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  	wasm_pc_f_loop_export *obj.LSym
   132  )
   133  
   134  const (
   135  	/* mark flags */
   136  	WasmImport = 1 << 0
   137  )
   138  
   139  const (
   140  	// This is a special wasm module name that when used as the module name
   141  	// in //go:wasmimport will cause the generated code to pass the stack pointer
   142  	// directly to the imported function. In other words, any function that
   143  	// uses the gojs module understands the internal Go WASM ABI directly.
   144  	GojsModule = "gojs"
   145  )
   146  
   147  func instinit(ctxt *obj.Link) {
   148  	morestack = ctxt.Lookup("runtime.morestack")
   149  	morestackNoCtxt = ctxt.Lookup("runtime.morestack_noctxt")
   150  	sigpanic = ctxt.LookupABI("runtime.sigpanic", obj.ABIInternal)
   151  	wasm_pc_f_loop_export = ctxt.Lookup("wasm_pc_f_loop_export")
   152  }
   153  
   154  func preprocess(ctxt *obj.Link, s *obj.LSym, newprog obj.ProgAlloc) {
   155  	appendp := func(p *obj.Prog, as obj.As, args ...obj.Addr) *obj.Prog {
   156  		if p.As != obj.ANOP {
   157  			p2 := obj.Appendp(p, newprog)
   158  			p2.Pc = p.Pc
   159  			p = p2
   160  		}
   161  		p.As = as
   162  		switch len(args) {
   163  		case 0:
   164  			p.From = obj.Addr{}
   165  			p.To = obj.Addr{}
   166  		case 1:
   167  			if unaryDst[as] {
   168  				p.From = obj.Addr{}
   169  				p.To = args[0]
   170  			} else {
   171  				p.From = args[0]
   172  				p.To = obj.Addr{}
   173  			}
   174  		case 2:
   175  			p.From = args[0]
   176  			p.To = args[1]
   177  		default:
   178  			panic("bad args")
   179  		}
   180  		return p
   181  	}
   182  
   183  	framesize := s.Func().Text.To.Offset
   184  	if framesize < 0 {
   185  		panic("bad framesize")
   186  	}
   187  	s.Func().Args = s.Func().Text.To.Val.(int32)
   188  	s.Func().Locals = int32(framesize)
   189  
   190  	// If the function exits just to call out to a wasmimport, then
   191  	// generate the code to translate from our internal Go-stack
   192  	// based call convention to the native webassembly call convention.
   193  	if s.Func().WasmImport != nil {
   194  		genWasmImportWrapper(s, appendp)
   195  
   196  		// It should be 0 already, but we'll set it to 0 anyway just to be sure
   197  		// that the code below which adds frame expansion code to the function body
   198  		// isn't run. We don't want the frame expansion code because our function
   199  		// body is just the code to translate and call the imported function.
   200  		framesize = 0
   201  	} else if s.Func().WasmExport != nil {
   202  		genWasmExportWrapper(s, appendp)
   203  	} else if s.Func().Text.From.Sym.Wrapper() {
   204  		// if g._panic != nil && g._panic.argp == FP {
   205  		//   g._panic.argp = bottom-of-frame
   206  		// }
   207  		//
   208  		// MOVD g_panic(g), R0
   209  		// Get R0
   210  		// I64Eqz
   211  		// Not
   212  		// If
   213  		//   Get SP
   214  		//   I64ExtendI32U
   215  		//   I64Const $framesize+8
   216  		//   I64Add
   217  		//   I64Load panic_argp(R0)
   218  		//   I64Eq
   219  		//   If
   220  		//     MOVD SP, panic_argp(R0)
   221  		//   End
   222  		// End
   223  
   224  		gpanic := obj.Addr{
   225  			Type:   obj.TYPE_MEM,
   226  			Reg:    REGG,
   227  			Offset: 4 * 8, // g_panic
   228  		}
   229  
   230  		panicargp := obj.Addr{
   231  			Type:   obj.TYPE_MEM,
   232  			Reg:    REG_R0,
   233  			Offset: 0, // panic.argp
   234  		}
   235  
   236  		p := s.Func().Text
   237  		p = appendp(p, AMOVD, gpanic, regAddr(REG_R0))
   238  
   239  		p = appendp(p, AGet, regAddr(REG_R0))
   240  		p = appendp(p, AI64Eqz)
   241  		p = appendp(p, ANot)
   242  		p = appendp(p, AIf)
   243  
   244  		p = appendp(p, AGet, regAddr(REG_SP))
   245  		p = appendp(p, AI64ExtendI32U)
   246  		p = appendp(p, AI64Const, constAddr(framesize+8))
   247  		p = appendp(p, AI64Add)
   248  		p = appendp(p, AI64Load, panicargp)
   249  
   250  		p = appendp(p, AI64Eq)
   251  		p = appendp(p, AIf)
   252  		p = appendp(p, AMOVD, regAddr(REG_SP), panicargp)
   253  		p = appendp(p, AEnd)
   254  
   255  		p = appendp(p, AEnd)
   256  	}
   257  
   258  	if framesize > 0 {
   259  		p := s.Func().Text
   260  		p = appendp(p, AGet, regAddr(REG_SP))
   261  		p = appendp(p, AI32Const, constAddr(framesize))
   262  		p = appendp(p, AI32Sub)
   263  		p = appendp(p, ASet, regAddr(REG_SP))
   264  		p.Spadj = int32(framesize)
   265  	}
   266  
   267  	// If the framesize is 0, then imply nosplit because it's a specially
   268  	// generated function.
   269  	needMoreStack := framesize > 0 && !s.Func().Text.From.Sym.NoSplit()
   270  
   271  	// If the maymorestack debug option is enabled, insert the
   272  	// call to maymorestack *before* processing resume points so
   273  	// we can construct a resume point after maymorestack for
   274  	// morestack to resume at.
   275  	var pMorestack = s.Func().Text
   276  	if needMoreStack && ctxt.Flag_maymorestack != "" {
   277  		p := pMorestack
   278  
   279  		// Save REGCTXT on the stack.
   280  		const tempFrame = 8
   281  		p = appendp(p, AGet, regAddr(REG_SP))
   282  		p = appendp(p, AI32Const, constAddr(tempFrame))
   283  		p = appendp(p, AI32Sub)
   284  		p = appendp(p, ASet, regAddr(REG_SP))
   285  		p.Spadj = tempFrame
   286  		ctxtp := obj.Addr{
   287  			Type:   obj.TYPE_MEM,
   288  			Reg:    REG_SP,
   289  			Offset: 0,
   290  		}
   291  		p = appendp(p, AMOVD, regAddr(REGCTXT), ctxtp)
   292  
   293  		// maymorestack must not itself preempt because we
   294  		// don't have full stack information, so this can be
   295  		// ACALLNORESUME.
   296  		p = appendp(p, ACALLNORESUME, constAddr(0))
   297  		// See ../x86/obj6.go
   298  		sym := ctxt.LookupABI(ctxt.Flag_maymorestack, s.ABI())
   299  		p.To = obj.Addr{Type: obj.TYPE_MEM, Name: obj.NAME_EXTERN, Sym: sym}
   300  
   301  		// Restore REGCTXT.
   302  		p = appendp(p, AMOVD, ctxtp, regAddr(REGCTXT))
   303  		p = appendp(p, AGet, regAddr(REG_SP))
   304  		p = appendp(p, AI32Const, constAddr(tempFrame))
   305  		p = appendp(p, AI32Add)
   306  		p = appendp(p, ASet, regAddr(REG_SP))
   307  		p.Spadj = -tempFrame
   308  
   309  		// Add an explicit ARESUMEPOINT after maymorestack for
   310  		// morestack to resume at.
