cmd/internal/obj: generate SEH aux symbols for windows/amd64

This CL updates the Go compiler so it generate SEH unwind info [1] as a
function auxiliary symbol when building for windows/amd64.

A follow up CL will teach the Go linker how to assemble these codes
into the PE .xdata section.

Updates #57302

[1] https://learn.microsoft.com/en-us/cpp/build/exception-handling-x64#struct-unwind_info

Change-Id: I40ae0437bfee326c1a67c2b5e1496f0bf3ecea17
Reviewed-on: https://go-review.googlesource.com/c/go/+/461749
Reviewed-by: Davis Goodin <[email protected]>
Reviewed-by: Michael Knyszek <[email protected]>
TryBot-Result: Gopher Robot <[email protected]>
Reviewed-by: Than McIntosh <[email protected]>
Run-TryBot: Quim Muntal <[email protected]>

src/cmd/internal/goobj/objfile.go

@@ -443,6 +443,7 @@ const (
 	AuxPcinline
 	AuxPcdata
 	AuxWasmImport
+	AuxSehUnwindInfo
 )
 
 func (a *Aux) Type() uint8 { return a[0] }

src/cmd/internal/obj/link.go

@@ -503,6 +503,8 @@ type FuncInfo struct {
 	FuncInfoSym   *LSym
 	WasmImportSym *LSym
 	WasmImport    *WasmImport
+
+	sehUnwindInfoSym *LSym
 }
 
 // JumpTable represents a table used for implementing multi-way
@@ -1072,6 +1074,7 @@ type LinkArch struct {
 	Preprocess     func(*Link, *LSym, ProgAlloc)
 	Assemble       func(*Link, *LSym, ProgAlloc)
 	Progedit       func(*Link, *Prog, ProgAlloc)
+	SEH            func(*Link, *LSym) *LSym
 	UnaryDst       map[As]bool // Instruction takes one operand, a destination.
 	DWARFRegisters map[int16]int16
 }

src/cmd/internal/obj/objfile.go

@@ -602,6 +602,9 @@ func (w *writer) Aux(s *LSym) {
 		if fn.Pcln.Pcinline != nil && fn.Pcln.Pcinline.Size != 0 {
 			w.aux1(goobj.AuxPcinline, fn.Pcln.Pcinline)
 		}
+		if fn.sehUnwindInfoSym != nil && fn.sehUnwindInfoSym.Size != 0 {
+			w.aux1(goobj.AuxSehUnwindInfo, fn.sehUnwindInfoSym)
+		}
 		for _, pcSym := range fn.Pcln.Pcdata {
 			w.aux1(goobj.AuxPcdata, pcSym)
 		}
@@ -707,6 +710,9 @@ func nAuxSym(s *LSym) int {
 		if fn.Pcln.Pcinline != nil && fn.Pcln.Pcinline.Size != 0 {
 			n++
 		}
+		if fn.sehUnwindInfoSym != nil && fn.sehUnwindInfoSym.Size != 0 {
+			n++
+		}
 		n += len(fn.Pcln.Pcdata)
 		if fn.WasmImport != nil {
 			if fn.WasmImportSym == nil || fn.WasmImportSym.Size == 0 {
@@ -770,7 +776,7 @@ func genFuncInfoSyms(ctxt *Link) {
 		fn.FuncInfoSym = isym
 		b.Reset()
 
-		auxsyms := []*LSym{fn.dwarfRangesSym, fn.dwarfLocSym, fn.dwarfDebugLinesSym, fn.dwarfInfoSym, fn.WasmImportSym}
+		auxsyms := []*LSym{fn.dwarfRangesSym, fn.dwarfLocSym, fn.dwarfDebugLinesSym, fn.dwarfInfoSym, fn.WasmImportSym, fn.sehUnwindInfoSym}
 		for _, s := range auxsyms {
 			if s == nil || s.Size == 0 {
 				continue

src/cmd/internal/obj/plist.go

@@ -157,6 +157,9 @@ func Flushplist(ctxt *Link, plist *Plist, newprog ProgAlloc, myimportpath string
 		if myimportpath != "" {
 			ctxt.populateDWARF(plist.Curfn, s, myimportpath)
 		}
+		if ctxt.Headtype == objabi.Hwindows && ctxt.Arch.SEH != nil {
+			s.Func().sehUnwindInfoSym = ctxt.Arch.SEH(ctxt, s)
+		}
 	}
 }
 

src/cmd/internal/obj/sym.go

@@ -416,7 +416,7 @@ func (ctxt *Link) traverseFuncAux(flag traverseFlag, fsym *LSym, fn func(parent
 		}
 	}
 
