cmd/internal/obj, cmd/link, runtime: a saner model for TLS on arm

this leaves lots of cruft behind, will delete that soon

Change-Id: I12d6b6192f89bcdd89b2b0873774bd3458373b8a
Reviewed-on: https://go-review.googlesource.com/14196Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

src/cmd/internal/obj/arm/a.out.go

@@ -155,6 +155,16 @@ const (
 	C_HREG
 
 	C_ADDR /* reference to relocatable address */
+
+	// TLS "var" in local exec mode: will become a constant offset from
+	// thread local base that is ultimately chosen by the program linker.
+	C_TLS_LE
+
+	// TLS "var" in initial exec mode: will become a memory address (chosen
+	// by the program linker) that the dynamic linker will fill with the
+	// offset from the thread local base.
+	C_TLS_IE
+
 	C_TEXTSIZE
 
 	C_GOK

src/cmd/internal/obj/arm/anames5.go

@@ -41,6 +41,8 @@ var cnames5 = []string{
 	"SP",
 	"HREG",
 	"ADDR",
+	"C_TLS_LE",
+	"C_TLS_IE",
 	"TEXTSIZE",
 	"GOK",
 	"NCLASS",

src/cmd/internal/obj/arm/asm5.go

@@ -104,6 +104,8 @@ var optab = []Optab{
 	Optab{AWORD, C_NONE, C_NONE, C_LCON, 11, 4, 0, 0, 0},
 	Optab{AWORD, C_NONE, C_NONE, C_LCONADDR, 11, 4, 0, 0, 0},
 	Optab{AWORD, C_NONE, C_NONE, C_ADDR, 11, 4, 0, 0, 0},
+	Optab{AWORD, C_NONE, C_NONE, C_TLS_LE, 103, 4, 0, 0, 0},
+	Optab{AWORD, C_NONE, C_NONE, C_TLS_IE, 104, 4, 0, 0, 0},
 	Optab{AMOVW, C_NCON, C_NONE, C_REG, 12, 4, 0, 0, 0},
 	Optab{AMOVW, C_LCON, C_NONE, C_REG, 12, 4, 0, LFROM, 0},
 	Optab{AMOVW, C_LCONADDR, C_NONE, C_REG, 12, 4, 0, LFROM | LPCREL, 4},
@@ -151,6 +153,8 @@ var optab = []Optab{
 	Optab{AMOVBU, C_REG, C_NONE, C_LAUTO, 30, 8, REGSP, LTO, 0},
 	Optab{AMOVBU, C_REG, C_NONE, C_LOREG, 30, 8, 0, LTO, 0},
 	Optab{AMOVBU, C_REG, C_NONE, C_ADDR, 64, 8, 0, LTO | LPCREL, 4},
+	Optab{AMOVW, C_TLS_LE, C_NONE, C_REG, 101, 4, 0, LFROM, 0},
+	Optab{AMOVW, C_TLS_IE, C_NONE, C_REG, 102, 8, 0, LFROM, 0},
 	Optab{AMOVW, C_LAUTO, C_NONE, C_REG, 31, 8, REGSP, LFROM, 0},
 	Optab{AMOVW, C_LOREG, C_NONE, C_REG, 31, 8, 0, LFROM, 0},
 	Optab{AMOVW, C_ADDR, C_NONE, C_REG, 65, 8, 0, LFROM | LPCREL, 4},
@@ -1016,6 +1020,14 @@ func aclass(ctxt *obj.Link, a *obj.Addr) int {
 			}
 
 			ctxt.Instoffset = 0 // s.b. unused but just in case
+			if a.Sym.Type == obj.STLSBSS {
+				if ctxt.Flag_shared != 0 {
+					return C_TLS_IE
+				} else {
+					return C_TLS_LE
+				}
+			}
+
 			return C_ADDR
 
 		case obj.NAME_AUTO:
@@ -2037,6 +2049,50 @@ func asmout(ctxt *obj.Link, p *obj.Prog, o *Optab, out []uint32) {
 			o2 = oprrr(ctxt, AADD, int(p.Scond)) | REGTMP&15 | (REGPC&15)<<16 | (REGTMP&15)<<12
 		}
 
