[dev.link] cmd/link: directly use loader.ExtReloc in ELF relocation generation

Convert the part that uses relocations to use loader.ExtReloc
directly. It still uses sym.Symbols for now, but not sym.Relocs.

This reduces some memory usage: linking cmd/compile with external
linking,

name             old allocs/op  new allocs/op  delta
Loadlibfull_GC     52.2MB ± 0%    13.9MB ± 0%  -73.40%  (p=0.008 n=5+5)

name             old live-B     new live-B     delta
Loadlibfull_GC      75.5M ± 0%     61.9M ± 0%  -18.02%  (p=0.008 n=5+5)

Change-Id: I317ecbf516063c42b255b2caba310ea6281342d7
Reviewed-on: https://go-review.googlesource.com/c/go/+/231319
Reviewed-by: Jeremy Faller <jeremy@golang.org>
Reviewed-by: Than McIntosh <thanm@google.com>

7 files changed, 144 insertions, 47 deletions

diff --git a/src/cmd/link/internal/amd64/asm.go b/src/cmd/link/internal/amd64/asm.go
index e2c33b8001..c2d54703c1 100644
--- a/src/cmd/link/internal/amd64/asm.go
+++ b/src/cmd/link/internal/amd64/asm.go
@@ -393,36 +393,38 @@ func adddynrel2(target *ld.Target, ldr *loader.Loader, syms *ld.ArchSyms, s load
 	return false
 }
 
-func elfreloc1(ctxt *ld.Link, r *sym.Reloc, sectoff int64) bool {
+func elfreloc2(ctxt *ld.Link, ldr *loader.Loader, s loader.Sym, r loader.ExtRelocView, sectoff int64) bool {
 	ctxt.Out.Write64(uint64(sectoff))
 
-	elfsym := ld.ElfSymForReloc(ctxt, r.Xsym)
-	switch r.Type {
+	xsym := ldr.Syms[r.Xsym]
+	elfsym := ld.ElfSymForReloc(ctxt, xsym)
+	siz := r.Siz()
+	switch r.Type() {
 	default:
 		return false
 	case objabi.R_ADDR, objabi.R_DWARFSECREF:
-		if r.Siz == 4 {
+		if siz == 4 {
 			ctxt.Out.Write64(uint64(elf.R_X86_64_32) | uint64(elfsym)<<32)
-		} else if r.Siz == 8 {
+		} else if siz == 8 {
 			ctxt.Out.Write64(uint64(elf.R_X86_64_64) | uint64(elfsym)<<32)
 		} else {
 			return false
 		}
 	case objabi.R_TLS_LE:
-		if r.Siz == 4 {
+		if siz == 4 {
 			ctxt.Out.Write64(uint64(elf.R_X86_64_TPOFF32) | uint64(elfsym)<<32)
 		} else {
 			return false
 		}
 	case objabi.R_TLS_IE:
-		if r.Siz == 4 {
+		if siz == 4 {
 			ctxt.Out.Write64(uint64(elf.R_X86_64_GOTTPOFF) | uint64(elfsym)<<32)
 		} else {
 			return false
 		}
 	case objabi.R_CALL:
-		if r.Siz == 4 {
-			if r.Xsym.Type == sym.SDYNIMPORT {
+		if siz == 4 {
+			if xsym.Type == sym.SDYNIMPORT {
 				if ctxt.DynlinkingGo() {
 					ctxt.Out.Write64(uint64(elf.R_X86_64_PLT32) | uint64(elfsym)<<32)
 				} else {
@@ -435,8 +437,8 @@ func elfreloc1(ctxt *ld.Link, r *sym.Reloc, sectoff int64) bool {
 			return false
 		}
 	case objabi.R_PCREL:
-		if r.Siz == 4 {
-			if r.Xsym.Type == sym.SDYNIMPORT && r.Xsym.ElfType() == elf.STT_FUNC {
+		if siz == 4 {
+			if xsym.Type == sym.SDYNIMPORT && xsym.ElfType() == elf.STT_FUNC {
 				ctxt.Out.Write64(uint64(elf.R_X86_64_PLT32) | uint64(elfsym)<<32)
 			} else {
 				ctxt.Out.Write64(uint64(elf.R_X86_64_PC32) | uint64(elfsym)<<32)
@@ -445,7 +447,7 @@ func elfreloc1(ctxt *ld.Link, r *sym.Reloc, sectoff int64) bool {
 			return false
 		}
 	case objabi.R_GOTPCREL:
-		if r.Siz == 4 {
+		if siz == 4 {
 			ctxt.Out.Write64(uint64(elf.R_X86_64_GOTPCREL) | uint64(elfsym)<<32)
 		} else {
 			return false
diff --git a/src/cmd/link/internal/amd64/obj.go b/src/cmd/link/internal/amd64/obj.go
index 1fbbf60366..645547cb3e 100644
--- a/src/cmd/link/internal/amd64/obj.go
+++ b/src/cmd/link/internal/amd64/obj.go
@@ -52,7 +52,7 @@ func Init() (*sys.Arch, ld.Arch) {
 		Archrelocvariant: archrelocvariant,
 		Asmb:             asmb,
 		Asmb2:            asmb2,
-		Elfreloc1:        elfreloc1,
+		Elfreloc2:        elfreloc2,
 		Elfsetupplt:      elfsetupplt,
 		Gentext2:         gentext2,
 		Machoreloc1:      machoreloc1,
diff --git a/src/cmd/link/internal/ld/elf.go b/src/cmd/link/internal/ld/elf.go
index 83f100c12b..78298beafe 100644
--- a/src/cmd/link/internal/ld/elf.go
+++ b/src/cmd/link/internal/ld/elf.go
@@ -1374,6 +1374,11 @@ func elfshreloc(arch *sys.Arch, sect *sym.Section) *ElfShdr {
 }
 
