cmd/compile, cmd/link: enable Duff's device on darwin/arm64

Duff's device was disabled on darwin/arm64 because the darwin
linker couldn't handle a branch relocation with non-zero addend.
This is no longer the case now. The darwin linker can handle it
just fine. So enable it.

Fixes golang#54189.

Change-Id: Ida7ebafe6eb01db1af5bb8ae60a62491da5eabdf
Reviewed-on: https://go-review.googlesource.com/c/go/+/420894
Reviewed-by: Eric Fang <eric.fang@arm.com>
Reviewed-by: Than McIntosh <thanm@google.com>
Run-TryBot: Cherry Mui <cherryyz@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>

src/cmd/compile/internal/arm64/ggen.go

@@ -10,11 +10,8 @@ import (
 	"cmd/compile/internal/types"
 	"cmd/internal/obj"
 	"cmd/internal/obj/arm64"
-	"internal/buildcfg"
 )
 
-var darwin = buildcfg.GOOS == "darwin" || buildcfg.GOOS == "ios"
-
 func padframe(frame int64) int64 {
 	// arm64 requires that the frame size (not counting saved FP&LR)
 	// be 16 bytes aligned. If not, pad it.
@@ -32,7 +29,7 @@ func zerorange(pp *objw.Progs, p *obj.Prog, off, cnt int64, _ *uint32) *obj.Prog
 		for i := int64(0); i < cnt; i += int64(types.PtrSize) {
 			p = pp.Append(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGZERO, 0, obj.TYPE_MEM, arm64.REGSP, 8+off+i)
 		}
-	} else if cnt <= int64(128*types.PtrSize) && !darwin { // darwin ld64 cannot handle BR26 reloc with non-zero addend
+	} else if cnt <= int64(128*types.PtrSize) {
 		if cnt%(2*int64(types.PtrSize)) != 0 {
 			p = pp.Append(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGZERO, 0, obj.TYPE_MEM, arm64.REGSP, 8+off)
 			off += int64(types.PtrSize)

src/cmd/compile/internal/ssa/config.go

@@ -230,7 +230,6 @@ func NewConfig(arch string, types Types, ctxt *obj.Link, optimize, softfloat boo
 		c.FPReg = framepointerRegARM64
 		c.LinkReg = linkRegARM64
 		c.hasGReg = true
-		c.noDuffDevice = buildcfg.GOOS == "darwin" || buildcfg.GOOS == "ios" // darwin linker cannot handle BR26 reloc with non-zero addend
 	case "ppc64":
 		c.BigEndian = true
 		fallthrough

src/cmd/link/internal/arm64/asm.go

@@ -539,7 +539,8 @@ func machoreloc1(arch *sys.Arch, out *ld.OutBuf, ldr *loader.Loader, s loader.Sy
 		v |= ld.MACHO_ARM64_RELOC_UNSIGNED << 28
 	case objabi.R_CALLARM64:
 		if xadd != 0 {
-			ldr.Errorf(s, "ld64 doesn't allow BR26 reloc with non-zero addend: %s+%d", ldr.SymName(rs), xadd)
+			out.Write32(uint32(sectoff))
+			out.Write32((ld.MACHO_ARM64_RELOC_ADDEND << 28) | (2 << 25) | uint32(xadd&0xffffff))
 		}
 
 		v |= 1 << 24 // pc-relative bit
@@ -719,15 +720,22 @@ func archreloc(target *ld.Target, ldr *loader.Loader, syms *ld.ArchSyms, r loade
 			}
 
 			return val, nExtReloc, isOk
-		case objabi.R_CALLARM64,
-			objabi.R_ARM64_TLS_LE,
-			objabi.R_ARM64_TLS_IE:
+
+		case objabi.R_CALLARM64:
 			nExtReloc = 1
-			if rt == objabi.R_ARM64_TLS_IE {
-				nExtReloc = 2 // need two ELF relocations. see elfreloc1
+			if target.IsDarwin() && r.Add() != 0 {
+				nExtReloc = 2 // need another relocation for addend
 			}
 			return val, nExtReloc, isOk
 
+		case objabi.R_ARM64_TLS_LE:
+			nExtReloc = 1
+			return val, nExtReloc, isOk
+
+		case objabi.R_ARM64_TLS_IE:
+			nExtReloc = 2 // need two ELF relocations. see elfreloc1
+			return val, nExtReloc, isOk
+
 		case objabi.R_ADDR:
 			if target.IsWindows() && r.Add() != 0 {
 				if r.Siz() == 8 {
@@ -946,20 +954,6 @@ func extreloc(target *ld.Target, ldr *loader.Loader, r loader.Reloc, s loader.Sy
 	case objabi.R_ARM64_GOTPCREL,
 		objabi.R_ADDRARM64:
 		rr := ld.ExtrelocViaOuterSym(ldr, r, s)
-
-		// Note: ld64 currently has a bug that any non-zero addend for BR26 relocation
-		// will make the linking fail because it thinks the code is not PIC even though
-		// the BR26 relocation should be fully resolved at link time.
-		// That is the reason why the next if block is disabled. When the bug in ld64
-		// is fixed, we can enable this block and also enable duff's device in cmd/7g.
-		if false && target.IsDarwin() {
-			// Mach-O wants the addend to be encoded in the instruction
-			// Note that although Mach-O supports ARM64_RELOC_ADDEND, it
-			// can only encode 24-bit of signed addend, but the instructions
-			// supports 33-bit of signed addend, so we always encode the
-			// addend in place.
-			rr.Xadd = 0
-		}
 		return rr, true
 	case objabi.R_CALLARM64,
 		objabi.R_ARM64_TLS_LE,