cmd/internal/ld, runtime: halve tlsoffset on ELF/intel

For OSes that use elf on intel, 2*Ptrsize bytes are reserved for TLS.
But only one pointer (g) has been stored in the TLS for a while now.
So we can set it to just Ptrsize, which happily matches what happens
when externally linking.

Fixes #9913

Change-Id: Ic816369d3a55a8cdcc23be349b1a1791d53f5f81
Reviewed-on: https://go-review.googlesource.com/6584
Run-TryBot: Minux Ma <minux@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>

src/cmd/internal/ld/sym.go

@@ -107,7 +107,7 @@ func linknew(arch *LinkArch) *Link {
 		Hopenbsd,
 		Hdragonfly,
 		Hsolaris:
-		ctxt.Tlsoffset = -2 * ctxt.Arch.Ptrsize
+		ctxt.Tlsoffset = -1 * ctxt.Arch.Ptrsize
 
 	case Hnacl:
 		switch ctxt.Arch.Thechar {

src/cmd/internal/obj/sym.go

@@ -170,7 +170,7 @@ func Linknew(arch *LinkArch) *Link {
 		Hopenbsd,
 		Hdragonfly,
 		Hsolaris:
-		ctxt.Tlsoffset = -2 * ctxt.Arch.Ptrsize
+		ctxt.Tlsoffset = -1 * ctxt.Arch.Ptrsize
 
 	case Hnacl:
 		switch ctxt.Arch.Thechar {

src/runtime/sys_dragonfly_386.s

@@ -295,9 +295,9 @@ TEXT runtime·setldt(SB),NOSPLIT,$4
 	RET
 
 TEXT runtime·settls(SB),NOSPLIT,$24
-	// adjust for ELF: wants to use -8(GS) and -4(GS) for g and m
+	// adjust for ELF: wants to use -4(GS) for g
 	MOVL	tlsbase+0(FP), CX
-	ADDL	$8, CX
+	ADDL	$4, CX
 
 	// Set up a struct tls_info - a size of -1 maps the whole address
 	// space and is required for direct-tls access of variable data

src/runtime/sys_dragonfly_amd64.s

@@ -272,7 +272,7 @@ TEXT runtime·usleep(SB),NOSPLIT,$16
 
 // set tls base to DI
 TEXT runtime·settls(SB),NOSPLIT,$16
-	ADDQ	$16, DI	// adjust for ELF: wants to use -16(FS) and -8(FS) for g and m
+	ADDQ	$8, DI	// adjust for ELF: wants to use -8(FS) for g
 	MOVQ	DI, 0(SP)
 	MOVQ	$16, 8(SP)
 	MOVQ	$0, DI			// arg 1 - which

src/runtime/sys_freebsd_386.s

@@ -287,7 +287,7 @@ int i386_set_ldt(int, const union ldt_entry *, int);
 TEXT runtime·setldt(SB),NOSPLIT,$32
 	MOVL	address+4(FP), BX	// aka base
 	// see comment in sys_linux_386.s; freebsd is similar
-	ADDL	$0x8, BX
+	ADDL	$0x4, BX
 
 	// set up data_desc
 	LEAL	16(SP), AX	// struct data_desc

src/runtime/sys_freebsd_amd64.s

@@ -262,7 +262,7 @@ TEXT runtime·usleep(SB),NOSPLIT,$16
 
 // set tls base to DI
 TEXT runtime·settls(SB),NOSPLIT,$8
-	ADDQ	$16, DI	// adjust for ELF: wants to use -16(FS) and -8(FS) for g and m
+	ADDQ	$8, DI	// adjust for ELF: wants to use -8(FS) for g and m
 	MOVQ	DI, 0(SP)
 	MOVQ	SP, SI
 	MOVQ	$129, DI	// AMD64_SET_FSBASE