 func elfrelocsect(ctxt *Link, sect *sym.Section, syms []*sym.Symbol) {
+	if !ctxt.IsAMD64() {
+		elfrelocsect2(ctxt, sect, syms)
+		return
+	}
+
 	// If main section is SHT_NOBITS, nothing to relocate.
 	// Also nothing to relocate in .shstrtab.
 	if sect.Vaddr >= sect.Seg.Vaddr+sect.Seg.Filelen {
@@ -1394,6 +1399,7 @@ func elfrelocsect(ctxt *Link, sect *sym.Section, syms []*sym.Symbol) {
 		}
 	}
 
+	ldr := ctxt.loader
 	eaddr := int32(sect.Vaddr + sect.Length)
 	for _, s := range syms {
 		if !s.Attr.Reachable() {
@@ -1402,24 +1408,23 @@ func elfrelocsect(ctxt *Link, sect *sym.Section, syms []*sym.Symbol) {
 		if s.Value >= int64(eaddr) {
 			break
 		}
-		for ri := range s.R {
-			r := &s.R[ri]
-			if r.Done {
-				continue
-			}
-			if r.Xsym == nil {
-				Errorf(s, "missing xsym in relocation %#v %#v", r.Sym.Name, s)
+		i := loader.Sym(s.SymIdx)
+		relocs := ldr.ExtRelocs(i)
+		for ri := 0; ri < relocs.Count(); ri++ {
+			r := relocs.At(ri)
+			if r.Xsym == 0 {
+				Errorf(s, "missing xsym in relocation %v", ldr.SymName(r.Sym()))
 				continue
 			}
-			esr := ElfSymForReloc(ctxt, r.Xsym)
+			esr := ElfSymForReloc(ctxt, ldr.Syms[r.Xsym])
 			if esr == 0 {
-				Errorf(s, "reloc %d (%s) to non-elf symbol %s (outer=%s) %d (%s)", r.Type, sym.RelocName(ctxt.Arch, r.Type), r.Sym.Name, r.Xsym.Name, r.Sym.Type, r.Sym.Type)
+				Errorf(s, "reloc %d (%s) to non-elf symbol %s (outer=%s) %d (%s)", r.Type(), sym.RelocName(ctxt.Arch, r.Type()), ldr.Syms[r.Sym()].Name, ldr.Syms[r.Xsym].Name, ldr.Syms[r.Sym()].Type, ldr.Syms[r.Sym()].Type)
 			}
-			if !r.Xsym.Attr.Reachable() {
-				Errorf(s, "unreachable reloc %d (%s) target %v", r.Type, sym.RelocName(ctxt.Arch, r.Type), r.Xsym.Name)
+			if !ldr.AttrReachable(r.Xsym) {
+				Errorf(s, "unreachable reloc %d (%s) target %v", r.Type(), sym.RelocName(ctxt.Arch, r.Type()), ldr.Syms[r.Xsym].Name)
 			}
-			if !thearch.Elfreloc1(ctxt, r, int64(uint64(s.Value+int64(r.Off))-sect.Vaddr)) {
-				Errorf(s, "unsupported obj reloc %d (%s)/%d to %s", r.Type, sym.RelocName(ctxt.Arch, r.Type), r.Siz, r.Sym.Name)
+			if !thearch.Elfreloc2(ctxt, ldr, i, r, int64(uint64(s.Value+int64(r.Off()))-sect.Vaddr)) {
+				Errorf(s, "unsupported obj reloc %d (%s)/%d to %s", r.Type(), sym.RelocName(ctxt.Arch, r.Type()), r.Siz(), ldr.Syms[r.Sym()].Name)
 			}
 		}
 	}
diff --git a/src/cmd/link/internal/ld/elf2.go b/src/cmd/link/internal/ld/elf2.go
index 3f7d72b310..07b64cfcb0 100644
--- a/src/cmd/link/internal/ld/elf2.go
+++ b/src/cmd/link/internal/ld/elf2.go
