cmd/compile, cmd/link, runtime: on ppc64x, maintain the TOC pointer in R2 when compiling PIC

The PowerPC ISA does not have a PC-relative load instruction, which poses
obvious challenges when generating position-independent code. The way the ELFv2
ABI addresses this is to specify that r2 points to a per "module" (shared
library or executable) TOC pointer. Maintaining this pointer requires
cooperation between codegen and the system linker:

 * Non-leaf functions leave space on the stack at r1+24 to save the TOC pointer.
 * A call to a function that *might* have to go via a PLT stub must be followed
   by a nop instruction that the system linker can replace with "ld r1, 24(r1)"
   to restore the TOC pointer (only when dynamically linking Go code).
 * When calling a function via a function pointer, the address of the function
   must be in r12, and the first couple of instructions (the "global entry
   point") of the called function use this to derive the address of the TOC
   for the module it is in.
 * When calling a function that is implemented in the same module, the system
   linker adjusts the call to skip over the instructions mentioned above (the
   "local entry point"), assuming that r2 is already correctly set.

So this changeset adds the global entry point instructions, sets the metadata so
the system linker knows where the local entry point is, inserts code to save the
TOC pointer at 24(r1), adds a nop after any call not known to be local and copes
with the odd non-local code transfer in the runtime (e.g. the stuff around
jmpdefer). It does not actually compile PIC yet.

Change-Id: I7522e22bdfd2f891745a900c60254fe9e372c854
Reviewed-on: https://go-review.googlesource.com/15967Reviewed-by: Russ Cox <rsc@golang.org>

src/cmd/compile/internal/gc/cgen.go

@@ -6,6 +6,7 @@ package gc
 
 import (
 	"cmd/internal/obj"
+	"cmd/internal/obj/ppc64"
 	"fmt"
 )
 
@@ -2339,15 +2340,39 @@ func Ginscall(f *Node, proc int) {
 		-1: // normal call but no return
 		if f.Op == ONAME && f.Class == PFUNC {
 			if f == Deferreturn {
-				// Deferred calls will appear to be returning to
-				// the CALL deferreturn(SB) that we are about to emit.
-				// However, the stack trace code will show the line
-				// of the instruction byte before the return PC.
-				// To avoid that being an unrelated instruction,
-				// insert an actual hardware NOP that will have the right line number.
-				// This is different from obj.ANOP, which is a virtual no-op
-				// that doesn't make it into the instruction stream.
+				// Deferred calls will appear to be returning to the CALL
+				// deferreturn(SB) that we are about to emit. However, the
+				// stack scanning code will think that the instruction
+				// before the CALL is executing. To avoid the scanning
+				// code making bad assumptions (both cosmetic such as
+				// showing the wrong line number and fatal, such as being
+				// confused over whether a stack slot contains a pointer
+				// or a scalar) insert an actual hardware NOP that will
+				// have the right line number. This is different from
+				// obj.ANOP, which is a virtual no-op that doesn't make it
+				// into the instruction stream.
 				Thearch.Ginsnop()
+
+				if Thearch.Thechar == '9' {
+					// On ppc64, when compiling Go into position
+					// independent code on ppc64le we insert an
+					// instruction to reload the TOC pointer from the
+					// stack as well. See the long comment near
+					// jmpdefer in runtime/asm_ppc64.s for why.
+					// If the MOVD is not needed, insert a hardware NOP
+					// so that the same number of instructions are used
+					// on ppc64 in both shared and non-shared modes.
+					if Ctxt.Flag_shared != 0 {
+						p := Thearch.Gins(ppc64.AMOVD, nil, nil)
+						p.From.Type = obj.TYPE_MEM
+						p.From.Offset = 24
+						p.From.Reg = ppc64.REGSP
+						p.To.Type = obj.TYPE_REG
+						p.To.Reg = ppc64.REG_R2
+					} else {
+						Thearch.Ginsnop()
+					}
+				}
 			}
 
 			p := Thearch.Gins(obj.ACALL, nil, f)

src/cmd/compile/internal/gc/plive.go

@@ -1426,7 +1426,14 @@ func livenessepilogue(lv *Liveness) {
 						// the PCDATA must begin one instruction early too.
 						// The instruction before a call to deferreturn is always a
 						// no-op, to keep PC-specific data unambiguous.
-						splicebefore(lv, bb, newpcdataprog(p.Opt.(*obj.Prog), pos), p.Opt.(*obj.Prog))
+						prev := p.Opt.(*obj.Prog)
+						if Ctxt.Arch.Thechar == '9' {
+							// On ppc64 there is an additional instruction
+							// (another no-op or reload of toc pointer) before
+							// the call.
+							prev = prev.Opt.(*obj.Prog)
+						}
+						splicebefore(lv, bb, newpcdataprog(prev, pos), prev)
 					} else {
 						splicebefore(lv, bb, newpcdataprog(p, pos), p)
 					}

