Source file src/cmd/internal/obj/wasm/wasmobj.go

     1  // Copyright 2018 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 wasm
     6  
     7  import (
     8  	"bytes"
     9  	"cmd/internal/obj"
    10  	"cmd/internal/objabi"
    11  	"cmd/internal/sys"
    12  	"encoding/binary"
    13  	"fmt"
    14  	"internal/abi"
    15  	"io"
    16  	"math"
    17  )
    18  
    19  var Register = map[string]int16{
    20  	"SP":    REG_SP,
    21  	"CTXT":  REG_CTXT,
    22  	"g":     REG_g,
    23  	"RET0":  REG_RET0,
    24  	"RET1":  REG_RET1,
    25  	"RET2":  REG_RET2,
    26  	"RET3":  REG_RET3,
    27  	"PAUSE": REG_PAUSE,
    28  
    29  	"R0":  REG_R0,
    30  	"R1":  REG_R1,
    31  	"R2":  REG_R2,
    32  	"R3":  REG_R3,
    33  	"R4":  REG_R4,
    34  	"R5":  REG_R5,
    35  	"R6":  REG_R6,
    36  	"R7":  REG_R7,
    37  	"R8":  REG_R8,
    38  	"R9":  REG_R9,
    39  	"R10": REG_R10,
    40  	"R11": REG_R11,
    41  	"R12": REG_R12,
    42  	"R13": REG_R13,
    43  	"R14": REG_R14,
    44  	"R15": REG_R15,
    45  
    46  	"F0":  REG_F0,
    47  	"F1":  REG_F1,
    48  	"F2":  REG_F2,
    49  	"F3":  REG_F3,
    50  	"F4":  REG_F4,
    51  	"F5":  REG_F5,
    52  	"F6":  REG_F6,
    53  	"F7":  REG_F7,
    54  	"F8":  REG_F8,
    55  	"F9":  REG_F9,
    56  	"F10": REG_F10,
    57  	"F11": REG_F11,
    58  	"F12": REG_F12,
    59  	"F13": REG_F13,
    60  	"F14": REG_F14,
    61  	"F15": REG_F15,
    62  
    63  	"F16": REG_F16,
    64  	"F17": REG_F17,
    65  	"F18": REG_F18,
    66  	"F19": REG_F19,
    67  	"F20": REG_F20,
    68  	"F21": REG_F21,
    69  	"F22": REG_F22,
    70  	"F23": REG_F23,
    71  	"F24": REG_F24,
    72  	"F25": REG_F25,
    73  	"F26": REG_F26,
    74  	"F27": REG_F27,
    75  	"F28": REG_F28,
    76  	"F29": REG_F29,
    77  	"F30": REG_F30,
    78  	"F31": REG_F31,
    79  
    80  	"PC_B": REG_PC_B,
    81  }
    82  
    83  var registerNames []string
    84  
    85  func init() {
    86  	obj.RegisterRegister(MINREG, MAXREG, rconv)
    87  	obj.RegisterOpcode(obj.ABaseWasm, Anames)
    88  
    89  	registerNames = make([]string, MAXREG-MINREG)
    90  	for name, reg := range Register {
    91  		registerNames[reg-MINREG] = name
    92  	}
    93  }
    94  
    95  func rconv(r int) string {
    96  	return registerNames[r-MINREG]
    97  }
    98  
    99  var unaryDst = map[obj.As]bool{
   100  	ASet:          true,
   101  	ATee:          true,
   102  	ACall:         true,
   103  	ACallIndirect: true,
   104  	ABr:           true,
   105  	ABrIf:         true,
   106  	ABrTable:      true,
   107  	AI32Store:     true,
   108  	AI64Store:     true,
   109  	AF32Store:     true,
   110  	AF64Store:     true,
   111  	AI32Store8:    true,
   112  	AI32Store16:   true,
   113  	AI64Store8:    true,
   114  	AI64Store16:   true,
   115  	AI64Store32:   true,
   116  	ACALLNORESUME: true,
   117  }
   118  
   119  var Linkwasm = obj.LinkArch{
   120  	Arch:       sys.ArchWasm,
   121  	Init:       instinit,
   122  	Preprocess: preprocess,
   123  	Assemble:   assemble,
   124  	UnaryDst:   unaryDst,
   125  }
   126  
   127  var (
   128  	morestack       *obj.LSym
   129  	morestackNoCtxt *obj.LSym
   130  	sigpanic        *obj.LSym
   131  )
   132  
   133  const (
   134  	/* mark flags */
   135  	WasmImport = 1 << 0
   136  )
   137  
   138  const (
   139  	// This is a special wasm module name that when used as the module name
   140  	// in //go:wasmimport will cause the generated code to pass the stack pointer
   141  	// directly to the imported function. In other words, any function that
   142  	// uses the gojs module understands the internal Go WASM ABI directly.
   143  	GojsModule = "gojs"
   144  )
   145  
   146  func instinit(ctxt *obj.Link) {
   147  	morestack = ctxt.Lookup("runtime.morestack")
   148  	morestackNoCtxt = ctxt.Lookup("runtime.morestack_noctxt")
   149  	sigpanic = ctxt.LookupABI("runtime.sigpanic", obj.ABIInternal)
   150  }
   151  
   152  func preprocess(ctxt *obj.Link, s *obj.LSym, newprog obj.ProgAlloc) {
   153  	appendp := func(p *obj.Prog, as obj.As, args ...obj.Addr) *obj.Prog {
   154  		if p.As != obj.ANOP {
   155  			p2 := obj.Appendp(p, newprog)
   156  			p2.Pc = p.Pc
   157  			p = p2
   158  		}
   159  		p.As = as
   160  		switch len(args) {
   161  		case 0:
   162  			p.From = obj.Addr{}
   163  			p.To = obj.Addr{}
   164  		case 1:
   165  			if unaryDst[as] {
   166  				p.From = obj.Addr{}
   167  				p.To = args[0]
   168  			} else {
   169  				p.From = args[0]
   170  				p.To = obj.Addr{}
   171  			}
   172  		case 2:
   173  			p.From = args[0]
   174  			p.To = args[1]
   175  		default:
   176  			panic("bad args")
   177  		}
   178  		return p
   179  	}
   180  
   181  	framesize := s.Func().Text.To.Offset
   182  	if framesize < 0 {
   183  		panic("bad framesize")
   184  	}
   185  	s.Func().Args = s.Func().Text.To.Val.(int32)
   186  	s.Func().Locals = int32(framesize)
   187  
   188  	// If the function exits just to call out to a wasmimport, then
   189  	// generate the code to translate from our internal Go-stack
   190  	// based call convention to the native webassembly call convention.