@@ -23,3 +23,57 @@ func elfsetstring(s *sym.Symbol, str string, off int) {
 	elfstr[nelfstr].off = off
 	nelfstr++
 }
+
+func elfrelocsect2(ctxt *Link, sect *sym.Section, syms []*sym.Symbol) {
+	// If main section is SHT_NOBITS, nothing to relocate.
+	// Also nothing to relocate in .shstrtab.
+	if sect.Vaddr >= sect.Seg.Vaddr+sect.Seg.Filelen {
+		return
+	}
+	if sect.Name == ".shstrtab" {
+		return
+	}
+
+	sect.Reloff = uint64(ctxt.Out.Offset())
+	for i, s := range syms {
+		if !s.Attr.Reachable() {
+			continue
+		}
+		if uint64(s.Value) >= sect.Vaddr {
+			syms = syms[i:]
+			break
+		}
+	}
+
+	eaddr := int32(sect.Vaddr + sect.Length)
+	for _, s := range syms {
+		if !s.Attr.Reachable() {
+			continue
+		}
+		if s.Value >= int64(eaddr) {
+			break
+		}
+		for ri := range s.R {
+			r := &s.R[ri]
+			if r.Done {
+				continue
+			}
+			if r.Xsym == nil {
+				Errorf(s, "missing xsym in relocation %#v %#v", r.Sym.Name, s)
+				continue
+			}
+			esr := ElfSymForReloc(ctxt, r.Xsym)
+			if esr == 0 {
+				Errorf(s, "reloc %d (%s) to non-elf symbol %s (outer=%s) %d (%s)", r.Type, sym.RelocName(ctxt.Arch, r.Type), r.Sym.Name, r.Xsym.Name, r.Sym.Type, r.Sym.Type)
+			}
+			if !r.Xsym.Attr.Reachable() {
+				Errorf(s, "unreachable reloc %d (%s) target %v", r.Type, sym.RelocName(ctxt.Arch, r.Type), r.Xsym.Name)
+			}
+			if !thearch.Elfreloc1(ctxt, r, int64(uint64(s.Value+int64(r.Off))-sect.Vaddr)) {
+				Errorf(s, "unsupported obj reloc %d (%s)/%d to %s", r.Type, sym.RelocName(ctxt.Arch, r.Type), r.Siz, r.Sym.Name)
+			}
+		}
+	}
+
+	sect.Rellen = uint64(ctxt.Out.Offset()) - sect.Reloff
+}
diff --git a/src/cmd/link/internal/ld/lib.go b/src/cmd/link/internal/ld/lib.go
index a328efd03b..3c60901124 100644
--- a/src/cmd/link/internal/ld/lib.go
+++ b/src/cmd/link/internal/ld/lib.go
@@ -270,6 +270,7 @@ type Arch struct {
 	Asmb2 func(*Link)
 
 	Elfreloc1   func(*Link, *sym.Reloc, int64) bool
+	Elfreloc2   func(*Link, *loader.Loader, loader.Sym, loader.ExtRelocView, int64) bool
 	Elfsetupplt func(ctxt *Link, plt, gotplt *loader.SymbolBuilder, dynamic loader.Sym)
 	Gentext     func(*Link)
 	Gentext2    func(*Link, *loader.Loader)
@@ -2825,9 +2826,9 @@ func addToTextp(ctxt *Link) {
 	ctxt.Textp = textp
 }
 
-func (ctxt *Link) loadlibfull(symGroupType []sym.SymKind, needReloc bool) {
+func (ctxt *Link) loadlibfull(symGroupType []sym.SymKind, needReloc, needExtReloc bool) {
 	// Load full symbol contents, resolve indexed references.
-	ctxt.loader.LoadFull(ctxt.Arch, ctxt.Syms, needReloc)
+	ctxt.loader.LoadFull(ctxt.Arch, ctxt.Syms, needReloc, needExtReloc)
 