   311  		pMorestack = appendp(p, ARESUMEPOINT)
   312  	}
   313  
   314  	// Introduce resume points for CALL instructions
   315  	// and collect other explicit resume points.
   316  	numResumePoints := 0
   317  	explicitBlockDepth := 0
   318  	pc := int64(0) // pc is only incremented when necessary, this avoids bloat of the BrTable instruction
   319  	var tableIdxs []uint64
   320  	tablePC := int64(0)
   321  	base := ctxt.PosTable.Pos(s.Func().Text.Pos).Base()
   322  	for p := s.Func().Text; p != nil; p = p.Link {
   323  		prevBase := base
   324  		base = ctxt.PosTable.Pos(p.Pos).Base()
   325  		switch p.As {
   326  		case ABlock, ALoop, AIf:
   327  			explicitBlockDepth++
   328  
   329  		case AEnd:
   330  			if explicitBlockDepth == 0 {
   331  				panic("End without block")
   332  			}
   333  			explicitBlockDepth--
   334  
   335  		case ARESUMEPOINT:
   336  			if explicitBlockDepth != 0 {
   337  				panic("RESUME can only be used on toplevel")
   338  			}
   339  			p.As = AEnd
   340  			for tablePC <= pc {
   341  				tableIdxs = append(tableIdxs, uint64(numResumePoints))
   342  				tablePC++
   343  			}
   344  			numResumePoints++
   345  			pc++
   346  
   347  		case obj.ACALL:
   348  			if explicitBlockDepth != 0 {
   349  				panic("CALL can only be used on toplevel, try CALLNORESUME instead")
   350  			}
   351  			appendp(p, ARESUMEPOINT)
   352  		}
   353  
   354  		p.Pc = pc
   355  
   356  		// Increase pc whenever some pc-value table needs a new entry. Don't increase it
   357  		// more often to avoid bloat of the BrTable instruction.
   358  		// The "base != prevBase" condition detects inlined instructions. They are an
   359  		// implicit call, so entering and leaving this section affects the stack trace.
   360  		if p.As == ACALLNORESUME || p.As == obj.ANOP || p.As == ANop || p.Spadj != 0 || base != prevBase {
   361  			pc++
   362  			if p.To.Sym == sigpanic {
   363  				// The panic stack trace expects the PC at the call of sigpanic,
   364  				// not the next one. However, runtime.Caller subtracts 1 from the
   365  				// PC. To make both PC and PC-1 work (have the same line number),
   366  				// we advance the PC by 2 at sigpanic.
   367  				pc++
   368  			}
   369  		}
   370  	}
   371  	tableIdxs = append(tableIdxs, uint64(numResumePoints))
   372  	s.Size = pc + 1
   373  
   374  	if needMoreStack {
   375  		p := pMorestack
   376  
   377  		if framesize <= abi.StackSmall {
   378  			// small stack: SP <= stackguard
   379  			// Get SP
   380  			// Get g
   381  			// I32WrapI64
   382  			// I32Load $stackguard0
   383  			// I32GtU
   384  
   385  			p = appendp(p, AGet, regAddr(REG_SP))
   386  			p = appendp(p, AGet, regAddr(REGG))
   387  			p = appendp(p, AI32WrapI64)
   388  			p = appendp(p, AI32Load, constAddr(2*int64(ctxt.Arch.PtrSize))) // G.stackguard0
   389  			p = appendp(p, AI32LeU)
   390  		} else {
   391  			// large stack: SP-framesize <= stackguard-StackSmall
   392  			//              SP <= stackguard+(framesize-StackSmall)
   393  			// Get SP
   394  			// Get g
   395  			// I32WrapI64
   396  			// I32Load $stackguard0
   397  			// I32Const $(framesize-StackSmall)
   398  			// I32Add
   399  			// I32GtU
   400  
   401  			p = appendp(p, AGet, regAddr(REG_SP))
   402  			p = appendp(p, AGet, regAddr(REGG))
   403  			p = appendp(p, AI32WrapI64)
   404  			p = appendp(p, AI32Load, constAddr(2*int64(ctxt.Arch.PtrSize))) // G.stackguard0
   405  			p = appendp(p, AI32Const, constAddr(framesize-abi.StackSmall))
   406  			p = appendp(p, AI32Add)
   407  			p = appendp(p, AI32LeU)
   408  		}
   409  		// TODO(neelance): handle wraparound case
   410  
   411  		p = appendp(p, AIf)
   412  		// This CALL does *not* have a resume point after it
   413  		// (we already inserted all of the resume points). As
   414  		// a result, morestack will resume at the *previous*
   415  		// resume point (typically, the beginning of the
   416  		// function) and perform the morestack check again.
   417  		// This is why we don't need an explicit loop like
   418  		// other architectures.
   419  		p = appendp(p, obj.ACALL, constAddr(0))
   420  		if s.Func().Text.From.Sym.NeedCtxt() {
   421  			p.To = obj.Addr{Type: obj.TYPE_MEM, Name: obj.NAME_EXTERN, Sym: morestack}
   422  		} else {
   423  			p.To = obj.Addr{Type: obj.TYPE_MEM, Name: obj.NAME_EXTERN, Sym: morestackNoCtxt}
   424  		}
   425  		p = appendp(p, AEnd)
   426  	}
   427  
   428  	// record the branches targeting the entry loop and the unwind exit,
   429  	// their targets with be filled in later
   430  	var entryPointLoopBranches []*obj.Prog
   431  	var unwindExitBranches []*obj.Prog
   432  	currentDepth := 0
   433  	for p := s.Func().Text; p != nil; p = p.Link {
   434  		switch p.As {
   435  		case ABlock, ALoop, AIf:
   436  			currentDepth++
   437  		case AEnd:
   438  			currentDepth--
   439  		}
   440  
   441  		switch p.As {
   442  		case obj.AJMP:
   443  			jmp := *p
   444  			p.As = obj.ANOP
   445  
   446  			if jmp.To.Type == obj.TYPE_BRANCH {
   447  				// jump to basic block
   448  				p = appendp(p, AI32Const, constAddr(jmp.To.Val.(*obj.Prog).Pc))
   449  				p = appendp(p, ASet, regAddr(REG_PC_B)) // write next basic block to PC_B
   450  				p = appendp(p, ABr)                     // jump to beginning of entryPointLoop
   451  				entryPointLoopBranches = append(entryPointLoopBranches, p)
   452  				break
   453  			}
   454  
   455  			// low-level WebAssembly call to function
   456  			switch jmp.To.Type {
   457  			case obj.TYPE_MEM:
   458  				if !notUsePC_B[jmp.To.Sym.Name] {
   459  					// Set PC_B parameter to function entry.
   460  					p = appendp(p, AI32Const, constAddr(0))
   461  				}
   462  				p = appendp(p, ACall, jmp.To)
   463  
   464  			case obj.TYPE_NONE:
   465  				// (target PC is on stack)
   466  				p = appendp(p, AI32WrapI64)
   467  				p = appendp(p, AI32Const, constAddr(16)) // only needs PC_F bits (16-31), PC_B bits (0-15) are zero
   468  				p = appendp(p, AI32ShrU)
   469  
   470  				// Set PC_B parameter to function entry.
