {
5416 case types.TINT8, types.TINT16, types.TINT32, types.TINT64, types.TINT:
5418 case types.TUINT8, types.TUINT16, types.TUINT32, types.TUINT64, types.TUINT, types.TUINTPTR, types.TUNSAFEPTR:
5424 // addr converts the address of the expression n to SSA, adds it to s and returns the SSA result.
5425 // The value that the returned Value represents is guaranteed to be non-nil.
5426 func (s *state) addr(n ir.Node) *ssa.Value {
5427 if n.Op() != ir.ONAME {
5433 s.Fatalf("addr of canSSA expression: %+v", n)
5436 t := types.NewPtr(n.Type())
5437 linksymOffset := func(lsym *obj.LSym, offset int64) *ssa.Value {
5438 v := s.entryNewValue1A(ssa.OpAddr, t, lsym, s.sb)
5439 // TODO: Make OpAddr use AuxInt as well as Aux.
5441 v = s.entryNewValue1I(ssa.OpOffPtr, v.Type, offset, v)
5446 case ir.OLINKSYMOFFSET:
5447 no := n.(*ir.LinksymOffsetExpr)
5448 return linksymOffset(no.Linksym, no.Offset_)
5451 if n.Heapaddr != nil {
5452 return s.expr(n.Heapaddr)
5457 return linksymOffset(n.Linksym(), 0)
5464 s.Fatalf("addr of undeclared ONAME %v. declared: %v", n, s.decladdrs)
5467 return s.newValue2Apos(ssa.OpLocalAddr, t, n, s.sp, s.mem(), !ir.IsAutoTmp(n))
5469 case ir.PPARAMOUT: // Same as PAUTO -- cannot generate LEA early.
5470 // ensure that we reuse symbols for out parameters so
5471 // that cse works on their addresses
5472 return s.newValue2Apos(ssa.OpLocalAddr, t, n, s.sp, s.mem(), true)
5474 s.Fatalf("variable address class %v not implemented", n.Class)
5478 // load return from callee
5479 n := n.(*ir.ResultExpr)
5480 return s.resultAddrOfCall(s.prevCall, n.Index, n.Type())
5482 n := n.(*ir.IndexExpr)
5483 if n.X.Type().IsSlice() {
5485 i := s.expr(n.Index)
5486 len := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], a)
5487 i = s.boundsCheck(i, len, ssa.BoundsIndex, n.Bounded())
5488 p := s.newValue1(ssa.OpSlicePtr, t, a)
5489 return s.newValue2(ssa.OpPtrIndex, t, p, i)
5492 i := s.expr(n.Index)
5493 len := s.constInt(types.Types[types.TINT], n.X.Type().NumElem())
5494 i = s.boundsCheck(i, len, ssa.BoundsIndex, n.Bounded())
5495 return s.newValue2(ssa.OpPtrIndex, types.NewPtr(n.X.Type().Elem()), a, i)
5498 n := n.(*ir.StarExpr)
5499 return s.exprPtr(n.X, n.Bounded(), n.Pos())
5501 n := n.(*ir.SelectorExpr)
5503 return s.newValue1I(ssa.OpOffPtr, t, n.Offset(), p)
5505 n := n.(*ir.SelectorExpr)
5506 p := s.exprPtr(n.X, n.Bounded(), n.Pos())
5507 return s.newValue1I(ssa.OpOffPtr, t, n.Offset(), p)
5509 n := n.(*ir.ConvExpr)
5510 if n.Type() == n.X.Type() {
5514 return s.newValue1(ssa.OpCopy, t, addr) // ensure that addr has the right type
5515 case ir.OCALLFUNC, ir.OCALLINTER:
5516 n := n.(*ir.CallExpr)
5517 return s.callAddr(n, callNormal)
5518 case ir.ODOTTYPE, ir.ODYNAMICDOTTYPE:
5520 if n.Op() == ir.ODOTTYPE {
5521 v, _ = s.dottype(n.(*ir.TypeAssertExpr), false)
5523 v, _ = s.dynamicDottype(n.(*ir.DynamicTypeAssertExpr), false)
5525 if v.Op != ssa.OpLoad {
5526 s.Fatalf("dottype of non-load")
5528 if v.Args[1] != s.mem() {
5529 s.Fatalf("memory no longer live from dottype load")
5533 s.Fatalf("unhandled addr %v", n.Op())
5538 // canSSA reports whether n is SSA-able.
5539 // n must be an ONAME (or an ODOT sequence with an ONAME base).
5540 func (s *state) canSSA(n ir.Node) bool {
5541 if base.Flag.N != 0 {
5546 if nn.Op() == ir.ODOT {
5547 nn := nn.(*ir.SelectorExpr)
5551 if nn.Op() == ir.OINDEX {
5552 nn := nn.(*ir.IndexExpr)
5553 if nn.X.Type().IsArray() {
5560 if n.Op() != ir.ONAME {
5563 return s.canSSAName(n.(*ir.Name)) && ssa.CanSSA(n.Type())
5566 func (s *state) canSSAName(name *ir.Name) bool {
5567 if name.Addrtaken() || !name.OnStack() {
5573 // TODO: handle this case? Named return values must be
5574 // in memory so that the deferred function can see them.
5575 // Maybe do: if !strings.HasPrefix(n.String(), "~") { return false }
5576 // Or maybe not, see issue 18860. Even unnamed return values
5577 // must be written back so if a defer recovers, the caller can see them.
5580 if s.cgoUnsafeArgs {
5581 // Cgo effectively takes the address of all result args,
5582 // but the compiler can't see that.
5587 // TODO: try to make more variables SSAable?
5590 // exprPtr evaluates n to a pointer and nil-checks it.
5591 func (s *state) exprPtr(n ir.Node, bounded bool, lineno src.XPos) *ssa.Value {
5593 if bounded || n.NonNil() {
5594 if s.f.Frontend().Debug_checknil() && lineno.Line() > 1 {
5595 s.f.Warnl(lineno, "removed nil check")
5603 // nilCheck generates nil pointer checking code.
5604 // Used only for automatically inserted nil checks,
5605 // not for user code like 'x != nil'.
5606 func (s *state) nilCheck(ptr *ssa.Value) {
5607 if base.Debug.DisableNil != 0 || s.curfn.NilCheckDisabled() {
5610 s.newValue2(ssa.OpNilCheck, types.TypeVoid, ptr, s.mem())
5613 // boundsCheck generates bounds checking code. Checks if 0 <= idx <[=] len, branches to exit if not.
5614 // Starts a new block on return.
5615 // On input, len must be converted to full int width and be nonnegative.
5616 // Returns idx converted to full int width.
5617 // If bounded is true then caller guarantees the index is not out of bounds
5618 // (but boundsCheck will still extend the index to full int width).
5619 func (s *state) boundsCheck(idx, len *ssa.Value, kind ssa.BoundsKind, bounded bool) *ssa.Value {
5620 idx = s.extendIndex(idx, len, kind, bounded)
5622 if bounded || base.Flag.B != 0 {
5623 // If bounded or bounds checking is flag-disabled, then no check necessary,
5624 // just return the extended index.
5626 // Here, bounded == true if the compiler generated the index itself,
5627 // such as in the expansion of a slice initializer. These indexes are
5628 // compiler-generated, not Go program variables, so they cannot be
5629 // attacker-controlled, so we can omit Spectre masking as well.
5631 // Note that we do not want to omit Spectre masking in code like:
5633 // if 0 <= i && i < len(x) {
5637 // Lucky for us, bounded==false for that code.
5638 // In that case (handled below), we emit a bound check (and Spectre mask)
5639 // and then the prove pass will remove the bounds check.
5640 // In theory the prove pass could potentially remove certain
5641 // Spectre masks, but it's very delicate and probably better
5642 // to be conservative and leave them all in.
5646 bNext := s.f.NewBlock(ssa.BlockPlain)
5647 bPanic := s.f.NewBlock(ssa.BlockExit)
5649 if !idx.Type.IsSigned() {
5651 case ssa.BoundsIndex:
5652 kind = ssa.BoundsIndexU
5653 case ssa.BoundsSliceAlen:
5654 kind = ssa.BoundsSliceAlenU
5655 case ssa.BoundsSliceAcap:
5656 kind = ssa.BoundsSliceAcapU
5657 case ssa.BoundsSliceB:
5658 kind = ssa.BoundsSliceBU
5659 case ssa.BoundsSlice3Alen:
5660 kind = ssa.BoundsSlice3AlenU
5661 case ssa.BoundsSlice3Acap:
5662 kind = ssa.BoundsSlice3AcapU
5663 case ssa.BoundsSlice3B:
5664 kind = ssa.BoundsSlice3BU
5665 case ssa.BoundsSlice3C:
5666 kind = ssa.BoundsSlice3CU
5671 if kind == ssa.BoundsIndex || kind == ssa.BoundsIndexU {
5672 cmp = s.newValue2(ssa.OpIsInBounds, types.Types[types.TBOOL], idx, len)
5674 cmp = s.newValue2(ssa.OpIsSliceInBounds, types.Types[types.TBOOL], idx, len)
5677 b.Kind = ssa.BlockIf
5679 b.Likely = ssa.BranchLikely
5683 s.startBlock(bPanic)
5684 if Arch.LinkArch.Family == sys.Wasm {
5685 // TODO(khr): figure out how to do "register" based calling convention for bounds checks.
5686 // Should be similar to gcWriteBarrier, but I can't make it work.
5687 s.rtcall(BoundsCheckFunc[kind], false, nil, idx, len)
5689 mem := s.newValue3I(ssa.OpPanicBounds, types.TypeMem, int64(kind), idx, len, s.mem())
5690 s.endBlock().SetControl(mem)
5694 // In Spectre index mode, apply an appropriate mask to avoid speculative out-of-bounds accesses.
5695 if base.Flag.Cfg.SpectreIndex {
5696 op := ssa.OpSpectreIndex
5697 if kind != ssa.BoundsIndex && kind != ssa.BoundsIndexU {
5698 op = ssa.OpSpectreSliceIndex
5700 idx = s.newValue2(op, types.Types[types.TINT], idx, len)
5706 // If cmp (a bool) is false, panic using the given function.
5707 func (s *state) check(cmp *ssa.Value, fn *obj.LSym) {
5709 b.Kind = ssa.BlockIf
5711 b.Likely = ssa.BranchLikely
5712 bNext := s.f.NewBlock(ssa.BlockPlain)
5714 pos := base.Ctxt.PosTable.Pos(line)
5715 fl := funcLine{f: fn, base: pos.Base(), line: pos.Line()}
5716 bPanic := s.panics[fl]
5718 bPanic = s.f.NewBlock(ssa.BlockPlain)
5719 s.panics[fl] = bPanic
5720 s.startBlock(bPanic)
5721 // The panic call takes/returns memory to ensure that the right
5722 // memory state is observed if the panic happens.
5723 s.rtcall(fn, false, nil)
5730 func (s *state) intDivide(n ir.Node, a, b *ssa.Value) *ssa.Value {
5733 case ssa.OpConst8, ssa.OpConst16, ssa.OpConst32, ssa.OpConst64:
5739 // do a size-appropriate check for zero
5740 cmp := s.newValue2(s.ssaOp(ir.ONE, n.Type()), types.Types[types.TBOOL], b, s.zeroVal(n.Type()))
5741 s.check(cmp, ir.Syms.Panicdivide)
5743 return s.newValue2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
5746 // rtcall issues a call to the given runtime function fn with the listed args.
5747 // Returns a slice of results of the given result types.
5748 // The call is added to the end of the current block.
5749 // If returns is false, the block is marked as an exit block.