+	case 101: /* movw tlsvar,R, local exec*/
+		if p.Scond&C_SCOND != C_SCOND_NONE {
+			ctxt.Diag("conditional tls")
+		}
+		o1 = omvl(ctxt, p, &p.From, int(p.To.Reg))
+
+	case 102: /* movw tlsvar,R, initial exec*/
+		if p.Scond&C_SCOND != C_SCOND_NONE {
+			ctxt.Diag("conditional tls")
+		}
+		o1 = omvl(ctxt, p, &p.From, int(p.To.Reg))
+		o2 = olrr(ctxt, int(p.To.Reg)&15, (REGPC & 15), int(p.To.Reg), int(p.Scond))
+
+	case 103: /* word tlsvar, local exec */
+		if p.To.Sym == nil {
+			ctxt.Diag("nil sym in tls %v", p)
+		}
+		if p.To.Offset != 0 {
+			ctxt.Diag("offset against tls var in %v", p)
+		}
+		// This case happens with words generated in the PC stream as part of
+		// the literal pool.
+		rel := obj.Addrel(ctxt.Cursym)
+
+		rel.Off = int32(ctxt.Pc)
+		rel.Siz = 4
+		rel.Sym = p.To.Sym
+		rel.Type = obj.R_TLS_LE
+		o1 = 0
+
+	case 104: /* word tlsvar, initial exec */
+		if p.To.Sym == nil {
+			ctxt.Diag("nil sym in tls %v", p)
+		}
+		if p.To.Offset != 0 {
+			ctxt.Diag("offset against tls var in %v", p)
+		}
+		rel := obj.Addrel(ctxt.Cursym)
+		rel.Off = int32(ctxt.Pc)
+		rel.Siz = 4
+		rel.Sym = p.To.Sym
+		rel.Type = obj.R_TLS_IE
+		rel.Add = ctxt.Pc - p.Rel.Pc - 8 - int64(rel.Siz)
+
 	case 68: /* floating point store -> ADDR */
 		o1 = omvl(ctxt, p, &p.To, REGTMP)
 

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

@@ -233,6 +233,12 @@ func elfreloc1(r *ld.Reloc, sectoff int64) int {
 		} else {
 			return -1
 		}
+
+	case obj.R_TLS_LE:
+		ld.Thearch.Lput(ld.R_ARM_TLS_LE32 | uint32(elfsym)<<8)
+
+	case obj.R_TLS_IE:
+		ld.Thearch.Lput(ld.R_ARM_TLS_IE32 | uint32(elfsym)<<8)
 	}
 
 	return 0

src/cmd/link/internal/ld/data.go

@@ -395,15 +395,7 @@ func relocsym(s *LSym) {
 				break
 			}
 			if Linkmode == LinkInternal && Iself && Thearch.Thechar == '5' {
-				// On ELF ARM, the thread pointer is 8 bytes before
-				// the start of the thread-local data block, so add 8
-				// to the actual TLS offset (r->sym->value).
-				// This 8 seems to be a fundamental constant of
-				// ELF on ARM (or maybe Glibc on ARM); it is not
-				// related to the fact that our own TLS storage happens
-				// to take up 8 bytes.
-				o = 8 + r.Sym.Value
-
+				panic("should no longer get here")
 				break
 			}
 
@@ -416,8 +408,10 @@ func relocsym(s *LSym) {
 		case obj.R_TLS_LE:
 			if Linkmode == LinkExternal && Iself && HEADTYPE != obj.Hopenbsd {
 				r.Done = 0
-				r.Sym = Ctxt.Tlsg
-				r.Xsym = Ctxt.Tlsg
+				if r.Sym == nil {
+					r.Sym = Ctxt.Tlsg
+				}
+				r.Xsym = r.Sym
 				r.Xadd = r.Add
 				o = 0
 				if Thearch.Thechar != '6' {
@@ -426,7 +420,16 @@ func relocsym(s *LSym) {
 				break
 			}
 
