▽

Source file src/cmd/compile/internal/ssagen/ssa.go

Documentation: cmd/compile/internal/ssagen

     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.
     4  
     5  package ssagen
     6  
     7  import (
     8  	"bufio"
     9  	"bytes"
    10  	"fmt"
    11  	"go/constant"
    12  	"html"
    13  	"internal/buildcfg"
    14  	"os"
    15  	"path/filepath"
    16  	"sort"
    17  	"strings"
    18  
    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/rttype"
    26  	"cmd/compile/internal/ssa"
    27  	"cmd/compile/internal/staticdata"
    28  	"cmd/compile/internal/typecheck"
    29  	"cmd/compile/internal/types"
    30  	"cmd/internal/obj"
    31  	"cmd/internal/objabi"
    32  	"cmd/internal/src"
    33  	"cmd/internal/sys"
    34  
    35  	rtabi "internal/abi"
    36  )
    37  
    38  var ssaConfig *ssa.Config
    39  var ssaCaches []ssa.Cache
    40  
    41  var ssaDump string     // early copy of $GOSSAFUNC; the func name to dump output for
    42  var ssaDir string      // optional destination for ssa dump file
    43  var ssaDumpStdout bool // whether to dump to stdout
    44  var ssaDumpCFG string  // generate CFGs for these phases
    45  const ssaDumpFile = "ssa.html"
    46  
    47  // ssaDumpInlined holds all inlined functions when ssaDump contains a function name.
    48  var ssaDumpInlined []*ir.Func
    49  
    50  func DumpInline(fn *ir.Func) {
    51  	if ssaDump != "" && ssaDump == ir.FuncName(fn) {
    52  		ssaDumpInlined = append(ssaDumpInlined, fn)
    53  	}
    54  }
    55  
    56  func InitEnv() {
    57  	ssaDump = os.Getenv("GOSSAFUNC")
    58  	ssaDir = os.Getenv("GOSSADIR")
    59  	if ssaDump != "" {
    60  		if strings.HasSuffix(ssaDump, "+") {
    61  			ssaDump = ssaDump[:len(ssaDump)-1]
    62  			ssaDumpStdout = true
    63  		}
    64  		spl := strings.Split(ssaDump, ":")
    65  		if len(spl) > 1 {
    66  			ssaDump = spl[0]
    67  			ssaDumpCFG = spl[1]
    68  		}
    69  	}
    70  }
    71  
    72  func InitConfig() {
    73  	types_ := ssa.NewTypes()
    74  
    75  	if Arch.SoftFloat {
    76  		softfloatInit()
    77  	}
    78  
    79  	// Generate a few pointer types that are uncommon in the frontend but common in the backend.
    80  	// Caching is disabled in the backend, so generating these here avoids allocations.
    81  	_ = types.NewPtr(types.Types[types.TINTER])                             // *interface{}
    82  	_ = types.NewPtr(types.NewPtr(types.Types[types.TSTRING]))              // **string
    83  	_ = types.NewPtr(types.NewSlice(types.Types[types.TINTER]))             // *[]interface{}
    84  	_ = types.NewPtr(types.NewPtr(types.ByteType))                          // **byte
    85  	_ = types.NewPtr(types.NewSlice(types.ByteType))                        // *[]byte
    86  	_ = types.NewPtr(types.NewSlice(types.Types[types.TSTRING]))            // *[]string
    87  	_ = types.NewPtr(types.NewPtr(types.NewPtr(types.Types[types.TUINT8]))) // ***uint8
    88  	_ = types.NewPtr(types.Types[types.TINT16])                             // *int16
    89  	_ = types.NewPtr(types.Types[types.TINT64])                             // *int64
    90  	_ = types.NewPtr(types.ErrorType)                                       // *error
    91  	_ = types.NewPtr(reflectdata.MapType())                                 // *runtime.hmap
    92  	_ = types.NewPtr(deferstruct())                                         // *runtime._defer
    93  	types.NewPtrCacheEnabled = false
    94  	ssaConfig = ssa.NewConfig(base.Ctxt.Arch.Name, *types_, base.Ctxt, base.Flag.N == 0, Arch.SoftFloat)
    95  	ssaConfig.Race = base.Flag.Race
    96  	ssaCaches = make([]ssa.Cache, base.Flag.LowerC)
    97  
    98  	// Set up some runtime functions we'll need to call.
    99  	ir.Syms.AssertE2I = typecheck.LookupRuntimeFunc("assertE2I")
   100  	ir.Syms.AssertE2I2 = typecheck.LookupRuntimeFunc("assertE2I2")
   101  	ir.Syms.AssertI2I = typecheck.LookupRuntimeFunc("assertI2I")
   102  	ir.Syms.AssertI2I2 = typecheck.LookupRuntimeFunc("assertI2I2")
   103  	ir.Syms.CgoCheckMemmove = typecheck.LookupRuntimeFunc("cgoCheckMemmove")
   104  	ir.Syms.CgoCheckPtrWrite = typecheck.LookupRuntimeFunc("cgoCheckPtrWrite")
   105  	ir.Syms.CheckPtrAlignment = typecheck.LookupRuntimeFunc("checkptrAlignment")
   106  	ir.Syms.Deferproc = typecheck.LookupRuntimeFunc("deferproc")
   107  	ir.Syms.Deferprocat = typecheck.LookupRuntimeFunc("deferprocat")
   108  	ir.Syms.DeferprocStack = typecheck.LookupRuntimeFunc("deferprocStack")
   109  	ir.Syms.Deferreturn = typecheck.LookupRuntimeFunc("deferreturn")
   110  	ir.Syms.Duffcopy = typecheck.LookupRuntimeFunc("duffcopy")
   111  	ir.Syms.Duffzero = typecheck.LookupRuntimeFunc("duffzero")
   112  	ir.Syms.GCWriteBarrier[0] = typecheck.LookupRuntimeFunc("gcWriteBarrier1")
   113  	ir.Syms.GCWriteBarrier[1] = typecheck.LookupRuntimeFunc("gcWriteBarrier2")
   114  	ir.Syms.GCWriteBarrier[2] = typecheck.LookupRuntimeFunc("gcWriteBarrier3")
   115  	ir.Syms.GCWriteBarrier[3] = typecheck.LookupRuntimeFunc("gcWriteBarrier4")
   116  	ir.Syms.GCWriteBarrier[4] = typecheck.LookupRuntimeFunc("gcWriteBarrier5")
   117  	ir.Syms.GCWriteBarrier[5] = typecheck.LookupRuntimeFunc("gcWriteBarrier6")
   118  	ir.Syms.GCWriteBarrier[6] = typecheck.LookupRuntimeFunc("gcWriteBarrier7")
   119  	ir.Syms.GCWriteBarrier[7] = typecheck.LookupRuntimeFunc("gcWriteBarrier8")
   120  	ir.Syms.Goschedguarded = typecheck.LookupRuntimeFunc("goschedguarded")
   121  	ir.Syms.Growslice = typecheck.LookupRuntimeFunc("growslice")
   122  	ir.Syms.InterfaceSwitch = typecheck.LookupRuntimeFunc("interfaceSwitch")
   123  	ir.Syms.Memmove = typecheck.LookupRuntimeFunc("memmove")
   124  	ir.Syms.Msanread = typecheck.LookupRuntimeFunc("msanread")
   125  	ir.Syms.Msanwrite = typecheck.LookupRuntimeFunc("msanwrite")
   126  	ir.Syms.Msanmove = typecheck.LookupRuntimeFunc("msanmove")
   127  	ir.Syms.Asanread = typecheck.LookupRuntimeFunc("asanread")
   128  	ir.Syms.Asanwrite = typecheck.LookupRuntimeFunc("asanwrite")
   129  	ir.Syms.Newobject = typecheck.LookupRuntimeFunc("newobject")
   130  	ir.Syms.Newproc = typecheck.LookupRuntimeFunc("newproc")
   131  	ir.Syms.Panicdivide = typecheck.LookupRuntimeFunc("panicdivide")
   132  	ir.Syms.PanicdottypeE = typecheck.LookupRuntimeFunc("panicdottypeE")
   133  	ir.Syms.PanicdottypeI = typecheck.LookupRuntimeFunc("panicdottypeI")
   134  	ir.Syms.Panicnildottype = typecheck.LookupRuntimeFunc("panicnildottype")
   135  	ir.Syms.Panicoverflow = typecheck.LookupRuntimeFunc("panicoverflow")
   136  	ir.Syms.Panicshift = typecheck.LookupRuntimeFunc("panicshift")
   137  	ir.Syms.Racefuncenter = typecheck.LookupRuntimeFunc("racefuncenter")
   138  	ir.Syms.Racefuncexit = typecheck.LookupRuntimeFunc("racefuncexit")
   139  	ir.Syms.Raceread = typecheck.LookupRuntimeFunc("raceread")
   140  	ir.Syms.Racereadrange = typecheck.LookupRuntimeFunc("racereadrange")
   141  	ir.Syms.Racewrite = typecheck.LookupRuntimeFunc("racewrite")
   142  	ir.Syms.Racewriterange = typecheck.LookupRuntimeFunc("racewriterange")
   143  	ir.Syms.TypeAssert = typecheck.LookupRuntimeFunc("typeAssert")
   144  	ir.Syms.WBZero = typecheck.LookupRuntimeFunc("wbZero")
   145  	ir.Syms.WBMove = typecheck.LookupRuntimeFunc("wbMove")
   146  	ir.Syms.X86HasPOPCNT = typecheck.LookupRuntimeVar("x86HasPOPCNT")       // bool
   147  	ir.Syms.X86HasSSE41 = typecheck.LookupRuntimeVar("x86HasSSE41")         // bool
   148  	ir.Syms.X86HasFMA = typecheck.LookupRuntimeVar("x86HasFMA")             // bool
   149  	ir.Syms.ARMHasVFPv4 = typecheck.LookupRuntimeVar("armHasVFPv4")         // bool
   150  	ir.Syms.ARM64HasATOMICS = typecheck.LookupRuntimeVar("arm64HasATOMICS") // bool
   151  	ir.Syms.Staticuint64s = typecheck.LookupRuntimeVar("staticuint64s")
   152  	ir.Syms.Typedmemmove = typecheck.LookupRuntimeFunc("typedmemmove")
   153  	ir.Syms.Udiv = typecheck.LookupRuntimeVar("udiv")                 // asm func with special ABI
   154  	ir.Syms.WriteBarrier = typecheck.LookupRuntimeVar("writeBarrier") // struct { bool; ... }
   155  	ir.Syms.Zerobase = typecheck.LookupRuntimeVar("zerobase")
   156  
   157  	if Arch.LinkArch.Family == sys.Wasm {
   158  		BoundsCheckFunc[ssa.BoundsIndex] = typecheck.LookupRuntimeFunc("goPanicIndex")
   159  		BoundsCheckFunc[ssa.BoundsIndexU] = typecheck.LookupRuntimeFunc("goPanicIndexU")
   160  		BoundsCheckFunc[ssa.BoundsSliceAlen] = typecheck.LookupRuntimeFunc("goPanicSliceAlen")
   161  		BoundsCheckFunc[ssa.BoundsSliceAlenU] = typecheck.LookupRuntimeFunc("goPanicSliceAlenU")
   162  		BoundsCheckFunc[ssa.BoundsSliceAcap] = typecheck.LookupRuntimeFunc("goPanicSliceAcap")
   163  		BoundsCheckFunc[ssa.BoundsSliceAcapU] = typecheck.LookupRuntimeFunc("goPanicSliceAcapU")
   164  		BoundsCheckFunc[ssa.BoundsSliceB] = typecheck.LookupRuntimeFunc("goPanicSliceB")
   165  		BoundsCheckFunc[ssa.BoundsSliceBU] = typecheck.LookupRuntimeFunc("goPanicSliceBU")
   166  		BoundsCheckFunc[ssa.BoundsSlice3Alen] = typecheck.LookupRuntimeFunc("goPanicSlice3Alen")
   167  		BoundsCheckFunc[ssa.BoundsSlice3AlenU] = typecheck.LookupRuntimeFunc("goPanicSlice3AlenU")
   168  		BoundsCheckFunc[ssa.BoundsSlice3Acap] = typecheck.LookupRuntimeFunc("goPanicSlice3Acap")
   169  		BoundsCheckFunc[ssa.BoundsSlice3AcapU] = typecheck.LookupRuntimeFunc("goPanicSlice3AcapU")
   170  		BoundsCheckFunc[ssa.BoundsSlice3B] = typecheck.LookupRuntimeFunc("goPanicSlice3B")
   171  		BoundsCheckFunc[ssa.BoundsSlice3BU] = typecheck.LookupRuntimeFunc("goPanicSlice3BU")
   172  		BoundsCheckFunc[ssa.BoundsSlice3C] = typecheck.LookupRuntimeFunc("goPanicSlice3C")
   173  		BoundsCheckFunc[ssa.BoundsSlice3CU] = typecheck.LookupRuntimeFunc("goPanicSlice3CU")
   174  		BoundsCheckFunc[ssa.BoundsConvert] = typecheck.LookupRuntimeFunc("goPanicSliceConvert")
   175  	} else {
   176  		BoundsCheckFunc[ssa.BoundsIndex] = typecheck.LookupRuntimeFunc("panicIndex")
   177  		BoundsCheckFunc[ssa.BoundsIndexU] = typecheck.LookupRuntimeFunc("panicIndexU")
   178  		BoundsCheckFunc[ssa.BoundsSliceAlen] = typecheck.LookupRuntimeFunc("panicSliceAlen")
   179  		BoundsCheckFunc[ssa.BoundsSliceAlenU] = typecheck.LookupRuntimeFunc("panicSliceAlenU")
   180  		BoundsCheckFunc[ssa.BoundsSliceAcap] = typecheck.LookupRuntimeFunc("panicSliceAcap")
   181  		BoundsCheckFunc[ssa.BoundsSliceAcapU] = typecheck.LookupRuntimeFunc("panicSliceAcapU")
   182  		BoundsCheckFunc[ssa.BoundsSliceB] = typecheck.LookupRuntimeFunc("panicSliceB")
   183  		BoundsCheckFunc[ssa.BoundsSliceBU] = typecheck.LookupRuntimeFunc("panicSliceBU")
   184  		BoundsCheckFunc[ssa.BoundsSlice3Alen] = typecheck.LookupRuntimeFunc("panicSlice3Alen")
   185  		BoundsCheckFunc[ssa.BoundsSlice3AlenU] = typecheck.LookupRuntimeFunc("panicSlice3AlenU")
   186  		BoundsCheckFunc[ssa.BoundsSlice3Acap] = typecheck.LookupRuntimeFunc("panicSlice3Acap")
   187  		BoundsCheckFunc[ssa.BoundsSlice3AcapU] = typecheck.LookupRuntimeFunc("panicSlice3AcapU")
   188  		BoundsCheckFunc[ssa.BoundsSlice3B] = typecheck.LookupRuntimeFunc("panicSlice3B")
   189  		BoundsCheckFunc[ssa.BoundsSlice3BU] = typecheck.LookupRuntimeFunc("panicSlice3BU")
   190  		BoundsCheckFunc[ssa.BoundsSlice3C] = typecheck.LookupRuntimeFunc("panicSlice3C")
   191  		BoundsCheckFunc[ssa.BoundsSlice3CU] = typecheck.LookupRuntimeFunc("panicSlice3CU")
   192  		BoundsCheckFunc[ssa.BoundsConvert] = typecheck.LookupRuntimeFunc("panicSliceConvert")
   193  	}
   194  	if Arch.LinkArch.PtrSize == 4 {
   195  		ExtendCheckFunc[ssa.BoundsIndex] = typecheck.LookupRuntimeVar("panicExtendIndex")
   196  		ExtendCheckFunc[ssa.BoundsIndexU] = typecheck.LookupRuntimeVar("panicExtendIndexU")
   197  		ExtendCheckFunc[ssa.BoundsSliceAlen] = typecheck.LookupRuntimeVar("panicExtendSliceAlen")
   198  		ExtendCheckFunc[ssa.BoundsSliceAlenU] = typecheck.LookupRuntimeVar("panicExtendSliceAlenU")
   199  		ExtendCheckFunc[ssa.BoundsSliceAcap] = typecheck.LookupRuntimeVar("panicExtendSliceAcap")
   200  		ExtendCheckFunc[ssa.BoundsSliceAcapU] = typecheck.LookupRuntimeVar("panicExtendSliceAcapU")
   201  		ExtendCheckFunc[ssa.BoundsSliceB] = typecheck.LookupRuntimeVar("panicExtendSliceB")
   202  		ExtendCheckFunc[ssa.BoundsSliceBU] = typecheck.LookupRuntimeVar("panicExtendSliceBU")
   203  		ExtendCheckFunc[ssa.BoundsSlice3Alen] = typecheck.LookupRuntimeVar("panicExtendSlice3Alen")
   204  		ExtendCheckFunc[ssa.BoundsSlice3AlenU] = typecheck.LookupRuntimeVar("panicExtendSlice3AlenU")
   205  		ExtendCheckFunc[ssa.BoundsSlice3Acap] = typecheck.LookupRuntimeVar("panicExtendSlice3Acap")
   206  		ExtendCheckFunc[ssa.BoundsSlice3AcapU] = typecheck.LookupRuntimeVar("panicExtendSlice3AcapU")
   207  		ExtendCheckFunc[ssa.BoundsSlice3B] = typecheck.LookupRuntimeVar("panicExtendSlice3B")
   208  		ExtendCheckFunc[ssa.BoundsSlice3BU] = typecheck.LookupRuntimeVar("panicExtendSlice3BU")
   209  		ExtendCheckFunc[ssa.BoundsSlice3C] = typecheck.LookupRuntimeVar("panicExtendSlice3C")
   210  		ExtendCheckFunc[ssa.BoundsSlice3CU] = typecheck.LookupRuntimeVar("panicExtendSlice3CU")
   211  	}
   212  
   213  	// Wasm (all asm funcs with special ABIs)
   214  	ir.Syms.WasmDiv = typecheck.LookupRuntimeVar("wasmDiv")
   215  	ir.Syms.WasmTruncS = typecheck.LookupRuntimeVar("wasmTruncS")
   216  	ir.Syms.WasmTruncU = typecheck.LookupRuntimeVar("wasmTruncU")
   217  	ir.Syms.SigPanic = typecheck.LookupRuntimeFunc("sigpanic")
   218  }
   219  
   220  // AbiForBodylessFuncStackMap returns the ABI for a bodyless function's stack map.
   221  // This is not necessarily the ABI used to call it.
   222  // Currently (1.17 dev) such a stack map is always ABI0;
   223  // any ABI wrapper that is present is nosplit, hence a precise
   224  // stack map is not needed there (the parameters survive only long
   225  // enough to call the wrapped assembly function).
   226  // This always returns a freshly copied ABI.
   227  func AbiForBodylessFuncStackMap(fn *ir.Func) *abi.ABIConfig {
   228  	return ssaConfig.ABI0.Copy() // No idea what races will result, be safe
   229  }
   230  
   231  // abiForFunc implements ABI policy for a function, but does not return a copy of the ABI.
   232  // Passing a nil function returns the default ABI based on experiment configuration.
   233  func abiForFunc(fn *ir.Func, abi0, abi1 *abi.ABIConfig) *abi.ABIConfig {
   234  	if buildcfg.Experiment.RegabiArgs {
   235  		// Select the ABI based on the function's defining ABI.
   236  		if fn == nil {
   237  			return abi1
   238  		}
   239  		switch fn.ABI {
   240  		case obj.ABI0:
   241  			return abi0
   242  		case obj.ABIInternal:
   243  			// TODO(austin): Clean up the nomenclature here.
   244  			// It's not clear that "abi1" is ABIInternal.
   245  			return abi1
   246  		}
   247  		base.Fatalf("function %v has unknown ABI %v", fn, fn.ABI)
   248  		panic("not reachable")
   249  	}
   250  
   251  	a := abi0
   252  	if fn != nil {
   253  		if fn.Pragma&ir.RegisterParams != 0 { // TODO(register args) remove after register abi is working
   254  			a = abi1
   255  		}
   256  	}
   257  	return a
   258  }
   259  
   260  // emitOpenDeferInfo emits FUNCDATA information about the defers in a function
   261  // that is using open-coded defers.  This funcdata is used to determine the active
   262  // defers in a function and execute those defers during panic processing.
   263  //
   264  // The funcdata is all encoded in varints (since values will almost always be less than
   265  // 128, but stack offsets could potentially be up to 2Gbyte). All "locations" (offsets)
   266  // for stack variables are specified as the number of bytes below varp (pointer to the
   267  // top of the local variables) for their starting address. The format is:
   268  //
   269  //   - Offset of the deferBits variable
   270  //   - Offset of the first closure slot (the rest are laid out consecutively).
   271  func (s *state) emitOpenDeferInfo() {
   272  	firstOffset := s.openDefers[0].closureNode.FrameOffset()
   273  
   274  	// Verify that cmpstackvarlt laid out the slots in order.
   275  	for i, r := range s.openDefers {
   276  		have := r.closureNode.FrameOffset()
   277  		want := firstOffset + int64(i)*int64(types.PtrSize)
   278  		if have != want {
   279  			base.FatalfAt(s.curfn.Pos(), "unexpected frame offset for open-coded defer slot #%v: have %v, want %v", i, have, want)
   280  		}
   281  	}
   282  
   283  	x := base.Ctxt.Lookup(s.curfn.LSym.Name + ".opendefer")
   284  	x.Set(obj.AttrContentAddressable, true)
   285  	s.curfn.LSym.Func().OpenCodedDeferInfo = x
   286  
   287  	off := 0
   288  	off = objw.Uvarint(x, off, uint64(-s.deferBitsTemp.FrameOffset()))
   289  	off = objw.Uvarint(x, off, uint64(-firstOffset))
   290  }
   291  
   292  // buildssa builds an SSA function for fn.
   293  // worker indicates which of the backend workers is doing the processing.
   294  func buildssa(fn *ir.Func, worker int, isPgoHot bool) *ssa.Func {
   295  	name := ir.FuncName(fn)
   296  
   297  	abiSelf := abiForFunc(fn, ssaConfig.ABI0, ssaConfig.ABI1)
   298  
   299  	printssa := false
   300  	// match either a simple name e.g. "(*Reader).Reset", package.name e.g. "compress/gzip.(*Reader).Reset", or subpackage name "gzip.(*Reader).Reset"
   301  	// optionally allows an ABI suffix specification in the GOSSAHASH, e.g. "(*Reader).Reset<0>" etc
   302  	if strings.Contains(ssaDump, name) { // in all the cases the function name is entirely contained within the GOSSAFUNC string.
   303  		nameOptABI := name
   304  		if strings.Contains(ssaDump, ",") { // ABI specification
   305  			nameOptABI = ssa.FuncNameABI(name, abiSelf.Which())
   306  		} else if strings.HasSuffix(ssaDump, ">") { // if they use the linker syntax instead....
   307  			l := len(ssaDump)
   308  			if l >= 3 && ssaDump[l-3] == '<' {
   309  				nameOptABI = ssa.FuncNameABI(name, abiSelf.Which())
   310  				ssaDump = ssaDump[:l-3] + "," + ssaDump[l-2:l-1]
   311  			}
   312  		}
   313  		pkgDotName := base.Ctxt.Pkgpath + "." + nameOptABI
   314  		printssa = nameOptABI == ssaDump || // "(*Reader).Reset"
   315  			pkgDotName == ssaDump || // "compress/gzip.(*Reader).Reset"
   316  			strings.HasSuffix(pkgDotName, ssaDump) && strings.HasSuffix(pkgDotName, "/"+ssaDump) // "gzip.(*Reader).Reset"
   317  	}
   318  
   319  	var astBuf *bytes.Buffer
   320  	if printssa {
   321  		astBuf = &bytes.Buffer{}
   322  		ir.FDumpList(astBuf, "buildssa-body", fn.Body)
   323  		if ssaDumpStdout {
   324  			fmt.Println("generating SSA for", name)
   325  			fmt.Print(astBuf.String())
   326  		}
   327  	}
   328  
   329  	var s state
   330  	s.pushLine(fn.Pos())
   331  	defer s.popLine()
   332  
   333  	s.hasdefer = fn.HasDefer()
   334  	if fn.Pragma&ir.CgoUnsafeArgs != 0 {
   335  		s.cgoUnsafeArgs = true
   336  	}
   337  	s.checkPtrEnabled = ir.ShouldCheckPtr(fn, 1)
   338  
   339  	if base.Flag.Cfg.Instrumenting && fn.Pragma&ir.Norace == 0 && !fn.Linksym().ABIWrapper() {
   340  		if !base.Flag.Race || !objabi.LookupPkgSpecial(fn.Sym().Pkg.Path).NoRaceFunc {
   341  			s.instrumentMemory = true
   342  		}
   343  		if base.Flag.Race {
   344  			s.instrumentEnterExit = true
   345  		}
   346  	}
   347  
   348  	fe := ssafn{
   349  		curfn: fn,
   350  		log:   printssa && ssaDumpStdout,
   351  	}
   352  	s.curfn = fn
   353  
   354  	cache := &ssaCaches[worker]
   355  	cache.Reset()
   356  
   357  	s.f = ssaConfig.NewFunc(&fe, cache)
   358  	s.config = ssaConfig
   359  	s.f.Type = fn.Type()
   360  	s.f.Name = name
   361  	s.f.PrintOrHtmlSSA = printssa
   362  	if fn.Pragma&ir.Nosplit != 0 {
   363  		s.f.NoSplit = true
   364  	}
   365  	s.f.ABI0 = ssaConfig.ABI0
   366  	s.f.ABI1 = ssaConfig.ABI1
   367  	s.f.ABIDefault = abiForFunc(nil, ssaConfig.ABI0, ssaConfig.ABI1)
   368  	s.f.ABISelf = abiSelf
   369  
   370  	s.panics = map[funcLine]*ssa.Block{}
   371  	s.softFloat = s.config.SoftFloat
   372  
   373  	// Allocate starting block
   374  	s.f.Entry = s.f.NewBlock(ssa.BlockPlain)
   375  	s.f.Entry.Pos = fn.Pos()
   376  	s.f.IsPgoHot = isPgoHot
   377  
   378  	if printssa {
   379  		ssaDF := ssaDumpFile
   380  		if ssaDir != "" {
   381  			ssaDF = filepath.Join(ssaDir, base.Ctxt.Pkgpath+"."+s.f.NameABI()+".html")
   382  			ssaD := filepath.Dir(ssaDF)
   383  			os.MkdirAll(ssaD, 0755)
   384  		}
   385  		s.f.HTMLWriter = ssa.NewHTMLWriter(ssaDF, s.f, ssaDumpCFG)
   386  		// TODO: generate and print a mapping from nodes to values and blocks
   387  		dumpSourcesColumn(s.f.HTMLWriter, fn)
   388  		s.f.HTMLWriter.WriteAST("AST", astBuf)
   389  	}
   390  
   391  	// Allocate starting values
   392  	s.labels = map[string]*ssaLabel{}
   393  	s.fwdVars = map[ir.Node]*ssa.Value{}
   394  	s.startmem = s.entryNewValue0(ssa.OpInitMem, types.TypeMem)
   395  
   396  	s.hasOpenDefers = base.Flag.N == 0 && s.hasdefer && !s.curfn.OpenCodedDeferDisallowed()
   397  	switch {
   398  	case base.Debug.NoOpenDefer != 0:
   399  		s.hasOpenDefers = false
   400  	case s.hasOpenDefers && (base.Ctxt.Flag_shared || base.Ctxt.Flag_dynlink) && base.Ctxt.Arch.Name == "386":
   401  		// Don't support open-coded defers for 386 ONLY when using shared
   402  		// libraries, because there is extra code (added by rewriteToUseGot())
   403  		// preceding the deferreturn/ret code that we don't track correctly.
   404  		s.hasOpenDefers = false
   405  	}
   406  	if s.hasOpenDefers && s.instrumentEnterExit {
   407  		// Skip doing open defers if we need to instrument function
   408  		// returns for the race detector, since we will not generate that
   409  		// code in the case of the extra deferreturn/ret segment.
   410  		s.hasOpenDefers = false
   411  	}
   412  	if s.hasOpenDefers {
   413  		// Similarly, skip if there are any heap-allocated result
   414  		// parameters that need to be copied back to their stack slots.
   415  		for _, f := range s.curfn.Type().Results() {
   416  			if !f.Nname.(*ir.Name).OnStack() {
   417  				s.hasOpenDefers = false
   418  				break
   419  			}
   420  		}
   421  	}
   422  	if s.hasOpenDefers &&
   423  		s.curfn.NumReturns*s.curfn.NumDefers > 15 {
   424  		// Since we are generating defer calls at every exit for
   425  		// open-coded defers, skip doing open-coded defers if there are
   426  		// too many returns (especially if there are multiple defers).
   427  		// Open-coded defers are most important for improving performance
   428  		// for smaller functions (which don't have many returns).
   429  		s.hasOpenDefers = false
   430  	}
   431  
   432  	s.sp = s.entryNewValue0(ssa.OpSP, types.Types[types.TUINTPTR]) // TODO: use generic pointer type (unsafe.Pointer?) instead
   433  	s.sb = s.entryNewValue0(ssa.OpSB, types.Types[types.TUINTPTR])
   434  
   435  	s.startBlock(s.f.Entry)
   436  	s.vars[memVar] = s.startmem
   437  	if s.hasOpenDefers {
   438  		// Create the deferBits variable and stack slot.  deferBits is a
   439  		// bitmask showing which of the open-coded defers in this function
   440  		// have been activated.
   441  		deferBitsTemp := typecheck.TempAt(src.NoXPos, s.curfn, types.Types[types.TUINT8])
   442  		deferBitsTemp.SetAddrtaken(true)
   443  		s.deferBitsTemp = deferBitsTemp
   444  		// For this value, AuxInt is initialized to zero by default
   445  		startDeferBits := s.entryNewValue0(ssa.OpConst8, types.Types[types.TUINT8])
   446  		s.vars[deferBitsVar] = startDeferBits
   447  		s.deferBitsAddr = s.addr(deferBitsTemp)
   448  		s.store(types.Types[types.TUINT8], s.deferBitsAddr, startDeferBits)
   449  		// Make sure that the deferBits stack slot is kept alive (for use
   450  		// by panics) and stores to deferBits are not eliminated, even if
   451  		// all checking code on deferBits in the function exit can be
   452  		// eliminated, because the defer statements were all
   453  		// unconditional.
   454  		s.vars[memVar] = s.newValue1Apos(ssa.OpVarLive, types.TypeMem, deferBitsTemp, s.mem(), false)
   455  	}
   456  
   457  	var params *abi.ABIParamResultInfo
   458  	params = s.f.ABISelf.ABIAnalyze(fn.Type(), true)
   459  
   460  	// The backend's stackframe pass prunes away entries from the fn's
   461  	// Dcl list, including PARAMOUT nodes that correspond to output
   462  	// params passed in registers. Walk the Dcl list and capture these
   463  	// nodes to a side list, so that we'll have them available during
   464  	// DWARF-gen later on. See issue 48573 for more details.
   465  	var debugInfo ssa.FuncDebug
   466  	for _, n := range fn.Dcl {
   467  		if n.Class == ir.PPARAMOUT && n.IsOutputParamInRegisters() {
   468  			debugInfo.RegOutputParams = append(debugInfo.RegOutputParams, n)
   469  		}
   470  	}
   471  	fn.DebugInfo = &debugInfo
   472  
   473  	// Generate addresses of local declarations
   474  	s.decladdrs = map[*ir.Name]*ssa.Value{}
   475  	for _, n := range fn.Dcl {
   476  		switch n.Class {
   477  		case ir.PPARAM:
   478  			// Be aware that blank and unnamed input parameters will not appear here, but do appear in the type
   479  			s.decladdrs[n] = s.entryNewValue2A(ssa.OpLocalAddr, types.NewPtr(n.Type()), n, s.sp, s.startmem)
   480  		case ir.PPARAMOUT:
   481  			s.decladdrs[n] = s.entryNewValue2A(ssa.OpLocalAddr, types.NewPtr(n.Type()), n, s.sp, s.startmem)
   482  		case ir.PAUTO:
   483  			// processed at each use, to prevent Addr coming
   484  			// before the decl.
   485  		default:
   486  			s.Fatalf("local variable with class %v unimplemented", n.Class)
   487  		}
   488  	}
   489  
   490  	s.f.OwnAux = ssa.OwnAuxCall(fn.LSym, params)
   491  
   492  	// Populate SSAable arguments.
   493  	for _, n := range fn.Dcl {
   494  		if n.Class == ir.PPARAM {
   495  			if s.canSSA(n) {
   496  				v := s.newValue0A(ssa.OpArg, n.Type(), n)
   497  				s.vars[n] = v
   498  				s.addNamedValue(n, v) // This helps with debugging information, not needed for compilation itself.
   499  			} else { // address was taken AND/OR too large for SSA
   500  				paramAssignment := ssa.ParamAssignmentForArgName(s.f, n)
   501  				if len(paramAssignment.Registers) > 0 {
   502  					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.
   503  						v := s.newValue0A(ssa.OpArg, n.Type(), n)
   504  						s.store(n.Type(), s.decladdrs[n], v)
   505  					} else { // Too big for SSA.
   506  						// Brute force, and early, do a bunch of stores from registers
   507  						// Note that expand calls knows about this and doesn't trouble itself with larger-than-SSA-able Args in registers.
   508  						s.storeParameterRegsToStack(s.f.ABISelf, paramAssignment, n, s.decladdrs[n], false)
   509  					}
   510  				}
   511  			}
   512  		}
   513  	}
   514  
   515  	// Populate closure variables.
   516  	if fn.Needctxt() {
   517  		clo := s.entryNewValue0(ssa.OpGetClosurePtr, s.f.Config.Types.BytePtr)
   518  		if fn.RangeParent != nil {
   519  			// For a range body closure, keep its closure pointer live on the
   520  			// stack with a special name, so the debugger can look for it and
   521  			// find the parent frame.
   522  			sym := &types.Sym{Name: ".closureptr", Pkg: types.LocalPkg}
   523  			cloSlot := s.curfn.NewLocal(src.NoXPos, sym, s.f.Config.Types.BytePtr)
   524  			cloSlot.SetUsed(true)
   525  			cloSlot.SetEsc(ir.EscNever)
   526  			cloSlot.SetAddrtaken(true)
   527  			s.f.CloSlot = cloSlot
   528  			s.vars[memVar] = s.newValue1Apos(ssa.OpVarDef, types.TypeMem, cloSlot, s.mem(), false)
   529  			addr := s.addr(cloSlot)
   530  			s.store(s.f.Config.Types.BytePtr, addr, clo)
   531  			// Keep it from being dead-store eliminated.
   532  			s.vars[memVar] = s.newValue1Apos(ssa.OpVarLive, types.TypeMem, cloSlot, s.mem(), false)
   533  		}
   534  		csiter := typecheck.NewClosureStructIter(fn.ClosureVars)
   535  		for {
   536  			n, typ, offset := csiter.Next()
   537  			if n == nil {
   538  				break
   539  			}
   540  
   541  			ptr := s.newValue1I(ssa.OpOffPtr, types.NewPtr(typ), offset, clo)
   542  
   543  			// If n is a small variable captured by value, promote
   544  			// it to PAUTO so it can be converted to SSA.
   545  			//
   546  			// Note: While we never capture a variable by value if
   547  			// the user took its address, we may have generated
   548  			// runtime calls that did (#43701). Since we don't
   549  			// convert Addrtaken variables to SSA anyway, no point
   550  			// in promoting them either.
   551  			if n.Byval() && !n.Addrtaken() && ssa.CanSSA(n.Type()) {
   552  				n.Class = ir.PAUTO
   553  				fn.Dcl = append(fn.Dcl, n)
   554  				s.assign(n, s.load(n.Type(), ptr), false, 0)
   555  				continue
   556  			}
   557  
   558  			if !n.Byval() {
   559  				ptr = s.load(typ, ptr)
   560  			}
   561  			s.setHeapaddr(fn.Pos(), n, ptr)
   562  		}
   563  	}
   564  
   565  	// Convert the AST-based IR to the SSA-based IR
   566  	if s.instrumentEnterExit {
   567  		s.rtcall(ir.Syms.Racefuncenter, true, nil, s.newValue0(ssa.OpGetCallerPC, types.Types[types.TUINTPTR]))
   568  	}
   569  	s.zeroResults()
   570  	s.paramsToHeap()
   571  	s.stmtList(fn.Body)
   572  
   573  	// fallthrough to exit
   574  	if s.curBlock != nil {
   575  		s.pushLine(fn.Endlineno)
   576  		s.exit()
   577  		s.popLine()
   578  	}
   579  
   580  	for _, b := range s.f.Blocks {
   581  		if b.Pos != src.NoXPos {
   582  			s.updateUnsetPredPos(b)
   583  		}
   584  	}
   585  
   586  	s.f.HTMLWriter.WritePhase("before insert phis", "before insert phis")
   587  
   588  	s.insertPhis()
   589  
   590  	// Main call to ssa package to compile function
   591  	ssa.Compile(s.f)
   592  
   593  	fe.AllocFrame(s.f)
   594  
   595  	if len(s.openDefers) != 0 {
   596  		s.emitOpenDeferInfo()
   597  	}
   598  
   599  	// Record incoming parameter spill information for morestack calls emitted in the assembler.
   600  	// This is done here, using all the parameters (used, partially used, and unused) because
   601  	// it mimics the behavior of the former ABI (everything stored) and because it's not 100%
   602  	// clear if naming conventions are respected in autogenerated code.
   603  	// TODO figure out exactly what's unused, don't spill it. Make liveness fine-grained, also.
   604  	for _, p := range params.InParams() {
   605  		typs, offs := p.RegisterTypesAndOffsets()
   606  		for i, t := range typs {
   607  			o := offs[i]                // offset within parameter
   608  			fo := p.FrameOffset(params) // offset of parameter in frame
   609  			reg := ssa.ObjRegForAbiReg(p.Registers[i], s.f.Config)
   610  			s.f.RegArgs = append(s.f.RegArgs, ssa.Spill{Reg: reg, Offset: fo + o, Type: t})
   611  		}
   612  	}
   613  
   614  	return s.f
   615  }
   616  
   617  func (s *state) storeParameterRegsToStack(abi *abi.ABIConfig, paramAssignment *abi.ABIParamAssignment, n *ir.Name, addr *ssa.Value, pointersOnly bool) {
   618  	typs, offs := paramAssignment.RegisterTypesAndOffsets()
   619  	for i, t := range typs {
   620  		if pointersOnly && !t.IsPtrShaped() {
   621  			continue
   622  		}
   623  		r := paramAssignment.Registers[i]
   624  		o := offs[i]
   625  		op, reg := ssa.ArgOpAndRegisterFor(r, abi)
   626  		aux := &ssa.AuxNameOffset{Name: n, Offset: o}
   627  		v := s.newValue0I(op, t, reg)
   628  		v.Aux = aux
   629  		p := s.newValue1I(ssa.OpOffPtr, types.NewPtr(t), o, addr)
   630  		s.store(t, p, v)
   631  	}
   632  }
   633  
   634  // zeroResults zeros the return values at the start of the function.
   635  // We need to do this very early in the function.  Defer might stop a
   636  // panic and show the return values as they exist at the time of
   637  // panic.  For precise stacks, the garbage collector assumes results
   638  // are always live, so we need to zero them before any allocations,
   639  // even allocations to move params/results to the heap.
   640  func (s *state) zeroResults() {
   641  	for _, f := range s.curfn.Type().Results() {
   642  		n := f.Nname.(*ir.Name)
   643  		if !n.OnStack() {
   644  			// The local which points to the return value is the
   645  			// thing that needs zeroing. This is already handled
   646  			// by a Needzero annotation in plive.go:(*liveness).epilogue.
   647  			continue
   648  		}
   649  		// Zero the stack location containing f.
   650  		if typ := n.Type(); ssa.CanSSA(typ) {
   651  			s.assign(n, s.zeroVal(typ), false, 0)
   652  		} else {
   653  			if typ.HasPointers() || ssa.IsMergeCandidate(n) {
   654  				s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, n, s.mem())
   655  			}
   656  			s.zero(n.Type(), s.decladdrs[n])
   657  		}
   658  	}
   659  }
   660  
   661  // paramsToHeap produces code to allocate memory for heap-escaped parameters
   662  // and to copy non-result parameters' values from the stack.
   663  func (s *state) paramsToHeap() {
   664  	do := func(params []*types.Field) {
   665  		for _, f := range params {
   666  			if f.Nname == nil {
   667  				continue // anonymous or blank parameter
   668  			}
   669  			n := f.Nname.(*ir.Name)
   670  			if ir.IsBlank(n) || n.OnStack() {
   671  				continue
   672  			}
   673  			s.newHeapaddr(n)
   674  			if n.Class == ir.PPARAM {
   675  				s.move(n.Type(), s.expr(n.Heapaddr), s.decladdrs[n])
   676  			}
   677  		}
   678  	}
   679  
   680  	typ := s.curfn.Type()
   681  	do(typ.Recvs())
   682  	do(typ.Params())
   683  	do(typ.Results())
   684  }
   685  
   686  // newHeapaddr allocates heap memory for n and sets its heap address.
   687  func (s *state) newHeapaddr(n *ir.Name) {
   688  	s.setHeapaddr(n.Pos(), n, s.newObject(n.Type(), nil))
   689  }
   690  
   691  // setHeapaddr allocates a new PAUTO variable to store ptr (which must be non-nil)
   692  // and then sets it as n's heap address.
   693  func (s *state) setHeapaddr(pos src.XPos, n *ir.Name, ptr *ssa.Value) {
   694  	if !ptr.Type.IsPtr() || !types.Identical(n.Type(), ptr.Type.Elem()) {
   695  		base.FatalfAt(n.Pos(), "setHeapaddr %L with type %v", n, ptr.Type)
   696  	}
   697  
   698  	// Declare variable to hold address.
   699  	sym := &types.Sym{Name: "&" + n.Sym().Name, Pkg: types.LocalPkg}
   700  	addr := s.curfn.NewLocal(pos, sym, types.NewPtr(n.Type()))
   701  	addr.SetUsed(true)
   702  	types.CalcSize(addr.Type())
   703  
   704  	if n.Class == ir.PPARAMOUT {
   705  		addr.SetIsOutputParamHeapAddr(true)
   706  	}
   707  
   708  	n.Heapaddr = addr
   709  	s.assign(addr, ptr, false, 0)
   710  }
   711  
   712  // newObject returns an SSA value denoting new(typ).
   713  func (s *state) newObject(typ *types.Type, rtype *ssa.Value) *ssa.Value {
   714  	if typ.Size() == 0 {
   715  		return s.newValue1A(ssa.OpAddr, types.NewPtr(typ), ir.Syms.Zerobase, s.sb)
   716  	}
   717  	if rtype == nil {
   718  		rtype = s.reflectType(typ)
   719  	}
   720  	return s.rtcall(ir.Syms.Newobject, true, []*types.Type{types.NewPtr(typ)}, rtype)[0]
   721  }
   722  
   723  func (s *state) checkPtrAlignment(n *ir.ConvExpr, v *ssa.Value, count *ssa.Value) {
   724  	if !n.Type().IsPtr() {
   725  		s.Fatalf("expected pointer type: %v", n.Type())
   726  	}
   727  	elem, rtypeExpr := n.Type().Elem(), n.ElemRType
   728  	if count != nil {
   729  		if !elem.IsArray() {
   730  			s.Fatalf("expected array type: %v", elem)
   731  		}
   732  		elem, rtypeExpr = elem.Elem(), n.ElemElemRType
   733  	}
   734  	size := elem.Size()
   735  	// Casting from larger type to smaller one is ok, so for smallest type, do nothing.
   736  	if elem.Alignment() == 1 && (size == 0 || size == 1 || count == nil) {
   737  		return
   738  	}
   739  	if count == nil {
   740  		count = s.constInt(types.Types[types.TUINTPTR], 1)
   741  	}
   742  	if count.Type.Size() != s.config.PtrSize {
   743  		s.Fatalf("expected count fit to a uintptr size, have: %d, want: %d", count.Type.Size(), s.config.PtrSize)
   744  	}
   745  	var rtype *ssa.Value
   746  	if rtypeExpr != nil {
   747  		rtype = s.expr(rtypeExpr)
   748  	} else {
   749  		rtype = s.reflectType(elem)
   750  	}
   751  	s.rtcall(ir.Syms.CheckPtrAlignment, true, nil, v, rtype, count)
   752  }
   753  
   754  // reflectType returns an SSA value representing a pointer to typ's
   755  // reflection type descriptor.
   756  func (s *state) reflectType(typ *types.Type) *ssa.Value {
   757  	// TODO(mdempsky): Make this Fatalf under Unified IR; frontend needs
   758  	// to supply RType expressions.
   759  	lsym := reflectdata.TypeLinksym(typ)
   760  	return s.entryNewValue1A(ssa.OpAddr, types.NewPtr(types.Types[types.TUINT8]), lsym, s.sb)
   761  }
   762  
   763  func dumpSourcesColumn(writer *ssa.HTMLWriter, fn *ir.Func) {
   764  	// Read sources of target function fn.
   765  	fname := base.Ctxt.PosTable.Pos(fn.Pos()).Filename()
   766  	targetFn, err := readFuncLines(fname, fn.Pos().Line(), fn.Endlineno.Line())
   767  	if err != nil {
   768  		writer.Logf("cannot read sources for function %v: %v", fn, err)
   769  	}
   770  
   771  	// Read sources of inlined functions.
   772  	var inlFns []*ssa.FuncLines
   773  	for _, fi := range ssaDumpInlined {
   774  		elno := fi.Endlineno
   775  		fname := base.Ctxt.PosTable.Pos(fi.Pos()).Filename()
   776  		fnLines, err := readFuncLines(fname, fi.Pos().Line(), elno.Line())
   777  		if err != nil {
   778  			writer.Logf("cannot read sources for inlined function %v: %v", fi, err)
   779  			continue
   780  		}
   781  		inlFns = append(inlFns, fnLines)
   782  	}
   783  
   784  	sort.Sort(ssa.ByTopo(inlFns))
   785  	if targetFn != nil {
   786  		inlFns = append([]*ssa.FuncLines{targetFn}, inlFns...)
   787  	}
   788  
   789  	writer.WriteSources("sources", inlFns)
   790  }
   791  
   792  func readFuncLines(file string, start, end uint) (*ssa.FuncLines, error) {
   793  	f, err := os.Open(os.ExpandEnv(file))
   794  	if err != nil {
   795  		return nil, err
   796  	}
   797  	defer f.Close()
   798  	var lines []string
   799  	ln := uint(1)
   800  	scanner := bufio.NewScanner(f)
   801  	for scanner.Scan() && ln <= end {
   802  		if ln >= start {
   803  			lines = append(lines, scanner.Text())
   804  		}
   805  		ln++
   806  	}
   807  	return &ssa.FuncLines{Filename: file, StartLineno: start, Lines: lines}, nil
   808  }
   809  
   810  // updateUnsetPredPos propagates the earliest-value position information for b
   811  // towards all of b's predecessors that need a position, and recurs on that
   812  // predecessor if its position is updated. B should have a non-empty position.
   813  func (s *state) updateUnsetPredPos(b *ssa.Block) {
   814  	if b.Pos == src.NoXPos {
   815  		s.Fatalf("Block %s should have a position", b)
   816  	}
   817  	bestPos := src.NoXPos
   818  	for _, e := range b.Preds {
   819  		p := e.Block()
   820  		if !p.LackingPos() {
   821  			continue
   822  		}
   823  		if bestPos == src.NoXPos {
   824  			bestPos = b.Pos
   825  			for _, v := range b.Values {
   826  				if v.LackingPos() {
   827  					continue
   828  				}
   829  				if v.Pos != src.NoXPos {
   830  					// Assume values are still in roughly textual order;
   831  					// TODO: could also seek minimum position?
   832  					bestPos = v.Pos
   833  					break
   834  				}
   835  			}
   836  		}
   837  		p.Pos = bestPos
   838  		s.updateUnsetPredPos(p) // We do not expect long chains of these, thus recursion is okay.
   839  	}
   840  }
   841  
   842  // Information about each open-coded defer.
   843  type openDeferInfo struct {
   844  	// The node representing the call of the defer
   845  	n *ir.CallExpr
   846  	// If defer call is closure call, the address of the argtmp where the
   847  	// closure is stored.
   848  	closure *ssa.Value
   849  	// The node representing the argtmp where the closure is stored - used for
   850  	// function, method, or interface call, to store a closure that panic
   851  	// processing can use for this defer.
   852  	closureNode *ir.Name
   853  }
   854  
   855  type state struct {
   856  	// configuration (arch) information
   857  	config *ssa.Config
   858  
   859  	// function we're building
   860  	f *ssa.Func
   861  
   862  	// Node for function
   863  	curfn *ir.Func
   864  
   865  	// labels in f
   866  	labels map[string]*ssaLabel
   867  
   868  	// unlabeled break and continue statement tracking
   869  	breakTo    *ssa.Block // current target for plain break statement
   870  	continueTo *ssa.Block // current target for plain continue statement
   871  
   872  	// current location where we're interpreting the AST
   873  	curBlock *ssa.Block
   874  
   875  	// variable assignments in the current block (map from variable symbol to ssa value)
   876  	// *Node is the unique identifier (an ONAME Node) for the variable.
   877  	// TODO: keep a single varnum map, then make all of these maps slices instead?
   878  	vars map[ir.Node]*ssa.Value
   879  
   880  	// fwdVars are variables that are used before they are defined in the current block.
   881  	// This map exists just to coalesce multiple references into a single FwdRef op.
   882  	// *Node is the unique identifier (an ONAME Node) for the variable.
   883  	fwdVars map[ir.Node]*ssa.Value
   884  
   885  	// all defined variables at the end of each block. Indexed by block ID.
   886  	defvars []map[ir.Node]*ssa.Value
   887  
   888  	// addresses of PPARAM and PPARAMOUT variables on the stack.
   889  	decladdrs map[*ir.Name]*ssa.Value
   890  
   891  	// starting values. Memory, stack pointer, and globals pointer
   892  	startmem *ssa.Value
   893  	sp       *ssa.Value
   894  	sb       *ssa.Value
   895  	// value representing address of where deferBits autotmp is stored
   896  	deferBitsAddr *ssa.Value
   897  	deferBitsTemp *ir.Name
   898  
   899  	// line number stack. The current line number is top of stack
   900  	line []src.XPos
   901  	// the last line number processed; it may have been popped
   902  	lastPos src.XPos
   903  
   904  	// list of panic calls by function name and line number.
   905  	// Used to deduplicate panic calls.
   906  	panics map[funcLine]*ssa.Block
   907  
   908  	cgoUnsafeArgs       bool
   909  	hasdefer            bool // whether the function contains a defer statement
   910  	softFloat           bool
   911  	hasOpenDefers       bool // whether we are doing open-coded defers
   912  	checkPtrEnabled     bool // whether to insert checkptr instrumentation
   913  	instrumentEnterExit bool // whether to instrument function enter/exit
   914  	instrumentMemory    bool // whether to instrument memory operations
   915  
   916  	// If doing open-coded defers, list of info about the defer calls in
   917  	// scanning order. Hence, at exit we should run these defers in reverse
   918  	// order of this list
   919  	openDefers []*openDeferInfo
   920  	// For open-coded defers, this is the beginning and end blocks of the last
   921  	// defer exit code that we have generated so far. We use these to share
   922  	// code between exits if the shareDeferExits option (disabled by default)
   923  	// is on.
   924  	lastDeferExit       *ssa.Block // Entry block of last defer exit code we generated
   925  	lastDeferFinalBlock *ssa.Block // Final block of last defer exit code we generated
   926  	lastDeferCount      int        // Number of defers encountered at that point
   927  
   928  	prevCall *ssa.Value // the previous call; use this to tie results to the call op.
   929  }
   930  
   931  type funcLine struct {
   932  	f    *obj.LSym
   933  	base *src.PosBase
   934  	line uint
   935  }
   936  
   937  type ssaLabel struct {
   938  	target         *ssa.Block // block identified by this label
   939  	breakTarget    *ssa.Block // block to break to in control flow node identified by this label
   940  	continueTarget *ssa.Block // block to continue to in control flow node identified by this label
   941  }
   942  
   943  // label returns the label associated with sym, creating it if necessary.
   944  func (s *state) label(sym *types.Sym) *ssaLabel {
   945  	lab := s.labels[sym.Name]
   946  	if lab == nil {
   947  		lab = new(ssaLabel)
   948  		s.labels[sym.Name] = lab
   949  	}
   950  	return lab
   951  }
   952  
   953  func (s *state) Logf(msg string, args ...interface{}) { s.f.Logf(msg, args...) }
   954  func (s *state) Log() bool                            { return s.f.Log() }
   955  func (s *state) Fatalf(msg string, args ...interface{}) {
   956  	s.f.Frontend().Fatalf(s.peekPos(), msg, args...)
   957  }
   958  func (s *state) Warnl(pos src.XPos, msg string, args ...interface{}) { s.f.Warnl(pos, msg, args...) }
   959  func (s *state) Debug_checknil() bool                                { return s.f.Frontend().Debug_checknil() }
   960  
   961  func ssaMarker(name string) *ir.Name {
   962  	return ir.NewNameAt(base.Pos, &types.Sym{Name: name}, nil)
   963  }
   964  
   965  var (
   966  	// marker node for the memory variable
   967  	memVar = ssaMarker("mem")
   968  
   969  	// marker nodes for temporary variables
   970  	ptrVar       = ssaMarker("ptr")
   971  	lenVar       = ssaMarker("len")
   972  	capVar       = ssaMarker("cap")
   973  	typVar       = ssaMarker("typ")
   974  	okVar        = ssaMarker("ok")
   975  	deferBitsVar = ssaMarker("deferBits")
   976  	hashVar      = ssaMarker("hash")
   977  )
   978  
   979  // startBlock sets the current block we're generating code in to b.
   980  func (s *state) startBlock(b *ssa.Block) {
   981  	if s.curBlock != nil {
   982  		s.Fatalf("starting block %v when block %v has not ended", b, s.curBlock)
   983  	}
   984  	s.curBlock = b
   985  	s.vars = map[ir.Node]*ssa.Value{}
   986  	for n := range s.fwdVars {
   987  		delete(s.fwdVars, n)
   988  	}
   989  }
   990  
   991  // endBlock marks the end of generating code for the current block.
   992  // Returns the (former) current block. Returns nil if there is no current
   993  // block, i.e. if no code flows to the current execution point.