5750 func (s *state) rtcall(fn *obj.LSym, returns bool, results []*types.Type, args ...*ssa.Value) []*ssa.Value {
5752 // Write args to the stack
5753 off := base.Ctxt.Arch.FixedFrameSize
5754 var callArgs []*ssa.Value
5755 var callArgTypes []*types.Type
5757 for _, arg := range args {
5759 off = types.RoundUp(off, t.Alignment())
5761 callArgs = append(callArgs, arg)
5762 callArgTypes = append(callArgTypes, t)
5765 off = types.RoundUp(off, int64(types.RegSize))
5769 aux := ssa.StaticAuxCall(fn, s.f.ABIDefault.ABIAnalyzeTypes(callArgTypes, results))
5770 callArgs = append(callArgs, s.mem())
5771 call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
5772 call.AddArgs(callArgs...)
5773 s.vars[memVar] = s.newValue1I(ssa.OpSelectN, types.TypeMem, int64(len(results)), call)
5778 b.Kind = ssa.BlockExit
5780 call.AuxInt = off - base.Ctxt.Arch.FixedFrameSize
5781 if len(results) > 0 {
5782 s.Fatalf("panic call can't have results")
5788 res := make([]*ssa.Value, len(results))
5789 for i, t := range results {
5790 off = types.RoundUp(off, t.Alignment())
5791 res[i] = s.resultOfCall(call, int64(i), t)
5794 off = types.RoundUp(off, int64(types.PtrSize))
5796 // Remember how much callee stack space we needed.
5802 // do *left = right for type t.
5803 func (s *state) storeType(t *types.Type, left, right *ssa.Value, skip skipMask, leftIsStmt bool) {
5804 s.instrument(t, left, instrumentWrite)
5806 if skip == 0 && (!t.HasPointers() || ssa.IsStackAddr(left)) {
5807 // Known to not have write barrier. Store the whole type.
5808 s.vars[memVar] = s.newValue3Apos(ssa.OpStore, types.TypeMem, t, left, right, s.mem(), leftIsStmt)
5812 // store scalar fields first, so write barrier stores for
5813 // pointer fields can be grouped together, and scalar values
5814 // don't need to be live across the write barrier call.
5815 // TODO: if the writebarrier pass knows how to reorder stores,
5816 // we can do a single store here as long as skip==0.
5817 s.storeTypeScalars(t, left, right, skip)
5818 if skip&skipPtr == 0 && t.HasPointers() {
5819 s.storeTypePtrs(t, left, right)
5823 // do *left = right for all scalar (non-pointer) parts of t.
5824 func (s *state) storeTypeScalars(t *types.Type, left, right *ssa.Value, skip skipMask) {
5826 case t.IsBoolean() || t.IsInteger() || t.IsFloat() || t.IsComplex():
5827 s.store(t, left, right)
5828 case t.IsPtrShaped():
5829 if t.IsPtr() && t.Elem().NotInHeap() {
5830 s.store(t, left, right) // see issue 42032
5832 // otherwise, no scalar fields.
5834 if skip&skipLen != 0 {
5837 len := s.newValue1(ssa.OpStringLen, types.Types[types.TINT], right)
5838 lenAddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, s.config.PtrSize, left)
5839 s.store(types.Types[types.TINT], lenAddr, len)
5841 if skip&skipLen == 0 {
5842 len := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], right)
5843 lenAddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, s.config.PtrSize, left)
5844 s.store(types.Types[types.TINT], lenAddr, len)
5846 if skip&skipCap == 0 {
5847 cap := s.newValue1(ssa.OpSliceCap, types.Types[types.TINT], right)
5848 capAddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, 2*s.config.PtrSize, left)
5849 s.store(types.Types[types.TINT], capAddr, cap)
5851 case t.IsInterface():
5852 // itab field doesn't need a write barrier (even though it is a pointer).
5853 itab := s.newValue1(ssa.OpITab, s.f.Config.Types.BytePtr, right)
5854 s.store(types.Types[types.TUINTPTR], left, itab)
5857 for i := 0; i < n; i++ {
5858 ft := t.FieldType(i)
5859 addr := s.newValue1I(ssa.OpOffPtr, ft.PtrTo(), t.FieldOff(i), left)
5860 val := s.newValue1I(ssa.OpStructSelect, ft, int64(i), right)
5861 s.storeTypeScalars(ft, addr, val, 0)
5863 case t.IsArray() && t.NumElem() == 0:
5865 case t.IsArray() && t.NumElem() == 1:
5866 s.storeTypeScalars(t.Elem(), left, s.newValue1I(ssa.OpArraySelect, t.Elem(), 0, right), 0)
5868 s.Fatalf("bad write barrier type %v", t)
5872 // do *left = right for all pointer parts of t.
5873 func (s *state) storeTypePtrs(t *types.Type, left, right *ssa.Value) {
5875 case t.IsPtrShaped():
5876 if t.IsPtr() && t.Elem().NotInHeap() {
5877 break // see issue 42032
5879 s.store(t, left, right)
5881 ptr := s.newValue1(ssa.OpStringPtr, s.f.Config.Types.BytePtr, right)
5882 s.store(s.f.Config.Types.BytePtr, left, ptr)
5884 elType := types.NewPtr(t.Elem())
5885 ptr := s.newValue1(ssa.OpSlicePtr, elType, right)
5886 s.store(elType, left, ptr)
5887 case t.IsInterface():
5888 // itab field is treated as a scalar.
5889 idata := s.newValue1(ssa.OpIData, s.f.Config.Types.BytePtr, right)
5890 idataAddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.BytePtrPtr, s.config.PtrSize, left)
5891 s.store(s.f.Config.Types.BytePtr, idataAddr, idata)
5894 for i := 0; i < n; i++ {
5895 ft := t.FieldType(i)
5896 if !ft.HasPointers() {
5899 addr := s.newValue1I(ssa.OpOffPtr, ft.PtrTo(), t.FieldOff(i), left)
5900 val := s.newValue1I(ssa.OpStructSelect, ft, int64(i), right)
5901 s.storeTypePtrs(ft, addr, val)
5903 case t.IsArray() && t.NumElem() == 0:
5905 case t.IsArray() && t.NumElem() == 1:
5906 s.storeTypePtrs(t.Elem(), left, s.newValue1I(ssa.OpArraySelect, t.Elem(), 0, right))
5908 s.Fatalf("bad write barrier type %v", t)
5912 // putArg evaluates n for the purpose of passing it as an argument to a function and returns the value for the call.
5913 func (s *state) putArg(n ir.Node, t *types.Type) *ssa.Value {
5916 a = s.newValue2(ssa.OpDereference, t, s.addr(n), s.mem())
5923 func (s *state) storeArgWithBase(n ir.Node, t *types.Type, base *ssa.Value, off int64) {
5924 pt := types.NewPtr(t)
5927 // Use special routine that avoids allocation on duplicate offsets.
5928 addr = s.constOffPtrSP(pt, off)
5930 addr = s.newValue1I(ssa.OpOffPtr, pt, off, base)
5940 s.storeType(t, addr, a, 0, false)
5943 // slice computes the slice v[i:j:k] and returns ptr, len, and cap of result.
5944 // i,j,k may be nil, in which case they are set to their default value.
5945 // v may be a slice, string or pointer to an array.
5946 func (s *state) slice(v, i, j, k *ssa.Value, bounded bool) (p, l, c *ssa.Value) {
5948 var ptr, len, cap *ssa.Value
5951 ptr = s.newValue1(ssa.OpSlicePtr, types.NewPtr(t.Elem()), v)
5952 len = s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], v)
5953 cap = s.newValue1(ssa.OpSliceCap, types.Types[types.TINT], v)
5955 ptr = s.newValue1(ssa.OpStringPtr, types.NewPtr(types.Types[types.TUINT8]), v)
5956 len = s.newValue1(ssa.OpStringLen, types.Types[types.TINT], v)
5959 if !t.Elem().IsArray() {
5960 s.Fatalf("bad ptr to array in slice %v\n", t)
5963 ptr = s.newValue1(ssa.OpCopy, types.NewPtr(t.Elem().Elem()), v)
5964 len = s.constInt(types.Types[types.TINT], t.Elem().NumElem())
5967 s.Fatalf("bad type in slice %v\n", t)
5970 // Set default values
5972 i = s.constInt(types.Types[types.TINT], 0)
5983 // Panic if slice indices are not in bounds.
5984 // Make sure we check these in reverse order so that we're always
5985 // comparing against a value known to be nonnegative. See issue 28797.
5988 kind := ssa.BoundsSlice3Alen
5990 kind = ssa.BoundsSlice3Acap
5992 k = s.boundsCheck(k, cap, kind, bounded)
5995 j = s.boundsCheck(j, k, ssa.BoundsSlice3B, bounded)
5997 i = s.boundsCheck(i, j, ssa.BoundsSlice3C, bounded)
6000 kind := ssa.BoundsSliceAlen
6002 kind = ssa.BoundsSliceAcap
6004 j = s.boundsCheck(j, k, kind, bounded)
6006 i = s.boundsCheck(i, j, ssa.BoundsSliceB, bounded)
6009 // Word-sized integer operations.
6010 subOp := s.ssaOp(ir.OSUB, types.Types[types.TINT])
6011 mulOp := s.ssaOp(ir.OMUL, types.Types[types.TINT])
6012 andOp := s.ssaOp(ir.OAND, types.Types[types.TINT])
6014 // Calculate the length (rlen) and capacity (rcap) of the new slice.
6015 // For strings the capacity of the result is unimportant. However,
6016 // we use rcap to test if we've generated a zero-length slice.
6017 // Use length of strings for that.
6018 rlen := s.newValue2(subOp, types.Types[types.TINT], j, i)
6020 if j != k && !t.IsString() {
6021 rcap = s.newValue2(subOp, types.Types[types.TINT], k, i)
6024 if (i.Op == ssa.OpConst64 || i.Op == ssa.OpConst32) && i.AuxInt == 0 {
6025 // No pointer arithmetic necessary.
6026 return ptr, rlen, rcap
6029 // Calculate the base pointer (rptr) for the new slice.
6031 // Generate the following code assuming that indexes are in bounds.
6032 // The masking is to make sure that we don't generate a slice
6033 // that points to the next object in memory. We cannot just set
6034 // the pointer to nil because then we would create a nil slice or
6039 // rptr = ptr + (mask(rcap) & (i * stride))
6041 // Where mask(x) is 0 if x==0 and -1 if x>0 and stride is the width
6042 // of the element type.
6043 stride := s.constInt(types.Types[types.TINT], ptr.Type.Elem().Size())
6045 // The delta is the number of bytes to offset ptr by.
6046 delta := s.newValue2(mulOp, types.Types[types.TINT], i, stride)
6048 // If we're slicing to the point where the capacity is zero,
6049 // zero out the delta.
6050 mask := s.newValue1(ssa.OpSlicemask, types.Types[types.TINT], rcap)
6051 delta = s.newValue2(andOp, types.Types[types.TINT], delta, mask)
6053 // Compute rptr = ptr + delta.
6054 rptr := s.newValue2(ssa.OpAddPtr, ptr.Type, ptr, delta)
6056 return rptr, rlen, rcap
6059 type u642fcvtTab struct {
6060 leq, cvt2F, and, rsh, or, add ssa.Op
6061 one func(*state, *types.Type, int64) *ssa.Value
6064 var u64_f64 = u642fcvtTab{
6066 cvt2F: ssa.OpCvt64to64F,
6068 rsh: ssa.OpRsh64Ux64,
6071 one: (*state).constInt64,
6074 var u64_f32 = u642fcvtTab{
6076 cvt2F: ssa.OpCvt64to32F,
6078 rsh: ssa.OpRsh64Ux64,
6081 one: (*state).constInt64,
6084 func (s *state) uint64Tofloat64(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6085 return s.uint64Tofloat(&u64_f64, n, x, ft, tt)
6088 func (s *state) uint64Tofloat32(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6089 return s.uint64Tofloat(&u64_f32, n, x, ft, tt)
6092 func (s *state) uint64Tofloat(cvttab *u642fcvtTab, n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6094 // result = (floatY) x
6096 // y = uintX(x) ; y = x & 1
6097 // z = uintX(x) ; z = z >> 1
6099 // result = floatY(z)
6100 // result = result + result
6103 // Code borrowed from old code generator.