src/cmd/compile/internal/ppc64/gsubr.go

@@ -580,9 +580,9 @@ func rawgins(as int, f *gc.Node, t *gc.Node) *obj.Prog {
 	case obj.ACALL:
 		if p.To.Type == obj.TYPE_REG && p.To.Reg != ppc64.REG_CTR {
 			// Allow front end to emit CALL REG, and rewrite into MOV REG, CTR; CALL CTR.
-			if gc.Ctxt.Flag_dynlink {
+			if gc.Ctxt.Flag_shared != 0 {
 				// Make sure function pointer is in R12 as well when
-				// dynamically linking Go.
+				// compiling Go into PIC.
 				// TODO(mwhudson): it would obviously be better to
 				// change the register allocation to put the value in
 				// R12 already, but I don't know how to do that.
@@ -602,6 +602,19 @@ func rawgins(as int, f *gc.Node, t *gc.Node) *obj.Prog {
 			p.To.Type = obj.TYPE_REG
 			p.To.Reg = ppc64.REG_CTR
 
+			if gc.Ctxt.Flag_shared != 0 {
+				// When compiling Go into PIC, the function we just
+				// called via pointer might have been implemented in
+				// a separate module and so overwritten the TOC
+				// pointer in R2; reload it.
+				q := gc.Prog(ppc64.AMOVD)
+				q.From.Type = obj.TYPE_MEM
+				q.From.Offset = 24
+				q.From.Reg = ppc64.REGSP
+				q.To.Type = obj.TYPE_REG
+				q.To.Reg = ppc64.REG_R2
+			}
+
 			if gc.Debug['g'] != 0 {
 				fmt.Printf("%v\n", p)
 				fmt.Printf("%v\n", pp)

src/cmd/compile/internal/ppc64/reg.go

@@ -113,8 +113,8 @@ func excludedregs() uint64 {
 	// Exclude registers with fixed functions
 	regbits := uint64(1<<0 | RtoB(ppc64.REGSP) | RtoB(ppc64.REGG) | RtoB(ppc64.REGTLS) | RtoB(ppc64.REGTMP))
 
-	if gc.Ctxt.Flag_dynlink {
-		// When dynamically linking Go, R2 is reserved to be the TOC pointer
+	if gc.Ctxt.Flag_shared != 0 {
+		// When compiling Go into PIC, R2 is reserved to be the TOC pointer
 		// and R12 so that calls via function pointer can stomp on it.
 		regbits |= RtoB(ppc64.REG_R2)
 		regbits |= RtoB(ppc64.REG_R12)

src/cmd/internal/obj/link.go

@@ -486,6 +486,11 @@ const (
 	// relocated symbol are inserted into this field; it is an error if the last two
 	// bits of the address are not 0.
 	R_ADDRPOWER_DS
+
+	// R_ADDRPOWER_PCREL relocates two D-form instructions like R_ADDRPOWER, but
+	// inserts the displacement from the place being relocated to the address of the
+	// the relocated symbol instead of just its address.
+	R_ADDRPOWER_PCREL
 )
 
 type Auto struct {

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

@@ -205,6 +205,7 @@ const (
 	C_DACON /* $n(REG) where int32 < n */
 	C_SBRA
 	C_LBRA
+	C_LBRAPIC
 	C_SAUTO
 	C_LAUTO
 	C_SEXT

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

@@ -281,6 +281,7 @@ var optab = []Optab{
 	{ABEQ, C_NONE, C_NONE, C_NONE, C_SBRA, 16, 4, 0},
 	{ABEQ, C_CREG, C_NONE, C_NONE, C_SBRA, 16, 4, 0},
 	{ABR, C_NONE, C_NONE, C_NONE, C_LBRA, 11, 4, 0},
+	{ABR, C_NONE, C_NONE, C_NONE, C_LBRAPIC, 11, 8, 0},
 	{ABC, C_SCON, C_REG, C_NONE, C_SBRA, 16, 4, 0},
 	{ABC, C_SCON, C_REG, C_NONE, C_LBRA, 17, 4, 0},
 	{ABR, C_NONE, C_NONE, C_NONE, C_LR, 18, 4, 0},
@@ -704,6 +705,9 @@ func aclass(ctxt *obj.Link, a *obj.Addr) int {
 		return C_DCON
 
 	case obj.TYPE_BRANCH:
+		if a.Sym != nil && ctxt.Flag_dynlink {
+			return C_LBRAPIC
+		}
 		return C_SBRA
 	}
 