   994  func (s *state) endBlock() *ssa.Block {
   995  	b := s.curBlock
   996  	if b == nil {
   997  		return nil
   998  	}
   999  	for len(s.defvars) <= int(b.ID) {
  1000  		s.defvars = append(s.defvars, nil)
  1001  	}
  1002  	s.defvars[b.ID] = s.vars
  1003  	s.curBlock = nil
  1004  	s.vars = nil
  1005  	if b.LackingPos() {
  1006  		// Empty plain blocks get the line of their successor (handled after all blocks created),
  1007  		// except for increment blocks in For statements (handled in ssa conversion of OFOR),
  1008  		// and for blocks ending in GOTO/BREAK/CONTINUE.
  1009  		b.Pos = src.NoXPos
  1010  	} else {
  1011  		b.Pos = s.lastPos
  1012  	}
  1013  	return b
  1014  }
  1015  
  1016  // pushLine pushes a line number on the line number stack.
  1017  func (s *state) pushLine(line src.XPos) {
  1018  	if !line.IsKnown() {
  1019  		// the frontend may emit node with line number missing,
  1020  		// use the parent line number in this case.
  1021  		line = s.peekPos()
  1022  		if base.Flag.K != 0 {
  1023  			base.Warn("buildssa: unknown position (line 0)")
  1024  		}
  1025  	} else {
  1026  		s.lastPos = line
  1027  	}
  1028  
  1029  	s.line = append(s.line, line)
  1030  }
  1031  
  1032  // popLine pops the top of the line number stack.
  1033  func (s *state) popLine() {
  1034  	s.line = s.line[:len(s.line)-1]
  1035  }
  1036  
  1037  // peekPos peeks the top of the line number stack.
  1038  func (s *state) peekPos() src.XPos {
  1039  	return s.line[len(s.line)-1]
  1040  }
  1041  
  1042  // newValue0 adds a new value with no arguments to the current block.
  1043  func (s *state) newValue0(op ssa.Op, t *types.Type) *ssa.Value {
  1044  	return s.curBlock.NewValue0(s.peekPos(), op, t)
  1045  }
  1046  
  1047  // newValue0A adds a new value with no arguments and an aux value to the current block.
  1048  func (s *state) newValue0A(op ssa.Op, t *types.Type, aux ssa.Aux) *ssa.Value {
  1049  	return s.curBlock.NewValue0A(s.peekPos(), op, t, aux)
  1050  }
  1051  
  1052  // newValue0I adds a new value with no arguments and an auxint value to the current block.
  1053  func (s *state) newValue0I(op ssa.Op, t *types.Type, auxint int64) *ssa.Value {
  1054  	return s.curBlock.NewValue0I(s.peekPos(), op, t, auxint)
  1055  }
  1056  
  1057  // newValue1 adds a new value with one argument to the current block.
  1058  func (s *state) newValue1(op ssa.Op, t *types.Type, arg *ssa.Value) *ssa.Value {
  1059  	return s.curBlock.NewValue1(s.peekPos(), op, t, arg)
  1060  }
  1061  
  1062  // newValue1A adds a new value with one argument and an aux value to the current block.
  1063  func (s *state) newValue1A(op ssa.Op, t *types.Type, aux ssa.Aux, arg *ssa.Value) *ssa.Value {
  1064  	return s.curBlock.NewValue1A(s.peekPos(), op, t, aux, arg)
  1065  }
  1066  
  1067  // newValue1Apos adds a new value with one argument and an aux value to the current block.
  1068  // isStmt determines whether the created values may be a statement or not
  1069  // (i.e., false means never, yes means maybe).
  1070  func (s *state) newValue1Apos(op ssa.Op, t *types.Type, aux ssa.Aux, arg *ssa.Value, isStmt bool) *ssa.Value {
  1071  	if isStmt {
  1072  		return s.curBlock.NewValue1A(s.peekPos(), op, t, aux, arg)
  1073  	}
  1074  	return s.curBlock.NewValue1A(s.peekPos().WithNotStmt(), op, t, aux, arg)
  1075  }
  1076  
  1077  // newValue1I adds a new value with one argument and an auxint value to the current block.
  1078  func (s *state) newValue1I(op ssa.Op, t *types.Type, aux int64, arg *ssa.Value) *ssa.Value {
  1079  	return s.curBlock.NewValue1I(s.peekPos(), op, t, aux, arg)
  1080  }
  1081  
  1082  // newValue2 adds a new value with two arguments to the current block.
  1083  func (s *state) newValue2(op ssa.Op, t *types.Type, arg0, arg1 *ssa.Value) *ssa.Value {
  1084  	return s.curBlock.NewValue2(s.peekPos(), op, t, arg0, arg1)
  1085  }
  1086  
  1087  // newValue2A adds a new value with two arguments and an aux value to the current block.
  1088  func (s *state) newValue2A(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1 *ssa.Value) *ssa.Value {
  1089  	return s.curBlock.NewValue2A(s.peekPos(), op, t, aux, arg0, arg1)
  1090  }
  1091  
  1092  // newValue2Apos adds a new value with two arguments and an aux value to the current block.
  1093  // isStmt determines whether the created values may be a statement or not
  1094  // (i.e., false means never, yes means maybe).
  1095  func (s *state) newValue2Apos(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1 *ssa.Value, isStmt bool) *ssa.Value {
  1096  	if isStmt {
  1097  		return s.curBlock.NewValue2A(s.peekPos(), op, t, aux, arg0, arg1)
  1098  	}
  1099  	return s.curBlock.NewValue2A(s.peekPos().WithNotStmt(), op, t, aux, arg0, arg1)
  1100  }
  1101  
  1102  // newValue2I adds a new value with two arguments and an auxint value to the current block.
  1103  func (s *state) newValue2I(op ssa.Op, t *types.Type, aux int64, arg0, arg1 *ssa.Value) *ssa.Value {
  1104  	return s.curBlock.NewValue2I(s.peekPos(), op, t, aux, arg0, arg1)
  1105  }
  1106  
  1107  // newValue3 adds a new value with three arguments to the current block.
  1108  func (s *state) newValue3(op ssa.Op, t *types.Type, arg0, arg1, arg2 *ssa.Value) *ssa.Value {
  1109  	return s.curBlock.NewValue3(s.peekPos(), op, t, arg0, arg1, arg2)
  1110  }
  1111  
  1112  // newValue3I adds a new value with three arguments and an auxint value to the current block.
  1113  func (s *state) newValue3I(op ssa.Op, t *types.Type, aux int64, arg0, arg1, arg2 *ssa.Value) *ssa.Value {
  1114  	return s.curBlock.NewValue3I(s.peekPos(), op, t, aux, arg0, arg1, arg2)
  1115  }
  1116  
  1117  // newValue3A adds a new value with three arguments and an aux value to the current block.
  1118  func (s *state) newValue3A(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1, arg2 *ssa.Value) *ssa.Value {
  1119  	return s.curBlock.NewValue3A(s.peekPos(), op, t, aux, arg0, arg1, arg2)
  1120  }
  1121  
  1122  // newValue3Apos adds a new value with three arguments and an aux value to the current block.
  1123  // isStmt determines whether the created values may be a statement or not
  1124  // (i.e., false means never, yes means maybe).
  1125  func (s *state) newValue3Apos(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1, arg2 *ssa.Value, isStmt bool) *ssa.Value {
  1126  	if isStmt {
  1127  		return s.curBlock.NewValue3A(s.peekPos(), op, t, aux, arg0, arg1, arg2)
  1128  	}
  1129  	return s.curBlock.NewValue3A(s.peekPos().WithNotStmt(), op, t, aux, arg0, arg1, arg2)
  1130  }
  1131  
  1132  // newValue4 adds a new value with four arguments to the current block.
  1133  func (s *state) newValue4(op ssa.Op, t *types.Type, arg0, arg1, arg2, arg3 *ssa.Value) *ssa.Value {
  1134  	return s.curBlock.NewValue4(s.peekPos(), op, t, arg0, arg1, arg2, arg3)
  1135  }
  1136  
  1137  // newValue4I adds a new value with four arguments and an auxint value to the current block.
  1138  func (s *state) newValue4I(op ssa.Op, t *types.Type, aux int64, arg0, arg1, arg2, arg3 *ssa.Value) *ssa.Value {
  1139  	return s.curBlock.NewValue4I(s.peekPos(), op, t, aux, arg0, arg1, arg2, arg3)
  1140  }
  1141  
  1142  func (s *state) entryBlock() *ssa.Block {
  1143  	b := s.f.Entry
  1144  	if base.Flag.N > 0 && s.curBlock != nil {
  1145  		// If optimizations are off, allocate in current block instead. Since with -N
  1146  		// we're not doing the CSE or tighten passes, putting lots of stuff in the
  1147  		// entry block leads to O(n^2) entries in the live value map during regalloc.
  1148  		// See issue 45897.
  1149  		b = s.curBlock
  1150  	}
  1151  	return b
  1152  }
  1153  
  1154  // entryNewValue0 adds a new value with no arguments to the entry block.
  1155  func (s *state) entryNewValue0(op ssa.Op, t *types.Type) *ssa.Value {
  1156  	return s.entryBlock().NewValue0(src.NoXPos, op, t)
  1157  }
  1158  
  1159  // entryNewValue0A adds a new value with no arguments and an aux value to the entry block.
  1160  func (s *state) entryNewValue0A(op ssa.Op, t *types.Type, aux ssa.Aux) *ssa.Value {
  1161  	return s.entryBlock().NewValue0A(src.NoXPos, op, t, aux)
  1162  }
  1163  
  1164  // entryNewValue1 adds a new value with one argument to the entry block.
  1165  func (s *state) entryNewValue1(op ssa.Op, t *types.Type, arg *ssa.Value) *ssa.Value {
  1166  	return s.entryBlock().NewValue1(src.NoXPos, op, t, arg)
  1167  }
  1168  
  1169  // entryNewValue1I adds a new value with one argument and an auxint value to the entry block.
  1170  func (s *state) entryNewValue1I(op ssa.Op, t *types.Type, auxint int64, arg *ssa.Value) *ssa.Value {
  1171  	return s.entryBlock().NewValue1I(src.NoXPos, op, t, auxint, arg)
  1172  }
  1173  
  1174  // entryNewValue1A adds a new value with one argument and an aux value to the entry block.
  1175  func (s *state) entryNewValue1A(op ssa.Op, t *types.Type, aux ssa.Aux, arg *ssa.Value) *ssa.Value {
  1176  	return s.entryBlock().NewValue1A(src.NoXPos, op, t, aux, arg)
  1177  }
  1178  
  1179  // entryNewValue2 adds a new value with two arguments to the entry block.
  1180  func (s *state) entryNewValue2(op ssa.Op, t *types.Type, arg0, arg1 *ssa.Value) *ssa.Value {
  1181  	return s.entryBlock().NewValue2(src.NoXPos, op, t, arg0, arg1)
  1182  }
  1183  
  1184  // entryNewValue2A adds a new value with two arguments and an aux value to the entry block.
  1185  func (s *state) entryNewValue2A(op ssa.Op, t *types.Type, aux ssa.Aux, arg0, arg1 *ssa.Value) *ssa.Value {
  1186  	return s.entryBlock().NewValue2A(src.NoXPos, op, t, aux, arg0, arg1)
  1187  }
  1188  
  1189  // const* routines add a new const value to the entry block.
  1190  func (s *state) constSlice(t *types.Type) *ssa.Value {
  1191  	return s.f.ConstSlice(t)
  1192  }
  1193  func (s *state) constInterface(t *types.Type) *ssa.Value {
  1194  	return s.f.ConstInterface(t)
  1195  }
  1196  func (s *state) constNil(t *types.Type) *ssa.Value { return s.f.ConstNil(t) }
  1197  func (s *state) constEmptyString(t *types.Type) *ssa.Value {
  1198  	return s.f.ConstEmptyString(t)
  1199  }
  1200  func (s *state) constBool(c bool) *ssa.Value {
  1201  	return s.f.ConstBool(types.Types[types.TBOOL], c)
  1202  }
  1203  func (s *state) constInt8(t *types.Type, c int8) *ssa.Value {
  1204  	return s.f.ConstInt8(t, c)
  1205  }
  1206  func (s *state) constInt16(t *types.Type, c int16) *ssa.Value {
  1207  	return s.f.ConstInt16(t, c)
  1208  }
  1209  func (s *state) constInt32(t *types.Type, c int32) *ssa.Value {
  1210  	return s.f.ConstInt32(t, c)
  1211  }
  1212  func (s *state) constInt64(t *types.Type, c int64) *ssa.Value {
  1213  	return s.f.ConstInt64(t, c)
  1214  }
  1215  func (s *state) constFloat32(t *types.Type, c float64) *ssa.Value {
  1216  	return s.f.ConstFloat32(t, c)
  1217  }
  1218  func (s *state) constFloat64(t *types.Type, c float64) *ssa.Value {
  1219  	return s.f.ConstFloat64(t, c)
  1220  }
  1221  func (s *state) constInt(t *types.Type, c int64) *ssa.Value {
  1222  	if s.config.PtrSize == 8 {
  1223  		return s.constInt64(t, c)
  1224  	}
  1225  	if int64(int32(c)) != c {
  1226  		s.Fatalf("integer constant too big %d", c)
  1227  	}
  1228  	return s.constInt32(t, int32(c))
  1229  }
  1230  func (s *state) constOffPtrSP(t *types.Type, c int64) *ssa.Value {
  1231  	return s.f.ConstOffPtrSP(t, c, s.sp)
  1232  }
  1233  
  1234  // newValueOrSfCall* are wrappers around newValue*, which may create a call to a
  1235  // soft-float runtime function instead (when emitting soft-float code).
  1236  func (s *state) newValueOrSfCall1(op ssa.Op, t *types.Type, arg *ssa.Value) *ssa.Value {
  1237  	if s.softFloat {
  1238  		if c, ok := s.sfcall(op, arg); ok {
  1239  			return c
  1240  		}
  1241  	}
  1242  	return s.newValue1(op, t, arg)
  1243  }
  1244  func (s *state) newValueOrSfCall2(op ssa.Op, t *types.Type, arg0, arg1 *ssa.Value) *ssa.Value {
  1245  	if s.softFloat {
  1246  		if c, ok := s.sfcall(op, arg0, arg1); ok {
  1247  			return c
  1248  		}
  1249  	}
  1250  	return s.newValue2(op, t, arg0, arg1)
  1251  }
  1252  
  1253  type instrumentKind uint8
  1254  
  1255  const (
  1256  	instrumentRead = iota
  1257  	instrumentWrite
  1258  	instrumentMove
  1259  )
  1260  
  1261  func (s *state) instrument(t *types.Type, addr *ssa.Value, kind instrumentKind) {
  1262  	s.instrument2(t, addr, nil, kind)
  1263  }
  1264  
  1265  // instrumentFields instruments a read/write operation on addr.
  1266  // If it is instrumenting for MSAN or ASAN and t is a struct type, it instruments
  1267  // operation for each field, instead of for the whole struct.
  1268  func (s *state) instrumentFields(t *types.Type, addr *ssa.Value, kind instrumentKind) {
  1269  	if !(base.Flag.MSan || base.Flag.ASan) || !t.IsStruct() {
  1270  		s.instrument(t, addr, kind)
  1271  		return
  1272  	}
  1273  	for _, f := range t.Fields() {
  1274  		if f.Sym.IsBlank() {
  1275  			continue
  1276  		}
  1277  		offptr := s.newValue1I(ssa.OpOffPtr, types.NewPtr(f.Type), f.Offset, addr)
  1278  		s.instrumentFields(f.Type, offptr, kind)
  1279  	}
  1280  }
  1281  
  1282  func (s *state) instrumentMove(t *types.Type, dst, src *ssa.Value) {
  1283  	if base.Flag.MSan {
  1284  		s.instrument2(t, dst, src, instrumentMove)
  1285  	} else {
  1286  		s.instrument(t, src, instrumentRead)
  1287  		s.instrument(t, dst, instrumentWrite)
  1288  	}
  1289  }
  1290  
  1291  func (s *state) instrument2(t *types.Type, addr, addr2 *ssa.Value, kind instrumentKind) {
  1292  	if !s.instrumentMemory {
  1293  		return
  1294  	}
  1295  
  1296  	w := t.Size()
  1297  	if w == 0 {
  1298  		return // can't race on zero-sized things
  1299  	}
  1300  
  1301  	if ssa.IsSanitizerSafeAddr(addr) {
  1302  		return
  1303  	}
  1304  
  1305  	var fn *obj.LSym
  1306  	needWidth := false
  1307  
  1308  	if addr2 != nil && kind != instrumentMove {
  1309  		panic("instrument2: non-nil addr2 for non-move instrumentation")
  1310  	}
  1311  
  1312  	if base.Flag.MSan {
  1313  		switch kind {
  1314  		case instrumentRead:
  1315  			fn = ir.Syms.Msanread
  1316  		case instrumentWrite:
  1317  			fn = ir.Syms.Msanwrite
  1318  		case instrumentMove:
  1319  			fn = ir.Syms.Msanmove
  1320  		default:
  1321  			panic("unreachable")
  1322  		}
  1323  		needWidth = true
  1324  	} else if base.Flag.Race && t.NumComponents(types.CountBlankFields) > 1 {
  1325  		// for composite objects we have to write every address
  1326  		// because a write might happen to any subobject.
  1327  		// composites with only one element don't have subobjects, though.
  1328  		switch kind {
  1329  		case instrumentRead:
  1330  			fn = ir.Syms.Racereadrange
  1331  		case instrumentWrite:
  1332  			fn = ir.Syms.Racewriterange
  1333  		default:
  1334  			panic("unreachable")
  1335  		}
  1336  		needWidth = true
  1337  	} else if base.Flag.Race {
  1338  		// for non-composite objects we can write just the start
  1339  		// address, as any write must write the first byte.
  1340  		switch kind {
  1341  		case instrumentRead:
  1342  			fn = ir.Syms.Raceread
  1343  		case instrumentWrite:
  1344  			fn = ir.Syms.Racewrite
  1345  		default:
  1346  			panic("unreachable")
  1347  		}
  1348  	} else if base.Flag.ASan {
  1349  		switch kind {
  1350  		case instrumentRead:
  1351  			fn = ir.Syms.Asanread
  1352  		case instrumentWrite:
  1353  			fn = ir.Syms.Asanwrite
  1354  		default:
  1355  			panic("unreachable")
  1356  		}
  1357  		needWidth = true
  1358  	} else {
  1359  		panic("unreachable")
  1360  	}
  1361  
  1362  	args := []*ssa.Value{addr}
  1363  	if addr2 != nil {
  1364  		args = append(args, addr2)
  1365  	}
  1366  	if needWidth {
  1367  		args = append(args, s.constInt(types.Types[types.TUINTPTR], w))
  1368  	}
  1369  	s.rtcall(fn, true, nil, args...)
  1370  }
  1371  
  1372  func (s *state) load(t *types.Type, src *ssa.Value) *ssa.Value {
  1373  	s.instrumentFields(t, src, instrumentRead)
  1374  	return s.rawLoad(t, src)
  1375  }
  1376  
  1377  func (s *state) rawLoad(t *types.Type, src *ssa.Value) *ssa.Value {
  1378  	return s.newValue2(ssa.OpLoad, t, src, s.mem())
  1379  }
  1380  
  1381  func (s *state) store(t *types.Type, dst, val *ssa.Value) {
  1382  	s.vars[memVar] = s.newValue3A(ssa.OpStore, types.TypeMem, t, dst, val, s.mem())
  1383  }
  1384  
  1385  func (s *state) zero(t *types.Type, dst *ssa.Value) {
  1386  	s.instrument(t, dst, instrumentWrite)
  1387  	store := s.newValue2I(ssa.OpZero, types.TypeMem, t.Size(), dst, s.mem())
  1388  	store.Aux = t
  1389  	s.vars[memVar] = store
  1390  }
  1391  
  1392  func (s *state) move(t *types.Type, dst, src *ssa.Value) {
  1393  	s.moveWhichMayOverlap(t, dst, src, false)
  1394  }
  1395  func (s *state) moveWhichMayOverlap(t *types.Type, dst, src *ssa.Value, mayOverlap bool) {
  1396  	s.instrumentMove(t, dst, src)
  1397  	if mayOverlap && t.IsArray() && t.NumElem() > 1 && !ssa.IsInlinableMemmove(dst, src, t.Size(), s.f.Config) {
  1398  		// Normally, when moving Go values of type T from one location to another,
  1399  		// we don't need to worry about partial overlaps. The two Ts must either be
  1400  		// in disjoint (nonoverlapping) memory or in exactly the same location.
  1401  		// There are 2 cases where this isn't true:
  1402  		//  1) Using unsafe you can arrange partial overlaps.
  1403  		//  2) Since Go 1.17, you can use a cast from a slice to a ptr-to-array.
  1404  		//     https://go.dev/ref/spec#Conversions_from_slice_to_array_pointer
  1405  		//     This feature can be used to construct partial overlaps of array types.
  1406  		//       var a [3]int
  1407  		//       p := (*[2]int)(a[:])
  1408  		//       q := (*[2]int)(a[1:])
  1409  		//       *p = *q
  1410  		// We don't care about solving 1. Or at least, we haven't historically
  1411  		// and no one has complained.
  1412  		// For 2, we need to ensure that if there might be partial overlap,
  1413  		// then we can't use OpMove; we must use memmove instead.
  1414  		// (memmove handles partial overlap by copying in the correct
  1415  		// direction. OpMove does not.)
  1416  		//
  1417  		// Note that we have to be careful here not to introduce a call when
  1418  		// we're marshaling arguments to a call or unmarshaling results from a call.
  1419  		// Cases where this is happening must pass mayOverlap to false.
  1420  		// (Currently this only happens when unmarshaling results of a call.)
  1421  		if t.HasPointers() {
  1422  			s.rtcall(ir.Syms.Typedmemmove, true, nil, s.reflectType(t), dst, src)
  1423  			// We would have otherwise implemented this move with straightline code,
  1424  			// including a write barrier. Pretend we issue a write barrier here,
  1425  			// so that the write barrier tests work. (Otherwise they'd need to know
  1426  			// the details of IsInlineableMemmove.)
  1427  			s.curfn.SetWBPos(s.peekPos())
  1428  		} else {
  1429  			s.rtcall(ir.Syms.Memmove, true, nil, dst, src, s.constInt(types.Types[types.TUINTPTR], t.Size()))
  1430  		}
  1431  		ssa.LogLargeCopy(s.f.Name, s.peekPos(), t.Size())
  1432  		return
  1433  	}
  1434  	store := s.newValue3I(ssa.OpMove, types.TypeMem, t.Size(), dst, src, s.mem())
  1435  	store.Aux = t
  1436  	s.vars[memVar] = store
  1437  }
  1438  
  1439  // stmtList converts the statement list n to SSA and adds it to s.
  1440  func (s *state) stmtList(l ir.Nodes) {
  1441  	for _, n := range l {
  1442  		s.stmt(n)
  1443  	}
  1444  }
  1445  
  1446  // stmt converts the statement n to SSA and adds it to s.
  1447  func (s *state) stmt(n ir.Node) {
  1448  	s.pushLine(n.Pos())
  1449  	defer s.popLine()
  1450  
  1451  	// If s.curBlock is nil, and n isn't a label (which might have an associated goto somewhere),
  1452  	// then this code is dead. Stop here.
  1453  	if s.curBlock == nil && n.Op() != ir.OLABEL {
  1454  		return
  1455  	}
  1456  
  1457  	s.stmtList(n.Init())
  1458  	switch n.Op() {
  1459  
  1460  	case ir.OBLOCK:
  1461  		n := n.(*ir.BlockStmt)
  1462  		s.stmtList(n.List)
  1463  
  1464  	case ir.OFALL: // no-op
  1465  
  1466  	// Expression statements
  1467  	case ir.OCALLFUNC:
  1468  		n := n.(*ir.CallExpr)
  1469  		if ir.IsIntrinsicCall(n) {
  1470  			s.intrinsicCall(n)
  1471  			return
  1472  		}
  1473  		fallthrough
  1474  
  1475  	case ir.OCALLINTER:
  1476  		n := n.(*ir.CallExpr)
  1477  		s.callResult(n, callNormal)
  1478  		if n.Op() == ir.OCALLFUNC && n.Fun.Op() == ir.ONAME && n.Fun.(*ir.Name).Class == ir.PFUNC {
  1479  			if fn := n.Fun.Sym().Name; base.Flag.CompilingRuntime && fn == "throw" ||
  1480  				n.Fun.Sym().Pkg == ir.Pkgs.Runtime &&
  1481  					(fn == "throwinit" || fn == "gopanic" || fn == "panicwrap" || fn == "block" ||
  1482  						fn == "panicmakeslicelen" || fn == "panicmakeslicecap" || fn == "panicunsafeslicelen" ||
  1483  						fn == "panicunsafeslicenilptr" || fn == "panicunsafestringlen" || fn == "panicunsafestringnilptr" ||
  1484  						fn == "panicrangestate") {
  1485  				m := s.mem()
  1486  				b := s.endBlock()
  1487  				b.Kind = ssa.BlockExit
  1488  				b.SetControl(m)
  1489  				// TODO: never rewrite OPANIC to OCALLFUNC in the
  1490  				// first place. Need to wait until all backends
  1491  				// go through SSA.
  1492  			}
  1493  		}
  1494  	case ir.ODEFER:
  1495  		n := n.(*ir.GoDeferStmt)
  1496  		if base.Debug.Defer > 0 {
  1497  			var defertype string
  1498  			if s.hasOpenDefers {
  1499  				defertype = "open-coded"
  1500  			} else if n.Esc() == ir.EscNever {
  1501  				defertype = "stack-allocated"
  1502  			} else {
  1503  				defertype = "heap-allocated"
  1504  			}
  1505  			base.WarnfAt(n.Pos(), "%s defer", defertype)
  1506  		}
  1507  		if s.hasOpenDefers {
  1508  			s.openDeferRecord(n.Call.(*ir.CallExpr))
  1509  		} else {
  1510  			d := callDefer
  1511  			if n.Esc() == ir.EscNever && n.DeferAt == nil {
  1512  				d = callDeferStack
  1513  			}
  1514  			s.call(n.Call.(*ir.CallExpr), d, false, n.DeferAt)
  1515  		}
  1516  	case ir.OGO:
  1517  		n := n.(*ir.GoDeferStmt)
  1518  		s.callResult(n.Call.(*ir.CallExpr), callGo)
  1519  
  1520  	case ir.OAS2DOTTYPE:
  1521  		n := n.(*ir.AssignListStmt)
  1522  		var res, resok *ssa.Value
  1523  		if n.Rhs[0].Op() == ir.ODOTTYPE2 {
  1524  			res, resok = s.dottype(n.Rhs[0].(*ir.TypeAssertExpr), true)
  1525  		} else {
  1526  			res, resok = s.dynamicDottype(n.Rhs[0].(*ir.DynamicTypeAssertExpr), true)
  1527  		}
  1528  		deref := false
  1529  		if !ssa.CanSSA(n.Rhs[0].Type()) {
  1530  			if res.Op != ssa.OpLoad {
  1531  				s.Fatalf("dottype of non-load")
  1532  			}
  1533  			mem := s.mem()
  1534  			if res.Args[1] != mem {
  1535  				s.Fatalf("memory no longer live from 2-result dottype load")
  1536  			}
  1537  			deref = true
  1538  			res = res.Args[0]
  1539  		}
  1540  		s.assign(n.Lhs[0], res, deref, 0)
  1541  		s.assign(n.Lhs[1], resok, false, 0)
  1542  		return
  1543  
  1544  	case ir.OAS2FUNC:
  1545  		// We come here only when it is an intrinsic call returning two values.
  1546  		n := n.(*ir.AssignListStmt)
  1547  		call := n.Rhs[0].(*ir.CallExpr)
  1548  		if !ir.IsIntrinsicCall(call) {
  1549  			s.Fatalf("non-intrinsic AS2FUNC not expanded %v", call)
  1550  		}
  1551  		v := s.intrinsicCall(call)
  1552  		v1 := s.newValue1(ssa.OpSelect0, n.Lhs[0].Type(), v)
  1553  		v2 := s.newValue1(ssa.OpSelect1, n.Lhs[1].Type(), v)
  1554  		s.assign(n.Lhs[0], v1, false, 0)
  1555  		s.assign(n.Lhs[1], v2, false, 0)
  1556  		return
  1557  
  1558  	case ir.ODCL:
  1559  		n := n.(*ir.Decl)
  1560  		if v := n.X; v.Esc() == ir.EscHeap {
  1561  			s.newHeapaddr(v)
  1562  		}
  1563  
  1564  	case ir.OLABEL:
  1565  		n := n.(*ir.LabelStmt)
  1566  		sym := n.Label
  1567  		if sym.IsBlank() {
  1568  			// Nothing to do because the label isn't targetable. See issue 52278.
  1569  			break
  1570  		}
  1571  		lab := s.label(sym)
  1572  
  1573  		// The label might already have a target block via a goto.
  1574  		if lab.target == nil {
  1575  			lab.target = s.f.NewBlock(ssa.BlockPlain)
  1576  		}
  1577  
  1578  		// Go to that label.
  1579  		// (We pretend "label:" is preceded by "goto label", unless the predecessor is unreachable.)
  1580  		if s.curBlock != nil {
  1581  			b := s.endBlock()
  1582  			b.AddEdgeTo(lab.target)
  1583  		}
  1584  		s.startBlock(lab.target)
  1585  
  1586  	case ir.OGOTO:
  1587  		n := n.(*ir.BranchStmt)
  1588  		sym := n.Label
  1589  
  1590  		lab := s.label(sym)
  1591  		if lab.target == nil {
  1592  			lab.target = s.f.NewBlock(ssa.BlockPlain)
  1593  		}
  1594  
  1595  		b := s.endBlock()
  1596  		b.Pos = s.lastPos.WithIsStmt() // Do this even if b is an empty block.
  1597  		b.AddEdgeTo(lab.target)
  1598  
  1599  	case ir.OAS:
  1600  		n := n.(*ir.AssignStmt)
  1601  		if n.X == n.Y && n.X.Op() == ir.ONAME {
  1602  			// An x=x assignment. No point in doing anything
  1603  			// here. In addition, skipping this assignment
  1604  			// prevents generating:
  1605  			//   VARDEF x
  1606  			//   COPY x -> x
  1607  			// which is bad because x is incorrectly considered
  1608  			// dead before the vardef. See issue #14904.
  1609  			return
  1610  		}
  1611  
  1612  		// mayOverlap keeps track of whether the LHS and RHS might
  1613  		// refer to partially overlapping memory. Partial overlapping can
  1614  		// only happen for arrays, see the comment in moveWhichMayOverlap.
  1615  		//
  1616  		// If both sides of the assignment are not dereferences, then partial
  1617  		// overlap can't happen. Partial overlap can only occur only when the
  1618  		// arrays referenced are strictly smaller parts of the same base array.
  1619  		// If one side of the assignment is a full array, then partial overlap
  1620  		// can't happen. (The arrays are either disjoint or identical.)
  1621  		mayOverlap := n.X.Op() == ir.ODEREF && (n.Y != nil && n.Y.Op() == ir.ODEREF)
  1622  		if n.Y != nil && n.Y.Op() == ir.ODEREF {
  1623  			p := n.Y.(*ir.StarExpr).X
  1624  			for p.Op() == ir.OCONVNOP {
  1625  				p = p.(*ir.ConvExpr).X
  1626  			}
  1627  			if p.Op() == ir.OSPTR && p.(*ir.UnaryExpr).X.Type().IsString() {
  1628  				// Pointer fields of strings point to unmodifiable memory.
  1629  				// That memory can't overlap with the memory being written.
  1630  				mayOverlap = false
  1631  			}
  1632  		}
  1633  
  1634  		// Evaluate RHS.
  1635  		rhs := n.Y
  1636  		if rhs != nil {
  1637  			switch rhs.Op() {
  1638  			case ir.OSTRUCTLIT, ir.OARRAYLIT, ir.OSLICELIT:
  1639  				// All literals with nonzero fields have already been
  1640  				// rewritten during walk. Any that remain are just T{}
  1641  				// or equivalents. Use the zero value.
  1642  				if !ir.IsZero(rhs) {
  1643  					s.Fatalf("literal with nonzero value in SSA: %v", rhs)
  1644  				}
  1645  				rhs = nil
  1646  			case ir.OAPPEND:
  1647  				rhs := rhs.(*ir.CallExpr)
  1648  				// Check whether we're writing the result of an append back to the same slice.
  1649  				// If so, we handle it specially to avoid write barriers on the fast
  1650  				// (non-growth) path.
  1651  				if !ir.SameSafeExpr(n.X, rhs.Args[0]) || base.Flag.N != 0 {
  1652  					break
  1653  				}
  1654  				// If the slice can be SSA'd, it'll be on the stack,
  1655  				// so there will be no write barriers,
  1656  				// so there's no need to attempt to prevent them.
  1657  				if s.canSSA(n.X) {
  1658  					if base.Debug.Append > 0 { // replicating old diagnostic message
  1659  						base.WarnfAt(n.Pos(), "append: len-only update (in local slice)")
  1660  					}
  1661  					break
  1662  				}
  1663  				if base.Debug.Append > 0 {
  1664  					base.WarnfAt(n.Pos(), "append: len-only update")
  1665  				}
  1666  				s.append(rhs, true)
  1667  				return
  1668  			}
  1669  		}
  1670  
  1671  		if ir.IsBlank(n.X) {
  1672  			// _ = rhs
  1673  			// Just evaluate rhs for side-effects.
  1674  			if rhs != nil {
  1675  				s.expr(rhs)
  1676  			}
  1677  			return
  1678  		}
  1679  
  1680  		var t *types.Type
  1681  		if n.Y != nil {
  1682  			t = n.Y.Type()
  1683  		} else {
  1684  			t = n.X.Type()
  1685  		}
  1686  
  1687  		var r *ssa.Value
  1688  		deref := !ssa.CanSSA(t)
  1689  		if deref {
  1690  			if rhs == nil {
  1691  				r = nil // Signal assign to use OpZero.
  1692  			} else {
  1693  				r = s.addr(rhs)
  1694  			}
  1695  		} else {
  1696  			if rhs == nil {
  1697  				r = s.zeroVal(t)
  1698  			} else {
  1699  				r = s.expr(rhs)
  1700  			}
  1701  		}
  1702  
  1703  		var skip skipMask
  1704  		if rhs != nil && (rhs.Op() == ir.OSLICE || rhs.Op() == ir.OSLICE3 || rhs.Op() == ir.OSLICESTR) && ir.SameSafeExpr(rhs.(*ir.SliceExpr).X, n.X) {
  1705  			// We're assigning a slicing operation back to its source.
  1706  			// Don't write back fields we aren't changing. See issue #14855.
  1707  			rhs := rhs.(*ir.SliceExpr)
  1708  			i, j, k := rhs.Low, rhs.High, rhs.Max
  1709  			if i != nil && (i.Op() == ir.OLITERAL && i.Val().Kind() == constant.Int && ir.Int64Val(i) == 0) {
  1710  				// [0:...] is the same as [:...]
  1711  				i = nil
  1712  			}
  1713  			// TODO: detect defaults for len/cap also.
  1714  			// Currently doesn't really work because (*p)[:len(*p)] appears here as:
  1715  			//    tmp = len(*p)
  1716  			//    (*p)[:tmp]
  1717  			// if j != nil && (j.Op == OLEN && SameSafeExpr(j.Left, n.Left)) {
  1718  			//      j = nil
  1719  			// }
  1720  			// if k != nil && (k.Op == OCAP && SameSafeExpr(k.Left, n.Left)) {
  1721  			//      k = nil
  1722  			// }
  1723  			if i == nil {
  1724  				skip |= skipPtr
  1725  				if j == nil {
  1726  					skip |= skipLen
  1727  				}
  1728  				if k == nil {
  1729  					skip |= skipCap
  1730  				}
  1731  			}
  1732  		}
  1733  
  1734  		s.assignWhichMayOverlap(n.X, r, deref, skip, mayOverlap)
  1735  
  1736  	case ir.OIF:
  1737  		n := n.(*ir.IfStmt)
  1738  		if ir.IsConst(n.Cond, constant.Bool) {
  1739  			s.stmtList(n.Cond.Init())
  1740  			if ir.BoolVal(n.Cond) {
  1741  				s.stmtList(n.Body)
  1742  			} else {
  1743  				s.stmtList(n.Else)
  1744  			}
  1745  			break
  1746  		}
  1747  
  1748  		bEnd := s.f.NewBlock(ssa.BlockPlain)
  1749  		var likely int8
  1750  		if n.Likely {
  1751  			likely = 1
  1752  		}
  1753  		var bThen *ssa.Block
  1754  		if len(n.Body) != 0 {
  1755  			bThen = s.f.NewBlock(ssa.BlockPlain)
  1756  		} else {
  1757  			bThen = bEnd
  1758  		}
  1759  		var bElse *ssa.Block
  1760  		if len(n.Else) != 0 {
  1761  			bElse = s.f.NewBlock(ssa.BlockPlain)
  1762  		} else {
  1763  			bElse = bEnd
  1764  		}
  1765  		s.condBranch(n.Cond, bThen, bElse, likely)
  1766  
  1767  		if len(n.Body) != 0 {
  1768  			s.startBlock(bThen)
  1769  			s.stmtList(n.Body)
  1770  			if b := s.endBlock(); b != nil {
  1771  				b.AddEdgeTo(bEnd)
  1772  			}
  1773  		}
  1774  		if len(n.Else) != 0 {
  1775  			s.startBlock(bElse)
  1776  			s.stmtList(n.Else)
  1777  			if b := s.endBlock(); b != nil {
  1778  				b.AddEdgeTo(bEnd)
  1779  			}
  1780  		}
  1781  		s.startBlock(bEnd)
  1782  
  1783  	case ir.ORETURN:
  1784  		n := n.(*ir.ReturnStmt)
  1785  		s.stmtList(n.Results)
  1786  		b := s.exit()
  1787  		b.Pos = s.lastPos.WithIsStmt()
  1788  
  1789  	case ir.OTAILCALL:
  1790  		n := n.(*ir.TailCallStmt)
  1791  		s.callResult(n.Call, callTail)
  1792  		call := s.mem()
  1793  		b := s.endBlock()
  1794  		b.Kind = ssa.BlockRetJmp // could use BlockExit. BlockRetJmp is mostly for clarity.
  1795  		b.SetControl(call)
  1796  
  1797  	case ir.OCONTINUE, ir.OBREAK:
  1798  		n := n.(*ir.BranchStmt)
  1799  		var to *ssa.Block
  1800  		if n.Label == nil {
  1801  			// plain break/continue
  1802  			switch n.Op() {
  1803  			case ir.OCONTINUE:
  1804  				to = s.continueTo
  1805  			case ir.OBREAK:
  1806  				to = s.breakTo
  1807  			}
  1808  		} else {
  1809  			// labeled break/continue; look up the target
  1810  			sym := n.Label
  1811  			lab := s.label(sym)
  1812  			switch n.Op() {
  1813  			case ir.OCONTINUE:
  1814  				to = lab.continueTarget
  1815  			case ir.OBREAK:
  1816  				to = lab.breakTarget
  1817  			}
  1818  		}
  1819  
  1820  		b := s.endBlock()
  1821  		b.Pos = s.lastPos.WithIsStmt() // Do this even if b is an empty block.
  1822  		b.AddEdgeTo(to)
  1823  
  1824  	case ir.OFOR:
  1825  		// OFOR: for Ninit; Left; Right { Nbody }
  1826  		// cond (Left); body (Nbody); incr (Right)
  1827  		n := n.(*ir.ForStmt)
  1828  		base.Assert(!n.DistinctVars) // Should all be rewritten before escape analysis
  1829  		bCond := s.f.NewBlock(ssa.BlockPlain)
  1830  		bBody := s.f.NewBlock(ssa.BlockPlain)
  1831  		bIncr := s.f.NewBlock(ssa.BlockPlain)
  1832  		bEnd := s.f.NewBlock(ssa.BlockPlain)
  1833  
  1834  		// ensure empty for loops have correct position; issue #30167
  1835  		bBody.Pos = n.Pos()
  1836  
  1837  		// first, jump to condition test
  1838  		b := s.endBlock()
  1839  		b.AddEdgeTo(bCond)
  1840  
  1841  		// generate code to test condition
  1842  		s.startBlock(bCond)
  1843  		if n.Cond != nil {
  1844  			s.condBranch(n.Cond, bBody, bEnd, 1)
  1845  		} else {
  1846  			b := s.endBlock()
  1847  			b.Kind = ssa.BlockPlain
  1848  			b.AddEdgeTo(bBody)
  1849  		}
  1850  
  1851  		// set up for continue/break in body
  1852  		prevContinue := s.continueTo
  1853  		prevBreak := s.breakTo
  1854  		s.continueTo = bIncr
  1855  		s.breakTo = bEnd
  1856  		var lab *ssaLabel
  1857  		if sym := n.Label; sym != nil {
  1858  			// labeled for loop
  1859  			lab = s.label(sym)
  1860  			lab.continueTarget = bIncr
  1861  			lab.breakTarget = bEnd
  1862  		}
  1863  
  1864  		// generate body
  1865  		s.startBlock(bBody)
  1866  		s.stmtList(n.Body)
  1867  
  1868  		// tear down continue/break
  1869  		s.continueTo = prevContinue
  1870  		s.breakTo = prevBreak
  1871  		if lab != nil {
  1872  			lab.continueTarget = nil
  1873  			lab.breakTarget = nil
  1874  		}
  1875  
  1876  		// done with body, goto incr
  1877  		if b := s.endBlock(); b != nil {
  1878  			b.AddEdgeTo(bIncr)
  1879  		}
  1880  
  1881  		// generate incr
  1882  		s.startBlock(bIncr)
  1883  		if n.Post != nil {
  1884  			s.stmt(n.Post)
  1885  		}
  1886  		if b := s.endBlock(); b != nil {
  1887  			b.AddEdgeTo(bCond)
  1888  			// It can happen that bIncr ends in a block containing only VARKILL,
  1889  			// and that muddles the debugging experience.
  1890  			if b.Pos == src.NoXPos {
  1891  				b.Pos = bCond.Pos
  1892  			}
  1893  		}
  1894  
  1895  		s.startBlock(bEnd)
  1896  
  1897  	case ir.OSWITCH, ir.OSELECT:
  1898  		// These have been mostly rewritten by the front end into their Nbody fields.
  1899  		// Our main task is to correctly hook up any break statements.
  1900  		bEnd := s.f.NewBlock(ssa.BlockPlain)
  1901  
  1902  		prevBreak := s.breakTo
  1903  		s.breakTo = bEnd
  1904  		var sym *types.Sym
  1905  		var body ir.Nodes
  1906  		if n.Op() == ir.OSWITCH {
  1907  			n := n.(*ir.SwitchStmt)
  1908  			sym = n.Label
  1909  			body = n.Compiled
  1910  		} else {
  1911  			n := n.(*ir.SelectStmt)
  1912  			sym = n.Label
  1913  			body = n.Compiled
  1914  		}
  1915  
  1916  		var lab *ssaLabel
  1917  		if sym != nil {
  1918  			// labeled
  1919  			lab = s.label(sym)
  1920  			lab.breakTarget = bEnd
  1921  		}
  1922  
  1923  		// generate body code
  1924  		s.stmtList(body)
  1925  
  1926  		s.breakTo = prevBreak
  1927  		if lab != nil {
  1928  			lab.breakTarget = nil
  1929  		}
  1930  
  1931  		// walk adds explicit OBREAK nodes to the end of all reachable code paths.
  1932  		// If we still have a current block here, then mark it unreachable.
  1933  		if s.curBlock != nil {
  1934  			m := s.mem()
  1935  			b := s.endBlock()
  1936  			b.Kind = ssa.BlockExit
  1937  			b.SetControl(m)
  1938  		}
  1939  		s.startBlock(bEnd)
  1940  
  1941  	case ir.OJUMPTABLE:
  1942  		n := n.(*ir.JumpTableStmt)
  1943  
  1944  		// Make blocks we'll need.
  1945  		jt := s.f.NewBlock(ssa.BlockJumpTable)
  1946  		bEnd := s.f.NewBlock(ssa.BlockPlain)
  1947  
  1948  		// The only thing that needs evaluating is the index we're looking up.
  1949  		idx := s.expr(n.Idx)
  1950  		unsigned := idx.Type.IsUnsigned()
  1951  
  1952  		// Extend so we can do everything in uintptr arithmetic.
  1953  		t := types.Types[types.TUINTPTR]
  1954  		idx = s.conv(nil, idx, idx.Type, t)
  1955  
  1956  		// The ending condition for the current block decides whether we'll use
  1957  		// the jump table at all.
  1958  		// We check that min <= idx <= max and jump around the jump table
  1959  		// if that test fails.
  1960  		// We implement min <= idx <= max with 0 <= idx-min <= max-min, because
  1961  		// we'll need idx-min anyway as the control value for the jump table.
  1962  		var min, max uint64
  1963  		if unsigned {
  1964  			min, _ = constant.Uint64Val(n.Cases[0])
  1965  			max, _ = constant.Uint64Val(n.Cases[len(n.Cases)-1])
  1966  		} else {
  1967  			mn, _ := constant.Int64Val(n.Cases[0])
  1968  			mx, _ := constant.Int64Val(n.Cases[len(n.Cases)-1])
  1969  			min = uint64(mn)
  1970  			max = uint64(mx)
  1971  		}
  1972  		// Compare idx-min with max-min, to see if we can use the jump table.
  1973  		idx = s.newValue2(s.ssaOp(ir.OSUB, t), t, idx, s.uintptrConstant(min))
  1974  		width := s.uintptrConstant(max - min)
  1975  		cmp := s.newValue2(s.ssaOp(ir.OLE, t), types.Types[types.TBOOL], idx, width)
  1976  		b := s.endBlock()
  1977  		b.Kind = ssa.BlockIf
  1978  		b.SetControl(cmp)
  1979  		b.AddEdgeTo(jt)             // in range - use jump table
  1980  		b.AddEdgeTo(bEnd)           // out of range - no case in the jump table will trigger
  1981  		b.Likely = ssa.BranchLikely // TODO: assumes missing the table entirely is unlikely. True?
  1982  
  1983  		// Build jump table block.
  1984  		s.startBlock(jt)
  1985  		jt.Pos = n.Pos()
  1986  		if base.Flag.Cfg.SpectreIndex {
  1987  			idx = s.newValue2(ssa.OpSpectreSliceIndex, t, idx, width)
  1988  		}
  1989  		jt.SetControl(idx)
  1990  
  1991  		// Figure out where we should go for each index in the table.
  1992  		table := make([]*ssa.Block, max-min+1)
  1993  		for i := range table {
  1994  			table[i] = bEnd // default target
  1995  		}
  1996  		for i := range n.Targets {
  1997  			c := n.Cases[i]
  1998  			lab := s.label(n.Targets[i])
  1999  			if lab.target == nil {
  2000  				lab.target = s.f.NewBlock(ssa.BlockPlain)
  2001  			}
  2002  			var val uint64
  2003  			if unsigned {
  2004  				val, _ = constant.Uint64Val(c)
  2005  			} else {
  2006  				vl, _ := constant.Int64Val(c)
  2007  				val = uint64(vl)
  2008  			}
  2009  			// Overwrite the default target.
  2010  			table[val-min] = lab.target
  2011  		}
  2012  		for _, t := range table {
  2013  			jt.AddEdgeTo(t)
  2014  		}
  2015  		s.endBlock()
  2016  
  2017  		s.startBlock(bEnd)
  2018  
  2019  	case ir.OINTERFACESWITCH:
  2020  		n := n.(*ir.InterfaceSwitchStmt)
  2021  		typs := s.f.Config.Types
  2022  
  2023  		t := s.expr(n.RuntimeType)
  2024  		h := s.expr(n.Hash)
  2025  		d := s.newValue1A(ssa.OpAddr, typs.BytePtr, n.Descriptor, s.sb)
  2026  
  2027  		// Check the cache first.
  2028  		var merge *ssa.Block
  2029  		if base.Flag.N == 0 && rtabi.UseInterfaceSwitchCache(Arch.LinkArch.Name) {
  2030  			// Note: we can only use the cache if we have the right atomic load instruction.
  2031  			// Double-check that here.
  2032  			if _, ok := intrinsics[intrinsicKey{Arch.LinkArch.Arch, "internal/runtime/atomic", "Loadp"}]; !ok {
  2033  				s.Fatalf("atomic load not available")
  2034  			}
  2035  			merge = s.f.NewBlock(ssa.BlockPlain)
  2036  			cacheHit := s.f.NewBlock(ssa.BlockPlain)
  2037  			cacheMiss := s.f.NewBlock(ssa.BlockPlain)
  2038  			loopHead := s.f.NewBlock(ssa.BlockPlain)
  2039  			loopBody := s.f.NewBlock(ssa.BlockPlain)
  2040  
  2041  			// Pick right size ops.
  2042  			var mul, and, add, zext ssa.Op
  2043  			if s.config.PtrSize == 4 {
  2044  				mul = ssa.OpMul32
  2045  				and = ssa.OpAnd32
  2046  				add = ssa.OpAdd32
  2047  				zext = ssa.OpCopy
  2048  			} else {
  2049  				mul = ssa.OpMul64
  2050  				and = ssa.OpAnd64
  2051  				add = ssa.OpAdd64
  2052  				zext = ssa.OpZeroExt32to64
  2053  			}
  2054  
  2055  			// Load cache pointer out of descriptor, with an atomic load so
  2056  			// we ensure that we see a fully written cache.
  2057  			atomicLoad := s.newValue2(ssa.OpAtomicLoadPtr, types.NewTuple(typs.BytePtr, types.TypeMem), d, s.mem())
  2058  			cache := s.newValue1(ssa.OpSelect0, typs.BytePtr, atomicLoad)
  2059  			s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, atomicLoad)
  2060  
  2061  			// Initialize hash variable.
  2062  			s.vars[hashVar] = s.newValue1(zext, typs.Uintptr, h)
  2063  
  2064  			// Load mask from cache.
  2065  			mask := s.newValue2(ssa.OpLoad, typs.Uintptr, cache, s.mem())
  2066  			// Jump to loop head.
  2067  			b := s.endBlock()
  2068  			b.AddEdgeTo(loopHead)
  2069  
  2070  			// At loop head, get pointer to the cache entry.
  2071  			//   e := &cache.Entries[hash&mask]
  2072  			s.startBlock(loopHead)
  2073  			entries := s.newValue2(ssa.OpAddPtr, typs.UintptrPtr, cache, s.uintptrConstant(uint64(s.config.PtrSize)))
  2074  			idx := s.newValue2(and, typs.Uintptr, s.variable(hashVar, typs.Uintptr), mask)
  2075  			idx = s.newValue2(mul, typs.Uintptr, idx, s.uintptrConstant(uint64(3*s.config.PtrSize)))
  2076  			e := s.newValue2(ssa.OpAddPtr, typs.UintptrPtr, entries, idx)
  2077  			//   hash++
  2078  			s.vars[hashVar] = s.newValue2(add, typs.Uintptr, s.variable(hashVar, typs.Uintptr), s.uintptrConstant(1))
  2079  
  2080  			// Look for a cache hit.
  2081  			//   if e.Typ == t { goto hit }
  2082  			eTyp := s.newValue2(ssa.OpLoad, typs.Uintptr, e, s.mem())
  2083  			cmp1 := s.newValue2(ssa.OpEqPtr, typs.Bool, t, eTyp)
  2084  			b = s.endBlock()
  2085  			b.Kind = ssa.BlockIf
  2086  			b.SetControl(cmp1)
  2087  			b.AddEdgeTo(cacheHit)
  2088  			b.AddEdgeTo(loopBody)
  2089  
  2090  			// Look for an empty entry, the tombstone for this hash table.
  2091  			//   if e.Typ == nil { goto miss }
  2092  			s.startBlock(loopBody)
  2093  			cmp2 := s.newValue2(ssa.OpEqPtr, typs.Bool, eTyp, s.constNil(typs.BytePtr))
  2094  			b = s.endBlock()
  2095  			b.Kind = ssa.BlockIf
  2096  			b.SetControl(cmp2)
  2097  			b.AddEdgeTo(cacheMiss)
  2098  			b.AddEdgeTo(loopHead)
  2099  
  2100  			// On a hit, load the data fields of the cache entry.
  2101  			//   Case = e.Case
  2102  			//   Itab = e.Itab
  2103  			s.startBlock(cacheHit)
  2104  			eCase := s.newValue2(ssa.OpLoad, typs.Int, s.newValue1I(ssa.OpOffPtr, typs.IntPtr, s.config.PtrSize, e), s.mem())
  2105  			eItab := s.newValue2(ssa.OpLoad, typs.BytePtr, s.newValue1I(ssa.OpOffPtr, typs.BytePtrPtr, 2*s.config.PtrSize, e), s.mem())
  2106  			s.assign(n.Case, eCase, false, 0)
  2107  			s.assign(n.Itab, eItab, false, 0)
  2108  			b = s.endBlock()
  2109  			b.AddEdgeTo(merge)
  2110  
  2111  			// On a miss, call into the runtime to get the answer.
  2112  			s.startBlock(cacheMiss)
  2113  		}
  2114  
  2115  		r := s.rtcall(ir.Syms.InterfaceSwitch, true, []*types.Type{typs.Int, typs.BytePtr}, d, t)
  2116  		s.assign(n.Case, r[0], false, 0)
  2117  		s.assign(n.Itab, r[1], false, 0)
  2118  
  2119  		if merge != nil {
  2120  			// Cache hits merge in here.
  2121  			b := s.endBlock()
  2122  			b.Kind = ssa.BlockPlain
  2123  			b.AddEdgeTo(merge)
  2124  			s.startBlock(merge)
  2125  		}
  2126  
  2127  	case ir.OCHECKNIL:
  2128  		n := n.(*ir.UnaryExpr)
  2129  		p := s.expr(n.X)
  2130  		_ = s.nilCheck(p)
  2131  		// TODO: check that throwing away the nilcheck result is ok.
  2132  
  2133  	case ir.OINLMARK:
  2134  		n := n.(*ir.InlineMarkStmt)
  2135  		s.newValue1I(ssa.OpInlMark, types.TypeVoid, n.Index, s.mem())
  2136  
  2137  	default:
  2138  		s.Fatalf("unhandled stmt %v", n.Op())
  2139  	}
  2140  }
  2141  
  2142  // If true, share as many open-coded defer exits as possible (with the downside of
  2143  // worse line-number information)
  2144  const shareDeferExits = false
  2145  
  2146  // exit processes any code that needs to be generated just before returning.