6104 // What's going on: large 64-bit "unsigned" looks like
6105 // negative number to hardware's integer-to-float
6106 // conversion. However, because the mantissa is only
6107 // 63 bits, we don't need the LSB, so instead we do an
6108 // unsigned right shift (divide by two), convert, and
6109 // double. However, before we do that, we need to be
6110 // sure that we do not lose a "1" if that made the
6111 // difference in the resulting rounding. Therefore, we
6112 // preserve it, and OR (not ADD) it back in. The case
6113 // that matters is when the eleven discarded bits are
6114 // equal to 10000000001; that rounds up, and the 1 cannot
6115 // be lost else it would round down if the LSB of the
6116 // candidate mantissa is 0.
6117 cmp := s.newValue2(cvttab.leq, types.Types[types.TBOOL], s.zeroVal(ft), x)
6119 b.Kind = ssa.BlockIf
6121 b.Likely = ssa.BranchLikely
6123 bThen := s.f.NewBlock(ssa.BlockPlain)
6124 bElse := s.f.NewBlock(ssa.BlockPlain)
6125 bAfter := s.f.NewBlock(ssa.BlockPlain)
6129 a0 := s.newValue1(cvttab.cvt2F, tt, x)
6132 bThen.AddEdgeTo(bAfter)
6136 one := cvttab.one(s, ft, 1)
6137 y := s.newValue2(cvttab.and, ft, x, one)
6138 z := s.newValue2(cvttab.rsh, ft, x, one)
6139 z = s.newValue2(cvttab.or, ft, z, y)
6140 a := s.newValue1(cvttab.cvt2F, tt, z)
6141 a1 := s.newValue2(cvttab.add, tt, a, a)
6144 bElse.AddEdgeTo(bAfter)
6146 s.startBlock(bAfter)
6147 return s.variable(n, n.Type())
6150 type u322fcvtTab struct {
6151 cvtI2F, cvtF2F ssa.Op
6154 var u32_f64 = u322fcvtTab{
6155 cvtI2F: ssa.OpCvt32to64F,
6159 var u32_f32 = u322fcvtTab{
6160 cvtI2F: ssa.OpCvt32to32F,
6161 cvtF2F: ssa.OpCvt64Fto32F,
6164 func (s *state) uint32Tofloat64(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6165 return s.uint32Tofloat(&u32_f64, n, x, ft, tt)
6168 func (s *state) uint32Tofloat32(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6169 return s.uint32Tofloat(&u32_f32, n, x, ft, tt)
6172 func (s *state) uint32Tofloat(cvttab *u322fcvtTab, n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6174 // result = floatY(x)
6176 // result = floatY(float64(x) + (1<<32))
6178 cmp := s.newValue2(ssa.OpLeq32, types.Types[types.TBOOL], s.zeroVal(ft), x)
6180 b.Kind = ssa.BlockIf
6182 b.Likely = ssa.BranchLikely
6184 bThen := s.f.NewBlock(ssa.BlockPlain)
6185 bElse := s.f.NewBlock(ssa.BlockPlain)
6186 bAfter := s.f.NewBlock(ssa.BlockPlain)
6190 a0 := s.newValue1(cvttab.cvtI2F, tt, x)
6193 bThen.AddEdgeTo(bAfter)
6197 a1 := s.newValue1(ssa.OpCvt32to64F, types.Types[types.TFLOAT64], x)
6198 twoToThe32 := s.constFloat64(types.Types[types.TFLOAT64], float64(1<<32))
6199 a2 := s.newValue2(ssa.OpAdd64F, types.Types[types.TFLOAT64], a1, twoToThe32)
6200 a3 := s.newValue1(cvttab.cvtF2F, tt, a2)
6204 bElse.AddEdgeTo(bAfter)
6206 s.startBlock(bAfter)
6207 return s.variable(n, n.Type())
6210 // referenceTypeBuiltin generates code for the len/cap builtins for maps and channels.
6211 func (s *state) referenceTypeBuiltin(n *ir.UnaryExpr, x *ssa.Value) *ssa.Value {
6212 if !n.X.Type().IsMap() && !n.X.Type().IsChan() {
6213 s.Fatalf("node must be a map or a channel")
6219 // return *((*int)n)
6221 // return *(((*int)n)+1)
6224 nilValue := s.constNil(types.Types[types.TUINTPTR])
6225 cmp := s.newValue2(ssa.OpEqPtr, types.Types[types.TBOOL], x, nilValue)
6227 b.Kind = ssa.BlockIf
6229 b.Likely = ssa.BranchUnlikely
6231 bThen := s.f.NewBlock(ssa.BlockPlain)
6232 bElse := s.f.NewBlock(ssa.BlockPlain)
6233 bAfter := s.f.NewBlock(ssa.BlockPlain)
6235 // length/capacity of a nil map/chan is zero
6238 s.vars[n] = s.zeroVal(lenType)
6240 bThen.AddEdgeTo(bAfter)
6246 // length is stored in the first word for map/chan
6247 s.vars[n] = s.load(lenType, x)
6249 // capacity is stored in the second word for chan
6250 sw := s.newValue1I(ssa.OpOffPtr, lenType.PtrTo(), lenType.Size(), x)
6251 s.vars[n] = s.load(lenType, sw)
6253 s.Fatalf("op must be OLEN or OCAP")
6256 bElse.AddEdgeTo(bAfter)
6258 s.startBlock(bAfter)
6259 return s.variable(n, lenType)
6262 type f2uCvtTab struct {
6263 ltf, cvt2U, subf, or ssa.Op
6264 floatValue func(*state, *types.Type, float64) *ssa.Value
6265 intValue func(*state, *types.Type, int64) *ssa.Value
6269 var f32_u64 = f2uCvtTab{
6271 cvt2U: ssa.OpCvt32Fto64,
6274 floatValue: (*state).constFloat32,
6275 intValue: (*state).constInt64,
6279 var f64_u64 = f2uCvtTab{
6281 cvt2U: ssa.OpCvt64Fto64,
6284 floatValue: (*state).constFloat64,
6285 intValue: (*state).constInt64,
6289 var f32_u32 = f2uCvtTab{
6291 cvt2U: ssa.OpCvt32Fto32,
6294 floatValue: (*state).constFloat32,
6295 intValue: func(s *state, t *types.Type, v int64) *ssa.Value { return s.constInt32(t, int32(v)) },
6299 var f64_u32 = f2uCvtTab{
6301 cvt2U: ssa.OpCvt64Fto32,
6304 floatValue: (*state).constFloat64,
6305 intValue: func(s *state, t *types.Type, v int64) *ssa.Value { return s.constInt32(t, int32(v)) },
6309 func (s *state) float32ToUint64(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6310 return s.floatToUint(&f32_u64, n, x, ft, tt)
6312 func (s *state) float64ToUint64(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6313 return s.floatToUint(&f64_u64, n, x, ft, tt)
6316 func (s *state) float32ToUint32(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6317 return s.floatToUint(&f32_u32, n, x, ft, tt)
6320 func (s *state) float64ToUint32(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6321 return s.floatToUint(&f64_u32, n, x, ft, tt)
6324 func (s *state) floatToUint(cvttab *f2uCvtTab, n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
6325 // cutoff:=1<<(intY_Size-1)
6326 // if x < floatX(cutoff) {
6327 // result = uintY(x)
6329 // y = x - floatX(cutoff)
6331 // result = z | -(cutoff)
6333 cutoff := cvttab.floatValue(s, ft, float64(cvttab.cutoff))
6334 cmp := s.newValue2(cvttab.ltf, types.Types[types.TBOOL], x, cutoff)
6336 b.Kind = ssa.BlockIf
6338 b.Likely = ssa.BranchLikely
6340 bThen := s.f.NewBlock(ssa.BlockPlain)
6341 bElse := s.f.NewBlock(ssa.BlockPlain)
6342 bAfter := s.f.NewBlock(ssa.BlockPlain)
6346 a0 := s.newValue1(cvttab.cvt2U, tt, x)
6349 bThen.AddEdgeTo(bAfter)
6353 y := s.newValue2(cvttab.subf, ft, x, cutoff)
6354 y = s.newValue1(cvttab.cvt2U, tt, y)
6355 z := cvttab.intValue(s, tt, int64(-cvttab.cutoff))
6356 a1 := s.newValue2(cvttab.or, tt, y, z)
6359 bElse.AddEdgeTo(bAfter)
6361 s.startBlock(bAfter)
6362 return s.variable(n, n.Type())
6365 // dottype generates SSA for a type assertion node.
6366 // commaok indicates whether to panic or return a bool.
6367 // If commaok is false, resok will be nil.
6368 func (s *state) dottype(n *ir.TypeAssertExpr, commaok bool) (res, resok *ssa.Value) {
6369 iface := s.expr(n.X) // input interface
6370 target := s.reflectType(n.Type()) // target type
6371 var targetItab *ssa.Value
6373 targetItab = s.expr(n.ITab)
6375 return s.dottype1(n.Pos(), n.X.Type(), n.Type(), iface, nil, target, targetItab, commaok)
6378 func (s *state) dynamicDottype(n *ir.DynamicTypeAssertExpr, commaok bool) (res, resok *ssa.Value) {
6379 iface := s.expr(n.X)
6380 var source, target, targetItab *ssa.Value
6381 if n.SrcRType != nil {
6382 source = s.expr(n.SrcRType)
6384 if !n.X.Type().IsEmptyInterface() && !n.Type().IsInterface() {
6385 byteptr := s.f.Config.Types.BytePtr
6386 targetItab = s.expr(n.ITab)
6387 // TODO(mdempsky): Investigate whether compiling n.RType could be
6388 // better than loading itab.typ.
6389 target = s.load(byteptr, s.newValue1I(ssa.OpOffPtr, byteptr, int64(types.PtrSize), targetItab)) // itab.typ
6391 target = s.expr(n.RType)
6393 return s.dottype1(n.Pos(), n.X.Type(), n.Type(), iface, source, target, targetItab, commaok)
6396 // dottype1 implements a x.(T) operation. iface is the argument (x), dst is the type we're asserting to (T)
6397 // and src is the type we're asserting from.
6398 // source is the *runtime._type of src
6399 // target is the *runtime._type of dst.
6400 // If src is a nonempty interface and dst is not an interface, targetItab is an itab representing (dst, src). Otherwise it is nil.
6401 // commaok is true if the caller wants a boolean success value. Otherwise, the generated code panics if the conversion fails.
6402 func (s *state) dottype1(pos src.XPos, src, dst *types.Type, iface, source, target, targetItab *ssa.Value, commaok bool) (res, resok *ssa.Value) {
6403 byteptr := s.f.Config.Types.BytePtr
6404 if dst.IsInterface() {
6405 if dst.IsEmptyInterface() {
6406 // Converting to an empty interface.
6407 // Input could be an empty or nonempty interface.
6408 if base.Debug.TypeAssert > 0 {
6409 base.WarnfAt(pos, "type assertion inlined")
6412 // Get itab/type field from input.
6413 itab := s.newValue1(ssa.OpITab, byteptr, iface)
6414 // Conversion succeeds iff that field is not nil.