   191  	if wi := s.Func().WasmImport; wi != nil {
   192  		s.Func().WasmImportSym = wi.CreateSym(ctxt)
   193  		p := s.Func().Text
   194  		if p.Link != nil {
   195  			panic("wrapper functions for WASM imports should not have a body")
   196  		}
   197  		to := obj.Addr{
   198  			Type: obj.TYPE_MEM,
   199  			Name: obj.NAME_EXTERN,
   200  			Sym:  s,
   201  		}
   202  
   203  		// If the module that the import is for is our magic "gojs" module, then this
   204  		// indicates that the called function understands the Go stack-based call convention
   205  		// so we just pass the stack pointer to it, knowing it will read the params directly
   206  		// off the stack and push the results into memory based on the stack pointer.
   207  		if wi.Module == GojsModule {
   208  			// The called function has a signature of 'func(sp int)'. It has access to the memory
   209  			// value somewhere to be able to address the memory based on the "sp" value.
   210  
   211  			p = appendp(p, AGet, regAddr(REG_SP))
   212  			p = appendp(p, ACall, to)
   213  
   214  			p.Mark = WasmImport
   215  		} else {
   216  			if len(wi.Results) > 1 {
   217  				// TODO(evanphx) implement support for the multi-value proposal:
   218  				// https://github.com/WebAssembly/multi-value/blob/master/proposals/multi-value/Overview.md
   219  				panic("invalid results type") // impossible until multi-value proposal has landed
   220  			}
   221  			if len(wi.Results) == 1 {
   222  				// If we have a result (rather than returning nothing at all), then
   223  				// we'll write the result to the Go stack relative to the current stack pointer.
   224  				// We cache the current stack pointer value on the wasm stack here and then use
   225  				// it after the Call instruction to store the result.
   226  				p = appendp(p, AGet, regAddr(REG_SP))
   227  			}
   228  			for _, f := range wi.Params {
   229  				// Each load instructions will consume the value of sp on the stack, so
   230  				// we need to read sp for each param. WASM appears to not have a stack dup instruction
   231  				// (a strange omission for a stack-based VM), if it did, we'd be using the dup here.
   232  				p = appendp(p, AGet, regAddr(REG_SP))
   233  
   234  				// Offset is the location of the param on the Go stack (ie relative to sp).
   235  				// Because of our call convention, the parameters are located an additional 8 bytes
   236  				// from sp because we store the return address as an int64 at the bottom of the stack.
   237  				// Ie the stack looks like [return_addr, param3, param2, param1, etc]
   238  
   239  				// Ergo, we add 8 to the true byte offset of the param to skip the return address.
   240  				loadOffset := f.Offset + 8
   241  
   242  				// We're reading the value from the Go stack onto the WASM stack and leaving it there
   243  				// for CALL to pick them up.
   244  				switch f.Type {
   245  				case obj.WasmI32:
   246  					p = appendp(p, AI32Load, constAddr(loadOffset))
   247  				case obj.WasmI64:
   248  					p = appendp(p, AI64Load, constAddr(loadOffset))
   249  				case obj.WasmF32:
   250  					p = appendp(p, AF32Load, constAddr(loadOffset))
   251  				case obj.WasmF64:
   252  					p = appendp(p, AF64Load, constAddr(loadOffset))
   253  				case obj.WasmPtr:
   254  					p = appendp(p, AI64Load, constAddr(loadOffset))
   255  					p = appendp(p, AI32WrapI64)
   256  				default:
   257  					panic("bad param type")
   258  				}
   259  			}
   260  
   261  			// The call instruction is marked as being for a wasm import so that a later phase
   262  			// will generate relocation information that allows us to patch this with then
   263  			// offset of the imported function in the wasm imports.
   264  			p = appendp(p, ACall, to)
   265  			p.Mark = WasmImport
   266  
   267  			if len(wi.Results) == 1 {
   268  				f := wi.Results[0]
   269  
   270  				// Much like with the params, we need to adjust the offset we store the result value
   271  				// to by 8 bytes to account for the return address on the Go stack.
   272  				storeOffset := f.Offset + 8
   273  
   274  				// This code is paired the code above that reads the stack pointer onto the wasm
   275  				// stack. We've done this so we have a consistent view of the sp value as it might
   276  				// be manipulated by the call and we want to ignore that manipulation here.
   277  				switch f.Type {
   278  				case obj.WasmI32:
   279  					p = appendp(p, AI32Store, constAddr(storeOffset))
   280  				case obj.WasmI64:
   281  					p = appendp(p, AI64Store, constAddr(storeOffset))
   282  				case obj.WasmF32:
   283  					p = appendp(p, AF32Store, constAddr(storeOffset))
   284  				case obj.WasmF64:
   285  					p = appendp(p, AF64Store, constAddr(storeOffset))
   286  				case obj.WasmPtr:
   287  					p = appendp(p, AI64ExtendI32U)
   288  					p = appendp(p, AI64Store, constAddr(storeOffset))
   289  				default:
   290  					panic("bad result type")
   291  				}
   292  			}
   293  		}
   294  
   295  		p = appendp(p, obj.ARET)
   296  
   297  		// It should be 0 already, but we'll set it to 0 anyway just to be sure
   298  		// that the code below which adds frame expansion code to the function body
   299  		// isn't run. We don't want the frame expansion code because our function
   300  		// body is just the code to translate and call the imported function.
   301  		framesize = 0
   302  	} else if s.Func().Text.From.Sym.Wrapper() {
   303  		// if g._panic != nil && g._panic.argp == FP {
   304  		//   g._panic.argp = bottom-of-frame
   305  		// }
   306  		//
   307  		// MOVD g_panic(g), R0
   308  		// Get R0
   309  		// I64Eqz
   310  		// Not
   311  		// If
   312  		//   Get SP
   313  		//   I64ExtendI32U
   314  		//   I64Const $framesize+8
   315  		//   I64Add
   316  		//   I64Load panic_argp(R0)
   317  		//   I64Eq
   318  		//   If
   319  		//     MOVD SP, panic_argp(R0)
   320  		//   End
   321  		// End
   322  
   323  		gpanic := obj.Addr{
   324  			Type:   obj.TYPE_MEM,
   325  			Reg:    REGG,
   326  			Offset: 4 * 8, // g_panic
   327  		}
   328  
   329  		panicargp := obj.Addr{
   330  			Type:   obj.TYPE_MEM,
   331  			Reg:    REG_R0,
   332  			Offset: 0, // panic.argp
   333  		}
   334  
   335  		p := s.Func().Text
   336  		p = appendp(p, AMOVD, gpanic, regAddr(REG_R0))
   337  
   338  		p = appendp(p, AGet, regAddr(REG_R0))
   339  		p = appendp(p, AI64Eqz)
   340  		p = appendp(p, ANot)
   341  		p = appendp(p, AIf)
   342  
   343  		p = appendp(p, AGet, regAddr(REG_SP))
   344  		p = appendp(p, AI64ExtendI32U)
   345  		p = appendp(p, AI64Const, constAddr(framesize+8))
   346  		p = appendp(p, AI64Add)
   347  		p = appendp(p, AI64Load, panicargp)
   348  
   349  		p = appendp(p, AI64Eq)
   350  		p = appendp(p, AIf)
   351  		p = appendp(p, AMOVD, regAddr(REG_SP), panicargp)
   352  		p = appendp(p, AEnd)
   353  
   354  		p = appendp(p, AEnd)
   355  	}
   356  
   357  	if framesize > 0 {
   358  		p := s.Func().Text
   359  		p = appendp(p, AGet, regAddr(REG_SP))
   360  		p = appendp(p, AI32Const, constAddr(framesize))
   361  		p = appendp(p, AI32Sub)
   362  		p = appendp(p, ASet, regAddr(REG_SP))
   363  		p.Spadj = int32(framesize)
   364  	}
   365  
   366  	// If the framesize is 0, then imply nosplit because it's a specially
   367  	// generated function.