  2147  // It returns a BlockRet block that ends the control flow. Its control value
  2148  // will be set to the final memory state.
  2149  func (s *state) exit() *ssa.Block {
  2150  	if s.hasdefer {
  2151  		if s.hasOpenDefers {
  2152  			if shareDeferExits && s.lastDeferExit != nil && len(s.openDefers) == s.lastDeferCount {
  2153  				if s.curBlock.Kind != ssa.BlockPlain {
  2154  					panic("Block for an exit should be BlockPlain")
  2155  				}
  2156  				s.curBlock.AddEdgeTo(s.lastDeferExit)
  2157  				s.endBlock()
  2158  				return s.lastDeferFinalBlock
  2159  			}
  2160  			s.openDeferExit()
  2161  		} else {
  2162  			s.rtcall(ir.Syms.Deferreturn, true, nil)
  2163  		}
  2164  	}
  2165  
  2166  	// Do actual return.
  2167  	// These currently turn into self-copies (in many cases).
  2168  	resultFields := s.curfn.Type().Results()
  2169  	results := make([]*ssa.Value, len(resultFields)+1, len(resultFields)+1)
  2170  	// Store SSAable and heap-escaped PPARAMOUT variables back to stack locations.
  2171  	for i, f := range resultFields {
  2172  		n := f.Nname.(*ir.Name)
  2173  		if s.canSSA(n) { // result is in some SSA variable
  2174  			if !n.IsOutputParamInRegisters() && n.Type().HasPointers() {
  2175  				// We are about to store to the result slot.
  2176  				s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, n, s.mem())
  2177  			}
  2178  			results[i] = s.variable(n, n.Type())
  2179  		} else if !n.OnStack() { // result is actually heap allocated
  2180  			// We are about to copy the in-heap result to the result slot.
  2181  			if n.Type().HasPointers() {
  2182  				s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, n, s.mem())
  2183  			}
  2184  			ha := s.expr(n.Heapaddr)
  2185  			s.instrumentFields(n.Type(), ha, instrumentRead)
  2186  			results[i] = s.newValue2(ssa.OpDereference, n.Type(), ha, s.mem())
  2187  		} else { // result is not SSA-able; not escaped, so not on heap, but too large for SSA.
  2188  			// Before register ABI this ought to be a self-move, home=dest,
  2189  			// With register ABI, it's still a self-move if parameter is on stack (i.e., too big or overflowed)
  2190  			// No VarDef, as the result slot is already holding live value.
  2191  			results[i] = s.newValue2(ssa.OpDereference, n.Type(), s.addr(n), s.mem())
  2192  		}
  2193  	}
  2194  
  2195  	// In -race mode, we need to call racefuncexit.
  2196  	// Note: This has to happen after we load any heap-allocated results,
  2197  	// otherwise races will be attributed to the caller instead.
  2198  	if s.instrumentEnterExit {
  2199  		s.rtcall(ir.Syms.Racefuncexit, true, nil)
  2200  	}
  2201  
  2202  	results[len(results)-1] = s.mem()
  2203  	m := s.newValue0(ssa.OpMakeResult, s.f.OwnAux.LateExpansionResultType())
  2204  	m.AddArgs(results...)
  2205  
  2206  	b := s.endBlock()
  2207  	b.Kind = ssa.BlockRet
  2208  	b.SetControl(m)
  2209  	if s.hasdefer && s.hasOpenDefers {
  2210  		s.lastDeferFinalBlock = b
  2211  	}
  2212  	return b
  2213  }
  2214  
  2215  type opAndType struct {
  2216  	op    ir.Op
  2217  	etype types.Kind
  2218  }
  2219  
  2220  var opToSSA = map[opAndType]ssa.Op{
  2221  	{ir.OADD, types.TINT8}:    ssa.OpAdd8,
  2222  	{ir.OADD, types.TUINT8}:   ssa.OpAdd8,
  2223  	{ir.OADD, types.TINT16}:   ssa.OpAdd16,
  2224  	{ir.OADD, types.TUINT16}:  ssa.OpAdd16,
  2225  	{ir.OADD, types.TINT32}:   ssa.OpAdd32,
  2226  	{ir.OADD, types.TUINT32}:  ssa.OpAdd32,
  2227  	{ir.OADD, types.TINT64}:   ssa.OpAdd64,
  2228  	{ir.OADD, types.TUINT64}:  ssa.OpAdd64,
  2229  	{ir.OADD, types.TFLOAT32}: ssa.OpAdd32F,
  2230  	{ir.OADD, types.TFLOAT64}: ssa.OpAdd64F,
  2231  
  2232  	{ir.OSUB, types.TINT8}:    ssa.OpSub8,
  2233  	{ir.OSUB, types.TUINT8}:   ssa.OpSub8,
  2234  	{ir.OSUB, types.TINT16}:   ssa.OpSub16,
  2235  	{ir.OSUB, types.TUINT16}:  ssa.OpSub16,
  2236  	{ir.OSUB, types.TINT32}:   ssa.OpSub32,
  2237  	{ir.OSUB, types.TUINT32}:  ssa.OpSub32,
  2238  	{ir.OSUB, types.TINT64}:   ssa.OpSub64,
  2239  	{ir.OSUB, types.TUINT64}:  ssa.OpSub64,
  2240  	{ir.OSUB, types.TFLOAT32}: ssa.OpSub32F,
  2241  	{ir.OSUB, types.TFLOAT64}: ssa.OpSub64F,
  2242  
  2243  	{ir.ONOT, types.TBOOL}: ssa.OpNot,
  2244  
  2245  	{ir.ONEG, types.TINT8}:    ssa.OpNeg8,
  2246  	{ir.ONEG, types.TUINT8}:   ssa.OpNeg8,
  2247  	{ir.ONEG, types.TINT16}:   ssa.OpNeg16,
  2248  	{ir.ONEG, types.TUINT16}:  ssa.OpNeg16,
  2249  	{ir.ONEG, types.TINT32}:   ssa.OpNeg32,
  2250  	{ir.ONEG, types.TUINT32}:  ssa.OpNeg32,
  2251  	{ir.ONEG, types.TINT64}:   ssa.OpNeg64,
  2252  	{ir.ONEG, types.TUINT64}:  ssa.OpNeg64,
  2253  	{ir.ONEG, types.TFLOAT32}: ssa.OpNeg32F,
  2254  	{ir.ONEG, types.TFLOAT64}: ssa.OpNeg64F,
  2255  
  2256  	{ir.OBITNOT, types.TINT8}:   ssa.OpCom8,
  2257  	{ir.OBITNOT, types.TUINT8}:  ssa.OpCom8,
  2258  	{ir.OBITNOT, types.TINT16}:  ssa.OpCom16,
  2259  	{ir.OBITNOT, types.TUINT16}: ssa.OpCom16,
  2260  	{ir.OBITNOT, types.TINT32}:  ssa.OpCom32,
  2261  	{ir.OBITNOT, types.TUINT32}: ssa.OpCom32,
  2262  	{ir.OBITNOT, types.TINT64}:  ssa.OpCom64,
  2263  	{ir.OBITNOT, types.TUINT64}: ssa.OpCom64,
  2264  
  2265  	{ir.OIMAG, types.TCOMPLEX64}:  ssa.OpComplexImag,
  2266  	{ir.OIMAG, types.TCOMPLEX128}: ssa.OpComplexImag,
  2267  	{ir.OREAL, types.TCOMPLEX64}:  ssa.OpComplexReal,
  2268  	{ir.OREAL, types.TCOMPLEX128}: ssa.OpComplexReal,
  2269  
  2270  	{ir.OMUL, types.TINT8}:    ssa.OpMul8,
  2271  	{ir.OMUL, types.TUINT8}:   ssa.OpMul8,
  2272  	{ir.OMUL, types.TINT16}:   ssa.OpMul16,
  2273  	{ir.OMUL, types.TUINT16}:  ssa.OpMul16,
  2274  	{ir.OMUL, types.TINT32}:   ssa.OpMul32,
  2275  	{ir.OMUL, types.TUINT32}:  ssa.OpMul32,
  2276  	{ir.OMUL, types.TINT64}:   ssa.OpMul64,
  2277  	{ir.OMUL, types.TUINT64}:  ssa.OpMul64,
  2278  	{ir.OMUL, types.TFLOAT32}: ssa.OpMul32F,
  2279  	{ir.OMUL, types.TFLOAT64}: ssa.OpMul64F,
  2280  
  2281  	{ir.ODIV, types.TFLOAT32}: ssa.OpDiv32F,
  2282  	{ir.ODIV, types.TFLOAT64}: ssa.OpDiv64F,
  2283  
  2284  	{ir.ODIV, types.TINT8}:   ssa.OpDiv8,
  2285  	{ir.ODIV, types.TUINT8}:  ssa.OpDiv8u,
  2286  	{ir.ODIV, types.TINT16}:  ssa.OpDiv16,
  2287  	{ir.ODIV, types.TUINT16}: ssa.OpDiv16u,
  2288  	{ir.ODIV, types.TINT32}:  ssa.OpDiv32,
  2289  	{ir.ODIV, types.TUINT32}: ssa.OpDiv32u,
  2290  	{ir.ODIV, types.TINT64}:  ssa.OpDiv64,
  2291  	{ir.ODIV, types.TUINT64}: ssa.OpDiv64u,
  2292  
  2293  	{ir.OMOD, types.TINT8}:   ssa.OpMod8,
  2294  	{ir.OMOD, types.TUINT8}:  ssa.OpMod8u,
  2295  	{ir.OMOD, types.TINT16}:  ssa.OpMod16,
  2296  	{ir.OMOD, types.TUINT16}: ssa.OpMod16u,
  2297  	{ir.OMOD, types.TINT32}:  ssa.OpMod32,
  2298  	{ir.OMOD, types.TUINT32}: ssa.OpMod32u,
  2299  	{ir.OMOD, types.TINT64}:  ssa.OpMod64,
  2300  	{ir.OMOD, types.TUINT64}: ssa.OpMod64u,
  2301  
  2302  	{ir.OAND, types.TINT8}:   ssa.OpAnd8,
  2303  	{ir.OAND, types.TUINT8}:  ssa.OpAnd8,
  2304  	{ir.OAND, types.TINT16}:  ssa.OpAnd16,
  2305  	{ir.OAND, types.TUINT16}: ssa.OpAnd16,
  2306  	{ir.OAND, types.TINT32}:  ssa.OpAnd32,
  2307  	{ir.OAND, types.TUINT32}: ssa.OpAnd32,
  2308  	{ir.OAND, types.TINT64}:  ssa.OpAnd64,
  2309  	{ir.OAND, types.TUINT64}: ssa.OpAnd64,
  2310  
  2311  	{ir.OOR, types.TINT8}:   ssa.OpOr8,
  2312  	{ir.OOR, types.TUINT8}:  ssa.OpOr8,
  2313  	{ir.OOR, types.TINT16}:  ssa.OpOr16,
  2314  	{ir.OOR, types.TUINT16}: ssa.OpOr16,
  2315  	{ir.OOR, types.TINT32}:  ssa.OpOr32,
  2316  	{ir.OOR, types.TUINT32}: ssa.OpOr32,
  2317  	{ir.OOR, types.TINT64}:  ssa.OpOr64,
  2318  	{ir.OOR, types.TUINT64}: ssa.OpOr64,
  2319  
  2320  	{ir.OXOR, types.TINT8}:   ssa.OpXor8,
  2321  	{ir.OXOR, types.TUINT8}:  ssa.OpXor8,
  2322  	{ir.OXOR, types.TINT16}:  ssa.OpXor16,
  2323  	{ir.OXOR, types.TUINT16}: ssa.OpXor16,
  2324  	{ir.OXOR, types.TINT32}:  ssa.OpXor32,
  2325  	{ir.OXOR, types.TUINT32}: ssa.OpXor32,
  2326  	{ir.OXOR, types.TINT64}:  ssa.OpXor64,
  2327  	{ir.OXOR, types.TUINT64}: ssa.OpXor64,
  2328  
  2329  	{ir.OEQ, types.TBOOL}:      ssa.OpEqB,
  2330  	{ir.OEQ, types.TINT8}:      ssa.OpEq8,
  2331  	{ir.OEQ, types.TUINT8}:     ssa.OpEq8,
  2332  	{ir.OEQ, types.TINT16}:     ssa.OpEq16,
  2333  	{ir.OEQ, types.TUINT16}:    ssa.OpEq16,
  2334  	{ir.OEQ, types.TINT32}:     ssa.OpEq32,
  2335  	{ir.OEQ, types.TUINT32}:    ssa.OpEq32,
  2336  	{ir.OEQ, types.TINT64}:     ssa.OpEq64,
  2337  	{ir.OEQ, types.TUINT64}:    ssa.OpEq64,
  2338  	{ir.OEQ, types.TINTER}:     ssa.OpEqInter,
  2339  	{ir.OEQ, types.TSLICE}:     ssa.OpEqSlice,
  2340  	{ir.OEQ, types.TFUNC}:      ssa.OpEqPtr,
  2341  	{ir.OEQ, types.TMAP}:       ssa.OpEqPtr,
  2342  	{ir.OEQ, types.TCHAN}:      ssa.OpEqPtr,
  2343  	{ir.OEQ, types.TPTR}:       ssa.OpEqPtr,
  2344  	{ir.OEQ, types.TUINTPTR}:   ssa.OpEqPtr,
  2345  	{ir.OEQ, types.TUNSAFEPTR}: ssa.OpEqPtr,
  2346  	{ir.OEQ, types.TFLOAT64}:   ssa.OpEq64F,
  2347  	{ir.OEQ, types.TFLOAT32}:   ssa.OpEq32F,
  2348  
  2349  	{ir.ONE, types.TBOOL}:      ssa.OpNeqB,
  2350  	{ir.ONE, types.TINT8}:      ssa.OpNeq8,
  2351  	{ir.ONE, types.TUINT8}:     ssa.OpNeq8,
  2352  	{ir.ONE, types.TINT16}:     ssa.OpNeq16,
  2353  	{ir.ONE, types.TUINT16}:    ssa.OpNeq16,
  2354  	{ir.ONE, types.TINT32}:     ssa.OpNeq32,
  2355  	{ir.ONE, types.TUINT32}:    ssa.OpNeq32,
  2356  	{ir.ONE, types.TINT64}:     ssa.OpNeq64,
  2357  	{ir.ONE, types.TUINT64}:    ssa.OpNeq64,
  2358  	{ir.ONE, types.TINTER}:     ssa.OpNeqInter,
  2359  	{ir.ONE, types.TSLICE}:     ssa.OpNeqSlice,
  2360  	{ir.ONE, types.TFUNC}:      ssa.OpNeqPtr,
  2361  	{ir.ONE, types.TMAP}:       ssa.OpNeqPtr,
  2362  	{ir.ONE, types.TCHAN}:      ssa.OpNeqPtr,
  2363  	{ir.ONE, types.TPTR}:       ssa.OpNeqPtr,
  2364  	{ir.ONE, types.TUINTPTR}:   ssa.OpNeqPtr,
  2365  	{ir.ONE, types.TUNSAFEPTR}: ssa.OpNeqPtr,
  2366  	{ir.ONE, types.TFLOAT64}:   ssa.OpNeq64F,
  2367  	{ir.ONE, types.TFLOAT32}:   ssa.OpNeq32F,
  2368  
  2369  	{ir.OLT, types.TINT8}:    ssa.OpLess8,
  2370  	{ir.OLT, types.TUINT8}:   ssa.OpLess8U,
  2371  	{ir.OLT, types.TINT16}:   ssa.OpLess16,
  2372  	{ir.OLT, types.TUINT16}:  ssa.OpLess16U,
  2373  	{ir.OLT, types.TINT32}:   ssa.OpLess32,
  2374  	{ir.OLT, types.TUINT32}:  ssa.OpLess32U,
  2375  	{ir.OLT, types.TINT64}:   ssa.OpLess64,
  2376  	{ir.OLT, types.TUINT64}:  ssa.OpLess64U,
  2377  	{ir.OLT, types.TFLOAT64}: ssa.OpLess64F,
  2378  	{ir.OLT, types.TFLOAT32}: ssa.OpLess32F,
  2379  
  2380  	{ir.OLE, types.TINT8}:    ssa.OpLeq8,
  2381  	{ir.OLE, types.TUINT8}:   ssa.OpLeq8U,
  2382  	{ir.OLE, types.TINT16}:   ssa.OpLeq16,
  2383  	{ir.OLE, types.TUINT16}:  ssa.OpLeq16U,
  2384  	{ir.OLE, types.TINT32}:   ssa.OpLeq32,
  2385  	{ir.OLE, types.TUINT32}:  ssa.OpLeq32U,
  2386  	{ir.OLE, types.TINT64}:   ssa.OpLeq64,
  2387  	{ir.OLE, types.TUINT64}:  ssa.OpLeq64U,
  2388  	{ir.OLE, types.TFLOAT64}: ssa.OpLeq64F,
  2389  	{ir.OLE, types.TFLOAT32}: ssa.OpLeq32F,
  2390  }
  2391  
  2392  func (s *state) concreteEtype(t *types.Type) types.Kind {
  2393  	e := t.Kind()
  2394  	switch e {
  2395  	default:
  2396  		return e
  2397  	case types.TINT:
  2398  		if s.config.PtrSize == 8 {
  2399  			return types.TINT64
  2400  		}
  2401  		return types.TINT32
  2402  	case types.TUINT:
  2403  		if s.config.PtrSize == 8 {
  2404  			return types.TUINT64
  2405  		}
  2406  		return types.TUINT32
  2407  	case types.TUINTPTR:
  2408  		if s.config.PtrSize == 8 {
  2409  			return types.TUINT64
  2410  		}
  2411  		return types.TUINT32
  2412  	}
  2413  }
  2414  
  2415  func (s *state) ssaOp(op ir.Op, t *types.Type) ssa.Op {
  2416  	etype := s.concreteEtype(t)
  2417  	x, ok := opToSSA[opAndType{op, etype}]
  2418  	if !ok {
  2419  		s.Fatalf("unhandled binary op %v %s", op, etype)
  2420  	}
  2421  	return x
  2422  }
  2423  
  2424  type opAndTwoTypes struct {
  2425  	op     ir.Op
  2426  	etype1 types.Kind
  2427  	etype2 types.Kind
  2428  }
  2429  
  2430  type twoTypes struct {
  2431  	etype1 types.Kind
  2432  	etype2 types.Kind
  2433  }
  2434  
  2435  type twoOpsAndType struct {
  2436  	op1              ssa.Op
  2437  	op2              ssa.Op
  2438  	intermediateType types.Kind
  2439  }
  2440  
  2441  var fpConvOpToSSA = map[twoTypes]twoOpsAndType{
  2442  
  2443  	{types.TINT8, types.TFLOAT32}:  {ssa.OpSignExt8to32, ssa.OpCvt32to32F, types.TINT32},
  2444  	{types.TINT16, types.TFLOAT32}: {ssa.OpSignExt16to32, ssa.OpCvt32to32F, types.TINT32},
  2445  	{types.TINT32, types.TFLOAT32}: {ssa.OpCopy, ssa.OpCvt32to32F, types.TINT32},
  2446  	{types.TINT64, types.TFLOAT32}: {ssa.OpCopy, ssa.OpCvt64to32F, types.TINT64},
  2447  
  2448  	{types.TINT8, types.TFLOAT64}:  {ssa.OpSignExt8to32, ssa.OpCvt32to64F, types.TINT32},
  2449  	{types.TINT16, types.TFLOAT64}: {ssa.OpSignExt16to32, ssa.OpCvt32to64F, types.TINT32},
  2450  	{types.TINT32, types.TFLOAT64}: {ssa.OpCopy, ssa.OpCvt32to64F, types.TINT32},
  2451  	{types.TINT64, types.TFLOAT64}: {ssa.OpCopy, ssa.OpCvt64to64F, types.TINT64},
  2452  
  2453  	{types.TFLOAT32, types.TINT8}:  {ssa.OpCvt32Fto32, ssa.OpTrunc32to8, types.TINT32},
  2454  	{types.TFLOAT32, types.TINT16}: {ssa.OpCvt32Fto32, ssa.OpTrunc32to16, types.TINT32},
  2455  	{types.TFLOAT32, types.TINT32}: {ssa.OpCvt32Fto32, ssa.OpCopy, types.TINT32},
  2456  	{types.TFLOAT32, types.TINT64}: {ssa.OpCvt32Fto64, ssa.OpCopy, types.TINT64},
  2457  
  2458  	{types.TFLOAT64, types.TINT8}:  {ssa.OpCvt64Fto32, ssa.OpTrunc32to8, types.TINT32},
  2459  	{types.TFLOAT64, types.TINT16}: {ssa.OpCvt64Fto32, ssa.OpTrunc32to16, types.TINT32},
  2460  	{types.TFLOAT64, types.TINT32}: {ssa.OpCvt64Fto32, ssa.OpCopy, types.TINT32},
  2461  	{types.TFLOAT64, types.TINT64}: {ssa.OpCvt64Fto64, ssa.OpCopy, types.TINT64},
  2462  	// unsigned
  2463  	{types.TUINT8, types.TFLOAT32}:  {ssa.OpZeroExt8to32, ssa.OpCvt32to32F, types.TINT32},
  2464  	{types.TUINT16, types.TFLOAT32}: {ssa.OpZeroExt16to32, ssa.OpCvt32to32F, types.TINT32},
  2465  	{types.TUINT32, types.TFLOAT32}: {ssa.OpZeroExt32to64, ssa.OpCvt64to32F, types.TINT64}, // go wide to dodge unsigned
  2466  	{types.TUINT64, types.TFLOAT32}: {ssa.OpCopy, ssa.OpInvalid, types.TUINT64},            // Cvt64Uto32F, branchy code expansion instead
  2467  
  2468  	{types.TUINT8, types.TFLOAT64}:  {ssa.OpZeroExt8to32, ssa.OpCvt32to64F, types.TINT32},
  2469  	{types.TUINT16, types.TFLOAT64}: {ssa.OpZeroExt16to32, ssa.OpCvt32to64F, types.TINT32},
  2470  	{types.TUINT32, types.TFLOAT64}: {ssa.OpZeroExt32to64, ssa.OpCvt64to64F, types.TINT64}, // go wide to dodge unsigned
  2471  	{types.TUINT64, types.TFLOAT64}: {ssa.OpCopy, ssa.OpInvalid, types.TUINT64},            // Cvt64Uto64F, branchy code expansion instead
  2472  
  2473  	{types.TFLOAT32, types.TUINT8}:  {ssa.OpCvt32Fto32, ssa.OpTrunc32to8, types.TINT32},
  2474  	{types.TFLOAT32, types.TUINT16}: {ssa.OpCvt32Fto32, ssa.OpTrunc32to16, types.TINT32},
  2475  	{types.TFLOAT32, types.TUINT32}: {ssa.OpCvt32Fto64, ssa.OpTrunc64to32, types.TINT64}, // go wide to dodge unsigned
  2476  	{types.TFLOAT32, types.TUINT64}: {ssa.OpInvalid, ssa.OpCopy, types.TUINT64},          // Cvt32Fto64U, branchy code expansion instead
  2477  
  2478  	{types.TFLOAT64, types.TUINT8}:  {ssa.OpCvt64Fto32, ssa.OpTrunc32to8, types.TINT32},
  2479  	{types.TFLOAT64, types.TUINT16}: {ssa.OpCvt64Fto32, ssa.OpTrunc32to16, types.TINT32},
  2480  	{types.TFLOAT64, types.TUINT32}: {ssa.OpCvt64Fto64, ssa.OpTrunc64to32, types.TINT64}, // go wide to dodge unsigned
  2481  	{types.TFLOAT64, types.TUINT64}: {ssa.OpInvalid, ssa.OpCopy, types.TUINT64},          // Cvt64Fto64U, branchy code expansion instead
  2482  
  2483  	// float
  2484  	{types.TFLOAT64, types.TFLOAT32}: {ssa.OpCvt64Fto32F, ssa.OpCopy, types.TFLOAT32},
  2485  	{types.TFLOAT64, types.TFLOAT64}: {ssa.OpRound64F, ssa.OpCopy, types.TFLOAT64},
  2486  	{types.TFLOAT32, types.TFLOAT32}: {ssa.OpRound32F, ssa.OpCopy, types.TFLOAT32},
  2487  	{types.TFLOAT32, types.TFLOAT64}: {ssa.OpCvt32Fto64F, ssa.OpCopy, types.TFLOAT64},
  2488  }
  2489  
  2490  // this map is used only for 32-bit arch, and only includes the difference
  2491  // on 32-bit arch, don't use int64<->float conversion for uint32
  2492  var fpConvOpToSSA32 = map[twoTypes]twoOpsAndType{
  2493  	{types.TUINT32, types.TFLOAT32}: {ssa.OpCopy, ssa.OpCvt32Uto32F, types.TUINT32},
  2494  	{types.TUINT32, types.TFLOAT64}: {ssa.OpCopy, ssa.OpCvt32Uto64F, types.TUINT32},
  2495  	{types.TFLOAT32, types.TUINT32}: {ssa.OpCvt32Fto32U, ssa.OpCopy, types.TUINT32},
  2496  	{types.TFLOAT64, types.TUINT32}: {ssa.OpCvt64Fto32U, ssa.OpCopy, types.TUINT32},
  2497  }
  2498  
  2499  // uint64<->float conversions, only on machines that have instructions for that
  2500  var uint64fpConvOpToSSA = map[twoTypes]twoOpsAndType{
  2501  	{types.TUINT64, types.TFLOAT32}: {ssa.OpCopy, ssa.OpCvt64Uto32F, types.TUINT64},
  2502  	{types.TUINT64, types.TFLOAT64}: {ssa.OpCopy, ssa.OpCvt64Uto64F, types.TUINT64},
  2503  	{types.TFLOAT32, types.TUINT64}: {ssa.OpCvt32Fto64U, ssa.OpCopy, types.TUINT64},
  2504  	{types.TFLOAT64, types.TUINT64}: {ssa.OpCvt64Fto64U, ssa.OpCopy, types.TUINT64},
  2505  }
  2506  
  2507  var shiftOpToSSA = map[opAndTwoTypes]ssa.Op{
  2508  	{ir.OLSH, types.TINT8, types.TUINT8}:   ssa.OpLsh8x8,
  2509  	{ir.OLSH, types.TUINT8, types.TUINT8}:  ssa.OpLsh8x8,
  2510  	{ir.OLSH, types.TINT8, types.TUINT16}:  ssa.OpLsh8x16,
  2511  	{ir.OLSH, types.TUINT8, types.TUINT16}: ssa.OpLsh8x16,
  2512  	{ir.OLSH, types.TINT8, types.TUINT32}:  ssa.OpLsh8x32,
  2513  	{ir.OLSH, types.TUINT8, types.TUINT32}: ssa.OpLsh8x32,
  2514  	{ir.OLSH, types.TINT8, types.TUINT64}:  ssa.OpLsh8x64,
  2515  	{ir.OLSH, types.TUINT8, types.TUINT64}: ssa.OpLsh8x64,
  2516  
  2517  	{ir.OLSH, types.TINT16, types.TUINT8}:   ssa.OpLsh16x8,
  2518  	{ir.OLSH, types.TUINT16, types.TUINT8}:  ssa.OpLsh16x8,
  2519  	{ir.OLSH, types.TINT16, types.TUINT16}:  ssa.OpLsh16x16,
  2520  	{ir.OLSH, types.TUINT16, types.TUINT16}: ssa.OpLsh16x16,
  2521  	{ir.OLSH, types.TINT16, types.TUINT32}:  ssa.OpLsh16x32,
  2522  	{ir.OLSH, types.TUINT16, types.TUINT32}: ssa.OpLsh16x32,
  2523  	{ir.OLSH, types.TINT16, types.TUINT64}:  ssa.OpLsh16x64,
  2524  	{ir.OLSH, types.TUINT16, types.TUINT64}: ssa.OpLsh16x64,
  2525  
  2526  	{ir.OLSH, types.TINT32, types.TUINT8}:   ssa.OpLsh32x8,
  2527  	{ir.OLSH, types.TUINT32, types.TUINT8}:  ssa.OpLsh32x8,
  2528  	{ir.OLSH, types.TINT32, types.TUINT16}:  ssa.OpLsh32x16,
  2529  	{ir.OLSH, types.TUINT32, types.TUINT16}: ssa.OpLsh32x16,
  2530  	{ir.OLSH, types.TINT32, types.TUINT32}:  ssa.OpLsh32x32,
  2531  	{ir.OLSH, types.TUINT32, types.TUINT32}: ssa.OpLsh32x32,
  2532  	{ir.OLSH, types.TINT32, types.TUINT64}:  ssa.OpLsh32x64,
  2533  	{ir.OLSH, types.TUINT32, types.TUINT64}: ssa.OpLsh32x64,
  2534  
  2535  	{ir.OLSH, types.TINT64, types.TUINT8}:   ssa.OpLsh64x8,
  2536  	{ir.OLSH, types.TUINT64, types.TUINT8}:  ssa.OpLsh64x8,
  2537  	{ir.OLSH, types.TINT64, types.TUINT16}:  ssa.OpLsh64x16,
  2538  	{ir.OLSH, types.TUINT64, types.TUINT16}: ssa.OpLsh64x16,
  2539  	{ir.OLSH, types.TINT64, types.TUINT32}:  ssa.OpLsh64x32,
  2540  	{ir.OLSH, types.TUINT64, types.TUINT32}: ssa.OpLsh64x32,
  2541  	{ir.OLSH, types.TINT64, types.TUINT64}:  ssa.OpLsh64x64,
  2542  	{ir.OLSH, types.TUINT64, types.TUINT64}: ssa.OpLsh64x64,
  2543  
  2544  	{ir.ORSH, types.TINT8, types.TUINT8}:   ssa.OpRsh8x8,
  2545  	{ir.ORSH, types.TUINT8, types.TUINT8}:  ssa.OpRsh8Ux8,
  2546  	{ir.ORSH, types.TINT8, types.TUINT16}:  ssa.OpRsh8x16,
  2547  	{ir.ORSH, types.TUINT8, types.TUINT16}: ssa.OpRsh8Ux16,
  2548  	{ir.ORSH, types.TINT8, types.TUINT32}:  ssa.OpRsh8x32,
  2549  	{ir.ORSH, types.TUINT8, types.TUINT32}: ssa.OpRsh8Ux32,
  2550  	{ir.ORSH, types.TINT8, types.TUINT64}:  ssa.OpRsh8x64,
  2551  	{ir.ORSH, types.TUINT8, types.TUINT64}: ssa.OpRsh8Ux64,
  2552  
  2553  	{ir.ORSH, types.TINT16, types.TUINT8}:   ssa.OpRsh16x8,
  2554  	{ir.ORSH, types.TUINT16, types.TUINT8}:  ssa.OpRsh16Ux8,
  2555  	{ir.ORSH, types.TINT16, types.TUINT16}:  ssa.OpRsh16x16,
  2556  	{ir.ORSH, types.TUINT16, types.TUINT16}: ssa.OpRsh16Ux16,
  2557  	{ir.ORSH, types.TINT16, types.TUINT32}:  ssa.OpRsh16x32,
  2558  	{ir.ORSH, types.TUINT16, types.TUINT32}: ssa.OpRsh16Ux32,
  2559  	{ir.ORSH, types.TINT16, types.TUINT64}:  ssa.OpRsh16x64,
  2560  	{ir.ORSH, types.TUINT16, types.TUINT64}: ssa.OpRsh16Ux64,
  2561  
  2562  	{ir.ORSH, types.TINT32, types.TUINT8}:   ssa.OpRsh32x8,
  2563  	{ir.ORSH, types.TUINT32, types.TUINT8}:  ssa.OpRsh32Ux8,
  2564  	{ir.ORSH, types.TINT32, types.TUINT16}:  ssa.OpRsh32x16,
  2565  	{ir.ORSH, types.TUINT32, types.TUINT16}: ssa.OpRsh32Ux16,
  2566  	{ir.ORSH, types.TINT32, types.TUINT32}:  ssa.OpRsh32x32,
  2567  	{ir.ORSH, types.TUINT32, types.TUINT32}: ssa.OpRsh32Ux32,
  2568  	{ir.ORSH, types.TINT32, types.TUINT64}:  ssa.OpRsh32x64,
  2569  	{ir.ORSH, types.TUINT32, types.TUINT64}: ssa.OpRsh32Ux64,
  2570  
  2571  	{ir.ORSH, types.TINT64, types.TUINT8}:   ssa.OpRsh64x8,
  2572  	{ir.ORSH, types.TUINT64, types.TUINT8}:  ssa.OpRsh64Ux8,
  2573  	{ir.ORSH, types.TINT64, types.TUINT16}:  ssa.OpRsh64x16,
  2574  	{ir.ORSH, types.TUINT64, types.TUINT16}: ssa.OpRsh64Ux16,
  2575  	{ir.ORSH, types.TINT64, types.TUINT32}:  ssa.OpRsh64x32,
  2576  	{ir.ORSH, types.TUINT64, types.TUINT32}: ssa.OpRsh64Ux32,
  2577  	{ir.ORSH, types.TINT64, types.TUINT64}:  ssa.OpRsh64x64,
  2578  	{ir.ORSH, types.TUINT64, types.TUINT64}: ssa.OpRsh64Ux64,
  2579  }
  2580  
  2581  func (s *state) ssaShiftOp(op ir.Op, t *types.Type, u *types.Type) ssa.Op {
  2582  	etype1 := s.concreteEtype(t)
  2583  	etype2 := s.concreteEtype(u)
  2584  	x, ok := shiftOpToSSA[opAndTwoTypes{op, etype1, etype2}]
  2585  	if !ok {
  2586  		s.Fatalf("unhandled shift op %v etype=%s/%s", op, etype1, etype2)
  2587  	}
  2588  	return x
  2589  }
  2590  
  2591  func (s *state) uintptrConstant(v uint64) *ssa.Value {
  2592  	if s.config.PtrSize == 4 {
  2593  		return s.newValue0I(ssa.OpConst32, types.Types[types.TUINTPTR], int64(v))
  2594  	}
  2595  	return s.newValue0I(ssa.OpConst64, types.Types[types.TUINTPTR], int64(v))
  2596  }
  2597  
  2598  func (s *state) conv(n ir.Node, v *ssa.Value, ft, tt *types.Type) *ssa.Value {
  2599  	if ft.IsBoolean() && tt.IsKind(types.TUINT8) {
  2600  		// Bool -> uint8 is generated internally when indexing into runtime.staticbyte.
  2601  		return s.newValue1(ssa.OpCvtBoolToUint8, tt, v)
  2602  	}
  2603  	if ft.IsInteger() && tt.IsInteger() {
  2604  		var op ssa.Op
  2605  		if tt.Size() == ft.Size() {
  2606  			op = ssa.OpCopy
  2607  		} else if tt.Size() < ft.Size() {
  2608  			// truncation
  2609  			switch 10*ft.Size() + tt.Size() {
  2610  			case 21:
  2611  				op = ssa.OpTrunc16to8
  2612  			case 41:
  2613  				op = ssa.OpTrunc32to8
  2614  			case 42:
  2615  				op = ssa.OpTrunc32to16
  2616  			case 81:
  2617  				op = ssa.OpTrunc64to8
  2618  			case 82:
  2619  				op = ssa.OpTrunc64to16
  2620  			case 84:
  2621  				op = ssa.OpTrunc64to32
  2622  			default:
  2623  				s.Fatalf("weird integer truncation %v -> %v", ft, tt)
  2624  			}
  2625  		} else if ft.IsSigned() {
  2626  			// sign extension
  2627  			switch 10*ft.Size() + tt.Size() {
  2628  			case 12:
  2629  				op = ssa.OpSignExt8to16
  2630  			case 14:
  2631  				op = ssa.OpSignExt8to32
  2632  			case 18:
  2633  				op = ssa.OpSignExt8to64
  2634  			case 24:
  2635  				op = ssa.OpSignExt16to32
  2636  			case 28:
  2637  				op = ssa.OpSignExt16to64
  2638  			case 48:
  2639  				op = ssa.OpSignExt32to64
  2640  			default:
  2641  				s.Fatalf("bad integer sign extension %v -> %v", ft, tt)
  2642  			}
  2643  		} else {
  2644  			// zero extension
  2645  			switch 10*ft.Size() + tt.Size() {
  2646  			case 12:
  2647  				op = ssa.OpZeroExt8to16
  2648  			case 14:
  2649  				op = ssa.OpZeroExt8to32
  2650  			case 18:
  2651  				op = ssa.OpZeroExt8to64
  2652  			case 24:
  2653  				op = ssa.OpZeroExt16to32
  2654  			case 28:
  2655  				op = ssa.OpZeroExt16to64
  2656  			case 48:
  2657  				op = ssa.OpZeroExt32to64
  2658  			default:
  2659  				s.Fatalf("weird integer sign extension %v -> %v", ft, tt)
  2660  			}
  2661  		}
  2662  		return s.newValue1(op, tt, v)
  2663  	}
  2664  
  2665  	if ft.IsComplex() && tt.IsComplex() {
  2666  		var op ssa.Op
  2667  		if ft.Size() == tt.Size() {
  2668  			switch ft.Size() {
  2669  			case 8:
  2670  				op = ssa.OpRound32F
  2671  			case 16:
  2672  				op = ssa.OpRound64F
  2673  			default:
  2674  				s.Fatalf("weird complex conversion %v -> %v", ft, tt)
  2675  			}
  2676  		} else if ft.Size() == 8 && tt.Size() == 16 {
  2677  			op = ssa.OpCvt32Fto64F
  2678  		} else if ft.Size() == 16 && tt.Size() == 8 {
  2679  			op = ssa.OpCvt64Fto32F
  2680  		} else {
  2681  			s.Fatalf("weird complex conversion %v -> %v", ft, tt)
  2682  		}
  2683  		ftp := types.FloatForComplex(ft)
  2684  		ttp := types.FloatForComplex(tt)
  2685  		return s.newValue2(ssa.OpComplexMake, tt,
  2686  			s.newValueOrSfCall1(op, ttp, s.newValue1(ssa.OpComplexReal, ftp, v)),
  2687  			s.newValueOrSfCall1(op, ttp, s.newValue1(ssa.OpComplexImag, ftp, v)))
  2688  	}
  2689  
  2690  	if tt.IsComplex() { // and ft is not complex
  2691  		// Needed for generics support - can't happen in normal Go code.
  2692  		et := types.FloatForComplex(tt)
  2693  		v = s.conv(n, v, ft, et)
  2694  		return s.newValue2(ssa.OpComplexMake, tt, v, s.zeroVal(et))
  2695  	}
  2696  
  2697  	if ft.IsFloat() || tt.IsFloat() {
  2698  		conv, ok := fpConvOpToSSA[twoTypes{s.concreteEtype(ft), s.concreteEtype(tt)}]
  2699  		if s.config.RegSize == 4 && Arch.LinkArch.Family != sys.MIPS && !s.softFloat {
  2700  			if conv1, ok1 := fpConvOpToSSA32[twoTypes{s.concreteEtype(ft), s.concreteEtype(tt)}]; ok1 {
  2701  				conv = conv1
  2702  			}
  2703  		}
  2704  		if Arch.LinkArch.Family == sys.ARM64 || Arch.LinkArch.Family == sys.Wasm || Arch.LinkArch.Family == sys.S390X || s.softFloat {
  2705  			if conv1, ok1 := uint64fpConvOpToSSA[twoTypes{s.concreteEtype(ft), s.concreteEtype(tt)}]; ok1 {
  2706  				conv = conv1
  2707  			}
  2708  		}
  2709  
  2710  		if Arch.LinkArch.Family == sys.MIPS && !s.softFloat {
  2711  			if ft.Size() == 4 && ft.IsInteger() && !ft.IsSigned() {
  2712  				// tt is float32 or float64, and ft is also unsigned
  2713  				if tt.Size() == 4 {
  2714  					return s.uint32Tofloat32(n, v, ft, tt)
  2715  				}
  2716  				if tt.Size() == 8 {
  2717  					return s.uint32Tofloat64(n, v, ft, tt)
  2718  				}
  2719  			} else if tt.Size() == 4 && tt.IsInteger() && !tt.IsSigned() {
  2720  				// ft is float32 or float64, and tt is unsigned integer
  2721  				if ft.Size() == 4 {
  2722  					return s.float32ToUint32(n, v, ft, tt)
  2723  				}
  2724  				if ft.Size() == 8 {
  2725  					return s.float64ToUint32(n, v, ft, tt)
  2726  				}
  2727  			}
  2728  		}
  2729  
  2730  		if !ok {
  2731  			s.Fatalf("weird float conversion %v -> %v", ft, tt)
  2732  		}
  2733  		op1, op2, it := conv.op1, conv.op2, conv.intermediateType
  2734  
  2735  		if op1 != ssa.OpInvalid && op2 != ssa.OpInvalid {
  2736  			// normal case, not tripping over unsigned 64
  2737  			if op1 == ssa.OpCopy {
  2738  				if op2 == ssa.OpCopy {
  2739  					return v
  2740  				}
  2741  				return s.newValueOrSfCall1(op2, tt, v)
  2742  			}
  2743  			if op2 == ssa.OpCopy {
  2744  				return s.newValueOrSfCall1(op1, tt, v)
  2745  			}
  2746  			return s.newValueOrSfCall1(op2, tt, s.newValueOrSfCall1(op1, types.Types[it], v))
  2747  		}
  2748  		// Tricky 64-bit unsigned cases.
  2749  		if ft.IsInteger() {
  2750  			// tt is float32 or float64, and ft is also unsigned
  2751  			if tt.Size() == 4 {
  2752  				return s.uint64Tofloat32(n, v, ft, tt)
  2753  			}
  2754  			if tt.Size() == 8 {
  2755  				return s.uint64Tofloat64(n, v, ft, tt)
  2756  			}
  2757  			s.Fatalf("weird unsigned integer to float conversion %v -> %v", ft, tt)
  2758  		}
  2759  		// ft is float32 or float64, and tt is unsigned integer
  2760  		if ft.Size() == 4 {
  2761  			return s.float32ToUint64(n, v, ft, tt)
  2762  		}
  2763  		if ft.Size() == 8 {
  2764  			return s.float64ToUint64(n, v, ft, tt)
  2765  		}
  2766  		s.Fatalf("weird float to unsigned integer conversion %v -> %v", ft, tt)
  2767  		return nil
  2768  	}
  2769  
  2770  	s.Fatalf("unhandled OCONV %s -> %s", ft.Kind(), tt.Kind())
  2771  	return nil
  2772  }
  2773  
  2774  // expr converts the expression n to ssa, adds it to s and returns the ssa result.
  2775  func (s *state) expr(n ir.Node) *ssa.Value {
  2776  	return s.exprCheckPtr(n, true)
  2777  }
  2778  
  2779  func (s *state) exprCheckPtr(n ir.Node, checkPtrOK bool) *ssa.Value {
  2780  	if ir.HasUniquePos(n) {
  2781  		// ONAMEs and named OLITERALs have the line number
  2782  		// of the decl, not the use. See issue 14742.
  2783  		s.pushLine(n.Pos())
  2784  		defer s.popLine()
  2785  	}
  2786  
  2787  	s.stmtList(n.Init())
  2788  	switch n.Op() {
  2789  	case ir.OBYTES2STRTMP:
  2790  		n := n.(*ir.ConvExpr)
  2791  		slice := s.expr(n.X)
  2792  		ptr := s.newValue1(ssa.OpSlicePtr, s.f.Config.Types.BytePtr, slice)
  2793  		len := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], slice)
  2794  		return s.newValue2(ssa.OpStringMake, n.Type(), ptr, len)
  2795  	case ir.OSTR2BYTESTMP:
  2796  		n := n.(*ir.ConvExpr)
  2797  		str := s.expr(n.X)
  2798  		ptr := s.newValue1(ssa.OpStringPtr, s.f.Config.Types.BytePtr, str)
  2799  		if !n.NonNil() {
  2800  			// We need to ensure []byte("") evaluates to []byte{}, and not []byte(nil).
  2801  			//
  2802  			// TODO(mdempsky): Investigate using "len != 0" instead of "ptr != nil".
  2803  			cond := s.newValue2(ssa.OpNeqPtr, types.Types[types.TBOOL], ptr, s.constNil(ptr.Type))
  2804  			zerobase := s.newValue1A(ssa.OpAddr, ptr.Type, ir.Syms.Zerobase, s.sb)
  2805  			ptr = s.ternary(cond, ptr, zerobase)
  2806  		}
  2807  		len := s.newValue1(ssa.OpStringLen, types.Types[types.TINT], str)
  2808  		return s.newValue3(ssa.OpSliceMake, n.Type(), ptr, len, len)
  2809  	case ir.OCFUNC:
  2810  		n := n.(*ir.UnaryExpr)
  2811  		aux := n.X.(*ir.Name).Linksym()
  2812  		// OCFUNC is used to build function values, which must
  2813  		// always reference ABIInternal entry points.
  2814  		if aux.ABI() != obj.ABIInternal {
  2815  			s.Fatalf("expected ABIInternal: %v", aux.ABI())
  2816  		}
  2817  		return s.entryNewValue1A(ssa.OpAddr, n.Type(), aux, s.sb)
  2818  	case ir.ONAME:
  2819  		n := n.(*ir.Name)
  2820  		if n.Class == ir.PFUNC {
  2821  			// "value" of a function is the address of the function's closure
  2822  			sym := staticdata.FuncLinksym(n)
  2823  			return s.entryNewValue1A(ssa.OpAddr, types.NewPtr(n.Type()), sym, s.sb)
  2824  		}
  2825  		if s.canSSA(n) {
  2826  			return s.variable(n, n.Type())
  2827  		}
  2828  		return s.load(n.Type(), s.addr(n))
  2829  	case ir.OLINKSYMOFFSET:
  2830  		n := n.(*ir.LinksymOffsetExpr)
  2831  		return s.load(n.Type(), s.addr(n))
  2832  	case ir.ONIL:
  2833  		n := n.(*ir.NilExpr)
  2834  		t := n.Type()
  2835  		switch {
  2836  		case t.IsSlice():
  2837  			return s.constSlice(t)
  2838  		case t.IsInterface():
  2839  			return s.constInterface(t)
  2840  		default:
  2841  			return s.constNil(t)
  2842  		}
  2843  	case ir.OLITERAL:
  2844  		switch u := n.Val(); u.Kind() {
  2845  		case constant.Int:
  2846  			i := ir.IntVal(n.Type(), u)
  2847  			switch n.Type().Size() {
  2848  			case 1:
  2849  				return s.constInt8(n.Type(), int8(i))
  2850  			case 2:
  2851  				return s.constInt16(n.Type(), int16(i))
  2852  			case 4:
  2853  				return s.constInt32(n.Type(), int32(i))
  2854  			case 8:
  2855  				return s.constInt64(n.Type(), i)
  2856  			default:
  2857  				s.Fatalf("bad integer size %d", n.Type().Size())
  2858  				return nil
  2859  			}
  2860  		case constant.String:
  2861  			i := constant.StringVal(u)
  2862  			if i == "" {
  2863  				return s.constEmptyString(n.Type())
  2864  			}
  2865  			return s.entryNewValue0A(ssa.OpConstString, n.Type(), ssa.StringToAux(i))
  2866  		case constant.Bool:
  2867  			return s.constBool(constant.BoolVal(u))
  2868  		case constant.Float:
  2869  			f, _ := constant.Float64Val(u)
  2870  			switch n.Type().Size() {
  2871  			case 4:
  2872  				return s.constFloat32(n.Type(), f)
  2873  			case 8:
  2874  				return s.constFloat64(n.Type(), f)
  2875  			default:
  2876  				s.Fatalf("bad float size %d", n.Type().Size())
  2877  				return nil
  2878  			}
  2879  		case constant.Complex:
  2880  			re, _ := constant.Float64Val(constant.Real(u))
  2881  			im, _ := constant.Float64Val(constant.Imag(u))
  2882  			switch n.Type().Size() {
  2883  			case 8:
  2884  				pt := types.Types[types.TFLOAT32]
  2885  				return s.newValue2(ssa.OpComplexMake, n.Type(),
  2886  					s.constFloat32(pt, re),
  2887  					s.constFloat32(pt, im))
  2888  			case 16:
  2889  				pt := types.Types[types.TFLOAT64]
  2890  				return s.newValue2(ssa.OpComplexMake, n.Type(),
  2891  					s.constFloat64(pt, re),
  2892  					s.constFloat64(pt, im))
  2893  			default:
  2894  				s.Fatalf("bad complex size %d", n.Type().Size())
  2895  				return nil
  2896  			}
  2897  		default:
  2898  			s.Fatalf("unhandled OLITERAL %v", u.Kind())
  2899  			return nil
  2900  		}
  2901  	case ir.OCONVNOP:
  2902  		n := n.(*ir.ConvExpr)
  2903  		to := n.Type()
  2904  		from := n.X.Type()
  2905  
  2906  		// Assume everything will work out, so set up our return value.
  2907  		// Anything interesting that happens from here is a fatal.
  2908  		x := s.expr(n.X)
  2909  		if to == from {
  2910  			return x
  2911  		}
  2912  
  2913  		// Special case for not confusing GC and liveness.
  2914  		// We don't want pointers accidentally classified
  2915  		// as not-pointers or vice-versa because of copy
  2916  		// elision.
  2917  		if to.IsPtrShaped() != from.IsPtrShaped() {
  2918  			return s.newValue2(ssa.OpConvert, to, x, s.mem())
  2919  		}
  2920  
  2921  		v := s.newValue1(ssa.OpCopy, to, x) // ensure that v has the right type
  2922  
  2923  		// CONVNOP closure
  2924  		if to.Kind() == types.TFUNC && from.IsPtrShaped() {
  2925  			return v
  2926  		}
  2927  
  2928  		// named <--> unnamed type or typed <--> untyped const
  2929  		if from.Kind() == to.Kind() {
  2930  			return v
  2931  		}
  2932  
  2933  		// unsafe.Pointer <--> *T
  2934  		if to.IsUnsafePtr() && from.IsPtrShaped() || from.IsUnsafePtr() && to.IsPtrShaped() {
  2935  			if s.checkPtrEnabled && checkPtrOK && to.IsPtr() && from.IsUnsafePtr() {
  2936  				s.checkPtrAlignment(n, v, nil)
  2937  			}
  2938  			return v
  2939  		}
  2940  
  2941  		// map <--> *hmap
  2942  		if to.Kind() == types.TMAP && from == types.NewPtr(reflectdata.MapType()) {
  2943  			return v
  2944  		}
  2945  
  2946  		types.CalcSize(from)
  2947  		types.CalcSize(to)
  2948  		if from.Size() != to.Size() {
  2949  			s.Fatalf("CONVNOP width mismatch %v (%d) -> %v (%d)\n", from, from.Size(), to, to.Size())
  2950  			return nil
  2951  		}
  2952  		if etypesign(from.Kind()) != etypesign(to.Kind()) {
  2953  			s.Fatalf("CONVNOP sign mismatch %v (%s) -> %v (%s)\n", from, from.Kind(), to, to.Kind())
  2954  			return nil
  2955  		}
  2956  
  2957  		if base.Flag.Cfg.Instrumenting {
  2958  			// These appear to be fine, but they fail the
  2959  			// integer constraint below, so okay them here.
  2960  			// Sample non-integer conversion: map[string]string -> *uint8
  2961  			return v
  2962  		}
  2963  
  2964  		if etypesign(from.Kind()) == 0 {
  2965  			s.Fatalf("CONVNOP unrecognized non-integer %v -> %v\n", from, to)
  2966  			return nil
  2967  		}
  2968  
  2969  		// integer, same width, same sign
  2970  		return v
  2971  
  2972  	case ir.OCONV:
  2973  		n := n.(*ir.ConvExpr)
  2974  		x := s.expr(n.X)
  2975  		return s.conv(n, x, n.X.Type(), n.Type())
  2976  
  2977  	case ir.ODOTTYPE:
  2978  		n := n.(*ir.TypeAssertExpr)
  2979  		res, _ := s.dottype(n, false)
  2980  		return res
  2981  
  2982  	case ir.ODYNAMICDOTTYPE:
  2983  		n := n.(*ir.DynamicTypeAssertExpr)
  2984  		res, _ := s.dynamicDottype(n, false)
  2985  		return res
  2986  
  2987  	// binary ops
  2988  	case ir.OLT, ir.OEQ, ir.ONE, ir.OLE, ir.OGE, ir.OGT:
  2989  		n := n.(*ir.BinaryExpr)
  2990  		a := s.expr(n.X)
  2991  		b := s.expr(n.Y)
  2992  		if n.X.Type().IsComplex() {
  2993  			pt := types.FloatForComplex(n.X.Type())
  2994  			op := s.ssaOp(ir.OEQ, pt)
  2995  			r := s.newValueOrSfCall2(op, types.Types[types.TBOOL], s.newValue1(ssa.OpComplexReal, pt, a), s.newValue1(ssa.OpComplexReal, pt, b))
  2996  			i := s.newValueOrSfCall2(op, types.Types[types.TBOOL], s.newValue1(ssa.OpComplexImag, pt, a), s.newValue1(ssa.OpComplexImag, pt, b))
  2997  			c := s.newValue2(ssa.OpAndB, types.Types[types.TBOOL], r, i)
  2998  			switch n.Op() {
  2999  			case ir.OEQ:
  3000  				return c
  3001  			case ir.ONE:
  3002  				return s.newValue1(ssa.OpNot, types.Types[types.TBOOL], c)
  3003  			default:
  3004  				s.Fatalf("ordered complex compare %v", n.Op())
  3005  			}
  3006  		}
  3007  
  3008  		// Convert OGE and OGT into OLE and OLT.