6415 cond := s.newValue2(ssa.OpNeqPtr, types.Types[types.TBOOL], itab, s.constNil(byteptr))
6417 if src.IsEmptyInterface() && commaok {
6418 // Converting empty interface to empty interface with ,ok is just a nil check.
6422 // Branch on nilness.
6424 b.Kind = ssa.BlockIf
6426 b.Likely = ssa.BranchLikely
6427 bOk := s.f.NewBlock(ssa.BlockPlain)
6428 bFail := s.f.NewBlock(ssa.BlockPlain)
6433 // On failure, panic by calling panicnildottype.
6435 s.rtcall(ir.Syms.Panicnildottype, false, nil, target)
6437 // On success, return (perhaps modified) input interface.
6439 if src.IsEmptyInterface() {
6440 res = iface // Use input interface unchanged.
6443 // Load type out of itab, build interface with existing idata.
6444 off := s.newValue1I(ssa.OpOffPtr, byteptr, int64(types.PtrSize), itab)
6445 typ := s.load(byteptr, off)
6446 idata := s.newValue1(ssa.OpIData, byteptr, iface)
6447 res = s.newValue2(ssa.OpIMake, dst, typ, idata)
6452 // nonempty -> empty
6453 // Need to load type from itab
6454 off := s.newValue1I(ssa.OpOffPtr, byteptr, int64(types.PtrSize), itab)
6455 s.vars[typVar] = s.load(byteptr, off)
6458 // itab is nil, might as well use that as the nil result.
6460 s.vars[typVar] = itab
6464 bEnd := s.f.NewBlock(ssa.BlockPlain)
6466 bFail.AddEdgeTo(bEnd)
6468 idata := s.newValue1(ssa.OpIData, byteptr, iface)
6469 res = s.newValue2(ssa.OpIMake, dst, s.variable(typVar, byteptr), idata)
6471 delete(s.vars, typVar) // no practical effect, just to indicate typVar is no longer live.
6474 // converting to a nonempty interface needs a runtime call.
6475 if base.Debug.TypeAssert > 0 {
6476 base.WarnfAt(pos, "type assertion not inlined")
6479 fn := ir.Syms.AssertI2I
6480 if src.IsEmptyInterface() {
6481 fn = ir.Syms.AssertE2I
6483 data := s.newValue1(ssa.OpIData, types.Types[types.TUNSAFEPTR], iface)
6484 tab := s.newValue1(ssa.OpITab, byteptr, iface)
6485 tab = s.rtcall(fn, true, []*types.Type{byteptr}, target, tab)[0]
6486 return s.newValue2(ssa.OpIMake, dst, tab, data), nil
6488 fn := ir.Syms.AssertI2I2
6489 if src.IsEmptyInterface() {
6490 fn = ir.Syms.AssertE2I2
6492 res = s.rtcall(fn, true, []*types.Type{dst}, target, iface)[0]
6493 resok = s.newValue2(ssa.OpNeqInter, types.Types[types.TBOOL], res, s.constInterface(dst))
6497 if base.Debug.TypeAssert > 0 {
6498 base.WarnfAt(pos, "type assertion inlined")
6501 // Converting to a concrete type.
6502 direct := types.IsDirectIface(dst)
6503 itab := s.newValue1(ssa.OpITab, byteptr, iface) // type word of interface
6504 if base.Debug.TypeAssert > 0 {
6505 base.WarnfAt(pos, "type assertion inlined")
6507 var wantedFirstWord *ssa.Value
6508 if src.IsEmptyInterface() {
6509 // Looking for pointer to target type.
6510 wantedFirstWord = target
6512 // Looking for pointer to itab for target type and source interface.
6513 wantedFirstWord = targetItab
6516 var tmp ir.Node // temporary for use with large types
6517 var addr *ssa.Value // address of tmp
6518 if commaok && !ssa.CanSSA(dst) {
6519 // unSSAable type, use temporary.
6520 // TODO: get rid of some of these temporaries.
6521 tmp, addr = s.temp(pos, dst)
6524 cond := s.newValue2(ssa.OpEqPtr, types.Types[types.TBOOL], itab, wantedFirstWord)
6526 b.Kind = ssa.BlockIf
6528 b.Likely = ssa.BranchLikely
6530 bOk := s.f.NewBlock(ssa.BlockPlain)
6531 bFail := s.f.NewBlock(ssa.BlockPlain)
6536 // on failure, panic by calling panicdottype
6540 taddr = s.reflectType(src)
6542 if src.IsEmptyInterface() {
6543 s.rtcall(ir.Syms.PanicdottypeE, false, nil, itab, target, taddr)
6545 s.rtcall(ir.Syms.PanicdottypeI, false, nil, itab, target, taddr)
6548 // on success, return data from interface
6551 return s.newValue1(ssa.OpIData, dst, iface), nil
6553 p := s.newValue1(ssa.OpIData, types.NewPtr(dst), iface)
6554 return s.load(dst, p), nil
6557 // commaok is the more complicated case because we have
6558 // a control flow merge point.
6559 bEnd := s.f.NewBlock(ssa.BlockPlain)
6560 // Note that we need a new valVar each time (unlike okVar where we can
6561 // reuse the variable) because it might have a different type every time.
6562 valVar := ssaMarker("val")
6564 // type assertion succeeded
6568 s.vars[valVar] = s.newValue1(ssa.OpIData, dst, iface)
6570 p := s.newValue1(ssa.OpIData, types.NewPtr(dst), iface)
6571 s.vars[valVar] = s.load(dst, p)
6574 p := s.newValue1(ssa.OpIData, types.NewPtr(dst), iface)
6575 s.move(dst, addr, p)
6577 s.vars[okVar] = s.constBool(true)
6581 // type assertion failed
6584 s.vars[valVar] = s.zeroVal(dst)
6588 s.vars[okVar] = s.constBool(false)
6590 bFail.AddEdgeTo(bEnd)
6595 res = s.variable(valVar, dst)
6596 delete(s.vars, valVar) // no practical effect, just to indicate typVar is no longer live.
6598 res = s.load(dst, addr)
6600 resok = s.variable(okVar, types.Types[types.TBOOL])
6601 delete(s.vars, okVar) // ditto
6605 // temp allocates a temp of type t at position pos
6606 func (s *state) temp(pos src.XPos, t *types.Type) (*ir.Name, *ssa.Value) {
6607 tmp := typecheck.TempAt(pos, s.curfn, t)
6608 if t.HasPointers() {
6609 s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, tmp, s.mem())
6615 // variable returns the value of a variable at the current location.
6616 func (s *state) variable(n ir.Node, t *types.Type) *ssa.Value {
6626 if s.curBlock == s.f.Entry {
6627 // No variable should be live at entry.
6628 s.f.Fatalf("value %v (%v) incorrectly live at entry", n, v)
6630 // Make a FwdRef, which records a value that's live on block input.
6631 // We'll find the matching definition as part of insertPhis.
6632 v = s.newValue0A(ssa.OpFwdRef, t, fwdRefAux{N: n})
6634 if n.Op() == ir.ONAME {
6635 s.addNamedValue(n.(*ir.Name), v)
6640 func (s *state) mem() *ssa.Value {
6641 return s.variable(memVar, types.TypeMem)
6644 func (s *state) addNamedValue(n *ir.Name, v *ssa.Value) {
6645 if n.Class == ir.Pxxx {
6646 // Don't track our marker nodes (memVar etc.).
6649 if ir.IsAutoTmp(n) {
6650 // Don't track temporary variables.
6653 if n.Class == ir.PPARAMOUT {
6654 // Don't track named output values. This prevents return values
6655 // from being assigned too early. See #14591 and #14762. TODO: allow this.
6658 loc := ssa.LocalSlot{N: n, Type: n.Type(), Off: 0}
6659 values, ok := s.f.NamedValues[loc]
6661 s.f.Names = append(s.f.Names, &loc)
6662 s.f.CanonicalLocalSlots[loc] = &loc
6664 s.f.NamedValues[loc] = append(values, v)
6667 // Branch is an unresolved branch.
6668 type Branch struct {
6669 P *obj.Prog // branch instruction
6670 B *ssa.Block // target
6673 // State contains state needed during Prog generation.
6679 // Branches remembers all the branch instructions we've seen
6680 // and where they would like to go.
6683 // JumpTables remembers all the jump tables we've seen.
6684 JumpTables []*ssa.Block
6686 // bstart remembers where each block starts (indexed by block ID)
6689 maxarg int64 // largest frame size for arguments to calls made by the function
6691 // Map from GC safe points to liveness index, generated by
6692 // liveness analysis.
6693 livenessMap liveness.Map
6695 // partLiveArgs includes arguments that may be partially live, for which we
6696 // need to generate instructions that spill the argument registers.
6697 partLiveArgs map[*ir.Name]bool
6699 // lineRunStart records the beginning of the current run of instructions
6700 // within a single block sharing the same line number
6701 // Used to move statement marks to the beginning of such runs.
6702 lineRunStart *obj.Prog
6704 // wasm: The number of values on the WebAssembly stack. This is only used as a safeguard.
6705 OnWasmStackSkipped int
6708 func (s *State) FuncInfo() *obj.FuncInfo {
6709 return s.pp.CurFunc.LSym.Func()
6712 // Prog appends a new Prog.
6713 func (s *State) Prog(as obj.As) *obj.Prog {
6715 if objw.LosesStmtMark(as) {
6718 // Float a statement start to the beginning of any same-line run.
6719 // lineRunStart is reset at block boundaries, which appears to work well.
6720 if s.lineRunStart == nil || s.lineRunStart.Pos.Line() != p.Pos.Line() {
6722 } else if p.Pos.IsStmt() == src.PosIsStmt {
6723 s.lineRunStart.Pos = s.lineRunStart.Pos.WithIsStmt()
6724 p.Pos = p.Pos.WithNotStmt()
6729 // Pc returns the current Prog.
6730 func (s *State) Pc() *obj.Prog {
6734 // SetPos sets the current source position.
6735 func (s *State) SetPos(pos src.XPos) {
6739 // Br emits a single branch instruction and returns the instruction.
6740 // Not all architectures need the returned instruction, but otherwise
6741 // the boilerplate is common to all.
6742 func (s *State) Br(op obj.As, target *ssa.Block) *obj.Prog {
6744 p.To.Type = obj.TYPE_BRANCH
6745 s.Branches = append(s.Branches, Branch{P: p, B: target})
6749 // DebugFriendlySetPosFrom adjusts Pos.IsStmt subject to heuristics
6750 // that reduce "jumpy" line number churn when debugging.
6751 // Spill/fill/copy instructions from the register allocator,
6752 // phi functions, and instructions with a no-pos position
6753 // are examples of instructions that can cause churn.
6754 func (s *State) DebugFriendlySetPosFrom(v *ssa.Value) {
6756 case ssa.OpPhi, ssa.OpCopy, ssa.OpLoadReg, ssa.OpStoreReg:
6757 // These are not statements
6758 s.SetPos(v.Pos.WithNotStmt())
6761 if p != src.NoXPos {
6762 // If the position is defined, update the position.
6763 // Also convert default IsStmt to NotStmt; only
6764 // explicit statement boundaries should appear
6765 // in the generated code.
6766 if p.IsStmt() != src.PosIsStmt {
6767 if s.pp.Pos.IsStmt() == src.PosIsStmt && s.pp.Pos.SameFileAndLine(p) {
6768 // If s.pp.Pos already has a statement mark, then it was set here (below) for
6769 // the previous value. If an actual instruction had been emitted for that
6770 // value, then the statement mark would have been reset. Since the statement
6771 // mark of s.pp.Pos was not reset, this position (file/line) still needs a
6772 // statement mark on an instruction. If file and line for this value are
6773 // the same as the previous value, then the first instruction for this
6774 // value will work to take the statement mark. Return early to avoid
6775 // resetting the statement mark.