   368  	needMoreStack := framesize > 0 && !s.Func().Text.From.Sym.NoSplit()
   369  
   370  	// If the maymorestack debug option is enabled, insert the
   371  	// call to maymorestack *before* processing resume points so
   372  	// we can construct a resume point after maymorestack for
   373  	// morestack to resume at.
   374  	var pMorestack = s.Func().Text
   375  	if needMoreStack && ctxt.Flag_maymorestack != "" {
   376  		p := pMorestack
   377  
   378  		// Save REGCTXT on the stack.
   379  		const tempFrame = 8
   380  		p = appendp(p, AGet, regAddr(REG_SP))
   381  		p = appendp(p, AI32Const, constAddr(tempFrame))
   382  		p = appendp(p, AI32Sub)
   383  		p = appendp(p, ASet, regAddr(REG_SP))
   384  		p.Spadj = tempFrame
   385  		ctxtp := obj.Addr{
   386  			Type:   obj.TYPE_MEM,
   387  			Reg:    REG_SP,
   388  			Offset: 0,
   389  		}
   390  		p = appendp(p, AMOVD, regAddr(REGCTXT), ctxtp)
   391  
   392  		// maymorestack must not itself preempt because we
   393  		// don't have full stack information, so this can be
   394  		// ACALLNORESUME.
   395  		p = appendp(p, ACALLNORESUME, constAddr(0))
   396  		// See ../x86/obj6.go
   397  		sym := ctxt.LookupABI(ctxt.Flag_maymorestack, s.ABI())
   398  		p.To = obj.Addr{Type: obj.TYPE_MEM, Name: obj.NAME_EXTERN, Sym: sym}
   399  
   400  		// Restore REGCTXT.
   401  		p = appendp(p, AMOVD, ctxtp, regAddr(REGCTXT))
   402  		p = appendp(p, AGet, regAddr(REG_SP))
   403  		p = appendp(p, AI32Const, constAddr(tempFrame))
   404  		p = appendp(p, AI32Add)
   405  		p = appendp(p, ASet, regAddr(REG_SP))
   406  		p.Spadj = -tempFrame
   407  
   408  		// Add an explicit ARESUMEPOINT after maymorestack for
   409  		// morestack to resume at.
   410  		pMorestack = appendp(p, ARESUMEPOINT)
   411  	}
   412  
   413  	// Introduce resume points for CALL instructions
   414  	// and collect other explicit resume points.
   415  	numResumePoints := 0
   416  	explicitBlockDepth := 0
   417  	pc := int64(0) // pc is only incremented when necessary, this avoids bloat of the BrTable instruction
   418  	var tableIdxs []uint64
   419  	tablePC := int64(0)
   420  	base := ctxt.PosTable.Pos(s.Func().Text.Pos).Base()
   421  	for p := s.Func().Text; p != nil; p = p.Link {
   422  		prevBase := base
   423  		base = ctxt.PosTable.Pos(p.Pos).Base()
   424  		switch p.As {
   425  		case ABlock, ALoop, AIf:
   426  			explicitBlockDepth++
   427  
   428  		case AEnd:
   429  			if explicitBlockDepth == 0 {
   430  				panic("End without block")
   431  			}
   432  			explicitBlockDepth--
   433  
   434  		case ARESUMEPOINT:
   435  			if explicitBlockDepth != 0 {
   436  				panic("RESUME can only be used on toplevel")
   437  			}
   438  			p.As = AEnd
   439  			for tablePC <= pc {
   440  				tableIdxs = append(tableIdxs, uint64(numResumePoints))
   441  				tablePC++
   442  			}
   443  			numResumePoints++
   444  			pc++
   445  
   446  		case obj.ACALL:
   447  			if explicitBlockDepth != 0 {
   448  				panic("CALL can only be used on toplevel, try CALLNORESUME instead")
   449  			}
   450  			appendp(p, ARESUMEPOINT)
   451  		}
   452  
   453  		p.Pc = pc
   454  
   455  		// Increase pc whenever some pc-value table needs a new entry. Don't increase it
   456  		// more often to avoid bloat of the BrTable instruction.
   457  		// The "base != prevBase" condition detects inlined instructions. They are an
   458  		// implicit call, so entering and leaving this section affects the stack trace.
   459  		if p.As == ACALLNORESUME || p.As == obj.ANOP || p.As == ANop || p.Spadj != 0 || base != prevBase {
   460  			pc++
   461  			if p.To.Sym == sigpanic {
   462  				// The panic stack trace expects the PC at the call of sigpanic,
   463  				// not the next one. However, runtime.Caller subtracts 1 from the
   464  				// PC. To make both PC and PC-1 work (have the same line number),
   465  				// we advance the PC by 2 at sigpanic.
   466  				pc++
   467  			}
   468  		}
   469  	}
   470  	tableIdxs = append(tableIdxs, uint64(numResumePoints))
   471  	s.Size = pc + 1
   472  
   473  	if needMoreStack {
   474  		p := pMorestack
   475  
   476  		if framesize <= abi.StackSmall {
   477  			// small stack: SP <= stackguard
   478  			// Get SP
   479  			// Get g
   480  			// I32WrapI64
   481  			// I32Load $stackguard0
   482  			// I32GtU
   483  
   484  			p = appendp(p, AGet, regAddr(REG_SP))
   485  			p = appendp(p, AGet, regAddr(REGG))
   486  			p = appendp(p, AI32WrapI64)
   487  			p = appendp(p, AI32Load, constAddr(2*int64(ctxt.Arch.PtrSize))) // G.stackguard0
   488  			p = appendp(p, AI32LeU)
   489  		} else {
   490  			// large stack: SP-framesize <= stackguard-StackSmall
   491  			//              SP <= stackguard+(framesize-StackSmall)
   492  			// Get SP
   493  			// Get g
   494  			// I32WrapI64
   495  			// I32Load $stackguard0
   496  			// I32Const $(framesize-StackSmall)
   497  			// I32Add
   498  			// I32GtU
   499  
   500  			p = appendp(p, AGet, regAddr(REG_SP))
   501  			p = appendp(p, AGet, regAddr(REGG))
   502  			p = appendp(p, AI32WrapI64)
   503  			p = appendp(p, AI32Load, constAddr(2*int64(ctxt.Arch.PtrSize))) // G.stackguard0
   504  			p = appendp(p, AI32Const, constAddr(framesize-abi.StackSmall))
   505  			p = appendp(p, AI32Add)
   506  			p = appendp(p, AI32LeU)
   507  		}
   508  		// TODO(neelance): handle wraparound case
   509  
   510  		p = appendp(p, AIf)
   511  		// This CALL does *not* have a resume point after it
   512  		// (we already inserted all of the resume points). As
   513  		// a result, morestack will resume at the *previous*
   514  		// resume point (typically, the beginning of the
   515  		// function) and perform the morestack check again.