 	// Convert ctxt.Moduledata2 to ctxt.Moduledata, etc
 	if ctxt.Moduledata2 != 0 {
diff --git a/src/cmd/link/internal/ld/main.go b/src/cmd/link/internal/ld/main.go
index 6bd6a8e467..1ed8ccb828 100644
--- a/src/cmd/link/internal/ld/main.go
+++ b/src/cmd/link/internal/ld/main.go
@@ -335,10 +335,13 @@ func Main(arch *sys.Arch, theArch Arch) {
 		// An exception is internal linking on Windows, see pe.go:addPEBaseRelocSym
 		// Wasm is another exception, where it applies text relocations in Asmb2.
 		needReloc := (ctxt.IsWindows() && ctxt.IsInternal()) || ctxt.IsWasm()
-		ctxt.loadlibfull(symGroupType, needReloc) // XXX do it here for now
+		// On AMD64 ELF, we directly use the loader's ExtRelocs, so we don't
+		// need conversion. Otherwise we do.
+		needExtReloc := ctxt.IsExternal() && !(ctxt.IsAMD64() && ctxt.IsELF)
+		ctxt.loadlibfull(symGroupType, needReloc, needExtReloc) // XXX do it here for now
 	} else {
 		bench.Start("loadlibfull")
-		ctxt.loadlibfull(symGroupType, true) // XXX do it here for now
+		ctxt.loadlibfull(symGroupType, true, false) // XXX do it here for now
 		bench.Start("reloc")
 		ctxt.reloc2()
 	}
diff --git a/src/cmd/link/internal/loader/loader.go b/src/cmd/link/internal/loader/loader.go
index f4d91644c9..2627218ced 100644
--- a/src/cmd/link/internal/loader/loader.go
+++ b/src/cmd/link/internal/loader/loader.go
@@ -56,6 +56,14 @@ type ExtReloc struct {
 	Xadd int64
 }
 
+// ExtRelocView is a view of an external relocation.
+// It is intended to be constructed on the fly, such as ExtRelocs.At.
+// It is not the data structure used to store the payload internally.
+type ExtRelocView struct {
+	Reloc2
+	*ExtReloc
+}
+
 // Reloc2 holds a "handle" to access a relocation record from an
 // object file.
 type Reloc2 struct {
@@ -1116,7 +1124,7 @@ func (l *Loader) InitOutData() {
 	l.outdata = make([][]byte, l.extStart)
 }
 
-// SetExtRelocs sets the section of the i-th symbol. i is global index.
+// SetExtRelocs sets the external relocations of the i-th symbol. i is global index.
 func (l *Loader) SetExtRelocs(i Sym, relocs []ExtReloc) {
 	l.extRelocs[i] = relocs
 }
@@ -1699,6 +1707,24 @@ func (l *Loader) relocs(r *oReader, li int) Relocs {
 	}
 }
 
+// ExtRelocs returns the external relocations of the i-th symbol.
+func (l *Loader) ExtRelocs(i Sym) ExtRelocs {
+	return ExtRelocs{l.Relocs(i), l.extRelocs[i]}
+}
+
+// ExtRelocs represents the set of external relocations of a symbol.
+type ExtRelocs struct {
+	rs Relocs
+	es []ExtReloc
+}
+
+func (ers ExtRelocs) Count() int { return len(ers.es) }
+
+func (ers ExtRelocs) At(j int) ExtRelocView {
+	i := ers.es[j].Idx
+	return ExtRelocView{ers.rs.At2(i), &ers.es[j]}
+}
+
 // RelocByOff implements sort.Interface for sorting relocations by offset.
 
 type RelocByOff []Reloc
@@ -2033,7 +2059,7 @@ func (l *Loader) preprocess(arch *sys.Arch, s Sym, name string) {
 }
 
 // Load full contents.
-func (l *Loader) LoadFull(arch *sys.Arch, syms *sym.Symbols, needReloc bool) {
+func (l *Loader) LoadFull(arch *sys.Arch, syms *sym.Symbols, needReloc, needExtReloc bool) {
 	// create all Symbols first.
 	l.growSyms(l.NSym())
 	l.growSects(l.NSym())
@@ -2047,7 +2073,7 @@ func (l *Loader) LoadFull(arch *sys.Arch, syms *sym.Symbols, needReloc bool) {
 
 	nr := 0 // total number of sym.Reloc's we'll need
 	for _, o := range l.objs[1:] {
-		nr += loadObjSyms(l, syms, o.r, needReloc)
+		nr += loadObjSyms(l, syms, o.r, needReloc, needExtReloc)
 	}
 