  3009  		op := n.Op()
  3010  		switch op {
  3011  		case ir.OGE:
  3012  			op, a, b = ir.OLE, b, a
  3013  		case ir.OGT:
  3014  			op, a, b = ir.OLT, b, a
  3015  		}
  3016  		if n.X.Type().IsFloat() {
  3017  			// float comparison
  3018  			return s.newValueOrSfCall2(s.ssaOp(op, n.X.Type()), types.Types[types.TBOOL], a, b)
  3019  		}
  3020  		// integer comparison
  3021  		return s.newValue2(s.ssaOp(op, n.X.Type()), types.Types[types.TBOOL], a, b)
  3022  	case ir.OMUL:
  3023  		n := n.(*ir.BinaryExpr)
  3024  		a := s.expr(n.X)
  3025  		b := s.expr(n.Y)
  3026  		if n.Type().IsComplex() {
  3027  			mulop := ssa.OpMul64F
  3028  			addop := ssa.OpAdd64F
  3029  			subop := ssa.OpSub64F
  3030  			pt := types.FloatForComplex(n.Type()) // Could be Float32 or Float64
  3031  			wt := types.Types[types.TFLOAT64]     // Compute in Float64 to minimize cancellation error
  3032  
  3033  			areal := s.newValue1(ssa.OpComplexReal, pt, a)
  3034  			breal := s.newValue1(ssa.OpComplexReal, pt, b)
  3035  			aimag := s.newValue1(ssa.OpComplexImag, pt, a)
  3036  			bimag := s.newValue1(ssa.OpComplexImag, pt, b)
  3037  
  3038  			if pt != wt { // Widen for calculation
  3039  				areal = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, areal)
  3040  				breal = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, breal)
  3041  				aimag = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, aimag)
  3042  				bimag = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, bimag)
  3043  			}
  3044  
  3045  			xreal := s.newValueOrSfCall2(subop, wt, s.newValueOrSfCall2(mulop, wt, areal, breal), s.newValueOrSfCall2(mulop, wt, aimag, bimag))
  3046  			ximag := s.newValueOrSfCall2(addop, wt, s.newValueOrSfCall2(mulop, wt, areal, bimag), s.newValueOrSfCall2(mulop, wt, aimag, breal))
  3047  
  3048  			if pt != wt { // Narrow to store back
  3049  				xreal = s.newValueOrSfCall1(ssa.OpCvt64Fto32F, pt, xreal)
  3050  				ximag = s.newValueOrSfCall1(ssa.OpCvt64Fto32F, pt, ximag)
  3051  			}
  3052  
  3053  			return s.newValue2(ssa.OpComplexMake, n.Type(), xreal, ximag)
  3054  		}
  3055  
  3056  		if n.Type().IsFloat() {
  3057  			return s.newValueOrSfCall2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
  3058  		}
  3059  
  3060  		return s.newValue2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
  3061  
  3062  	case ir.ODIV:
  3063  		n := n.(*ir.BinaryExpr)
  3064  		a := s.expr(n.X)
  3065  		b := s.expr(n.Y)
  3066  		if n.Type().IsComplex() {
  3067  			// TODO this is not executed because the front-end substitutes a runtime call.
  3068  			// That probably ought to change; with modest optimization the widen/narrow
  3069  			// conversions could all be elided in larger expression trees.
  3070  			mulop := ssa.OpMul64F
  3071  			addop := ssa.OpAdd64F
  3072  			subop := ssa.OpSub64F
  3073  			divop := ssa.OpDiv64F
  3074  			pt := types.FloatForComplex(n.Type()) // Could be Float32 or Float64
  3075  			wt := types.Types[types.TFLOAT64]     // Compute in Float64 to minimize cancellation error
  3076  
  3077  			areal := s.newValue1(ssa.OpComplexReal, pt, a)
  3078  			breal := s.newValue1(ssa.OpComplexReal, pt, b)
  3079  			aimag := s.newValue1(ssa.OpComplexImag, pt, a)
  3080  			bimag := s.newValue1(ssa.OpComplexImag, pt, b)
  3081  
  3082  			if pt != wt { // Widen for calculation
  3083  				areal = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, areal)
  3084  				breal = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, breal)
  3085  				aimag = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, aimag)
  3086  				bimag = s.newValueOrSfCall1(ssa.OpCvt32Fto64F, wt, bimag)
  3087  			}
  3088  
  3089  			denom := s.newValueOrSfCall2(addop, wt, s.newValueOrSfCall2(mulop, wt, breal, breal), s.newValueOrSfCall2(mulop, wt, bimag, bimag))
  3090  			xreal := s.newValueOrSfCall2(addop, wt, s.newValueOrSfCall2(mulop, wt, areal, breal), s.newValueOrSfCall2(mulop, wt, aimag, bimag))
  3091  			ximag := s.newValueOrSfCall2(subop, wt, s.newValueOrSfCall2(mulop, wt, aimag, breal), s.newValueOrSfCall2(mulop, wt, areal, bimag))
  3092  
  3093  			// TODO not sure if this is best done in wide precision or narrow
  3094  			// Double-rounding might be an issue.
  3095  			// Note that the pre-SSA implementation does the entire calculation
  3096  			// in wide format, so wide is compatible.
  3097  			xreal = s.newValueOrSfCall2(divop, wt, xreal, denom)
  3098  			ximag = s.newValueOrSfCall2(divop, wt, ximag, denom)
  3099  
  3100  			if pt != wt { // Narrow to store back
  3101  				xreal = s.newValueOrSfCall1(ssa.OpCvt64Fto32F, pt, xreal)
  3102  				ximag = s.newValueOrSfCall1(ssa.OpCvt64Fto32F, pt, ximag)
  3103  			}
  3104  			return s.newValue2(ssa.OpComplexMake, n.Type(), xreal, ximag)
  3105  		}
  3106  		if n.Type().IsFloat() {
  3107  			return s.newValueOrSfCall2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
  3108  		}
  3109  		return s.intDivide(n, a, b)
  3110  	case ir.OMOD:
  3111  		n := n.(*ir.BinaryExpr)
  3112  		a := s.expr(n.X)
  3113  		b := s.expr(n.Y)
  3114  		return s.intDivide(n, a, b)
  3115  	case ir.OADD, ir.OSUB:
  3116  		n := n.(*ir.BinaryExpr)
  3117  		a := s.expr(n.X)
  3118  		b := s.expr(n.Y)
  3119  		if n.Type().IsComplex() {
  3120  			pt := types.FloatForComplex(n.Type())
  3121  			op := s.ssaOp(n.Op(), pt)
  3122  			return s.newValue2(ssa.OpComplexMake, n.Type(),
  3123  				s.newValueOrSfCall2(op, pt, s.newValue1(ssa.OpComplexReal, pt, a), s.newValue1(ssa.OpComplexReal, pt, b)),
  3124  				s.newValueOrSfCall2(op, pt, s.newValue1(ssa.OpComplexImag, pt, a), s.newValue1(ssa.OpComplexImag, pt, b)))
  3125  		}
  3126  		if n.Type().IsFloat() {
  3127  			return s.newValueOrSfCall2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
  3128  		}
  3129  		return s.newValue2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
  3130  	case ir.OAND, ir.OOR, ir.OXOR:
  3131  		n := n.(*ir.BinaryExpr)
  3132  		a := s.expr(n.X)
  3133  		b := s.expr(n.Y)
  3134  		return s.newValue2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
  3135  	case ir.OANDNOT:
  3136  		n := n.(*ir.BinaryExpr)
  3137  		a := s.expr(n.X)
  3138  		b := s.expr(n.Y)
  3139  		b = s.newValue1(s.ssaOp(ir.OBITNOT, b.Type), b.Type, b)
  3140  		return s.newValue2(s.ssaOp(ir.OAND, n.Type()), a.Type, a, b)
  3141  	case ir.OLSH, ir.ORSH:
  3142  		n := n.(*ir.BinaryExpr)
  3143  		a := s.expr(n.X)
  3144  		b := s.expr(n.Y)
  3145  		bt := b.Type
  3146  		if bt.IsSigned() {
  3147  			cmp := s.newValue2(s.ssaOp(ir.OLE, bt), types.Types[types.TBOOL], s.zeroVal(bt), b)
  3148  			s.check(cmp, ir.Syms.Panicshift)
  3149  			bt = bt.ToUnsigned()
  3150  		}
  3151  		return s.newValue2(s.ssaShiftOp(n.Op(), n.Type(), bt), a.Type, a, b)
  3152  	case ir.OANDAND, ir.OOROR:
  3153  		// To implement OANDAND (and OOROR), we introduce a
  3154  		// new temporary variable to hold the result. The
  3155  		// variable is associated with the OANDAND node in the
  3156  		// s.vars table (normally variables are only
  3157  		// associated with ONAME nodes). We convert
  3158  		//     A && B
  3159  		// to
  3160  		//     var = A
  3161  		//     if var {
  3162  		//         var = B
  3163  		//     }
  3164  		// Using var in the subsequent block introduces the
  3165  		// necessary phi variable.
  3166  		n := n.(*ir.LogicalExpr)
  3167  		el := s.expr(n.X)
  3168  		s.vars[n] = el
  3169  
  3170  		b := s.endBlock()
  3171  		b.Kind = ssa.BlockIf
  3172  		b.SetControl(el)
  3173  		// In theory, we should set b.Likely here based on context.
  3174  		// However, gc only gives us likeliness hints
  3175  		// in a single place, for plain OIF statements,
  3176  		// and passing around context is finicky, so don't bother for now.
  3177  
  3178  		bRight := s.f.NewBlock(ssa.BlockPlain)
  3179  		bResult := s.f.NewBlock(ssa.BlockPlain)
  3180  		if n.Op() == ir.OANDAND {
  3181  			b.AddEdgeTo(bRight)
  3182  			b.AddEdgeTo(bResult)
  3183  		} else if n.Op() == ir.OOROR {
  3184  			b.AddEdgeTo(bResult)
  3185  			b.AddEdgeTo(bRight)
  3186  		}
  3187  
  3188  		s.startBlock(bRight)
  3189  		er := s.expr(n.Y)
  3190  		s.vars[n] = er
  3191  
  3192  		b = s.endBlock()
  3193  		b.AddEdgeTo(bResult)
  3194  
  3195  		s.startBlock(bResult)
  3196  		return s.variable(n, types.Types[types.TBOOL])
  3197  	case ir.OCOMPLEX:
  3198  		n := n.(*ir.BinaryExpr)
  3199  		r := s.expr(n.X)
  3200  		i := s.expr(n.Y)
  3201  		return s.newValue2(ssa.OpComplexMake, n.Type(), r, i)
  3202  
  3203  	// unary ops
  3204  	case ir.ONEG:
  3205  		n := n.(*ir.UnaryExpr)
  3206  		a := s.expr(n.X)
  3207  		if n.Type().IsComplex() {
  3208  			tp := types.FloatForComplex(n.Type())
  3209  			negop := s.ssaOp(n.Op(), tp)
  3210  			return s.newValue2(ssa.OpComplexMake, n.Type(),
  3211  				s.newValue1(negop, tp, s.newValue1(ssa.OpComplexReal, tp, a)),
  3212  				s.newValue1(negop, tp, s.newValue1(ssa.OpComplexImag, tp, a)))
  3213  		}
  3214  		return s.newValue1(s.ssaOp(n.Op(), n.Type()), a.Type, a)
  3215  	case ir.ONOT, ir.OBITNOT:
  3216  		n := n.(*ir.UnaryExpr)
  3217  		a := s.expr(n.X)
  3218  		return s.newValue1(s.ssaOp(n.Op(), n.Type()), a.Type, a)
  3219  	case ir.OIMAG, ir.OREAL:
  3220  		n := n.(*ir.UnaryExpr)
  3221  		a := s.expr(n.X)
  3222  		return s.newValue1(s.ssaOp(n.Op(), n.X.Type()), n.Type(), a)
  3223  	case ir.OPLUS:
  3224  		n := n.(*ir.UnaryExpr)
  3225  		return s.expr(n.X)
  3226  
  3227  	case ir.OADDR:
  3228  		n := n.(*ir.AddrExpr)
  3229  		return s.addr(n.X)
  3230  
  3231  	case ir.ORESULT:
  3232  		n := n.(*ir.ResultExpr)
  3233  		if s.prevCall == nil || s.prevCall.Op != ssa.OpStaticLECall && s.prevCall.Op != ssa.OpInterLECall && s.prevCall.Op != ssa.OpClosureLECall {
  3234  			panic("Expected to see a previous call")
  3235  		}
  3236  		which := n.Index
  3237  		if which == -1 {
  3238  			panic(fmt.Errorf("ORESULT %v does not match call %s", n, s.prevCall))
  3239  		}
  3240  		return s.resultOfCall(s.prevCall, which, n.Type())
  3241  
  3242  	case ir.ODEREF:
  3243  		n := n.(*ir.StarExpr)
  3244  		p := s.exprPtr(n.X, n.Bounded(), n.Pos())
  3245  		return s.load(n.Type(), p)
  3246  
  3247  	case ir.ODOT:
  3248  		n := n.(*ir.SelectorExpr)
  3249  		if n.X.Op() == ir.OSTRUCTLIT {
  3250  			// All literals with nonzero fields have already been
  3251  			// rewritten during walk. Any that remain are just T{}
  3252  			// or equivalents. Use the zero value.
  3253  			if !ir.IsZero(n.X) {
  3254  				s.Fatalf("literal with nonzero value in SSA: %v", n.X)
  3255  			}
  3256  			return s.zeroVal(n.Type())
  3257  		}
  3258  		// If n is addressable and can't be represented in
  3259  		// SSA, then load just the selected field. This
  3260  		// prevents false memory dependencies in race/msan/asan
  3261  		// instrumentation.
  3262  		if ir.IsAddressable(n) && !s.canSSA(n) {
  3263  			p := s.addr(n)
  3264  			return s.load(n.Type(), p)
  3265  		}
  3266  		v := s.expr(n.X)
  3267  		return s.newValue1I(ssa.OpStructSelect, n.Type(), int64(fieldIdx(n)), v)
  3268  
  3269  	case ir.ODOTPTR:
  3270  		n := n.(*ir.SelectorExpr)
  3271  		p := s.exprPtr(n.X, n.Bounded(), n.Pos())
  3272  		p = s.newValue1I(ssa.OpOffPtr, types.NewPtr(n.Type()), n.Offset(), p)
  3273  		return s.load(n.Type(), p)
  3274  
  3275  	case ir.OINDEX:
  3276  		n := n.(*ir.IndexExpr)
  3277  		switch {
  3278  		case n.X.Type().IsString():
  3279  			if n.Bounded() && ir.IsConst(n.X, constant.String) && ir.IsConst(n.Index, constant.Int) {
  3280  				// Replace "abc"[1] with 'b'.
  3281  				// Delayed until now because "abc"[1] is not an ideal constant.
  3282  				// See test/fixedbugs/issue11370.go.
  3283  				return s.newValue0I(ssa.OpConst8, types.Types[types.TUINT8], int64(int8(ir.StringVal(n.X)[ir.Int64Val(n.Index)])))
  3284  			}
  3285  			a := s.expr(n.X)
  3286  			i := s.expr(n.Index)
  3287  			len := s.newValue1(ssa.OpStringLen, types.Types[types.TINT], a)
  3288  			i = s.boundsCheck(i, len, ssa.BoundsIndex, n.Bounded())
  3289  			ptrtyp := s.f.Config.Types.BytePtr
  3290  			ptr := s.newValue1(ssa.OpStringPtr, ptrtyp, a)
  3291  			if ir.IsConst(n.Index, constant.Int) {
  3292  				ptr = s.newValue1I(ssa.OpOffPtr, ptrtyp, ir.Int64Val(n.Index), ptr)
  3293  			} else {
  3294  				ptr = s.newValue2(ssa.OpAddPtr, ptrtyp, ptr, i)
  3295  			}
  3296  			return s.load(types.Types[types.TUINT8], ptr)
  3297  		case n.X.Type().IsSlice():
  3298  			p := s.addr(n)
  3299  			return s.load(n.X.Type().Elem(), p)
  3300  		case n.X.Type().IsArray():
  3301  			if ssa.CanSSA(n.X.Type()) {
  3302  				// SSA can handle arrays of length at most 1.
  3303  				bound := n.X.Type().NumElem()
  3304  				a := s.expr(n.X)
  3305  				i := s.expr(n.Index)
  3306  				if bound == 0 {
  3307  					// Bounds check will never succeed.  Might as well
  3308  					// use constants for the bounds check.
  3309  					z := s.constInt(types.Types[types.TINT], 0)
  3310  					s.boundsCheck(z, z, ssa.BoundsIndex, false)
  3311  					// The return value won't be live, return junk.
  3312  					// But not quite junk, in case bounds checks are turned off. See issue 48092.
  3313  					return s.zeroVal(n.Type())
  3314  				}
  3315  				len := s.constInt(types.Types[types.TINT], bound)
  3316  				s.boundsCheck(i, len, ssa.BoundsIndex, n.Bounded()) // checks i == 0
  3317  				return s.newValue1I(ssa.OpArraySelect, n.Type(), 0, a)
  3318  			}
  3319  			p := s.addr(n)
  3320  			return s.load(n.X.Type().Elem(), p)
  3321  		default:
  3322  			s.Fatalf("bad type for index %v", n.X.Type())
  3323  			return nil
  3324  		}
  3325  
  3326  	case ir.OLEN, ir.OCAP:
  3327  		n := n.(*ir.UnaryExpr)
  3328  		switch {
  3329  		case n.X.Type().IsSlice():
  3330  			op := ssa.OpSliceLen
  3331  			if n.Op() == ir.OCAP {
  3332  				op = ssa.OpSliceCap
  3333  			}
  3334  			return s.newValue1(op, types.Types[types.TINT], s.expr(n.X))
  3335  		case n.X.Type().IsString(): // string; not reachable for OCAP
  3336  			return s.newValue1(ssa.OpStringLen, types.Types[types.TINT], s.expr(n.X))
  3337  		case n.X.Type().IsMap(), n.X.Type().IsChan():
  3338  			return s.referenceTypeBuiltin(n, s.expr(n.X))
  3339  		default: // array
  3340  			return s.constInt(types.Types[types.TINT], n.X.Type().NumElem())
  3341  		}
  3342  
  3343  	case ir.OSPTR:
  3344  		n := n.(*ir.UnaryExpr)
  3345  		a := s.expr(n.X)
  3346  		if n.X.Type().IsSlice() {
  3347  			if n.Bounded() {
  3348  				return s.newValue1(ssa.OpSlicePtr, n.Type(), a)
  3349  			}
  3350  			return s.newValue1(ssa.OpSlicePtrUnchecked, n.Type(), a)
  3351  		} else {
  3352  			return s.newValue1(ssa.OpStringPtr, n.Type(), a)
  3353  		}
  3354  
  3355  	case ir.OITAB:
  3356  		n := n.(*ir.UnaryExpr)
  3357  		a := s.expr(n.X)
  3358  		return s.newValue1(ssa.OpITab, n.Type(), a)
  3359  
  3360  	case ir.OIDATA:
  3361  		n := n.(*ir.UnaryExpr)
  3362  		a := s.expr(n.X)
  3363  		return s.newValue1(ssa.OpIData, n.Type(), a)
  3364  
  3365  	case ir.OMAKEFACE:
  3366  		n := n.(*ir.BinaryExpr)
  3367  		tab := s.expr(n.X)
  3368  		data := s.expr(n.Y)
  3369  		return s.newValue2(ssa.OpIMake, n.Type(), tab, data)
  3370  
  3371  	case ir.OSLICEHEADER:
  3372  		n := n.(*ir.SliceHeaderExpr)
  3373  		p := s.expr(n.Ptr)
  3374  		l := s.expr(n.Len)
  3375  		c := s.expr(n.Cap)
  3376  		return s.newValue3(ssa.OpSliceMake, n.Type(), p, l, c)
  3377  
  3378  	case ir.OSTRINGHEADER:
  3379  		n := n.(*ir.StringHeaderExpr)
  3380  		p := s.expr(n.Ptr)
  3381  		l := s.expr(n.Len)
  3382  		return s.newValue2(ssa.OpStringMake, n.Type(), p, l)
  3383  
  3384  	case ir.OSLICE, ir.OSLICEARR, ir.OSLICE3, ir.OSLICE3ARR:
  3385  		n := n.(*ir.SliceExpr)
  3386  		check := s.checkPtrEnabled && n.Op() == ir.OSLICE3ARR && n.X.Op() == ir.OCONVNOP && n.X.(*ir.ConvExpr).X.Type().IsUnsafePtr()
  3387  		v := s.exprCheckPtr(n.X, !check)
  3388  		var i, j, k *ssa.Value
  3389  		if n.Low != nil {
  3390  			i = s.expr(n.Low)
  3391  		}
  3392  		if n.High != nil {
  3393  			j = s.expr(n.High)
  3394  		}
  3395  		if n.Max != nil {
  3396  			k = s.expr(n.Max)
  3397  		}
  3398  		p, l, c := s.slice(v, i, j, k, n.Bounded())
  3399  		if check {
  3400  			// Emit checkptr instrumentation after bound check to prevent false positive, see #46938.
  3401  			s.checkPtrAlignment(n.X.(*ir.ConvExpr), v, s.conv(n.Max, k, k.Type, types.Types[types.TUINTPTR]))
  3402  		}
  3403  		return s.newValue3(ssa.OpSliceMake, n.Type(), p, l, c)
  3404  
  3405  	case ir.OSLICESTR:
  3406  		n := n.(*ir.SliceExpr)
  3407  		v := s.expr(n.X)
  3408  		var i, j *ssa.Value
  3409  		if n.Low != nil {
  3410  			i = s.expr(n.Low)
  3411  		}
  3412  		if n.High != nil {
  3413  			j = s.expr(n.High)
  3414  		}
  3415  		p, l, _ := s.slice(v, i, j, nil, n.Bounded())
  3416  		return s.newValue2(ssa.OpStringMake, n.Type(), p, l)
  3417  
  3418  	case ir.OSLICE2ARRPTR:
  3419  		// if arrlen > slice.len {
  3420  		//   panic(...)
  3421  		// }
  3422  		// slice.ptr
  3423  		n := n.(*ir.ConvExpr)
  3424  		v := s.expr(n.X)
  3425  		nelem := n.Type().Elem().NumElem()
  3426  		arrlen := s.constInt(types.Types[types.TINT], nelem)
  3427  		cap := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], v)
  3428  		s.boundsCheck(arrlen, cap, ssa.BoundsConvert, false)
  3429  		op := ssa.OpSlicePtr
  3430  		if nelem == 0 {
  3431  			op = ssa.OpSlicePtrUnchecked
  3432  		}
  3433  		return s.newValue1(op, n.Type(), v)
  3434  
  3435  	case ir.OCALLFUNC:
  3436  		n := n.(*ir.CallExpr)
  3437  		if ir.IsIntrinsicCall(n) {
  3438  			return s.intrinsicCall(n)
  3439  		}
  3440  		fallthrough
  3441  
  3442  	case ir.OCALLINTER:
  3443  		n := n.(*ir.CallExpr)
  3444  		return s.callResult(n, callNormal)
  3445  
  3446  	case ir.OGETG:
  3447  		n := n.(*ir.CallExpr)
  3448  		return s.newValue1(ssa.OpGetG, n.Type(), s.mem())
  3449  
  3450  	case ir.OGETCALLERPC:
  3451  		n := n.(*ir.CallExpr)
  3452  		return s.newValue0(ssa.OpGetCallerPC, n.Type())
  3453  
  3454  	case ir.OGETCALLERSP:
  3455  		n := n.(*ir.CallExpr)
  3456  		return s.newValue1(ssa.OpGetCallerSP, n.Type(), s.mem())
  3457  
  3458  	case ir.OAPPEND:
  3459  		return s.append(n.(*ir.CallExpr), false)
  3460  
  3461  	case ir.OMIN, ir.OMAX:
  3462  		return s.minMax(n.(*ir.CallExpr))
  3463  
  3464  	case ir.OSTRUCTLIT, ir.OARRAYLIT:
  3465  		// All literals with nonzero fields have already been
  3466  		// rewritten during walk. Any that remain are just T{}
  3467  		// or equivalents. Use the zero value.
  3468  		n := n.(*ir.CompLitExpr)
  3469  		if !ir.IsZero(n) {
  3470  			s.Fatalf("literal with nonzero value in SSA: %v", n)
  3471  		}
  3472  		return s.zeroVal(n.Type())
  3473  
  3474  	case ir.ONEW:
  3475  		n := n.(*ir.UnaryExpr)
  3476  		var rtype *ssa.Value
  3477  		if x, ok := n.X.(*ir.DynamicType); ok && x.Op() == ir.ODYNAMICTYPE {
  3478  			rtype = s.expr(x.RType)
  3479  		}
  3480  		return s.newObject(n.Type().Elem(), rtype)
  3481  
  3482  	case ir.OUNSAFEADD:
  3483  		n := n.(*ir.BinaryExpr)
  3484  		ptr := s.expr(n.X)
  3485  		len := s.expr(n.Y)
  3486  
  3487  		// Force len to uintptr to prevent misuse of garbage bits in the
  3488  		// upper part of the register (#48536).
  3489  		len = s.conv(n, len, len.Type, types.Types[types.TUINTPTR])
  3490  
  3491  		return s.newValue2(ssa.OpAddPtr, n.Type(), ptr, len)
  3492  
  3493  	default:
  3494  		s.Fatalf("unhandled expr %v", n.Op())
  3495  		return nil
  3496  	}
  3497  }
  3498  
  3499  func (s *state) resultOfCall(c *ssa.Value, which int64, t *types.Type) *ssa.Value {
  3500  	aux := c.Aux.(*ssa.AuxCall)
  3501  	pa := aux.ParamAssignmentForResult(which)
  3502  	// TODO(register args) determine if in-memory TypeOK is better loaded early from SelectNAddr or later when SelectN is expanded.
  3503  	// SelectN is better for pattern-matching and possible call-aware analysis we might want to do in the future.
  3504  	if len(pa.Registers) == 0 && !ssa.CanSSA(t) {
  3505  		addr := s.newValue1I(ssa.OpSelectNAddr, types.NewPtr(t), which, c)
  3506  		return s.rawLoad(t, addr)
  3507  	}
  3508  	return s.newValue1I(ssa.OpSelectN, t, which, c)
  3509  }
  3510  
  3511  func (s *state) resultAddrOfCall(c *ssa.Value, which int64, t *types.Type) *ssa.Value {
  3512  	aux := c.Aux.(*ssa.AuxCall)
  3513  	pa := aux.ParamAssignmentForResult(which)
  3514  	if len(pa.Registers) == 0 {
  3515  		return s.newValue1I(ssa.OpSelectNAddr, types.NewPtr(t), which, c)
  3516  	}
  3517  	_, addr := s.temp(c.Pos, t)
  3518  	rval := s.newValue1I(ssa.OpSelectN, t, which, c)
  3519  	s.vars[memVar] = s.newValue3Apos(ssa.OpStore, types.TypeMem, t, addr, rval, s.mem(), false)
  3520  	return addr
  3521  }
  3522  
  3523  // append converts an OAPPEND node to SSA.
  3524  // If inplace is false, it converts the OAPPEND expression n to an ssa.Value,
  3525  // adds it to s, and returns the Value.
  3526  // If inplace is true, it writes the result of the OAPPEND expression n
  3527  // back to the slice being appended to, and returns nil.
  3528  // inplace MUST be set to false if the slice can be SSA'd.
  3529  // Note: this code only handles fixed-count appends. Dotdotdot appends
  3530  // have already been rewritten at this point (by walk).
  3531  func (s *state) append(n *ir.CallExpr, inplace bool) *ssa.Value {
  3532  	// If inplace is false, process as expression "append(s, e1, e2, e3)":
  3533  	//
  3534  	// ptr, len, cap := s
  3535  	// len += 3
  3536  	// if uint(len) > uint(cap) {
  3537  	//     ptr, len, cap = growslice(ptr, len, cap, 3, typ)
  3538  	//     Note that len is unmodified by growslice.
  3539  	// }
  3540  	// // with write barriers, if needed:
  3541  	// *(ptr+(len-3)) = e1
  3542  	// *(ptr+(len-2)) = e2
  3543  	// *(ptr+(len-1)) = e3
  3544  	// return makeslice(ptr, len, cap)
  3545  	//
  3546  	//
  3547  	// If inplace is true, process as statement "s = append(s, e1, e2, e3)":
  3548  	//
  3549  	// a := &s
  3550  	// ptr, len, cap := s
  3551  	// len += 3
  3552  	// if uint(len) > uint(cap) {
  3553  	//    ptr, len, cap = growslice(ptr, len, cap, 3, typ)
  3554  	//    vardef(a)    // if necessary, advise liveness we are writing a new a
  3555  	//    *a.cap = cap // write before ptr to avoid a spill
  3556  	//    *a.ptr = ptr // with write barrier
  3557  	// }
  3558  	// *a.len = len
  3559  	// // with write barriers, if needed:
  3560  	// *(ptr+(len-3)) = e1
  3561  	// *(ptr+(len-2)) = e2
  3562  	// *(ptr+(len-1)) = e3
  3563  
  3564  	et := n.Type().Elem()
  3565  	pt := types.NewPtr(et)
  3566  
  3567  	// Evaluate slice
  3568  	sn := n.Args[0] // the slice node is the first in the list
  3569  	var slice, addr *ssa.Value
  3570  	if inplace {
  3571  		addr = s.addr(sn)
  3572  		slice = s.load(n.Type(), addr)
  3573  	} else {
  3574  		slice = s.expr(sn)
  3575  	}
  3576  
  3577  	// Allocate new blocks
  3578  	grow := s.f.NewBlock(ssa.BlockPlain)
  3579  	assign := s.f.NewBlock(ssa.BlockPlain)
  3580  
  3581  	// Decomposse input slice.
  3582  	p := s.newValue1(ssa.OpSlicePtr, pt, slice)
  3583  	l := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], slice)
  3584  	c := s.newValue1(ssa.OpSliceCap, types.Types[types.TINT], slice)
  3585  
  3586  	// Add number of new elements to length.
  3587  	nargs := s.constInt(types.Types[types.TINT], int64(len(n.Args)-1))
  3588  	l = s.newValue2(s.ssaOp(ir.OADD, types.Types[types.TINT]), types.Types[types.TINT], l, nargs)
  3589  
  3590  	// Decide if we need to grow
  3591  	cmp := s.newValue2(s.ssaOp(ir.OLT, types.Types[types.TUINT]), types.Types[types.TBOOL], c, l)
  3592  
  3593  	// Record values of ptr/len/cap before branch.
  3594  	s.vars[ptrVar] = p
  3595  	s.vars[lenVar] = l
  3596  	if !inplace {
  3597  		s.vars[capVar] = c
  3598  	}
  3599  
  3600  	b := s.endBlock()
  3601  	b.Kind = ssa.BlockIf
  3602  	b.Likely = ssa.BranchUnlikely
  3603  	b.SetControl(cmp)
  3604  	b.AddEdgeTo(grow)
  3605  	b.AddEdgeTo(assign)
  3606  
  3607  	// Call growslice
  3608  	s.startBlock(grow)
  3609  	taddr := s.expr(n.Fun)
  3610  	r := s.rtcall(ir.Syms.Growslice, true, []*types.Type{n.Type()}, p, l, c, nargs, taddr)
  3611  
  3612  	// Decompose output slice
  3613  	p = s.newValue1(ssa.OpSlicePtr, pt, r[0])
  3614  	l = s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], r[0])
  3615  	c = s.newValue1(ssa.OpSliceCap, types.Types[types.TINT], r[0])
  3616  
  3617  	s.vars[ptrVar] = p
  3618  	s.vars[lenVar] = l
  3619  	s.vars[capVar] = c
  3620  	if inplace {
  3621  		if sn.Op() == ir.ONAME {
  3622  			sn := sn.(*ir.Name)
  3623  			if sn.Class != ir.PEXTERN {
  3624  				// Tell liveness we're about to build a new slice
  3625  				s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, sn, s.mem())
  3626  			}
  3627  		}
  3628  		capaddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, types.SliceCapOffset, addr)
  3629  		s.store(types.Types[types.TINT], capaddr, c)
  3630  		s.store(pt, addr, p)
  3631  	}
  3632  
  3633  	b = s.endBlock()
  3634  	b.AddEdgeTo(assign)
  3635  
  3636  	// assign new elements to slots
  3637  	s.startBlock(assign)
  3638  	p = s.variable(ptrVar, pt)                      // generates phi for ptr
  3639  	l = s.variable(lenVar, types.Types[types.TINT]) // generates phi for len
  3640  	if !inplace {
  3641  		c = s.variable(capVar, types.Types[types.TINT]) // generates phi for cap
  3642  	}
  3643  
  3644  	if inplace {
  3645  		// Update length in place.
  3646  		// We have to wait until here to make sure growslice succeeded.
  3647  		lenaddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, types.SliceLenOffset, addr)
  3648  		s.store(types.Types[types.TINT], lenaddr, l)
  3649  	}
  3650  
  3651  	// Evaluate args
  3652  	type argRec struct {
  3653  		// if store is true, we're appending the value v.  If false, we're appending the
  3654  		// value at *v.
  3655  		v     *ssa.Value
  3656  		store bool
  3657  	}
  3658  	args := make([]argRec, 0, len(n.Args[1:]))
  3659  	for _, n := range n.Args[1:] {
  3660  		if ssa.CanSSA(n.Type()) {
  3661  			args = append(args, argRec{v: s.expr(n), store: true})
  3662  		} else {
  3663  			v := s.addr(n)
  3664  			args = append(args, argRec{v: v})
  3665  		}
  3666  	}
  3667  
  3668  	// Write args into slice.
  3669  	oldLen := s.newValue2(s.ssaOp(ir.OSUB, types.Types[types.TINT]), types.Types[types.TINT], l, nargs)
  3670  	p2 := s.newValue2(ssa.OpPtrIndex, pt, p, oldLen)
  3671  	for i, arg := range args {
  3672  		addr := s.newValue2(ssa.OpPtrIndex, pt, p2, s.constInt(types.Types[types.TINT], int64(i)))
  3673  		if arg.store {
  3674  			s.storeType(et, addr, arg.v, 0, true)
  3675  		} else {
  3676  			s.move(et, addr, arg.v)
  3677  		}
  3678  	}
  3679  
  3680  	// The following deletions have no practical effect at this time
  3681  	// because state.vars has been reset by the preceding state.startBlock.
  3682  	// They only enforce the fact that these variables are no longer need in
  3683  	// the current scope.
  3684  	delete(s.vars, ptrVar)
  3685  	delete(s.vars, lenVar)
  3686  	if !inplace {
  3687  		delete(s.vars, capVar)
  3688  	}
  3689  
  3690  	// make result
  3691  	if inplace {
  3692  		return nil
  3693  	}
  3694  	return s.newValue3(ssa.OpSliceMake, n.Type(), p, l, c)
  3695  }
  3696  
  3697  // minMax converts an OMIN/OMAX builtin call into SSA.
  3698  func (s *state) minMax(n *ir.CallExpr) *ssa.Value {
  3699  	// The OMIN/OMAX builtin is variadic, but its semantics are
  3700  	// equivalent to left-folding a binary min/max operation across the
  3701  	// arguments list.
  3702  	fold := func(op func(x, a *ssa.Value) *ssa.Value) *ssa.Value {
  3703  		x := s.expr(n.Args[0])
  3704  		for _, arg := range n.Args[1:] {
  3705  			x = op(x, s.expr(arg))
  3706  		}
  3707  		return x
  3708  	}
  3709  
  3710  	typ := n.Type()
  3711  
  3712  	if typ.IsFloat() || typ.IsString() {
  3713  		// min/max semantics for floats are tricky because of NaNs and
  3714  		// negative zero. Some architectures have instructions which
  3715  		// we can use to generate the right result. For others we must
  3716  		// call into the runtime instead.
  3717  		//
  3718  		// Strings are conceptually simpler, but we currently desugar
  3719  		// string comparisons during walk, not ssagen.
  3720  
  3721  		if typ.IsFloat() {
  3722  			hasIntrinsic := false
  3723  			switch Arch.LinkArch.Family {
  3724  			case sys.AMD64, sys.ARM64, sys.RISCV64:
  3725  				hasIntrinsic = true
  3726  			case sys.PPC64:
  3727  				hasIntrinsic = buildcfg.GOPPC64 >= 9
  3728  			}
  3729  
  3730  			if hasIntrinsic {
  3731  				var op ssa.Op
  3732  				switch {
  3733  				case typ.Kind() == types.TFLOAT64 && n.Op() == ir.OMIN:
  3734  					op = ssa.OpMin64F
  3735  				case typ.Kind() == types.TFLOAT64 && n.Op() == ir.OMAX:
  3736  					op = ssa.OpMax64F
  3737  				case typ.Kind() == types.TFLOAT32 && n.Op() == ir.OMIN:
  3738  					op = ssa.OpMin32F
  3739  				case typ.Kind() == types.TFLOAT32 && n.Op() == ir.OMAX:
  3740  					op = ssa.OpMax32F
  3741  				}
  3742  				return fold(func(x, a *ssa.Value) *ssa.Value {
  3743  					return s.newValue2(op, typ, x, a)
  3744  				})
  3745  			}
  3746  		}
  3747  		var name string
  3748  		switch typ.Kind() {
  3749  		case types.TFLOAT32:
  3750  			switch n.Op() {
  3751  			case ir.OMIN:
  3752  				name = "fmin32"
  3753  			case ir.OMAX:
  3754  				name = "fmax32"
  3755  			}
  3756  		case types.TFLOAT64:
  3757  			switch n.Op() {
  3758  			case ir.OMIN:
  3759  				name = "fmin64"
  3760  			case ir.OMAX:
  3761  				name = "fmax64"
  3762  			}
  3763  		case types.TSTRING:
  3764  			switch n.Op() {
  3765  			case ir.OMIN:
  3766  				name = "strmin"
  3767  			case ir.OMAX:
  3768  				name = "strmax"
  3769  			}
  3770  		}
  3771  		fn := typecheck.LookupRuntimeFunc(name)
  3772  
  3773  		return fold(func(x, a *ssa.Value) *ssa.Value {
  3774  			return s.rtcall(fn, true, []*types.Type{typ}, x, a)[0]
  3775  		})
  3776  	}
  3777  
  3778  	lt := s.ssaOp(ir.OLT, typ)
  3779  
  3780  	return fold(func(x, a *ssa.Value) *ssa.Value {
  3781  		switch n.Op() {
  3782  		case ir.OMIN:
  3783  			// a < x ? a : x
  3784  			return s.ternary(s.newValue2(lt, types.Types[types.TBOOL], a, x), a, x)
  3785  		case ir.OMAX:
  3786  			// x < a ? a : x
  3787  			return s.ternary(s.newValue2(lt, types.Types[types.TBOOL], x, a), a, x)
  3788  		}
  3789  		panic("unreachable")
  3790  	})
  3791  }
  3792  
  3793  // ternary emits code to evaluate cond ? x : y.
  3794  func (s *state) ternary(cond, x, y *ssa.Value) *ssa.Value {
  3795  	// Note that we need a new ternaryVar each time (unlike okVar where we can
  3796  	// reuse the variable) because it might have a different type every time.
  3797  	ternaryVar := ssaMarker("ternary")
  3798  
  3799  	bThen := s.f.NewBlock(ssa.BlockPlain)
  3800  	bElse := s.f.NewBlock(ssa.BlockPlain)
  3801  	bEnd := s.f.NewBlock(ssa.BlockPlain)
  3802  
  3803  	b := s.endBlock()
  3804  	b.Kind = ssa.BlockIf
  3805  	b.SetControl(cond)
  3806  	b.AddEdgeTo(bThen)
  3807  	b.AddEdgeTo(bElse)
  3808  
  3809  	s.startBlock(bThen)
  3810  	s.vars[ternaryVar] = x
  3811  	s.endBlock().AddEdgeTo(bEnd)
  3812  
  3813  	s.startBlock(bElse)
  3814  	s.vars[ternaryVar] = y
  3815  	s.endBlock().AddEdgeTo(bEnd)
  3816  
  3817  	s.startBlock(bEnd)
  3818  	r := s.variable(ternaryVar, x.Type)
  3819  	delete(s.vars, ternaryVar)
  3820  	return r
  3821  }
  3822  
  3823  // condBranch evaluates the boolean expression cond and branches to yes
  3824  // if cond is true and no if cond is false.
  3825  // This function is intended to handle && and || better than just calling
  3826  // s.expr(cond) and branching on the result.
  3827  func (s *state) condBranch(cond ir.Node, yes, no *ssa.Block, likely int8) {
  3828  	switch cond.Op() {
  3829  	case ir.OANDAND:
  3830  		cond := cond.(*ir.LogicalExpr)
  3831  		mid := s.f.NewBlock(ssa.BlockPlain)
  3832  		s.stmtList(cond.Init())
  3833  		s.condBranch(cond.X, mid, no, max8(likely, 0))
  3834  		s.startBlock(mid)
  3835  		s.condBranch(cond.Y, yes, no, likely)
  3836  		return
  3837  		// Note: if likely==1, then both recursive calls pass 1.
  3838  		// If likely==-1, then we don't have enough information to decide
  3839  		// whether the first branch is likely or not. So we pass 0 for
  3840  		// the likeliness of the first branch.
  3841  		// TODO: have the frontend give us branch prediction hints for
  3842  		// OANDAND and OOROR nodes (if it ever has such info).
  3843  	case ir.OOROR:
  3844  		cond := cond.(*ir.LogicalExpr)
  3845  		mid := s.f.NewBlock(ssa.BlockPlain)
  3846  		s.stmtList(cond.Init())
  3847  		s.condBranch(cond.X, yes, mid, min8(likely, 0))
  3848  		s.startBlock(mid)
  3849  		s.condBranch(cond.Y, yes, no, likely)
  3850  		return
  3851  		// Note: if likely==-1, then both recursive calls pass -1.
  3852  		// If likely==1, then we don't have enough info to decide
  3853  		// the likelihood of the first branch.
  3854  	case ir.ONOT:
  3855  		cond := cond.(*ir.UnaryExpr)
  3856  		s.stmtList(cond.Init())
  3857  		s.condBranch(cond.X, no, yes, -likely)
  3858  		return
  3859  	case ir.OCONVNOP:
  3860  		cond := cond.(*ir.ConvExpr)
  3861  		s.stmtList(cond.Init())
  3862  		s.condBranch(cond.X, yes, no, likely)
  3863  		return
  3864  	}
  3865  	c := s.expr(cond)
  3866  	b := s.endBlock()
  3867  	b.Kind = ssa.BlockIf
  3868  	b.SetControl(c)
  3869  	b.Likely = ssa.BranchPrediction(likely) // gc and ssa both use -1/0/+1 for likeliness
  3870  	b.AddEdgeTo(yes)
  3871  	b.AddEdgeTo(no)
  3872  }
  3873  
  3874  type skipMask uint8
  3875  
  3876  const (
  3877  	skipPtr skipMask = 1 << iota
  3878  	skipLen
  3879  	skipCap
  3880  )
  3881  
  3882  // assign does left = right.
  3883  // Right has already been evaluated to ssa, left has not.
  3884  // If deref is true, then we do left = *right instead (and right has already been nil-checked).
  3885  // If deref is true and right == nil, just do left = 0.
  3886  // skip indicates assignments (at the top level) that can be avoided.
  3887  // mayOverlap indicates whether left&right might partially overlap in memory. Default is false.
  3888  func (s *state) assign(left ir.Node, right *ssa.Value, deref bool, skip skipMask) {
  3889  	s.assignWhichMayOverlap(left, right, deref, skip, false)
  3890  }
  3891  func (s *state) assignWhichMayOverlap(left ir.Node, right *ssa.Value, deref bool, skip skipMask, mayOverlap bool) {
  3892  	if left.Op() == ir.ONAME && ir.IsBlank(left) {
  3893  		return
  3894  	}
  3895  	t := left.Type()
  3896  	types.CalcSize(t)
  3897  	if s.canSSA(left) {
  3898  		if deref {
  3899  			s.Fatalf("can SSA LHS %v but not RHS %s", left, right)
  3900  		}
  3901  		if left.Op() == ir.ODOT {
  3902  			// We're assigning to a field of an ssa-able value.
  3903  			// We need to build a new structure with the new value for the
  3904  			// field we're assigning and the old values for the other fields.
  3905  			// For instance:
  3906  			//   type T struct {a, b, c int}
  3907  			//   var T x
  3908  			//   x.b = 5
  3909  			// For the x.b = 5 assignment we want to generate x = T{x.a, 5, x.c}
  3910  
  3911  			// Grab information about the structure type.
  3912  			left := left.(*ir.SelectorExpr)
  3913  			t := left.X.Type()
  3914  			nf := t.NumFields()
  3915  			idx := fieldIdx(left)
  3916  
  3917  			// Grab old value of structure.
  3918  			old := s.expr(left.X)
  3919  
  3920  			// Make new structure.
  3921  			new := s.newValue0(ssa.StructMakeOp(t.NumFields()), t)
  3922  
  3923  			// Add fields as args.
  3924  			for i := 0; i < nf; i++ {
  3925  				if i == idx {
  3926  					new.AddArg(right)
  3927  				} else {
  3928  					new.AddArg(s.newValue1I(ssa.OpStructSelect, t.FieldType(i), int64(i), old))
  3929  				}
  3930  			}
  3931  
  3932  			// Recursively assign the new value we've made to the base of the dot op.
  3933  			s.assign(left.X, new, false, 0)
  3934  			// TODO: do we need to update named values here?
  3935  			return
  3936  		}
  3937  		if left.Op() == ir.OINDEX && left.(*ir.IndexExpr).X.Type().IsArray() {
  3938  			left := left.(*ir.IndexExpr)
  3939  			s.pushLine(left.Pos())
  3940  			defer s.popLine()
  3941  			// We're assigning to an element of an ssa-able array.
  3942  			// a[i] = v
  3943  			t := left.X.Type()
  3944  			n := t.NumElem()
  3945  
  3946  			i := s.expr(left.Index) // index
  3947  			if n == 0 {
  3948  				// The bounds check must fail.  Might as well
  3949  				// ignore the actual index and just use zeros.
  3950  				z := s.constInt(types.Types[types.TINT], 0)
  3951  				s.boundsCheck(z, z, ssa.BoundsIndex, false)
  3952  				return
  3953  			}
  3954  			if n != 1 {
  3955  				s.Fatalf("assigning to non-1-length array")
  3956  			}
  3957  			// Rewrite to a = [1]{v}
  3958  			len := s.constInt(types.Types[types.TINT], 1)
  3959  			s.boundsCheck(i, len, ssa.BoundsIndex, false) // checks i == 0
  3960  			v := s.newValue1(ssa.OpArrayMake1, t, right)
  3961  			s.assign(left.X, v, false, 0)
  3962  			return
  3963  		}
  3964  		left := left.(*ir.Name)
  3965  		// Update variable assignment.
  3966  		s.vars[left] = right
  3967  		s.addNamedValue(left, right)
  3968  		return
  3969  	}
  3970  
  3971  	// If this assignment clobbers an entire local variable, then emit
  3972  	// OpVarDef so liveness analysis knows the variable is redefined.
  3973  	if base, ok := clobberBase(left).(*ir.Name); ok && base.OnStack() && skip == 0 && (t.HasPointers() || ssa.IsMergeCandidate(base)) {
  3974  		s.vars[memVar] = s.newValue1Apos(ssa.OpVarDef, types.TypeMem, base, s.mem(), !ir.IsAutoTmp(base))
  3975  	}
  3976  
  3977  	// Left is not ssa-able. Compute its address.
  3978  	addr := s.addr(left)
  3979  	if ir.IsReflectHeaderDataField(left) {
  3980  		// Package unsafe's documentation says storing pointers into
  3981  		// reflect.SliceHeader and reflect.StringHeader's Data fields
  3982  		// is valid, even though they have type uintptr (#19168).
  3983  		// Mark it pointer type to signal the writebarrier pass to
  3984  		// insert a write barrier.
  3985  		t = types.Types[types.TUNSAFEPTR]
  3986  	}
  3987  	if deref {
  3988  		// Treat as a mem->mem move.
  3989  		if right == nil {
  3990  			s.zero(t, addr)
  3991  		} else {
  3992  			s.moveWhichMayOverlap(t, addr, right, mayOverlap)
  3993  		}
  3994  		return
  3995  	}
  3996  	// Treat as a store.
  3997  	s.storeType(t, addr, right, skip, !ir.IsAutoTmp(left))
  3998  }
  3999  
  4000  // zeroVal returns the zero value for type t.
  4001  func (s *state) zeroVal(t *types.Type) *ssa.Value {
  4002  	switch {
  4003  	case t.IsInteger():
  4004  		switch t.Size() {
  4005  		case 1:
  4006  			return s.constInt8(t, 0)
  4007  		case 2:
  4008  			return s.constInt16(t, 0)
  4009  		case 4:
  4010  			return s.constInt32(t, 0)
  4011  		case 8:
  4012  			return s.constInt64(t, 0)
  4013  		default:
  4014  			s.Fatalf("bad sized integer type %v", t)
  4015  		}
  4016  	case t.IsFloat():
  4017  		switch t.Size() {
  4018  		case 4:
  4019  			return s.constFloat32(t, 0)
  4020  		case 8:
  4021  			return s.constFloat64(t, 0)
  4022  		default:
  4023  			s.Fatalf("bad sized float type %v", t)
  4024  		}
  4025  	case t.IsComplex():
  4026  		switch t.Size() {
  4027  		case 8:
  4028  			z := s.constFloat32(types.Types[types.TFLOAT32], 0)
  4029  			return s.entryNewValue2(ssa.OpComplexMake, t, z, z)
  4030  		case 16:
  4031  			z := s.constFloat64(types.Types[types.TFLOAT64], 0)
  4032  			return s.entryNewValue2(ssa.OpComplexMake, t, z, z)
  4033  		default:
  4034  			s.Fatalf("bad sized complex type %v", t)
  4035  		}
  4036  
  4037  	case t.IsString():
  4038  		return s.constEmptyString(t)
  4039  	case t.IsPtrShaped():
  4040  		return s.constNil(t)
  4041  	case t.IsBoolean():
  4042  		return s.constBool(false)
  4043  	case t.IsInterface():
  4044  		return s.constInterface(t)
  4045  	case t.IsSlice():
  4046  		return s.constSlice(t)
  4047  	case t.IsStruct():
  4048  		n := t.NumFields()
  4049  		v := s.entryNewValue0(ssa.StructMakeOp(t.NumFields()), t)
  4050  		for i := 0; i < n; i++ {
  4051  			v.AddArg(s.zeroVal(t.FieldType(i)))
  4052  		}
  4053  		return v
  4054  	case t.IsArray():
  4055  		switch t.NumElem() {
  4056  		case 0:
  4057  			return s.entryNewValue0(ssa.OpArrayMake0, t)
  4058  		case 1:
  4059  			return s.entryNewValue1(ssa.OpArrayMake1, t, s.zeroVal(t.Elem()))
  4060  		}
  4061  	}
  4062  	s.Fatalf("zero for type %v not implemented", t)
  4063  	return nil
  4064  }
  4065  
  4066  type callKind int8
  4067  
  4068  const (
  4069  	callNormal callKind = iota
  4070  	callDefer
  4071  	callDeferStack
  4072  	callGo
  4073  	callTail
  4074  )
  4075  
  4076  type sfRtCallDef struct {
  4077  	rtfn  *obj.LSym
  4078  	rtype types.Kind
  4079  }
  4080  
  4081  var softFloatOps map[ssa.Op]sfRtCallDef
  4082  
  4083  func softfloatInit() {
  4084  	// Some of these operations get transformed by sfcall.
  4085  	softFloatOps = map[ssa.Op]sfRtCallDef{
  4086  		ssa.OpAdd32F: {typecheck.LookupRuntimeFunc("fadd32"), types.TFLOAT32},
  4087  		ssa.OpAdd64F: {typecheck.LookupRuntimeFunc("fadd64"), types.TFLOAT64},
  4088  		ssa.OpSub32F: {typecheck.LookupRuntimeFunc("fadd32"), types.TFLOAT32},
  4089  		ssa.OpSub64F: {typecheck.LookupRuntimeFunc("fadd64"), types.TFLOAT64},
  4090  		ssa.OpMul32F: {typecheck.LookupRuntimeFunc("fmul32"), types.TFLOAT32},
  4091  		ssa.OpMul64F: {typecheck.LookupRuntimeFunc("fmul64"), types.TFLOAT64},
  4092  		ssa.OpDiv32F: {typecheck.LookupRuntimeFunc("fdiv32"), types.TFLOAT32},
  4093  		ssa.OpDiv64F: {typecheck.LookupRuntimeFunc("fdiv64"), types.TFLOAT64},
  4094  
  4095  		ssa.OpEq64F:   {typecheck.LookupRuntimeFunc("feq64"), types.TBOOL},
  4096  		ssa.OpEq32F:   {typecheck.LookupRuntimeFunc("feq32"), types.TBOOL},
  4097  		ssa.OpNeq64F:  {typecheck.LookupRuntimeFunc("feq64"), types.TBOOL},
  4098  		ssa.OpNeq32F:  {typecheck.LookupRuntimeFunc("feq32"), types.TBOOL},
  4099  		ssa.OpLess64F: {typecheck.LookupRuntimeFunc("fgt64"), types.TBOOL},
  4100  		ssa.OpLess32F: {typecheck.LookupRuntimeFunc("fgt32"), types.TBOOL},
  4101  		ssa.OpLeq64F:  {typecheck.LookupRuntimeFunc("fge64"), types.TBOOL},
  4102  		ssa.OpLeq32F:  {typecheck.LookupRuntimeFunc("fge32"), types.TBOOL},
  4103  
  4104  		ssa.OpCvt32to32F:  {typecheck.LookupRuntimeFunc("fint32to32"), types.TFLOAT32},
  4105  		ssa.OpCvt32Fto32:  {typecheck.LookupRuntimeFunc("f32toint32"), types.TINT32},
  4106  		ssa.OpCvt64to32F:  {typecheck.LookupRuntimeFunc("fint64to32"), types.TFLOAT32},
  4107  		ssa.OpCvt32Fto64:  {typecheck.LookupRuntimeFunc("f32toint64"), types.TINT64},
  4108  		ssa.OpCvt64Uto32F: {typecheck.LookupRuntimeFunc("fuint64to32"), types.TFLOAT32},
  4109  		ssa.OpCvt32Fto64U: {typecheck.LookupRuntimeFunc("f32touint64"), types.TUINT64},
  4110  		ssa.OpCvt32to64F:  {typecheck.LookupRuntimeFunc("fint32to64"), types.TFLOAT64},
  4111  		ssa.OpCvt64Fto32:  {typecheck.LookupRuntimeFunc("f64toint32"), types.TINT32},
  4112  		ssa.OpCvt64to64F:  {typecheck.LookupRuntimeFunc("fint64to64"), types.TFLOAT64},
  4113  		ssa.OpCvt64Fto64:  {typecheck.LookupRuntimeFunc("f64toint64"), types.TINT64},
  4114  		ssa.OpCvt64Uto64F: {typecheck.LookupRuntimeFunc("fuint64to64"), types.TFLOAT64},
  4115  		ssa.OpCvt64Fto64U: {typecheck.LookupRuntimeFunc("f64touint64"), types.TUINT64},
  4116  		ssa.OpCvt32Fto64F: {typecheck.LookupRuntimeFunc("f32to64"), types.TFLOAT64},
  4117  		ssa.OpCvt64Fto32F: {typecheck.LookupRuntimeFunc("f64to32"), types.TFLOAT32},
  4118  	}
  4119  }
  4120  
  4121  // TODO: do not emit sfcall if operation can be optimized to constant in later
  4122  // opt phase
  4123  func (s *state) sfcall(op ssa.Op, args ...*ssa.Value) (*ssa.Value, bool) {
  4124  	f2i := func(t *types.Type) *types.Type {
  4125  		switch t.Kind() {
  4126  		case types.TFLOAT32:
  4127  			return types.Types[types.TUINT32]
  4128  		case types.TFLOAT64:
  4129  			return types.Types[types.TUINT64]
  4130  		}
  4131  		return t
  4132  	}
  4133  
  4134  	if callDef, ok := softFloatOps[op]; ok {
  4135  		switch op {
  4136  		case ssa.OpLess32F,
  4137  			ssa.OpLess64F,
  4138  			ssa.OpLeq32F,
  4139  			ssa.OpLeq64F:
  4140  			args[0], args[1] = args[1], args[0]
  4141  		case ssa.OpSub32F,
  4142  			ssa.OpSub64F:
  4143  			args[1] = s.newValue1(s.ssaOp(ir.ONEG, types.Types[callDef.rtype]), args[1].Type, args[1])
  4144  		}
  4145  
  4146  		// runtime functions take uints for floats and returns uints.