   516  		// This is why we don't need an explicit loop like
   517  		// other architectures.
   518  		p = appendp(p, obj.ACALL, constAddr(0))
   519  		if s.Func().Text.From.Sym.NeedCtxt() {
   520  			p.To = obj.Addr{Type: obj.TYPE_MEM, Name: obj.NAME_EXTERN, Sym: morestack}
   521  		} else {
   522  			p.To = obj.Addr{Type: obj.TYPE_MEM, Name: obj.NAME_EXTERN, Sym: morestackNoCtxt}
   523  		}
   524  		p = appendp(p, AEnd)
   525  	}
   526  
   527  	// record the branches targeting the entry loop and the unwind exit,
   528  	// their targets with be filled in later
   529  	var entryPointLoopBranches []*obj.Prog
   530  	var unwindExitBranches []*obj.Prog
   531  	currentDepth := 0
   532  	for p := s.Func().Text; p != nil; p = p.Link {
   533  		switch p.As {
   534  		case ABlock, ALoop, AIf:
   535  			currentDepth++
   536  		case AEnd:
   537  			currentDepth--
   538  		}
   539  
   540  		switch p.As {
   541  		case obj.AJMP:
   542  			jmp := *p
   543  			p.As = obj.ANOP
   544  
   545  			if jmp.To.Type == obj.TYPE_BRANCH {
   546  				// jump to basic block
   547  				p = appendp(p, AI32Const, constAddr(jmp.To.Val.(*obj.Prog).Pc))
   548  				p = appendp(p, ASet, regAddr(REG_PC_B)) // write next basic block to PC_B
   549  				p = appendp(p, ABr)                     // jump to beginning of entryPointLoop
   550  				entryPointLoopBranches = append(entryPointLoopBranches, p)
   551  				break
   552  			}
   553  
   554  			// low-level WebAssembly call to function
   555  			switch jmp.To.Type {
   556  			case obj.TYPE_MEM:
   557  				if !notUsePC_B[jmp.To.Sym.Name] {
   558  					// Set PC_B parameter to function entry.
   559  					p = appendp(p, AI32Const, constAddr(0))
   560  				}
   561  				p = appendp(p, ACall, jmp.To)
   562  
   563  			case obj.TYPE_NONE:
   564  				// (target PC is on stack)
   565  				p = appendp(p, AI32WrapI64)
   566  				p = appendp(p, AI32Const, constAddr(16)) // only needs PC_F bits (16-31), PC_B bits (0-15) are zero
   567  				p = appendp(p, AI32ShrU)
   568  
   569  				// Set PC_B parameter to function entry.
   570  				// We need to push this before pushing the target PC_F,
   571  				// so temporarily pop PC_F, using our REG_PC_B as a
   572  				// scratch register, and push it back after pushing 0.
   573  				p = appendp(p, ASet, regAddr(REG_PC_B))
   574  				p = appendp(p, AI32Const, constAddr(0))
   575  				p = appendp(p, AGet, regAddr(REG_PC_B))
   576  
   577  				p = appendp(p, ACallIndirect)
   578  
   579  			default:
   580  				panic("bad target for JMP")
   581  			}
   582  
   583  			p = appendp(p, AReturn)
   584  
   585  		case obj.ACALL, ACALLNORESUME:
   586  			call := *p
   587  			p.As = obj.ANOP
   588  
   589  			pcAfterCall := call.Link.Pc
   590  			if call.To.Sym == sigpanic {
   591  				pcAfterCall-- // sigpanic expects to be called without advancing the pc
   592  			}
   593  
   594  			// SP -= 8
   595  			p = appendp(p, AGet, regAddr(REG_SP))
   596  			p = appendp(p, AI32Const, constAddr(8))
   597  			p = appendp(p, AI32Sub)
   598  			p = appendp(p, ASet, regAddr(REG_SP))
   599  
   600  			// write return address to Go stack
   601  			p = appendp(p, AGet, regAddr(REG_SP))
   602  			p = appendp(p, AI64Const, obj.Addr{
   603  				Type:   obj.TYPE_ADDR,
   604  				Name:   obj.NAME_EXTERN,
   605  				Sym:    s,           // PC_F
   606  				Offset: pcAfterCall, // PC_B
   607  			})
   608  			p = appendp(p, AI64Store, constAddr(0))
   609  
   610  			// low-level WebAssembly call to function
   611  			switch call.To.Type {
   612  			case obj.TYPE_MEM:
   613  				if !notUsePC_B[call.To.Sym.Name] {
   614  					// Set PC_B parameter to function entry.
   615  					p = appendp(p, AI32Const, constAddr(0))
   616  				}
   617  				p = appendp(p, ACall, call.To)
   618  
   619  			case obj.TYPE_NONE:
   620  				// (target PC is on stack)
   621  				p = appendp(p, AI32WrapI64)
   622  				p = appendp(p, AI32Const, constAddr(16)) // only needs PC_F bits (16-31), PC_B bits (0-15) are zero
   623  				p = appendp(p, AI32ShrU)
   624  
   625  				// Set PC_B parameter to function entry.
   626  				// We need to push this before pushing the target PC_F,
   627  				// so temporarily pop PC_F, using our PC_B as a
   628  				// scratch register, and push it back after pushing 0.
   629  				p = appendp(p, ASet, regAddr(REG_PC_B))
   630  				p = appendp(p, AI32Const, constAddr(0))
   631  				p = appendp(p, AGet, regAddr(REG_PC_B))
   632  
   633  				p = appendp(p, ACallIndirect)
   634  
   635  			default:
   636  				panic("bad target for CALL")
   637  			}
   638  
   639  			// return value of call is on the top of the stack, indicating whether to unwind the WebAssembly stack
   640  			if call.As == ACALLNORESUME && call.To.Sym != sigpanic { // sigpanic unwinds the stack, but it never resumes
   641  				// trying to unwind WebAssembly stack but call has no resume point, terminate with error
   642  				p = appendp(p, AIf)
   643  				p = appendp(p, obj.AUNDEF)
   644  				p = appendp(p, AEnd)
   645  			} else {
   646  				// unwinding WebAssembly stack to switch goroutine, return 1
   647  				p = appendp(p, ABrIf)
   648  				unwindExitBranches = append(unwindExitBranches, p)
   649  			}
   650  
   651  		case obj.ARET, ARETUNWIND:
   652  			ret := *p
   653  			p.As = obj.ANOP
   654  
   655  			if framesize > 0 {
   656  				// SP += framesize
   657  				p = appendp(p, AGet, regAddr(REG_SP))
   658  				p = appendp(p, AI32Const, constAddr(framesize))
   659  				p = appendp(p, AI32Add)
   660  				p = appendp(p, ASet, regAddr(REG_SP))
   661  				// TODO(neelance): This should theoretically set Spadj, but it only works without.