-	auxsyms := []*LSym{fninfo.dwarfRangesSym, fninfo.dwarfLocSym, fninfo.dwarfDebugLinesSym, fninfo.dwarfInfoSym, fninfo.WasmImportSym}
+	auxsyms := []*LSym{fninfo.dwarfRangesSym, fninfo.dwarfLocSym, fninfo.dwarfDebugLinesSym, fninfo.dwarfInfoSym, fninfo.WasmImportSym, fninfo.sehUnwindInfoSym}
 	for _, s := range auxsyms {
 		if s == nil || s.Size == 0 {
 			continue

src/cmd/internal/obj/x86/obj6.go

@@ -628,6 +628,7 @@ func preprocess(ctxt *obj.Link, cursym *obj.LSym, newprog obj.ProgAlloc) {
 		p.To.Offset += int64(bpsize)
 	} else {
 		bpsize = 0
+		p.From.Sym.Set(obj.AttrNoFrame, true)
 	}
 
 	textarg := int64(p.To.Val.(int32))
@@ -1526,6 +1527,7 @@ var Linkamd64 = obj.LinkArch{
 	Preprocess:     preprocess,
 	Assemble:       span6,
 	Progedit:       progedit,
+	SEH:            populateSeh,
 	UnaryDst:       unaryDst,
 	DWARFRegisters: AMD64DWARFRegisters,
 }