   471  				// We need to push this before pushing the target PC_F,
   472  				// so temporarily pop PC_F, using our REG_PC_B as a
   473  				// scratch register, and push it back after pushing 0.
   474  				p = appendp(p, ASet, regAddr(REG_PC_B))
   475  				p = appendp(p, AI32Const, constAddr(0))
   476  				p = appendp(p, AGet, regAddr(REG_PC_B))
   477  
   478  				p = appendp(p, ACallIndirect)
   479  
   480  			default:
   481  				panic("bad target for JMP")
   482  			}
   483  
   484  			p = appendp(p, AReturn)
   485  
   486  		case obj.ACALL, ACALLNORESUME:
   487  			call := *p
   488  			p.As = obj.ANOP
   489  
   490  			pcAfterCall := call.Link.Pc
   491  			if call.To.Sym == sigpanic {
   492  				pcAfterCall-- // sigpanic expects to be called without advancing the pc
   493  			}
   494  
   495  			// SP -= 8
   496  			p = appendp(p, AGet, regAddr(REG_SP))
   497  			p = appendp(p, AI32Const, constAddr(8))
   498  			p = appendp(p, AI32Sub)
   499  			p = appendp(p, ASet, regAddr(REG_SP))
   500  
   501  			// write return address to Go stack
   502  			p = appendp(p, AGet, regAddr(REG_SP))
   503  			p = appendp(p, AI64Const, obj.Addr{
   504  				Type:   obj.TYPE_ADDR,
   505  				Name:   obj.NAME_EXTERN,
   506  				Sym:    s,           // PC_F
   507  				Offset: pcAfterCall, // PC_B
   508  			})
   509  			p = appendp(p, AI64Store, constAddr(0))
   510  
   511  			// low-level WebAssembly call to function
   512  			switch call.To.Type {
   513  			case obj.TYPE_MEM:
   514  				if !notUsePC_B[call.To.Sym.Name] {
   515  					// Set PC_B parameter to function entry.
   516  					p = appendp(p, AI32Const, constAddr(0))
   517  				}
   518  				p = appendp(p, ACall, call.To)
   519  
   520  			case obj.TYPE_NONE:
   521  				// (target PC is on stack)
   522  				p = appendp(p, AI32WrapI64)
   523  				p = appendp(p, AI32Const, constAddr(16)) // only needs PC_F bits (16-31), PC_B bits (0-15) are zero
   524  				p = appendp(p, AI32ShrU)
   525  
   526  				// Set PC_B parameter to function entry.
   527  				// We need to push this before pushing the target PC_F,
   528  				// so temporarily pop PC_F, using our PC_B as a
   529  				// scratch register, and push it back after pushing 0.
   530  				p = appendp(p, ASet, regAddr(REG_PC_B))
   531  				p = appendp(p, AI32Const, constAddr(0))
   532  				p = appendp(p, AGet, regAddr(REG_PC_B))
   533  
   534  				p = appendp(p, ACallIndirect)
   535  
   536  			default:
   537  				panic("bad target for CALL")
   538  			}
   539  
   540  			// return value of call is on the top of the stack, indicating whether to unwind the WebAssembly stack
   541  			if call.As == ACALLNORESUME && call.To.Sym != sigpanic { // sigpanic unwinds the stack, but it never resumes
   542  				// trying to unwind WebAssembly stack but call has no resume point, terminate with error
   543  				p = appendp(p, AIf)
   544  				p = appendp(p, obj.AUNDEF)
   545  				p = appendp(p, AEnd)
   546  			} else {
   547  				// unwinding WebAssembly stack to switch goroutine, return 1
   548  				p = appendp(p, ABrIf)
   549  				unwindExitBranches = append(unwindExitBranches, p)
   550  			}
   551  
   552  		case obj.ARET, ARETUNWIND:
   553  			ret := *p
   554  			p.As = obj.ANOP
   555  
   556  			if framesize > 0 {
   557  				// SP += framesize
   558  				p = appendp(p, AGet, regAddr(REG_SP))
   559  				p = appendp(p, AI32Const, constAddr(framesize))
   560  				p = appendp(p, AI32Add)
   561  				p = appendp(p, ASet, regAddr(REG_SP))
   562  				// TODO(neelance): This should theoretically set Spadj, but it only works without.
   563  				// p.Spadj = int32(-framesize)
   564  			}
   565  
   566  			if ret.To.Type == obj.TYPE_MEM {
   567  				// Set PC_B parameter to function entry.
   568  				p = appendp(p, AI32Const, constAddr(0))
   569  
   570  				// low-level WebAssembly call to function
   571  				p = appendp(p, ACall, ret.To)
   572  				p = appendp(p, AReturn)
   573  				break
   574  			}
   575  
   576  			// SP += 8
   577  			p = appendp(p, AGet, regAddr(REG_SP))
   578  			p = appendp(p, AI32Const, constAddr(8))
   579  			p = appendp(p, AI32Add)
   580  			p = appendp(p, ASet, regAddr(REG_SP))
   581  
   582  			if ret.As == ARETUNWIND {
   583  				// function needs to unwind the WebAssembly stack, return 1
   584  				p = appendp(p, AI32Const, constAddr(1))
   585  				p = appendp(p, AReturn)
   586  				break
   587  			}
   588  
   589  			// not unwinding the WebAssembly stack, return 0
   590  			p = appendp(p, AI32Const, constAddr(0))
   591  			p = appendp(p, AReturn)
   592  		}
   593  	}
   594  
   595  	for p := s.Func().Text; p != nil; p = p.Link {
   596  		switch p.From.Name {
   597  		case obj.NAME_AUTO:
   598  			p.From.Offset += framesize
   599  		case obj.NAME_PARAM:
   600  			p.From.Reg = REG_SP
   601  			p.From.Offset += framesize + 8 // parameters are after the frame and the 8-byte return address
   602  		}
   603  
   604  		switch p.To.Name {
   605  		case obj.NAME_AUTO:
   606  			p.To.Offset += framesize
   607  		case obj.NAME_PARAM:
   608  			p.To.Reg = REG_SP
   609  			p.To.Offset += framesize + 8 // parameters are after the frame and the 8-byte return address
   610  		}
   611  
   612  		switch p.As {
   613  		case AGet:
   614  			if p.From.Type == obj.TYPE_ADDR {
   615  				get := *p
   616  				p.As = obj.ANOP
   617  
   618  				switch get.From.Name {
   619  				case obj.NAME_EXTERN:
   620  					p = appendp(p, AI64Const, get.From)
   621  				case obj.NAME_AUTO, obj.NAME_PARAM:
   622  					p = appendp(p, AGet, regAddr(get.From.Reg))
   623  					if get.From.Reg == REG_SP {
   624  						p = appendp(p, AI64ExtendI32U)
   625  					}
   626  					if get.From.Offset != 0 {
   627  						p = appendp(p, AI64Const, constAddr(get.From.Offset))
   628  						p = appendp(p, AI64Add)
   629  					}
   630  				default:
   631  					panic("bad Get: invalid name")
   632  				}
   633  			}
   634  
   635  		case AI32Load, AI64Load, AF32Load, AF64Load, AI32Load8S, AI32Load8U, AI32Load16S, AI32Load16U, AI64Load8S, AI64Load8U, AI64Load16S, AI64Load16U, AI64Load32S, AI64Load32U:
   636  			if p.From.Type == obj.TYPE_MEM {