6777 // The reset of s.pp.Pos occurs in (*Progs).Prog() -- if it emits
6778 // an instruction, and the instruction's statement mark was set,
6779 // and it is not one of the LosesStmtMark instructions,
6780 // then Prog() resets the statement mark on the (*Progs).Pos.
6784 // Calls use the pos attached to v, but copy the statement mark from State
6788 s.SetPos(s.pp.Pos.WithNotStmt())
6793 // emit argument info (locations on stack) for traceback.
6794 func emitArgInfo(e *ssafn, f *ssa.Func, pp *objw.Progs) {
6795 ft := e.curfn.Type()
6796 if ft.NumRecvs() == 0 && ft.NumParams() == 0 {
6800 x := EmitArgInfo(e.curfn, f.OwnAux.ABIInfo())
6801 x.Set(obj.AttrContentAddressable, true)
6802 e.curfn.LSym.Func().ArgInfo = x
6804 // Emit a funcdata pointing at the arg info data.
6805 p := pp.Prog(obj.AFUNCDATA)
6806 p.From.SetConst(rtabi.FUNCDATA_ArgInfo)
6807 p.To.Type = obj.TYPE_MEM
6808 p.To.Name = obj.NAME_EXTERN
6812 // emit argument info (locations on stack) of f for traceback.
6813 func EmitArgInfo(f *ir.Func, abiInfo *abi.ABIParamResultInfo) *obj.LSym {
6814 x := base.Ctxt.Lookup(fmt.Sprintf("%s.arginfo%d", f.LSym.Name, f.ABI))
6815 // NOTE: do not set ContentAddressable here. This may be referenced from
6816 // assembly code by name (in this case f is a declaration).
6817 // Instead, set it in emitArgInfo above.
6819 PtrSize := int64(types.PtrSize)
6820 uintptrTyp := types.Types[types.TUINTPTR]
6822 isAggregate := func(t *types.Type) bool {
6823 return t.IsStruct() || t.IsArray() || t.IsComplex() || t.IsInterface() || t.IsString() || t.IsSlice()
6826 // Populate the data.
6827 // The data is a stream of bytes, which contains the offsets and sizes of the
6828 // non-aggregate arguments or non-aggregate fields/elements of aggregate-typed
6829 // arguments, along with special "operators". Specifically,
6830 // - for each non-aggrgate arg/field/element, its offset from FP (1 byte) and
6832 // - special operators:
6833 // - 0xff - end of sequence
6834 // - 0xfe - print { (at the start of an aggregate-typed argument)
6835 // - 0xfd - print } (at the end of an aggregate-typed argument)
6836 // - 0xfc - print ... (more args/fields/elements)
6837 // - 0xfb - print _ (offset too large)
6838 // These constants need to be in sync with runtime.traceback.go:printArgs.
6844 _offsetTooLarge = 0xfb
6845 _special = 0xf0 // above this are operators, below this are ordinary offsets
6849 limit = 10 // print no more than 10 args/components
6850 maxDepth = 5 // no more than 5 layers of nesting
6852 // maxLen is a (conservative) upper bound of the byte stream length. For
6853 // each arg/component, it has no more than 2 bytes of data (size, offset),
6854 // and no more than one {, }, ... at each level (it cannot have both the
6855 // data and ... unless it is the last one, just be conservative). Plus 1
6857 maxLen = (maxDepth*3+2)*limit + 1
6862 writebyte := func(o uint8) { wOff = objw.Uint8(x, wOff, o) }
6864 // Write one non-aggrgate arg/field/element.
6865 write1 := func(sz, offset int64) {
6866 if offset >= _special {
6867 writebyte(_offsetTooLarge)
6869 writebyte(uint8(offset))
6870 writebyte(uint8(sz))
6875 // Visit t recursively and write it out.
6876 // Returns whether to continue visiting.
6877 var visitType func(baseOffset int64, t *types.Type, depth int) bool
6878 visitType = func(baseOffset int64, t *types.Type, depth int) bool {
6880 writebyte(_dotdotdot)
6883 if !isAggregate(t) {
6884 write1(t.Size(), baseOffset)
6887 writebyte(_startAgg)
6889 if depth >= maxDepth {
6890 writebyte(_dotdotdot)
6896 case t.IsInterface(), t.IsString():
6897 _ = visitType(baseOffset, uintptrTyp, depth) &&
6898 visitType(baseOffset+PtrSize, uintptrTyp, depth)
6900 _ = visitType(baseOffset, uintptrTyp, depth) &&
6901 visitType(baseOffset+PtrSize, uintptrTyp, depth) &&
6902 visitType(baseOffset+PtrSize*2, uintptrTyp, depth)
6904 _ = visitType(baseOffset, types.FloatForComplex(t), depth) &&
6905 visitType(baseOffset+t.Size()/2, types.FloatForComplex(t), depth)
6907 if t.NumElem() == 0 {
6908 n++ // {} counts as a component
6911 for i := int64(0); i < t.NumElem(); i++ {
6912 if !visitType(baseOffset, t.Elem(), depth) {
6915 baseOffset += t.Elem().Size()
6918 if t.NumFields() == 0 {
6919 n++ // {} counts as a component
6922 for _, field := range t.Fields() {
6923 if !visitType(baseOffset+field.Offset, field.Type, depth) {
6933 if strings.Contains(f.LSym.Name, "[") {
6934 // Skip the dictionary argument - it is implicit and the user doesn't need to see it.
6938 for _, a := range abiInfo.InParams()[start:] {
6939 if !visitType(a.FrameOffset(abiInfo), a.Type, 0) {
6945 base.Fatalf("ArgInfo too large")
6951 // for wrapper, emit info of wrapped function.
6952 func emitWrappedFuncInfo(e *ssafn, pp *objw.Progs) {
6953 if base.Ctxt.Flag_linkshared {
6954 // Relative reference (SymPtrOff) to another shared object doesn't work.
6959 wfn := e.curfn.WrappedFunc
6964 wsym := wfn.Linksym()
6965 x := base.Ctxt.LookupInit(fmt.Sprintf("%s.wrapinfo", wsym.Name), func(x *obj.LSym) {
6966 objw.SymPtrOff(x, 0, wsym)
6967 x.Set(obj.AttrContentAddressable, true)
6969 e.curfn.LSym.Func().WrapInfo = x
6971 // Emit a funcdata pointing at the wrap info data.
6972 p := pp.Prog(obj.AFUNCDATA)
6973 p.From.SetConst(rtabi.FUNCDATA_WrapInfo)
6974 p.To.Type = obj.TYPE_MEM
6975 p.To.Name = obj.NAME_EXTERN
6979 // genssa appends entries to pp for each instruction in f.
6980 func genssa(f *ssa.Func, pp *objw.Progs) {
6982 s.ABI = f.OwnAux.Fn.ABI()
6984 e := f.Frontend().(*ssafn)
6986 s.livenessMap, s.partLiveArgs = liveness.Compute(e.curfn, f, e.stkptrsize, pp)
6987 emitArgInfo(e, f, pp)
6988 argLiveBlockMap, argLiveValueMap := liveness.ArgLiveness(e.curfn, f, pp)
6990 openDeferInfo := e.curfn.LSym.Func().OpenCodedDeferInfo
6991 if openDeferInfo != nil {
6992 // This function uses open-coded defers -- write out the funcdata
6993 // info that we computed at the end of genssa.
6994 p := pp.Prog(obj.AFUNCDATA)
6995 p.From.SetConst(rtabi.FUNCDATA_OpenCodedDeferInfo)
6996 p.To.Type = obj.TYPE_MEM
6997 p.To.Name = obj.NAME_EXTERN
6998 p.To.Sym = openDeferInfo
7001 emitWrappedFuncInfo(e, pp)
7003 // Remember where each block starts.
7004 s.bstart = make([]*obj.Prog, f.NumBlocks())
7006 var progToValue map[*obj.Prog]*ssa.Value
7007 var progToBlock map[*obj.Prog]*ssa.Block
7008 var valueToProgAfter []*obj.Prog // The first Prog following computation of a value v; v is visible at this point.
7009 gatherPrintInfo := f.PrintOrHtmlSSA || ssa.GenssaDump[f.Name]
7010 if gatherPrintInfo {
7011 progToValue = make(map[*obj.Prog]*ssa.Value, f.NumValues())
7012 progToBlock = make(map[*obj.Prog]*ssa.Block, f.NumBlocks())
7013 f.Logf("genssa %s\n", f.Name)
7014 progToBlock[s.pp.Next] = f.Blocks[0]
7017 if base.Ctxt.Flag_locationlists {
7018 if cap(f.Cache.ValueToProgAfter) < f.NumValues() {
7019 f.Cache.ValueToProgAfter = make([]*obj.Prog, f.NumValues())
7021 valueToProgAfter = f.Cache.ValueToProgAfter[:f.NumValues()]
7022 for i := range valueToProgAfter {
7023 valueToProgAfter[i] = nil
7027 // If the very first instruction is not tagged as a statement,
7028 // debuggers may attribute it to previous function in program.
7029 firstPos := src.NoXPos
7030 for _, v := range f.Entry.Values {
7031 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 {
7033 v.Pos = firstPos.WithDefaultStmt()
7038 // inlMarks has an entry for each Prog that implements an inline mark.
7039 // It maps from that Prog to the global inlining id of the inlined body
7040 // which should unwind to this Prog's location.
7041 var inlMarks map[*obj.Prog]int32
7042 var inlMarkList []*obj.Prog
7044 // inlMarksByPos maps from a (column 1) source position to the set of
7045 // Progs that are in the set above and have that source position.
7046 var inlMarksByPos map[src.XPos][]*obj.Prog
7048 var argLiveIdx int = -1 // argument liveness info index
7050 // Emit basic blocks
7051 for i, b := range f.Blocks {
7052 s.bstart[b.ID] = s.pp.Next
7053 s.lineRunStart = nil
7054 s.SetPos(s.pp.Pos.WithNotStmt()) // It needs a non-empty Pos, but cannot be a statement boundary (yet).