-			if Iself || Ctxt.Headtype == obj.Hplan9 || Ctxt.Headtype == obj.Hdarwin {
+			if Iself && Thearch.Thechar == '5' {
+				// On ELF ARM, the thread pointer is 8 bytes before
+				// the start of the thread-local data block, so add 8
+				// to the actual TLS offset (r->sym->value).
+				// This 8 seems to be a fundamental constant of
+				// ELF on ARM (or maybe Glibc on ARM); it is not
+				// related to the fact that our own TLS storage happens
+				// to take up 8 bytes.
+				o = 8 + r.Sym.Value
+			} else if Iself || Ctxt.Headtype == obj.Hplan9 || Ctxt.Headtype == obj.Hdarwin {
 				o = int64(Ctxt.Tlsoffset) + r.Add
 			} else if Ctxt.Headtype == obj.Hwindows {
 				o = r.Add
@@ -437,8 +440,10 @@ func relocsym(s *LSym) {
 		case obj.R_TLS_IE:
 			if Linkmode == LinkExternal && Iself && HEADTYPE != obj.Hopenbsd {
 				r.Done = 0
-				r.Sym = Ctxt.Tlsg
-				r.Xsym = Ctxt.Tlsg
+				if r.Sym == nil {
+					r.Sym = Ctxt.Tlsg
+				}
+				r.Xsym = r.Sym
 				r.Xadd = r.Add
 				o = 0
 				if Thearch.Thechar != '6' {

src/runtime/tls_arm.s

@@ -15,7 +15,7 @@
 // Note: both functions will clobber R0 and R11 and
 // can be called from 5c ABI code.
 
-// On android and darwin, runtime.tlsg is a normal variable.
+// On android and darwin, runtime.tls_g is a normal variable.
 // TLS offset is computed in x_cgo_inittls.
 #ifdef GOOS_android
 #define TLSG_IS_VARIABLE
@@ -41,14 +41,7 @@ TEXT runtime·save_g(SB),NOSPLIT,$-4
 	// The replacement function saves LR in R11 over the call to read_tls_fallback.
 	MRC	15, 0, R0, C13, C0, 3 // fetch TLS base pointer
 	BIC $3, R0 // Darwin/ARM might return unaligned pointer
-	// $runtime.tlsg(SB) is a special linker symbol.
-	// It is the offset from the TLS base pointer to our
-	// thread-local storage for g.
-#ifdef TLSG_IS_VARIABLE
-	MOVW	runtime·tlsg(SB), R11
-#else
-	MOVW	$runtime·tlsg(SB), R11
-#endif
+	MOVW	runtime·tls_g(SB), R11
 	ADD	R11, R0
 	MOVW	g, 0(R0)
 	MOVW	g, R0 // preserve R0 across call to setg<>
@@ -68,11 +61,7 @@ TEXT runtime·load_g(SB),NOSPLIT,$0
 	// $runtime.tlsg(SB) is a special linker symbol.
 	// It is the offset from the TLS base pointer to our
 	// thread-local storage for g.
-#ifdef TLSG_IS_VARIABLE
-	MOVW	runtime·tlsg(SB), R11
-#else
-	MOVW	$runtime·tlsg(SB), R11
-#endif
+	MOVW	runtime·tls_g(SB), R11
 	ADD	R11, R0
 	MOVW	0(R0), g
 	RET
@@ -95,7 +84,11 @@ TEXT runtime·_initcgo(SB),NOSPLIT,$4
 	B.EQ	nocgo
 	MRC     15, 0, R0, C13, C0, 3 	// load TLS base pointer
 	MOVW 	R0, R3 			// arg 3: TLS base pointer
-	MOVW 	$runtime·tlsg(SB), R2 	// arg 2: tlsg
+#ifdef TLSG_IS_VARIABLE
+	MOVW 	$runtime·tls_g(SB), R2 	// arg 2: &tls_g
+#else
+        MOVW	$0, R2			// arg 2: not used when using platform tls
+#endif
 	MOVW	$setg_gcc<>(SB), R1 	// arg 1: setg
 	MOVW	g, R0 			// arg 0: G
 	BL	(R4) // will clobber R0-R3
@@ -109,5 +102,7 @@ TEXT setg_gcc<>(SB),NOSPLIT,$0
 	B		runtime·save_g(SB)
 
 #ifdef TLSG_IS_VARIABLE
-GLOBL runtime·tlsg+0(SB), NOPTR, $4
+GLOBL runtime·tls_g+0(SB), NOPTR, $4
+#else
+GLOBL runtime·tls_g+0(SB), TLSBSS, $4
 #endif