   637  				as := p.As
   638  				from := p.From
   639  
   640  				p.As = AGet
   641  				p.From = regAddr(from.Reg)
   642  
   643  				if from.Reg != REG_SP {
   644  					p = appendp(p, AI32WrapI64)
   645  				}
   646  
   647  				p = appendp(p, as, constAddr(from.Offset))
   648  			}
   649  
   650  		case AMOVB, AMOVH, AMOVW, AMOVD:
   651  			mov := *p
   652  			p.As = obj.ANOP
   653  
   654  			var loadAs obj.As
   655  			var storeAs obj.As
   656  			switch mov.As {
   657  			case AMOVB:
   658  				loadAs = AI64Load8U
   659  				storeAs = AI64Store8
   660  			case AMOVH:
   661  				loadAs = AI64Load16U
   662  				storeAs = AI64Store16
   663  			case AMOVW:
   664  				loadAs = AI64Load32U
   665  				storeAs = AI64Store32
   666  			case AMOVD:
   667  				loadAs = AI64Load
   668  				storeAs = AI64Store
   669  			}
   670  
   671  			appendValue := func() {
   672  				switch mov.From.Type {
   673  				case obj.TYPE_CONST:
   674  					p = appendp(p, AI64Const, constAddr(mov.From.Offset))
   675  
   676  				case obj.TYPE_ADDR:
   677  					switch mov.From.Name {
   678  					case obj.NAME_NONE, obj.NAME_PARAM, obj.NAME_AUTO:
   679  						p = appendp(p, AGet, regAddr(mov.From.Reg))
   680  						if mov.From.Reg == REG_SP {
   681  							p = appendp(p, AI64ExtendI32U)
   682  						}
   683  						p = appendp(p, AI64Const, constAddr(mov.From.Offset))
   684  						p = appendp(p, AI64Add)
   685  					case obj.NAME_EXTERN:
   686  						p = appendp(p, AI64Const, mov.From)
   687  					default:
   688  						panic("bad name for MOV")
   689  					}
   690  
   691  				case obj.TYPE_REG:
   692  					p = appendp(p, AGet, mov.From)
   693  					if mov.From.Reg == REG_SP {
   694  						p = appendp(p, AI64ExtendI32U)
   695  					}
   696  
   697  				case obj.TYPE_MEM:
   698  					p = appendp(p, AGet, regAddr(mov.From.Reg))
   699  					if mov.From.Reg != REG_SP {
   700  						p = appendp(p, AI32WrapI64)
   701  					}
   702  					p = appendp(p, loadAs, constAddr(mov.From.Offset))
   703  
   704  				default:
   705  					panic("bad MOV type")
   706  				}
   707  			}
   708  
   709  			switch mov.To.Type {
   710  			case obj.TYPE_REG:
   711  				appendValue()
   712  				if mov.To.Reg == REG_SP {
   713  					p = appendp(p, AI32WrapI64)
   714  				}
   715  				p = appendp(p, ASet, mov.To)
   716  
   717  			case obj.TYPE_MEM:
   718  				switch mov.To.Name {
   719  				case obj.NAME_NONE, obj.NAME_PARAM:
   720  					p = appendp(p, AGet, regAddr(mov.To.Reg))
   721  					if mov.To.Reg != REG_SP {
   722  						p = appendp(p, AI32WrapI64)
   723  					}
   724  				case obj.NAME_EXTERN:
   725  					p = appendp(p, AI32Const, obj.Addr{Type: obj.TYPE_ADDR, Name: obj.NAME_EXTERN, Sym: mov.To.Sym})
   726  				default:
   727  					panic("bad MOV name")
   728  				}
   729  				appendValue()
   730  				p = appendp(p, storeAs, constAddr(mov.To.Offset))
   731  
   732  			default:
   733  				panic("bad MOV type")
   734  			}
   735  		}
   736  	}
   737  
   738  	{
   739  		p := s.Func().Text
   740  		if len(unwindExitBranches) > 0 {
   741  			p = appendp(p, ABlock) // unwindExit, used to return 1 when unwinding the stack
   742  			for _, b := range unwindExitBranches {
   743  				b.To = obj.Addr{Type: obj.TYPE_BRANCH, Val: p}
   744  			}
   745  		}
   746  		if len(entryPointLoopBranches) > 0 {
   747  			p = appendp(p, ALoop) // entryPointLoop, used to jump between basic blocks
   748  			for _, b := range entryPointLoopBranches {
   749  				b.To = obj.Addr{Type: obj.TYPE_BRANCH, Val: p}
   750  			}
   751  		}
   752  		if numResumePoints > 0 {
   753  			// Add Block instructions for resume points and BrTable to jump to selected resume point.
   754  			for i := 0; i < numResumePoints+1; i++ {
   755  				p = appendp(p, ABlock)
   756  			}
   757  			p = appendp(p, AGet, regAddr(REG_PC_B)) // read next basic block from PC_B
   758  			p = appendp(p, ABrTable, obj.Addr{Val: tableIdxs})
   759  			p = appendp(p, AEnd) // end of Block
   760  		}
   761  		for p.Link != nil {
   762  			p = p.Link // function instructions
   763  		}
   764  		if len(entryPointLoopBranches) > 0 {
   765  			p = appendp(p, AEnd) // end of entryPointLoop
   766  		}
   767  		p = appendp(p, obj.AUNDEF)
   768  		if len(unwindExitBranches) > 0 {
   769  			p = appendp(p, AEnd) // end of unwindExit
   770  			p = appendp(p, AI32Const, constAddr(1))
   771  		}
   772  	}
   773  
   774  	currentDepth = 0
   775  	blockDepths := make(map[*obj.Prog]int)
   776  	for p := s.Func().Text; p != nil; p = p.Link {
   777  		switch p.As {
   778  		case ABlock, ALoop, AIf:
   779  			currentDepth++
   780  			blockDepths[p] = currentDepth
   781  		case AEnd:
   782  			currentDepth--
   783  		}
   784  
   785  		switch p.As {
   786  		case ABr, ABrIf:
   787  			if p.To.Type == obj.TYPE_BRANCH {
   788  				blockDepth, ok := blockDepths[p.To.Val.(*obj.Prog)]
   789  				if !ok {
   790  					panic("label not at block")
   791  				}
   792  				p.To = constAddr(int64(currentDepth - blockDepth))
   793  			}
   794  		}
   795  	}
   796  }
   797  
   798  // Generate function body for wasmimport wrapper function.
   799  func genWasmImportWrapper(s *obj.LSym, appendp func(p *obj.Prog, as obj.As, args ...obj.Addr) *obj.Prog) {
   800  	wi := s.Func().WasmImport
   801  	wi.CreateAuxSym()
   802  	p := s.Func().Text
   803  	if p.Link != nil {
   804  		panic("wrapper functions for WASM imports should not have a body")
   805  	}
   806  	to := obj.Addr{
   807  		Type: obj.TYPE_MEM,
   808  		Name: obj.NAME_EXTERN,
   809  		Sym:  s,
   810  	}
   811  
   812  	// If the module that the import is for is our magic "gojs" module, then this
   813  	// indicates that the called function understands the Go stack-based call convention
   814  	// so we just pass the stack pointer to it, knowing it will read the params directly
   815  	// off the stack and push the results into memory based on the stack pointer.
   816  	if wi.Module == GojsModule {
   817  		// The called function has a signature of 'func(sp int)'. It has access to the memory
   818  		// value somewhere to be able to address the memory based on the "sp" value.
