windows: use ArbitraryUserPointer as TLS slot

R=hectorchu, alex.brainman
CC=golang-dev
https://golang.org/cl/5519054

src/cmd/6l/pass.c

@@ -276,7 +276,7 @@ patch(void)
 			// Convert
 			//   op	  n(GS), reg
 			// to
-			//   MOVL 0x58(GS), reg
+			//   MOVL 0x28(GS), reg
 			//   op	  n(reg), reg
 			// The purpose of this patch is to fix some accesses
 			// to extern register variables (TLS) on Windows, as
@@ -291,7 +291,7 @@ patch(void)
 				q->as = p->as;
 				p->as = AMOVQ;
 				p->from.type = D_INDIR+D_GS;
-				p->from.offset = 0x58;
+				p->from.offset = 0x28;
 			}
 		}
 		if(HEADTYPE == Hlinux || HEADTYPE == Hfreebsd
@@ -428,11 +428,11 @@ dostkoff(void)
 			p->from.offset = tlsoffset+0;
 			p->to.type = D_CX;
 			if(HEADTYPE == Hwindows) {
-				// movq %gs:0x58, %rcx
+				// movq %gs:0x28, %rcx
 				// movq (%rcx), %rcx
 				p->as = AMOVQ;
 				p->from.type = D_INDIR+D_GS;
-				p->from.offset = 0x58;
+				p->from.offset = 0x28;
 				p->to.type = D_CX;
 
 			

src/cmd/8l/pass.c

@@ -259,7 +259,7 @@ patch(void)
 				// Convert
 				//   op	  n(GS), reg
 				// to
-				//   MOVL 0x2C(FS), reg
+				//   MOVL 0x14(FS), reg
 				//   op	  n(reg), reg
 				// The purpose of this patch is to fix some accesses
 				// to extern register variables (TLS) on Windows, as
@@ -273,7 +273,7 @@ patch(void)
 					q->as = p->as;
 					p->as = AMOVL;
 					p->from.type = D_INDIR+D_FS;
-					p->from.offset = 0x2C;
+					p->from.offset = 0x14;
 				}
 			}
 			if(HEADTYPE == Hlinux) {
@@ -424,7 +424,7 @@ dostkoff(void)
 			case Hwindows:
 				p->as = AMOVL;
 				p->from.type = D_INDIR+D_FS;
-				p->from.offset = 0x2c;
+				p->from.offset = 0x14;
 				p->to.type = D_CX;
 
 				p = appendp(p);

src/pkg/runtime/cgo/gcc_windows_386.c

@@ -45,8 +45,8 @@ threadentry(void *v)
 	 */
 	tls0 = (void*)LocalAlloc(LPTR, 32);
 	asm volatile (
-		"movl %0, %%fs:0x2c\n"	// MOVL tls0, 0x2c(FS)
-		"movl %%fs:0x2c, %%eax\n"	// MOVL 0x2c(FS), tmp
+		"movl %0, %%fs:0x14\n"	// MOVL tls0, 0x14(FS)
+		"movl %%fs:0x14, %%eax\n"	// MOVL 0x14(FS), tmp
 		"movl %1, 0(%%eax)\n"	// MOVL g, 0(FS)
 		"movl %2, 4(%%eax)\n"	// MOVL m, 4(FS)
 		:: "r"(tls0), "r"(ts.g), "r"(ts.m) : "%eax"

src/pkg/runtime/cgo/gcc_windows_amd64.c

@@ -45,8 +45,8 @@ threadentry(void *v)
 	 */
 	tls0 = (void*)LocalAlloc(LPTR, 64);
 	asm volatile (
-	  "movq %0, %%gs:0x58\n"	// MOVL tls0, 0x58(GS)
-	  "movq %%gs:0x58, %%rax\n" // MOVQ 0x58(GS), tmp
+	  "movq %0, %%gs:0x28\n"	// MOVL tls0, 0x28(GS)
+	  "movq %%gs:0x28, %%rax\n" // MOVQ 0x28(GS), tmp
 	  "movq %1, 0(%%rax)\n" // MOVQ g, 0(GS)
 	  "movq %2, 8(%%rax)\n" // MOVQ m, 8(GS)
 	  :: "r"(tls0), "r"(ts.g), "r"(ts.m) : "%rax"