  4147  		// Convert to uints so we use the right calling convention.
  4148  		for i, a := range args {
  4149  			if a.Type.IsFloat() {
  4150  				args[i] = s.newValue1(ssa.OpCopy, f2i(a.Type), a)
  4151  			}
  4152  		}
  4153  
  4154  		rt := types.Types[callDef.rtype]
  4155  		result := s.rtcall(callDef.rtfn, true, []*types.Type{f2i(rt)}, args...)[0]
  4156  		if rt.IsFloat() {
  4157  			result = s.newValue1(ssa.OpCopy, rt, result)
  4158  		}
  4159  		if op == ssa.OpNeq32F || op == ssa.OpNeq64F {
  4160  			result = s.newValue1(ssa.OpNot, result.Type, result)
  4161  		}
  4162  		return result, true
  4163  	}
  4164  	return nil, false
  4165  }
  4166  
  4167  var intrinsics map[intrinsicKey]intrinsicBuilder
  4168  
  4169  // An intrinsicBuilder converts a call node n into an ssa value that
  4170  // implements that call as an intrinsic. args is a list of arguments to the func.
  4171  type intrinsicBuilder func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value
  4172  
  4173  type intrinsicKey struct {
  4174  	arch *sys.Arch
  4175  	pkg  string
  4176  	fn   string
  4177  }
  4178  
  4179  func InitTables() {
  4180  	intrinsics = map[intrinsicKey]intrinsicBuilder{}
  4181  
  4182  	var all []*sys.Arch
  4183  	var p4 []*sys.Arch
  4184  	var p8 []*sys.Arch
  4185  	var lwatomics []*sys.Arch
  4186  	for _, a := range &sys.Archs {
  4187  		all = append(all, a)
  4188  		if a.PtrSize == 4 {
  4189  			p4 = append(p4, a)
  4190  		} else {
  4191  			p8 = append(p8, a)
  4192  		}
  4193  		if a.Family != sys.PPC64 {
  4194  			lwatomics = append(lwatomics, a)
  4195  		}
  4196  	}
  4197  
  4198  	// add adds the intrinsic b for pkg.fn for the given list of architectures.
  4199  	add := func(pkg, fn string, b intrinsicBuilder, archs ...*sys.Arch) {
  4200  		for _, a := range archs {
  4201  			intrinsics[intrinsicKey{a, pkg, fn}] = b
  4202  		}
  4203  	}
  4204  	// addF does the same as add but operates on architecture families.
  4205  	addF := func(pkg, fn string, b intrinsicBuilder, archFamilies ...sys.ArchFamily) {
  4206  		m := 0
  4207  		for _, f := range archFamilies {
  4208  			if f >= 32 {
  4209  				panic("too many architecture families")
  4210  			}
  4211  			m |= 1 << uint(f)
  4212  		}
  4213  		for _, a := range all {
  4214  			if m>>uint(a.Family)&1 != 0 {
  4215  				intrinsics[intrinsicKey{a, pkg, fn}] = b
  4216  			}
  4217  		}
  4218  	}
  4219  	// alias defines pkg.fn = pkg2.fn2 for all architectures in archs for which pkg2.fn2 exists.
  4220  	alias := func(pkg, fn, pkg2, fn2 string, archs ...*sys.Arch) {
  4221  		aliased := false
  4222  		for _, a := range archs {
  4223  			if b, ok := intrinsics[intrinsicKey{a, pkg2, fn2}]; ok {
  4224  				intrinsics[intrinsicKey{a, pkg, fn}] = b
  4225  				aliased = true
  4226  			}
  4227  		}
  4228  		if !aliased {
  4229  			panic(fmt.Sprintf("attempted to alias undefined intrinsic: %s.%s", pkg, fn))
  4230  		}
  4231  	}
  4232  
  4233  	/******** runtime ********/
  4234  	if !base.Flag.Cfg.Instrumenting {
  4235  		add("runtime", "slicebytetostringtmp",
  4236  			func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4237  				// Compiler frontend optimizations emit OBYTES2STRTMP nodes
  4238  				// for the backend instead of slicebytetostringtmp calls
  4239  				// when not instrumenting.
  4240  				return s.newValue2(ssa.OpStringMake, n.Type(), args[0], args[1])
  4241  			},
  4242  			all...)
  4243  	}
  4244  	addF("runtime/internal/math", "MulUintptr",
  4245  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4246  			if s.config.PtrSize == 4 {
  4247  				return s.newValue2(ssa.OpMul32uover, types.NewTuple(types.Types[types.TUINT], types.Types[types.TUINT]), args[0], args[1])
  4248  			}
  4249  			return s.newValue2(ssa.OpMul64uover, types.NewTuple(types.Types[types.TUINT], types.Types[types.TUINT]), args[0], args[1])
  4250  		},
  4251  		sys.AMD64, sys.I386, sys.Loong64, sys.MIPS64, sys.RISCV64, sys.ARM64)
  4252  	add("runtime", "KeepAlive",
  4253  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4254  			data := s.newValue1(ssa.OpIData, s.f.Config.Types.BytePtr, args[0])
  4255  			s.vars[memVar] = s.newValue2(ssa.OpKeepAlive, types.TypeMem, data, s.mem())
  4256  			return nil
  4257  		},
  4258  		all...)
  4259  	add("runtime", "getclosureptr",
  4260  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4261  			return s.newValue0(ssa.OpGetClosurePtr, s.f.Config.Types.Uintptr)
  4262  		},
  4263  		all...)
  4264  
  4265  	add("runtime", "getcallerpc",
  4266  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4267  			return s.newValue0(ssa.OpGetCallerPC, s.f.Config.Types.Uintptr)
  4268  		},
  4269  		all...)
  4270  
  4271  	add("runtime", "getcallersp",
  4272  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4273  			return s.newValue1(ssa.OpGetCallerSP, s.f.Config.Types.Uintptr, s.mem())
  4274  		},
  4275  		all...)
  4276  
  4277  	addF("runtime", "publicationBarrier",
  4278  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4279  			s.vars[memVar] = s.newValue1(ssa.OpPubBarrier, types.TypeMem, s.mem())
  4280  			return nil
  4281  		},
  4282  		sys.ARM64, sys.PPC64, sys.RISCV64)
  4283  
  4284  	brev_arch := []sys.ArchFamily{sys.AMD64, sys.I386, sys.ARM64, sys.ARM, sys.S390X}
  4285  	if buildcfg.GOPPC64 >= 10 {
  4286  		// Use only on Power10 as the new byte reverse instructions that Power10 provide
  4287  		// make it worthwhile as an intrinsic
  4288  		brev_arch = append(brev_arch, sys.PPC64)
  4289  	}
  4290  	/******** runtime/internal/sys ********/
  4291  	addF("runtime/internal/sys", "Bswap32",
  4292  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4293  			return s.newValue1(ssa.OpBswap32, types.Types[types.TUINT32], args[0])
  4294  		},
  4295  		brev_arch...)
  4296  	addF("runtime/internal/sys", "Bswap64",
  4297  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4298  			return s.newValue1(ssa.OpBswap64, types.Types[types.TUINT64], args[0])
  4299  		},
  4300  		brev_arch...)
  4301  
  4302  	/****** Prefetch ******/
  4303  	makePrefetchFunc := func(op ssa.Op) func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4304  		return func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4305  			s.vars[memVar] = s.newValue2(op, types.TypeMem, args[0], s.mem())
  4306  			return nil
  4307  		}
  4308  	}
  4309  
  4310  	// Make Prefetch intrinsics for supported platforms
  4311  	// On the unsupported platforms stub function will be eliminated
  4312  	addF("runtime/internal/sys", "Prefetch", makePrefetchFunc(ssa.OpPrefetchCache),
  4313  		sys.AMD64, sys.ARM64, sys.PPC64)
  4314  	addF("runtime/internal/sys", "PrefetchStreamed", makePrefetchFunc(ssa.OpPrefetchCacheStreamed),
  4315  		sys.AMD64, sys.ARM64, sys.PPC64)
  4316  
  4317  	/******** internal/runtime/atomic ********/
  4318  	addF("internal/runtime/atomic", "Load",
  4319  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4320  			v := s.newValue2(ssa.OpAtomicLoad32, types.NewTuple(types.Types[types.TUINT32], types.TypeMem), args[0], s.mem())
  4321  			s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
  4322  			return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT32], v)
  4323  		},
  4324  		sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
  4325  	addF("internal/runtime/atomic", "Load8",
  4326  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4327  			v := s.newValue2(ssa.OpAtomicLoad8, types.NewTuple(types.Types[types.TUINT8], types.TypeMem), args[0], s.mem())
  4328  			s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
  4329  			return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT8], v)
  4330  		},
  4331  		sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
  4332  	addF("internal/runtime/atomic", "Load64",
  4333  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4334  			v := s.newValue2(ssa.OpAtomicLoad64, types.NewTuple(types.Types[types.TUINT64], types.TypeMem), args[0], s.mem())
  4335  			s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
  4336  			return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT64], v)
  4337  		},
  4338  		sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
  4339  	addF("internal/runtime/atomic", "LoadAcq",
  4340  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4341  			v := s.newValue2(ssa.OpAtomicLoadAcq32, types.NewTuple(types.Types[types.TUINT32], types.TypeMem), args[0], s.mem())
  4342  			s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
  4343  			return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT32], v)
  4344  		},
  4345  		sys.PPC64, sys.S390X)
  4346  	addF("internal/runtime/atomic", "LoadAcq64",
  4347  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4348  			v := s.newValue2(ssa.OpAtomicLoadAcq64, types.NewTuple(types.Types[types.TUINT64], types.TypeMem), args[0], s.mem())
  4349  			s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
  4350  			return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT64], v)
  4351  		},
  4352  		sys.PPC64)
  4353  	addF("internal/runtime/atomic", "Loadp",
  4354  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4355  			v := s.newValue2(ssa.OpAtomicLoadPtr, types.NewTuple(s.f.Config.Types.BytePtr, types.TypeMem), args[0], s.mem())
  4356  			s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
  4357  			return s.newValue1(ssa.OpSelect0, s.f.Config.Types.BytePtr, v)
  4358  		},
  4359  		sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
  4360  
  4361  	addF("internal/runtime/atomic", "Store",
  4362  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4363  			s.vars[memVar] = s.newValue3(ssa.OpAtomicStore32, types.TypeMem, args[0], args[1], s.mem())
  4364  			return nil
  4365  		},
  4366  		sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
  4367  	addF("internal/runtime/atomic", "Store8",
  4368  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4369  			s.vars[memVar] = s.newValue3(ssa.OpAtomicStore8, types.TypeMem, args[0], args[1], s.mem())
  4370  			return nil
  4371  		},
  4372  		sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
  4373  	addF("internal/runtime/atomic", "Store64",
  4374  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4375  			s.vars[memVar] = s.newValue3(ssa.OpAtomicStore64, types.TypeMem, args[0], args[1], s.mem())
  4376  			return nil
  4377  		},
  4378  		sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
  4379  	addF("internal/runtime/atomic", "StorepNoWB",
  4380  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4381  			s.vars[memVar] = s.newValue3(ssa.OpAtomicStorePtrNoWB, types.TypeMem, args[0], args[1], s.mem())
  4382  			return nil
  4383  		},
  4384  		sys.AMD64, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.RISCV64, sys.S390X)
  4385  	addF("internal/runtime/atomic", "StoreRel",
  4386  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4387  			s.vars[memVar] = s.newValue3(ssa.OpAtomicStoreRel32, types.TypeMem, args[0], args[1], s.mem())
  4388  			return nil
  4389  		},
  4390  		sys.PPC64, sys.S390X)
  4391  	addF("internal/runtime/atomic", "StoreRel64",
  4392  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4393  			s.vars[memVar] = s.newValue3(ssa.OpAtomicStoreRel64, types.TypeMem, args[0], args[1], s.mem())
  4394  			return nil
  4395  		},
  4396  		sys.PPC64)
  4397  
  4398  	addF("internal/runtime/atomic", "Xchg",
  4399  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4400  			v := s.newValue3(ssa.OpAtomicExchange32, types.NewTuple(types.Types[types.TUINT32], types.TypeMem), args[0], args[1], s.mem())
  4401  			s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
  4402  			return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT32], v)
  4403  		},
  4404  		sys.AMD64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
  4405  	addF("internal/runtime/atomic", "Xchg64",
  4406  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4407  			v := s.newValue3(ssa.OpAtomicExchange64, types.NewTuple(types.Types[types.TUINT64], types.TypeMem), args[0], args[1], s.mem())
  4408  			s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
  4409  			return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT64], v)
  4410  		},
  4411  		sys.AMD64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
  4412  
  4413  	type atomicOpEmitter func(s *state, n *ir.CallExpr, args []*ssa.Value, op ssa.Op, typ types.Kind)
  4414  
  4415  	makeAtomicGuardedIntrinsicARM64 := func(op0, op1 ssa.Op, typ types.Kind, emit atomicOpEmitter) intrinsicBuilder {
  4416  
  4417  		return func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4418  			if buildcfg.GOARM64.LSE {
  4419  				emit(s, n, args, op1, typ)
  4420  			} else {
  4421  				// Target Atomic feature is identified by dynamic detection
  4422  				addr := s.entryNewValue1A(ssa.OpAddr, types.Types[types.TBOOL].PtrTo(), ir.Syms.ARM64HasATOMICS, s.sb)
  4423  				v := s.load(types.Types[types.TBOOL], addr)
  4424  				b := s.endBlock()
  4425  				b.Kind = ssa.BlockIf
  4426  				b.SetControl(v)
  4427  				bTrue := s.f.NewBlock(ssa.BlockPlain)
  4428  				bFalse := s.f.NewBlock(ssa.BlockPlain)
  4429  				bEnd := s.f.NewBlock(ssa.BlockPlain)
  4430  				b.AddEdgeTo(bTrue)
  4431  				b.AddEdgeTo(bFalse)
  4432  				b.Likely = ssa.BranchLikely
  4433  
  4434  				// We have atomic instructions - use it directly.
  4435  				s.startBlock(bTrue)
  4436  				emit(s, n, args, op1, typ)
  4437  				s.endBlock().AddEdgeTo(bEnd)
  4438  
  4439  				// Use original instruction sequence.
  4440  				s.startBlock(bFalse)
  4441  				emit(s, n, args, op0, typ)
  4442  				s.endBlock().AddEdgeTo(bEnd)
  4443  
  4444  				// Merge results.
  4445  				s.startBlock(bEnd)
  4446  			}
  4447  			if typ == types.TNIL {
  4448  				return nil
  4449  			} else {
  4450  				return s.variable(n, types.Types[typ])
  4451  			}
  4452  		}
  4453  	}
  4454  
  4455  	atomicEmitterARM64 := func(s *state, n *ir.CallExpr, args []*ssa.Value, op ssa.Op, typ types.Kind) {
  4456  		v := s.newValue3(op, types.NewTuple(types.Types[typ], types.TypeMem), args[0], args[1], s.mem())
  4457  		s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
  4458  		s.vars[n] = s.newValue1(ssa.OpSelect0, types.Types[typ], v)
  4459  	}
  4460  	addF("internal/runtime/atomic", "Xchg",
  4461  		makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicExchange32, ssa.OpAtomicExchange32Variant, types.TUINT32, atomicEmitterARM64),
  4462  		sys.ARM64)
  4463  	addF("internal/runtime/atomic", "Xchg64",
  4464  		makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicExchange64, ssa.OpAtomicExchange64Variant, types.TUINT64, atomicEmitterARM64),
  4465  		sys.ARM64)
  4466  
  4467  	addF("internal/runtime/atomic", "Xadd",
  4468  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4469  			v := s.newValue3(ssa.OpAtomicAdd32, types.NewTuple(types.Types[types.TUINT32], types.TypeMem), args[0], args[1], s.mem())
  4470  			s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
  4471  			return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT32], v)
  4472  		},
  4473  		sys.AMD64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
  4474  	addF("internal/runtime/atomic", "Xadd64",
  4475  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4476  			v := s.newValue3(ssa.OpAtomicAdd64, types.NewTuple(types.Types[types.TUINT64], types.TypeMem), args[0], args[1], s.mem())
  4477  			s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
  4478  			return s.newValue1(ssa.OpSelect0, types.Types[types.TUINT64], v)
  4479  		},
  4480  		sys.AMD64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
  4481  
  4482  	addF("internal/runtime/atomic", "Xadd",
  4483  		makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAdd32, ssa.OpAtomicAdd32Variant, types.TUINT32, atomicEmitterARM64),
  4484  		sys.ARM64)
  4485  	addF("internal/runtime/atomic", "Xadd64",
  4486  		makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAdd64, ssa.OpAtomicAdd64Variant, types.TUINT64, atomicEmitterARM64),
  4487  		sys.ARM64)
  4488  
  4489  	addF("internal/runtime/atomic", "Cas",
  4490  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4491  			v := s.newValue4(ssa.OpAtomicCompareAndSwap32, types.NewTuple(types.Types[types.TBOOL], types.TypeMem), args[0], args[1], args[2], s.mem())
  4492  			s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
  4493  			return s.newValue1(ssa.OpSelect0, types.Types[types.TBOOL], v)
  4494  		},
  4495  		sys.AMD64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
  4496  	addF("internal/runtime/atomic", "Cas64",
  4497  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4498  			v := s.newValue4(ssa.OpAtomicCompareAndSwap64, types.NewTuple(types.Types[types.TBOOL], types.TypeMem), args[0], args[1], args[2], s.mem())
  4499  			s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
  4500  			return s.newValue1(ssa.OpSelect0, types.Types[types.TBOOL], v)
  4501  		},
  4502  		sys.AMD64, sys.Loong64, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
  4503  	addF("internal/runtime/atomic", "CasRel",
  4504  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4505  			v := s.newValue4(ssa.OpAtomicCompareAndSwap32, types.NewTuple(types.Types[types.TBOOL], types.TypeMem), args[0], args[1], args[2], s.mem())
  4506  			s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
  4507  			return s.newValue1(ssa.OpSelect0, types.Types[types.TBOOL], v)
  4508  		},
  4509  		sys.PPC64)
  4510  
  4511  	atomicCasEmitterARM64 := func(s *state, n *ir.CallExpr, args []*ssa.Value, op ssa.Op, typ types.Kind) {
  4512  		v := s.newValue4(op, types.NewTuple(types.Types[types.TBOOL], types.TypeMem), args[0], args[1], args[2], s.mem())
  4513  		s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
  4514  		s.vars[n] = s.newValue1(ssa.OpSelect0, types.Types[typ], v)
  4515  	}
  4516  
  4517  	addF("internal/runtime/atomic", "Cas",
  4518  		makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicCompareAndSwap32, ssa.OpAtomicCompareAndSwap32Variant, types.TBOOL, atomicCasEmitterARM64),
  4519  		sys.ARM64)
  4520  	addF("internal/runtime/atomic", "Cas64",
  4521  		makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicCompareAndSwap64, ssa.OpAtomicCompareAndSwap64Variant, types.TBOOL, atomicCasEmitterARM64),
  4522  		sys.ARM64)
  4523  
  4524  	addF("internal/runtime/atomic", "And8",
  4525  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4526  			s.vars[memVar] = s.newValue3(ssa.OpAtomicAnd8, types.TypeMem, args[0], args[1], s.mem())
  4527  			return nil
  4528  		},
  4529  		sys.AMD64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
  4530  	addF("internal/runtime/atomic", "And",
  4531  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4532  			s.vars[memVar] = s.newValue3(ssa.OpAtomicAnd32, types.TypeMem, args[0], args[1], s.mem())
  4533  			return nil
  4534  		},
  4535  		sys.AMD64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
  4536  	addF("internal/runtime/atomic", "Or8",
  4537  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4538  			s.vars[memVar] = s.newValue3(ssa.OpAtomicOr8, types.TypeMem, args[0], args[1], s.mem())
  4539  			return nil
  4540  		},
  4541  		sys.AMD64, sys.ARM64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
  4542  	addF("internal/runtime/atomic", "Or",
  4543  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4544  			s.vars[memVar] = s.newValue3(ssa.OpAtomicOr32, types.TypeMem, args[0], args[1], s.mem())
  4545  			return nil
  4546  		},
  4547  		sys.AMD64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X)
  4548  
  4549  	addF("internal/runtime/atomic", "And8",
  4550  		makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAnd8, ssa.OpAtomicAnd8Variant, types.TUINT8, atomicEmitterARM64),
  4551  		sys.ARM64)
  4552  	addF("internal/runtime/atomic", "Or8",
  4553  		makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicOr8, ssa.OpAtomicOr8Variant, types.TUINT8, atomicEmitterARM64),
  4554  		sys.ARM64)
  4555  	addF("internal/runtime/atomic", "And64",
  4556  		makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAnd64, ssa.OpAtomicAnd64Variant, types.TUINT64, atomicEmitterARM64),
  4557  		sys.ARM64)
  4558  	addF("internal/runtime/atomic", "And32",
  4559  		makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAnd32, ssa.OpAtomicAnd32Variant, types.TUINT32, atomicEmitterARM64),
  4560  		sys.ARM64)
  4561  	addF("internal/runtime/atomic", "And",
  4562  		makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicAnd32, ssa.OpAtomicAnd32Variant, types.TUINT32, atomicEmitterARM64),
  4563  		sys.ARM64)
  4564  	addF("internal/runtime/atomic", "Or64",
  4565  		makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicOr64, ssa.OpAtomicOr64Variant, types.TUINT64, atomicEmitterARM64),
  4566  		sys.ARM64)
  4567  	addF("internal/runtime/atomic", "Or32",
  4568  		makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicOr32, ssa.OpAtomicOr32Variant, types.TUINT32, atomicEmitterARM64),
  4569  		sys.ARM64)
  4570  	addF("internal/runtime/atomic", "Or",
  4571  		makeAtomicGuardedIntrinsicARM64(ssa.OpAtomicOr32, ssa.OpAtomicOr32Variant, types.TUINT32, atomicEmitterARM64),
  4572  		sys.ARM64)
  4573  
  4574  	// Aliases for atomic load operations
  4575  	alias("internal/runtime/atomic", "Loadint32", "internal/runtime/atomic", "Load", all...)
  4576  	alias("internal/runtime/atomic", "Loadint64", "internal/runtime/atomic", "Load64", all...)
  4577  	alias("internal/runtime/atomic", "Loaduintptr", "internal/runtime/atomic", "Load", p4...)
  4578  	alias("internal/runtime/atomic", "Loaduintptr", "internal/runtime/atomic", "Load64", p8...)
  4579  	alias("internal/runtime/atomic", "Loaduint", "internal/runtime/atomic", "Load", p4...)
  4580  	alias("internal/runtime/atomic", "Loaduint", "internal/runtime/atomic", "Load64", p8...)
  4581  	alias("internal/runtime/atomic", "LoadAcq", "internal/runtime/atomic", "Load", lwatomics...)
  4582  	alias("internal/runtime/atomic", "LoadAcq64", "internal/runtime/atomic", "Load64", lwatomics...)
  4583  	alias("internal/runtime/atomic", "LoadAcquintptr", "internal/runtime/atomic", "LoadAcq", p4...)
  4584  	alias("sync", "runtime_LoadAcquintptr", "internal/runtime/atomic", "LoadAcq", p4...) // linknamed
  4585  	alias("internal/runtime/atomic", "LoadAcquintptr", "internal/runtime/atomic", "LoadAcq64", p8...)
  4586  	alias("sync", "runtime_LoadAcquintptr", "internal/runtime/atomic", "LoadAcq64", p8...) // linknamed
  4587  
  4588  	// Aliases for atomic store operations
  4589  	alias("internal/runtime/atomic", "Storeint32", "internal/runtime/atomic", "Store", all...)
  4590  	alias("internal/runtime/atomic", "Storeint64", "internal/runtime/atomic", "Store64", all...)
  4591  	alias("internal/runtime/atomic", "Storeuintptr", "internal/runtime/atomic", "Store", p4...)
  4592  	alias("internal/runtime/atomic", "Storeuintptr", "internal/runtime/atomic", "Store64", p8...)
  4593  	alias("internal/runtime/atomic", "StoreRel", "internal/runtime/atomic", "Store", lwatomics...)
  4594  	alias("internal/runtime/atomic", "StoreRel64", "internal/runtime/atomic", "Store64", lwatomics...)
  4595  	alias("internal/runtime/atomic", "StoreReluintptr", "internal/runtime/atomic", "StoreRel", p4...)
  4596  	alias("sync", "runtime_StoreReluintptr", "internal/runtime/atomic", "StoreRel", p4...) // linknamed
  4597  	alias("internal/runtime/atomic", "StoreReluintptr", "internal/runtime/atomic", "StoreRel64", p8...)
  4598  	alias("sync", "runtime_StoreReluintptr", "internal/runtime/atomic", "StoreRel64", p8...) // linknamed
  4599  
  4600  	// Aliases for atomic swap operations
  4601  	alias("internal/runtime/atomic", "Xchgint32", "internal/runtime/atomic", "Xchg", all...)
  4602  	alias("internal/runtime/atomic", "Xchgint64", "internal/runtime/atomic", "Xchg64", all...)
  4603  	alias("internal/runtime/atomic", "Xchguintptr", "internal/runtime/atomic", "Xchg", p4...)
  4604  	alias("internal/runtime/atomic", "Xchguintptr", "internal/runtime/atomic", "Xchg64", p8...)
  4605  
  4606  	// Aliases for atomic add operations
  4607  	alias("internal/runtime/atomic", "Xaddint32", "internal/runtime/atomic", "Xadd", all...)
  4608  	alias("internal/runtime/atomic", "Xaddint64", "internal/runtime/atomic", "Xadd64", all...)
  4609  	alias("internal/runtime/atomic", "Xadduintptr", "internal/runtime/atomic", "Xadd", p4...)
  4610  	alias("internal/runtime/atomic", "Xadduintptr", "internal/runtime/atomic", "Xadd64", p8...)
  4611  
  4612  	// Aliases for atomic CAS operations
  4613  	alias("internal/runtime/atomic", "Casint32", "internal/runtime/atomic", "Cas", all...)
  4614  	alias("internal/runtime/atomic", "Casint64", "internal/runtime/atomic", "Cas64", all...)
  4615  	alias("internal/runtime/atomic", "Casuintptr", "internal/runtime/atomic", "Cas", p4...)
  4616  	alias("internal/runtime/atomic", "Casuintptr", "internal/runtime/atomic", "Cas64", p8...)
  4617  	alias("internal/runtime/atomic", "Casp1", "internal/runtime/atomic", "Cas", p4...)
  4618  	alias("internal/runtime/atomic", "Casp1", "internal/runtime/atomic", "Cas64", p8...)
  4619  	alias("internal/runtime/atomic", "CasRel", "internal/runtime/atomic", "Cas", lwatomics...)
  4620  
  4621  	// Aliases for atomic And/Or operations
  4622  	alias("internal/runtime/atomic", "Anduintptr", "internal/runtime/atomic", "And64", sys.ArchARM64)
  4623  	alias("internal/runtime/atomic", "Oruintptr", "internal/runtime/atomic", "Or64", sys.ArchARM64)
  4624  
  4625  	/******** math ********/
  4626  	addF("math", "sqrt",
  4627  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4628  			return s.newValue1(ssa.OpSqrt, types.Types[types.TFLOAT64], args[0])
  4629  		},
  4630  		sys.I386, sys.AMD64, sys.ARM, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64, sys.PPC64, sys.RISCV64, sys.S390X, sys.Wasm)
  4631  	addF("math", "Trunc",
  4632  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4633  			return s.newValue1(ssa.OpTrunc, types.Types[types.TFLOAT64], args[0])
  4634  		},
  4635  		sys.ARM64, sys.PPC64, sys.S390X, sys.Wasm)
  4636  	addF("math", "Ceil",
  4637  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4638  			return s.newValue1(ssa.OpCeil, types.Types[types.TFLOAT64], args[0])
  4639  		},
  4640  		sys.ARM64, sys.PPC64, sys.S390X, sys.Wasm)
  4641  	addF("math", "Floor",
  4642  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4643  			return s.newValue1(ssa.OpFloor, types.Types[types.TFLOAT64], args[0])
  4644  		},
  4645  		sys.ARM64, sys.PPC64, sys.S390X, sys.Wasm)
  4646  	addF("math", "Round",
  4647  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4648  			return s.newValue1(ssa.OpRound, types.Types[types.TFLOAT64], args[0])
  4649  		},
  4650  		sys.ARM64, sys.PPC64, sys.S390X)
  4651  	addF("math", "RoundToEven",
  4652  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4653  			return s.newValue1(ssa.OpRoundToEven, types.Types[types.TFLOAT64], args[0])
  4654  		},
  4655  		sys.ARM64, sys.S390X, sys.Wasm)
  4656  	addF("math", "Abs",
  4657  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4658  			return s.newValue1(ssa.OpAbs, types.Types[types.TFLOAT64], args[0])
  4659  		},
  4660  		sys.ARM64, sys.ARM, sys.PPC64, sys.RISCV64, sys.Wasm, sys.MIPS, sys.MIPS64)
  4661  	addF("math", "Copysign",
  4662  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4663  			return s.newValue2(ssa.OpCopysign, types.Types[types.TFLOAT64], args[0], args[1])
  4664  		},
  4665  		sys.PPC64, sys.RISCV64, sys.Wasm)
  4666  	addF("math", "FMA",
  4667  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4668  			return s.newValue3(ssa.OpFMA, types.Types[types.TFLOAT64], args[0], args[1], args[2])
  4669  		},
  4670  		sys.ARM64, sys.PPC64, sys.RISCV64, sys.S390X)
  4671  	addF("math", "FMA",
  4672  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4673  			if !s.config.UseFMA {
  4674  				s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
  4675  				return s.variable(n, types.Types[types.TFLOAT64])
  4676  			}
  4677  
  4678  			if buildcfg.GOAMD64 >= 3 {
  4679  				return s.newValue3(ssa.OpFMA, types.Types[types.TFLOAT64], args[0], args[1], args[2])
  4680  			}
  4681  
  4682  			v := s.entryNewValue0A(ssa.OpHasCPUFeature, types.Types[types.TBOOL], ir.Syms.X86HasFMA)
  4683  			b := s.endBlock()
  4684  			b.Kind = ssa.BlockIf
  4685  			b.SetControl(v)
  4686  			bTrue := s.f.NewBlock(ssa.BlockPlain)
  4687  			bFalse := s.f.NewBlock(ssa.BlockPlain)
  4688  			bEnd := s.f.NewBlock(ssa.BlockPlain)
  4689  			b.AddEdgeTo(bTrue)
  4690  			b.AddEdgeTo(bFalse)
  4691  			b.Likely = ssa.BranchLikely // >= haswell cpus are common
  4692  
  4693  			// We have the intrinsic - use it directly.
  4694  			s.startBlock(bTrue)
  4695  			s.vars[n] = s.newValue3(ssa.OpFMA, types.Types[types.TFLOAT64], args[0], args[1], args[2])
  4696  			s.endBlock().AddEdgeTo(bEnd)
  4697  
  4698  			// Call the pure Go version.
  4699  			s.startBlock(bFalse)
  4700  			s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
  4701  			s.endBlock().AddEdgeTo(bEnd)
  4702  
  4703  			// Merge results.
  4704  			s.startBlock(bEnd)
  4705  			return s.variable(n, types.Types[types.TFLOAT64])
  4706  		},
  4707  		sys.AMD64)
  4708  	addF("math", "FMA",
  4709  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4710  			if !s.config.UseFMA {
  4711  				s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
  4712  				return s.variable(n, types.Types[types.TFLOAT64])
  4713  			}
  4714  			addr := s.entryNewValue1A(ssa.OpAddr, types.Types[types.TBOOL].PtrTo(), ir.Syms.ARMHasVFPv4, s.sb)
  4715  			v := s.load(types.Types[types.TBOOL], addr)
  4716  			b := s.endBlock()
  4717  			b.Kind = ssa.BlockIf
  4718  			b.SetControl(v)
  4719  			bTrue := s.f.NewBlock(ssa.BlockPlain)
  4720  			bFalse := s.f.NewBlock(ssa.BlockPlain)
  4721  			bEnd := s.f.NewBlock(ssa.BlockPlain)
  4722  			b.AddEdgeTo(bTrue)
  4723  			b.AddEdgeTo(bFalse)
  4724  			b.Likely = ssa.BranchLikely
  4725  
  4726  			// We have the intrinsic - use it directly.
  4727  			s.startBlock(bTrue)
  4728  			s.vars[n] = s.newValue3(ssa.OpFMA, types.Types[types.TFLOAT64], args[0], args[1], args[2])
  4729  			s.endBlock().AddEdgeTo(bEnd)
  4730  
  4731  			// Call the pure Go version.
  4732  			s.startBlock(bFalse)
  4733  			s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
  4734  			s.endBlock().AddEdgeTo(bEnd)
  4735  
  4736  			// Merge results.
  4737  			s.startBlock(bEnd)
  4738  			return s.variable(n, types.Types[types.TFLOAT64])
  4739  		},
  4740  		sys.ARM)
  4741  
  4742  	makeRoundAMD64 := func(op ssa.Op) func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4743  		return func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4744  			if buildcfg.GOAMD64 >= 2 {
  4745  				return s.newValue1(op, types.Types[types.TFLOAT64], args[0])
  4746  			}
  4747  
  4748  			v := s.entryNewValue0A(ssa.OpHasCPUFeature, types.Types[types.TBOOL], ir.Syms.X86HasSSE41)
  4749  			b := s.endBlock()
  4750  			b.Kind = ssa.BlockIf
  4751  			b.SetControl(v)
  4752  			bTrue := s.f.NewBlock(ssa.BlockPlain)
  4753  			bFalse := s.f.NewBlock(ssa.BlockPlain)
  4754  			bEnd := s.f.NewBlock(ssa.BlockPlain)
  4755  			b.AddEdgeTo(bTrue)
  4756  			b.AddEdgeTo(bFalse)
  4757  			b.Likely = ssa.BranchLikely // most machines have sse4.1 nowadays
  4758  
  4759  			// We have the intrinsic - use it directly.
  4760  			s.startBlock(bTrue)
  4761  			s.vars[n] = s.newValue1(op, types.Types[types.TFLOAT64], args[0])
  4762  			s.endBlock().AddEdgeTo(bEnd)
  4763  
  4764  			// Call the pure Go version.
  4765  			s.startBlock(bFalse)
  4766  			s.vars[n] = s.callResult(n, callNormal) // types.Types[TFLOAT64]
  4767  			s.endBlock().AddEdgeTo(bEnd)
  4768  
  4769  			// Merge results.
  4770  			s.startBlock(bEnd)
  4771  			return s.variable(n, types.Types[types.TFLOAT64])
  4772  		}
  4773  	}
  4774  	addF("math", "RoundToEven",
  4775  		makeRoundAMD64(ssa.OpRoundToEven),
  4776  		sys.AMD64)
  4777  	addF("math", "Floor",
  4778  		makeRoundAMD64(ssa.OpFloor),
  4779  		sys.AMD64)
  4780  	addF("math", "Ceil",
  4781  		makeRoundAMD64(ssa.OpCeil),
  4782  		sys.AMD64)
  4783  	addF("math", "Trunc",
  4784  		makeRoundAMD64(ssa.OpTrunc),
  4785  		sys.AMD64)
  4786  
  4787  	/******** math/bits ********/
  4788  	addF("math/bits", "TrailingZeros64",
  4789  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4790  			return s.newValue1(ssa.OpCtz64, types.Types[types.TINT], args[0])
  4791  		},
  4792  		sys.AMD64, sys.I386, sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
  4793  	addF("math/bits", "TrailingZeros32",
  4794  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4795  			return s.newValue1(ssa.OpCtz32, types.Types[types.TINT], args[0])
  4796  		},
  4797  		sys.AMD64, sys.I386, sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
  4798  	addF("math/bits", "TrailingZeros16",
  4799  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4800  			x := s.newValue1(ssa.OpZeroExt16to32, types.Types[types.TUINT32], args[0])
  4801  			c := s.constInt32(types.Types[types.TUINT32], 1<<16)
  4802  			y := s.newValue2(ssa.OpOr32, types.Types[types.TUINT32], x, c)
  4803  			return s.newValue1(ssa.OpCtz32, types.Types[types.TINT], y)
  4804  		},
  4805  		sys.MIPS)
  4806  	addF("math/bits", "TrailingZeros16",
  4807  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4808  			return s.newValue1(ssa.OpCtz16, types.Types[types.TINT], args[0])
  4809  		},
  4810  		sys.AMD64, sys.I386, sys.ARM, sys.ARM64, sys.Wasm)
  4811  	addF("math/bits", "TrailingZeros16",
  4812  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4813  			x := s.newValue1(ssa.OpZeroExt16to64, types.Types[types.TUINT64], args[0])
  4814  			c := s.constInt64(types.Types[types.TUINT64], 1<<16)
  4815  			y := s.newValue2(ssa.OpOr64, types.Types[types.TUINT64], x, c)
  4816  			return s.newValue1(ssa.OpCtz64, types.Types[types.TINT], y)
  4817  		},
  4818  		sys.S390X, sys.PPC64)
  4819  	addF("math/bits", "TrailingZeros8",
  4820  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4821  			x := s.newValue1(ssa.OpZeroExt8to32, types.Types[types.TUINT32], args[0])
  4822  			c := s.constInt32(types.Types[types.TUINT32], 1<<8)
  4823  			y := s.newValue2(ssa.OpOr32, types.Types[types.TUINT32], x, c)
  4824  			return s.newValue1(ssa.OpCtz32, types.Types[types.TINT], y)
  4825  		},
  4826  		sys.MIPS)
  4827  	addF("math/bits", "TrailingZeros8",
  4828  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4829  			return s.newValue1(ssa.OpCtz8, types.Types[types.TINT], args[0])
  4830  		},
  4831  		sys.AMD64, sys.I386, sys.ARM, sys.ARM64, sys.Wasm)
  4832  	addF("math/bits", "TrailingZeros8",
  4833  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4834  			x := s.newValue1(ssa.OpZeroExt8to64, types.Types[types.TUINT64], args[0])
  4835  			c := s.constInt64(types.Types[types.TUINT64], 1<<8)
  4836  			y := s.newValue2(ssa.OpOr64, types.Types[types.TUINT64], x, c)
  4837  			return s.newValue1(ssa.OpCtz64, types.Types[types.TINT], y)
  4838  		},
  4839  		sys.S390X)
  4840  	alias("math/bits", "ReverseBytes64", "runtime/internal/sys", "Bswap64", all...)
  4841  	alias("math/bits", "ReverseBytes32", "runtime/internal/sys", "Bswap32", all...)
  4842  	// ReverseBytes inlines correctly, no need to intrinsify it.
  4843  	// Nothing special is needed for targets where ReverseBytes16 lowers to a rotate
  4844  	// On Power10, 16-bit rotate is not available so use BRH instruction
  4845  	if buildcfg.GOPPC64 >= 10 {
  4846  		addF("math/bits", "ReverseBytes16",
  4847  			func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4848  				return s.newValue1(ssa.OpBswap16, types.Types[types.TUINT], args[0])
  4849  			},
  4850  			sys.PPC64)
  4851  	}
  4852  
  4853  	addF("math/bits", "Len64",
  4854  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4855  			return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], args[0])
  4856  		},
  4857  		sys.AMD64, sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
  4858  	addF("math/bits", "Len32",
  4859  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4860  			return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], args[0])
  4861  		},
  4862  		sys.AMD64, sys.ARM64, sys.PPC64)
  4863  	addF("math/bits", "Len32",
  4864  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4865  			if s.config.PtrSize == 4 {
  4866  				return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], args[0])
  4867  			}
  4868  			x := s.newValue1(ssa.OpZeroExt32to64, types.Types[types.TUINT64], args[0])
  4869  			return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], x)
  4870  		},
  4871  		sys.ARM, sys.S390X, sys.MIPS, sys.Wasm)
  4872  	addF("math/bits", "Len16",
  4873  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4874  			if s.config.PtrSize == 4 {
  4875  				x := s.newValue1(ssa.OpZeroExt16to32, types.Types[types.TUINT32], args[0])
  4876  				return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], x)
  4877  			}
  4878  			x := s.newValue1(ssa.OpZeroExt16to64, types.Types[types.TUINT64], args[0])
  4879  			return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], x)
  4880  		},
  4881  		sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
  4882  	addF("math/bits", "Len16",
  4883  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4884  			return s.newValue1(ssa.OpBitLen16, types.Types[types.TINT], args[0])
  4885  		},
  4886  		sys.AMD64)
  4887  	addF("math/bits", "Len8",
  4888  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4889  			if s.config.PtrSize == 4 {
  4890  				x := s.newValue1(ssa.OpZeroExt8to32, types.Types[types.TUINT32], args[0])
  4891  				return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], x)
  4892  			}
  4893  			x := s.newValue1(ssa.OpZeroExt8to64, types.Types[types.TUINT64], args[0])
  4894  			return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], x)
  4895  		},
  4896  		sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
  4897  	addF("math/bits", "Len8",
  4898  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4899  			return s.newValue1(ssa.OpBitLen8, types.Types[types.TINT], args[0])
  4900  		},
  4901  		sys.AMD64)
  4902  	addF("math/bits", "Len",
  4903  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4904  			if s.config.PtrSize == 4 {
  4905  				return s.newValue1(ssa.OpBitLen32, types.Types[types.TINT], args[0])
  4906  			}
  4907  			return s.newValue1(ssa.OpBitLen64, types.Types[types.TINT], args[0])
  4908  		},
  4909  		sys.AMD64, sys.ARM64, sys.ARM, sys.S390X, sys.MIPS, sys.PPC64, sys.Wasm)
  4910  	// LeadingZeros is handled because it trivially calls Len.
  4911  	addF("math/bits", "Reverse64",
  4912  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4913  			return s.newValue1(ssa.OpBitRev64, types.Types[types.TINT], args[0])
  4914  		},
  4915  		sys.ARM64)
  4916  	addF("math/bits", "Reverse32",
  4917  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4918  			return s.newValue1(ssa.OpBitRev32, types.Types[types.TINT], args[0])
  4919  		},
  4920  		sys.ARM64)
  4921  	addF("math/bits", "Reverse16",
  4922  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4923  			return s.newValue1(ssa.OpBitRev16, types.Types[types.TINT], args[0])
  4924  		},
  4925  		sys.ARM64)
  4926  	addF("math/bits", "Reverse8",
  4927  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4928  			return s.newValue1(ssa.OpBitRev8, types.Types[types.TINT], args[0])
  4929  		},
  4930  		sys.ARM64)
  4931  	addF("math/bits", "Reverse",
  4932  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4933  			return s.newValue1(ssa.OpBitRev64, types.Types[types.TINT], args[0])
  4934  		},
  4935  		sys.ARM64)
  4936  	addF("math/bits", "RotateLeft8",
  4937  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4938  			return s.newValue2(ssa.OpRotateLeft8, types.Types[types.TUINT8], args[0], args[1])
  4939  		},
  4940  		sys.AMD64, sys.RISCV64)
  4941  	addF("math/bits", "RotateLeft16",
  4942  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4943  			return s.newValue2(ssa.OpRotateLeft16, types.Types[types.TUINT16], args[0], args[1])
  4944  		},
  4945  		sys.AMD64, sys.RISCV64)
  4946  	addF("math/bits", "RotateLeft32",
  4947  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4948  			return s.newValue2(ssa.OpRotateLeft32, types.Types[types.TUINT32], args[0], args[1])
  4949  		},
  4950  		sys.AMD64, sys.ARM, sys.ARM64, sys.Loong64, sys.PPC64, sys.RISCV64, sys.S390X, sys.Wasm)
  4951  	addF("math/bits", "RotateLeft64",
  4952  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4953  			return s.newValue2(ssa.OpRotateLeft64, types.Types[types.TUINT64], args[0], args[1])
  4954  		},
  4955  		sys.AMD64, sys.ARM64, sys.Loong64, sys.PPC64, sys.RISCV64, sys.S390X, sys.Wasm)
  4956  	alias("math/bits", "RotateLeft", "math/bits", "RotateLeft64", p8...)
  4957  
  4958  	makeOnesCountAMD64 := func(op ssa.Op) func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4959  		return func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4960  			if buildcfg.GOAMD64 >= 2 {
  4961  				return s.newValue1(op, types.Types[types.TINT], args[0])
  4962  			}
  4963  
  4964  			v := s.entryNewValue0A(ssa.OpHasCPUFeature, types.Types[types.TBOOL], ir.Syms.X86HasPOPCNT)
  4965  			b := s.endBlock()
  4966  			b.Kind = ssa.BlockIf
  4967  			b.SetControl(v)
  4968  			bTrue := s.f.NewBlock(ssa.BlockPlain)
  4969  			bFalse := s.f.NewBlock(ssa.BlockPlain)
  4970  			bEnd := s.f.NewBlock(ssa.BlockPlain)
  4971  			b.AddEdgeTo(bTrue)
  4972  			b.AddEdgeTo(bFalse)
  4973  			b.Likely = ssa.BranchLikely // most machines have popcnt nowadays
  4974  
  4975  			// We have the intrinsic - use it directly.
  4976  			s.startBlock(bTrue)
  4977  			s.vars[n] = s.newValue1(op, types.Types[types.TINT], args[0])
  4978  			s.endBlock().AddEdgeTo(bEnd)
  4979  
  4980  			// Call the pure Go version.
  4981  			s.startBlock(bFalse)
  4982  			s.vars[n] = s.callResult(n, callNormal) // types.Types[TINT]
  4983  			s.endBlock().AddEdgeTo(bEnd)
  4984  
  4985  			// Merge results.
  4986  			s.startBlock(bEnd)
  4987  			return s.variable(n, types.Types[types.TINT])
  4988  		}
  4989  	}
  4990  	addF("math/bits", "OnesCount64",
  4991  		makeOnesCountAMD64(ssa.OpPopCount64),
  4992  		sys.AMD64)
  4993  	addF("math/bits", "OnesCount64",
  4994  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  4995  			return s.newValue1(ssa.OpPopCount64, types.Types[types.TINT], args[0])
  4996  		},
  4997  		sys.PPC64, sys.ARM64, sys.S390X, sys.Wasm)
  4998  	addF("math/bits", "OnesCount32",
  4999  		makeOnesCountAMD64(ssa.OpPopCount32),
  5000  		sys.AMD64)
  5001  	addF("math/bits", "OnesCount32",
  5002  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  5003  			return s.newValue1(ssa.OpPopCount32, types.Types[types.TINT], args[0])
  5004  		},
  5005  		sys.PPC64, sys.ARM64, sys.S390X, sys.Wasm)
  5006  	addF("math/bits", "OnesCount16",
  5007  		makeOnesCountAMD64(ssa.OpPopCount16),
  5008  		sys.AMD64)
  5009  	addF("math/bits", "OnesCount16",
  5010  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  5011  			return s.newValue1(ssa.OpPopCount16, types.Types[types.TINT], args[0])
  5012  		},
  5013  		sys.ARM64, sys.S390X, sys.PPC64, sys.Wasm)
  5014  	addF("math/bits", "OnesCount8",
  5015  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  5016  			return s.newValue1(ssa.OpPopCount8, types.Types[types.TINT], args[0])
  5017  		},
  5018  		sys.S390X, sys.PPC64, sys.Wasm)
  5019  	addF("math/bits", "OnesCount",
  5020  		makeOnesCountAMD64(ssa.OpPopCount64),
  5021  		sys.AMD64)
  5022  	addF("math/bits", "Mul64",
  5023  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  5024  			return s.newValue2(ssa.OpMul64uhilo, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1])
  5025  		},
  5026  		sys.AMD64, sys.ARM64, sys.PPC64, sys.S390X, sys.MIPS64, sys.RISCV64, sys.Loong64)
  5027  	alias("math/bits", "Mul", "math/bits", "Mul64", p8...)
  5028  	alias("runtime/internal/math", "Mul64", "math/bits", "Mul64", p8...)
  5029  	addF("math/bits", "Add64",
  5030  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  5031  			return s.newValue3(ssa.OpAdd64carry, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1], args[2])
  5032  		},
  5033  		sys.AMD64, sys.ARM64, sys.PPC64, sys.S390X, sys.RISCV64, sys.Loong64, sys.MIPS64)
  5034  	alias("math/bits", "Add", "math/bits", "Add64", p8...)
  5035  	alias("runtime/internal/math", "Add64", "math/bits", "Add64", all...)
  5036  	addF("math/bits", "Sub64",
  5037  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  5038  			return s.newValue3(ssa.OpSub64borrow, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1], args[2])
  5039  		},
  5040  		sys.AMD64, sys.ARM64, sys.PPC64, sys.S390X, sys.RISCV64, sys.Loong64, sys.MIPS64)
  5041  	alias("math/bits", "Sub", "math/bits", "Sub64", p8...)
  5042  	addF("math/bits", "Div64",
  5043  		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
  5044  			// check for divide-by-zero/overflow and panic with appropriate message
  5045  			cmpZero := s.newValue2(s.ssaOp(ir.ONE, types.Types[types.TUINT64]), types.Types[types.TBOOL], args[2], s.zeroVal(types.Types[types.TUINT64]))
  5046  			s.check(cmpZero, ir.Syms.Panicdivide)
  5047  			cmpOverflow := s.newValue2(s.ssaOp(ir.OLT, types.Types[types.TUINT64]), types.Types[types.TBOOL], args[0], args[2])
  5048  			s.check(cmpOverflow, ir.Syms.Panicoverflow)
  5049  			return s.newValue3(ssa.OpDiv128u, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1], args[2])
  5050  		},
  5051  		sys.AMD64)
  5052  	alias("math/bits", "Div", "math/bits", "Div64", sys.ArchAMD64)
  5053  
  5054  	alias("runtime/internal/sys", "TrailingZeros8", "math/bits", "TrailingZeros8", all...)
  5055  	alias("runtime/internal/sys", "TrailingZeros32", "math/bits", "TrailingZeros32", all...)
  5056  	alias("runtime/internal/sys", "TrailingZeros64", "math/bits", "TrailingZeros64", all...)
  5057  	alias("runtime/internal/sys", "Len8", "math/bits", "Len8", all...)
  5058  	alias("runtime/internal/sys", "Len64", "math/bits", "Len64", all...)
  5059  	alias("runtime/internal/sys", "OnesCount64", "math/bits", "OnesCount64", all...)
  5060  
  5061  	/******** sync/atomic ********/
  5062  
  5063  	// Note: these are disabled by flag_race in findIntrinsic below.
  5064  	alias("sync/atomic", "LoadInt32", "internal/runtime/atomic", "Load", all...)
  5065  	alias("sync/atomic", "LoadInt64", "internal/runtime/atomic", "Load64", all...)
  5066  	alias("sync/atomic", "LoadPointer", "internal/runtime/atomic", "Loadp", all...)
  5067  	alias("sync/atomic", "LoadUint32", "internal/runtime/atomic", "Load", all...)
  5068  	alias("sync/atomic", "LoadUint64", "internal/runtime/atomic", "Load64", all...)
  5069  	alias("sync/atomic", "LoadUintptr", "internal/runtime/atomic", "Load", p4...)
  5070  	alias("sync/atomic", "LoadUintptr", "internal/runtime/atomic", "Load64", p8...)
  5071  
  5072  	alias("sync/atomic", "StoreInt32", "internal/runtime/atomic", "Store", all...)
  5073  	alias("sync/atomic", "StoreInt64", "internal/runtime/atomic", "Store64", all...)
  5074  	// Note: not StorePointer, that needs a write barrier.  Same below for {CompareAnd}Swap.
  5075  	alias("sync/atomic", "StoreUint32", "internal/runtime/atomic", "Store", all...)
  5076  	alias("sync/atomic", "StoreUint64", "internal/runtime/atomic", "Store64", all...)
  5077  	alias("sync/atomic", "StoreUintptr", "internal/runtime/atomic", "Store", p4...)
  5078  	alias("sync/atomic", "StoreUintptr", "internal/runtime/atomic", "Store64", p8...)
  5079  
  5080  	alias("sync/atomic", "SwapInt32", "internal/runtime/atomic", "Xchg", all...)
  5081  	alias("sync/atomic", "SwapInt64", "internal/runtime/atomic", "Xchg64", all...)
  5082  	alias("sync/atomic", "SwapUint32", "internal/runtime/atomic", "Xchg", all...)
  5083  	alias("sync/atomic", "SwapUint64", "internal/runtime/atomic", "Xchg64", all...)
  5084  	alias("sync/atomic", "SwapUintptr", "internal/runtime/atomic", "Xchg", p4...)
  5085  	alias("sync/atomic", "SwapUintptr", "internal/runtime/atomic", "Xchg64", p8...)
  5086  
  5087  	alias("sync/atomic", "CompareAndSwapInt32", "internal/runtime/atomic", "Cas", all...)
  5088  	alias("sync/atomic", "CompareAndSwapInt64", "internal/runtime/atomic", "Cas64", all...)
  5089  	alias("sync/atomic", "CompareAndSwapUint32", "internal/runtime/atomic", "Cas", all...)
  5090  	alias("sync/atomic", "CompareAndSwapUint64", "internal/runtime/atomic", "Cas64", all...)