   819  
   820  		p = appendp(p, AGet, regAddr(REG_SP))
   821  		p = appendp(p, ACall, to)
   822  
   823  		p.Mark = WasmImport
   824  	} else {
   825  		if len(wi.Results) > 1 {
   826  			// TODO(evanphx) implement support for the multi-value proposal:
   827  			// https://github.com/WebAssembly/multi-value/blob/master/proposals/multi-value/Overview.md
   828  			panic("invalid results type") // impossible until multi-value proposal has landed
   829  		}
   830  		for _, f := range wi.Params {
   831  			// Each load instructions will consume the value of sp on the stack, so
   832  			// we need to read sp for each param. WASM appears to not have a stack dup instruction
   833  			// (a strange omission for a stack-based VM), if it did, we'd be using the dup here.
   834  			p = appendp(p, AGet, regAddr(REG_SP))
   835  
   836  			// Offset is the location of the param on the Go stack (ie relative to sp).
   837  			// Because of our call convention, the parameters are located an additional 8 bytes
   838  			// from sp because we store the return address as an int64 at the bottom of the stack.
   839  			// Ie the stack looks like [return_addr, param3, param2, param1, etc]
   840  
   841  			// Ergo, we add 8 to the true byte offset of the param to skip the return address.
   842  			loadOffset := f.Offset + 8
   843  
   844  			// We're reading the value from the Go stack onto the WASM stack and leaving it there
   845  			// for CALL to pick them up.
   846  			switch f.Type {
   847  			case obj.WasmI32:
   848  				p = appendp(p, AI32Load, constAddr(loadOffset))
   849  			case obj.WasmI64:
   850  				p = appendp(p, AI64Load, constAddr(loadOffset))
   851  			case obj.WasmF32:
   852  				p = appendp(p, AF32Load, constAddr(loadOffset))
   853  			case obj.WasmF64:
   854  				p = appendp(p, AF64Load, constAddr(loadOffset))
   855  			case obj.WasmPtr:
   856  				p = appendp(p, AI32Load, constAddr(loadOffset))
   857  			case obj.WasmBool:
   858  				p = appendp(p, AI32Load8U, constAddr(loadOffset))
   859  			default:
   860  				panic("bad param type")
   861  			}
   862  		}
   863  
   864  		// The call instruction is marked as being for a wasm import so that a later phase
   865  		// will generate relocation information that allows us to patch this with then
   866  		// offset of the imported function in the wasm imports.
   867  		p = appendp(p, ACall, to)
   868  		p.Mark = WasmImport
   869  
   870  		if len(wi.Results) == 1 {
   871  			f := wi.Results[0]
   872  
   873  			// Much like with the params, we need to adjust the offset we store the result value
   874  			// to by 8 bytes to account for the return address on the Go stack.
   875  			storeOffset := f.Offset + 8
   876  
   877  			// We need to push SP on the Wasm stack for the Store instruction, which needs to
   878  			// be pushed before the value (call result). So we pop the value into a register,
   879  			// push SP, and push the value back.
   880  			// We cannot get the SP onto the stack before the call, as if the host function
   881  			// calls back into Go, the Go stack may have moved.
   882  			switch f.Type {
   883  			case obj.WasmI32:
   884  				p = appendp(p, AI64ExtendI32U) // the register is 64-bit, so we have to extend
   885  				p = appendp(p, ASet, regAddr(REG_R0))
   886  				p = appendp(p, AGet, regAddr(REG_SP))
   887  				p = appendp(p, AGet, regAddr(REG_R0))
   888  				p = appendp(p, AI64Store32, constAddr(storeOffset))
   889  			case obj.WasmI64:
   890  				p = appendp(p, ASet, regAddr(REG_R0))
   891  				p = appendp(p, AGet, regAddr(REG_SP))
   892  				p = appendp(p, AGet, regAddr(REG_R0))
   893  				p = appendp(p, AI64Store, constAddr(storeOffset))
   894  			case obj.WasmF32:
   895  				p = appendp(p, ASet, regAddr(REG_F0))
   896  				p = appendp(p, AGet, regAddr(REG_SP))
   897  				p = appendp(p, AGet, regAddr(REG_F0))
   898  				p = appendp(p, AF32Store, constAddr(storeOffset))
   899  			case obj.WasmF64:
   900  				p = appendp(p, ASet, regAddr(REG_F16))
   901  				p = appendp(p, AGet, regAddr(REG_SP))
   902  				p = appendp(p, AGet, regAddr(REG_F16))
   903  				p = appendp(p, AF64Store, constAddr(storeOffset))
   904  			case obj.WasmPtr:
   905  				p = appendp(p, AI64ExtendI32U)
   906  				p = appendp(p, ASet, regAddr(REG_R0))
   907  				p = appendp(p, AGet, regAddr(REG_SP))
   908  				p = appendp(p, AGet, regAddr(REG_R0))
   909  				p = appendp(p, AI64Store, constAddr(storeOffset))
   910  			case obj.WasmBool:
   911  				p = appendp(p, AI64ExtendI32U)
   912  				p = appendp(p, ASet, regAddr(REG_R0))
   913  				p = appendp(p, AGet, regAddr(REG_SP))
   914  				p = appendp(p, AGet, regAddr(REG_R0))
   915  				p = appendp(p, AI64Store8, constAddr(storeOffset))
   916  			default:
   917  				panic("bad result type")
   918  			}
   919  		}
   920  	}
   921  
   922  	p = appendp(p, obj.ARET)
   923  }
   924  
   925  // Generate function body for wasmexport wrapper function.
   926  func genWasmExportWrapper(s *obj.LSym, appendp func(p *obj.Prog, as obj.As, args ...obj.Addr) *obj.Prog) {
   927  	we := s.Func().WasmExport
   928  	we.CreateAuxSym()
   929  	p := s.Func().Text
   930  	framesize := p.To.Offset
   931  	for p.Link != nil && p.Link.As == obj.AFUNCDATA {
   932  		p = p.Link
   933  	}
   934  	if p.Link != nil {
   935  		panic("wrapper functions for WASM export should not have a body")
   936  	}
   937  
   938  	// Store args
   939  	for i, f := range we.Params {
   940  		p = appendp(p, AGet, regAddr(REG_SP))
   941  		p = appendp(p, AGet, regAddr(REG_R0+int16(i)))
   942  		switch f.Type {
   943  		case obj.WasmI32:
   944  			p = appendp(p, AI32Store, constAddr(f.Offset))
   945  		case obj.WasmI64:
   946  			p = appendp(p, AI64Store, constAddr(f.Offset))
   947  		case obj.WasmF32:
   948  			p = appendp(p, AF32Store, constAddr(f.Offset))
   949  		case obj.WasmF64:
   950  			p = appendp(p, AF64Store, constAddr(f.Offset))
   951  		case obj.WasmPtr:
   952  			p = appendp(p, AI64ExtendI32U)
   953  			p = appendp(p, AI64Store, constAddr(f.Offset))
   954  		case obj.WasmBool:
   955  			p = appendp(p, AI32Store8, constAddr(f.Offset))
   956  		default:
   957  			panic("bad param type")
   958  		}
   959  	}
   960  
   961  	// Call the Go function.
   962  	// XXX maybe use ACALL and let later phase expand? But we don't use PC_B. Maybe we should?
   963  	// Go calling convention expects we push a return PC before call.