@@ -1714,6 +1718,7 @@ func asmout(ctxt *obj.Link, p *obj.Prog, o *Optab, out []uint32) {
 			rel.Add = int64(v)
 			rel.Type = obj.R_CALLPOWER
 		}
+		o2 = 0x60000000 // nop, sometimes overwritten by ld r2, 24(r1) when dynamic linking
 
 	case 12: /* movb r,r (extsb); movw r,r (extsw) */
 		if p.To.Reg == REGZERO && p.From.Type == obj.TYPE_CONST {

src/cmd/internal/obj/ppc64/obj9.go

@@ -342,11 +342,45 @@ func preprocess(ctxt *obj.Link, cursym *obj.LSym) {
 
 			p.To.Offset = int64(autosize)
 
-			if p.From3.Offset&obj.NOSPLIT == 0 {
-				p = stacksplit(ctxt, p, autosize) // emit split check
+			q = p
+
+			if ctxt.Flag_shared != 0 && cursym.Name != "runtime.duffzero" && cursym.Name != "runtime.duffcopy" {
+				// When compiling Go into PIC, all functions must start
+				// with instructions to load the TOC pointer into r2:
+				//
+				//	addis r2, r12, .TOC.-func@ha
+				//	addi r2, r2, .TOC.-func@l+4
+				//
+				// We could probably skip this prologue in some situations
+				// but it's a bit subtle. However, it is both safe and
+				// necessary to leave the prologue off duffzero and
+				// duffcopy as we rely on being able to jump to a specific
+				// instruction offset for them.
+				//
+				// These are AWORDS because there is no (afaict) way to
+				// generate the addis instruction except as part of the
+				// load of a large constant, and in that case there is no
+				// way to use r12 as the source.
+				q = obj.Appendp(ctxt, q)
+				q.As = AWORD
+				q.Lineno = p.Lineno
+				q.From.Type = obj.TYPE_CONST
+				q.From.Offset = 0x3c4c0000
+				q = obj.Appendp(ctxt, q)
+				q.As = AWORD
+				q.Lineno = p.Lineno
+				q.From.Type = obj.TYPE_CONST
+				q.From.Offset = 0x38420000
+				rel := obj.Addrel(ctxt.Cursym)
+				rel.Off = 0
+				rel.Siz = 8
+				rel.Sym = obj.Linklookup(ctxt, ".TOC.", 0)
+				rel.Type = obj.R_ADDRPOWER_PCREL
 			}
 
-			q = p
+			if cursym.Text.From3.Offset&obj.NOSPLIT == 0 {
+				q = stacksplit(ctxt, q, autosize) // emit split check
+			}
 
 			if autosize != 0 {
 				/* use MOVDU to adjust R1 when saving R31, if autosize is small */
@@ -354,7 +388,7 @@ func preprocess(ctxt *obj.Link, cursym *obj.LSym) {
 					mov = AMOVDU
 					aoffset = int(-autosize)
 				} else {
-					q = obj.Appendp(ctxt, p)
+					q = obj.Appendp(ctxt, q)
 					q.As = AADD
 					q.Lineno = p.Lineno
 					q.From.Type = obj.TYPE_CONST
@@ -395,6 +429,17 @@ func preprocess(ctxt *obj.Link, cursym *obj.LSym) {
 				q.Spadj = int32(-aoffset)
 			}
 
+			if ctxt.Flag_shared != 0 {
+				q = obj.Appendp(ctxt, q)
+				q.As = AMOVD
+				q.Lineno = p.Lineno
+				q.From.Type = obj.TYPE_REG
+				q.From.Reg = REG_R2
+				q.To.Type = obj.TYPE_MEM
+				q.To.Reg = REGSP
+				q.To.Offset = 24
+			}
+
 			if cursym.Text.From3.Offset&obj.WRAPPER != 0 {
 				// if(g->panic != nil && g->panic->argp == FP) g->panic->argp = bottom-of-frame
 				//

