From: Michael Pratt <mpratt@google.com>
Date: Fri, 6 Dec 2019 22:25:51 +0000 (-0500)
Subject: cmd/link: generate relative relocs for PIE
X-Git-Tag: go1.15beta1~1145
X-Git-Url: http://www.git.cypherpunks.su/?a=commitdiff_plain;h=675a2e70780208dce18a0bc4b3e420de71fe0178;p=gostls13.git

cmd/link: generate relative relocs for PIE

Go's PIE binaries have tons of relocations, all R_X86_64_64 [1] when
internally linked. R_X86_64_64 relocations require symbol lookup in the
dynamic linker, which can be quite slow. The simple Go HTTP server
in #36028 takes over 1s to complete dynamic linking!

The external linker generates R_X86_64_RELATIVE [2] relocations, which
are significantly more efficient. It turns out that generating these
relocations internally is quite simple, so lets do it.

Rather than referencing targ.Dynid in r_info and having the dynamic
linker do a symbol lookup and then add (final targ address) + r.Add, use
AddAddrPlus to generate another R_ADDR to have the linker compute (targ
address + r.Add). The dynamic linker is then only left with base address
+ r_addend.

Since we don't reference the symbol in the final relocation, Adddynsym
is no longer necessary, saving ~1MB (of ~9MB) from the binary size of
the example in #36028.

[1] R_AARCH64_ABS64 on arm64.
[2] R_AARCH64_RELATIVE on arm64.

Fixes #36028

Change-Id: Ie0f4bb1d911f2e05ab46a0a0454c0f46ac01b43b
Reviewed-on: https://go-review.googlesource.com/c/go/+/210181
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
---

diff --git a/src/cmd/link/internal/amd64/asm.go b/src/cmd/link/internal/amd64/asm.go
index 991f5523ed..74fa8dbb90 100644
--- a/src/cmd/link/internal/amd64/asm.go
+++ b/src/cmd/link/internal/amd64/asm.go
@@ -322,19 +322,31 @@ func adddynrel(ctxt *ld.Link, s *sym.Symbol, r *sym.Reloc) bool {
 		}
 
 		if ctxt.IsELF {
-			// TODO: We generate a R_X86_64_64 relocation for every R_ADDR, even
-			// though it would be more efficient (for the dynamic linker) if we
-			// generated R_X86_RELATIVE instead.
-			ld.Adddynsym(ctxt, targ)
+			// Generate R_X86_64_RELATIVE relocations for best
+			// efficiency in the dynamic linker.
+			//
+			// As noted above, symbol addresses have not been
+			// assigned yet, so we can't generate the final reloc
+			// entry yet. We ultimately want:
+			//
+			// r_offset = s + r.Off
+			// r_info = R_X86_64_RELATIVE
+			// r_addend = targ + r.Add
+			//
+			// The dynamic linker will set *offset = base address +
+			// addend.
+			//
+			// AddAddrPlus is used for r_offset and r_addend to
+			// generate new R_ADDR relocations that will update
+			// these fields in the 'reloc' phase.
 			rela := ctxt.Syms.Lookup(".rela", 0)
 			rela.AddAddrPlus(ctxt.Arch, s, int64(r.Off))
 			if r.Siz == 8 {
-				rela.AddUint64(ctxt.Arch, ld.ELF64_R_INFO(uint32(targ.Dynid), uint32(elf.R_X86_64_64)))
+				rela.AddUint64(ctxt.Arch, ld.ELF64_R_INFO(0, uint32(elf.R_X86_64_RELATIVE)))
 			} else {
-				// TODO: never happens, remove.
-				rela.AddUint64(ctxt.Arch, ld.ELF64_R_INFO(uint32(targ.Dynid), uint32(elf.R_X86_64_32)))
+				ld.Errorf(s, "unexpected relocation for dynamic symbol %s", targ.Name)
 			}
-			rela.AddUint64(ctxt.Arch, uint64(r.Add))
+			rela.AddAddrPlus(ctxt.Arch, targ, int64(r.Add))
 			r.Type = objabi.ElfRelocOffset // ignore during relocsym
 			return true
 		}
diff --git a/src/cmd/link/internal/arm64/asm.go b/src/cmd/link/internal/arm64/asm.go
index 9fccf73a59..690116de01 100644
--- a/src/cmd/link/internal/arm64/asm.go
+++ b/src/cmd/link/internal/arm64/asm.go
@@ -278,18 +278,31 @@ func adddynrel(ctxt *ld.Link, s *sym.Symbol, r *sym.Reloc) bool {
 		}
 
 		if ctxt.IsELF {
-			// TODO: We generate a R_AARCH64_ABS64 relocation for every R_ADDR, even
-			// though it would be more efficient (for the dynamic linker) if we
-			// generated R_AARCH64_RELATIVE instead.
-			ld.Adddynsym(ctxt, targ)
+			// Generate R_AARCH64_RELATIVE relocations for best
+			// efficiency in the dynamic linker.
+			//
+			// As noted above, symbol addresses have not been
+			// assigned yet, so we can't generate the final reloc
+			// entry yet. We ultimately want:
+			//
+			// r_offset = s + r.Off
+			// r_info = R_AARCH64_RELATIVE
+			// r_addend = targ + r.Add
+			//
+			// The dynamic linker will set *offset = base address +
+			// addend.
+			//
+			// AddAddrPlus is used for r_offset and r_addend to
+			// generate new R_ADDR relocations that will update
+			// these fields in the 'reloc' phase.
 			rela := ctxt.Syms.Lookup(".rela", 0)
 			rela.AddAddrPlus(ctxt.Arch, s, int64(r.Off))
 			if r.Siz == 8 {
-				rela.AddUint64(ctxt.Arch, ld.ELF64_R_INFO(uint32(targ.Dynid), uint32(elf.R_AARCH64_ABS64)))
+				rela.AddUint64(ctxt.Arch, ld.ELF64_R_INFO(0, uint32(elf.R_AARCH64_RELATIVE)))
 			} else {
 				ld.Errorf(s, "unexpected relocation for dynamic symbol %s", targ.Name)
 			}
-			rela.AddUint64(ctxt.Arch, uint64(r.Add))
+			rela.AddAddrPlus(ctxt.Arch, targ, int64(r.Add))
 			r.Type = objabi.ElfRelocOffset // ignore during relocsym
 			return true
 		}