   964  	// SP -= 8
   965  	p = appendp(p, AGet, regAddr(REG_SP))
   966  	p = appendp(p, AI32Const, constAddr(8))
   967  	p = appendp(p, AI32Sub)
   968  	p = appendp(p, ASet, regAddr(REG_SP))
   969  	// write return address to Go stack
   970  	p = appendp(p, AGet, regAddr(REG_SP))
   971  	retAddr := obj.Addr{
   972  		Type:   obj.TYPE_ADDR,
   973  		Name:   obj.NAME_EXTERN,
   974  		Sym:    s, // PC_F
   975  		Offset: 1, // PC_B=1, past the prologue, so we have the right SP delta
   976  	}
   977  	if framesize == 0 {
   978  		// Frameless function, no prologue.
   979  		retAddr.Offset = 0
   980  	}
   981  	p = appendp(p, AI64Const, retAddr)
   982  	p = appendp(p, AI64Store, constAddr(0))
   983  	// Set PC_B parameter to function entry
   984  	p = appendp(p, AI32Const, constAddr(0))
   985  	p = appendp(p, ACall, obj.Addr{Type: obj.TYPE_MEM, Name: obj.NAME_EXTERN, Sym: we.WrappedSym})
   986  	// Return value is on the top of the stack, indicating whether to unwind the Wasm stack.
   987  	// In the unwinding case, we call wasm_pc_f_loop_export to handle stack switch and rewinding,
   988  	// until a normal return (non-unwinding) back to this function.
   989  	p = appendp(p, AIf)
   990  	p = appendp(p, AI32Const, retAddr)
   991  	p = appendp(p, AI32Const, constAddr(16))
   992  	p = appendp(p, AI32ShrU)
   993  	p = appendp(p, ACall, obj.Addr{Type: obj.TYPE_MEM, Name: obj.NAME_EXTERN, Sym: wasm_pc_f_loop_export})
   994  	p = appendp(p, AEnd)
   995  
   996  	// Load result
   997  	if len(we.Results) > 1 {
   998  		panic("invalid results type")
   999  	} else if len(we.Results) == 1 {
  1000  		p = appendp(p, AGet, regAddr(REG_SP))
  1001  		f := we.Results[0]
  1002  		switch f.Type {
  1003  		case obj.WasmI32:
  1004  			p = appendp(p, AI32Load, constAddr(f.Offset))
  1005  		case obj.WasmI64:
  1006  			p = appendp(p, AI64Load, constAddr(f.Offset))
  1007  		case obj.WasmF32:
  1008  			p = appendp(p, AF32Load, constAddr(f.Offset))
  1009  		case obj.WasmF64:
  1010  			p = appendp(p, AF64Load, constAddr(f.Offset))
  1011  		case obj.WasmPtr:
  1012  			p = appendp(p, AI32Load, constAddr(f.Offset))
  1013  		case obj.WasmBool:
  1014  			p = appendp(p, AI32Load8U, constAddr(f.Offset))
  1015  		default:
  1016  			panic("bad result type")
  1017  		}
  1018  	}
  1019  
  1020  	// Epilogue. Cannot use ARET as we don't follow Go calling convention.
  1021  	if framesize > 0 {
  1022  		// SP += framesize
  1023  		p = appendp(p, AGet, regAddr(REG_SP))
  1024  		p = appendp(p, AI32Const, constAddr(framesize))
  1025  		p = appendp(p, AI32Add)
  1026  		p = appendp(p, ASet, regAddr(REG_SP))
  1027  	}
  1028  	p = appendp(p, AReturn)
  1029  }
  1030  
  1031  func constAddr(value int64) obj.Addr {
  1032  	return obj.Addr{Type: obj.TYPE_CONST, Offset: value}
  1033  }
  1034  
  1035  func regAddr(reg int16) obj.Addr {
  1036  	return obj.Addr{Type: obj.TYPE_REG, Reg: reg}
  1037  }
  1038  
  1039  // Most of the Go functions has a single parameter (PC_B) in
  1040  // Wasm ABI. This is a list of exceptions.
  1041  var notUsePC_B = map[string]bool{
  1042  	"_rt0_wasm_js":            true,
  1043  	"_rt0_wasm_wasip1":        true,
  1044  	"_rt0_wasm_wasip1_lib":    true,
  1045  	"wasm_export_run":         true,
  1046  	"wasm_export_resume":      true,
  1047  	"wasm_export_getsp":       true,
  1048  	"wasm_pc_f_loop":          true,
  1049  	"wasm_pc_f_loop_export":   true,
  1050  	"gcWriteBarrier":          true,
  1051  	"runtime.gcWriteBarrier1": true,
  1052  	"runtime.gcWriteBarrier2": true,
  1053  	"runtime.gcWriteBarrier3": true,
  1054  	"runtime.gcWriteBarrier4": true,
  1055  	"runtime.gcWriteBarrier5": true,
  1056  	"runtime.gcWriteBarrier6": true,
  1057  	"runtime.gcWriteBarrier7": true,
  1058  	"runtime.gcWriteBarrier8": true,
  1059  	"runtime.wasmDiv":         true,
  1060  	"runtime.wasmTruncS":      true,
  1061  	"runtime.wasmTruncU":      true,
  1062  	"cmpbody":                 true,
  1063  	"memeqbody":               true,
  1064  	"memcmp":                  true,
  1065  	"memchr":                  true,
  1066  }
  1067  
  1068  func assemble(ctxt *obj.Link, s *obj.LSym, newprog obj.ProgAlloc) {
  1069  	type regVar struct {
  1070  		global bool
  1071  		index  uint64
  1072  	}
  1073  
  1074  	type varDecl struct {
  1075  		count uint64
  1076  		typ   valueType
  1077  	}
  1078  
  1079  	hasLocalSP := false
  1080  	regVars := [MAXREG - MINREG]*regVar{
  1081  		REG_SP - MINREG:    {true, 0},
  1082  		REG_CTXT - MINREG:  {true, 1},
  1083  		REG_g - MINREG:     {true, 2},
  1084  		REG_RET0 - MINREG:  {true, 3},
  1085  		REG_RET1 - MINREG:  {true, 4},
  1086  		REG_RET2 - MINREG:  {true, 5},
  1087  		REG_RET3 - MINREG:  {true, 6},
  1088  		REG_PAUSE - MINREG: {true, 7},
  1089  	}
  1090  	var varDecls []*varDecl
  1091  	useAssemblyRegMap := func() {
  1092  		for i := int16(0); i < 16; i++ {
  1093  			regVars[REG_R0+i-MINREG] = &regVar{false, uint64(i)}
  1094  		}
  1095  	}
  1096  
  1097  	// Function starts with declaration of locals: numbers and types.
  1098  	// Some functions use a special calling convention.