src/pkg/runtime/mkasmh.sh

@@ -38,7 +38,7 @@ case "$GOARCH" in
 	#	../../libcgo/darwin_386.c:/^threadentry
 	case "$GOOS" in
 	windows)
-		echo '#define	get_tls(r)	MOVL 0x2c(FS), r'
+		echo '#define	get_tls(r)	MOVL 0x14(FS), r'
 		echo '#define	g(r)	0(r)'
 		echo '#define	m(r)	4(r)'
 		;;
@@ -82,7 +82,7 @@ case "$GOARCH" in
 amd64)
 	case "$GOOS" in
 	windows)
-		echo '#define	get_tls(r) MOVQ 0x58(GS), r'
+		echo '#define	get_tls(r) MOVQ 0x28(GS), r'
 		echo '#define	g(r) 0(r)'
 		echo '#define	m(r) 8(r)'
 		;;

src/pkg/runtime/sys_windows_386.s

@@ -98,7 +98,7 @@ TEXT runtime·externalthreadhandler(SB),7,$0
 	PUSHL	BX
 	PUSHL	SI
 	PUSHL	DI
-	PUSHL	0x2c(FS)
+	PUSHL	0x14(FS)
 	MOVL	SP, DX
 
 	// setup dummy m, g
@@ -108,7 +108,7 @@ TEXT runtime·externalthreadhandler(SB),7,$0
 	CALL	runtime·memclr(SB)	// smashes AX,BX,CX
 
 	LEAL	m_tls(SP), CX
-	MOVL	CX, 0x2c(FS)
+	MOVL	CX, 0x14(FS)
 	MOVL	SP, m(CX)
 	MOVL	SP, BX
 	SUBL	$g_end, SP		// space for G
@@ -129,7 +129,7 @@ TEXT runtime·externalthreadhandler(SB),7,$0
 	get_tls(CX)
 	MOVL	g(CX), CX
 	MOVL	g_stackbase(CX), SP
-	POPL	0x2c(FS)
+	POPL	0x14(FS)
 	POPL	DI
 	POPL	SI
 	POPL	BX
@@ -208,7 +208,7 @@ TEXT runtime·tstart(SB),7,$0
 
 	// Set up tls.
 	LEAL	m_tls(CX), SI
-	MOVL	SI, 0x2c(FS)
+	MOVL	SI, 0x14(FS)
 	MOVL	CX, m(SI)
 	MOVL	DX, g(SI)
 
@@ -246,5 +246,5 @@ TEXT runtime·tstart_stdcall(SB),7,$0
 // setldt(int entry, int address, int limit)
 TEXT runtime·setldt(SB),7,$0
 	MOVL	address+4(FP), CX
-	MOVL	CX, 0x2c(FS)
+	MOVL	CX, 0x14(FS)
 	RET

src/pkg/runtime/sys_windows_amd64.s

@@ -121,7 +121,7 @@ TEXT runtime·externalthreadhandler(SB),7,$0
 	PUSHQ	BX
 	PUSHQ	SI
 	PUSHQ	DI
-	PUSHQ	0x58(GS)
+	PUSHQ	0x28(GS)
 	MOVQ	SP, DX
 
 	// setup dummy m, g
@@ -131,7 +131,7 @@ TEXT runtime·externalthreadhandler(SB),7,$0
 	CALL	runtime·memclr(SB)	// smashes AX,BX,CX
 
 	LEAQ	m_tls(SP), CX
-	MOVQ	CX, 0x58(GS)
+	MOVQ	CX, 0x28(GS)
 	MOVQ	SP, m(CX)
 	MOVQ	SP, BX
 	SUBQ	$g_end, SP		// space for G
@@ -152,7 +152,7 @@ TEXT runtime·externalthreadhandler(SB),7,$0
 	get_tls(CX)
 	MOVQ	g(CX), CX
 	MOVQ	g_stackbase(CX), SP
-	POPQ	0x58(GS)
+	POPQ	0x28(GS)
 	POPQ	DI
 	POPQ	SI
 	POPQ	BX
@@ -254,7 +254,7 @@ TEXT runtime·tstart_stdcall(SB),7,$0
 
 	// Set up tls.
 	LEAQ	m_tls(CX), SI
-	MOVQ	SI, 0x58(GS)
+	MOVQ	SI, 0x28(GS)
 	MOVQ	CX, m(SI)
 	MOVQ	DX, g(SI)
 
@@ -276,5 +276,5 @@ TEXT runtime·notok(SB),7,$0
 // set tls base to DI
 TEXT runtime·settls(SB),7,$0
 	CALL	runtime·setstacklimits(SB)
-	MOVQ	DI, 0x58(GS)
+	MOVQ	DI, 0x28(GS)
 	RET