 	// Make a first pass through the external symbols, making
@@ -2063,7 +2089,7 @@ func (l *Loader) LoadFull(arch *sys.Arch, syms *sym.Symbols, needReloc bool) {
 		if needReloc {
 			nr += len(pp.relocs)
 		}
-		if int(i) < len(l.extRelocs) {
+		if needExtReloc && int(i) < len(l.extRelocs) {
 			nr += len(l.extRelocs[i])
 		}
 		// create and install the sym.Symbol here so that l.Syms will
@@ -2079,7 +2105,7 @@ func (l *Loader) LoadFull(arch *sys.Arch, syms *sym.Symbols, needReloc bool) {
 	// allocate a single large slab of relocations for all live symbols
 	if nr != 0 {
 		l.relocBatch = make([]sym.Reloc, nr)
-		if len(l.extRelocs) != 0 {
+		if needExtReloc {
 			l.relocExtBatch = make([]sym.RelocExt, nr)
 		}
 	}
@@ -2102,8 +2128,9 @@ func (l *Loader) LoadFull(arch *sys.Arch, syms *sym.Symbols, needReloc bool) {
 			relocs := l.Relocs(i)
 			l.convertRelocations(i, &relocs, s, false)
 		}
-
-		l.convertExtRelocs(s, i)
+		if needExtReloc {
+			l.convertExtRelocs(s, i)
+		}
 
 		// Copy data
 		s.P = pp.data
@@ -2114,7 +2141,7 @@ func (l *Loader) LoadFull(arch *sys.Arch, syms *sym.Symbols, needReloc bool) {
 
 	// load contents of defined symbols
 	for _, o := range l.objs[1:] {
-		loadObjFull(l, o.r, needReloc)
+		loadObjFull(l, o.r, needReloc, needExtReloc)
 	}
 
 	// Sanity check: we should have consumed all batched allocations.
@@ -2170,17 +2197,21 @@ func (l *Loader) LoadFull(arch *sys.Arch, syms *sym.Symbols, needReloc bool) {
 	l.plt = nil
 	l.got = nil
 	l.dynid = nil
-	l.relocVariant = nil
-	l.extRelocs = nil
+	if needExtReloc { // converted to sym.Relocs, drop loader references
+		l.relocVariant = nil
+		l.extRelocs = nil
+	}
 
 	// Drop fields that are no longer needed.
 	for _, i := range l.extReader.syms {
 		pp := l.getPayload(i)
 		pp.name = ""
-		pp.relocs = nil
-		pp.reltypes = nil
 		pp.auxs = nil
 		pp.data = nil
+		if needExtReloc {
+			pp.relocs = nil
+			pp.reltypes = nil
+		}
 	}
 }
 
@@ -2450,7 +2481,7 @@ func topLevelSym(sname string, skind sym.SymKind) bool {
 // loadObjSyms creates sym.Symbol objects for the live Syms in the
 // object corresponding to object reader "r". Return value is the
 // number of sym.Reloc entries required for all the new symbols.
-func loadObjSyms(l *Loader, syms *sym.Symbols, r *oReader, needReloc bool) int {
+func loadObjSyms(l *Loader, syms *sym.Symbols, r *oReader, needReloc, needExtReloc bool) int {
 	nr := 0
 	for i, n := 0, r.NSym()+r.NNonpkgdef(); i < n; i++ {
 		gi := r.syms[i]
@@ -2483,7 +2514,7 @@ func loadObjSyms(l *Loader, syms *sym.Symbols, r *oReader, needReloc bool) int {
 		if needReloc {
 			nr += r.NReloc(i)
 		}
-		if int(gi) < len(l.extRelocs) {
+		if needExtReloc && int(gi) < len(l.extRelocs) {
 			nr += len(l.extRelocs[gi])
 		}
 	}
@@ -2690,7 +2721,7 @@ func (l *Loader) FreeSym(i Sym) {
 	}
 }
 
-func loadObjFull(l *Loader, r *oReader, needReloc bool) {
+func loadObjFull(l *Loader, r *oReader, needReloc, needExtReloc bool) {
 	for i, n := 0, r.NSym()+r.NNonpkgdef(); i < n; i++ {
 		// A symbol may be a dup or overwritten. In this case, its
 		// content will actually be provided by a different object
@@ -2722,8 +2753,9 @@ func loadObjFull(l *Loader, r *oReader, needReloc bool) {
 			l.relocBatch = batch[relocs.Count():]
 			l.convertRelocations(gi, &relocs, s, false)
 		}
-
-		l.convertExtRelocs(s, gi)
+		if needExtReloc {
+			l.convertExtRelocs(s, gi)
+		}
 
 		// Aux symbol info
 		auxs := r.Auxs(i)