  1099  	switch s.Name {
  1100  	case "_rt0_wasm_js", "_rt0_wasm_wasip1", "_rt0_wasm_wasip1_lib",
  1101  		"wasm_export_run", "wasm_export_resume", "wasm_export_getsp",
  1102  		"wasm_pc_f_loop", "runtime.wasmDiv", "runtime.wasmTruncS", "runtime.wasmTruncU", "memeqbody":
  1103  		varDecls = []*varDecl{}
  1104  		useAssemblyRegMap()
  1105  	case "wasm_pc_f_loop_export":
  1106  		varDecls = []*varDecl{{count: 2, typ: i32}}
  1107  		useAssemblyRegMap()
  1108  	case "memchr", "memcmp":
  1109  		varDecls = []*varDecl{{count: 2, typ: i32}}
  1110  		useAssemblyRegMap()
  1111  	case "cmpbody":
  1112  		varDecls = []*varDecl{{count: 2, typ: i64}}
  1113  		useAssemblyRegMap()
  1114  	case "gcWriteBarrier":
  1115  		varDecls = []*varDecl{{count: 5, typ: i64}}
  1116  		useAssemblyRegMap()
  1117  	case "runtime.gcWriteBarrier1",
  1118  		"runtime.gcWriteBarrier2",
  1119  		"runtime.gcWriteBarrier3",
  1120  		"runtime.gcWriteBarrier4",
  1121  		"runtime.gcWriteBarrier5",
  1122  		"runtime.gcWriteBarrier6",
  1123  		"runtime.gcWriteBarrier7",
  1124  		"runtime.gcWriteBarrier8":
  1125  		// no locals
  1126  		useAssemblyRegMap()
  1127  	default:
  1128  		if s.Func().WasmExport != nil {
  1129  			// no local SP, not following Go calling convention
  1130  			useAssemblyRegMap()
  1131  			break
  1132  		}
  1133  
  1134  		// Normal calling convention: PC_B as WebAssembly parameter. First local variable is local SP cache.
  1135  		regVars[REG_PC_B-MINREG] = &regVar{false, 0}
  1136  		hasLocalSP = true
  1137  
  1138  		var regUsed [MAXREG - MINREG]bool
  1139  		for p := s.Func().Text; p != nil; p = p.Link {
  1140  			if p.From.Reg != 0 {
  1141  				regUsed[p.From.Reg-MINREG] = true
  1142  			}
  1143  			if p.To.Reg != 0 {
  1144  				regUsed[p.To.Reg-MINREG] = true
  1145  			}
  1146  		}
  1147  
  1148  		regs := []int16{REG_SP}
  1149  		for reg := int16(REG_R0); reg <= REG_F31; reg++ {
  1150  			if regUsed[reg-MINREG] {
  1151  				regs = append(regs, reg)
  1152  			}
  1153  		}
  1154  
  1155  		var lastDecl *varDecl
  1156  		for i, reg := range regs {
  1157  			t := regType(reg)
  1158  			if lastDecl == nil || lastDecl.typ != t {
  1159  				lastDecl = &varDecl{
  1160  					count: 0,
  1161  					typ:   t,
  1162  				}
  1163  				varDecls = append(varDecls, lastDecl)
  1164  			}
  1165  			lastDecl.count++
  1166  			if reg != REG_SP {
  1167  				regVars[reg-MINREG] = &regVar{false, 1 + uint64(i)}
  1168  			}
  1169  		}
  1170  	}
  1171  
  1172  	w := new(bytes.Buffer)
  1173  
  1174  	writeUleb128(w, uint64(len(varDecls)))
  1175  	for _, decl := range varDecls {
  1176  		writeUleb128(w, decl.count)
  1177  		w.WriteByte(byte(decl.typ))
  1178  	}
  1179  
  1180  	if hasLocalSP {
  1181  		// Copy SP from its global variable into a local variable. Accessing a local variable is more efficient.
  1182  		updateLocalSP(w)
  1183  	}
  1184  
  1185  	for p := s.Func().Text; p != nil; p = p.Link {
  1186  		switch p.As {
  1187  		case AGet:
  1188  			if p.From.Type != obj.TYPE_REG {
  1189  				panic("bad Get: argument is not a register")
  1190  			}
  1191  			reg := p.From.Reg
  1192  			v := regVars[reg-MINREG]
  1193  			if v == nil {
  1194  				panic("bad Get: invalid register")
  1195  			}
  1196  			if reg == REG_SP && hasLocalSP {
  1197  				writeOpcode(w, ALocalGet)
  1198  				writeUleb128(w, 1) // local SP
  1199  				continue
  1200  			}
  1201  			if v.global {
  1202  				writeOpcode(w, AGlobalGet)
  1203  			} else {
  1204  				writeOpcode(w, ALocalGet)
  1205  			}
  1206  			writeUleb128(w, v.index)
  1207  			continue
  1208  
  1209  		case ASet:
  1210  			if p.To.Type != obj.TYPE_REG {
  1211  				panic("bad Set: argument is not a register")
  1212  			}
  1213  			reg := p.To.Reg
  1214  			v := regVars[reg-MINREG]
  1215  			if v == nil {
  1216  				panic("bad Set: invalid register")
  1217  			}
  1218  			if reg == REG_SP && hasLocalSP {
  1219  				writeOpcode(w, ALocalTee)
  1220  				writeUleb128(w, 1) // local SP
  1221  			}
  1222  			if v.global {
  1223  				writeOpcode(w, AGlobalSet)
  1224  			} else {
  1225  				if p.Link.As == AGet && p.Link.From.Reg == reg {
  1226  					writeOpcode(w, ALocalTee)
  1227  					p = p.Link
  1228  				} else {
  1229  					writeOpcode(w, ALocalSet)
  1230  				}
  1231  			}
  1232  			writeUleb128(w, v.index)
  1233  			continue
  1234  
  1235  		case ATee:
  1236  			if p.To.Type != obj.TYPE_REG {
  1237  				panic("bad Tee: argument is not a register")
  1238  			}
  1239  			reg := p.To.Reg
  1240  			v := regVars[reg-MINREG]
  1241  			if v == nil {
  1242  				panic("bad Tee: invalid register")
  1243  			}
  1244  			writeOpcode(w, ALocalTee)
  1245  			writeUleb128(w, v.index)
  1246  			continue
  1247  
  1248  		case ANot:
  1249  			writeOpcode(w, AI32Eqz)
  1250  			continue
  1251  
  1252  		case obj.AUNDEF:
  1253  			writeOpcode(w, AUnreachable)
  1254  			continue
  1255  
  1256  		case obj.ANOP, obj.ATEXT, obj.AFUNCDATA, obj.APCDATA:
  1257  			// ignore
  1258  			continue
  1259  		}
  1260  
  1261  		writeOpcode(w, p.As)
  1262  
  1263  		switch p.As {
  1264  		case ABlock, ALoop, AIf:
  1265  			if p.From.Offset != 0 {
  1266  				// block type, rarely used, e.g. for code compiled with emscripten
  1267  				w.WriteByte(0x80 - byte(p.From.Offset))
  1268  				continue
  1269  			}
  1270  			w.WriteByte(0x40)
  1271  
  1272  		case ABr, ABrIf:
  1273  			if p.To.Type != obj.TYPE_CONST {
  1274  				panic("bad Br/BrIf")
  1275  			}
  1276  			writeUleb128(w, uint64(p.To.Offset))
  1277  
  1278  		case ABrTable:
  1279  			idxs := p.To.Val.([]uint64)
  1280  			writeUleb128(w, uint64(len(idxs)-1))
  1281  			for _, idx := range idxs {
  1282  				writeUleb128(w, idx)
  1283  			}
  1284  
  1285  		case ACall:
  1286  			switch p.To.Type {
  1287  			case obj.TYPE_CONST:
  1288  				writeUleb128(w, uint64(p.To.Offset))
  1289  
  1290  			case obj.TYPE_MEM:
  1291  				if p.To.Name != obj.NAME_EXTERN && p.To.Name != obj.NAME_STATIC {
  1292  					fmt.Println(p.To)
  1293  					panic("bad name for Call")
  1294  				}
  1295  				typ := objabi.R_CALL
  1296  				if p.Mark&WasmImport != 0 {
  1297  					typ = objabi.R_WASMIMPORT
  1298  				}
  1299  				s.AddRel(ctxt, obj.Reloc{
  1300  					Type: typ,
  1301  					Off:  int32(w.Len()),
  1302  					Siz:  1, // actually variable sized
  1303  					Sym:  p.To.Sym,
  1304  				})
  1305  				if hasLocalSP {
  1306  					// The stack may have moved, which changes SP. Update the local SP variable.
