cmd/link, runtime: add initial cgo support for ppc64

We should be able to build docker after this get applied.

Updates #13192

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

@@ -374,6 +374,13 @@ func adddynrel(ctxt *ld.Link, s *sym.Symbol, r *sym.Reloc) bool {
 }
 
 func elfreloc1(ctxt *ld.Link, r *sym.Reloc, sectoff int64) bool {
+	// Beware that bit0~bit15 start from the third byte of a instruction in Big-Endian machines.
+	if r.Type == objabi.R_ADDR || r.Type == objabi.R_POWER_TLS ||  r.Type == objabi.R_CALLPOWER {
+	} else {
+		if ctxt.Arch.ByteOrder == binary.BigEndian {
+			sectoff += 2
+		}
+	}
 	ctxt.Out.Write64(uint64(sectoff))
 
 	elfsym := r.Xsym.ElfsymForReloc()

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

@@ -93,9 +93,6 @@ func archinit(ctxt *ld.Link) {
 		}
 
 	case objabi.Hlinux: /* ppc64 elf */
-		if ctxt.Arch == sys.ArchPPC64 {
-			*ld.FlagD = true // TODO(austin): ELF ABI v1 not supported yet
-		}
 		ld.Elfinit(ctxt)
 		ld.HEADR = ld.ELFRESERVE
 		if *ld.FlagTextAddr == -1 {

src/runtime/asm_ppc64x.s

@@ -36,6 +36,12 @@ TEXT runtime·rt0_go(SB),NOSPLIT,$0
 	MOVD	_cgo_init(SB), R12
 	CMP	R0, R12
 	BEQ	nocgo
+#ifdef GOARCH_ppc64
+	// ppc64 use elf ABI v1. we must get the real entry address from
+	// first slot of the function descriptor before call.
+	MOVD	8(R12), R2
+	MOVD	(R12), R12
+#endif
 	MOVD	R12, CTR		// r12 = "global function entry point"
 	MOVD	R13, R5			// arg 2: TLS base pointer
 	MOVD	$setg_gcc<>(SB), R4 	// arg 1: setg
@@ -564,6 +570,16 @@ g0:
 	MOVD	R0, 0(R1)	// clear back chain pointer (TODO can we give it real back trace information?)
 	// This is a "global call", so put the global entry point in r12
 	MOVD	R3, R12
+
+#ifdef GOARCH_ppc64
+	// ppc64 use elf ABI v1. we must get the real entry address from
+	// first slot of the function descriptor before call.
+#ifndef GOOS_aix
+	// aix just passes the function pointer for the moment, see golang.org/cl/146898 for details.
+	MOVD	8(R12), R2
+	MOVD	(R12), R12
+#endif
+#endif
 	MOVD	R12, CTR
 	MOVD	R4, R3		// arg in r3
 	BL	(CTR)
@@ -722,9 +738,20 @@ TEXT runtime·setg(SB), NOSPLIT, $0-8
 	BL	runtime·save_g(SB)
 	RET
 
+#ifdef GOARCH_ppc64
+TEXT setg_gcc<>(SB),NOSPLIT|NOFRAME,$0-0
+	DWORD	$_setg_gcc<>(SB)
+	DWORD	$0
+	DWORD	$0
+#endif
+
 // void setg_gcc(G*); set g in C TLS.
 // Must obey the gcc calling convention.
+#ifdef GOARCH_ppc64le
 TEXT setg_gcc<>(SB),NOSPLIT|NOFRAME,$0-0
+#else
+TEXT _setg_gcc<>(SB),NOSPLIT|NOFRAME,$0-0
+#endif
 	// The standard prologue clobbers R31, which is callee-save in
 	// the C ABI, so we have to use $-8-0 and save LR ourselves.
 	MOVD	LR, R4

src/runtime/rt0_linux_ppc64.s

@@ -6,6 +6,11 @@ TEXT _rt0_ppc64_linux(SB),NOSPLIT,$0
 	DWORD $0
 	DWORD $0
 
+TEXT main(SB),NOSPLIT,$0
+	DWORD $_main<>(SB)
+	DWORD $0
+	DWORD $0
+
 TEXT _main<>(SB),NOSPLIT,$-8
 	// In a statically linked binary, the stack contains argc,
 	// argv as argc string pointers followed by a NULL, envv as a
@@ -13,11 +18,13 @@ TEXT _main<>(SB),NOSPLIT,$-8
 	// There is no TLS base pointer.
 	//
 	// TODO(austin): Support ABI v1 dynamic linking entry point
-	MOVD 0(R1), R3 // argc
-	ADD $8, R1, R4 // argv
-	BR main(SB)
-
-TEXT main(SB),NOSPLIT,$-8
 	MOVD	$runtime·rt0_go(SB), R12
 	MOVD	R12, CTR
+	MOVBZ	runtime·iscgo(SB), R5
+	CMP	R5, $0
+	BEQ	nocgo
+	BR	(CTR)
+nocgo:
+	MOVD	0(R1), R3 // argc
+	ADD	$8, R1, R4 // argv
 	BR	(CTR)

src/runtime/sys_linux_ppc64x.s

@@ -414,7 +414,7 @@ TEXT runtime·cgoSigtramp(SB),NOSPLIT|NOFRAME,$0
 	DWORD	$0
 	DWORD	$0
 TEXT runtime·_cgoSigtramp(SB),NOSPLIT,$0
-	JMP	runtime·sigtramp(SB)
+	JMP	runtime·_sigtramp(SB)
 #endif
 
 TEXT runtime·sigprofNonGoWrapper<>(SB),NOSPLIT,$0