7056 if idx, ok := argLiveBlockMap[b.ID]; ok && idx != argLiveIdx {
7058 p := s.pp.Prog(obj.APCDATA)
7059 p.From.SetConst(rtabi.PCDATA_ArgLiveIndex)
7060 p.To.SetConst(int64(idx))
7063 // Emit values in block
7064 Arch.SSAMarkMoves(&s, b)
7065 for _, v := range b.Values {
7067 s.DebugFriendlySetPosFrom(v)
7069 if v.Op.ResultInArg0() && v.ResultReg() != v.Args[0].Reg() {
7070 v.Fatalf("input[0] and output not in same register %s", v.LongString())
7075 // memory arg needs no code
7077 // input args need no code
7078 case ssa.OpSP, ssa.OpSB:
7080 case ssa.OpSelect0, ssa.OpSelect1, ssa.OpSelectN, ssa.OpMakeResult:
7083 // nothing to do when there's a g register,
7084 // and checkLower complains if there's not
7085 case ssa.OpVarDef, ssa.OpVarLive, ssa.OpKeepAlive, ssa.OpWBend:
7086 // nothing to do; already used by liveness
7090 // nothing to do; no-op conversion for liveness
7091 if v.Args[0].Reg() != v.Reg() {
7092 v.Fatalf("OpConvert should be a no-op: %s; %s", v.Args[0].LongString(), v.LongString())
7095 p := Arch.Ginsnop(s.pp)
7096 if inlMarks == nil {
7097 inlMarks = map[*obj.Prog]int32{}
7098 inlMarksByPos = map[src.XPos][]*obj.Prog{}
7100 inlMarks[p] = v.AuxInt32()
7101 inlMarkList = append(inlMarkList, p)
7102 pos := v.Pos.AtColumn1()
7103 inlMarksByPos[pos] = append(inlMarksByPos[pos], p)
7104 firstPos = src.NoXPos
7107 // Special case for first line in function; move it to the start (which cannot be a register-valued instruction)
7108 if firstPos != src.NoXPos && v.Op != ssa.OpArgIntReg && v.Op != ssa.OpArgFloatReg && v.Op != ssa.OpLoadReg && v.Op != ssa.OpStoreReg {
7110 firstPos = src.NoXPos
7112 // Attach this safe point to the next
7114 s.pp.NextLive = s.livenessMap.Get(v)
7115 s.pp.NextUnsafe = s.livenessMap.GetUnsafe(v)
7117 // let the backend handle it
7118 Arch.SSAGenValue(&s, v)
7121 if idx, ok := argLiveValueMap[v.ID]; ok && idx != argLiveIdx {
7123 p := s.pp.Prog(obj.APCDATA)
7124 p.From.SetConst(rtabi.PCDATA_ArgLiveIndex)
7125 p.To.SetConst(int64(idx))
7128 if base.Ctxt.Flag_locationlists {
7129 valueToProgAfter[v.ID] = s.pp.Next
7132 if gatherPrintInfo {
7133 for ; x != s.pp.Next; x = x.Link {
7138 // If this is an empty infinite loop, stick a hardware NOP in there so that debuggers are less confused.
7139 if s.bstart[b.ID] == s.pp.Next && len(b.Succs) == 1 && b.Succs[0].Block() == b {
7140 p := Arch.Ginsnop(s.pp)
7141 p.Pos = p.Pos.WithIsStmt()
7142 if b.Pos == src.NoXPos {
7143 b.Pos = p.Pos // It needs a file, otherwise a no-file non-zero line causes confusion. See #35652.
7144 if b.Pos == src.NoXPos {
7145 b.Pos = pp.Text.Pos // Sometimes p.Pos is empty. See #35695.
7148 b.Pos = b.Pos.WithBogusLine() // Debuggers are not good about infinite loops, force a change in line number
7151 // Set unsafe mark for any end-of-block generated instructions
7152 // (normally, conditional or unconditional branches).
7153 // This is particularly important for empty blocks, as there
7154 // are no values to inherit the unsafe mark from.
7155 s.pp.NextUnsafe = s.livenessMap.GetUnsafeBlock(b)
7157 // Emit control flow instructions for block
7159 if i < len(f.Blocks)-1 && base.Flag.N == 0 {
7160 // If -N, leave next==nil so every block with successors
7161 // ends in a JMP (except call blocks - plive doesn't like
7162 // select{send,recv} followed by a JMP call). Helps keep
7163 // line numbers for otherwise empty blocks.
7164 next = f.Blocks[i+1]
7168 Arch.SSAGenBlock(&s, b, next)
7169 if gatherPrintInfo {
7170 for ; x != s.pp.Next; x = x.Link {
7175 if f.Blocks[len(f.Blocks)-1].Kind == ssa.BlockExit {
7176 // We need the return address of a panic call to
7177 // still be inside the function in question. So if
7178 // it ends in a call which doesn't return, add a
7179 // nop (which will never execute) after the call.
7182 if openDeferInfo != nil {
7183 // When doing open-coded defers, generate a disconnected call to
7184 // deferreturn and a return. This will be used to during panic
7185 // recovery to unwind the stack and return back to the runtime.
7186 s.pp.NextLive = s.livenessMap.DeferReturn
7187 p := pp.Prog(obj.ACALL)
7188 p.To.Type = obj.TYPE_MEM
7189 p.To.Name = obj.NAME_EXTERN
7190 p.To.Sym = ir.Syms.Deferreturn
7192 // Load results into registers. So when a deferred function
7193 // recovers a panic, it will return to caller with right results.
7194 // The results are already in memory, because they are not SSA'd
7195 // when the function has defers (see canSSAName).
7196 for _, o := range f.OwnAux.ABIInfo().OutParams() {
7198 rts, offs := o.RegisterTypesAndOffsets()
7199 for i := range o.Registers {
7200 Arch.LoadRegResult(&s, f, rts[i], ssa.ObjRegForAbiReg(o.Registers[i], f.Config), n, offs[i])
7207 if inlMarks != nil {
7210 // We have some inline marks. Try to find other instructions we're
7211 // going to emit anyway, and use those instructions instead of the
7213 for p := pp.Text; p != nil; p = p.Link {
7214 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 {
7215 // Don't use 0-sized instructions as inline marks, because we need
7216 // to identify inline mark instructions by pc offset.
7217 // (Some of these instructions are sometimes zero-sized, sometimes not.
7218 // We must not use anything that even might be zero-sized.)
7219 // TODO: are there others?
7222 if _, ok := inlMarks[p]; ok {
7223 // Don't use inline marks themselves. We don't know
7224 // whether they will be zero-sized or not yet.
7227 if p.As == obj.ACALL || p.As == obj.ADUFFCOPY || p.As == obj.ADUFFZERO {
7230 pos := p.Pos.AtColumn1()
7231 s := inlMarksByPos[pos]
7235 for _, m := range s {
7236 // We found an instruction with the same source position as
7237 // some of the inline marks.
7238 // Use this instruction instead.
7239 p.Pos = p.Pos.WithIsStmt() // promote position to a statement
7240 pp.CurFunc.LSym.Func().AddInlMark(p, inlMarks[m])
7241 // Make the inline mark a real nop, so it doesn't generate any code.
7247 delete(inlMarksByPos, pos)
7249 // Any unmatched inline marks now need to be added to the inlining tree (and will generate a nop instruction).
7250 for _, p := range inlMarkList {
7251 if p.As != obj.ANOP {
7252 pp.CurFunc.LSym.Func().AddInlMark(p, inlMarks[p])
7256 if e.stksize == 0 && !hasCall {
7257 // Frameless leaf function. It doesn't need any preamble,
7258 // so make sure its first instruction isn't from an inlined callee.
7259 // If it is, add a nop at the start of the function with a position
7260 // equal to the start of the function.
7261 // This ensures that runtime.FuncForPC(uintptr(reflect.ValueOf(fn).Pointer())).Name()
7262 // returns the right answer. See issue 58300.
7263 for p := pp.Text; p != nil; p = p.Link {
7264 if p.As == obj.AFUNCDATA || p.As == obj.APCDATA || p.As == obj.ATEXT || p.As == obj.ANOP {
7267 if base.Ctxt.PosTable.Pos(p.Pos).Base().InliningIndex() >= 0 {
7268 // Make a real (not 0-sized) nop.
7269 nop := Arch.Ginsnop(pp)
7270 nop.Pos = e.curfn.Pos().WithIsStmt()
7272 // Unfortunately, Ginsnop puts the instruction at the
7273 // end of the list. Move it up to just before p.
7275 // Unlink from the current list.
7276 for x := pp.Text; x != nil; x = x.Link {
7282 // Splice in right before p.
7283 for x := pp.Text; x != nil; x = x.Link {
7296 if base.Ctxt.Flag_locationlists {
7297 var debugInfo *ssa.FuncDebug
7298 debugInfo = e.curfn.DebugInfo.(*ssa.FuncDebug)
7299 if e.curfn.ABI == obj.ABIInternal && base.Flag.N != 0 {
7300 ssa.BuildFuncDebugNoOptimized(base.Ctxt, f, base.Debug.LocationLists > 1, StackOffset, debugInfo)
7302 ssa.BuildFuncDebug(base.Ctxt, f, base.Debug.LocationLists, StackOffset, debugInfo)
7305 idToIdx := make([]int, f.NumBlocks())
7306 for i, b := range f.Blocks {
7309 // Note that at this moment, Prog.Pc is a sequence number; it's
7310 // not a real PC until after assembly, so this mapping has to
7312 debugInfo.GetPC = func(b, v ssa.ID) int64 {
7314 case ssa.BlockStart.ID:
7315 if b == f.Entry.ID {
7316 return 0 // Start at the very beginning, at the assembler-generated prologue.
7317 // this should only happen for function args (ssa.OpArg)
7320 case ssa.BlockEnd.ID:
7321 blk := f.Blocks[idToIdx[b]]
7322 nv := len(blk.Values)
7323 return valueToProgAfter[blk.Values[nv-1].ID].Pc
7324 case ssa.FuncEnd.ID:
7325 return e.curfn.LSym.Size
7327 return valueToProgAfter[v].Pc
7332 // Resolve branches, and relax DefaultStmt into NotStmt
7333 for _, br := range s.Branches {
7334 br.P.To.SetTarget(s.bstart[br.B.ID])
7335 if br.P.Pos.IsStmt() != src.PosIsStmt {
7336 br.P.Pos = br.P.Pos.WithNotStmt()
7337 } else if v0 := br.B.FirstPossibleStmtValue(); v0 != nil && v0.Pos.Line() == br.P.Pos.Line() && v0.Pos.IsStmt() == src.PosIsStmt {
7338 br.P.Pos = br.P.Pos.WithNotStmt()
7343 // Resolve jump table destinations.
7344 for _, jt := range s.JumpTables {
7345 // Convert from *Block targets to *Prog targets.
7346 targets := make([]*obj.Prog, len(jt.Succs))
7347 for i, e := range jt.Succs {
7348 targets[i] = s.bstart[e.Block().ID]
7350 // Add to list of jump tables to be resolved at assembly time.
7351 // The assembler converts from *Prog entries to absolute addresses
7352 // once it knows instruction byte offsets.
7353 fi := pp.CurFunc.LSym.Func()
7354 fi.JumpTables = append(fi.JumpTables, obj.JumpTable{Sym: jt.Aux.(*obj.LSym), Targets: targets})
7357 if e.log { // spew to stdout
7359 for p := pp.Text; p != nil; p = p.Link {
7360 if p.Pos.IsKnown() && p.InnermostFilename() != filename {
7361 filename = p.InnermostFilename()
7362 f.Logf("# %s\n", filename)
7366 if v, ok := progToValue[p]; ok {
7368 } else if b, ok := progToBlock[p]; ok {
7371 s = " " // most value and branch strings are 2-3 characters long
7373 f.Logf(" %-6s\t%.5d (%s)\t%s\n", s, p.Pc, p.InnermostLineNumber(), p.InstructionString())
7376 if f.HTMLWriter != nil { // spew to ssa.html
7377 var buf strings.Builder
7378 buf.WriteString("<code>")
7379 buf.WriteString("<dl class=\"ssa-gen\">")
7381 for p := pp.Text; p != nil; p = p.Link {
7382 // Don't spam every line with the file name, which is often huge.
7383 // Only print changes, and "unknown" is not a change.
7384 if p.Pos.IsKnown() && p.InnermostFilename() != filename {
7385 filename = p.InnermostFilename()
7386 buf.WriteString("<dt class=\"ssa-prog-src\"></dt><dd class=\"ssa-prog\">")
7387 buf.WriteString(html.EscapeString("# " + filename))
7388 buf.WriteString("</dd>")
7391 buf.WriteString("<dt class=\"ssa-prog-src\">")
7392 if v, ok := progToValue[p]; ok {
7393 buf.WriteString(v.HTML())
7394 } else if b, ok := progToBlock[p]; ok {
7395 buf.WriteString("<b>" + b.HTML() + "</b>")
7397 buf.WriteString("</dt>")
7398 buf.WriteString("<dd class=\"ssa-prog\">")
7399 fmt.Fprintf(&buf, "%.5d <span class=\"l%v line-number\">(%s)</span> %s", p.Pc, p.InnermostLineNumber(), p.InnermostLineNumberHTML(), html.EscapeString(p.InstructionString()))
7400 buf.WriteString("</dd>")
7402 buf.WriteString("</dl>")
7403 buf.WriteString("</code>")
7404 f.HTMLWriter.WriteColumn("genssa", "genssa", "ssa-prog", buf.String())
7406 if ssa.GenssaDump[f.Name] {
7407 fi := f.DumpFileForPhase("genssa")
7410 // inliningDiffers if any filename changes or if any line number except the innermost (last index) changes.