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

@@ -356,7 +356,9 @@ func relocsym(s *LSym) {
 		// We need to be able to reference dynimport symbols when linking against
 		// shared libraries, and Solaris needs it always
 		if HEADTYPE != obj.Hsolaris && r.Sym != nil && r.Sym.Type == obj.SDYNIMPORT && !DynlinkingGo() {
-			Diag("unhandled relocation for %s (type %d rtype %d)", r.Sym.Name, r.Sym.Type, r.Type)
+			if !(Thearch.Thechar == '9' && Linkmode == LinkExternal && r.Sym.Name == ".TOC.") {
+				Diag("unhandled relocation for %s (type %d rtype %d)", r.Sym.Name, r.Sym.Type, r.Type)
+			}
 		}
 		if r.Sym != nil && r.Sym.Type != obj.STLSBSS && !r.Sym.Reachable {
 			Diag("unreachable sym in relocation: %s %s", s.Name, r.Sym.Name)

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

@@ -1838,7 +1838,7 @@ func genasmsym(put func(*LSym, string, int, int64, int64, int, *LSym)) {
 	}
 
 	for s := Ctxt.Allsym; s != nil; s = s.Allsym {
-		if s.Hide != 0 || ((s.Name == "" || s.Name[0] == '.') && s.Version == 0 && s.Name != ".rathole") {
+		if s.Hide != 0 || ((s.Name == "" || s.Name[0] == '.') && s.Version == 0 && s.Name != ".rathole" && s.Name != ".TOC.") {
 			continue
 		}
 		switch s.Type & obj.SMASK {

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

@@ -162,6 +162,12 @@ func putelfsym(x *LSym, s string, t int, addr int64, size int64, ver int, go_ *L
 	if x.Type&obj.SHIDDEN != 0 {
 		other = STV_HIDDEN
 	}
+	if Buildmode == BuildmodePIE && Thearch.Thechar == '9' && type_ == STT_FUNC && x.Name != "runtime.duffzero" && x.Name != "runtime.duffcopy" {
+		// On ppc64 the top three bits of the st_other field indicate how
+		// many instructions separate the global and local entry points. In
+		// our case it is two instructions, indicated by the value 3.
+		other |= 3 << 5
+	}
 
 	if DynlinkingGo() && bind == STB_GLOBAL && elfbind == STB_LOCAL && x.Type == obj.STEXT {
 		// When dynamically linking, we want references to functions defined

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

@@ -325,6 +325,13 @@ func elfreloc1(r *ld.Reloc, sectoff int64) int {
 		ld.Thearch.Vput(uint64(sectoff + 4))
 		ld.Thearch.Vput(ld.R_PPC64_ADDR16_LO_DS | uint64(elfsym)<<32)
 
+	case obj.R_ADDRPOWER_PCREL:
+		ld.Thearch.Vput(ld.R_PPC64_REL16_HA | uint64(elfsym)<<32)
+		ld.Thearch.Vput(uint64(r.Xadd))
+		ld.Thearch.Vput(uint64(sectoff + 4))
+		ld.Thearch.Vput(ld.R_PPC64_REL16_LO | uint64(elfsym)<<32)
+		r.Xadd += 4
+
 	case obj.R_CALLPOWER:
 		if r.Siz != 4 {
 			return -1
@@ -433,7 +440,8 @@ func archreloc(r *ld.Reloc, s *ld.LSym, val *int64) int {
 			return 0
 
 		case obj.R_ADDRPOWER,
-			obj.R_ADDRPOWER_DS:
+			obj.R_ADDRPOWER_DS,
+			obj.R_ADDRPOWER_PCREL:
 			r.Done = 0
 
 			// set up addend for eventual relocation via outer symbol.

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

@@ -103,6 +103,11 @@ func archinit() {
 		ld.Linkmode = ld.LinkExternal
 	}
 
+	if ld.Linkmode == ld.LinkExternal {
+		toc := ld.Linklookup(ld.Ctxt, ".TOC.", 0)
+		toc.Type = obj.SDYNIMPORT
+	}
+
 	switch ld.HEADTYPE {
 	default:
 		if ld.Linkmode == ld.LinkAuto {

src/runtime/asm_ppc64x.s

@@ -17,8 +17,9 @@ TEXT runtime·rt0_go(SB),NOSPLIT,$0
 	BL	runtime·reginit(SB)
 
 	SUB	$(FIXED_FRAME+16), R1
-	MOVW	R3, FIXED_FRAME+0(R1) // argc
-	MOVD	R4, FIXED_FRAME+8(R1) // argv
+	MOVD	R2, 24(R1)		// stash the TOC pointer away again now we've created a new frame
+	MOVW	R3, FIXED_FRAME+0(R1)	// argc
+	MOVD	R4, FIXED_FRAME+8(R1)	// argv
 