  5091  	alias("sync/atomic", "CompareAndSwapUintptr", "internal/runtime/atomic", "Cas", p4...)
  5092  	alias("sync/atomic", "CompareAndSwapUintptr", "internal/runtime/atomic", "Cas64", p8...)
  5093  
  5094  	alias("sync/atomic", "AddInt32", "internal/runtime/atomic", "Xadd", all...)
  5095  	alias("sync/atomic", "AddInt64", "internal/runtime/atomic", "Xadd64", all...)
  5096  	alias("sync/atomic", "AddUint32", "internal/runtime/atomic", "Xadd", all...)
  5097  	alias("sync/atomic", "AddUint64", "internal/runtime/atomic", "Xadd64", all...)
  5098  	alias("sync/atomic", "AddUintptr", "internal/runtime/atomic", "Xadd", p4...)
  5099  	alias("sync/atomic", "AddUintptr", "internal/runtime/atomic", "Xadd64", p8...)
  5100  
  5101  	alias("sync/atomic", "AndInt32", "internal/runtime/atomic", "And32", sys.ArchARM64)
  5102  	alias("sync/atomic", "AndUint32", "internal/runtime/atomic", "And32", sys.ArchARM64)
  5103  	alias("sync/atomic", "AndInt64", "internal/runtime/atomic", "And64", sys.ArchARM64)
  5104  	alias("sync/atomic", "AndUint64", "internal/runtime/atomic", "And64", sys.ArchARM64)
  5105  	alias("sync/atomic", "AndUintptr", "internal/runtime/atomic", "And64", sys.ArchARM64)
  5106  	alias("sync/atomic", "OrInt32", "internal/runtime/atomic", "Or32", sys.ArchARM64)
  5107  	alias("sync/atomic", "OrUint32", "internal/runtime/atomic", "Or32", sys.ArchARM64)
  5108  	alias("sync/atomic", "OrInt64", "internal/runtime/atomic", "Or64", sys.ArchARM64)
  5109  	alias("sync/atomic", "OrUint64", "internal/runtime/atomic", "Or64", sys.ArchARM64)
  5110  	alias("sync/atomic", "OrUintptr", "internal/runtime/atomic", "Or64", sys.ArchARM64)
  5111  
  5112  	/******** math/big ********/
  5113  	alias("math/big", "mulWW", "math/bits", "Mul64", p8...)
  5114  }
  5115  
  5116  // findIntrinsic returns a function which builds the SSA equivalent of the
  5117  // function identified by the symbol sym.  If sym is not an intrinsic call, returns nil.
  5118  func findIntrinsic(sym *types.Sym) intrinsicBuilder {
  5119  	if sym == nil || sym.Pkg == nil {
  5120  		return nil
  5121  	}
  5122  	pkg := sym.Pkg.Path
  5123  	if sym.Pkg == ir.Pkgs.Runtime {
  5124  		pkg = "runtime"
  5125  	}
  5126  	if base.Flag.Race && pkg == "sync/atomic" {
  5127  		// The race detector needs to be able to intercept these calls.
  5128  		// We can't intrinsify them.
  5129  		return nil
  5130  	}
  5131  	// Skip intrinsifying math functions (which may contain hard-float
  5132  	// instructions) when soft-float
  5133  	if Arch.SoftFloat && pkg == "math" {
  5134  		return nil
  5135  	}
  5136  
  5137  	fn := sym.Name
  5138  	if ssa.IntrinsicsDisable {
  5139  		if pkg == "runtime" && (fn == "getcallerpc" || fn == "getcallersp" || fn == "getclosureptr") {
  5140  			// These runtime functions don't have definitions, must be intrinsics.
  5141  		} else {
  5142  			return nil
  5143  		}
  5144  	}
  5145  	return intrinsics[intrinsicKey{Arch.LinkArch.Arch, pkg, fn}]
  5146  }
  5147  
  5148  func IsIntrinsicCall(n *ir.CallExpr) bool {
  5149  	if n == nil {
  5150  		return false
  5151  	}
  5152  	name, ok := n.Fun.(*ir.Name)
  5153  	if !ok {
  5154  		return false
  5155  	}
  5156  	return findIntrinsic(name.Sym()) != nil
  5157  }
  5158  
  5159  // intrinsicCall converts a call to a recognized intrinsic function into the intrinsic SSA operation.
  5160  func (s *state) intrinsicCall(n *ir.CallExpr) *ssa.Value {
  5161  	v := findIntrinsic(n.Fun.Sym())(s, n, s.intrinsicArgs(n))
  5162  	if ssa.IntrinsicsDebug > 0 {
  5163  		x := v
  5164  		if x == nil {
  5165  			x = s.mem()
  5166  		}
  5167  		if x.Op == ssa.OpSelect0 || x.Op == ssa.OpSelect1 {
  5168  			x = x.Args[0]
  5169  		}
  5170  		base.WarnfAt(n.Pos(), "intrinsic substitution for %v with %s", n.Fun.Sym().Name, x.LongString())
  5171  	}
  5172  	return v
  5173  }
  5174  
  5175  // intrinsicArgs extracts args from n, evaluates them to SSA values, and returns them.
  5176  func (s *state) intrinsicArgs(n *ir.CallExpr) []*ssa.Value {
  5177  	args := make([]*ssa.Value, len(n.Args))
  5178  	for i, n := range n.Args {
  5179  		args[i] = s.expr(n)
  5180  	}
  5181  	return args
  5182  }
  5183  
  5184  // openDeferRecord adds code to evaluate and store the function for an open-code defer
  5185  // call, and records info about the defer, so we can generate proper code on the
  5186  // exit paths. n is the sub-node of the defer node that is the actual function
  5187  // call. We will also record funcdata information on where the function is stored
  5188  // (as well as the deferBits variable), and this will enable us to run the proper
  5189  // defer calls during panics.
  5190  func (s *state) openDeferRecord(n *ir.CallExpr) {
  5191  	if len(n.Args) != 0 || n.Op() != ir.OCALLFUNC || n.Fun.Type().NumResults() != 0 {
  5192  		s.Fatalf("defer call with arguments or results: %v", n)
  5193  	}
  5194  
  5195  	opendefer := &openDeferInfo{
  5196  		n: n,
  5197  	}
  5198  	fn := n.Fun
  5199  	// We must always store the function value in a stack slot for the
  5200  	// runtime panic code to use. But in the defer exit code, we will
  5201  	// call the function directly if it is a static function.
  5202  	closureVal := s.expr(fn)
  5203  	closure := s.openDeferSave(fn.Type(), closureVal)
  5204  	opendefer.closureNode = closure.Aux.(*ir.Name)
  5205  	if !(fn.Op() == ir.ONAME && fn.(*ir.Name).Class == ir.PFUNC) {
  5206  		opendefer.closure = closure
  5207  	}
  5208  	index := len(s.openDefers)
  5209  	s.openDefers = append(s.openDefers, opendefer)
  5210  
  5211  	// Update deferBits only after evaluation and storage to stack of
  5212  	// the function is successful.
  5213  	bitvalue := s.constInt8(types.Types[types.TUINT8], 1<<uint(index))
  5214  	newDeferBits := s.newValue2(ssa.OpOr8, types.Types[types.TUINT8], s.variable(deferBitsVar, types.Types[types.TUINT8]), bitvalue)
  5215  	s.vars[deferBitsVar] = newDeferBits
  5216  	s.store(types.Types[types.TUINT8], s.deferBitsAddr, newDeferBits)
  5217  }
  5218  
  5219  // openDeferSave generates SSA nodes to store a value (with type t) for an
  5220  // open-coded defer at an explicit autotmp location on the stack, so it can be
  5221  // reloaded and used for the appropriate call on exit. Type t must be a function type
  5222  // (therefore SSAable). val is the value to be stored. The function returns an SSA
  5223  // value representing a pointer to the autotmp location.
  5224  func (s *state) openDeferSave(t *types.Type, val *ssa.Value) *ssa.Value {
  5225  	if !ssa.CanSSA(t) {
  5226  		s.Fatalf("openDeferSave of non-SSA-able type %v val=%v", t, val)
  5227  	}
  5228  	if !t.HasPointers() {
  5229  		s.Fatalf("openDeferSave of pointerless type %v val=%v", t, val)
  5230  	}
  5231  	pos := val.Pos
  5232  	temp := typecheck.TempAt(pos.WithNotStmt(), s.curfn, t)
  5233  	temp.SetOpenDeferSlot(true)
  5234  	temp.SetFrameOffset(int64(len(s.openDefers))) // so cmpstackvarlt can order them
  5235  	var addrTemp *ssa.Value
  5236  	// Use OpVarLive to make sure stack slot for the closure is not removed by
  5237  	// dead-store elimination
  5238  	if s.curBlock.ID != s.f.Entry.ID {
  5239  		// Force the tmp storing this defer function to be declared in the entry
  5240  		// block, so that it will be live for the defer exit code (which will
  5241  		// actually access it only if the associated defer call has been activated).
  5242  		if t.HasPointers() {
  5243  			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])
  5244  		}
  5245  		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])
  5246  		addrTemp = s.f.Entry.NewValue2A(src.NoXPos, ssa.OpLocalAddr, types.NewPtr(temp.Type()), temp, s.sp, s.defvars[s.f.Entry.ID][memVar])
  5247  	} else {
  5248  		// Special case if we're still in the entry block. We can't use
  5249  		// the above code, since s.defvars[s.f.Entry.ID] isn't defined
  5250  		// until we end the entry block with s.endBlock().
  5251  		if t.HasPointers() {
  5252  			s.vars[memVar] = s.newValue1Apos(ssa.OpVarDef, types.TypeMem, temp, s.mem(), false)
  5253  		}
  5254  		s.vars[memVar] = s.newValue1Apos(ssa.OpVarLive, types.TypeMem, temp, s.mem(), false)
  5255  		addrTemp = s.newValue2Apos(ssa.OpLocalAddr, types.NewPtr(temp.Type()), temp, s.sp, s.mem(), false)
  5256  	}
  5257  	// Since we may use this temp during exit depending on the
  5258  	// deferBits, we must define it unconditionally on entry.
  5259  	// Therefore, we must make sure it is zeroed out in the entry
  5260  	// block if it contains pointers, else GC may wrongly follow an
  5261  	// uninitialized pointer value.
  5262  	temp.SetNeedzero(true)
  5263  	// We are storing to the stack, hence we can avoid the full checks in
  5264  	// storeType() (no write barrier) and do a simple store().
  5265  	s.store(t, addrTemp, val)
  5266  	return addrTemp
  5267  }
  5268  
  5269  // openDeferExit generates SSA for processing all the open coded defers at exit.
  5270  // The code involves loading deferBits, and checking each of the bits to see if
  5271  // the corresponding defer statement was executed. For each bit that is turned
  5272  // on, the associated defer call is made.
  5273  func (s *state) openDeferExit() {
  5274  	deferExit := s.f.NewBlock(ssa.BlockPlain)
  5275  	s.endBlock().AddEdgeTo(deferExit)
  5276  	s.startBlock(deferExit)
  5277  	s.lastDeferExit = deferExit
  5278  	s.lastDeferCount = len(s.openDefers)
  5279  	zeroval := s.constInt8(types.Types[types.TUINT8], 0)
  5280  	// Test for and run defers in reverse order
  5281  	for i := len(s.openDefers) - 1; i >= 0; i-- {
  5282  		r := s.openDefers[i]
  5283  		bCond := s.f.NewBlock(ssa.BlockPlain)
  5284  		bEnd := s.f.NewBlock(ssa.BlockPlain)
  5285  
  5286  		deferBits := s.variable(deferBitsVar, types.Types[types.TUINT8])
  5287  		// Generate code to check if the bit associated with the current
  5288  		// defer is set.
  5289  		bitval := s.constInt8(types.Types[types.TUINT8], 1<<uint(i))
  5290  		andval := s.newValue2(ssa.OpAnd8, types.Types[types.TUINT8], deferBits, bitval)
  5291  		eqVal := s.newValue2(ssa.OpEq8, types.Types[types.TBOOL], andval, zeroval)
  5292  		b := s.endBlock()
  5293  		b.Kind = ssa.BlockIf
  5294  		b.SetControl(eqVal)
  5295  		b.AddEdgeTo(bEnd)
  5296  		b.AddEdgeTo(bCond)
  5297  		bCond.AddEdgeTo(bEnd)
  5298  		s.startBlock(bCond)
  5299  
  5300  		// Clear this bit in deferBits and force store back to stack, so
  5301  		// we will not try to re-run this defer call if this defer call panics.
  5302  		nbitval := s.newValue1(ssa.OpCom8, types.Types[types.TUINT8], bitval)
  5303  		maskedval := s.newValue2(ssa.OpAnd8, types.Types[types.TUINT8], deferBits, nbitval)
  5304  		s.store(types.Types[types.TUINT8], s.deferBitsAddr, maskedval)
  5305  		// Use this value for following tests, so we keep previous
  5306  		// bits cleared.
  5307  		s.vars[deferBitsVar] = maskedval
  5308  
  5309  		// Generate code to call the function call of the defer, using the
  5310  		// closure that were stored in argtmps at the point of the defer
  5311  		// statement.
  5312  		fn := r.n.Fun
  5313  		stksize := fn.Type().ArgWidth()
  5314  		var callArgs []*ssa.Value
  5315  		var call *ssa.Value
  5316  		if r.closure != nil {
  5317  			v := s.load(r.closure.Type.Elem(), r.closure)
  5318  			s.maybeNilCheckClosure(v, callDefer)
  5319  			codeptr := s.rawLoad(types.Types[types.TUINTPTR], v)
  5320  			aux := ssa.ClosureAuxCall(s.f.ABIDefault.ABIAnalyzeTypes(nil, nil))
  5321  			call = s.newValue2A(ssa.OpClosureLECall, aux.LateExpansionResultType(), aux, codeptr, v)
  5322  		} else {
  5323  			aux := ssa.StaticAuxCall(fn.(*ir.Name).Linksym(), s.f.ABIDefault.ABIAnalyzeTypes(nil, nil))
  5324  			call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
  5325  		}
  5326  		callArgs = append(callArgs, s.mem())
  5327  		call.AddArgs(callArgs...)
  5328  		call.AuxInt = stksize
  5329  		s.vars[memVar] = s.newValue1I(ssa.OpSelectN, types.TypeMem, 0, call)
  5330  		// Make sure that the stack slots with pointers are kept live
  5331  		// through the call (which is a pre-emption point). Also, we will
  5332  		// use the first call of the last defer exit to compute liveness
  5333  		// for the deferreturn, so we want all stack slots to be live.
  5334  		if r.closureNode != nil {
  5335  			s.vars[memVar] = s.newValue1Apos(ssa.OpVarLive, types.TypeMem, r.closureNode, s.mem(), false)
  5336  		}
  5337  
  5338  		s.endBlock()
  5339  		s.startBlock(bEnd)
  5340  	}
  5341  }
  5342  
  5343  func (s *state) callResult(n *ir.CallExpr, k callKind) *ssa.Value {
  5344  	return s.call(n, k, false, nil)
  5345  }
  5346  
  5347  func (s *state) callAddr(n *ir.CallExpr, k callKind) *ssa.Value {
  5348  	return s.call(n, k, true, nil)
  5349  }
  5350  
  5351  // Calls the function n using the specified call type.
  5352  // Returns the address of the return value (or nil if none).
  5353  func (s *state) call(n *ir.CallExpr, k callKind, returnResultAddr bool, deferExtra ir.Expr) *ssa.Value {
  5354  	s.prevCall = nil
  5355  	var calleeLSym *obj.LSym // target function (if static)
  5356  	var closure *ssa.Value   // ptr to closure to run (if dynamic)
  5357  	var codeptr *ssa.Value   // ptr to target code (if dynamic)
  5358  	var dextra *ssa.Value    // defer extra arg
  5359  	var rcvr *ssa.Value      // receiver to set
  5360  	fn := n.Fun
  5361  	var ACArgs []*types.Type    // AuxCall args
  5362  	var ACResults []*types.Type // AuxCall results
  5363  	var callArgs []*ssa.Value   // For late-expansion, the args themselves (not stored, args to the call instead).
  5364  
  5365  	callABI := s.f.ABIDefault
  5366  
  5367  	if k != callNormal && k != callTail && (len(n.Args) != 0 || n.Op() == ir.OCALLINTER || n.Fun.Type().NumResults() != 0) {
  5368  		s.Fatalf("go/defer call with arguments: %v", n)
  5369  	}
  5370  
  5371  	switch n.Op() {
  5372  	case ir.OCALLFUNC:
  5373  		if (k == callNormal || k == callTail) && fn.Op() == ir.ONAME && fn.(*ir.Name).Class == ir.PFUNC {
  5374  			fn := fn.(*ir.Name)
  5375  			calleeLSym = callTargetLSym(fn)
  5376  			if buildcfg.Experiment.RegabiArgs {
  5377  				// This is a static call, so it may be
  5378  				// a direct call to a non-ABIInternal
  5379  				// function. fn.Func may be nil for
  5380  				// some compiler-generated functions,
  5381  				// but those are all ABIInternal.
  5382  				if fn.Func != nil {
  5383  					callABI = abiForFunc(fn.Func, s.f.ABI0, s.f.ABI1)
  5384  				}
  5385  			} else {
  5386  				// TODO(register args) remove after register abi is working
  5387  				inRegistersImported := fn.Pragma()&ir.RegisterParams != 0
  5388  				inRegistersSamePackage := fn.Func != nil && fn.Func.Pragma&ir.RegisterParams != 0
  5389  				if inRegistersImported || inRegistersSamePackage {
  5390  					callABI = s.f.ABI1
  5391  				}
  5392  			}
  5393  			break
  5394  		}
  5395  		closure = s.expr(fn)
  5396  		if k != callDefer && k != callDeferStack {
  5397  			// Deferred nil function needs to panic when the function is invoked,
  5398  			// not the point of defer statement.
  5399  			s.maybeNilCheckClosure(closure, k)
  5400  		}
  5401  	case ir.OCALLINTER:
  5402  		if fn.Op() != ir.ODOTINTER {
  5403  			s.Fatalf("OCALLINTER: n.Left not an ODOTINTER: %v", fn.Op())
  5404  		}
  5405  		fn := fn.(*ir.SelectorExpr)
  5406  		var iclosure *ssa.Value
  5407  		iclosure, rcvr = s.getClosureAndRcvr(fn)
  5408  		if k == callNormal {
  5409  			codeptr = s.load(types.Types[types.TUINTPTR], iclosure)
  5410  		} else {
  5411  			closure = iclosure
  5412  		}
  5413  	}
  5414  	if deferExtra != nil {
  5415  		dextra = s.expr(deferExtra)
  5416  	}
  5417  
  5418  	params := callABI.ABIAnalyze(n.Fun.Type(), false /* Do not set (register) nNames from caller side -- can cause races. */)
  5419  	types.CalcSize(fn.Type())
  5420  	stksize := params.ArgWidth() // includes receiver, args, and results
  5421  
  5422  	res := n.Fun.Type().Results()
  5423  	if k == callNormal || k == callTail {
  5424  		for _, p := range params.OutParams() {
  5425  			ACResults = append(ACResults, p.Type)
  5426  		}
  5427  	}
  5428  
  5429  	var call *ssa.Value
  5430  	if k == callDeferStack {
  5431  		if stksize != 0 {
  5432  			s.Fatalf("deferprocStack with non-zero stack size %d: %v", stksize, n)
  5433  		}
  5434  		// Make a defer struct on the stack.
  5435  		t := deferstruct()
  5436  		n, addr := s.temp(n.Pos(), t)
  5437  		n.SetNonMergeable(true)
  5438  		s.store(closure.Type,
  5439  			s.newValue1I(ssa.OpOffPtr, closure.Type.PtrTo(), t.FieldOff(deferStructFnField), addr),
  5440  			closure)
  5441  
  5442  		// Call runtime.deferprocStack with pointer to _defer record.
  5443  		ACArgs = append(ACArgs, types.Types[types.TUINTPTR])
  5444  		aux := ssa.StaticAuxCall(ir.Syms.DeferprocStack, s.f.ABIDefault.ABIAnalyzeTypes(ACArgs, ACResults))
  5445  		callArgs = append(callArgs, addr, s.mem())
  5446  		call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
  5447  		call.AddArgs(callArgs...)
  5448  		call.AuxInt = int64(types.PtrSize) // deferprocStack takes a *_defer arg
  5449  	} else {
  5450  		// Store arguments to stack, including defer/go arguments and receiver for method calls.
  5451  		// These are written in SP-offset order.
  5452  		argStart := base.Ctxt.Arch.FixedFrameSize
  5453  		// Defer/go args.
  5454  		if k != callNormal && k != callTail {
  5455  			// Write closure (arg to newproc/deferproc).
  5456  			ACArgs = append(ACArgs, types.Types[types.TUINTPTR]) // not argExtra
  5457  			callArgs = append(callArgs, closure)
  5458  			stksize += int64(types.PtrSize)
  5459  			argStart += int64(types.PtrSize)
  5460  			if dextra != nil {
  5461  				// Extra token of type any for deferproc
  5462  				ACArgs = append(ACArgs, types.Types[types.TINTER])
  5463  				callArgs = append(callArgs, dextra)
  5464  				stksize += 2 * int64(types.PtrSize)
  5465  				argStart += 2 * int64(types.PtrSize)
  5466  			}
  5467  		}
  5468  
  5469  		// Set receiver (for interface calls).
  5470  		if rcvr != nil {
  5471  			callArgs = append(callArgs, rcvr)
  5472  		}
  5473  
  5474  		// Write args.
  5475  		t := n.Fun.Type()
  5476  		args := n.Args
  5477  
  5478  		for _, p := range params.InParams() { // includes receiver for interface calls
  5479  			ACArgs = append(ACArgs, p.Type)
  5480  		}
  5481  
  5482  		// Split the entry block if there are open defers, because later calls to
  5483  		// openDeferSave may cause a mismatch between the mem for an OpDereference
  5484  		// and the call site which uses it. See #49282.
  5485  		if s.curBlock.ID == s.f.Entry.ID && s.hasOpenDefers {
  5486  			b := s.endBlock()
  5487  			b.Kind = ssa.BlockPlain
  5488  			curb := s.f.NewBlock(ssa.BlockPlain)
  5489  			b.AddEdgeTo(curb)
  5490  			s.startBlock(curb)
  5491  		}
  5492  
  5493  		for i, n := range args {
  5494  			callArgs = append(callArgs, s.putArg(n, t.Param(i).Type))
  5495  		}
  5496  
  5497  		callArgs = append(callArgs, s.mem())
  5498  
  5499  		// call target
  5500  		switch {
  5501  		case k == callDefer:
  5502  			sym := ir.Syms.Deferproc
  5503  			if dextra != nil {
  5504  				sym = ir.Syms.Deferprocat
  5505  			}
  5506  			aux := ssa.StaticAuxCall(sym, s.f.ABIDefault.ABIAnalyzeTypes(ACArgs, ACResults)) // TODO paramResultInfo for Deferproc(at)
  5507  			call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
  5508  		case k == callGo:
  5509  			aux := ssa.StaticAuxCall(ir.Syms.Newproc, s.f.ABIDefault.ABIAnalyzeTypes(ACArgs, ACResults))
  5510  			call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux) // TODO paramResultInfo for Newproc
  5511  		case closure != nil:
  5512  			// rawLoad because loading the code pointer from a
  5513  			// closure is always safe, but IsSanitizerSafeAddr
  5514  			// can't always figure that out currently, and it's
  5515  			// critical that we not clobber any arguments already
  5516  			// stored onto the stack.
  5517  			codeptr = s.rawLoad(types.Types[types.TUINTPTR], closure)
  5518  			aux := ssa.ClosureAuxCall(callABI.ABIAnalyzeTypes(ACArgs, ACResults))
  5519  			call = s.newValue2A(ssa.OpClosureLECall, aux.LateExpansionResultType(), aux, codeptr, closure)
  5520  		case codeptr != nil:
  5521  			// Note that the "receiver" parameter is nil because the actual receiver is the first input parameter.
  5522  			aux := ssa.InterfaceAuxCall(params)
  5523  			call = s.newValue1A(ssa.OpInterLECall, aux.LateExpansionResultType(), aux, codeptr)
  5524  		case calleeLSym != nil:
  5525  			aux := ssa.StaticAuxCall(calleeLSym, params)
  5526  			call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
  5527  			if k == callTail {
  5528  				call.Op = ssa.OpTailLECall
  5529  				stksize = 0 // Tail call does not use stack. We reuse caller's frame.
  5530  			}
  5531  		default:
  5532  			s.Fatalf("bad call type %v %v", n.Op(), n)
  5533  		}
  5534  		call.AddArgs(callArgs...)
  5535  		call.AuxInt = stksize // Call operations carry the argsize of the callee along with them
  5536  	}
  5537  	s.prevCall = call
  5538  	s.vars[memVar] = s.newValue1I(ssa.OpSelectN, types.TypeMem, int64(len(ACResults)), call)
  5539  	// Insert VarLive opcodes.
  5540  	for _, v := range n.KeepAlive {
  5541  		if !v.Addrtaken() {
  5542  			s.Fatalf("KeepAlive variable %v must have Addrtaken set", v)
  5543  		}
  5544  		switch v.Class {
  5545  		case ir.PAUTO, ir.PPARAM, ir.PPARAMOUT:
  5546  		default:
  5547  			s.Fatalf("KeepAlive variable %v must be Auto or Arg", v)
  5548  		}
  5549  		s.vars[memVar] = s.newValue1A(ssa.OpVarLive, types.TypeMem, v, s.mem())
  5550  	}
  5551  
  5552  	// Finish block for defers
  5553  	if k == callDefer || k == callDeferStack {
  5554  		b := s.endBlock()
  5555  		b.Kind = ssa.BlockDefer
  5556  		b.SetControl(call)
  5557  		bNext := s.f.NewBlock(ssa.BlockPlain)
  5558  		b.AddEdgeTo(bNext)
  5559  		// Add recover edge to exit code.
  5560  		r := s.f.NewBlock(ssa.BlockPlain)
  5561  		s.startBlock(r)
  5562  		s.exit()
  5563  		b.AddEdgeTo(r)
  5564  		b.Likely = ssa.BranchLikely
  5565  		s.startBlock(bNext)
  5566  	}
  5567  
  5568  	if len(res) == 0 || k != callNormal {
  5569  		// call has no return value. Continue with the next statement.
  5570  		return nil
  5571  	}
  5572  	fp := res[0]
  5573  	if returnResultAddr {
  5574  		return s.resultAddrOfCall(call, 0, fp.Type)
  5575  	}
  5576  	return s.newValue1I(ssa.OpSelectN, fp.Type, 0, call)
  5577  }
  5578  
  5579  // maybeNilCheckClosure checks if a nil check of a closure is needed in some
  5580  // architecture-dependent situations and, if so, emits the nil check.
  5581  func (s *state) maybeNilCheckClosure(closure *ssa.Value, k callKind) {
  5582  	if Arch.LinkArch.Family == sys.Wasm || buildcfg.GOOS == "aix" && k != callGo {
  5583  		// 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.
  5584  		// TODO(neelance): On other architectures this should be eliminated by the optimization steps
  5585  		s.nilCheck(closure)
  5586  	}
  5587  }
  5588  
  5589  // getClosureAndRcvr returns values for the appropriate closure and receiver of an
  5590  // interface call
  5591  func (s *state) getClosureAndRcvr(fn *ir.SelectorExpr) (*ssa.Value, *ssa.Value) {
  5592  	i := s.expr(fn.X)
  5593  	itab := s.newValue1(ssa.OpITab, types.Types[types.TUINTPTR], i)
  5594  	s.nilCheck(itab)
  5595  	itabidx := fn.Offset() + rttype.ITab.OffsetOf("Fun")
  5596  	closure := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.UintptrPtr, itabidx, itab)
  5597  	rcvr := s.newValue1(ssa.OpIData, s.f.Config.Types.BytePtr, i)
  5598  	return closure, rcvr
  5599  }
  5600  
  5601  // etypesign returns the signed-ness of e, for integer/pointer etypes.
  5602  // -1 means signed, +1 means unsigned, 0 means non-integer/non-pointer.
  5603  func etypesign(e types.Kind) int8 {
  5604  	switch e {
  5605  	case types.TINT8, types.TINT16, types.TINT32, types.TINT64, types.TINT:
  5606  		return -1
  5607  	case types.TUINT8, types.TUINT16, types.TUINT32, types.TUINT64, types.TUINT, types.TUINTPTR, types.TUNSAFEPTR:
  5608  		return +1
  5609  	}
  5610  	return 0
  5611  }
  5612  
  5613  // addr converts the address of the expression n to SSA, adds it to s and returns the SSA result.
  5614  // The value that the returned Value represents is guaranteed to be non-nil.
  5615  func (s *state) addr(n ir.Node) *ssa.Value {
  5616  	if n.Op() != ir.ONAME {
  5617  		s.pushLine(n.Pos())
  5618  		defer s.popLine()
  5619  	}
  5620  
  5621  	if s.canSSA(n) {
  5622  		s.Fatalf("addr of canSSA expression: %+v", n)
  5623  	}
  5624  
  5625  	t := types.NewPtr(n.Type())
  5626  	linksymOffset := func(lsym *obj.LSym, offset int64) *ssa.Value {
  5627  		v := s.entryNewValue1A(ssa.OpAddr, t, lsym, s.sb)
  5628  		// TODO: Make OpAddr use AuxInt as well as Aux.
  5629  		if offset != 0 {
  5630  			v = s.entryNewValue1I(ssa.OpOffPtr, v.Type, offset, v)
  5631  		}
  5632  		return v
  5633  	}
  5634  	switch n.Op() {
  5635  	case ir.OLINKSYMOFFSET:
  5636  		no := n.(*ir.LinksymOffsetExpr)
  5637  		return linksymOffset(no.Linksym, no.Offset_)
  5638  	case ir.ONAME:
  5639  		n := n.(*ir.Name)
  5640  		if n.Heapaddr != nil {
  5641  			return s.expr(n.Heapaddr)
  5642  		}
  5643  		switch n.Class {
  5644  		case ir.PEXTERN:
  5645  			// global variable
  5646  			return linksymOffset(n.Linksym(), 0)
  5647  		case ir.PPARAM:
  5648  			// parameter slot
  5649  			v := s.decladdrs[n]
  5650  			if v != nil {
  5651  				return v
  5652  			}
  5653  			s.Fatalf("addr of undeclared ONAME %v. declared: %v", n, s.decladdrs)
  5654  			return nil
  5655  		case ir.PAUTO:
  5656  			return s.newValue2Apos(ssa.OpLocalAddr, t, n, s.sp, s.mem(), !ir.IsAutoTmp(n))
  5657  
  5658  		case ir.PPARAMOUT: // Same as PAUTO -- cannot generate LEA early.
  5659  			// ensure that we reuse symbols for out parameters so
  5660  			// that cse works on their addresses
  5661  			return s.newValue2Apos(ssa.OpLocalAddr, t, n, s.sp, s.mem(), true)
  5662  		default:
  5663  			s.Fatalf("variable address class %v not implemented", n.Class)
  5664  			return nil
  5665  		}
  5666  	case ir.ORESULT:
  5667  		// load return from callee
  5668  		n := n.(*ir.ResultExpr)
  5669  		return s.resultAddrOfCall(s.prevCall, n.Index, n.Type())
  5670  	case ir.OINDEX:
  5671  		n := n.(*ir.IndexExpr)
  5672  		if n.X.Type().IsSlice() {
  5673  			a := s.expr(n.X)
  5674  			i := s.expr(n.Index)
  5675  			len := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], a)
  5676  			i = s.boundsCheck(i, len, ssa.BoundsIndex, n.Bounded())
  5677  			p := s.newValue1(ssa.OpSlicePtr, t, a)
  5678  			return s.newValue2(ssa.OpPtrIndex, t, p, i)
  5679  		} else { // array
  5680  			a := s.addr(n.X)
  5681  			i := s.expr(n.Index)
  5682  			len := s.constInt(types.Types[types.TINT], n.X.Type().NumElem())
  5683  			i = s.boundsCheck(i, len, ssa.BoundsIndex, n.Bounded())
  5684  			return s.newValue2(ssa.OpPtrIndex, types.NewPtr(n.X.Type().Elem()), a, i)
  5685  		}
  5686  	case ir.ODEREF:
  5687  		n := n.(*ir.StarExpr)
  5688  		return s.exprPtr(n.X, n.Bounded(), n.Pos())
  5689  	case ir.ODOT:
  5690  		n := n.(*ir.SelectorExpr)
  5691  		p := s.addr(n.X)
  5692  		return s.newValue1I(ssa.OpOffPtr, t, n.Offset(), p)
  5693  	case ir.ODOTPTR:
  5694  		n := n.(*ir.SelectorExpr)
  5695  		p := s.exprPtr(n.X, n.Bounded(), n.Pos())
  5696  		return s.newValue1I(ssa.OpOffPtr, t, n.Offset(), p)
  5697  	case ir.OCONVNOP:
  5698  		n := n.(*ir.ConvExpr)
  5699  		if n.Type() == n.X.Type() {
  5700  			return s.addr(n.X)
  5701  		}
  5702  		addr := s.addr(n.X)
  5703  		return s.newValue1(ssa.OpCopy, t, addr) // ensure that addr has the right type
  5704  	case ir.OCALLFUNC, ir.OCALLINTER:
  5705  		n := n.(*ir.CallExpr)
  5706  		return s.callAddr(n, callNormal)
  5707  	case ir.ODOTTYPE, ir.ODYNAMICDOTTYPE:
  5708  		var v *ssa.Value
  5709  		if n.Op() == ir.ODOTTYPE {
  5710  			v, _ = s.dottype(n.(*ir.TypeAssertExpr), false)
  5711  		} else {
  5712  			v, _ = s.dynamicDottype(n.(*ir.DynamicTypeAssertExpr), false)
  5713  		}
  5714  		if v.Op != ssa.OpLoad {
  5715  			s.Fatalf("dottype of non-load")
  5716  		}
  5717  		if v.Args[1] != s.mem() {
  5718  			s.Fatalf("memory no longer live from dottype load")
  5719  		}
  5720  		return v.Args[0]
  5721  	default:
  5722  		s.Fatalf("unhandled addr %v", n.Op())
  5723  		return nil
  5724  	}
  5725  }
  5726  
  5727  // canSSA reports whether n is SSA-able.
  5728  // n must be an ONAME (or an ODOT sequence with an ONAME base).
  5729  func (s *state) canSSA(n ir.Node) bool {
  5730  	if base.Flag.N != 0 {
  5731  		return false
  5732  	}
  5733  	for {
  5734  		nn := n
  5735  		if nn.Op() == ir.ODOT {
  5736  			nn := nn.(*ir.SelectorExpr)
  5737  			n = nn.X
  5738  			continue
  5739  		}
  5740  		if nn.Op() == ir.OINDEX {
  5741  			nn := nn.(*ir.IndexExpr)
  5742  			if nn.X.Type().IsArray() {
  5743  				n = nn.X
  5744  				continue
  5745  			}
  5746  		}
  5747  		break
  5748  	}
  5749  	if n.Op() != ir.ONAME {
  5750  		return false
  5751  	}
  5752  	return s.canSSAName(n.(*ir.Name)) && ssa.CanSSA(n.Type())
  5753  }
  5754  
  5755  func (s *state) canSSAName(name *ir.Name) bool {
  5756  	if name.Addrtaken() || !name.OnStack() {
  5757  		return false
  5758  	}
  5759  	switch name.Class {
  5760  	case ir.PPARAMOUT:
  5761  		if s.hasdefer {
  5762  			// TODO: handle this case? Named return values must be
  5763  			// in memory so that the deferred function can see them.
  5764  			// Maybe do: if !strings.HasPrefix(n.String(), "~") { return false }
  5765  			// Or maybe not, see issue 18860.  Even unnamed return values
  5766  			// must be written back so if a defer recovers, the caller can see them.
  5767  			return false
  5768  		}
  5769  		if s.cgoUnsafeArgs {
  5770  			// Cgo effectively takes the address of all result args,
  5771  			// but the compiler can't see that.
  5772  			return false
  5773  		}
  5774  	}
  5775  	return true
  5776  	// TODO: try to make more variables SSAable?
  5777  }
  5778  
  5779  // exprPtr evaluates n to a pointer and nil-checks it.
  5780  func (s *state) exprPtr(n ir.Node, bounded bool, lineno src.XPos) *ssa.Value {
  5781  	p := s.expr(n)
  5782  	if bounded || n.NonNil() {
  5783  		if s.f.Frontend().Debug_checknil() && lineno.Line() > 1 {
  5784  			s.f.Warnl(lineno, "removed nil check")
  5785  		}
  5786  		return p
  5787  	}
  5788  	p = s.nilCheck(p)
  5789  	return p
  5790  }
  5791  
  5792  // nilCheck generates nil pointer checking code.
  5793  // Used only for automatically inserted nil checks,
  5794  // not for user code like 'x != nil'.
  5795  // Returns a "definitely not nil" copy of x to ensure proper ordering
  5796  // of the uses of the post-nilcheck pointer.
  5797  func (s *state) nilCheck(ptr *ssa.Value) *ssa.Value {
  5798  	if base.Debug.DisableNil != 0 || s.curfn.NilCheckDisabled() {
  5799  		return ptr
  5800  	}
  5801  	return s.newValue2(ssa.OpNilCheck, ptr.Type, ptr, s.mem())
  5802  }
  5803  
  5804  // boundsCheck generates bounds checking code. Checks if 0 <= idx <[=] len, branches to exit if not.
  5805  // Starts a new block on return.
  5806  // On input, len must be converted to full int width and be nonnegative.
  5807  // Returns idx converted to full int width.
  5808  // If bounded is true then caller guarantees the index is not out of bounds
  5809  // (but boundsCheck will still extend the index to full int width).
  5810  func (s *state) boundsCheck(idx, len *ssa.Value, kind ssa.BoundsKind, bounded bool) *ssa.Value {
  5811  	idx = s.extendIndex(idx, len, kind, bounded)
  5812  
  5813  	if bounded || base.Flag.B != 0 {
  5814  		// If bounded or bounds checking is flag-disabled, then no check necessary,
  5815  		// just return the extended index.
  5816  		//
  5817  		// Here, bounded == true if the compiler generated the index itself,
  5818  		// such as in the expansion of a slice initializer. These indexes are
  5819  		// compiler-generated, not Go program variables, so they cannot be
  5820  		// attacker-controlled, so we can omit Spectre masking as well.
  5821  		//
  5822  		// Note that we do not want to omit Spectre masking in code like:
  5823  		//
  5824  		//	if 0 <= i && i < len(x) {
  5825  		//		use(x[i])
  5826  		//	}
  5827  		//
  5828  		// Lucky for us, bounded==false for that code.
  5829  		// In that case (handled below), we emit a bound check (and Spectre mask)
  5830  		// and then the prove pass will remove the bounds check.
  5831  		// In theory the prove pass could potentially remove certain
  5832  		// Spectre masks, but it's very delicate and probably better
  5833  		// to be conservative and leave them all in.
  5834  		return idx
  5835  	}
  5836  
  5837  	bNext := s.f.NewBlock(ssa.BlockPlain)
  5838  	bPanic := s.f.NewBlock(ssa.BlockExit)
  5839  
  5840  	if !idx.Type.IsSigned() {
  5841  		switch kind {
  5842  		case ssa.BoundsIndex:
  5843  			kind = ssa.BoundsIndexU
  5844  		case ssa.BoundsSliceAlen:
  5845  			kind = ssa.BoundsSliceAlenU
  5846  		case ssa.BoundsSliceAcap:
  5847  			kind = ssa.BoundsSliceAcapU
  5848  		case ssa.BoundsSliceB:
  5849  			kind = ssa.BoundsSliceBU
  5850  		case ssa.BoundsSlice3Alen:
  5851  			kind = ssa.BoundsSlice3AlenU
  5852  		case ssa.BoundsSlice3Acap:
  5853  			kind = ssa.BoundsSlice3AcapU
  5854  		case ssa.BoundsSlice3B:
  5855  			kind = ssa.BoundsSlice3BU
  5856  		case ssa.BoundsSlice3C:
  5857  			kind = ssa.BoundsSlice3CU
  5858  		}
  5859  	}
  5860  
  5861  	var cmp *ssa.Value
  5862  	if kind == ssa.BoundsIndex || kind == ssa.BoundsIndexU {
  5863  		cmp = s.newValue2(ssa.OpIsInBounds, types.Types[types.TBOOL], idx, len)
  5864  	} else {
  5865  		cmp = s.newValue2(ssa.OpIsSliceInBounds, types.Types[types.TBOOL], idx, len)
  5866  	}
  5867  	b := s.endBlock()
  5868  	b.Kind = ssa.BlockIf
  5869  	b.SetControl(cmp)
  5870  	b.Likely = ssa.BranchLikely
  5871  	b.AddEdgeTo(bNext)
  5872  	b.AddEdgeTo(bPanic)
  5873  
  5874  	s.startBlock(bPanic)
  5875  	if Arch.LinkArch.Family == sys.Wasm {
  5876  		// TODO(khr): figure out how to do "register" based calling convention for bounds checks.
  5877  		// Should be similar to gcWriteBarrier, but I can't make it work.
  5878  		s.rtcall(BoundsCheckFunc[kind], false, nil, idx, len)
  5879  	} else {
  5880  		mem := s.newValue3I(ssa.OpPanicBounds, types.TypeMem, int64(kind), idx, len, s.mem())
  5881  		s.endBlock().SetControl(mem)
  5882  	}
  5883  	s.startBlock(bNext)
  5884  
  5885  	// In Spectre index mode, apply an appropriate mask to avoid speculative out-of-bounds accesses.
  5886  	if base.Flag.Cfg.SpectreIndex {
  5887  		op := ssa.OpSpectreIndex
  5888  		if kind != ssa.BoundsIndex && kind != ssa.BoundsIndexU {
  5889  			op = ssa.OpSpectreSliceIndex
  5890  		}
  5891  		idx = s.newValue2(op, types.Types[types.TINT], idx, len)
  5892  	}
  5893  
  5894  	return idx
  5895  }
  5896  
  5897  // If cmp (a bool) is false, panic using the given function.
  5898  func (s *state) check(cmp *ssa.Value, fn *obj.LSym) {
  5899  	b := s.endBlock()
  5900  	b.Kind = ssa.BlockIf
  5901  	b.SetControl(cmp)
  5902  	b.Likely = ssa.BranchLikely
  5903  	bNext := s.f.NewBlock(ssa.BlockPlain)
  5904  	line := s.peekPos()
  5905  	pos := base.Ctxt.PosTable.Pos(line)
  5906  	fl := funcLine{f: fn, base: pos.Base(), line: pos.Line()}
  5907  	bPanic := s.panics[fl]
  5908  	if bPanic == nil {
  5909  		bPanic = s.f.NewBlock(ssa.BlockPlain)
  5910  		s.panics[fl] = bPanic
  5911  		s.startBlock(bPanic)
  5912  		// The panic call takes/returns memory to ensure that the right
  5913  		// memory state is observed if the panic happens.
  5914  		s.rtcall(fn, false, nil)
  5915  	}
  5916  	b.AddEdgeTo(bNext)
  5917  	b.AddEdgeTo(bPanic)
  5918  	s.startBlock(bNext)
  5919  }
  5920  
  5921  func (s *state) intDivide(n ir.Node, a, b *ssa.Value) *ssa.Value {
  5922  	needcheck := true
  5923  	switch b.Op {
  5924  	case ssa.OpConst8, ssa.OpConst16, ssa.OpConst32, ssa.OpConst64:
  5925  		if b.AuxInt != 0 {
  5926  			needcheck = false
  5927  		}
  5928  	}
  5929  	if needcheck {
  5930  		// do a size-appropriate check for zero
  5931  		cmp := s.newValue2(s.ssaOp(ir.ONE, n.Type()), types.Types[types.TBOOL], b, s.zeroVal(n.Type()))
  5932  		s.check(cmp, ir.Syms.Panicdivide)
  5933  	}
  5934  	return s.newValue2(s.ssaOp(n.Op(), n.Type()), a.Type, a, b)
  5935  }
  5936  
  5937  // rtcall issues a call to the given runtime function fn with the listed args.
  5938  // Returns a slice of results of the given result types.
  5939  // The call is added to the end of the current block.
  5940  // If returns is false, the block is marked as an exit block.
  5941  func (s *state) rtcall(fn *obj.LSym, returns bool, results []*types.Type, args ...*ssa.Value) []*ssa.Value {
  5942  	s.prevCall = nil
  5943  	// Write args to the stack
  5944  	off := base.Ctxt.Arch.FixedFrameSize
  5945  	var callArgs []*ssa.Value
  5946  	var callArgTypes []*types.Type
  5947  
  5948  	for _, arg := range args {
  5949  		t := arg.Type
  5950  		off = types.RoundUp(off, t.Alignment())
  5951  		size := t.Size()
  5952  		callArgs = append(callArgs, arg)
  5953  		callArgTypes = append(callArgTypes, t)
  5954  		off += size
  5955  	}
  5956  	off = types.RoundUp(off, int64(types.RegSize))
  5957  
  5958  	// Issue call
  5959  	var call *ssa.Value
  5960  	aux := ssa.StaticAuxCall(fn, s.f.ABIDefault.ABIAnalyzeTypes(callArgTypes, results))
  5961  	callArgs = append(callArgs, s.mem())
  5962  	call = s.newValue0A(ssa.OpStaticLECall, aux.LateExpansionResultType(), aux)
  5963  	call.AddArgs(callArgs...)
  5964  	s.vars[memVar] = s.newValue1I(ssa.OpSelectN, types.TypeMem, int64(len(results)), call)
  5965  
  5966  	if !returns {
  5967  		// Finish block
  5968  		b := s.endBlock()
  5969  		b.Kind = ssa.BlockExit
  5970  		b.SetControl(call)
  5971  		call.AuxInt = off - base.Ctxt.Arch.FixedFrameSize
  5972  		if len(results) > 0 {
  5973  			s.Fatalf("panic call can't have results")
  5974  		}
  5975  		return nil
  5976  	}
  5977  
  5978  	// Load results
  5979  	res := make([]*ssa.Value, len(results))
  5980  	for i, t := range results {
  5981  		off = types.RoundUp(off, t.Alignment())
  5982  		res[i] = s.resultOfCall(call, int64(i), t)
  5983  		off += t.Size()
  5984  	}
  5985  	off = types.RoundUp(off, int64(types.PtrSize))
  5986  
  5987  	// Remember how much callee stack space we needed.
  5988  	call.AuxInt = off
  5989  
  5990  	return res
  5991  }
  5992  
  5993  // do *left = right for type t.
  5994  func (s *state) storeType(t *types.Type, left, right *ssa.Value, skip skipMask, leftIsStmt bool) {
  5995  	s.instrument(t, left, instrumentWrite)
  5996  
  5997  	if skip == 0 && (!t.HasPointers() || ssa.IsStackAddr(left)) {
  5998  		// Known to not have write barrier. Store the whole type.
  5999  		s.vars[memVar] = s.newValue3Apos(ssa.OpStore, types.TypeMem, t, left, right, s.mem(), leftIsStmt)
  6000  		return
  6001  	}
  6002  
  6003  	// store scalar fields first, so write barrier stores for
  6004  	// pointer fields can be grouped together, and scalar values
  6005  	// don't need to be live across the write barrier call.
  6006  	// TODO: if the writebarrier pass knows how to reorder stores,
  6007  	// we can do a single store here as long as skip==0.
  6008  	s.storeTypeScalars(t, left, right, skip)
  6009  	if skip&skipPtr == 0 && t.HasPointers() {
  6010  		s.storeTypePtrs(t, left, right)
  6011  	}
  6012  }
  6013  
  6014  // do *left = right for all scalar (non-pointer) parts of t.
  6015  func (s *state) storeTypeScalars(t *types.Type, left, right *ssa.Value, skip skipMask) {
  6016  	switch {
  6017  	case t.IsBoolean() || t.IsInteger() || t.IsFloat() || t.IsComplex():
  6018  		s.store(t, left, right)
  6019  	case t.IsPtrShaped():
  6020  		if t.IsPtr() && t.Elem().NotInHeap() {
  6021  			s.store(t, left, right) // see issue 42032
  6022  		}
  6023  		// otherwise, no scalar fields.
  6024  	case t.IsString():
  6025  		if skip&skipLen != 0 {
  6026  			return
  6027  		}
  6028  		len := s.newValue1(ssa.OpStringLen, types.Types[types.TINT], right)
  6029  		lenAddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, s.config.PtrSize, left)
  6030  		s.store(types.Types[types.TINT], lenAddr, len)
  6031  	case t.IsSlice():
  6032  		if skip&skipLen == 0 {
  6033  			len := s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], right)
  6034  			lenAddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, s.config.PtrSize, left)
  6035  			s.store(types.Types[types.TINT], lenAddr, len)
  6036  		}
  6037  		if skip&skipCap == 0 {
  6038  			cap := s.newValue1(ssa.OpSliceCap, types.Types[types.TINT], right)
  6039  			capAddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.IntPtr, 2*s.config.PtrSize, left)
  6040  			s.store(types.Types[types.TINT], capAddr, cap)
  6041  		}
  6042  	case t.IsInterface():
  6043  		// itab field doesn't need a write barrier (even though it is a pointer).
  6044  		itab := s.newValue1(ssa.OpITab, s.f.Config.Types.BytePtr, right)
  6045  		s.store(types.Types[types.TUINTPTR], left, itab)
  6046  	case t.IsStruct():
  6047  		n := t.NumFields()
  6048  		for i := 0; i < n; i++ {
  6049  			ft := t.FieldType(i)
  6050  			addr := s.newValue1I(ssa.OpOffPtr, ft.PtrTo(), t.FieldOff(i), left)
  6051  			val := s.newValue1I(ssa.OpStructSelect, ft, int64(i), right)
  6052  			s.storeTypeScalars(ft, addr, val, 0)
  6053  		}
  6054  	case t.IsArray() && t.NumElem() == 0:
  6055  		// nothing
  6056  	case t.IsArray() && t.NumElem() == 1:
  6057  		s.storeTypeScalars(t.Elem(), left, s.newValue1I(ssa.OpArraySelect, t.Elem(), 0, right), 0)
  6058  	default:
  6059  		s.Fatalf("bad write barrier type %v", t)
  6060  	}
  6061  }
  6062  
  6063  // do *left = right for all pointer parts of t.
  6064  func (s *state) storeTypePtrs(t *types.Type, left, right *ssa.Value) {
  6065  	switch {
  6066  	case t.IsPtrShaped():
  6067  		if t.IsPtr() && t.Elem().NotInHeap() {
  6068  			break // see issue 42032
  6069  		}
  6070  		s.store(t, left, right)
  6071  	case t.IsString():
  6072  		ptr := s.newValue1(ssa.OpStringPtr, s.f.Config.Types.BytePtr, right)
  6073  		s.store(s.f.Config.Types.BytePtr, left, ptr)
  6074  	case t.IsSlice():
  6075  		elType := types.NewPtr(t.Elem())
  6076  		ptr := s.newValue1(ssa.OpSlicePtr, elType, right)
  6077  		s.store(elType, left, ptr)
  6078  	case t.IsInterface():
  6079  		// itab field is treated as a scalar.
  6080  		idata := s.newValue1(ssa.OpIData, s.f.Config.Types.BytePtr, right)
  6081  		idataAddr := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.BytePtrPtr, s.config.PtrSize, left)
  6082  		s.store(s.f.Config.Types.BytePtr, idataAddr, idata)
  6083  	case t.IsStruct():
  6084  		n := t.NumFields()
  6085  		for i := 0; i < n; i++ {
  6086  			ft := t.FieldType(i)
  6087  			if !ft.HasPointers() {
  6088  				continue
  6089  			}
  6090  			addr := s.newValue1I(ssa.OpOffPtr, ft.PtrTo(), t.FieldOff(i), left)
  6091  			val := s.newValue1I(ssa.OpStructSelect, ft, int64(i), right)
  6092  			s.storeTypePtrs(ft, addr, val)
  6093  		}
  6094  	case t.IsArray() && t.NumElem() == 0:
  6095  		// nothing
  6096  	case t.IsArray() && t.NumElem() == 1:
  6097  		s.storeTypePtrs(t.Elem(), left, s.newValue1I(ssa.OpArraySelect, t.Elem(), 0, right))
  6098  	default:
  6099  		s.Fatalf("bad write barrier type %v", t)
  6100  	}
  6101  }
  6102  
  6103  // putArg evaluates n for the purpose of passing it as an argument to a function and returns the value for the call.
  6104  func (s *state) putArg(n ir.Node, t *types.Type) *ssa.Value {
  6105  	var a *ssa.Value
  6106  	if !ssa.CanSSA(t) {
  6107  		a = s.newValue2(ssa.OpDereference, t, s.addr(n), s.mem())
  6108  	} else {
  6109  		a = s.expr(n)
  6110  	}
  6111  	return a
  6112  }
  6113  
  6114  func (s *state) storeArgWithBase(n ir.Node, t *types.Type, base *ssa.Value, off int64) {
  6115  	pt := types.NewPtr(t)
  6116  	var addr *ssa.Value
  6117  	if base == s.sp {
  6118  		// Use special routine that avoids allocation on duplicate offsets.
  6119  		addr = s.constOffPtrSP(pt, off)
  6120  	} else {
  6121  		addr = s.newValue1I(ssa.OpOffPtr, pt, off, base)
  6122  	}
  6123  
  6124  	if !ssa.CanSSA(t) {
  6125  		a := s.addr(n)
  6126  		s.move(t, addr, a)
  6127  		return
  6128  	}
  6129  
  6130  	a := s.expr(n)
  6131  	s.storeType(t, addr, a, 0, false)
  6132  }
  6133  
  6134  // slice computes the slice v[i:j:k] and returns ptr, len, and cap of result.
  6135  // i,j,k may be nil, in which case they are set to their default value.
  6136  // v may be a slice, string or pointer to an array.
