cmd/internal/obj, runtime: implement IE model TLS on ppc64le

This requires changing the tls access code to match the patterns documented in
the ABI documentation or the system linker will "optimize" it into ridiculousness.

With this change, -buildmode=pie works, although as it is tested in testshared,
the tests are not run yet.

Change-Id: I1efa6687af0a5b8db3385b10f6542a49056b2eb3
Reviewed-on: https://go-review.googlesource.com/15971Reviewed-by: Russ Cox <rsc@golang.org>

src/cmd/internal/obj/link.go

@@ -480,6 +480,19 @@ const (
 	// instruction word.
 	R_POWER_TLS_LE
 
+	// R_POWER_TLS_IE is used to implement the "initial exec" model for tls access. It
+	// relocates a D-form, DS-form instruction sequence like R_ADDRPOWER_DS. It
+	// inserts to the offset of GOT slot for the thread-local symbol from the TOC (the
+	// GOT slot is filled by the dynamic linker with the offset of the thread-local
+	// symbol from the thread pointer (R13)).
+	R_POWER_TLS_IE
+
+	// R_POWER_TLS marks an X-form instruction such as "MOVD 0(R13)(R31*1), g" as
+	// accessing a particular thread-local symbol. It does not affect code generation
+	// but is used by the system linker when relaxing "initial exec" model code to
+	// "local exec" model code.
+	R_POWER_TLS
+
 	// R_ADDRPOWER_DS is similar to R_ADDRPOWER above, but assumes the second
 	// instruction is a "DS-form" instruction, which has an immediate field occupying
 	// bits [15:2] of the instruction word. Bits [15:2] of the address of the

src/cmd/internal/obj/ppc64/a.out.go

@@ -222,6 +222,7 @@ const (
 	C_GOK
 	C_ADDR
 	C_TLS_LE
+	C_TLS_IE
 	C_TEXTSIZE
 
 	C_NCLASS /* must be the last */

src/cmd/internal/obj/ppc64/anames9.go

@@ -36,6 +36,7 @@ var cnames9 = []string{
 	"GOK",
 	"ADDR",
 	"TLS_LE",
+	"TLS_IE",
 	"TEXTSIZE",
 	"NCLASS",
 }

src/cmd/internal/obj/ppc64/asm9.go

@@ -246,6 +246,7 @@ var optab = []Optab{
 	{AMOVB, C_ADDR, C_NONE, C_NONE, C_REG, 76, 12, 0},
 
 	{AMOVD, C_TLS_LE, C_NONE, C_NONE, C_REG, 79, 4, 0},
+	{AMOVD, C_TLS_IE, C_NONE, C_NONE, C_REG, 80, 8, 0},
 
 	/* load constant */
 	{AMOVD, C_SECON, C_NONE, C_NONE, C_REG, 3, 4, REGSB},
@@ -587,7 +588,11 @@ func aclass(ctxt *obj.Link, a *obj.Addr) int {
 			ctxt.Instoffset = a.Offset
 			if a.Sym != nil { // use relocation
 				if a.Sym.Type == obj.STLSBSS {
-					return C_TLS_LE
+					if ctxt.Flag_shared != 0 {
+						return C_TLS_IE
+					} else {
+						return C_TLS_LE
+					}
 				}
 				return C_ADDR
 			}
@@ -1652,6 +1657,18 @@ func asmout(ctxt *obj.Link, p *obj.Prog, o *Optab, out []uint32) {
 			if v != 0 {
 				ctxt.Diag("illegal indexed instruction\n%v", p)
 			}
+			if ctxt.Flag_shared != 0 && r == REG_R13 {
+				rel := obj.Addrel(ctxt.Cursym)
+				rel.Off = int32(ctxt.Pc)
+				rel.Siz = 4
+				// This (and the matching part in the load case
+				// below) are the only places in the ppc64 toolchain
+				// that knows the name of the tls variable. Possibly
+				// we could add some assembly syntax so that the name
+				// of the variable does not have to be assumed.
+				rel.Sym = obj.Linklookup(ctxt, "runtime.tls_g", 0)
+				rel.Type = obj.R_POWER_TLS
+			}
 			o1 = AOP_RRR(uint32(opstorex(ctxt, int(p.As))), uint32(p.From.Reg), uint32(p.To.Index), uint32(r))
 		} else {
 			if int32(int16(v)) != v {
@@ -1671,6 +1688,13 @@ func asmout(ctxt *obj.Link, p *obj.Prog, o *Optab, out []uint32) {
 			if v != 0 {
 				ctxt.Diag("illegal indexed instruction\n%v", p)
 			}
+			if ctxt.Flag_shared != 0 && r == REG_R13 {
+				rel := obj.Addrel(ctxt.Cursym)
+				rel.Off = int32(ctxt.Pc)
+				rel.Siz = 4
+				rel.Sym = obj.Linklookup(ctxt, "runtime.tls_g", 0)
+				rel.Type = obj.R_POWER_TLS
+			}
 			o1 = AOP_RRR(uint32(oploadx(ctxt, int(p.As))), uint32(p.To.Reg), uint32(p.From.Index), uint32(r))
 		} else {
 			if int32(int16(v)) != v {
@@ -2467,6 +2491,18 @@ func asmout(ctxt *obj.Link, p *obj.Prog, o *Optab, out []uint32) {
 		rel.Sym = p.From.Sym
 		rel.Type = obj.R_POWER_TLS_LE
 