  1307  					updateLocalSP(w)
  1308  				}
  1309  
  1310  			default:
  1311  				panic("bad type for Call")
  1312  			}
  1313  
  1314  		case ACallIndirect:
  1315  			writeUleb128(w, uint64(p.To.Offset))
  1316  			w.WriteByte(0x00) // reserved value
  1317  			if hasLocalSP {
  1318  				// The stack may have moved, which changes SP. Update the local SP variable.
  1319  				updateLocalSP(w)
  1320  			}
  1321  
  1322  		case AI32Const, AI64Const:
  1323  			if p.From.Name == obj.NAME_EXTERN {
  1324  				s.AddRel(ctxt, obj.Reloc{
  1325  					Type: objabi.R_ADDR,
  1326  					Off:  int32(w.Len()),
  1327  					Siz:  1, // actually variable sized
  1328  					Sym:  p.From.Sym,
  1329  					Add:  p.From.Offset,
  1330  				})
  1331  				break
  1332  			}
  1333  			writeSleb128(w, p.From.Offset)
  1334  
  1335  		case AF32Const:
  1336  			b := make([]byte, 4)
  1337  			binary.LittleEndian.PutUint32(b, math.Float32bits(float32(p.From.Val.(float64))))
  1338  			w.Write(b)
  1339  
  1340  		case AF64Const:
  1341  			b := make([]byte, 8)
  1342  			binary.LittleEndian.PutUint64(b, math.Float64bits(p.From.Val.(float64)))
  1343  			w.Write(b)
  1344  
  1345  		case AI32Load, AI64Load, AF32Load, AF64Load, AI32Load8S, AI32Load8U, AI32Load16S, AI32Load16U, AI64Load8S, AI64Load8U, AI64Load16S, AI64Load16U, AI64Load32S, AI64Load32U:
  1346  			if p.From.Offset < 0 {
  1347  				panic("negative offset for *Load")
  1348  			}
  1349  			if p.From.Type != obj.TYPE_CONST {
  1350  				panic("bad type for *Load")
  1351  			}
  1352  			if p.From.Offset > math.MaxUint32 {
  1353  				ctxt.Diag("bad offset in %v", p)
  1354  			}
  1355  			writeUleb128(w, align(p.As))
  1356  			writeUleb128(w, uint64(p.From.Offset))
  1357  
  1358  		case AI32Store, AI64Store, AF32Store, AF64Store, AI32Store8, AI32Store16, AI64Store8, AI64Store16, AI64Store32:
  1359  			if p.To.Offset < 0 {
  1360  				panic("negative offset")
  1361  			}
  1362  			if p.From.Offset > math.MaxUint32 {
  1363  				ctxt.Diag("bad offset in %v", p)
  1364  			}
  1365  			writeUleb128(w, align(p.As))
  1366  			writeUleb128(w, uint64(p.To.Offset))
  1367  
  1368  		case ACurrentMemory, AGrowMemory, AMemoryFill:
  1369  			w.WriteByte(0x00)
  1370  
  1371  		case AMemoryCopy:
  1372  			w.WriteByte(0x00)
  1373  			w.WriteByte(0x00)
  1374  
  1375  		}
  1376  	}
  1377  
  1378  	w.WriteByte(0x0b) // end
  1379  
  1380  	s.P = w.Bytes()
  1381  }
  1382  
  1383  func updateLocalSP(w *bytes.Buffer) {
  1384  	writeOpcode(w, AGlobalGet)
  1385  	writeUleb128(w, 0) // global SP
  1386  	writeOpcode(w, ALocalSet)
  1387  	writeUleb128(w, 1) // local SP
  1388  }
  1389  
  1390  func writeOpcode(w *bytes.Buffer, as obj.As) {
  1391  	switch {
  1392  	case as < AUnreachable:
  1393  		panic(fmt.Sprintf("unexpected assembler op: %s", as))
  1394  	case as < AEnd:
  1395  		w.WriteByte(byte(as - AUnreachable + 0x00))
  1396  	case as < ADrop:
  1397  		w.WriteByte(byte(as - AEnd + 0x0B))
  1398  	case as < ALocalGet:
  1399  		w.WriteByte(byte(as - ADrop + 0x1A))
  1400  	case as < AI32Load:
  1401  		w.WriteByte(byte(as - ALocalGet + 0x20))
  1402  	case as < AI32TruncSatF32S:
  1403  		w.WriteByte(byte(as - AI32Load + 0x28))
  1404  	case as < ALast:
  1405  		w.WriteByte(0xFC)
  1406  		w.WriteByte(byte(as - AI32TruncSatF32S + 0x00))
  1407  	default:
  1408  		panic(fmt.Sprintf("unexpected assembler op: %s", as))
  1409  	}
  1410  }
  1411  
  1412  type valueType byte
  1413  
  1414  const (
  1415  	i32 valueType = 0x7F
  1416  	i64 valueType = 0x7E
  1417  	f32 valueType = 0x7D
  1418  	f64 valueType = 0x7C
  1419  )
  1420  
  1421  func regType(reg int16) valueType {
  1422  	switch {
  1423  	case reg == REG_SP:
  1424  		return i32
  1425  	case reg >= REG_R0 && reg <= REG_R15:
  1426  		return i64
  1427  	case reg >= REG_F0 && reg <= REG_F15:
  1428  		return f32
  1429  	case reg >= REG_F16 && reg <= REG_F31:
  1430  		return f64
  1431  	default:
  1432  		panic("invalid register")
  1433  	}
  1434  }
  1435  
  1436  func align(as obj.As) uint64 {
  1437  	switch as {
  1438  	case AI32Load8S, AI32Load8U, AI64Load8S, AI64Load8U, AI32Store8, AI64Store8:
  1439  		return 0
  1440  	case AI32Load16S, AI32Load16U, AI64Load16S, AI64Load16U, AI32Store16, AI64Store16:
  1441  		return 1
  1442  	case AI32Load, AF32Load, AI64Load32S, AI64Load32U, AI32Store, AF32Store, AI64Store32:
  1443  		return 2
  1444  	case AI64Load, AF64Load, AI64Store, AF64Store:
  1445  		return 3
  1446  	default:
  1447  		panic("align: bad op")
  1448  	}
  1449  }
  1450  
  1451  func writeUleb128(w io.ByteWriter, v uint64) {
  1452  	if v < 128 {
  1453  		w.WriteByte(uint8(v))
  1454  		return
  1455  	}
  1456  	more := true
  1457  	for more {
  1458  		c := uint8(v & 0x7f)
  1459  		v >>= 7
  1460  		more = v != 0
  1461  		if more {
  1462  			c |= 0x80
  1463  		}
  1464  		w.WriteByte(c)
  1465  	}
  1466  }
  1467  
  1468  func writeSleb128(w io.ByteWriter, v int64) {
  1469  	more := true
  1470  	for more {
  1471  		c := uint8(v & 0x7f)
  1472  		s := uint8(v & 0x40)
  1473  		v >>= 7
  1474  		more = !((v == 0 && s == 0) || (v == -1 && s != 0))
  1475  		if more {
  1476  			c |= 0x80
  1477  		}
  1478  		w.WriteByte(c)
  1479  	}
  1480  }
  1481
View as plain text