  6137  func (s *state) slice(v, i, j, k *ssa.Value, bounded bool) (p, l, c *ssa.Value) {
  6138  	t := v.Type
  6139  	var ptr, len, cap *ssa.Value
  6140  	switch {
  6141  	case t.IsSlice():
  6142  		ptr = s.newValue1(ssa.OpSlicePtr, types.NewPtr(t.Elem()), v)
  6143  		len = s.newValue1(ssa.OpSliceLen, types.Types[types.TINT], v)
  6144  		cap = s.newValue1(ssa.OpSliceCap, types.Types[types.TINT], v)
  6145  	case t.IsString():
  6146  		ptr = s.newValue1(ssa.OpStringPtr, types.NewPtr(types.Types[types.TUINT8]), v)
  6147  		len = s.newValue1(ssa.OpStringLen, types.Types[types.TINT], v)
  6148  		cap = len
  6149  	case t.IsPtr():
  6150  		if !t.Elem().IsArray() {
  6151  			s.Fatalf("bad ptr to array in slice %v\n", t)
  6152  		}
  6153  		nv := s.nilCheck(v)
  6154  		ptr = s.newValue1(ssa.OpCopy, types.NewPtr(t.Elem().Elem()), nv)
  6155  		len = s.constInt(types.Types[types.TINT], t.Elem().NumElem())
  6156  		cap = len
  6157  	default:
  6158  		s.Fatalf("bad type in slice %v\n", t)
  6159  	}
  6160  
  6161  	// Set default values
  6162  	if i == nil {
  6163  		i = s.constInt(types.Types[types.TINT], 0)
  6164  	}
  6165  	if j == nil {
  6166  		j = len
  6167  	}
  6168  	three := true
  6169  	if k == nil {
  6170  		three = false
  6171  		k = cap
  6172  	}
  6173  
  6174  	// Panic if slice indices are not in bounds.
  6175  	// Make sure we check these in reverse order so that we're always
  6176  	// comparing against a value known to be nonnegative. See issue 28797.
  6177  	if three {
  6178  		if k != cap {
  6179  			kind := ssa.BoundsSlice3Alen
  6180  			if t.IsSlice() {
  6181  				kind = ssa.BoundsSlice3Acap
  6182  			}
  6183  			k = s.boundsCheck(k, cap, kind, bounded)
  6184  		}
  6185  		if j != k {
  6186  			j = s.boundsCheck(j, k, ssa.BoundsSlice3B, bounded)
  6187  		}
  6188  		i = s.boundsCheck(i, j, ssa.BoundsSlice3C, bounded)
  6189  	} else {
  6190  		if j != k {
  6191  			kind := ssa.BoundsSliceAlen
  6192  			if t.IsSlice() {
  6193  				kind = ssa.BoundsSliceAcap
  6194  			}
  6195  			j = s.boundsCheck(j, k, kind, bounded)
  6196  		}
  6197  		i = s.boundsCheck(i, j, ssa.BoundsSliceB, bounded)
  6198  	}
  6199  
  6200  	// Word-sized integer operations.
  6201  	subOp := s.ssaOp(ir.OSUB, types.Types[types.TINT])
  6202  	mulOp := s.ssaOp(ir.OMUL, types.Types[types.TINT])
  6203  	andOp := s.ssaOp(ir.OAND, types.Types[types.TINT])
  6204  
  6205  	// Calculate the length (rlen) and capacity (rcap) of the new slice.
  6206  	// For strings the capacity of the result is unimportant. However,
  6207  	// we use rcap to test if we've generated a zero-length slice.
  6208  	// Use length of strings for that.
  6209  	rlen := s.newValue2(subOp, types.Types[types.TINT], j, i)
  6210  	rcap := rlen
  6211  	if j != k && !t.IsString() {
  6212  		rcap = s.newValue2(subOp, types.Types[types.TINT], k, i)
  6213  	}
  6214  
  6215  	if (i.Op == ssa.OpConst64 || i.Op == ssa.OpConst32) && i.AuxInt == 0 {
  6216  		// No pointer arithmetic necessary.
  6217  		return ptr, rlen, rcap
  6218  	}
  6219  
  6220  	// Calculate the base pointer (rptr) for the new slice.
  6221  	//
  6222  	// Generate the following code assuming that indexes are in bounds.
  6223  	// The masking is to make sure that we don't generate a slice
  6224  	// that points to the next object in memory. We cannot just set
  6225  	// the pointer to nil because then we would create a nil slice or
  6226  	// string.
  6227  	//
  6228  	//     rcap = k - i
  6229  	//     rlen = j - i
  6230  	//     rptr = ptr + (mask(rcap) & (i * stride))
  6231  	//
  6232  	// Where mask(x) is 0 if x==0 and -1 if x>0 and stride is the width
  6233  	// of the element type.
  6234  	stride := s.constInt(types.Types[types.TINT], ptr.Type.Elem().Size())
  6235  
  6236  	// The delta is the number of bytes to offset ptr by.
  6237  	delta := s.newValue2(mulOp, types.Types[types.TINT], i, stride)
  6238  
  6239  	// If we're slicing to the point where the capacity is zero,
  6240  	// zero out the delta.
  6241  	mask := s.newValue1(ssa.OpSlicemask, types.Types[types.TINT], rcap)
  6242  	delta = s.newValue2(andOp, types.Types[types.TINT], delta, mask)
  6243  
  6244  	// Compute rptr = ptr + delta.
  6245  	rptr := s.newValue2(ssa.OpAddPtr, ptr.Type, ptr, delta)
  6246  
  6247  	return rptr, rlen, rcap
  6248  }
  6249  
  6250  type u642fcvtTab struct {
  6251  	leq, cvt2F, and, rsh, or, add ssa.Op
  6252  	one                           func(*state, *types.Type, int64) *ssa.Value
  6253  }
  6254  
  6255  var u64_f64 = u642fcvtTab{
  6256  	leq:   ssa.OpLeq64,
  6257  	cvt2F: ssa.OpCvt64to64F,
  6258  	and:   ssa.OpAnd64,
  6259  	rsh:   ssa.OpRsh64Ux64,
  6260  	or:    ssa.OpOr64,
  6261  	add:   ssa.OpAdd64F,
  6262  	one:   (*state).constInt64,
  6263  }
  6264  
  6265  var u64_f32 = u642fcvtTab{
  6266  	leq:   ssa.OpLeq64,
  6267  	cvt2F: ssa.OpCvt64to32F,
  6268  	and:   ssa.OpAnd64,
  6269  	rsh:   ssa.OpRsh64Ux64,
  6270  	or:    ssa.OpOr64,
  6271  	add:   ssa.OpAdd32F,
  6272  	one:   (*state).constInt64,
  6273  }
  6274  
  6275  func (s *state) uint64Tofloat64(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
  6276  	return s.uint64Tofloat(&u64_f64, n, x, ft, tt)
  6277  }
  6278  
  6279  func (s *state) uint64Tofloat32(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
  6280  	return s.uint64Tofloat(&u64_f32, n, x, ft, tt)
  6281  }
  6282  
  6283  func (s *state) uint64Tofloat(cvttab *u642fcvtTab, n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
  6284  	// if x >= 0 {
  6285  	//    result = (floatY) x
  6286  	// } else {
  6287  	// 	  y = uintX(x) ; y = x & 1
  6288  	// 	  z = uintX(x) ; z = z >> 1
  6289  	// 	  z = z | y
  6290  	// 	  result = floatY(z)
  6291  	// 	  result = result + result
  6292  	// }
  6293  	//
  6294  	// Code borrowed from old code generator.
  6295  	// What's going on: large 64-bit "unsigned" looks like
  6296  	// negative number to hardware's integer-to-float
  6297  	// conversion. However, because the mantissa is only
  6298  	// 63 bits, we don't need the LSB, so instead we do an
  6299  	// unsigned right shift (divide by two), convert, and
  6300  	// double. However, before we do that, we need to be
  6301  	// sure that we do not lose a "1" if that made the
  6302  	// difference in the resulting rounding. Therefore, we
  6303  	// preserve it, and OR (not ADD) it back in. The case
  6304  	// that matters is when the eleven discarded bits are
  6305  	// equal to 10000000001; that rounds up, and the 1 cannot
  6306  	// be lost else it would round down if the LSB of the
  6307  	// candidate mantissa is 0.
  6308  	cmp := s.newValue2(cvttab.leq, types.Types[types.TBOOL], s.zeroVal(ft), x)
  6309  	b := s.endBlock()
  6310  	b.Kind = ssa.BlockIf
  6311  	b.SetControl(cmp)
  6312  	b.Likely = ssa.BranchLikely
  6313  
  6314  	bThen := s.f.NewBlock(ssa.BlockPlain)
  6315  	bElse := s.f.NewBlock(ssa.BlockPlain)
  6316  	bAfter := s.f.NewBlock(ssa.BlockPlain)
  6317  
  6318  	b.AddEdgeTo(bThen)
  6319  	s.startBlock(bThen)
  6320  	a0 := s.newValue1(cvttab.cvt2F, tt, x)
  6321  	s.vars[n] = a0
  6322  	s.endBlock()
  6323  	bThen.AddEdgeTo(bAfter)
  6324  
  6325  	b.AddEdgeTo(bElse)
  6326  	s.startBlock(bElse)
  6327  	one := cvttab.one(s, ft, 1)
  6328  	y := s.newValue2(cvttab.and, ft, x, one)
  6329  	z := s.newValue2(cvttab.rsh, ft, x, one)
  6330  	z = s.newValue2(cvttab.or, ft, z, y)
  6331  	a := s.newValue1(cvttab.cvt2F, tt, z)
  6332  	a1 := s.newValue2(cvttab.add, tt, a, a)
  6333  	s.vars[n] = a1
  6334  	s.endBlock()
  6335  	bElse.AddEdgeTo(bAfter)
  6336  
  6337  	s.startBlock(bAfter)
  6338  	return s.variable(n, n.Type())
  6339  }
  6340  
  6341  type u322fcvtTab struct {
  6342  	cvtI2F, cvtF2F ssa.Op
  6343  }
  6344  
  6345  var u32_f64 = u322fcvtTab{
  6346  	cvtI2F: ssa.OpCvt32to64F,
  6347  	cvtF2F: ssa.OpCopy,
  6348  }
  6349  
  6350  var u32_f32 = u322fcvtTab{
  6351  	cvtI2F: ssa.OpCvt32to32F,
  6352  	cvtF2F: ssa.OpCvt64Fto32F,
  6353  }
  6354  
  6355  func (s *state) uint32Tofloat64(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
  6356  	return s.uint32Tofloat(&u32_f64, n, x, ft, tt)
  6357  }
  6358  
  6359  func (s *state) uint32Tofloat32(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
  6360  	return s.uint32Tofloat(&u32_f32, n, x, ft, tt)
  6361  }
  6362  
  6363  func (s *state) uint32Tofloat(cvttab *u322fcvtTab, n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
  6364  	// if x >= 0 {
  6365  	// 	result = floatY(x)
  6366  	// } else {
  6367  	// 	result = floatY(float64(x) + (1<<32))
  6368  	// }
  6369  	cmp := s.newValue2(ssa.OpLeq32, types.Types[types.TBOOL], s.zeroVal(ft), x)
  6370  	b := s.endBlock()
  6371  	b.Kind = ssa.BlockIf
  6372  	b.SetControl(cmp)
  6373  	b.Likely = ssa.BranchLikely
  6374  
  6375  	bThen := s.f.NewBlock(ssa.BlockPlain)
  6376  	bElse := s.f.NewBlock(ssa.BlockPlain)
  6377  	bAfter := s.f.NewBlock(ssa.BlockPlain)
  6378  
  6379  	b.AddEdgeTo(bThen)
  6380  	s.startBlock(bThen)
  6381  	a0 := s.newValue1(cvttab.cvtI2F, tt, x)
  6382  	s.vars[n] = a0
  6383  	s.endBlock()
  6384  	bThen.AddEdgeTo(bAfter)
  6385  
  6386  	b.AddEdgeTo(bElse)
  6387  	s.startBlock(bElse)
  6388  	a1 := s.newValue1(ssa.OpCvt32to64F, types.Types[types.TFLOAT64], x)
  6389  	twoToThe32 := s.constFloat64(types.Types[types.TFLOAT64], float64(1<<32))
  6390  	a2 := s.newValue2(ssa.OpAdd64F, types.Types[types.TFLOAT64], a1, twoToThe32)
  6391  	a3 := s.newValue1(cvttab.cvtF2F, tt, a2)
  6392  
  6393  	s.vars[n] = a3
  6394  	s.endBlock()
  6395  	bElse.AddEdgeTo(bAfter)
  6396  
  6397  	s.startBlock(bAfter)
  6398  	return s.variable(n, n.Type())
  6399  }
  6400  
  6401  // referenceTypeBuiltin generates code for the len/cap builtins for maps and channels.
  6402  func (s *state) referenceTypeBuiltin(n *ir.UnaryExpr, x *ssa.Value) *ssa.Value {
  6403  	if !n.X.Type().IsMap() && !n.X.Type().IsChan() {
  6404  		s.Fatalf("node must be a map or a channel")
  6405  	}
  6406  	if n.X.Type().IsChan() && n.Op() == ir.OLEN {
  6407  		s.Fatalf("cannot inline len(chan)") // must use runtime.chanlen now
  6408  	}
  6409  	if n.X.Type().IsChan() && n.Op() == ir.OCAP {
  6410  		s.Fatalf("cannot inline cap(chan)") // must use runtime.chancap now
  6411  	}
  6412  	// if n == nil {
  6413  	//   return 0
  6414  	// } else {
  6415  	//   // len
  6416  	//   return *((*int)n)
  6417  	//   // cap
  6418  	//   return *(((*int)n)+1)
  6419  	// }
  6420  	lenType := n.Type()
  6421  	nilValue := s.constNil(types.Types[types.TUINTPTR])
  6422  	cmp := s.newValue2(ssa.OpEqPtr, types.Types[types.TBOOL], x, nilValue)
  6423  	b := s.endBlock()
  6424  	b.Kind = ssa.BlockIf
  6425  	b.SetControl(cmp)
  6426  	b.Likely = ssa.BranchUnlikely
  6427  
  6428  	bThen := s.f.NewBlock(ssa.BlockPlain)
  6429  	bElse := s.f.NewBlock(ssa.BlockPlain)
  6430  	bAfter := s.f.NewBlock(ssa.BlockPlain)
  6431  
  6432  	// length/capacity of a nil map/chan is zero
  6433  	b.AddEdgeTo(bThen)
  6434  	s.startBlock(bThen)
  6435  	s.vars[n] = s.zeroVal(lenType)
  6436  	s.endBlock()
  6437  	bThen.AddEdgeTo(bAfter)
  6438  
  6439  	b.AddEdgeTo(bElse)
  6440  	s.startBlock(bElse)
  6441  	switch n.Op() {
  6442  	case ir.OLEN:
  6443  		// length is stored in the first word for map/chan
  6444  		s.vars[n] = s.load(lenType, x)
  6445  	case ir.OCAP:
  6446  		// capacity is stored in the second word for chan
  6447  		sw := s.newValue1I(ssa.OpOffPtr, lenType.PtrTo(), lenType.Size(), x)
  6448  		s.vars[n] = s.load(lenType, sw)
  6449  	default:
  6450  		s.Fatalf("op must be OLEN or OCAP")
  6451  	}
  6452  	s.endBlock()
  6453  	bElse.AddEdgeTo(bAfter)
  6454  
  6455  	s.startBlock(bAfter)
  6456  	return s.variable(n, lenType)
  6457  }
  6458  
  6459  type f2uCvtTab struct {
  6460  	ltf, cvt2U, subf, or ssa.Op
  6461  	floatValue           func(*state, *types.Type, float64) *ssa.Value
  6462  	intValue             func(*state, *types.Type, int64) *ssa.Value
  6463  	cutoff               uint64
  6464  }
  6465  
  6466  var f32_u64 = f2uCvtTab{
  6467  	ltf:        ssa.OpLess32F,
  6468  	cvt2U:      ssa.OpCvt32Fto64,
  6469  	subf:       ssa.OpSub32F,
  6470  	or:         ssa.OpOr64,
  6471  	floatValue: (*state).constFloat32,
  6472  	intValue:   (*state).constInt64,
  6473  	cutoff:     1 << 63,
  6474  }
  6475  
  6476  var f64_u64 = f2uCvtTab{
  6477  	ltf:        ssa.OpLess64F,
  6478  	cvt2U:      ssa.OpCvt64Fto64,
  6479  	subf:       ssa.OpSub64F,
  6480  	or:         ssa.OpOr64,
  6481  	floatValue: (*state).constFloat64,
  6482  	intValue:   (*state).constInt64,
  6483  	cutoff:     1 << 63,
  6484  }
  6485  
  6486  var f32_u32 = f2uCvtTab{
  6487  	ltf:        ssa.OpLess32F,
  6488  	cvt2U:      ssa.OpCvt32Fto32,
  6489  	subf:       ssa.OpSub32F,
  6490  	or:         ssa.OpOr32,
  6491  	floatValue: (*state).constFloat32,
  6492  	intValue:   func(s *state, t *types.Type, v int64) *ssa.Value { return s.constInt32(t, int32(v)) },
  6493  	cutoff:     1 << 31,
  6494  }
  6495  
  6496  var f64_u32 = f2uCvtTab{
  6497  	ltf:        ssa.OpLess64F,
  6498  	cvt2U:      ssa.OpCvt64Fto32,
  6499  	subf:       ssa.OpSub64F,
  6500  	or:         ssa.OpOr32,
  6501  	floatValue: (*state).constFloat64,
  6502  	intValue:   func(s *state, t *types.Type, v int64) *ssa.Value { return s.constInt32(t, int32(v)) },
  6503  	cutoff:     1 << 31,
  6504  }
  6505  
  6506  func (s *state) float32ToUint64(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
  6507  	return s.floatToUint(&f32_u64, n, x, ft, tt)
  6508  }
  6509  func (s *state) float64ToUint64(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
  6510  	return s.floatToUint(&f64_u64, n, x, ft, tt)
  6511  }
  6512  
  6513  func (s *state) float32ToUint32(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
  6514  	return s.floatToUint(&f32_u32, n, x, ft, tt)
  6515  }
  6516  
  6517  func (s *state) float64ToUint32(n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
  6518  	return s.floatToUint(&f64_u32, n, x, ft, tt)
  6519  }
  6520  
  6521  func (s *state) floatToUint(cvttab *f2uCvtTab, n ir.Node, x *ssa.Value, ft, tt *types.Type) *ssa.Value {
  6522  	// cutoff:=1<<(intY_Size-1)
  6523  	// if x < floatX(cutoff) {
  6524  	// 	result = uintY(x)
  6525  	// } else {
  6526  	// 	y = x - floatX(cutoff)
  6527  	// 	z = uintY(y)
  6528  	// 	result = z | -(cutoff)
  6529  	// }
  6530  	cutoff := cvttab.floatValue(s, ft, float64(cvttab.cutoff))
  6531  	cmp := s.newValue2(cvttab.ltf, types.Types[types.TBOOL], x, cutoff)
  6532  	b := s.endBlock()
  6533  	b.Kind = ssa.BlockIf
  6534  	b.SetControl(cmp)
  6535  	b.Likely = ssa.BranchLikely
  6536  
  6537  	bThen := s.f.NewBlock(ssa.BlockPlain)
  6538  	bElse := s.f.NewBlock(ssa.BlockPlain)
  6539  	bAfter := s.f.NewBlock(ssa.BlockPlain)
  6540  
  6541  	b.AddEdgeTo(bThen)
  6542  	s.startBlock(bThen)
  6543  	a0 := s.newValue1(cvttab.cvt2U, tt, x)
  6544  	s.vars[n] = a0
  6545  	s.endBlock()
  6546  	bThen.AddEdgeTo(bAfter)
  6547  
  6548  	b.AddEdgeTo(bElse)
  6549  	s.startBlock(bElse)
  6550  	y := s.newValue2(cvttab.subf, ft, x, cutoff)
  6551  	y = s.newValue1(cvttab.cvt2U, tt, y)
  6552  	z := cvttab.intValue(s, tt, int64(-cvttab.cutoff))
  6553  	a1 := s.newValue2(cvttab.or, tt, y, z)
  6554  	s.vars[n] = a1
  6555  	s.endBlock()
  6556  	bElse.AddEdgeTo(bAfter)
  6557  
  6558  	s.startBlock(bAfter)
  6559  	return s.variable(n, n.Type())
  6560  }
  6561  
  6562  // dottype generates SSA for a type assertion node.
  6563  // commaok indicates whether to panic or return a bool.
  6564  // If commaok is false, resok will be nil.
  6565  func (s *state) dottype(n *ir.TypeAssertExpr, commaok bool) (res, resok *ssa.Value) {
  6566  	iface := s.expr(n.X)              // input interface
  6567  	target := s.reflectType(n.Type()) // target type
  6568  	var targetItab *ssa.Value
  6569  	if n.ITab != nil {
  6570  		targetItab = s.expr(n.ITab)
  6571  	}
  6572  	return s.dottype1(n.Pos(), n.X.Type(), n.Type(), iface, nil, target, targetItab, commaok, n.Descriptor)
  6573  }
  6574  
  6575  func (s *state) dynamicDottype(n *ir.DynamicTypeAssertExpr, commaok bool) (res, resok *ssa.Value) {
  6576  	iface := s.expr(n.X)
  6577  	var source, target, targetItab *ssa.Value
  6578  	if n.SrcRType != nil {
  6579  		source = s.expr(n.SrcRType)
  6580  	}
  6581  	if !n.X.Type().IsEmptyInterface() && !n.Type().IsInterface() {
  6582  		byteptr := s.f.Config.Types.BytePtr
  6583  		targetItab = s.expr(n.ITab)
  6584  		// TODO(mdempsky): Investigate whether compiling n.RType could be
  6585  		// better than loading itab.typ.
  6586  		target = s.load(byteptr, s.newValue1I(ssa.OpOffPtr, byteptr, rttype.ITab.OffsetOf("Type"), targetItab))
  6587  	} else {
  6588  		target = s.expr(n.RType)
  6589  	}
  6590  	return s.dottype1(n.Pos(), n.X.Type(), n.Type(), iface, source, target, targetItab, commaok, nil)
  6591  }
  6592  
  6593  // dottype1 implements a x.(T) operation. iface is the argument (x), dst is the type we're asserting to (T)
  6594  // and src is the type we're asserting from.
  6595  // source is the *runtime._type of src
  6596  // target is the *runtime._type of dst.
  6597  // If src is a nonempty interface and dst is not an interface, targetItab is an itab representing (dst, src). Otherwise it is nil.
  6598  // commaok is true if the caller wants a boolean success value. Otherwise, the generated code panics if the conversion fails.
  6599  // descriptor is a compiler-allocated internal/abi.TypeAssert whose address is passed to runtime.typeAssert when
  6600  // the target type is a compile-time-known non-empty interface. It may be nil.
  6601  func (s *state) dottype1(pos src.XPos, src, dst *types.Type, iface, source, target, targetItab *ssa.Value, commaok bool, descriptor *obj.LSym) (res, resok *ssa.Value) {
  6602  	typs := s.f.Config.Types
  6603  	byteptr := typs.BytePtr
  6604  	if dst.IsInterface() {
  6605  		if dst.IsEmptyInterface() {
  6606  			// Converting to an empty interface.
  6607  			// Input could be an empty or nonempty interface.
  6608  			if base.Debug.TypeAssert > 0 {
  6609  				base.WarnfAt(pos, "type assertion inlined")
  6610  			}
  6611  
  6612  			// Get itab/type field from input.
  6613  			itab := s.newValue1(ssa.OpITab, byteptr, iface)
  6614  			// Conversion succeeds iff that field is not nil.
  6615  			cond := s.newValue2(ssa.OpNeqPtr, types.Types[types.TBOOL], itab, s.constNil(byteptr))
  6616  
  6617  			if src.IsEmptyInterface() && commaok {
  6618  				// Converting empty interface to empty interface with ,ok is just a nil check.
  6619  				return iface, cond
  6620  			}
  6621  
  6622  			// Branch on nilness.
  6623  			b := s.endBlock()
  6624  			b.Kind = ssa.BlockIf
  6625  			b.SetControl(cond)
  6626  			b.Likely = ssa.BranchLikely
  6627  			bOk := s.f.NewBlock(ssa.BlockPlain)
  6628  			bFail := s.f.NewBlock(ssa.BlockPlain)
  6629  			b.AddEdgeTo(bOk)
  6630  			b.AddEdgeTo(bFail)
  6631  
  6632  			if !commaok {
  6633  				// On failure, panic by calling panicnildottype.
  6634  				s.startBlock(bFail)
  6635  				s.rtcall(ir.Syms.Panicnildottype, false, nil, target)
  6636  
  6637  				// On success, return (perhaps modified) input interface.
  6638  				s.startBlock(bOk)
  6639  				if src.IsEmptyInterface() {
  6640  					res = iface // Use input interface unchanged.
  6641  					return
  6642  				}
  6643  				// Load type out of itab, build interface with existing idata.
  6644  				off := s.newValue1I(ssa.OpOffPtr, byteptr, rttype.ITab.OffsetOf("Type"), itab)
  6645  				typ := s.load(byteptr, off)
  6646  				idata := s.newValue1(ssa.OpIData, byteptr, iface)
  6647  				res = s.newValue2(ssa.OpIMake, dst, typ, idata)
  6648  				return
  6649  			}
  6650  
  6651  			s.startBlock(bOk)
  6652  			// nonempty -> empty
  6653  			// Need to load type from itab
  6654  			off := s.newValue1I(ssa.OpOffPtr, byteptr, rttype.ITab.OffsetOf("Type"), itab)
  6655  			s.vars[typVar] = s.load(byteptr, off)
  6656  			s.endBlock()
  6657  
  6658  			// itab is nil, might as well use that as the nil result.
  6659  			s.startBlock(bFail)
  6660  			s.vars[typVar] = itab
  6661  			s.endBlock()
  6662  
  6663  			// Merge point.
  6664  			bEnd := s.f.NewBlock(ssa.BlockPlain)
  6665  			bOk.AddEdgeTo(bEnd)
  6666  			bFail.AddEdgeTo(bEnd)
  6667  			s.startBlock(bEnd)
  6668  			idata := s.newValue1(ssa.OpIData, byteptr, iface)
  6669  			res = s.newValue2(ssa.OpIMake, dst, s.variable(typVar, byteptr), idata)
  6670  			resok = cond
  6671  			delete(s.vars, typVar) // no practical effect, just to indicate typVar is no longer live.
  6672  			return
  6673  		}
  6674  		// converting to a nonempty interface needs a runtime call.
  6675  		if base.Debug.TypeAssert > 0 {
  6676  			base.WarnfAt(pos, "type assertion not inlined")
  6677  		}
  6678  
  6679  		itab := s.newValue1(ssa.OpITab, byteptr, iface)
  6680  		data := s.newValue1(ssa.OpIData, types.Types[types.TUNSAFEPTR], iface)
  6681  
  6682  		// First, check for nil.
  6683  		bNil := s.f.NewBlock(ssa.BlockPlain)
  6684  		bNonNil := s.f.NewBlock(ssa.BlockPlain)
  6685  		bMerge := s.f.NewBlock(ssa.BlockPlain)
  6686  		cond := s.newValue2(ssa.OpNeqPtr, types.Types[types.TBOOL], itab, s.constNil(byteptr))
  6687  		b := s.endBlock()
  6688  		b.Kind = ssa.BlockIf
  6689  		b.SetControl(cond)
  6690  		b.Likely = ssa.BranchLikely
  6691  		b.AddEdgeTo(bNonNil)
  6692  		b.AddEdgeTo(bNil)
  6693  
  6694  		s.startBlock(bNil)
  6695  		if commaok {
  6696  			s.vars[typVar] = itab // which will be nil
  6697  			b := s.endBlock()
  6698  			b.AddEdgeTo(bMerge)
  6699  		} else {
  6700  			// Panic if input is nil.
  6701  			s.rtcall(ir.Syms.Panicnildottype, false, nil, target)
  6702  		}
  6703  
  6704  		// Get typ, possibly by loading out of itab.
  6705  		s.startBlock(bNonNil)
  6706  		typ := itab
  6707  		if !src.IsEmptyInterface() {
  6708  			typ = s.load(byteptr, s.newValue1I(ssa.OpOffPtr, byteptr, rttype.ITab.OffsetOf("Type"), itab))
  6709  		}
  6710  
  6711  		// Check the cache first.
  6712  		var d *ssa.Value
  6713  		if descriptor != nil {
  6714  			d = s.newValue1A(ssa.OpAddr, byteptr, descriptor, s.sb)
  6715  			if base.Flag.N == 0 && rtabi.UseInterfaceSwitchCache(Arch.LinkArch.Name) {
  6716  				// Note: we can only use the cache if we have the right atomic load instruction.
  6717  				// Double-check that here.
  6718  				if _, ok := intrinsics[intrinsicKey{Arch.LinkArch.Arch, "internal/runtime/atomic", "Loadp"}]; !ok {
  6719  					s.Fatalf("atomic load not available")
  6720  				}
  6721  				// Pick right size ops.
  6722  				var mul, and, add, zext ssa.Op
  6723  				if s.config.PtrSize == 4 {
  6724  					mul = ssa.OpMul32
  6725  					and = ssa.OpAnd32
  6726  					add = ssa.OpAdd32
  6727  					zext = ssa.OpCopy
  6728  				} else {
  6729  					mul = ssa.OpMul64
  6730  					and = ssa.OpAnd64
  6731  					add = ssa.OpAdd64
  6732  					zext = ssa.OpZeroExt32to64
  6733  				}
  6734  
  6735  				loopHead := s.f.NewBlock(ssa.BlockPlain)
  6736  				loopBody := s.f.NewBlock(ssa.BlockPlain)
  6737  				cacheHit := s.f.NewBlock(ssa.BlockPlain)
  6738  				cacheMiss := s.f.NewBlock(ssa.BlockPlain)
  6739  
  6740  				// Load cache pointer out of descriptor, with an atomic load so
  6741  				// we ensure that we see a fully written cache.
  6742  				atomicLoad := s.newValue2(ssa.OpAtomicLoadPtr, types.NewTuple(typs.BytePtr, types.TypeMem), d, s.mem())
  6743  				cache := s.newValue1(ssa.OpSelect0, typs.BytePtr, atomicLoad)
  6744  				s.vars[memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, atomicLoad)
  6745  
  6746  				// Load hash from type or itab.
  6747  				var hash *ssa.Value
  6748  				if src.IsEmptyInterface() {
  6749  					hash = s.newValue2(ssa.OpLoad, typs.UInt32, s.newValue1I(ssa.OpOffPtr, typs.UInt32Ptr, rttype.Type.OffsetOf("Hash"), typ), s.mem())
  6750  				} else {
  6751  					hash = s.newValue2(ssa.OpLoad, typs.UInt32, s.newValue1I(ssa.OpOffPtr, typs.UInt32Ptr, rttype.ITab.OffsetOf("Hash"), itab), s.mem())
  6752  				}
  6753  				hash = s.newValue1(zext, typs.Uintptr, hash)
  6754  				s.vars[hashVar] = hash
  6755  				// Load mask from cache.
  6756  				mask := s.newValue2(ssa.OpLoad, typs.Uintptr, cache, s.mem())
  6757  				// Jump to loop head.
  6758  				b := s.endBlock()
  6759  				b.AddEdgeTo(loopHead)
  6760  
  6761  				// At loop head, get pointer to the cache entry.
  6762  				//   e := &cache.Entries[hash&mask]
  6763  				s.startBlock(loopHead)
  6764  				idx := s.newValue2(and, typs.Uintptr, s.variable(hashVar, typs.Uintptr), mask)
  6765  				idx = s.newValue2(mul, typs.Uintptr, idx, s.uintptrConstant(uint64(2*s.config.PtrSize)))
  6766  				idx = s.newValue2(add, typs.Uintptr, idx, s.uintptrConstant(uint64(s.config.PtrSize)))
  6767  				e := s.newValue2(ssa.OpAddPtr, typs.UintptrPtr, cache, idx)
  6768  				//   hash++
  6769  				s.vars[hashVar] = s.newValue2(add, typs.Uintptr, s.variable(hashVar, typs.Uintptr), s.uintptrConstant(1))
  6770  
  6771  				// Look for a cache hit.
  6772  				//   if e.Typ == typ { goto hit }
  6773  				eTyp := s.newValue2(ssa.OpLoad, typs.Uintptr, e, s.mem())
  6774  				cmp1 := s.newValue2(ssa.OpEqPtr, typs.Bool, typ, eTyp)
  6775  				b = s.endBlock()
  6776  				b.Kind = ssa.BlockIf
  6777  				b.SetControl(cmp1)
  6778  				b.AddEdgeTo(cacheHit)
  6779  				b.AddEdgeTo(loopBody)
  6780  
  6781  				// Look for an empty entry, the tombstone for this hash table.
  6782  				//   if e.Typ == nil { goto miss }
  6783  				s.startBlock(loopBody)
  6784  				cmp2 := s.newValue2(ssa.OpEqPtr, typs.Bool, eTyp, s.constNil(typs.BytePtr))
  6785  				b = s.endBlock()
  6786  				b.Kind = ssa.BlockIf
  6787  				b.SetControl(cmp2)
  6788  				b.AddEdgeTo(cacheMiss)
  6789  				b.AddEdgeTo(loopHead)
  6790  
  6791  				// On a hit, load the data fields of the cache entry.
  6792  				//   Itab = e.Itab
  6793  				s.startBlock(cacheHit)
  6794  				eItab := s.newValue2(ssa.OpLoad, typs.BytePtr, s.newValue1I(ssa.OpOffPtr, typs.BytePtrPtr, s.config.PtrSize, e), s.mem())
  6795  				s.vars[typVar] = eItab
  6796  				b = s.endBlock()
  6797  				b.AddEdgeTo(bMerge)
  6798  
  6799  				// On a miss, call into the runtime to get the answer.
  6800  				s.startBlock(cacheMiss)
  6801  			}
  6802  		}
  6803  
  6804  		// Call into runtime to get itab for result.
  6805  		if descriptor != nil {
  6806  			itab = s.rtcall(ir.Syms.TypeAssert, true, []*types.Type{byteptr}, d, typ)[0]
  6807  		} else {
  6808  			var fn *obj.LSym
  6809  			if commaok {
  6810  				fn = ir.Syms.AssertE2I2
  6811  			} else {
  6812  				fn = ir.Syms.AssertE2I
  6813  			}
  6814  			itab = s.rtcall(fn, true, []*types.Type{byteptr}, target, typ)[0]
  6815  		}
  6816  		s.vars[typVar] = itab
  6817  		b = s.endBlock()
  6818  		b.AddEdgeTo(bMerge)
  6819  
  6820  		// Build resulting interface.
  6821  		s.startBlock(bMerge)
  6822  		itab = s.variable(typVar, byteptr)
  6823  		var ok *ssa.Value
  6824  		if commaok {
  6825  			ok = s.newValue2(ssa.OpNeqPtr, types.Types[types.TBOOL], itab, s.constNil(byteptr))
  6826  		}
  6827  		return s.newValue2(ssa.OpIMake, dst, itab, data), ok
  6828  	}
  6829  
  6830  	if base.Debug.TypeAssert > 0 {
  6831  		base.WarnfAt(pos, "type assertion inlined")
  6832  	}
  6833  
  6834  	// Converting to a concrete type.
  6835  	direct := types.IsDirectIface(dst)
  6836  	itab := s.newValue1(ssa.OpITab, byteptr, iface) // type word of interface
  6837  	if base.Debug.TypeAssert > 0 {
  6838  		base.WarnfAt(pos, "type assertion inlined")
  6839  	}
  6840  	var wantedFirstWord *ssa.Value
  6841  	if src.IsEmptyInterface() {
  6842  		// Looking for pointer to target type.
  6843  		wantedFirstWord = target
  6844  	} else {
  6845  		// Looking for pointer to itab for target type and source interface.
  6846  		wantedFirstWord = targetItab
  6847  	}
  6848  
  6849  	var tmp ir.Node     // temporary for use with large types
  6850  	var addr *ssa.Value // address of tmp
  6851  	if commaok && !ssa.CanSSA(dst) {
  6852  		// unSSAable type, use temporary.
  6853  		// TODO: get rid of some of these temporaries.
  6854  		tmp, addr = s.temp(pos, dst)
  6855  	}
  6856  
  6857  	cond := s.newValue2(ssa.OpEqPtr, types.Types[types.TBOOL], itab, wantedFirstWord)
  6858  	b := s.endBlock()
  6859  	b.Kind = ssa.BlockIf
  6860  	b.SetControl(cond)
  6861  	b.Likely = ssa.BranchLikely
  6862  
  6863  	bOk := s.f.NewBlock(ssa.BlockPlain)
  6864  	bFail := s.f.NewBlock(ssa.BlockPlain)
  6865  	b.AddEdgeTo(bOk)
  6866  	b.AddEdgeTo(bFail)
  6867  
  6868  	if !commaok {
  6869  		// on failure, panic by calling panicdottype
  6870  		s.startBlock(bFail)
  6871  		taddr := source
  6872  		if taddr == nil {
  6873  			taddr = s.reflectType(src)
  6874  		}
  6875  		if src.IsEmptyInterface() {
  6876  			s.rtcall(ir.Syms.PanicdottypeE, false, nil, itab, target, taddr)
  6877  		} else {
  6878  			s.rtcall(ir.Syms.PanicdottypeI, false, nil, itab, target, taddr)
  6879  		}
  6880  
  6881  		// on success, return data from interface
  6882  		s.startBlock(bOk)
  6883  		if direct {
  6884  			return s.newValue1(ssa.OpIData, dst, iface), nil
  6885  		}
  6886  		p := s.newValue1(ssa.OpIData, types.NewPtr(dst), iface)
  6887  		return s.load(dst, p), nil
  6888  	}
  6889  
  6890  	// commaok is the more complicated case because we have
  6891  	// a control flow merge point.
  6892  	bEnd := s.f.NewBlock(ssa.BlockPlain)
  6893  	// Note that we need a new valVar each time (unlike okVar where we can
  6894  	// reuse the variable) because it might have a different type every time.
  6895  	valVar := ssaMarker("val")
  6896  
  6897  	// type assertion succeeded
  6898  	s.startBlock(bOk)
  6899  	if tmp == nil {
  6900  		if direct {
  6901  			s.vars[valVar] = s.newValue1(ssa.OpIData, dst, iface)
  6902  		} else {
  6903  			p := s.newValue1(ssa.OpIData, types.NewPtr(dst), iface)
  6904  			s.vars[valVar] = s.load(dst, p)
  6905  		}
  6906  	} else {
  6907  		p := s.newValue1(ssa.OpIData, types.NewPtr(dst), iface)
  6908  		s.move(dst, addr, p)
  6909  	}
  6910  	s.vars[okVar] = s.constBool(true)
  6911  	s.endBlock()
  6912  	bOk.AddEdgeTo(bEnd)
  6913  
  6914  	// type assertion failed
  6915  	s.startBlock(bFail)
  6916  	if tmp == nil {
  6917  		s.vars[valVar] = s.zeroVal(dst)
  6918  	} else {
  6919  		s.zero(dst, addr)
  6920  	}
  6921  	s.vars[okVar] = s.constBool(false)
  6922  	s.endBlock()
  6923  	bFail.AddEdgeTo(bEnd)
  6924  
  6925  	// merge point
  6926  	s.startBlock(bEnd)
  6927  	if tmp == nil {
  6928  		res = s.variable(valVar, dst)
  6929  		delete(s.vars, valVar) // no practical effect, just to indicate typVar is no longer live.
  6930  	} else {
  6931  		res = s.load(dst, addr)
  6932  	}
  6933  	resok = s.variable(okVar, types.Types[types.TBOOL])
  6934  	delete(s.vars, okVar) // ditto
  6935  	return res, resok
  6936  }
  6937  
  6938  // temp allocates a temp of type t at position pos
  6939  func (s *state) temp(pos src.XPos, t *types.Type) (*ir.Name, *ssa.Value) {
  6940  	tmp := typecheck.TempAt(pos, s.curfn, t)
  6941  	if t.HasPointers() || (ssa.IsMergeCandidate(tmp) && t != deferstruct()) {
  6942  		s.vars[memVar] = s.newValue1A(ssa.OpVarDef, types.TypeMem, tmp, s.mem())
  6943  	}
  6944  	addr := s.addr(tmp)
  6945  	return tmp, addr
  6946  }
  6947  
  6948  // variable returns the value of a variable at the current location.
  6949  func (s *state) variable(n ir.Node, t *types.Type) *ssa.Value {
  6950  	v := s.vars[n]
  6951  	if v != nil {
  6952  		return v
  6953  	}
  6954  	v = s.fwdVars[n]
  6955  	if v != nil {
  6956  		return v
  6957  	}
  6958  
  6959  	if s.curBlock == s.f.Entry {
  6960  		// No variable should be live at entry.
  6961  		s.f.Fatalf("value %v (%v) incorrectly live at entry", n, v)
  6962  	}
  6963  	// Make a FwdRef, which records a value that's live on block input.
  6964  	// We'll find the matching definition as part of insertPhis.
  6965  	v = s.newValue0A(ssa.OpFwdRef, t, fwdRefAux{N: n})
  6966  	s.fwdVars[n] = v
  6967  	if n.Op() == ir.ONAME {
  6968  		s.addNamedValue(n.(*ir.Name), v)
  6969  	}
  6970  	return v
  6971  }
  6972  
  6973  func (s *state) mem() *ssa.Value {
  6974  	return s.variable(memVar, types.TypeMem)
  6975  }
  6976  
  6977  func (s *state) addNamedValue(n *ir.Name, v *ssa.Value) {
  6978  	if n.Class == ir.Pxxx {
  6979  		// Don't track our marker nodes (memVar etc.).
  6980  		return
  6981  	}
  6982  	if ir.IsAutoTmp(n) {
  6983  		// Don't track temporary variables.
  6984  		return
  6985  	}
  6986  	if n.Class == ir.PPARAMOUT {
  6987  		// Don't track named output values.  This prevents return values
  6988  		// from being assigned too early. See #14591 and #14762. TODO: allow this.
  6989  		return
  6990  	}
  6991  	loc := ssa.LocalSlot{N: n, Type: n.Type(), Off: 0}
  6992  	values, ok := s.f.NamedValues[loc]
  6993  	if !ok {
  6994  		s.f.Names = append(s.f.Names, &loc)
  6995  		s.f.CanonicalLocalSlots[loc] = &loc
  6996  	}
  6997  	s.f.NamedValues[loc] = append(values, v)
  6998  }
  6999  
  7000  // Branch is an unresolved branch.
  7001  type Branch struct {
  7002  	P *obj.Prog  // branch instruction
  7003  	B *ssa.Block // target
  7004  }
  7005  
  7006  // State contains state needed during Prog generation.
  7007  type State struct {
  7008  	ABI obj.ABI
  7009  
  7010  	pp *objw.Progs
  7011  
  7012  	// Branches remembers all the branch instructions we've seen
  7013  	// and where they would like to go.
  7014  	Branches []Branch
  7015  
  7016  	// JumpTables remembers all the jump tables we've seen.
  7017  	JumpTables []*ssa.Block
  7018  
  7019  	// bstart remembers where each block starts (indexed by block ID)
  7020  	bstart []*obj.Prog
  7021  
  7022  	maxarg int64 // largest frame size for arguments to calls made by the function
  7023  
  7024  	// Map from GC safe points to liveness index, generated by
  7025  	// liveness analysis.
  7026  	livenessMap liveness.Map
  7027  
  7028  	// partLiveArgs includes arguments that may be partially live, for which we
  7029  	// need to generate instructions that spill the argument registers.
  7030  	partLiveArgs map[*ir.Name]bool
  7031  
  7032  	// lineRunStart records the beginning of the current run of instructions
  7033  	// within a single block sharing the same line number
  7034  	// Used to move statement marks to the beginning of such runs.
  7035  	lineRunStart *obj.Prog
  7036  
  7037  	// wasm: The number of values on the WebAssembly stack. This is only used as a safeguard.
  7038  	OnWasmStackSkipped int
  7039  }
  7040  
  7041  func (s *State) FuncInfo() *obj.FuncInfo {
  7042  	return s.pp.CurFunc.LSym.Func()
  7043  }
  7044  
  7045  // Prog appends a new Prog.
  7046  func (s *State) Prog(as obj.As) *obj.Prog {
  7047  	p := s.pp.Prog(as)
  7048  	if objw.LosesStmtMark(as) {
  7049  		return p
  7050  	}
  7051  	// Float a statement start to the beginning of any same-line run.
  7052  	// lineRunStart is reset at block boundaries, which appears to work well.
  7053  	if s.lineRunStart == nil || s.lineRunStart.Pos.Line() != p.Pos.Line() {
  7054  		s.lineRunStart = p
  7055  	} else if p.Pos.IsStmt() == src.PosIsStmt {
  7056  		s.lineRunStart.Pos = s.lineRunStart.Pos.WithIsStmt()
  7057  		p.Pos = p.Pos.WithNotStmt()
  7058  	}
  7059  	return p
  7060  }
  7061  
  7062  // Pc returns the current Prog.
  7063  func (s *State) Pc() *obj.Prog {
  7064  	return s.pp.Next
  7065  }
  7066  
  7067  // SetPos sets the current source position.
  7068  func (s *State) SetPos(pos src.XPos) {
  7069  	s.pp.Pos = pos
  7070  }
  7071  
  7072  // Br emits a single branch instruction and returns the instruction.
  7073  // Not all architectures need the returned instruction, but otherwise
  7074  // the boilerplate is common to all.
  7075  func (s *State) Br(op obj.As, target *ssa.Block) *obj.Prog {
  7076  	p := s.Prog(op)
  7077  	p.To.Type = obj.TYPE_BRANCH
  7078  	s.Branches = append(s.Branches, Branch{P: p, B: target})
  7079  	return p
  7080  }
  7081  
  7082  // DebugFriendlySetPosFrom adjusts Pos.IsStmt subject to heuristics
  7083  // that reduce "jumpy" line number churn when debugging.
  7084  // Spill/fill/copy instructions from the register allocator,
  7085  // phi functions, and instructions with a no-pos position
  7086  // are examples of instructions that can cause churn.
  7087  func (s *State) DebugFriendlySetPosFrom(v *ssa.Value) {
  7088  	switch v.Op {
  7089  	case ssa.OpPhi, ssa.OpCopy, ssa.OpLoadReg, ssa.OpStoreReg:
  7090  		// These are not statements
  7091  		s.SetPos(v.Pos.WithNotStmt())
  7092  	default:
  7093  		p := v.Pos
  7094  		if p != src.NoXPos {
  7095  			// If the position is defined, update the position.
  7096  			// Also convert default IsStmt to NotStmt; only
  7097  			// explicit statement boundaries should appear
  7098  			// in the generated code.
  7099  			if p.IsStmt() != src.PosIsStmt {
  7100  				if s.pp.Pos.IsStmt() == src.PosIsStmt && s.pp.Pos.SameFileAndLine(p) {
  7101  					// If s.pp.Pos already has a statement mark, then it was set here (below) for
  7102  					// the previous value.  If an actual instruction had been emitted for that
  7103  					// value, then the statement mark would have been reset.  Since the statement
  7104  					// mark of s.pp.Pos was not reset, this position (file/line) still needs a
  7105  					// statement mark on an instruction.  If file and line for this value are
  7106  					// the same as the previous value, then the first instruction for this
  7107  					// value will work to take the statement mark.  Return early to avoid
  7108  					// resetting the statement mark.
  7109  					//
  7110  					// The reset of s.pp.Pos occurs in (*Progs).Prog() -- if it emits
  7111  					// an instruction, and the instruction's statement mark was set,
  7112  					// and it is not one of the LosesStmtMark instructions,
  7113  					// then Prog() resets the statement mark on the (*Progs).Pos.
  7114  					return
  7115  				}
  7116  				p = p.WithNotStmt()
  7117  				// Calls use the pos attached to v, but copy the statement mark from State
  7118  			}
  7119  			s.SetPos(p)
  7120  		} else {
  7121  			s.SetPos(s.pp.Pos.WithNotStmt())
  7122  		}
  7123  	}
  7124  }
  7125  
  7126  // emit argument info (locations on stack) for traceback.
  7127  func emitArgInfo(e *ssafn, f *ssa.Func, pp *objw.Progs) {
  7128  	ft := e.curfn.Type()
  7129  	if ft.NumRecvs() == 0 && ft.NumParams() == 0 {
  7130  		return
  7131  	}
  7132  
  7133  	x := EmitArgInfo(e.curfn, f.OwnAux.ABIInfo())
  7134  	x.Set(obj.AttrContentAddressable, true)
  7135  	e.curfn.LSym.Func().ArgInfo = x
  7136  
  7137  	// Emit a funcdata pointing at the arg info data.
  7138  	p := pp.Prog(obj.AFUNCDATA)
  7139  	p.From.SetConst(rtabi.FUNCDATA_ArgInfo)
  7140  	p.To.Type = obj.TYPE_MEM
  7141  	p.To.Name = obj.NAME_EXTERN
  7142  	p.To.Sym = x
  7143  }
  7144  
  7145  // emit argument info (locations on stack) of f for traceback.
  7146  func EmitArgInfo(f *ir.Func, abiInfo *abi.ABIParamResultInfo) *obj.LSym {
  7147  	x := base.Ctxt.Lookup(fmt.Sprintf("%s.arginfo%d", f.LSym.Name, f.ABI))
  7148  	// NOTE: do not set ContentAddressable here. This may be referenced from
  7149  	// assembly code by name (in this case f is a declaration).
  7150  	// Instead, set it in emitArgInfo above.
  7151  
  7152  	PtrSize := int64(types.PtrSize)
  7153  	uintptrTyp := types.Types[types.TUINTPTR]
  7154  
  7155  	isAggregate := func(t *types.Type) bool {
  7156  		return t.IsStruct() || t.IsArray() || t.IsComplex() || t.IsInterface() || t.IsString() || t.IsSlice()
  7157  	}
  7158  
  7159  	wOff := 0
  7160  	n := 0
  7161  	writebyte := func(o uint8) { wOff = objw.Uint8(x, wOff, o) }
  7162  
  7163  	// Write one non-aggregate arg/field/element.
  7164  	write1 := func(sz, offset int64) {
  7165  		if offset >= rtabi.TraceArgsSpecial {
  7166  			writebyte(rtabi.TraceArgsOffsetTooLarge)
  7167  		} else {
  7168  			writebyte(uint8(offset))
  7169  			writebyte(uint8(sz))
  7170  		}
  7171  		n++
  7172  	}
  7173  
  7174  	// Visit t recursively and write it out.
  7175  	// Returns whether to continue visiting.
  7176  	var visitType func(baseOffset int64, t *types.Type, depth int) bool
  7177  	visitType = func(baseOffset int64, t *types.Type, depth int) bool {
  7178  		if n >= rtabi.TraceArgsLimit {
  7179  			writebyte(rtabi.TraceArgsDotdotdot)
  7180  			return false
  7181  		}
  7182  		if !isAggregate(t) {
  7183  			write1(t.Size(), baseOffset)
  7184  			return true
  7185  		}
  7186  		writebyte(rtabi.TraceArgsStartAgg)
  7187  		depth++
  7188  		if depth >= rtabi.TraceArgsMaxDepth {
  7189  			writebyte(rtabi.TraceArgsDotdotdot)
  7190  			writebyte(rtabi.TraceArgsEndAgg)
  7191  			n++
  7192  			return true
  7193  		}
  7194  		switch {
  7195  		case t.IsInterface(), t.IsString():
  7196  			_ = visitType(baseOffset, uintptrTyp, depth) &&
  7197  				visitType(baseOffset+PtrSize, uintptrTyp, depth)
  7198  		case t.IsSlice():
  7199  			_ = visitType(baseOffset, uintptrTyp, depth) &&
  7200  				visitType(baseOffset+PtrSize, uintptrTyp, depth) &&
  7201  				visitType(baseOffset+PtrSize*2, uintptrTyp, depth)
  7202  		case t.IsComplex():
  7203  			_ = visitType(baseOffset, types.FloatForComplex(t), depth) &&
  7204  				visitType(baseOffset+t.Size()/2, types.FloatForComplex(t), depth)
  7205  		case t.IsArray():
  7206  			if t.NumElem() == 0 {
  7207  				n++ // {} counts as a component
  7208  				break
  7209  			}
  7210  			for i := int64(0); i < t.NumElem(); i++ {
  7211  				if !visitType(baseOffset, t.Elem(), depth) {
  7212  					break
  7213  				}
  7214  				baseOffset += t.Elem().Size()
  7215  			}
  7216  		case t.IsStruct():
  7217  			if t.NumFields() == 0 {
  7218  				n++ // {} counts as a component
  7219  				break
  7220  			}
  7221  			for _, field := range t.Fields() {
  7222  				if !visitType(baseOffset+field.Offset, field.Type, depth) {
  7223  					break
  7224  				}
  7225  			}
  7226  		}
  7227  		writebyte(rtabi.TraceArgsEndAgg)
  7228  		return true
  7229  	}
  7230  
  7231  	start := 0
  7232  	if strings.Contains(f.LSym.Name, "[") {
  7233  		// Skip the dictionary argument - it is implicit and the user doesn't need to see it.
  7234  		start = 1
  7235  	}
  7236  
  7237  	for _, a := range abiInfo.InParams()[start:] {
  7238  		if !visitType(a.FrameOffset(abiInfo), a.Type, 0) {
  7239  			break
  7240  		}
  7241  	}
  7242  	writebyte(rtabi.TraceArgsEndSeq)
  7243  	if wOff > rtabi.TraceArgsMaxLen {
  7244  		base.Fatalf("ArgInfo too large")
  7245  	}
  7246  
  7247  	return x
  7248  }
  7249  
  7250  // for wrapper, emit info of wrapped function.
  7251  func emitWrappedFuncInfo(e *ssafn, pp *objw.Progs) {
  7252  	if base.Ctxt.Flag_linkshared {
  7253  		// Relative reference (SymPtrOff) to another shared object doesn't work.
  7254  		// Unfortunate.
  7255  		return
  7256  	}
  7257  
  7258  	wfn := e.curfn.WrappedFunc
  7259  	if wfn == nil {
  7260  		return
  7261  	}
  7262  
  7263  	wsym := wfn.Linksym()
  7264  	x := base.Ctxt.LookupInit(fmt.Sprintf("%s.wrapinfo", wsym.Name), func(x *obj.LSym) {
  7265  		objw.SymPtrOff(x, 0, wsym)
  7266  		x.Set(obj.AttrContentAddressable, true)
  7267  	})
  7268  	e.curfn.LSym.Func().WrapInfo = x
  7269  
  7270  	// Emit a funcdata pointing at the wrap info data.
  7271  	p := pp.Prog(obj.AFUNCDATA)
  7272  	p.From.SetConst(rtabi.FUNCDATA_WrapInfo)
  7273  	p.To.Type = obj.TYPE_MEM
  7274  	p.To.Name = obj.NAME_EXTERN
  7275  	p.To.Sym = x
  7276  }
  7277  
  7278  // genssa appends entries to pp for each instruction in f.