src/runtime/sys_linux_386.s

@@ -410,16 +410,16 @@ TEXT runtime·setldt(SB),NOSPLIT,$32
 	 * When linking against the system libraries,
 	 * we use its pthread_create and let it set up %gs
 	 * for us.  When we do that, the private storage
-	 * we get is not at 0(GS), 4(GS), but -8(GS), -4(GS).
+	 * we get is not at 0(GS), but -4(GS).
 	 * To insulate the rest of the tool chain from this
-	 * ugliness, 8l rewrites 0(TLS) into -8(GS) for us.
+	 * ugliness, 8l rewrites 0(TLS) into -4(GS) for us.
 	 * To accommodate that rewrite, we translate
 	 * the address here and bump the limit to 0xffffffff (no limit)
-	 * so that -8(GS) maps to 0(address).
-	 * Also, the final 0(GS) (current 8(CX)) has to point
+	 * so that -4(GS) maps to 0(address).
+	 * Also, the final 0(GS) (current 4(CX)) has to point
 	 * to itself, to mimic ELF.
 	 */
-	ADDL	$0x8, CX	// address
+	ADDL	$0x4, CX	// address
 	MOVL	CX, 0(CX)
 
 	// set up user_desc

src/runtime/sys_linux_amd64.s

@@ -359,7 +359,7 @@ TEXT runtime·sigaltstack(SB),NOSPLIT,$-8
 
 // set tls base to DI
 TEXT runtime·settls(SB),NOSPLIT,$32
-	ADDQ	$16, DI	// ELF wants to use -16(FS), -8(FS)
+	ADDQ	$8, DI	// ELF wants to use -8(FS)
 
 	MOVQ	DI, SI
 	MOVQ	$0x1002, DI	// ARCH_SET_FS

src/runtime/sys_netbsd_386.s

@@ -307,9 +307,9 @@ TEXT runtime·setldt(SB),NOSPLIT,$8
 	RET
 
 TEXT runtime·settls(SB),NOSPLIT,$16
-	// adjust for ELF: wants to use -8(GS) and -4(GS) for g and m
+	// adjust for ELF: wants to use -4(GS) for g
 	MOVL	base+0(FP), CX
-	ADDL	$8, CX
+	ADDL	$4, CX
 	MOVL	$0, 0(SP)		// syscall gap
 	MOVL	CX, 4(SP)		// arg 1 - ptr
 	MOVL	$317, AX		// sys__lwp_setprivate

src/runtime/sys_netbsd_amd64.s

@@ -307,8 +307,8 @@ TEXT runtime·sigaltstack(SB),NOSPLIT,$-8
 
 // set tls base to DI
 TEXT runtime·settls(SB),NOSPLIT,$8
-	// adjust for ELF: wants to use -16(FS) and -8(FS) for g and m
-	ADDQ	$16, DI			// arg 1 - ptr
+	// adjust for ELF: wants to use -8(FS) for g
+	ADDQ	$8, DI			// arg 1 - ptr
 	MOVQ	$317, AX		// sys__lwp_setprivate
 	SYSCALL
 	JCC	2(PC)

src/runtime/sys_openbsd_386.s

@@ -325,9 +325,9 @@ TEXT runtime·setldt(SB),NOSPLIT,$4
 	RET
 
 TEXT runtime·settls(SB),NOSPLIT,$8
-	// adjust for ELF: wants to use -8(GS) and -4(GS) for g and m
+	// adjust for ELF: wants to use -4(GS) for g
 	MOVL	tlsbase+0(FP), CX
-	ADDL	$8, CX
+	ADDL	$4, CX
 	MOVL	$0, 0(SP)		// syscall gap
 	MOVL	CX, 4(SP)		// arg 1 - tcb
 	MOVL	$329, AX		// sys___set_tcb

src/runtime/sys_openbsd_amd64.s

@@ -298,7 +298,7 @@ TEXT runtime·sigaltstack(SB),NOSPLIT,$-8
 // set tls base to DI
 TEXT runtime·settls(SB),NOSPLIT,$0
 	// adjust for ELF: wants to use -16(FS) and -8(FS) for g and m
-	ADDQ	$16, DI
+	ADDQ	$8, DI
 	MOVQ	$329, AX		// sys___settcb
 	SYSCALL
 	JCC	2(PC)