 	// create istack out of the given (operating system) stack.
 	// _cgo_init may update stackguard.
@@ -45,6 +46,7 @@ TEXT runtime·rt0_go(SB),NOSPLIT,$0
 	RLDCR	$0, R1, $~15, R1	// 16-byte align
 	BL	(CTR)			// may clobber R0, R3-R12
 	MOVD	R14, R1			// restore stack
+	MOVD	24(R1), R2
 	XOR	R0, R0			// fix R0
 
 nocgo:
@@ -181,6 +183,7 @@ TEXT runtime·mcall(SB), NOSPLIT|NOFRAME, $0-8
 	MOVDU	R0, -8(R1)
 	MOVDU	R0, -8(R1)
 	BL	(CTR)
+	MOVD	24(R1), R2
 	BR	runtime·badmcall2(SB)
 
 // systemstack_switch is a dummy routine that systemstack leaves at the bottom
@@ -189,6 +192,11 @@ TEXT runtime·mcall(SB), NOSPLIT|NOFRAME, $0-8
 // at the top of the system stack because the one at the top of
 // the system stack terminates the stack walk (see topofstack()).
 TEXT runtime·systemstack_switch(SB), NOSPLIT, $0-0
+	// We have several undefs here so that 16 bytes past
+	// $runtime·systemstack_switch lies within them whether or not the
+        // instructions that derive r2 from r12 are there.
+	UNDEF
+	UNDEF
 	UNDEF
 	BL	(LR)	// make sure this function is not leaf
 	RET
@@ -221,7 +229,7 @@ switch:
 	// save our state in g->sched.  Pretend to
 	// be systemstack_switch if the G stack is scanned.
 	MOVD	$runtime·systemstack_switch(SB), R6
-	ADD	$8, R6	// get past prologue
+	ADD     $16, R6 // get past prologue (including r2-setting instructions when they're there)
 	MOVD	R6, (g_sched+gobuf_pc)(g)
 	MOVD	R1, (g_sched+gobuf_sp)(g)
 	MOVD	R0, (g_sched+gobuf_lr)(g)
@@ -242,6 +250,13 @@ switch:
 	MOVD	R12, CTR
 	BL	(CTR)
 
+	// restore TOC pointer. It seems unlikely that we will use systemstack
+	// to call a function defined in another module, but the results of
+	// doing so would be so confusing that it's worth doing this.
+	MOVD	g_m(g), R3
+	MOVD	m_curg(R3), g
+	MOVD	(g_sched+gobuf_sp)(g), R3
+	MOVD	24(R3), R2
 	// switch back to g
 	MOVD	g_m(g), R3
 	MOVD	m_curg(R3), g
@@ -255,6 +270,7 @@ noswitch:
 	MOVD	0(R11), R12	// code pointer
 	MOVD	R12, CTR
 	BL	(CTR)
+	MOVD	24(R1), R2
 	RET
 
 /*
@@ -402,6 +418,7 @@ TEXT NAME(SB), WRAPPER, $MAXSIZE-24;		\
 	MOVD	R12, CTR;			\
 	PCDATA  $PCDATA_StackMapIndex, $0;	\
 	BL	(CTR);				\
+	MOVD	24(R1), R2;			\
 	/* copy return values back */		\
 	MOVD	arg+16(FP), R3;			\
 	MOVWZ	n+24(FP), R4;			\
@@ -465,11 +482,15 @@ TEXT runtime·procyield(SB),NOSPLIT,$0-0
 // void jmpdefer(fv, sp);
 // called from deferreturn.
 // 1. grab stored LR for caller
-// 2. sub 4 bytes to get back to BL deferreturn
+// 2. sub 8 bytes to get back to either nop or toc reload before deferreturn
 // 3. BR to fn
+// When dynamically linking Go, it is not sufficient to rewind to the BL
+// deferreturn -- we might be jumping between modules and so we need to reset
+// the TOC pointer in r2. To do this, codegen inserts MOVD 24(R1), R2 *before*
+// the BL deferreturn and jmpdefer rewinds to that.
 TEXT runtime·jmpdefer(SB), NOSPLIT|NOFRAME, $0-16
 	MOVD	0(R1), R31
-	SUB	$4, R31
+	SUB     $8, R31
 	MOVD	R31, LR
 