   662  				// p.Spadj = int32(-framesize)
   663  			}
   664  
   665  			if ret.To.Type == obj.TYPE_MEM {
   666  				// Set PC_B parameter to function entry.
   667  				p = appendp(p, AI32Const, constAddr(0))
   668  
   669  				// low-level WebAssembly call to function
   670  				p = appendp(p, ACall, ret.To)
   671  				p = appendp(p, AReturn)
   672  				break
   673  			}
   674  
   675  			// SP += 8
   676  			p = appendp(p, AGet, regAddr(REG_SP))
   677  			p = appendp(p, AI32Const, constAddr(8))
   678  			p = appendp(p, AI32Add)
   679  			p = appendp(p, ASet, regAddr(REG_SP))
   680  
   681  			if ret.As == ARETUNWIND {
   682  				// function needs to unwind the WebAssembly stack, return 1
   683  				p = appendp(p, AI32Const, constAddr(1))
   684  				p = appendp(p, AReturn)
   685  				break
   686  			}
   687  
   688  			// not unwinding the WebAssembly stack, return 0
   689  			p = appendp(p, AI32Const, constAddr(0))
   690  			p = appendp(p, AReturn)
   691  		}
   692  	}
   693  
   694  	for p := s.Func().Text; p != nil; p = p.Link {
   695  		switch p.From.Name {
   696  		case obj.NAME_AUTO:
   697  			p.From.Offset += framesize
   698  		case obj.NAME_PARAM:
   699  			p.From.Reg = REG_SP
   700  			p.From.Offset += framesize + 8 // parameters are after the frame and the 8-byte return address
   701  		}
   702  
   703  		switch p.To.Name {
   704  		case obj.NAME_AUTO:
   705  			p.To.Offset += framesize
   706  		case obj.NAME_PARAM:
   707  			p.To.Reg = REG_SP
   708  			p.To.Offset += framesize + 8 // parameters are after the frame and the 8-byte return address
   709  		}
   710  
   711  		switch p.As {
   712  		case AGet:
   713  			if p.From.Type == obj.TYPE_ADDR {
   714  				get := *p
   715  				p.As = obj.ANOP
   716  
   717  				switch get.From.Name {
   718  				case obj.NAME_EXTERN:
   719  					p = appendp(p, AI64Const, get.From)
   720  				case obj.NAME_AUTO, obj.NAME_PARAM:
   721  					p = appendp(p, AGet, regAddr(get.From.Reg))
   722  					if get.From.Reg == REG_SP {
   723  						p = appendp(p, AI64ExtendI32U)
   724  					}
   725  					if get.From.Offset != 0 {
   726  						p = appendp(p, AI64Const, constAddr(get.From.Offset))
   727  						p = appendp(p, AI64Add)
   728  					}
   729  				default:
   730  					panic("bad Get: invalid name")
   731  				}
   732  			}
   733  
   734  		case AI32Load, AI64Load, AF32Load, AF64Load, AI32Load8S, AI32Load8U, AI32Load16S, AI32Load16U, AI64Load8S, AI64Load8U, AI64Load16S, AI64Load16U, AI64Load32S, AI64Load32U:
   735  			if p.From.Type == obj.TYPE_MEM {
   736  				as := p.As
   737  				from := p.From
   738  
   739  				p.As = AGet
   740  				p.From = regAddr(from.Reg)
   741  
   742  				if from.Reg != REG_SP {
   743  					p = appendp(p, AI32WrapI64)
   744  				}
   745  
   746  				p = appendp(p, as, constAddr(from.Offset))
   747  			}
   748  
   749  		case AMOVB, AMOVH, AMOVW, AMOVD:
   750  			mov := *p
   751  			p.As = obj.ANOP
   752  
   753  			var loadAs obj.As
   754  			var storeAs obj.As
   755  			switch mov.As {
   756  			case AMOVB:
   757  				loadAs = AI64Load8U
   758  				storeAs = AI64Store8
   759  			case AMOVH:
   760  				loadAs = AI64Load16U
   761  				storeAs = AI64Store16
   762  			case AMOVW:
   763  				loadAs = AI64Load32U
   764  				storeAs = AI64Store32
   765  			case AMOVD:
   766  				loadAs = AI64Load
   767  				storeAs = AI64Store
   768  			}
   769  
   770  			appendValue := func() {
   771  				switch mov.From.Type {
   772  				case obj.TYPE_CONST:
   773  					p = appendp(p, AI64Const, constAddr(mov.From.Offset))
   774  
   775  				case obj.TYPE_ADDR:
   776  					switch mov.From.Name {
   777  					case obj.NAME_NONE, obj.NAME_PARAM, obj.NAME_AUTO:
   778  						p = appendp(p, AGet, regAddr(mov.From.Reg))
   779  						if mov.From.Reg == REG_SP {
   780  							p = appendp(p, AI64ExtendI32U)
   781  						}
   782  						p = appendp(p, AI64Const, constAddr(mov.From.Offset))
   783  						p = appendp(p, AI64Add)
   784  					case obj.NAME_EXTERN:
   785  						p = appendp(p, AI64Const, mov.From)
   786  					default:
   787  						panic("bad name for MOV")
   788  					}
   789  
   790  				case obj.TYPE_REG:
   791  					p = appendp(p, AGet, mov.From)
   792  					if mov.From.Reg == REG_SP {
   793  						p = appendp(p, AI64ExtendI32U)
   794  					}
   795  
   796  				case obj.TYPE_MEM:
   797  					p = appendp(p, AGet, regAddr(mov.From.Reg))
   798  					if mov.From.Reg != REG_SP {
   799  						p = appendp(p, AI32WrapI64)
   800  					}
   801  					p = appendp(p, loadAs, constAddr(mov.From.Offset))
   802  
   803  				default:
   804  					panic("bad MOV type")
   805  				}
   806  			}
   807  
   808  			switch mov.To.Type {
   809  			case obj.TYPE_REG:
   810  				appendValue()
   811  				if mov.To.Reg == REG_SP {
   812  					p = appendp(p, AI32WrapI64)
   813  				}
   814  				p = appendp(p, ASet, mov.To)
   815  
   816  			case obj.TYPE_MEM:
   817  				switch mov.To.Name {
   818  				case obj.NAME_NONE, obj.NAME_PARAM:
   819  					p = appendp(p, AGet, regAddr(mov.To.Reg))
   820  					if mov.To.Reg != REG_SP {
   821  						p = appendp(p, AI32WrapI64)
   822  					}
   823  				case obj.NAME_EXTERN:
   824  					p = appendp(p, AI32Const, obj.Addr{Type: obj.TYPE_ADDR, Name: obj.NAME_EXTERN, Sym: mov.To.Sym})
   825  				default:
   826  					panic("bad MOV name")
   827  				}
   828  				appendValue()
   829  				p = appendp(p, storeAs, constAddr(mov.To.Offset))
   830  
   831  			default:
   832  				panic("bad MOV type")
   833  			}
   834  		}
   835  	}
   836  
   837  	{
   838  		p := s.Func().Text
   839  		if len(unwindExitBranches) > 0 {
   840  			p = appendp(p, ABlock) // unwindExit, used to return 1 when unwinding the stack
   841  			for _, b := range unwindExitBranches {
   842  				b.To = obj.Addr{Type: obj.TYPE_BRANCH, Val: p}
   843  			}
   844  		}
   845  		if len(entryPointLoopBranches) > 0 {
   846  			p = appendp(p, ALoop) // entryPointLoop, used to jump between basic blocks
   847  			for _, b := range entryPointLoopBranches {
   848  				b.To = obj.Addr{Type: obj.TYPE_BRANCH, Val: p}
   849  			}
   850  		}
   851  		if numResumePoints > 0 {
   852  			// Add Block instructions for resume points and BrTable to jump to selected resume point.