+	case 80:
+		if p.From.Offset != 0 {
+			ctxt.Diag("invalid offset against tls var %v", p)
+		}
+		o1 = AOP_IRR(OP_ADDIS, uint32(p.To.Reg), REG_R2, 0)
+		o2 = AOP_IRR(uint32(opload(ctxt, AMOVD)), uint32(p.To.Reg), uint32(p.To.Reg), 0)
+		rel := obj.Addrel(ctxt.Cursym)
+		rel.Off = int32(ctxt.Pc)
+		rel.Siz = 8
+		rel.Sym = p.From.Sym
+		rel.Type = obj.R_POWER_TLS_IE
+
 	}
 
 	out[0] = o1

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

@@ -564,23 +564,28 @@ const (
 	R_PPC_EMB_BIT_FLD     = 115
 	R_PPC_EMB_RELSDA      = 116
 
-	R_PPC64_ADDR32       = R_PPC_ADDR32
-	R_PPC64_ADDR16_LO    = R_PPC_ADDR16_LO
-	R_PPC64_ADDR16_HA    = R_PPC_ADDR16_HA
-	R_PPC64_REL24        = R_PPC_REL24
-	R_PPC64_JMP_SLOT     = R_PPC_JMP_SLOT
-	R_PPC64_TPREL16      = R_PPC_TPREL16
-	R_PPC64_ADDR64       = 38
-	R_PPC64_TOC16        = 47
-	R_PPC64_TOC16_LO     = 48
-	R_PPC64_TOC16_HI     = 49
-	R_PPC64_TOC16_HA     = 50
-	R_PPC64_ADDR16_LO_DS = 57
-	R_PPC64_TOC16_DS     = 63
-	R_PPC64_TOC16_LO_DS  = 64
-	R_PPC64_REL16_LO     = 250
-	R_PPC64_REL16_HI     = 251
-	R_PPC64_REL16_HA     = 252
+	R_PPC64_ADDR32            = R_PPC_ADDR32
+	R_PPC64_ADDR16_LO         = R_PPC_ADDR16_LO
+	R_PPC64_ADDR16_HA         = R_PPC_ADDR16_HA
+	R_PPC64_REL24             = R_PPC_REL24
+	R_PPC64_GOT16_HA          = R_PPC_GOT16_HA
+	R_PPC64_JMP_SLOT          = R_PPC_JMP_SLOT
+	R_PPC64_TPREL16           = R_PPC_TPREL16
+	R_PPC64_ADDR64            = 38
+	R_PPC64_TOC16             = 47
+	R_PPC64_TOC16_LO          = 48
+	R_PPC64_TOC16_HI          = 49
+	R_PPC64_TOC16_HA          = 50
+	R_PPC64_ADDR16_LO_DS      = 57
+	R_PPC64_GOT16_LO_DS       = 59
+	R_PPC64_TOC16_DS          = 63
+	R_PPC64_TOC16_LO_DS       = 64
+	R_PPC64_TLS               = 67
+	R_PPC64_GOT_TPREL16_LO_DS = 88
+	R_PPC64_GOT_TPREL16_HA    = 90
+	R_PPC64_REL16_LO          = 250
+	R_PPC64_REL16_HI          = 251
+	R_PPC64_REL16_HA          = 252
 
 	R_SPARC_NONE     = 0
 	R_SPARC_8        = 1

src/cmd/link/internal/ppc64/asm.go

@@ -310,9 +310,18 @@ func elfreloc1(r *ld.Reloc, sectoff int64) int {
 			return -1
 		}
 
+	case obj.R_POWER_TLS:
+		ld.Thearch.Vput(ld.R_PPC64_TLS | uint64(elfsym)<<32)
+
 	case obj.R_POWER_TLS_LE:
 		ld.Thearch.Vput(ld.R_PPC64_TPREL16 | uint64(elfsym)<<32)
 
+	case obj.R_POWER_TLS_IE:
+		ld.Thearch.Vput(ld.R_PPC64_GOT_TPREL16_HA | uint64(elfsym)<<32)
+		ld.Thearch.Vput(uint64(r.Xadd))
+		ld.Thearch.Vput(uint64(sectoff + 4))
+		ld.Thearch.Vput(ld.R_PPC64_GOT_TPREL16_LO_DS | uint64(elfsym)<<32)
+
 	case obj.R_ADDRPOWER:
 		ld.Thearch.Vput(ld.R_PPC64_ADDR16_HA | uint64(elfsym)<<32)
 		ld.Thearch.Vput(uint64(r.Xadd))
@@ -444,7 +453,7 @@ func archreloc(r *ld.Reloc, s *ld.LSym, val *int64) int {
 		default:
 			return -1
 
-		case obj.R_POWER_TLS_LE:
+		case obj.R_POWER_TLS, obj.R_POWER_TLS_LE, obj.R_POWER_TLS_IE:
 			r.Done = 0
 			// check Outer is nil, Type is TLSBSS?
 			r.Xadd = r.Add

src/runtime/tls_ppc64x.s

@@ -27,8 +27,7 @@ TEXT runtime·save_g(SB),NOSPLIT|NOFRAME,$0-0
 	CMP	R31, $0
 	BEQ	nocgo
 	MOVD	runtime·tls_g(SB), R31
-	ADD	R13, R31
-	MOVD	g, 0(R31)
+	MOVD	g, 0(R13)(R31*1)
 
 nocgo:
 	RET
@@ -44,8 +43,7 @@ nocgo:
 // NOTE: _cgo_topofstack assumes this only clobbers g (R30), and R31.
 TEXT runtime·load_g(SB),NOSPLIT|NOFRAME,$0-0
 	MOVD	runtime·tls_g(SB), R31
-	ADD	R13, R31
-	MOVD	0(R31), g
+	MOVD	0(R13)(R31*1), g
 	RET
 
 GLOBL runtime·tls_g+0(SB), TLSBSS, $8