7411 inliningDiffers := func(a, b []src.Pos) bool {
7412 if len(a) != len(b) {
7416 if a[i].Filename() != b[i].Filename() {
7419 if i != len(a)-1 && a[i].Line() != b[i].Line() {
7426 var allPosOld []src.Pos
7427 var allPos []src.Pos
7429 for p := pp.Text; p != nil; p = p.Link {
7430 if p.Pos.IsKnown() {
7432 p.Ctxt.AllPos(p.Pos, func(pos src.Pos) { allPos = append(allPos, pos) })
7433 if inliningDiffers(allPos, allPosOld) {
7434 for _, pos := range allPos {
7435 fmt.Fprintf(fi, "# %s:%d\n", pos.Filename(), pos.Line())
7437 allPos, allPosOld = allPosOld, allPos // swap, not copy, so that they do not share slice storage.
7442 if v, ok := progToValue[p]; ok {
7444 } else if b, ok := progToBlock[p]; ok {
7447 s = " " // most value and branch strings are 2-3 characters long
7449 fmt.Fprintf(fi, " %-6s\t%.5d %s\t%s\n", s, p.Pc, ssa.StmtString(p.Pos), p.InstructionString())
7457 f.HTMLWriter.Close()
7461 func defframe(s *State, e *ssafn, f *ssa.Func) {
7464 s.maxarg = types.RoundUp(s.maxarg, e.stkalign)
7465 frame := s.maxarg + e.stksize
7466 if Arch.PadFrame != nil {
7467 frame = Arch.PadFrame(frame)
7470 // Fill in argument and frame size.
7471 pp.Text.To.Type = obj.TYPE_TEXTSIZE
7472 pp.Text.To.Val = int32(types.RoundUp(f.OwnAux.ArgWidth(), int64(types.RegSize)))
7473 pp.Text.To.Offset = frame
7477 // Insert code to spill argument registers if the named slot may be partially
7478 // live. That is, the named slot is considered live by liveness analysis,
7479 // (because a part of it is live), but we may not spill all parts into the
7480 // slot. This can only happen with aggregate-typed arguments that are SSA-able
7481 // and not address-taken (for non-SSA-able or address-taken arguments we always
7483 // Note: spilling is unnecessary in the -N/no-optimize case, since all values
7484 // will be considered non-SSAable and spilled up front.
7485 // TODO(register args) Make liveness more fine-grained to that partial spilling is okay.
7486 if f.OwnAux.ABIInfo().InRegistersUsed() != 0 && base.Flag.N == 0 {
7487 // First, see if it is already spilled before it may be live. Look for a spill
7488 // in the entry block up to the first safepoint.
7489 type nameOff struct {
7493 partLiveArgsSpilled := make(map[nameOff]bool)
7494 for _, v := range f.Entry.Values {
7498 if v.Op != ssa.OpStoreReg || v.Args[0].Op != ssa.OpArgIntReg {
7501 n, off := ssa.AutoVar(v)
7502 if n.Class != ir.PPARAM || n.Addrtaken() || !ssa.CanSSA(n.Type()) || !s.partLiveArgs[n] {
7505 partLiveArgsSpilled[nameOff{n, off}] = true
7508 // Then, insert code to spill registers if not already.
7509 for _, a := range f.OwnAux.ABIInfo().InParams() {
7511 if n == nil || n.Addrtaken() || !ssa.CanSSA(n.Type()) || !s.partLiveArgs[n] || len(a.Registers) <= 1 {
7514 rts, offs := a.RegisterTypesAndOffsets()
7515 for i := range a.Registers {
7516 if !rts[i].HasPointers() {
7519 if partLiveArgsSpilled[nameOff{n, offs[i]}] {
7520 continue // already spilled
7522 reg := ssa.ObjRegForAbiReg(a.Registers[i], f.Config)
7523 p = Arch.SpillArgReg(pp, p, f, rts[i], reg, n, offs[i])
7528 // Insert code to zero ambiguously live variables so that the
7529 // garbage collector only sees initialized values when it
7530 // looks for pointers.
7533 // Opaque state for backend to use. Current backends use it to
7534 // keep track of which helper registers have been zeroed.
7537 // Iterate through declarations. Autos are sorted in decreasing
7538 // frame offset order.
7539 for _, n := range e.curfn.Dcl {
7543 if n.Class != ir.PAUTO {
7544 e.Fatalf(n.Pos(), "needzero class %d", n.Class)
7546 if n.Type().Size()%int64(types.PtrSize) != 0 || n.FrameOffset()%int64(types.PtrSize) != 0 || n.Type().Size() == 0 {
7547 e.Fatalf(n.Pos(), "var %L has size %d offset %d", n, n.Type().Size(), n.Offset_)
7550 if lo != hi && n.FrameOffset()+n.Type().Size() >= lo-int64(2*types.RegSize) {
7551 // Merge with range we already have.
7552 lo = n.FrameOffset()
7557 p = Arch.ZeroRange(pp, p, frame+lo, hi-lo, &state)
7560 lo = n.FrameOffset()
7561 hi = lo + n.Type().Size()
7564 // Zero final range.
7565 Arch.ZeroRange(pp, p, frame+lo, hi-lo, &state)
7568 // For generating consecutive jump instructions to model a specific branching
7569 type IndexJump struct {
7574 func (s *State) oneJump(b *ssa.Block, jump *IndexJump) {
7575 p := s.Br(jump.Jump, b.Succs[jump.Index].Block())
7579 // CombJump generates combinational instructions (2 at present) for a block jump,
7580 // thereby the behaviour of non-standard condition codes could be simulated
7581 func (s *State) CombJump(b, next *ssa.Block, jumps *[2][2]IndexJump) {
7583 case b.Succs[0].Block():
7584 s.oneJump(b, &jumps[0][0])
7585 s.oneJump(b, &jumps[0][1])
7586 case b.Succs[1].Block():
7587 s.oneJump(b, &jumps[1][0])
7588 s.oneJump(b, &jumps[1][1])
7591 if b.Likely != ssa.BranchUnlikely {
7592 s.oneJump(b, &jumps[1][0])
7593 s.oneJump(b, &jumps[1][1])
7594 q = s.Br(obj.AJMP, b.Succs[1].Block())
7596 s.oneJump(b, &jumps[0][0])
7597 s.oneJump(b, &jumps[0][1])
7598 q = s.Br(obj.AJMP, b.Succs[0].Block())
7604 // AddAux adds the offset in the aux fields (AuxInt and Aux) of v to a.
7605 func AddAux(a *obj.Addr, v *ssa.Value) {
7606 AddAux2(a, v, v.AuxInt)
7608 func AddAux2(a *obj.Addr, v *ssa.Value, offset int64) {
7609 if a.Type != obj.TYPE_MEM && a.Type != obj.TYPE_ADDR {
7610 v.Fatalf("bad AddAux addr %v", a)
7612 // add integer offset
7615 // If no additional symbol offset, we're done.
7619 // Add symbol's offset from its base register.
7620 switch n := v.Aux.(type) {
7622 a.Name = obj.NAME_EXTERN
7625 a.Name = obj.NAME_EXTERN
7628 if n.Class == ir.PPARAM || (n.Class == ir.PPARAMOUT && !n.IsOutputParamInRegisters()) {
7629 a.Name = obj.NAME_PARAM
7631 a.Name = obj.NAME_AUTO
7634 a.Offset += n.FrameOffset()
7636 v.Fatalf("aux in %s not implemented %#v", v, v.Aux)
7640 // extendIndex extends v to a full int width.
7641 // panic with the given kind if v does not fit in an int (only on 32-bit archs).
7642 func (s *state) extendIndex(idx, len *ssa.Value, kind ssa.BoundsKind, bounded bool) *ssa.Value {
7643 size := idx.Type.Size()
7644 if size == s.config.PtrSize {
7647 if size > s.config.PtrSize {
7648 // truncate 64-bit indexes on 32-bit pointer archs. Test the
7649 // high word and branch to out-of-bounds failure if it is not 0.
7651 if idx.Type.IsSigned() {
7652 lo = s.newValue1(ssa.OpInt64Lo, types.Types[types.TINT], idx)
7654 lo = s.newValue1(ssa.OpInt64Lo, types.Types[types.TUINT], idx)
7656 if bounded || base.Flag.B != 0 {
7659 bNext := s.f.NewBlock(ssa.BlockPlain)
7660 bPanic := s.f.NewBlock(ssa.BlockExit)
7661 hi := s.newValue1(ssa.OpInt64Hi, types.Types[types.TUINT32], idx)
7662 cmp := s.newValue2(ssa.OpEq32, types.Types[types.TBOOL], hi, s.constInt32(types.Types[types.TUINT32], 0))
7663 if !idx.Type.IsSigned() {
7665 case ssa.BoundsIndex:
7666 kind = ssa.BoundsIndexU
7667 case ssa.BoundsSliceAlen:
7668 kind = ssa.BoundsSliceAlenU
7669 case ssa.BoundsSliceAcap:
7670 kind = ssa.BoundsSliceAcapU
7671 case ssa.BoundsSliceB:
7672 kind = ssa.BoundsSliceBU
7673 case ssa.BoundsSlice3Alen:
7674 kind = ssa.BoundsSlice3AlenU
7675 case ssa.BoundsSlice3Acap:
7676 kind = ssa.BoundsSlice3AcapU
7677 case ssa.BoundsSlice3B:
7678 kind = ssa.BoundsSlice3BU
7679 case ssa.BoundsSlice3C:
7680 kind = ssa.BoundsSlice3CU
7684 b.Kind = ssa.BlockIf
7686 b.Likely = ssa.BranchLikely
7690 s.startBlock(bPanic)
7691 mem := s.newValue4I(ssa.OpPanicExtend, types.TypeMem, int64(kind), hi, lo, len, s.mem())
7692 s.endBlock().SetControl(mem)
7698 // Extend value to the required size
7700 if idx.Type.IsSigned() {
7701 switch 10*size + s.config.PtrSize {
7703 op = ssa.OpSignExt8to32
7705 op = ssa.OpSignExt8to64
7707 op = ssa.OpSignExt16to32
7709 op = ssa.OpSignExt16to64
7711 op = ssa.OpSignExt32to64
7713 s.Fatalf("bad signed index extension %s", idx.Type)
7716 switch 10*size + s.config.PtrSize {
7718 op = ssa.OpZeroExt8to32
7720 op = ssa.OpZeroExt8to64
7722 op = ssa.OpZeroExt16to32
7724 op = ssa.OpZeroExt16to64
7726 op = ssa.OpZeroExt32to64
7728 s.Fatalf("bad unsigned index extension %s", idx.Type)
7731 return s.newValue1(op, types.Types[types.TINT], idx)
7734 // CheckLoweredPhi checks that regalloc and stackalloc correctly handled phi values.
7735 // Called during ssaGenValue.