 	MOVD	fv+0(FP), R11
@@ -532,8 +553,13 @@ g0:
 	// C code can clobber R0, so set it back to 0.  F27-F31 are
 	// callee save, so we don't need to recover those.
 	XOR	R0, R0
-	// Restore g, stack pointer.  R3 is errno, so don't touch it
+	// Restore g, stack pointer, toc pointer.
+	// R3 is errno, so don't touch it
 	MOVD	40(R1), g
+	MOVD    (g_stack+stack_hi)(g), R5
+	MOVD    32(R1), R6
+	SUB     R6, R5
+	MOVD    24(R5), R2
 	BL	runtime·save_g(SB)
 	MOVD	(g_stack+stack_hi)(g), R5
 	MOVD	32(R1), R6
@@ -1007,8 +1033,16 @@ TEXT _cgo_topofstack(SB),NOSPLIT|NOFRAME,$0
 
 // The top-most function running on a goroutine
 // returns to goexit+PCQuantum.
+//
+// When dynamically linking Go, it can be returned to from a function
+// implemented in a different module and so needs to reload the TOC pointer
+// from the stack (although this function declares that it does not set up x-a
+// frame, newproc1 does in fact allocate one for goexit and saves the TOC
+// pointer in the correct place).
+// goexit+_PCQuantum is halfway through the usual global entry point prologue
+// that derives r2 from r12 which is a bit silly, but not harmful.
 TEXT runtime·goexit(SB),NOSPLIT|NOFRAME,$0-0
-	MOVD	R0, R0	// NOP
+	MOVD	24(R1), R2
 	BL	runtime·goexit1(SB)	// does not return
 	// traceback from goexit1 must hit code range of goexit
 	MOVD	R0, R0	// NOP
@@ -1027,3 +1061,12 @@ TEXT runtime·prefetchnta(SB),NOSPLIT,$0-8
 
 TEXT runtime·sigreturn(SB),NOSPLIT,$0-8
         RET
+
+// prepGoExitFrame saves the current TOC pointer (i.e. the TOC pointer for the
+// module containing runtime) to the frame that goexit will execute in when
+// the goroutine exits. It's implemented in assembly mainly because that's the
+// easiest way to get access to R2.
+TEXT runtime·prepGoExitFrame(SB),NOSPLIT,$0-8
+      MOVD    sp+0(FP), R3
+      MOVD    R2, 24(R3)
+      RET

src/runtime/cgo/gcc_ppc64x.S

@@ -31,6 +31,7 @@ EXT(crosscall_ppc64):
 
 	// Set up Go ABI constant registers
 	bl	_cgo_reginit
+	nop
 
 	// Restore g pointer (r30 in Go ABI, which may have been clobbered by C)
 	mr	%r30, %r4

src/runtime/proc.go

@@ -2588,6 +2588,7 @@ func newproc1(fn *funcval, argp *uint8, narg int32, nret int32, callerpc uintptr
 	if usesLR {
 		// caller's LR
 		*(*unsafe.Pointer)(unsafe.Pointer(sp)) = nil
+		prepGoExitFrame(sp)
 		spArg += sys.MinFrameSize
 	}
 	memmove(unsafe.Pointer(spArg), unsafe.Pointer(argp), uintptr(narg))

src/runtime/sys_linux_ppc64x.s

@@ -209,6 +209,7 @@ TEXT runtime·sigfwd(SB),NOSPLIT,$0-32
 	MOVD	fn+0(FP), R12
 	MOVD	R12, CTR
 	BL	(CTR)
+	MOVD	24(R1), R2
 	RET
 
 #ifdef GOARCH_ppc64le
@@ -238,6 +239,7 @@ TEXT runtime·_sigtramp(SB),NOSPLIT,$64
 	MOVD	$runtime·sigtrampgo(SB), R12
 	MOVD	R12, CTR
 	BL	(CTR)
+	MOVD	24(R1), R2
 	RET
 
 TEXT runtime·mmap(SB),NOSPLIT|NOFRAME,$0

src/runtime/sys_nonppc64x.go

+// Copyright 2015 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build !ppc64,!ppc64le
+
+package runtime
+
+func prepGoExitFrame(sp uintptr) {
+}

src/runtime/sys_ppc64x.go

@@ -35,3 +35,5 @@ func rewindmorestack(buf *gobuf) {
 	print("runtime: pc=", hex(buf.pc), " ", hex(inst), "\n")
 	throw("runtime: misuse of rewindmorestack")
 }
+
+func prepGoExitFrame(sp uintptr)