   853  			for i := 0; i < numResumePoints+1; i++ {
   854  				p = appendp(p, ABlock)
   855  			}
   856  			p = appendp(p, AGet, regAddr(REG_PC_B)) // read next basic block from PC_B
   857  			p = appendp(p, ABrTable, obj.Addr{Val: tableIdxs})
   858  			p = appendp(p, AEnd) // end of Block
   859  		}
   860  		for p.Link != nil {
   861  			p = p.Link // function instructions
   862  		}
   863  		if len(entryPointLoopBranches) > 0 {
   864  			p = appendp(p, AEnd) // end of entryPointLoop
   865  		}
   866  		p = appendp(p, obj.AUNDEF)
   867  		if len(unwindExitBranches) > 0 {
   868  			p = appendp(p, AEnd) // end of unwindExit
   869  			p = appendp(p, AI32Const, constAddr(1))
   870  		}
   871  	}
   872  
   873  	currentDepth = 0
   874  	blockDepths := make(map[*obj.Prog]int)
   875  	for p := s.Func().Text; p != nil; p = p.Link {
   876  		switch p.As {
   877  		case ABlock, ALoop, AIf:
   878  			currentDepth++
   879  			blockDepths[p] = currentDepth
   880  		case AEnd:
   881  			currentDepth--
   882  		}
   883  
   884  		switch p.As {
   885  		case ABr, ABrIf:
   886  			if p.To.Type == obj.TYPE_BRANCH {
   887  				blockDepth, ok := blockDepths[p.To.Val.(*obj.Prog)]
   888  				if !ok {
   889  					panic("label not at block")
   890  				}
   891  				p.To = constAddr(int64(currentDepth - blockDepth))
   892  			}
   893  		}
   894  	}
   895  }
   896  
   897  func constAddr(value int64) obj.Addr {
   898  	return obj.Addr{Type: obj.TYPE_CONST, Offset: value}
   899  }
   900  
   901  func regAddr(reg int16) obj.Addr {
   902  	return obj.Addr{Type: obj.TYPE_REG, Reg: reg}
   903  }
   904  
   905  // Most of the Go functions has a single parameter (PC_B) in
   906  // Wasm ABI. This is a list of exceptions.
   907  var notUsePC_B = map[string]bool{
   908  	"_rt0_wasm_js":            true,
   909  	"_rt0_wasm_wasip1":        true,
   910  	"wasm_export_run":         true,
   911  	"wasm_export_resume":      true,
   912  	"wasm_export_getsp":       true,
   913  	"wasm_pc_f_loop":          true,
   914  	"gcWriteBarrier":          true,
   915  	"runtime.gcWriteBarrier1": true,
   916  	"runtime.gcWriteBarrier2": true,
   917  	"runtime.gcWriteBarrier3": true,
   918  	"runtime.gcWriteBarrier4": true,
   919  	"runtime.gcWriteBarrier5": true,
   920  	"runtime.gcWriteBarrier6": true,
   921  	"runtime.gcWriteBarrier7": true,
   922  	"runtime.gcWriteBarrier8": true,
   923  	"runtime.wasmDiv":         true,
   924  	"runtime.wasmTruncS":      true,
   925  	"runtime.wasmTruncU":      true,
   926  	"cmpbody":                 true,
   927  	"memeqbody":               true,
   928  	"memcmp":                  true,
   929  	"memchr":                  true,
   930  }
   931  
   932  func assemble(ctxt *obj.Link, s *obj.LSym, newprog obj.ProgAlloc) {
   933  	type regVar struct {
   934  		global bool
   935  		index  uint64
   936  	}
   937  
   938  	type varDecl struct {
   939  		count uint64
   940  		typ   valueType
   941  	}
   942  
   943  	hasLocalSP := false
   944  	regVars := [MAXREG - MINREG]*regVar{
   945  		REG_SP - MINREG:    {true, 0},
   946  		REG_CTXT - MINREG:  {true, 1},
   947  		REG_g - MINREG:     {true, 2},
   948  		REG_RET0 - MINREG:  {true, 3},
   949  		REG_RET1 - MINREG:  {true, 4},
   950  		REG_RET2 - MINREG:  {true, 5},
   951  		REG_RET3 - MINREG:  {true, 6},
   952  		REG_PAUSE - MINREG: {true, 7},
   953  	}
   954  	var varDecls []*varDecl
   955  	useAssemblyRegMap := func() {
   956  		for i := int16(0); i < 16; i++ {
   957  			regVars[REG_R0+i-MINREG] = &regVar{false, uint64(i)}
   958  		}
   959  	}
   960  
   961  	// Function starts with declaration of locals: numbers and types.
   962  	// Some functions use a special calling convention.
   963  	switch s.Name {
   964  	case "_rt0_wasm_js", "_rt0_wasm_wasip1", "wasm_export_run", "wasm_export_resume", "wasm_export_getsp",
   965  		"wasm_pc_f_loop", "runtime.wasmDiv", "runtime.wasmTruncS", "runtime.wasmTruncU", "memeqbody":
   966  		varDecls = []*varDecl{}
   967  		useAssemblyRegMap()
   968  	case "memchr", "memcmp":
   969  		varDecls = []*varDecl{{count: 2, typ: i32}}
   970  		useAssemblyRegMap()
   971  	case "cmpbody":
   972  		varDecls = []*varDecl{{count: 2, typ: i64}}
   973  		useAssemblyRegMap()
   974  	case "gcWriteBarrier":
   975  		varDecls = []*varDecl{{count: 5, typ: i64}}
   976  		useAssemblyRegMap()
   977  	case "runtime.gcWriteBarrier1",
   978  		"runtime.gcWriteBarrier2",
   979  		"runtime.gcWriteBarrier3",
   980  		"runtime.gcWriteBarrier4",
   981  		"runtime.gcWriteBarrier5",
   982  		"runtime.gcWriteBarrier6",
   983  		"runtime.gcWriteBarrier7",
   984  		"runtime.gcWriteBarrier8":
   985  		// no locals
   986  		useAssemblyRegMap()
   987  	default:
   988  		// Normal calling convention: PC_B as WebAssembly parameter. First local variable is local SP cache.