  7279  func genssa(f *ssa.Func, pp *objw.Progs) {
  7280  	var s State
  7281  	s.ABI = f.OwnAux.Fn.ABI()
  7282  
  7283  	e := f.Frontend().(*ssafn)
  7284  
  7285  	s.livenessMap, s.partLiveArgs = liveness.Compute(e.curfn, f, e.stkptrsize, pp)
  7286  	emitArgInfo(e, f, pp)
  7287  	argLiveBlockMap, argLiveValueMap := liveness.ArgLiveness(e.curfn, f, pp)
  7288  
  7289  	openDeferInfo := e.curfn.LSym.Func().OpenCodedDeferInfo
  7290  	if openDeferInfo != nil {
  7291  		// This function uses open-coded defers -- write out the funcdata
  7292  		// info that we computed at the end of genssa.
  7293  		p := pp.Prog(obj.AFUNCDATA)
  7294  		p.From.SetConst(rtabi.FUNCDATA_OpenCodedDeferInfo)
  7295  		p.To.Type = obj.TYPE_MEM
  7296  		p.To.Name = obj.NAME_EXTERN
  7297  		p.To.Sym = openDeferInfo
  7298  	}
  7299  
  7300  	emitWrappedFuncInfo(e, pp)
  7301  
  7302  	// Remember where each block starts.
  7303  	s.bstart = make([]*obj.Prog, f.NumBlocks())
  7304  	s.pp = pp
  7305  	var progToValue map[*obj.Prog]*ssa.Value
  7306  	var progToBlock map[*obj.Prog]*ssa.Block
  7307  	var valueToProgAfter []*obj.Prog // The first Prog following computation of a value v; v is visible at this point.
  7308  	gatherPrintInfo := f.PrintOrHtmlSSA || ssa.GenssaDump[f.Name]
  7309  	if gatherPrintInfo {
  7310  		progToValue = make(map[*obj.Prog]*ssa.Value, f.NumValues())
  7311  		progToBlock = make(map[*obj.Prog]*ssa.Block, f.NumBlocks())
  7312  		f.Logf("genssa %s\n", f.Name)
  7313  		progToBlock[s.pp.Next] = f.Blocks[0]
  7314  	}
  7315  
  7316  	if base.Ctxt.Flag_locationlists {
  7317  		if cap(f.Cache.ValueToProgAfter) < f.NumValues() {
  7318  			f.Cache.ValueToProgAfter = make([]*obj.Prog, f.NumValues())
  7319  		}
  7320  		valueToProgAfter = f.Cache.ValueToProgAfter[:f.NumValues()]
  7321  		for i := range valueToProgAfter {
  7322  			valueToProgAfter[i] = nil
  7323  		}
  7324  	}
  7325  
  7326  	// If the very first instruction is not tagged as a statement,
  7327  	// debuggers may attribute it to previous function in program.
  7328  	firstPos := src.NoXPos
  7329  	for _, v := range f.Entry.Values {
  7330  		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 {
  7331  			firstPos = v.Pos
  7332  			v.Pos = firstPos.WithDefaultStmt()
  7333  			break
  7334  		}
  7335  	}
  7336  
  7337  	// inlMarks has an entry for each Prog that implements an inline mark.
  7338  	// It maps from that Prog to the global inlining id of the inlined body
  7339  	// which should unwind to this Prog's location.
  7340  	var inlMarks map[*obj.Prog]int32
  7341  	var inlMarkList []*obj.Prog
  7342  
  7343  	// inlMarksByPos maps from a (column 1) source position to the set of
  7344  	// Progs that are in the set above and have that source position.
  7345  	var inlMarksByPos map[src.XPos][]*obj.Prog
  7346  
  7347  	var argLiveIdx int = -1 // argument liveness info index
  7348  
  7349  	// These control cache line alignment; if the required portion of
  7350  	// a cache line is not available, then pad to obtain cache line
  7351  	// alignment.  Not implemented on all architectures, may not be
  7352  	// useful on all architectures.
  7353  	var hotAlign, hotRequire int64
  7354  
  7355  	if base.Debug.AlignHot > 0 {
  7356  		switch base.Ctxt.Arch.Name {
  7357  		// enable this on a case-by-case basis, with benchmarking.
  7358  		// currently shown:
  7359  		//   good for amd64
  7360  		//   not helpful for Apple Silicon
  7361  		//
  7362  		case "amd64", "386":
  7363  			// Align to 64 if 31 or fewer bytes remain in a cache line
  7364  			// benchmarks a little better than always aligning, and also
  7365  			// adds slightly less to the (PGO-compiled) binary size.
  7366  			hotAlign = 64
  7367  			hotRequire = 31
  7368  		}
  7369  	}
  7370  
  7371  	// Emit basic blocks
  7372  	for i, b := range f.Blocks {
  7373  
  7374  		s.lineRunStart = nil
  7375  		s.SetPos(s.pp.Pos.WithNotStmt()) // It needs a non-empty Pos, but cannot be a statement boundary (yet).
  7376  
  7377  		if hotAlign > 0 && b.Hotness&ssa.HotPgoInitial == ssa.HotPgoInitial {
  7378  			// So far this has only been shown profitable for PGO-hot loop headers.
  7379  			// The Hotness values allows distinctions betwen initial blocks that are "hot" or not, and "flow-in" or not.
  7380  			// Currently only the initial blocks of loops are tagged in this way;
  7381  			// there are no blocks tagged "pgo-hot" that are not also tagged "initial".
  7382  			// TODO more heuristics, more architectures.
  7383  			p := s.pp.Prog(obj.APCALIGNMAX)
  7384  			p.From.SetConst(hotAlign)
  7385  			p.To.SetConst(hotRequire)
  7386  		}
  7387  
  7388  		s.bstart[b.ID] = s.pp.Next
  7389  
  7390  		if idx, ok := argLiveBlockMap[b.ID]; ok && idx != argLiveIdx {
  7391  			argLiveIdx = idx
  7392  			p := s.pp.Prog(obj.APCDATA)
  7393  			p.From.SetConst(rtabi.PCDATA_ArgLiveIndex)
  7394  			p.To.SetConst(int64(idx))
  7395  		}
  7396  
  7397  		// Emit values in block
  7398  		Arch.SSAMarkMoves(&s, b)
  7399  		for _, v := range b.Values {
  7400  			x := s.pp.Next
  7401  			s.DebugFriendlySetPosFrom(v)
  7402  
  7403  			if v.Op.ResultInArg0() && v.ResultReg() != v.Args[0].Reg() {
  7404  				v.Fatalf("input[0] and output not in same register %s", v.LongString())
  7405  			}
  7406  
  7407  			switch v.Op {
  7408  			case ssa.OpInitMem:
  7409  				// memory arg needs no code
  7410  			case ssa.OpArg:
  7411  				// input args need no code
  7412  			case ssa.OpSP, ssa.OpSB:
  7413  				// nothing to do
  7414  			case ssa.OpSelect0, ssa.OpSelect1, ssa.OpSelectN, ssa.OpMakeResult:
  7415  				// nothing to do
  7416  			case ssa.OpGetG:
  7417  				// nothing to do when there's a g register,
  7418  				// and checkLower complains if there's not
  7419  			case ssa.OpVarDef, ssa.OpVarLive, ssa.OpKeepAlive, ssa.OpWBend:
  7420  				// nothing to do; already used by liveness
  7421  			case ssa.OpPhi:
  7422  				CheckLoweredPhi(v)
  7423  			case ssa.OpConvert:
  7424  				// nothing to do; no-op conversion for liveness
  7425  				if v.Args[0].Reg() != v.Reg() {
  7426  					v.Fatalf("OpConvert should be a no-op: %s; %s", v.Args[0].LongString(), v.LongString())
  7427  				}
  7428  			case ssa.OpInlMark:
  7429  				p := Arch.Ginsnop(s.pp)
  7430  				if inlMarks == nil {
  7431  					inlMarks = map[*obj.Prog]int32{}
  7432  					inlMarksByPos = map[src.XPos][]*obj.Prog{}
  7433  				}
  7434  				inlMarks[p] = v.AuxInt32()
  7435  				inlMarkList = append(inlMarkList, p)
  7436  				pos := v.Pos.AtColumn1()
  7437  				inlMarksByPos[pos] = append(inlMarksByPos[pos], p)
  7438  				firstPos = src.NoXPos
  7439  
  7440  			default:
  7441  				// Special case for first line in function; move it to the start (which cannot be a register-valued instruction)
  7442  				if firstPos != src.NoXPos && v.Op != ssa.OpArgIntReg && v.Op != ssa.OpArgFloatReg && v.Op != ssa.OpLoadReg && v.Op != ssa.OpStoreReg {
  7443  					s.SetPos(firstPos)
  7444  					firstPos = src.NoXPos
  7445  				}
  7446  				// Attach this safe point to the next
  7447  				// instruction.
  7448  				s.pp.NextLive = s.livenessMap.Get(v)
  7449  				s.pp.NextUnsafe = s.livenessMap.GetUnsafe(v)
  7450  
  7451  				// let the backend handle it
  7452  				Arch.SSAGenValue(&s, v)
  7453  			}
  7454  
  7455  			if idx, ok := argLiveValueMap[v.ID]; ok && idx != argLiveIdx {
  7456  				argLiveIdx = idx
  7457  				p := s.pp.Prog(obj.APCDATA)
  7458  				p.From.SetConst(rtabi.PCDATA_ArgLiveIndex)
  7459  				p.To.SetConst(int64(idx))
  7460  			}
  7461  
  7462  			if base.Ctxt.Flag_locationlists {
  7463  				valueToProgAfter[v.ID] = s.pp.Next
  7464  			}
  7465  
  7466  			if gatherPrintInfo {
  7467  				for ; x != s.pp.Next; x = x.Link {
  7468  					progToValue[x] = v
  7469  				}
  7470  			}
  7471  		}
  7472  		// If this is an empty infinite loop, stick a hardware NOP in there so that debuggers are less confused.
  7473  		if s.bstart[b.ID] == s.pp.Next && len(b.Succs) == 1 && b.Succs[0].Block() == b {
  7474  			p := Arch.Ginsnop(s.pp)
  7475  			p.Pos = p.Pos.WithIsStmt()
  7476  			if b.Pos == src.NoXPos {
  7477  				b.Pos = p.Pos // It needs a file, otherwise a no-file non-zero line causes confusion.  See #35652.
  7478  				if b.Pos == src.NoXPos {
  7479  					b.Pos = s.pp.Text.Pos // Sometimes p.Pos is empty.  See #35695.
  7480  				}
  7481  			}
  7482  			b.Pos = b.Pos.WithBogusLine() // Debuggers are not good about infinite loops, force a change in line number
  7483  		}
  7484  
  7485  		// Set unsafe mark for any end-of-block generated instructions
  7486  		// (normally, conditional or unconditional branches).
  7487  		// This is particularly important for empty blocks, as there
  7488  		// are no values to inherit the unsafe mark from.
  7489  		s.pp.NextUnsafe = s.livenessMap.GetUnsafeBlock(b)
  7490  
  7491  		// Emit control flow instructions for block
  7492  		var next *ssa.Block
  7493  		if i < len(f.Blocks)-1 && base.Flag.N == 0 {
  7494  			// If -N, leave next==nil so every block with successors
  7495  			// ends in a JMP (except call blocks - plive doesn't like
  7496  			// select{send,recv} followed by a JMP call).  Helps keep
  7497  			// line numbers for otherwise empty blocks.
  7498  			next = f.Blocks[i+1]
  7499  		}
  7500  		x := s.pp.Next
  7501  		s.SetPos(b.Pos)
  7502  		Arch.SSAGenBlock(&s, b, next)
  7503  		if gatherPrintInfo {
  7504  			for ; x != s.pp.Next; x = x.Link {
  7505  				progToBlock[x] = b
  7506  			}
  7507  		}
  7508  	}
  7509  	if f.Blocks[len(f.Blocks)-1].Kind == ssa.BlockExit {
  7510  		// We need the return address of a panic call to
  7511  		// still be inside the function in question. So if
  7512  		// it ends in a call which doesn't return, add a
  7513  		// nop (which will never execute) after the call.
  7514  		Arch.Ginsnop(s.pp)
  7515  	}
  7516  	if openDeferInfo != nil {
  7517  		// When doing open-coded defers, generate a disconnected call to
  7518  		// deferreturn and a return. This will be used to during panic
  7519  		// recovery to unwind the stack and return back to the runtime.
  7520  		s.pp.NextLive = s.livenessMap.DeferReturn
  7521  		p := s.pp.Prog(obj.ACALL)
  7522  		p.To.Type = obj.TYPE_MEM
  7523  		p.To.Name = obj.NAME_EXTERN
  7524  		p.To.Sym = ir.Syms.Deferreturn
  7525  
  7526  		// Load results into registers. So when a deferred function
  7527  		// recovers a panic, it will return to caller with right results.
  7528  		// The results are already in memory, because they are not SSA'd
  7529  		// when the function has defers (see canSSAName).
  7530  		for _, o := range f.OwnAux.ABIInfo().OutParams() {
  7531  			n := o.Name
  7532  			rts, offs := o.RegisterTypesAndOffsets()
  7533  			for i := range o.Registers {
  7534  				Arch.LoadRegResult(&s, f, rts[i], ssa.ObjRegForAbiReg(o.Registers[i], f.Config), n, offs[i])
  7535  			}
  7536  		}
  7537  
  7538  		s.pp.Prog(obj.ARET)
  7539  	}
  7540  
  7541  	if inlMarks != nil {
  7542  		hasCall := false
  7543  
  7544  		// We have some inline marks. Try to find other instructions we're
  7545  		// going to emit anyway, and use those instructions instead of the
  7546  		// inline marks.
  7547  		for p := s.pp.Text; p != nil; p = p.Link {
  7548  			if p.As == obj.ANOP || p.As == obj.AFUNCDATA || p.As == obj.APCDATA || p.As == obj.ATEXT ||
  7549  				p.As == obj.APCALIGN || p.As == obj.APCALIGNMAX || Arch.LinkArch.Family == sys.Wasm {
  7550  				// Don't use 0-sized instructions as inline marks, because we need
  7551  				// to identify inline mark instructions by pc offset.
  7552  				// (Some of these instructions are sometimes zero-sized, sometimes not.
  7553  				// We must not use anything that even might be zero-sized.)
  7554  				// TODO: are there others?
  7555  				continue
  7556  			}
  7557  			if _, ok := inlMarks[p]; ok {
  7558  				// Don't use inline marks themselves. We don't know
  7559  				// whether they will be zero-sized or not yet.
  7560  				continue
  7561  			}
  7562  			if p.As == obj.ACALL || p.As == obj.ADUFFCOPY || p.As == obj.ADUFFZERO {
  7563  				hasCall = true
  7564  			}
  7565  			pos := p.Pos.AtColumn1()
  7566  			marks := inlMarksByPos[pos]
  7567  			if len(marks) == 0 {
  7568  				continue
  7569  			}
  7570  			for _, m := range marks {
  7571  				// We found an instruction with the same source position as
  7572  				// some of the inline marks.
  7573  				// Use this instruction instead.
  7574  				p.Pos = p.Pos.WithIsStmt() // promote position to a statement
  7575  				s.pp.CurFunc.LSym.Func().AddInlMark(p, inlMarks[m])
  7576  				// Make the inline mark a real nop, so it doesn't generate any code.
  7577  				m.As = obj.ANOP
  7578  				m.Pos = src.NoXPos
  7579  				m.From = obj.Addr{}
  7580  				m.To = obj.Addr{}
  7581  			}
  7582  			delete(inlMarksByPos, pos)
  7583  		}
  7584  		// Any unmatched inline marks now need to be added to the inlining tree (and will generate a nop instruction).
  7585  		for _, p := range inlMarkList {
  7586  			if p.As != obj.ANOP {
  7587  				s.pp.CurFunc.LSym.Func().AddInlMark(p, inlMarks[p])
  7588  			}
  7589  		}
  7590  
  7591  		if e.stksize == 0 && !hasCall {
  7592  			// Frameless leaf function. It doesn't need any preamble,
  7593  			// so make sure its first instruction isn't from an inlined callee.
  7594  			// If it is, add a nop at the start of the function with a position
  7595  			// equal to the start of the function.
  7596  			// This ensures that runtime.FuncForPC(uintptr(reflect.ValueOf(fn).Pointer())).Name()
  7597  			// returns the right answer. See issue 58300.
  7598  			for p := s.pp.Text; p != nil; p = p.Link {
  7599  				if p.As == obj.AFUNCDATA || p.As == obj.APCDATA || p.As == obj.ATEXT || p.As == obj.ANOP {
  7600  					continue
  7601  				}
  7602  				if base.Ctxt.PosTable.Pos(p.Pos).Base().InliningIndex() >= 0 {
  7603  					// Make a real (not 0-sized) nop.
  7604  					nop := Arch.Ginsnop(s.pp)
  7605  					nop.Pos = e.curfn.Pos().WithIsStmt()
  7606  
  7607  					// Unfortunately, Ginsnop puts the instruction at the
  7608  					// end of the list. Move it up to just before p.
  7609  
  7610  					// Unlink from the current list.
  7611  					for x := s.pp.Text; x != nil; x = x.Link {
  7612  						if x.Link == nop {
  7613  							x.Link = nop.Link
  7614  							break
  7615  						}
  7616  					}
  7617  					// Splice in right before p.
  7618  					for x := s.pp.Text; x != nil; x = x.Link {
  7619  						if x.Link == p {
  7620  							nop.Link = p
  7621  							x.Link = nop
  7622  							break
  7623  						}
  7624  					}
  7625  				}
  7626  				break
  7627  			}
  7628  		}
  7629  	}
  7630  
  7631  	if base.Ctxt.Flag_locationlists {
  7632  		var debugInfo *ssa.FuncDebug
  7633  		debugInfo = e.curfn.DebugInfo.(*ssa.FuncDebug)
  7634  		if e.curfn.ABI == obj.ABIInternal && base.Flag.N != 0 {
  7635  			ssa.BuildFuncDebugNoOptimized(base.Ctxt, f, base.Debug.LocationLists > 1, StackOffset, debugInfo)
  7636  		} else {
  7637  			ssa.BuildFuncDebug(base.Ctxt, f, base.Debug.LocationLists, StackOffset, debugInfo)
  7638  		}
  7639  		bstart := s.bstart
  7640  		idToIdx := make([]int, f.NumBlocks())
  7641  		for i, b := range f.Blocks {
  7642  			idToIdx[b.ID] = i
  7643  		}
  7644  		// Register a callback that will be used later to fill in PCs into location
  7645  		// lists. At the moment, Prog.Pc is a sequence number; it's not a real PC
  7646  		// until after assembly, so the translation needs to be deferred.
  7647  		debugInfo.GetPC = func(b, v ssa.ID) int64 {
  7648  			switch v {
  7649  			case ssa.BlockStart.ID:
  7650  				if b == f.Entry.ID {
  7651  					return 0 // Start at the very beginning, at the assembler-generated prologue.
  7652  					// this should only happen for function args (ssa.OpArg)
  7653  				}
  7654  				return bstart[b].Pc
  7655  			case ssa.BlockEnd.ID:
  7656  				blk := f.Blocks[idToIdx[b]]
  7657  				nv := len(blk.Values)
  7658  				return valueToProgAfter[blk.Values[nv-1].ID].Pc
  7659  			case ssa.FuncEnd.ID:
  7660  				return e.curfn.LSym.Size
  7661  			default:
  7662  				return valueToProgAfter[v].Pc
  7663  			}
  7664  		}
  7665  	}
  7666  
  7667  	// Resolve branches, and relax DefaultStmt into NotStmt
  7668  	for _, br := range s.Branches {
  7669  		br.P.To.SetTarget(s.bstart[br.B.ID])
  7670  		if br.P.Pos.IsStmt() != src.PosIsStmt {
  7671  			br.P.Pos = br.P.Pos.WithNotStmt()
  7672  		} else if v0 := br.B.FirstPossibleStmtValue(); v0 != nil && v0.Pos.Line() == br.P.Pos.Line() && v0.Pos.IsStmt() == src.PosIsStmt {
  7673  			br.P.Pos = br.P.Pos.WithNotStmt()
  7674  		}
  7675  
  7676  	}
  7677  
  7678  	// Resolve jump table destinations.
  7679  	for _, jt := range s.JumpTables {
  7680  		// Convert from *Block targets to *Prog targets.
  7681  		targets := make([]*obj.Prog, len(jt.Succs))
  7682  		for i, e := range jt.Succs {
  7683  			targets[i] = s.bstart[e.Block().ID]
  7684  		}
  7685  		// Add to list of jump tables to be resolved at assembly time.
  7686  		// The assembler converts from *Prog entries to absolute addresses
  7687  		// once it knows instruction byte offsets.
  7688  		fi := s.pp.CurFunc.LSym.Func()
  7689  		fi.JumpTables = append(fi.JumpTables, obj.JumpTable{Sym: jt.Aux.(*obj.LSym), Targets: targets})
  7690  	}
  7691  
  7692  	if e.log { // spew to stdout
  7693  		filename := ""
  7694  		for p := s.pp.Text; p != nil; p = p.Link {
  7695  			if p.Pos.IsKnown() && p.InnermostFilename() != filename {
  7696  				filename = p.InnermostFilename()
  7697  				f.Logf("# %s\n", filename)
  7698  			}
  7699  
  7700  			var s string
  7701  			if v, ok := progToValue[p]; ok {
  7702  				s = v.String()
  7703  			} else if b, ok := progToBlock[p]; ok {
  7704  				s = b.String()
  7705  			} else {
  7706  				s = "   " // most value and branch strings are 2-3 characters long
  7707  			}
  7708  			f.Logf(" %-6s\t%.5d (%s)\t%s\n", s, p.Pc, p.InnermostLineNumber(), p.InstructionString())
  7709  		}
  7710  	}
  7711  	if f.HTMLWriter != nil { // spew to ssa.html
  7712  		var buf strings.Builder
  7713  		buf.WriteString("<code>")
  7714  		buf.WriteString("<dl class=\"ssa-gen\">")
  7715  		filename := ""
  7716  		for p := s.pp.Text; p != nil; p = p.Link {
  7717  			// Don't spam every line with the file name, which is often huge.
  7718  			// Only print changes, and "unknown" is not a change.
  7719  			if p.Pos.IsKnown() && p.InnermostFilename() != filename {
  7720  				filename = p.InnermostFilename()
  7721  				buf.WriteString("<dt class=\"ssa-prog-src\"></dt><dd class=\"ssa-prog\">")
  7722  				buf.WriteString(html.EscapeString("# " + filename))
  7723  				buf.WriteString("</dd>")
  7724  			}
  7725  
  7726  			buf.WriteString("<dt class=\"ssa-prog-src\">")
  7727  			if v, ok := progToValue[p]; ok {
  7728  				buf.WriteString(v.HTML())
  7729  			} else if b, ok := progToBlock[p]; ok {
  7730  				buf.WriteString("<b>" + b.HTML() + "</b>")
  7731  			}
  7732  			buf.WriteString("</dt>")
  7733  			buf.WriteString("<dd class=\"ssa-prog\">")
  7734  			fmt.Fprintf(&buf, "%.5d <span class=\"l%v line-number\">(%s)</span> %s", p.Pc, p.InnermostLineNumber(), p.InnermostLineNumberHTML(), html.EscapeString(p.InstructionString()))
  7735  			buf.WriteString("</dd>")
  7736  		}
  7737  		buf.WriteString("</dl>")
  7738  		buf.WriteString("</code>")
  7739  		f.HTMLWriter.WriteColumn("genssa", "genssa", "ssa-prog", buf.String())
  7740  	}
  7741  	if ssa.GenssaDump[f.Name] {
  7742  		fi := f.DumpFileForPhase("genssa")
  7743  		if fi != nil {
  7744  
  7745  			// inliningDiffers if any filename changes or if any line number except the innermost (last index) changes.
  7746  			inliningDiffers := func(a, b []src.Pos) bool {
  7747  				if len(a) != len(b) {
  7748  					return true
  7749  				}
  7750  				for i := range a {
  7751  					if a[i].Filename() != b[i].Filename() {
  7752  						return true
  7753  					}
  7754  					if i != len(a)-1 && a[i].Line() != b[i].Line() {
  7755  						return true
  7756  					}
  7757  				}
  7758  				return false
  7759  			}
  7760  
  7761  			var allPosOld []src.Pos
  7762  			var allPos []src.Pos
  7763  
  7764  			for p := s.pp.Text; p != nil; p = p.Link {
  7765  				if p.Pos.IsKnown() {
  7766  					allPos = allPos[:0]
  7767  					p.Ctxt.AllPos(p.Pos, func(pos src.Pos) { allPos = append(allPos, pos) })
  7768  					if inliningDiffers(allPos, allPosOld) {
  7769  						for _, pos := range allPos {
  7770  							fmt.Fprintf(fi, "# %s:%d\n", pos.Filename(), pos.Line())
  7771  						}
  7772  						allPos, allPosOld = allPosOld, allPos // swap, not copy, so that they do not share slice storage.
  7773  					}
  7774  				}
  7775  
  7776  				var s string
  7777  				if v, ok := progToValue[p]; ok {
  7778  					s = v.String()
  7779  				} else if b, ok := progToBlock[p]; ok {
  7780  					s = b.String()
  7781  				} else {
  7782  					s = "   " // most value and branch strings are 2-3 characters long
  7783  				}
  7784  				fmt.Fprintf(fi, " %-6s\t%.5d %s\t%s\n", s, p.Pc, ssa.StmtString(p.Pos), p.InstructionString())
  7785  			}
  7786  			fi.Close()
  7787  		}
  7788  	}
  7789  
  7790  	defframe(&s, e, f)
  7791  
  7792  	f.HTMLWriter.Close()
  7793  	f.HTMLWriter = nil
  7794  }
  7795  
  7796  func defframe(s *State, e *ssafn, f *ssa.Func) {
  7797  	pp := s.pp
  7798  
  7799  	s.maxarg = types.RoundUp(s.maxarg, e.stkalign)
  7800  	frame := s.maxarg + e.stksize
  7801  	if Arch.PadFrame != nil {
  7802  		frame = Arch.PadFrame(frame)
  7803  	}
  7804  
  7805  	// Fill in argument and frame size.
  7806  	pp.Text.To.Type = obj.TYPE_TEXTSIZE
  7807  	pp.Text.To.Val = int32(types.RoundUp(f.OwnAux.ArgWidth(), int64(types.RegSize)))
  7808  	pp.Text.To.Offset = frame
  7809  
  7810  	p := pp.Text
  7811  
  7812  	// Insert code to spill argument registers if the named slot may be partially
  7813  	// live. That is, the named slot is considered live by liveness analysis,
  7814  	// (because a part of it is live), but we may not spill all parts into the
  7815  	// slot. This can only happen with aggregate-typed arguments that are SSA-able
  7816  	// and not address-taken (for non-SSA-able or address-taken arguments we always
  7817  	// spill upfront).
  7818  	// Note: spilling is unnecessary in the -N/no-optimize case, since all values
  7819  	// will be considered non-SSAable and spilled up front.
  7820  	// TODO(register args) Make liveness more fine-grained to that partial spilling is okay.
  7821  	if f.OwnAux.ABIInfo().InRegistersUsed() != 0 && base.Flag.N == 0 {
  7822  		// First, see if it is already spilled before it may be live. Look for a spill
  7823  		// in the entry block up to the first safepoint.
  7824  		type nameOff struct {
  7825  			n   *ir.Name
  7826  			off int64
  7827  		}
  7828  		partLiveArgsSpilled := make(map[nameOff]bool)
  7829  		for _, v := range f.Entry.Values {
  7830  			if v.Op.IsCall() {
  7831  				break
  7832  			}
  7833  			if v.Op != ssa.OpStoreReg || v.Args[0].Op != ssa.OpArgIntReg {
  7834  				continue
  7835  			}
  7836  			n, off := ssa.AutoVar(v)
  7837  			if n.Class != ir.PPARAM || n.Addrtaken() || !ssa.CanSSA(n.Type()) || !s.partLiveArgs[n] {
  7838  				continue
  7839  			}
  7840  			partLiveArgsSpilled[nameOff{n, off}] = true
  7841  		}
  7842  
  7843  		// Then, insert code to spill registers if not already.
  7844  		for _, a := range f.OwnAux.ABIInfo().InParams() {
  7845  			n := a.Name
  7846  			if n == nil || n.Addrtaken() || !ssa.CanSSA(n.Type()) || !s.partLiveArgs[n] || len(a.Registers) <= 1 {
  7847  				continue
  7848  			}
  7849  			rts, offs := a.RegisterTypesAndOffsets()
  7850  			for i := range a.Registers {
  7851  				if !rts[i].HasPointers() {
  7852  					continue
  7853  				}
  7854  				if partLiveArgsSpilled[nameOff{n, offs[i]}] {
  7855  					continue // already spilled
  7856  				}
  7857  				reg := ssa.ObjRegForAbiReg(a.Registers[i], f.Config)
  7858  				p = Arch.SpillArgReg(pp, p, f, rts[i], reg, n, offs[i])
  7859  			}
  7860  		}
  7861  	}
  7862  
  7863  	// Insert code to zero ambiguously live variables so that the
  7864  	// garbage collector only sees initialized values when it
  7865  	// looks for pointers.
  7866  	var lo, hi int64
  7867  
  7868  	// Opaque state for backend to use. Current backends use it to
  7869  	// keep track of which helper registers have been zeroed.
  7870  	var state uint32
  7871  
  7872  	// Iterate through declarations. Autos are sorted in decreasing
  7873  	// frame offset order.
  7874  	for _, n := range e.curfn.Dcl {
  7875  		if !n.Needzero() {
  7876  			continue
  7877  		}
  7878  		if n.Class != ir.PAUTO {
  7879  			e.Fatalf(n.Pos(), "needzero class %d", n.Class)
  7880  		}
  7881  		if n.Type().Size()%int64(types.PtrSize) != 0 || n.FrameOffset()%int64(types.PtrSize) != 0 || n.Type().Size() == 0 {
  7882  			e.Fatalf(n.Pos(), "var %L has size %d offset %d", n, n.Type().Size(), n.Offset_)
  7883  		}
  7884  
  7885  		if lo != hi && n.FrameOffset()+n.Type().Size() >= lo-int64(2*types.RegSize) {
  7886  			// Merge with range we already have.
  7887  			lo = n.FrameOffset()
  7888  			continue
  7889  		}
  7890  
  7891  		// Zero old range
  7892  		p = Arch.ZeroRange(pp, p, frame+lo, hi-lo, &state)
  7893  
  7894  		// Set new range.
  7895  		lo = n.FrameOffset()
  7896  		hi = lo + n.Type().Size()
  7897  	}
  7898  
  7899  	// Zero final range.
  7900  	Arch.ZeroRange(pp, p, frame+lo, hi-lo, &state)
  7901  }
  7902  
  7903  // For generating consecutive jump instructions to model a specific branching
  7904  type IndexJump struct {
  7905  	Jump  obj.As
  7906  	Index int
  7907  }
  7908  
  7909  func (s *State) oneJump(b *ssa.Block, jump *IndexJump) {
  7910  	p := s.Br(jump.Jump, b.Succs[jump.Index].Block())
  7911  	p.Pos = b.Pos
  7912  }
  7913  
  7914  // CombJump generates combinational instructions (2 at present) for a block jump,
  7915  // thereby the behaviour of non-standard condition codes could be simulated
  7916  func (s *State) CombJump(b, next *ssa.Block, jumps *[2][2]IndexJump) {
  7917  	switch next {
  7918  	case b.Succs[0].Block():
  7919  		s.oneJump(b, &jumps[0][0])
  7920  		s.oneJump(b, &jumps[0][1])
  7921  	case b.Succs[1].Block():
  7922  		s.oneJump(b, &jumps[1][0])
  7923  		s.oneJump(b, &jumps[1][1])
  7924  	default:
  7925  		var q *obj.Prog
  7926  		if b.Likely != ssa.BranchUnlikely {
  7927  			s.oneJump(b, &jumps[1][0])
  7928  			s.oneJump(b, &jumps[1][1])
  7929  			q = s.Br(obj.AJMP, b.Succs[1].Block())
  7930  		} else {
  7931  			s.oneJump(b, &jumps[0][0])
  7932  			s.oneJump(b, &jumps[0][1])
  7933  			q = s.Br(obj.AJMP, b.Succs[0].Block())
  7934  		}
  7935  		q.Pos = b.Pos
  7936  	}
  7937  }
  7938  
  7939  // AddAux adds the offset in the aux fields (AuxInt and Aux) of v to a.
  7940  func AddAux(a *obj.Addr, v *ssa.Value) {
  7941  	AddAux2(a, v, v.AuxInt)
  7942  }
  7943  func AddAux2(a *obj.Addr, v *ssa.Value, offset int64) {
  7944  	if a.Type != obj.TYPE_MEM && a.Type != obj.TYPE_ADDR {
  7945  		v.Fatalf("bad AddAux addr %v", a)
  7946  	}
  7947  	// add integer offset
  7948  	a.Offset += offset
  7949  
  7950  	// If no additional symbol offset, we're done.
  7951  	if v.Aux == nil {
  7952  		return
  7953  	}
  7954  	// Add symbol's offset from its base register.
  7955  	switch n := v.Aux.(type) {
  7956  	case *ssa.AuxCall:
  7957  		a.Name = obj.NAME_EXTERN
  7958  		a.Sym = n.Fn
  7959  	case *obj.LSym:
  7960  		a.Name = obj.NAME_EXTERN
  7961  		a.Sym = n
  7962  	case *ir.Name:
  7963  		if n.Class == ir.PPARAM || (n.Class == ir.PPARAMOUT && !n.IsOutputParamInRegisters()) {
  7964  			a.Name = obj.NAME_PARAM
  7965  		} else {
  7966  			a.Name = obj.NAME_AUTO
  7967  		}
  7968  		a.Sym = n.Linksym()
  7969  		a.Offset += n.FrameOffset()
  7970  	default:
  7971  		v.Fatalf("aux in %s not implemented %#v", v, v.Aux)
  7972  	}
  7973  }
  7974  
  7975  // extendIndex extends v to a full int width.
  7976  // panic with the given kind if v does not fit in an int (only on 32-bit archs).
  7977  func (s *state) extendIndex(idx, len *ssa.Value, kind ssa.BoundsKind, bounded bool) *ssa.Value {
  7978  	size := idx.Type.Size()
  7979  	if size == s.config.PtrSize {
  7980  		return idx
  7981  	}
  7982  	if size > s.config.PtrSize {
  7983  		// truncate 64-bit indexes on 32-bit pointer archs. Test the
  7984  		// high word and branch to out-of-bounds failure if it is not 0.
  7985  		var lo *ssa.Value
  7986  		if idx.Type.IsSigned() {
  7987  			lo = s.newValue1(ssa.OpInt64Lo, types.Types[types.TINT], idx)
  7988  		} else {
  7989  			lo = s.newValue1(ssa.OpInt64Lo, types.Types[types.TUINT], idx)
  7990  		}
  7991  		if bounded || base.Flag.B != 0 {
  7992  			return lo
  7993  		}
  7994  		bNext := s.f.NewBlock(ssa.BlockPlain)
  7995  		bPanic := s.f.NewBlock(ssa.BlockExit)
  7996  		hi := s.newValue1(ssa.OpInt64Hi, types.Types[types.TUINT32], idx)
  7997  		cmp := s.newValue2(ssa.OpEq32, types.Types[types.TBOOL], hi, s.constInt32(types.Types[types.TUINT32], 0))
  7998  		if !idx.Type.IsSigned() {
  7999  			switch kind {
  8000  			case ssa.BoundsIndex:
  8001  				kind = ssa.BoundsIndexU
  8002  			case ssa.BoundsSliceAlen:
  8003  				kind = ssa.BoundsSliceAlenU
  8004  			case ssa.BoundsSliceAcap:
  8005  				kind = ssa.BoundsSliceAcapU
  8006  			case ssa.BoundsSliceB:
  8007  				kind = ssa.BoundsSliceBU
  8008  			case ssa.BoundsSlice3Alen:
  8009  				kind = ssa.BoundsSlice3AlenU
  8010  			case ssa.BoundsSlice3Acap:
  8011  				kind = ssa.BoundsSlice3AcapU
  8012  			case ssa.BoundsSlice3B:
  8013  				kind = ssa.BoundsSlice3BU
  8014  			case ssa.BoundsSlice3C:
  8015  				kind = ssa.BoundsSlice3CU
  8016  			}
  8017  		}
  8018  		b := s.endBlock()
  8019  		b.Kind = ssa.BlockIf
  8020  		b.SetControl(cmp)
  8021  		b.Likely = ssa.BranchLikely
  8022  		b.AddEdgeTo(bNext)
  8023  		b.AddEdgeTo(bPanic)
  8024  
  8025  		s.startBlock(bPanic)
  8026  		mem := s.newValue4I(ssa.OpPanicExtend, types.TypeMem, int64(kind), hi, lo, len, s.mem())
  8027  		s.endBlock().SetControl(mem)
  8028  		s.startBlock(bNext)
  8029  
  8030  		return lo
  8031  	}
  8032  
  8033  	// Extend value to the required size
  8034  	var op ssa.Op
  8035  	if idx.Type.IsSigned() {
  8036  		switch 10*size + s.config.PtrSize {
  8037  		case 14:
  8038  			op = ssa.OpSignExt8to32
  8039  		case 18:
  8040  			op = ssa.OpSignExt8to64
  8041  		case 24:
  8042  			op = ssa.OpSignExt16to32
  8043  		case 28:
  8044  			op = ssa.OpSignExt16to64
  8045  		case 48:
  8046  			op = ssa.OpSignExt32to64
  8047  		default:
  8048  			s.Fatalf("bad signed index extension %s", idx.Type)
  8049  		}
  8050  	} else {
  8051  		switch 10*size + s.config.PtrSize {
  8052  		case 14:
  8053  			op = ssa.OpZeroExt8to32
  8054  		case 18:
  8055  			op = ssa.OpZeroExt8to64
  8056  		case 24:
  8057  			op = ssa.OpZeroExt16to32
  8058  		case 28:
  8059  			op = ssa.OpZeroExt16to64
  8060  		case 48:
  8061  			op = ssa.OpZeroExt32to64
  8062  		default:
  8063  			s.Fatalf("bad unsigned index extension %s", idx.Type)
  8064  		}
  8065  	}
  8066  	return s.newValue1(op, types.Types[types.TINT], idx)
  8067  }
  8068  
  8069  // CheckLoweredPhi checks that regalloc and stackalloc correctly handled phi values.
  8070  // Called during ssaGenValue.
  8071  func CheckLoweredPhi(v *ssa.Value) {
  8072  	if v.Op != ssa.OpPhi {
  8073  		v.Fatalf("CheckLoweredPhi called with non-phi value: %v", v.LongString())
  8074  	}
  8075  	if v.Type.IsMemory() {
  8076  		return
  8077  	}
  8078  	f := v.Block.Func
  8079  	loc := f.RegAlloc[v.ID]
  8080  	for _, a := range v.Args {
  8081  		if aloc := f.RegAlloc[a.ID]; aloc != loc { // TODO: .Equal() instead?
  8082  			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)
  8083  		}
  8084  	}
  8085  }
  8086  
  8087  // CheckLoweredGetClosurePtr checks that v is the first instruction in the function's entry block,
  8088  // except for incoming in-register arguments.
  8089  // The output of LoweredGetClosurePtr is generally hardwired to the correct register.
  8090  // That register contains the closure pointer on closure entry.
  8091  func CheckLoweredGetClosurePtr(v *ssa.Value) {
  8092  	entry := v.Block.Func.Entry
  8093  	if entry != v.Block {
  8094  		base.Fatalf("in %s, badly placed LoweredGetClosurePtr: %v %v", v.Block.Func.Name, v.Block, v)
  8095  	}
  8096  	for _, w := range entry.Values {
  8097  		if w == v {
  8098  			break
  8099  		}
  8100  		switch w.Op {
  8101  		case ssa.OpArgIntReg, ssa.OpArgFloatReg:
  8102  			// okay
  8103  		default:
  8104  			base.Fatalf("in %s, badly placed LoweredGetClosurePtr: %v %v", v.Block.Func.Name, v.Block, v)
  8105  		}
  8106  	}
  8107  }
  8108  
  8109  // CheckArgReg ensures that v is in the function's entry block.
  8110  func CheckArgReg(v *ssa.Value) {
  8111  	entry := v.Block.Func.Entry
  8112  	if entry != v.Block {
  8113  		base.Fatalf("in %s, badly placed ArgIReg or ArgFReg: %v %v", v.Block.Func.Name, v.Block, v)
  8114  	}
  8115  }
  8116  
  8117  func AddrAuto(a *obj.Addr, v *ssa.Value) {
  8118  	n, off := ssa.AutoVar(v)
  8119  	a.Type = obj.TYPE_MEM
  8120  	a.Sym = n.Linksym()
  8121  	a.Reg = int16(Arch.REGSP)
  8122  	a.Offset = n.FrameOffset() + off
  8123  	if n.Class == ir.PPARAM || (n.Class == ir.PPARAMOUT && !n.IsOutputParamInRegisters()) {
  8124  		a.Name = obj.NAME_PARAM
  8125  	} else {
  8126  		a.Name = obj.NAME_AUTO
  8127  	}
  8128  }
  8129  
  8130  // Call returns a new CALL instruction for the SSA value v.
  8131  // It uses PrepareCall to prepare the call.
  8132  func (s *State) Call(v *ssa.Value) *obj.Prog {
  8133  	pPosIsStmt := s.pp.Pos.IsStmt() // The statement-ness fo the call comes from ssaGenState
  8134  	s.PrepareCall(v)
  8135  
  8136  	p := s.Prog(obj.ACALL)
  8137  	if pPosIsStmt == src.PosIsStmt {
  8138  		p.Pos = v.Pos.WithIsStmt()
  8139  	} else {
  8140  		p.Pos = v.Pos.WithNotStmt()
  8141  	}
  8142  	if sym, ok := v.Aux.(*ssa.AuxCall); ok && sym.Fn != nil {
  8143  		p.To.Type = obj.TYPE_MEM
  8144  		p.To.Name = obj.NAME_EXTERN
  8145  		p.To.Sym = sym.Fn
  8146  	} else {
  8147  		// TODO(mdempsky): Can these differences be eliminated?
  8148  		switch Arch.LinkArch.Family {
  8149  		case sys.AMD64, sys.I386, sys.PPC64, sys.RISCV64, sys.S390X, sys.Wasm:
  8150  			p.To.Type = obj.TYPE_REG
  8151  		case sys.ARM, sys.ARM64, sys.Loong64, sys.MIPS, sys.MIPS64:
  8152  			p.To.Type = obj.TYPE_MEM
  8153  		default:
  8154  			base.Fatalf("unknown indirect call family")
  8155  		}
  8156  		p.To.Reg = v.Args[0].Reg()
  8157  	}
  8158  	return p
  8159  }
  8160  
  8161  // TailCall returns a new tail call instruction for the SSA value v.
  8162  // It is like Call, but for a tail call.
  8163  func (s *State) TailCall(v *ssa.Value) *obj.Prog {
  8164  	p := s.Call(v)
  8165  	p.As = obj.ARET
  8166  	return p
  8167  }
  8168  
  8169  // PrepareCall prepares to emit a CALL instruction for v and does call-related bookkeeping.
  8170  // It must be called immediately before emitting the actual CALL instruction,
  8171  // since it emits PCDATA for the stack map at the call (calls are safe points).
  8172  func (s *State) PrepareCall(v *ssa.Value) {
  8173  	idx := s.livenessMap.Get(v)
  8174  	if !idx.StackMapValid() {
  8175  		// See Liveness.hasStackMap.
  8176  		if sym, ok := v.Aux.(*ssa.AuxCall); !ok || !(sym.Fn == ir.Syms.WBZero || sym.Fn == ir.Syms.WBMove) {
  8177  			base.Fatalf("missing stack map index for %v", v.LongString())
  8178  		}
  8179  	}
  8180  
  8181  	call, ok := v.Aux.(*ssa.AuxCall)
  8182  
  8183  	if ok {
  8184  		// Record call graph information for nowritebarrierrec
  8185  		// analysis.
  8186  		if nowritebarrierrecCheck != nil {
  8187  			nowritebarrierrecCheck.recordCall(s.pp.CurFunc, call.Fn, v.Pos)
  8188  		}
  8189  	}
  8190  
  8191  	if s.maxarg < v.AuxInt {
  8192  		s.maxarg = v.AuxInt
  8193  	}
  8194  }
  8195  
  8196  // UseArgs records the fact that an instruction needs a certain amount of
  8197  // callee args space for its use.
  8198  func (s *State) UseArgs(n int64) {
  8199  	if s.maxarg < n {
  8200  		s.maxarg = n
  8201  	}
  8202  }
  8203  
  8204  // fieldIdx finds the index of the field referred to by the ODOT node n.
  8205  func fieldIdx(n *ir.SelectorExpr) int {
  8206  	t := n.X.Type()
  8207  	if !t.IsStruct() {
  8208  		panic("ODOT's LHS is not a struct")
  8209  	}
  8210  
  8211  	for i, f := range t.Fields() {
  8212  		if f.Sym == n.Sel {
  8213  			if f.Offset != n.Offset() {
  8214  				panic("field offset doesn't match")
  8215  			}
  8216  			return i
  8217  		}
  8218  	}
  8219  	panic(fmt.Sprintf("can't find field in expr %v\n", n))
  8220  
  8221  	// TODO: keep the result of this function somewhere in the ODOT Node
  8222  	// so we don't have to recompute it each time we need it.
  8223  }
  8224  
  8225  // ssafn holds frontend information about a function that the backend is processing.
  8226  // It also exports a bunch of compiler services for the ssa backend.
  8227  type ssafn struct {
  8228  	curfn      *ir.Func
  8229  	strings    map[string]*obj.LSym // map from constant string to data symbols
  8230  	stksize    int64                // stack size for current frame
  8231  	stkptrsize int64                // prefix of stack containing pointers
  8232  
  8233  	// alignment for current frame.
  8234  	// NOTE: when stkalign > PtrSize, currently this only ensures the offsets of
  8235  	// objects in the stack frame are aligned. The stack pointer is still aligned
  8236  	// only PtrSize.
  8237  	stkalign int64
  8238  
  8239  	log bool // print ssa debug to the stdout
  8240  }
  8241  
  8242  // StringData returns a symbol which
  8243  // is the data component of a global string constant containing s.
  8244  func (e *ssafn) StringData(s string) *obj.LSym {
  8245  	if aux, ok := e.strings[s]; ok {
  8246  		return aux
  8247  	}
  8248  	if e.strings == nil {
  8249  		e.strings = make(map[string]*obj.LSym)
  8250  	}
  8251  	data := staticdata.StringSym(e.curfn.Pos(), s)
  8252  	e.strings[s] = data
  8253  	return data
  8254  }
  8255  
  8256  // SplitSlot returns a slot representing the data of parent starting at offset.
  8257  func (e *ssafn) SplitSlot(parent *ssa.LocalSlot, suffix string, offset int64, t *types.Type) ssa.LocalSlot {
  8258  	node := parent.N
  8259  
  8260  	if node.Class != ir.PAUTO || node.Addrtaken() {
  8261  		// addressed things and non-autos retain their parents (i.e., cannot truly be split)
  8262  		return ssa.LocalSlot{N: node, Type: t, Off: parent.Off + offset}
  8263  	}
  8264  
  8265  	sym := &types.Sym{Name: node.Sym().Name + suffix, Pkg: types.LocalPkg}
  8266  	n := e.curfn.NewLocal(parent.N.Pos(), sym, t)
  8267  	n.SetUsed(true)
  8268  	n.SetEsc(ir.EscNever)
  8269  	types.CalcSize(t)
  8270  	return ssa.LocalSlot{N: n, Type: t, Off: 0, SplitOf: parent, SplitOffset: offset}
  8271  }
  8272  
  8273  // Logf logs a message from the compiler.
  8274  func (e *ssafn) Logf(msg string, args ...interface{}) {
  8275  	if e.log {
  8276  		fmt.Printf(msg, args...)
  8277  	}
  8278  }
  8279  
  8280  func (e *ssafn) Log() bool {
  8281  	return e.log
  8282  }
  8283  
  8284  // Fatalf reports a compiler error and exits.
  8285  func (e *ssafn) Fatalf(pos src.XPos, msg string, args ...interface{}) {
  8286  	base.Pos = pos
  8287  	nargs := append([]interface{}{ir.FuncName(e.curfn)}, args...)
  8288  	base.Fatalf("'%s': "+msg, nargs...)
  8289  }
  8290  
  8291  // Warnl reports a "warning", which is usually flag-triggered
  8292  // logging output for the benefit of tests.
  8293  func (e *ssafn) Warnl(pos src.XPos, fmt_ string, args ...interface{}) {
  8294  	base.WarnfAt(pos, fmt_, args...)
  8295  }
  8296  
  8297  func (e *ssafn) Debug_checknil() bool {
  8298  	return base.Debug.Nil != 0
  8299  }
  8300  
  8301  func (e *ssafn) UseWriteBarrier() bool {
  8302  	return base.Flag.WB
  8303  }
  8304  
  8305  func (e *ssafn) Syslook(name string) *obj.LSym {
  8306  	switch name {
  8307  	case "goschedguarded":
  8308  		return ir.Syms.Goschedguarded
  8309  	case "writeBarrier":
  8310  		return ir.Syms.WriteBarrier
  8311  	case "wbZero":
  8312  		return ir.Syms.WBZero
  8313  	case "wbMove":
  8314  		return ir.Syms.WBMove
  8315  	case "cgoCheckMemmove":
  8316  		return ir.Syms.CgoCheckMemmove
  8317  	case "cgoCheckPtrWrite":
  8318  		return ir.Syms.CgoCheckPtrWrite
  8319  	}
  8320  	e.Fatalf(src.NoXPos, "unknown Syslook func %v", name)
  8321  	return nil
  8322  }
  8323  
  8324  func (e *ssafn) Func() *ir.Func {
  8325  	return e.curfn
  8326  }
  8327  
  8328  func clobberBase(n ir.Node) ir.Node {
  8329  	if n.Op() == ir.ODOT {
  8330  		n := n.(*ir.SelectorExpr)
  8331  		if n.X.Type().NumFields() == 1 {
  8332  			return clobberBase(n.X)
  8333  		}
  8334  	}
  8335  	if n.Op() == ir.OINDEX {
  8336  		n := n.(*ir.IndexExpr)
  8337  		if n.X.Type().IsArray() && n.X.Type().NumElem() == 1 {
  8338  			return clobberBase(n.X)
  8339  		}
  8340  	}
  8341  	return n
  8342  }
  8343  
  8344  // callTargetLSym returns the correct LSym to call 'callee' using its ABI.
  8345  func callTargetLSym(callee *ir.Name) *obj.LSym {
  8346  	if callee.Func == nil {
  8347  		// TODO(austin): This happens in case of interface method I.M from imported package.
  8348  		// It's ABIInternal, and would be better if callee.Func was never nil and we didn't
  8349  		// need this case.
  8350  		return callee.Linksym()
  8351  	}
  8352  
  8353  	return callee.LinksymABI(callee.Func.ABI)
  8354  }
  8355  
  8356  func min8(a, b int8) int8 {
  8357  	if a < b {
  8358  		return a
  8359  	}
  8360  	return b
  8361  }
  8362  
  8363  func max8(a, b int8) int8 {
  8364  	if a > b {
  8365  		return a
  8366  	}
  8367  	return b
  8368  }
  8369  
  8370  // deferStructFnField is the field index of _defer.fn.
  8371  const deferStructFnField = 4
  8372  
  8373  var deferType *types.Type
  8374  
  8375  // deferstruct returns a type interchangeable with runtime._defer.
  8376  // Make sure this stays in sync with runtime/runtime2.go:_defer.
  8377  func deferstruct() *types.Type {
  8378  	if deferType != nil {
  8379  		return deferType
  8380  	}
  8381  
  8382  	makefield := func(name string, t *types.Type) *types.Field {
  8383  		sym := (*types.Pkg)(nil).Lookup(name)
  8384  		return types.NewField(src.NoXPos, sym, t)
  8385  	}
  8386  
  8387  	fields := []*types.Field{
  8388  		makefield("heap", types.Types[types.TBOOL]),
  8389  		makefield("rangefunc", types.Types[types.TBOOL]),
  8390  		makefield("sp", types.Types[types.TUINTPTR]),
  8391  		makefield("pc", types.Types[types.TUINTPTR]),
  8392  		// Note: the types here don't really matter. Defer structures
  8393  		// are always scanned explicitly during stack copying and GC,
  8394  		// so we make them uintptr type even though they are real pointers.
  8395  		makefield("fn", types.Types[types.TUINTPTR]),
  8396  		makefield("link", types.Types[types.TUINTPTR]),
  8397  		makefield("head", types.Types[types.TUINTPTR]),
  8398  	}
  8399  	if name := fields[deferStructFnField].Sym.Name; name != "fn" {
  8400  		base.Fatalf("deferStructFnField is %q, not fn", name)
  8401  	}
  8402  
  8403  	n := ir.NewDeclNameAt(src.NoXPos, ir.OTYPE, ir.Pkgs.Runtime.Lookup("_defer"))
  8404  	typ := types.NewNamed(n)
  8405  	n.SetType(typ)
  8406  	n.SetTypecheck(1)
  8407  
  8408  	// build struct holding the above fields
  8409  	typ.SetUnderlying(types.NewStruct(fields))
  8410  	types.CalcStructSize(typ)
  8411  
  8412  	deferType = typ
  8413  	return typ
  8414  }
  8415  
  8416  // SpillSlotAddr uses LocalSlot information to initialize an obj.Addr
  8417  // The resulting addr is used in a non-standard context -- in the prologue
  8418  // of a function, before the frame has been constructed, so the standard
  8419  // addressing for the parameters will be wrong.
  8420  func SpillSlotAddr(spill ssa.Spill, baseReg int16, extraOffset int64) obj.Addr {
  8421  	return obj.Addr{
  8422  		Name:   obj.NAME_NONE,
  8423  		Type:   obj.TYPE_MEM,
  8424  		Reg:    baseReg,
  8425  		Offset: spill.Offset + extraOffset,
  8426  	}
  8427  }
  8428  
  8429  var (
  8430  	BoundsCheckFunc [ssa.BoundsKindCount]*obj.LSym
  8431  	ExtendCheckFunc [ssa.BoundsKindCount]*obj.LSym
  8432  )
  8433  

View as plain text