src/cmd/internal/obj/x86/seh.go

@@ -0,0 +1,141 @@
+// Copyright 2023 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package x86
+
+import (
+	"cmd/internal/obj"
+	"cmd/internal/objabi"
+	"cmd/internal/src"
+	"encoding/base64"
+	"fmt"
+	"math"
+)
+
+type sehbuf struct {
+	ctxt *obj.Link
+	data []byte
+	off  int
+}
+
+func newsehbuf(ctxt *obj.Link, nodes uint8) sehbuf {
+	// - 8 bytes for the header
+	// - 2 bytes for each node
+	// - 2 bytes in case nodes is not even
+	size := 8 + nodes*2
+	if nodes%2 != 0 {
+		size += 2
+	}
+	return sehbuf{ctxt, make([]byte, size), 0}
+}
+
+func (b *sehbuf) write8(v uint8) {
+	b.data[b.off] = v
+	b.off++
+}
+
+func (b *sehbuf) write32(v uint32) {
+	b.ctxt.Arch.ByteOrder.PutUint32(b.data[b.off:], v)
+	b.off += 4
+}
+
+func (b *sehbuf) writecode(op, value uint8) {
+	b.write8(value<<4 | op)
+}
+
+// populateSeh generates the SEH unwind information for s.
+func populateSeh(ctxt *obj.Link, s *obj.LSym) (sehsym *obj.LSym) {
+	if s.NoFrame() {
+		return
+	}
+
+	// This implementation expects the following function prologue layout:
+	// - Stack split code (optional)
+	// - PUSHQ	BP
+	// - MOVQ	SP,	BP
+	//
+	// If the prologue layout change, the unwind information should be updated
+	// accordingly.
+
+	// Search for the PUSHQ BP instruction inside the prologue.
+	var pushbp *obj.Prog
+	for p := s.Func().Text; p != nil; p = p.Link {
+		if p.As == APUSHQ && p.From.Type == obj.TYPE_REG && p.From.Reg == REG_BP {
+			pushbp = p
+			break
+		}
+		if p.Pos.Xlogue() == src.PosPrologueEnd {
+			break
+		}
+	}
+	if pushbp == nil {
+		ctxt.Diag("missing frame pointer instruction: PUSHQ BP")
+		return
+	}
+
+	// It must be followed by a MOVQ SP, BP.
+	movbp := pushbp.Link
+	if movbp == nil {
+		ctxt.Diag("missing frame pointer instruction: MOVQ SP, BP")
+		return
+	}
+	if !(movbp.As == AMOVQ && movbp.From.Type == obj.TYPE_REG && movbp.From.Reg == REG_SP &&
+		movbp.To.Type == obj.TYPE_REG && movbp.To.Reg == REG_BP && movbp.From.Offset == 0) {
+		ctxt.Diag("unexpected frame pointer instruction\n%v", movbp)
+		return
+	}
+	if movbp.Link.Pc > math.MaxUint8 {
+		// SEH unwind information don't support prologues that are more than 255 bytes long.
+		// These are very rare, but still possible, e.g., when compiling functions with many
+		// parameters with -gcflags=-d=maymorestack=runtime.mayMoreStackPreempt.
+		// Return without reporting an error.
+		return
+	}
+
+	// Reference:
+	// https://learn.microsoft.com/en-us/cpp/build/exception-handling-x64#struct-unwind_info
+
+	const (
+		UWOP_PUSH_NONVOL = 0
+		UWOP_SET_FPREG   = 3
+		SEH_REG_BP       = 5
+	)
+
+	// Fow now we only support operations which are encoded
+	// using a single 2-byte node, so the number of nodes
+	// is the number of operations.
+	nodes := uint8(2)
+	buf := newsehbuf(ctxt, nodes)
+	buf.write8(1)                    // Flags + version
+	buf.write8(uint8(movbp.Link.Pc)) // Size of prolog
+	buf.write8(nodes)                // Count of nodes
+	buf.write8(SEH_REG_BP)           // FP register
+
+	// Notes are written in reverse order of appearance.
+	buf.write8(uint8(movbp.Link.Pc))
+	buf.writecode(UWOP_SET_FPREG, 0)
+
+	buf.write8(uint8(pushbp.Link.Pc))
+	buf.writecode(UWOP_PUSH_NONVOL, SEH_REG_BP)
+
+	// The following 4 bytes reference the RVA of the exception handler,
+	// in case the function has one. We don't use it for now.
+	buf.write32(0)
+
+	// The list of unwind infos in a PE binary have very low cardinality
+	// as each info only contains frame pointer operations,
+	// which are very similar across functions.
+	// Dedup them when possible.
+	hash := base64.StdEncoding.EncodeToString(buf.data)
+	symname := fmt.Sprintf("%d.%s", len(buf.data), hash)
+	return ctxt.LookupInit("go:sehuw."+symname, func(s *obj.LSym) {
+		s.WriteBytes(ctxt, 0, buf.data)
+		s.Type = objabi.SSEHUNWINDINFO
+		s.Set(obj.AttrDuplicateOK, true)
+		s.Set(obj.AttrLocal, true)
+		// Note: AttrContentAddressable cannot be set here,
+		// because the content-addressable-handling code
+		// does not know about aux symbols.
+	})
+}

src/cmd/internal/objabi/symkind.go

@@ -72,5 +72,6 @@ const (
 	SCOVERAGE_COUNTER
 	SCOVERAGE_AUXVAR
 
+	SSEHUNWINDINFO
 	// Update cmd/link/internal/sym/AbiSymKindToSymKind for new SymKind values.
 )

src/cmd/link/internal/sym/symkind.go

@@ -123,6 +123,9 @@ const (
 	SDWARFRANGE
 	SDWARFLOC
 	SDWARFLINES
+
+	// SEH symbol types
+	SSEHUNWINDINFO
 )
 
 // AbiSymKindToSymKind maps values read from object files (which are
@@ -148,6 +151,7 @@ var AbiSymKindToSymKind = [...]SymKind{
 	objabi.SLIBFUZZER_8BIT_COUNTER: SLIBFUZZER_8BIT_COUNTER,
 	objabi.SCOVERAGE_COUNTER:       SCOVERAGE_COUNTER,
 	objabi.SCOVERAGE_AUXVAR:        SCOVERAGE_AUXVAR,
+	objabi.SSEHUNWINDINFO:          SSEHUNWINDINFO,
 }
 
 // ReadOnly are the symbol kinds that form read-only sections. In some