7736 func CheckLoweredPhi(v *ssa.Value) {
7737 if v.Op != ssa.OpPhi {
7738 v.Fatalf("CheckLoweredPhi called with non-phi value: %v", v.LongString())
7740 if v.Type.IsMemory() {
7744 loc := f.RegAlloc[v.ID]
7745 for _, a := range v.Args {
7746 if aloc := f.RegAlloc[a.ID]; aloc != loc { // TODO: .Equal() instead?
7747 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)
7752 // CheckLoweredGetClosurePtr checks that v is the first instruction in the function's entry block,
7753 // except for incoming in-register arguments.
7754 // The output of LoweredGetClosurePtr is generally hardwired to the correct register.
7755 // That register contains the closure pointer on closure entry.
7756 func CheckLoweredGetClosurePtr(v *ssa.Value) {
7757 entry := v.Block.Func.Entry
7758 if entry != v.Block {
7759 base.Fatalf("in %s, badly placed LoweredGetClosurePtr: %v %v", v.Block.Func.Name, v.Block, v)
7761 for _, w := range entry.Values {
7766 case ssa.OpArgIntReg, ssa.OpArgFloatReg:
7769 base.Fatalf("in %s, badly placed LoweredGetClosurePtr: %v %v", v.Block.Func.Name, v.Block, v)
7774 // CheckArgReg ensures that v is in the function's entry block.
7775 func CheckArgReg(v *ssa.Value) {
7776 entry := v.Block.Func.Entry
7777 if entry != v.Block {
7778 base.Fatalf("in %s, badly placed ArgIReg or ArgFReg: %v %v", v.Block.Func.Name, v.Block, v)
7782 func AddrAuto(a *obj.Addr, v *ssa.Value) {
7783 n, off := ssa.AutoVar(v)
7784 a.Type = obj.TYPE_MEM
7786 a.Reg = int16(Arch.REGSP)
7787 a.Offset = n.FrameOffset() + off
7788 if n.Class == ir.PPARAM || (n.Class == ir.PPARAMOUT && !n.IsOutputParamInRegisters()) {
7789 a.Name = obj.NAME_PARAM
7791 a.Name = obj.NAME_AUTO
7795 // Call returns a new CALL instruction for the SSA value v.
7796 // It uses PrepareCall to prepare the call.
7797 func (s *State) Call(v *ssa.Value) *obj.Prog {
7798 pPosIsStmt := s.pp.Pos.IsStmt() // The statement-ness fo the call comes from ssaGenState
7801 p := s.Prog(obj.ACALL)
7802 if pPosIsStmt == src.PosIsStmt {
7803 p.Pos = v.Pos.WithIsStmt()
7805 p.Pos = v.Pos.WithNotStmt()
7807 if sym, ok := v.Aux.(*ssa.AuxCall); ok && sym.Fn != nil {
7808 p.To.Type = obj.TYPE_MEM
7809 p.To.Name = obj.NAME_EXTERN
7812 // TODO(mdempsky): Can these differences be eliminated?
7813 switch Arch.LinkArch.Family {
7814 case sys.AMD64, sys.I386, sys.PPC64, sys.RISCV64, sys.S390X, sys.Wasm:
7815 p.To.Type = obj.TYPE_REG
7816 case sys.ARM, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64:
7817 p.To.Type = obj.TYPE_MEM
7819 base.Fatalf("unknown indirect call family")
7821 p.To.Reg = v.Args[0].Reg()
7826 // TailCall returns a new tail call instruction for the SSA value v.
7827 // It is like Call, but for a tail call.
7828 func (s *State) TailCall(v *ssa.Value) *obj.Prog {
7834 // PrepareCall prepares to emit a CALL instruction for v and does call-related bookkeeping.
7835 // It must be called immediately before emitting the actual CALL instruction,
7836 // since it emits PCDATA for the stack map at the call (calls are safe points).
7837 func (s *State) PrepareCall(v *ssa.Value) {
7838 idx := s.livenessMap.Get(v)
7839 if !idx.StackMapValid() {
7840 // See Liveness.hasStackMap.
7841 if sym, ok := v.Aux.(*ssa.AuxCall); !ok || !(sym.Fn == ir.Syms.WBZero || sym.Fn == ir.Syms.WBMove) {
7842 base.Fatalf("missing stack map index for %v", v.LongString())
7846 call, ok := v.Aux.(*ssa.AuxCall)
7849 // Record call graph information for nowritebarrierrec
7851 if nowritebarrierrecCheck != nil {
7852 nowritebarrierrecCheck.recordCall(s.pp.CurFunc, call.Fn, v.Pos)
7856 if s.maxarg < v.AuxInt {
7861 // UseArgs records the fact that an instruction needs a certain amount of
7862 // callee args space for its use.
7863 func (s *State) UseArgs(n int64) {
7869 // fieldIdx finds the index of the field referred to by the ODOT node n.
7870 func fieldIdx(n *ir.SelectorExpr) int {
7873 panic("ODOT's LHS is not a struct")
7876 for i, f := range t.Fields() {
7878 if f.Offset != n.Offset() {
7879 panic("field offset doesn't match")
7884 panic(fmt.Sprintf("can't find field in expr %v\n", n))
7886 // TODO: keep the result of this function somewhere in the ODOT Node
7887 // so we don't have to recompute it each time we need it.
7890 // ssafn holds frontend information about a function that the backend is processing.
7891 // It also exports a bunch of compiler services for the ssa backend.
7894 strings map[string]*obj.LSym // map from constant string to data symbols
7895 stksize int64 // stack size for current frame
7896 stkptrsize int64 // prefix of stack containing pointers
7898 // alignment for current frame.
7899 // NOTE: when stkalign > PtrSize, currently this only ensures the offsets of
7900 // objects in the stack frame are aligned. The stack pointer is still aligned
7904 log bool // print ssa debug to the stdout
7907 // StringData returns a symbol which
7908 // is the data component of a global string constant containing s.
7909 func (e *ssafn) StringData(s string) *obj.LSym {
7910 if aux, ok := e.strings[s]; ok {
7913 if e.strings == nil {
7914 e.strings = make(map[string]*obj.LSym)
7916 data := staticdata.StringSym(e.curfn.Pos(), s)
7921 // SplitSlot returns a slot representing the data of parent starting at offset.
7922 func (e *ssafn) SplitSlot(parent *ssa.LocalSlot, suffix string, offset int64, t *types.Type) ssa.LocalSlot {
7925 if node.Class != ir.PAUTO || node.Addrtaken() {
7926 // addressed things and non-autos retain their parents (i.e., cannot truly be split)
7927 return ssa.LocalSlot{N: node, Type: t, Off: parent.Off + offset}
7930 sym := &types.Sym{Name: node.Sym().Name + suffix, Pkg: types.LocalPkg}
7931 n := e.curfn.NewLocal(parent.N.Pos(), sym, t)
7933 n.SetEsc(ir.EscNever)
7935 return ssa.LocalSlot{N: n, Type: t, Off: 0, SplitOf: parent, SplitOffset: offset}
7938 // Logf logs a message from the compiler.
7939 func (e *ssafn) Logf(msg string, args ...interface{}) {
7941 fmt.Printf(msg, args...)
7945 func (e *ssafn) Log() bool {
7949 // Fatalf reports a compiler error and exits.
7950 func (e *ssafn) Fatalf(pos src.XPos, msg string, args ...interface{}) {
7952 nargs := append([]interface{}{ir.FuncName(e.curfn)}, args...)
7953 base.Fatalf("'%s': "+msg, nargs...)
7956 // Warnl reports a "warning", which is usually flag-triggered
7957 // logging output for the benefit of tests.
7958 func (e *ssafn) Warnl(pos src.XPos, fmt_ string, args ...interface{}) {
7959 base.WarnfAt(pos, fmt_, args...)
7962 func (e *ssafn) Debug_checknil() bool {
7963 return base.Debug.Nil != 0
7966 func (e *ssafn) UseWriteBarrier() bool {
7970 func (e *ssafn) Syslook(name string) *obj.LSym {
7972 case "goschedguarded":
7973 return ir.Syms.Goschedguarded
7974 case "writeBarrier":
7975 return ir.Syms.WriteBarrier
7977 return ir.Syms.WBZero
7979 return ir.Syms.WBMove
7980 case "cgoCheckMemmove":
7981 return ir.Syms.CgoCheckMemmove
7982 case "cgoCheckPtrWrite":
7983 return ir.Syms.CgoCheckPtrWrite
7985 e.Fatalf(src.NoXPos, "unknown Syslook func %v", name)
7989 func (e *ssafn) Func() *ir.Func {
7993 func clobberBase(n ir.Node) ir.Node {
7994 if n.Op() == ir.ODOT {
7995 n := n.(*ir.SelectorExpr)
7996 if n.X.Type().NumFields() == 1 {
7997 return clobberBase(n.X)
8000 if n.Op() == ir.OINDEX {
8001 n := n.(*ir.IndexExpr)
8002 if n.X.Type().IsArray() && n.X.Type().NumElem() == 1 {
8003 return clobberBase(n.X)
8009 // callTargetLSym returns the correct LSym to call 'callee' using its ABI.
8010 func callTargetLSym(callee *ir.Name) *obj.LSym {
8011 if callee.Func == nil {
8012 // TODO(austin): This happens in case of interface method I.M from imported package.
8013 // It's ABIInternal, and would be better if callee.Func was never nil and we didn't
8015 return callee.Linksym()
8018 return callee.LinksymABI(callee.Func.ABI)
8021 func min8(a, b int8) int8 {
8028 func max8(a, b int8) int8 {
8035 // deferStructFnField is the field index of _defer.fn.
8036 const deferStructFnField = 4
8038 var deferType *types.Type
8040 // deferstruct returns a type interchangeable with runtime._defer.
8041 // Make sure this stays in sync with runtime/runtime2.go:_defer.
8042 func deferstruct() *types.Type {
8043 if deferType != nil {
8047 makefield := func(name string, t *types.Type) *types.Field {
8048 sym := (*types.Pkg)(nil).Lookup(name)
8049 return types.NewField(src.NoXPos, sym, t)
8052 fields := []*types.Field{
8053 makefield("heap", types.Types[types.TBOOL]),
8054 makefield("rangefunc", types.Types[types.TBOOL]),
8055 makefield("sp", types.Types[types.TUINTPTR]),
8056 makefield("pc", types.Types[types.TUINTPTR]),
8057 // Note: the types here don't really matter. Defer structures
8058 // are always scanned explicitly during stack copying and GC,
8059 // so we make them uintptr type even though they are real pointers.
8060 makefield("fn", types.Types[types.TUINTPTR]),
8061 makefield("link", types.Types[types.TUINTPTR]),
8062 makefield("head", types.Types[types.TUINTPTR]),
8064 if name := fields[deferStructFnField].Sym.Name; name != "fn" {
8065 base.Fatalf("deferStructFnField is %q, not fn", name)
8068 n := ir.NewDeclNameAt(src.NoXPos, ir.OTYPE, ir.Pkgs.Runtime.Lookup("_defer"))
8069 typ := types.NewNamed(n)
8073 // build struct holding the above fields
8074 typ.SetUnderlying(types.NewStruct(fields))
8075 types.CalcStructSize(typ)
8081 // SpillSlotAddr uses LocalSlot information to initialize an obj.Addr
8082 // The resulting addr is used in a non-standard context -- in the prologue
8083 // of a function, before the frame has been constructed, so the standard
8084 // addressing for the parameters will be wrong.
8085 func SpillSlotAddr(spill ssa.Spill, baseReg int16, extraOffset int64) obj.Addr {
8087 Name: obj.NAME_NONE,
8090 Offset: spill.Offset + extraOffset,
8095 BoundsCheckFunc [ssa.BoundsKindCount]*obj.LSym
8096 ExtendCheckFunc [ssa.BoundsKindCount]*obj.LSym