   989  		regVars[REG_PC_B-MINREG] = &regVar{false, 0}
   990  		hasLocalSP = true
   991  
   992  		var regUsed [MAXREG - MINREG]bool
   993  		for p := s.Func().Text; p != nil; p = p.Link {
   994  			if p.From.Reg != 0 {
   995  				regUsed[p.From.Reg-MINREG] = true
   996  			}
   997  			if p.To.Reg != 0 {
   998  				regUsed[p.To.Reg-MINREG] = true
   999  			}
  1000  		}
  1001  
  1002  		regs := []int16{REG_SP}
  1003  		for reg := int16(REG_R0); reg <= REG_F31; reg++ {
  1004  			if regUsed[reg-MINREG] {
  1005  				regs = append(regs, reg)
  1006  			}
  1007  		}
  1008  
  1009  		var lastDecl *varDecl
  1010  		for i, reg := range regs {
  1011  			t := regType(reg)
  1012  			if lastDecl == nil || lastDecl.typ != t {
  1013  				lastDecl = &varDecl{
  1014  					count: 0,
  1015  					typ:   t,
  1016  				}
  1017  				varDecls = append(varDecls, lastDecl)
  1018  			}
  1019  			lastDecl.count++
  1020  			if reg != REG_SP {
  1021  				regVars[reg-MINREG] = &regVar{false, 1 + uint64(i)}
  1022  			}
  1023  		}
  1024  	}
  1025  
  1026  	w := new(bytes.Buffer)
  1027  
  1028  	writeUleb128(w, uint64(len(varDecls)))
  1029  	for _, decl := range varDecls {
  1030  		writeUleb128(w, decl.count)
  1031  		w.WriteByte(byte(decl.typ))
  1032  	}
  1033  
  1034  	if hasLocalSP {
  1035  		// Copy SP from its global variable into a local variable. Accessing a local variable is more efficient.
  1036  		updateLocalSP(w)
  1037  	}
  1038  
  1039  	for p := s.Func().Text; p != nil; p = p.Link {
  1040  		switch p.As {
  1041  		case AGet:
  1042  			if p.From.Type != obj.TYPE_REG {
  1043  				panic("bad Get: argument is not a register")
  1044  			}
  1045  			reg := p.From.Reg
  1046  			v := regVars[reg-MINREG]
  1047  			if v == nil {
  1048  				panic("bad Get: invalid register")
  1049  			}
  1050  			if reg == REG_SP && hasLocalSP {
  1051  				writeOpcode(w, ALocalGet)
  1052  				writeUleb128(w, 1) // local SP
  1053  				continue
  1054  			}
  1055  			if v.global {
  1056  				writeOpcode(w, AGlobalGet)
  1057  			} else {
  1058  				writeOpcode(w, ALocalGet)
  1059  			}
  1060  			writeUleb128(w, v.index)
  1061  			continue
  1062  
  1063  		case ASet:
  1064  			if p.To.Type != obj.TYPE_REG {
  1065  				panic("bad Set: argument is not a register")
  1066  			}
  1067  			reg := p.To.Reg
  1068  			v := regVars[reg-MINREG]
  1069  			if v == nil {
  1070  				panic("bad Set: invalid register")
  1071  			}
  1072  			if reg == REG_SP && hasLocalSP {
  1073  				writeOpcode(w, ALocalTee)
  1074  				writeUleb128(w, 1) // local SP
  1075  			}
  1076  			if v.global {
  1077  				writeOpcode(w, AGlobalSet)
  1078  			} else {
  1079  				if p.Link.As == AGet && p.Link.From.Reg == reg {
  1080  					writeOpcode(w, ALocalTee)
  1081  					p = p.Link
  1082  				} else {
  1083  					writeOpcode(w, ALocalSet)
  1084  				}
  1085  			}
  1086  			writeUleb128(w, v.index)
  1087  			continue
  1088  
  1089  		case ATee:
  1090  			if p.To.Type != obj.TYPE_REG {
  1091  				panic("bad Tee: argument is not a register")
  1092  			}
  1093  			reg := p.To.Reg
  1094  			v := regVars[reg-MINREG]
  1095  			if v == nil {
  1096  				panic("bad Tee: invalid register")
  1097  			}
  1098  			writeOpcode(w, ALocalTee)
  1099  			writeUleb128(w, v.index)
  1100  			continue
  1101  
  1102  		case ANot:
  1103  			writeOpcode(w, AI32Eqz)
  1104  			continue
  1105  
  1106  		case obj.AUNDEF:
  1107  			writeOpcode(w, AUnreachable)
  1108  			continue
  1109  
  1110  		case obj.ANOP, obj.ATEXT, obj.AFUNCDATA, obj.APCDATA:
  1111  			// ignore
  1112  			continue
  1113  		}
  1114  
  1115  		writeOpcode(w, p.As)
  1116  
  1117  		switch p.As {
  1118  		case ABlock, ALoop, AIf:
  1119  			if p.From.Offset != 0 {
  1120  				// block type, rarely used, e.g. for code compiled with emscripten
  1121  				w.WriteByte(0x80 - byte(p.From.Offset))
  1122  				continue
  1123  			}
  1124  			w.WriteByte(0x40)
  1125  
  1126  		case ABr, ABrIf:
  1127  			if p.To.Type != obj.TYPE_CONST {
  1128  				panic("bad Br/BrIf")
  1129  			}
  1130  			writeUleb128(w, uint64(p.To.Offset))
  1131  
  1132  		case ABrTable:
  1133  			idxs := p.To.Val.([]uint64)
  1134  			writeUleb128(w, uint64(len(idxs)-1))
  1135  			for _, idx := range idxs {
  1136  				writeUleb128(w, idx)
  1137  			}
  1138  
  1139  		case ACall:
  1140  			switch p.To.Type {
  1141  			case obj.TYPE_CONST:
  1142  				writeUleb128(w, uint64(p.To.Offset))
  1143  
  1144  			case obj.TYPE_MEM:
  1145  				if p.To.Name != obj.NAME_EXTERN && p.To.Name != obj.NAME_STATIC {
  1146  					fmt.Println(p.To)
  1147  					panic("bad name for Call")
  1148  				}
  1149  				r := obj.Addrel(s)
  1150  				r.Siz = 1 // actually variable sized
  1151  				r.Off = int32(w.Len())
  1152  				r.Type = objabi.R_CALL
  1153  				if p.Mark&WasmImport != 0 {
  1154  					r.Type = objabi.R_WASMIMPORT
  1155  				}
  1156  				r.Sym = p.To.Sym
  1157  				if hasLocalSP {
  1158  					// The stack may have moved, which changes SP. Update the local SP variable.
