cgo: windows amd64 port

R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/4725041

src/cmd/6l/asm.c

@@ -524,7 +524,7 @@ adddynsym(Sym *s)
 			adduint64(d, 0);	// value
 		else
 			addaddr(d, s);
-	} else {
+	} else if(HEADTYPE != Hwindows) {
 		diag("adddynsym: unsupported binary format");
 	}
 }

src/cmd/ld/data.c

@@ -36,6 +36,7 @@
 #include	"../ld/pe.h"
 
 void	dynreloc(void);
+static vlong addaddrplus4(Sym *s, Sym *t, int32 add);
 
 /*
  * divide-and-conquer list-link
@@ -255,11 +256,19 @@ dynrelocsym(Sym *s)
 				r->add = targ->plt;
 				
 				// jmp *addr
+				if(thechar == '8') {
 					adduint8(rel, 0xff);
 					adduint8(rel, 0x25);
 					addaddr(rel, targ);
 					adduint8(rel, 0x90);
 					adduint8(rel, 0x90);
+				} else {
+					adduint8(rel, 0xff);
+					adduint8(rel, 0x24);
+					adduint8(rel, 0x25);
+					addaddrplus4(rel, targ, 0);
+					adduint8(rel, 0x90);
+				}
 			} else if(r->sym->plt >= 0) {
 				r->sym = rel;
 				r->add = targ->plt;
@@ -678,6 +687,27 @@ addaddrplus(Sym *s, Sym *t, int32 add)
 	return i;
 }
 
+vlong
+addaddrplus4(Sym *s, Sym *t, int32 add)
+{
+	vlong i;
+	Reloc *r;
+
+	if(s->type == 0)
+		s->type = SDATA;
+	s->reachable = 1;
+	i = s->size;
+	s->size += 4;
+	symgrow(s, s->size);
+	r = addrel(s);
+	r->sym = t;
+	r->off = i;
+	r->siz = 4;
+	r->type = D_ADDR;
+	r->add = add;
+	return i;
+}
+
 vlong
 addpcrelplus(Sym *s, Sym *t, int32 add)
 {

src/cmd/ld/ldpe.c

@@ -73,6 +73,24 @@
 #define IMAGE_REL_I386_SECREL7	0x000D
 #define IMAGE_REL_I386_REL32	0x0014
 
+#define IMAGE_REL_AMD64_ABSOLUTE 0x0000
+#define IMAGE_REL_AMD64_ADDR64 0x0001 // R_X86_64_64
+#define IMAGE_REL_AMD64_ADDR32 0x0002 // R_X86_64_PC32
+#define IMAGE_REL_AMD64_ADDR32NB 0x0003
+#define IMAGE_REL_AMD64_REL32 0x0004 
+#define IMAGE_REL_AMD64_REL32_1 0x0005
+#define IMAGE_REL_AMD64_REL32_2 0x0006
+#define IMAGE_REL_AMD64_REL32_3 0x0007
+#define IMAGE_REL_AMD64_REL32_4 0x0008
+#define IMAGE_REL_AMD64_REL32_5 0x0009
+#define IMAGE_REL_AMD64_SECTION 0x000A
+#define IMAGE_REL_AMD64_SECREL 0x000B
+#define IMAGE_REL_AMD64_SECREL7 0x000C
+#define IMAGE_REL_AMD64_TOKEN 0x000D
+#define IMAGE_REL_AMD64_SREL32 0x000E
+#define IMAGE_REL_AMD64_PAIR 0x000F
+#define IMAGE_REL_AMD64_SSPAN32 0x0010
+
 typedef struct PeSym PeSym;
 typedef struct PeSect PeSect;
 typedef struct PeObj PeObj;
@@ -261,6 +279,7 @@ ldpe(Biobuf *f, char *pkg, int64 len, char *pn)
 				default:
 					diag("%s: unknown relocation type %d;", pn, type);
 				case IMAGE_REL_I386_REL32:
+				case IMAGE_REL_AMD64_REL32:
 					rp->type = D_PCREL;
 					rp->add = 0;
 					break;
@@ -270,6 +289,16 @@ ldpe(Biobuf *f, char *pkg, int64 len, char *pn)
 					// load addend from image
 					rp->add = le32(rsect->base+rp->off);
 					break;
+				case IMAGE_REL_AMD64_ADDR32: // R_X86_64_PC32
+					rp->type = D_PCREL;
+					rp->add += 4;
+					break;
+				case IMAGE_REL_AMD64_ADDR64: // R_X86_64_64
+					rp->siz = 8;
+					rp->type = D_ADDR;
+					// load addend from image
+					rp->add = le64(rsect->base+rp->off);
+					break;
 			}
 		}
 		qsort(r, rsect->sh.NumberOfRelocations, sizeof r[0], rbyoff);

src/pkg/runtime/amd64/asm.s

@@ -18,7 +18,8 @@ TEXT _rt0_amd64(SB),7,$-8
 	TESTQ	AX, AX
 	JZ	needtls
 	CALL	AX
-	JMP ok
+	CMPL	runtime·iswindows(SB), $0
+	JEQ ok
 
 needtls:
 	LEAQ	runtime·tls0(SB), DI
@@ -432,6 +433,7 @@ TEXT runtime·asmcgocall(SB),7,$0
 	MOVQ	DI, 16(SP)	// save g
 	MOVQ	DX, 8(SP)	// save SP
 	MOVQ	BX, DI		// DI = first argument in AMD64 ABI
+	MOVQ	BX, CX		// CX = first argument in Win64
 	CALL	AX
 
 	// Restore registers, g, stack pointer.

src/pkg/runtime/cgo/windows_amd64.c

@@ -30,6 +30,7 @@ static void*
 threadentry(void *v)
 {
 	ThreadStart ts;
+	void *tls0;
 
 	ts = *(ThreadStart*)v;
 	free(v);
@@ -45,11 +46,13 @@ threadentry(void *v)
 	/*
 	 * Set specific keys in thread local storage.
 	 */
+	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, 0(%%rax)\n" // MOVQ g, 0(GS)
-	  "movq %1, 8(%%rax)\n" // MOVQ m, 8(GS)
-	  :: "r"(ts.g), "r"(ts.m) : "%rax"
+	  "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"
 	);
 
 	crosscall_amd64(ts.fn);

src/pkg/runtime/windows/amd64/rt0.s

@@ -8,3 +8,6 @@ TEXT	_rt0_amd64_windows(SB),7,$-8
 	MOVQ	$_rt0_amd64(SB), AX
 	MOVQ	SP, DI
 	JMP	AX
+
+DATA  runtime·iswindows(SB)/4, $1
+GLOBL runtime·iswindows(SB), $4