  1159  					updateLocalSP(w)
  1160  				}
  1161  
  1162  			default:
  1163  				panic("bad type for Call")
  1164  			}
  1165  
  1166  		case ACallIndirect:
  1167  			writeUleb128(w, uint64(p.To.Offset))
  1168  			w.WriteByte(0x00) // reserved value
  1169  			if hasLocalSP {
  1170  				// The stack may have moved, which changes SP. Update the local SP variable.
  1171  				updateLocalSP(w)
  1172  			}
  1173  
  1174  		case AI32Const, AI64Const:
  1175  			if p.From.Name == obj.NAME_EXTERN {
  1176  				r := obj.Addrel(s)
  1177  				r.Siz = 1 // actually variable sized
  1178  				r.Off = int32(w.Len())
  1179  				r.Type = objabi.R_ADDR
  1180  				r.Sym = p.From.Sym
  1181  				r.Add = p.From.Offset
  1182  				break
  1183  			}
  1184  			writeSleb128(w, p.From.Offset)
  1185  
  1186  		case AF32Const:
  1187  			b := make([]byte, 4)
  1188  			binary.LittleEndian.PutUint32(b, math.Float32bits(float32(p.From.Val.(float64))))
  1189  			w.Write(b)
  1190  
  1191  		case AF64Const:
  1192  			b := make([]byte, 8)
  1193  			binary.LittleEndian.PutUint64(b, math.Float64bits(p.From.Val.(float64)))
  1194  			w.Write(b)
  1195  
  1196  		case AI32Load, AI64Load, AF32Load, AF64Load, AI32Load8S, AI32Load8U, AI32Load16S, AI32Load16U, AI64Load8S, AI64Load8U, AI64Load16S, AI64Load16U, AI64Load32S, AI64Load32U:
  1197  			if p.From.Offset < 0 {
  1198  				panic("negative offset for *Load")
  1199  			}
  1200  			if p.From.Type != obj.TYPE_CONST {
  1201  				panic("bad type for *Load")
  1202  			}
  1203  			if p.From.Offset > math.MaxUint32 {
  1204  				ctxt.Diag("bad offset in %v", p)
  1205  			}
  1206  			writeUleb128(w, align(p.As))
  1207  			writeUleb128(w, uint64(p.From.Offset))
  1208  
  1209  		case AI32Store, AI64Store, AF32Store, AF64Store, AI32Store8, AI32Store16, AI64Store8, AI64Store16, AI64Store32:
  1210  			if p.To.Offset < 0 {
  1211  				panic("negative offset")
  1212  			}
  1213  			if p.From.Offset > math.MaxUint32 {
  1214  				ctxt.Diag("bad offset in %v", p)
  1215  			}
  1216  			writeUleb128(w, align(p.As))
  1217  			writeUleb128(w, uint64(p.To.Offset))
  1218  
  1219  		case ACurrentMemory, AGrowMemory, AMemoryFill:
  1220  			w.WriteByte(0x00)
  1221  
  1222  		case AMemoryCopy:
  1223  			w.WriteByte(0x00)
  1224  			w.WriteByte(0x00)
  1225  
  1226  		}
  1227  	}
  1228  
  1229  	w.WriteByte(0x0b) // end
  1230  
  1231  	s.P = w.Bytes()
  1232  }
  1233  
  1234  func updateLocalSP(w *bytes.Buffer) {
  1235  	writeOpcode(w, AGlobalGet)
  1236  	writeUleb128(w, 0) // global SP
  1237  	writeOpcode(w, ALocalSet)
  1238  	writeUleb128(w, 1) // local SP
  1239  }
  1240  
  1241  func writeOpcode(w *bytes.Buffer, as obj.As) {
  1242  	switch {
  1243  	case as < AUnreachable:
  1244  		panic(fmt.Sprintf("unexpected assembler op: %s", as))
  1245  	case as < AEnd:
  1246  		w.WriteByte(byte(as - AUnreachable + 0x00))
  1247  	case as < ADrop:
  1248  		w.WriteByte(byte(as - AEnd + 0x0B))
  1249  	case as < ALocalGet:
  1250  		w.WriteByte(byte(as - ADrop + 0x1A))
  1251  	case as < AI32Load:
  1252  		w.WriteByte(byte(as - ALocalGet + 0x20))
  1253  	case as < AI32TruncSatF32S:
  1254  		w.WriteByte(byte(as - AI32Load + 0x28))
  1255  	case as < ALast:
  1256  		w.WriteByte(0xFC)
  1257  		w.WriteByte(byte(as - AI32TruncSatF32S + 0x00))
  1258  	default:
  1259  		panic(fmt.Sprintf("unexpected assembler op: %s", as))
  1260  	}
  1261  }
  1262  
  1263  type valueType byte
  1264  
  1265  const (
  1266  	i32 valueType = 0x7F
  1267  	i64 valueType = 0x7E
  1268  	f32 valueType = 0x7D
  1269  	f64 valueType = 0x7C
  1270  )
  1271  
  1272  func regType(reg int16) valueType {
  1273  	switch {
  1274  	case reg == REG_SP:
  1275  		return i32
  1276  	case reg >= REG_R0 && reg <= REG_R15:
  1277  		return i64
  1278  	case reg >= REG_F0 && reg <= REG_F15:
  1279  		return f32
  1280  	case reg >= REG_F16 && reg <= REG_F31:
  1281  		return f64
  1282  	default:
  1283  		panic("invalid register")
  1284  	}
  1285  }
  1286  
  1287  func align(as obj.As) uint64 {
  1288  	switch as {
  1289  	case AI32Load8S, AI32Load8U, AI64Load8S, AI64Load8U, AI32Store8, AI64Store8:
  1290  		return 0
  1291  	case AI32Load16S, AI32Load16U, AI64Load16S, AI64Load16U, AI32Store16, AI64Store16:
  1292  		return 1
  1293  	case AI32Load, AF32Load, AI64Load32S, AI64Load32U, AI32Store, AF32Store, AI64Store32:
  1294  		return 2
  1295  	case AI64Load, AF64Load, AI64Store, AF64Store:
  1296  		return 3
  1297  	default:
  1298  		panic("align: bad op")
  1299  	}
  1300  }
  1301  
  1302  func writeUleb128(w io.ByteWriter, v uint64) {
  1303  	if v < 128 {
  1304  		w.WriteByte(uint8(v))
  1305  		return
  1306  	}
  1307  	more := true
  1308  	for more {
  1309  		c := uint8(v & 0x7f)
  1310  		v >>= 7
  1311  		more = v != 0
  1312  		if more {
  1313  			c |= 0x80
  1314  		}
  1315  		w.WriteByte(c)
  1316  	}
  1317  }
  1318  
  1319  func writeSleb128(w io.ByteWriter, v int64) {
  1320  	more := true
  1321  	for more {
  1322  		c := uint8(v & 0x7f)
  1323  		s := uint8(v & 0x40)
  1324  		v >>= 7
  1325  		more = !((v == 0 && s == 0) || (v == -1 && s != 0))
  1326  		if more {
  1327  			c |= 0x80
  1328  		}
  1329  		w.WriteByte(c)
  1330  	}
  1331  }
  1332  

View as plain text